diff --git "a/reports.jsonl" "b/reports.jsonl"
new file mode 100644--- /dev/null
+++ "b/reports.jsonl"
@@ -0,0 +1,76 @@
+{"source": "reports", "source_filetype": "pdf", "abstract": "an informal survey was circulated among participants, asking them to make their best guess at the chance that there will be disasters of different types before 2100. This report summarizes the main results. The median extinction risk estimates were: \n Risk At least 1 million dead \n At least 1 billion dead \n Human extinction Number killed by molecular nanotech weapons.", "authors": ["Anders Sandberg", "Nick Bostrom"], "title": "GLOBAL CATASTROPHIC RISKS SURVEY", "text": "25% 10% 5% Total killed by superintelligent AI. \n 10% 5% 5% Total killed in all wars (including civil wars). \n 98% 30% 4% Number killed in the single biggest engineered pandemic. \n 30% 10% 2% Total killed in all nuclear wars. These results should be taken with a grain of salt. Non-responses have been omitted, although some might represent a statement of zero probability rather than no opinion. \n 30% There are likely to be many cognitive biases that affect the result, such as unpacking bias and the availability heuristic--well as old-fashioned optimism and pessimism. In appendix A the results are plotted with individual response distributions visible. \n Other Risks The list of risks was not intended to be inclusive of all the biggest risks. Respondents were invited to contribute their own global catastrophic risks, showing risks they considered significant. Several suggested totalitarian world government, climate-induced disasters, ecological/resource crunches and \"other risks\"--specified or unknowable threats. Other suggestions were asteroid/comet impacts, bad crisis management, hightech asymmetric war attacking brittle IT-based societies, back-contamination from space probes, electromagnetic pulses, genocide/democides, risks from physics research and degradation of quality assurance. \n Suggestions Respondents were also asked to suggest what they would recommend to policymakers. Several argued for nuclear disarmament, or at least lowering the number of weapons under the threshold for existential catastrophe, as well as reducing stocks of highly enriched uranium and making nuclear arsenals harder to accidentally launch. One option discussed was formation of global biotech-related governance, legislation and enforcement, or even a global body like the IPCC or UNFCCC to study and act on catastrophic risk. At the very least there was much interest in developing defenses against misuses of biotechnology, and a recognition for the need of unbiased early detection systems for a variety of risks, be they near Earth objects or actors with WMD capabilities. Views on emerging technologies such as nanotech, AI, and cognition enhancement were mixed: some proposed avoiding funding them; others deliberate crash programs to ensure they would be in the right hands, the risks understood, and the technologies able to be used against other catastrophic risks. Other suggestions included raising awareness of the problem, more research on cyber security issues, the need to build societal resiliency in depth, prepare for categories of disasters rather than individual types, building refuges and change energy consumption patterns. \n Appendix A Below are the individual results, shown as grey dots (jittered for distinguishability) and with the median as a bar.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/2008-1.tei.xml", "id": "80b29863675604ca65e2bd6ac3466467"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/1602.04019.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Robot Hacking Games ( , RHG) are government-backed competitions that China uses to advance automatic software vulnerability discovery, patching, and exploitation technologies. 1 These tools offer both offensive and defensive capabilities that promise to increase the scale and pace of vulnerability discovery. If successful, countries could use these tools to find software vulnerabilities quicker than their adversaries. A fully developed capability would allow defenders to patch vulnerabilities as quickly as they are found; attackers could build new exploits equally fast. The Defense Advanced Research Project Agency's Cyber Grand Challenge in 2016 spurred China's interest in this area. The DARPA effort resulted in the creation of state-ofthe-art tools in each of these areas, which have since been siloed into separate programs. China, by contrast, has hosted at least seven competitions since 2017. China's competition structure embodies its military-civil fusion strategy, attracting a collection of academic, military, and privatesector teams. Just two years after the People's Liberation Army's National University of Defense Technology won the first competition in 2017, the military started managing competitions of its own. 2 By 2021, a laboratory run by the PLA Equipment Development Department hosted its first RHG competition. 3 These management and oversight roles situate the PLA in an ideal position to evaluate and attract the best tools and talent. Other state hacking teams, like those of the Ministry of State Security (MSS), will benefit from the technology's development, too. Leading Chinese cybersecurity experts and government strategy documents tie automated software vulnerability discovery, patching, and exploitation tools to Chinese President Xi Jinping's goal for China to become a \"cyber powerhouse\" ( ). 4 These policy documents create a de facto political mandate for China's cybersecurity community to develop the desired tools. Although they will not make China a \"cyber powerhouse\" on their own, their development illustrates one important capability that China has chosen in pursuit of its goal.", "authors": ["Dakota Cary"], "title": "Robot Hacking Games China's Competitions to Automate the Software Vulnerability Lifecycle CSET Issue Brief", "text": "Introduction A collection of seven server racks, each with their team's color and logo splashed on the covers of the bulky boxes, stood on stage in a Las Vegas conference room in August 2016.  5  Professional commentators narrated as each team's Cyber Reasoning System (CRS) hacked away on their own code, and their competitors' servers, trying to find vulnerabilities. Over the course of the competition, these machines earned points by patching their own vulnerabilities while maintaining system performance and submitting successful attacks against opposing teams' servers.  6  But the event lacked the chatter of fingers furiously striking keys that normally accompanies hacking competitions. A few feet away from the flowing bits and bytes, a collection of PhDs, researchers, and private-sector innovators who created the CRSs watched the scoreboard update after every five-minute round. Like coaches at a swim meet, all they could do was sit back and watch. Source: DARPA DARPA hoped to show that software vulnerability discovery, patching, and exploitation could be automated. Together, these three phases constitute the \"vulnerability lifecycle.\" Once a software vulnerability is found, what happens next depends on who found it. Attackers exploit those vulnerabilities, allowing them to access protected systems. Defenders patch those same vulnerabilities to prevent compromise. Both offense and defense want to automate software vulnerability discovery, a welldeveloped field of research consisting of corporate developers and cybersecurity experts using tools to find software flaws. Automated patching and exploitation are relatively less-developed and not as widely used. DARPA's CGC and China's RHGs lump these three distinct phases together because they rely on similar technical processes and techniques. This paper refers to these capabilities as \"tools to automate the vulnerabilities lifecycle,\" or AVL tools. Currently, software vulnerability discovery, exploitation, and patching can be labor-intensive. 8 Software developers, often with years of experience, must pore over code looking for ways it can break. Even with existing tools and techniques, such as symbolic execution, it is impossible to consider all possible avenues of failure. Mistakes leave behind vulnerabilities that attackers may exploit. Open source fuzzing tools, such as American Fuzzy Lop, help researchers locate cracks in their code by generating inputs to cause software crashes. But the time dedicated to this process during software development is constrained by economics. Corporate requirements and shareholder value dictate the amount of time spent securing products, often resulting in insufficient attention. High labor costs and slow product development dent profits. AVL tools would pay huge dividends to companies and governments able to deploy the technology. The Cyber Grand Challenge provided a glimpse of the future by automatically identifying vulnerabilities, building and applying patches, and exploiting vulnerable programs. Although the event targeted relatively simple software compared to more widely-used programs, it demonstrated that AVL tools are viable. The day after CGC's machine-only event concluded, another important event unfolded. DEF CON, a conference for hackers that was co-located with DARPA's event, invited CGC's winning team to enter their system in a capture-the-flag game against DEF CON's human finalists.  9  ForAllSecure, the CGC's winning team from Carnegie Mellon University, agreed to submit their CRS, Mayhem. In the end, Mayhem lost to all of the competition's 14 human teams.  10  But it was not a resounding defeat. In the first 10 hours of the CTF, Mayhem led some of the human teams. In the hours that followed, the teams overtook the machine. By the end of the CTF, humans remained undefeated in hacking competitions. The events in Las Vegas set off a firestorm of articles touting the impact automation would have on cybersecurity.  11  The articles were optimistic and half right. ForAllSecure eventually received a contract to deploy their CRS on DOD systems.  12  Other competitors sold their systems to cybersecurity firms.  13  But AVL tools are still only deployed piecemeal-as specialized vulnerability discovery tools, not as fully developed vulnerability lifecycle products. On this front, the article's predictions were wrong. The technology is not trustworthy enough to automatically patch software, and most exploit generation requires a hands-on approach. Still, the competition and its results were so consequential that the Smithsonian National Museum of American History displayed Mayhem in an exhibition on innovation in defense technology.  14  CGC changed some fundamental assumptions about software and security that underpin the cyber domain. All the hard work spent engineering and fine-tuning the CRSs to do a human's job was sure to go somewhere impressive. But as far as is publicly known, the technology has not. DARPA never planned a second CGC. The agency pushed its research on automated exploitation and automated patching into siloed DARPA programs, reducing the public incentive to assemble such systems while simultaneously removing the technology's focal point for the cybersecurity community. The grand challenge model used for CGC isn't intended to support annual competitions, but rather tries to spur innovation and evaluate the best technology for a particular field at a single point in time. For Chinese Communist Party (CCP) policymakers, CGC did just that. \n China's Robot Hacking Games ( ) China's cybersecurity policymakers began monitoring the development of DARPA's CGC when it was first announced in 2014. Chinese policy publications and industry magazines hyped up the importance of the competition for cybersecurity.  15  That same year, Xi signaled that the party wanted China to become a \"cyber powerhouse\" ( ), an intentionally vague term meant to inspire. 16 Following Xi's announcement, CCP policymakers began releasing strategy documents to define what the political objective meant in technical terms. \n Xu Guibao ( ) was one of those policymakers. Xu served as a senior manager at a government think tank under China's Ministry of Industry and Information Technology. 17 Throughout his time in government service, he authored numerous policy documents related to China's 13th Five-Year Plan and received 10 patents for his technical innovations. When policymakers needed someone to serve as lead author for China's 2015 \"Internet + Artificial Intelligence Three-Year Action Plan,\" to support China's AI-related technology development, he was a perfect fit. In the span of a single year, Xu witnessed three events that shaped his perspective on the technologies China needed to achieve its \"cyber powerhouse\" ambitions. The first event drew global attention. In early 2016, DeepMind's AlphaGo beat Lee Sedol, a world-champion Go player, in four of five games.  18  The event concentrated minds around the world on the potential impact of machine learning. A few months later in late 2016, DARPA's CGC concluded with a human versus machine competition, where ForAllSecure's Mayhem led two of the fourteen human teams from DEF CON before ultimately losing.  19  For Xu, Mayhem's short-lived lead over humans reminded him of Lee Sedol's loss. The influential academic referred to Mayhem's performance against those teams as the \"AlphaGo incident in the field of cybersecurity\"-a bit hyperbolic, but indicative of his thinking at the time.  20  As winter turned to spring, Xu watched as the WannaCry ransomware tore its way across networks and pillaged computers around the world, including networks in China. WannaCry shook Chinese policymakers. For a regime that prizes stability and the government's ability to solve society's problems, the WannaCry incident concentrated minds on the powerful and uncontrollable effects malware could have. The trend in cybersecurity and AI was clear to Xu-automation promised scale and capabilities that could best humans. Xu published an influential article in 2017 titled \"U.S. Intelligent Cyberattack and Cyber Defense: Inspiration for China's Cyber Powerhouse Strategy.\"  21  The title alone was a clear indication Xu thought AVL tools should be one part of China's journey to becoming a \"cyber powerhouse.\" Xu argued that China must \"accelerate the development of a networked system of autonomous repair and offensive and defensive robots\" to achieve its cyber powerhouse strategy.  The International Robot Hacking Game ( , RHG) attempted to recreate DARPA's CGC.  26  Teams even decorated their server racks with the same combination of lights, colored trim, and logos. Despite differences in scoring and structure, China's first RHG tested the same types of technologies as the CGC: automated vulnerability discovery, patching, and exploitation.  27  In the same way that DARPA oversaw the CGC, the Ministry of State Security's 13th Bureau, the Central Cyberspace Administration of China, and the Ministry of Education supervised the International RHG competition. 28 Despite including \"international\" in its name, China's first RHG attracted only three of the competition's 22 teams from abroad-one was a CGC finalist.  29  In a humorous twist, a source familiar with the competition claimed one Chinese team just copied open-source code published by a CGC participant and hoped for the best. 30 They didn't win. In the end, China's National University of Defense Technology, a PLA military academy, beat the CGC finalist and other entrants to win the competition.  31  The \"international\" component of RHG competitions has since been dropped. Chinese policymakers saw what they needed to see. In the months following its first RHG, China doubled down on Xu's recommendations in its \"Internet +\" Artificial Intelligence Three-Year Action Plan covering the 2018 to 2020 time period.  32  The plan stated that \"in order to solve the security technology problems such as vulnerability discovery, security testing, threat warning, attack detection, and emergency response, enterprises should promote the advanced application of advanced AI technology in the field of cybersecurity.\" 33 By 2018, AVL tools had solidified their place in China's technology development strategies. The subsequent promulgation and standardization of RHG competitions was swift. Including preliminary rounds, China has hosted at least a dozen competitions for AVL technology since DARPA's CGC in 2016.  34  As an indication of its now prominent role, a 2019 article published by Civil-Military Integration in Cyberspace promoted the RHG model as a new standard for cybersecurity competitions in China, joining classic cyber games like capture the flag and jeopardy.  35  Center for Security and Emerging Technology | 8  \n Implications of China's RHGs China's Pursuit Will Endure Xi wants China to become a \"cyber powerhouse.\" Strategic policy documents signal that AVL tools are key to achieving Xi's ambitions.  37  As a result, the Party expects organizations able to research the technology to do so.  38  Efforts to develop AVL tools will persist until new strategic documents redefine what it means to be a \"cyber powerhouse\" or the technology meets the needs of the government. The widespread adoption of the RHG competition model provides strong incentives for Chinese academics, firms, and PLA laboratories to develop the technology. Although the prize money for winners of RHG competitions is paltry compared to private-sector competitions ($50K vs. $250K), party committees at universities and companies are able to encourage their organization's participation. In the United States, such small awards would fall short of the costs for just one researcher to work on AVL tools. In China, the CCP's political mandate to pursue the technology ensures that the competitions and technology remain a focal point for the cybersecurity community, regardless of the rewards offered. Organizations that are able to support the technology's development but choose not to would be out of step with the party-a politically untenable position. 39 China's crackdown on tech firms will concentrate minds on the need to be on the same team as party policymakers. The strong political signal by the CCP mobilizes resources across China to focus on the technologies' development. \n Increasing PLA Involvement The Ministry of State Security 13th Bureau and Ministry of Education served as government \"steering organizations\" ( ) responsible for managing the first three RHG competitions. 40 Some regional offices of the MSS 13th Bureau run cyber operations in partnership with regional State Security Bureaus. 41 But the 13th Bureau is also responsible for general cybersecurity issues within government agencies. The motivation behind the bureau's involvement in the first three RHGs is unclear-the Ministry of Education's involvement may suggest benign intentions. Although the MSS 13th Bureau has not hosted an RHG competition since late 2018, research on AVL tools may have been moved in-house. The technology's offensive and defensive uses, combined with the bureau's dual-purpose missions, obfuscate the nature of its interest. Few questions remain about the interest of the PLA, however. The Third Annual Qiangwang Cup ( ), which is self-described as having \"a natural tendency towards military-civil fusion ( ),\" marked the shift towards PLA involvement. Qiangwang Cup was the first competition overseen by the Central Cyberspace Administration of China and PLA Information Engineering University.  42  The shift from MOE and MSS 13th Bureau oversight suggests increased military interest in the technology. PLA Information Engineering University is part of the PLA Strategic Support Force's Network Systems Department, which is responsible for military hacking operations.  43  The university's oversight of the RHG may reflect an interest in recruiting students with knowledge of AVL tools, since the Qiangwang Cup is a competition for college students.  \n RHGs Are Evolving As long as AVL tools are central to the competition, hosts can change game structures and experiment with operational concepts. The Zongheng Cup introduced human-machine team competitions, where an automated AVL system supports two people in a 3-vs.-3 capture-the-flag style competition.  46  This human-in-the-loop concept is behind one of DARPA's follow-on programs to the CGC-Computers and Humans Exploring Software Security (CHESS).  47  Overseen by a lab affiliated with the PLA Equipment Development Department, the Zongheng Cup demonstrates converging operational concepts between the United States and China. RHGs are no longer changing their structures to match those of the Ministry of Education, but instead those of the PLA. \n Experience and Collaboration China's system of competitions attracts new participants, facilitates hands-on experience, and fosters relationships between institutions and competing teams. \"Promot[ing] the training and selection of talents in the field of AI-based cybersecurity\" was a key objective for China's first RHG and remains a goal of each subsequent competition.  48  Although automated software vulnerability discovery, patching, and exploitation promise to be more efficient than human professionals alone, these systems still require specialized knowledge to deploy. Operators with experience using the technology can more easily diagnose and solve errors as they arise during deployment. Competitions also encourage relationships between participants. These relationships can be formal, such as teams representing multiple institutions, or informal-social gatherings after the competition. Having a cohort of researchers familiar with the technology is crucial to its successful deployment. Close professional connections could provide networks for troubleshooting technical issues or helping the PLA deploy the technology. \n Conclusion China's state hacking teams, which involve the PLA and Ministry of State Security, stand ready to adopt AVL tools. A report from MIT Technology Review detailed how China's government monitored cybersecurity competitions for new tools and techniques, then rapidly acquired and deployed them against domestic surveillance targets in Xinjiang.  49  RHGs are likely no different. But a full lifecycle AVL tool has not been compiled yet. Instead, individual parts of the tools-like fuzzers, symbolic execution, or automatic exploit generation-may progress in a piecemeal fashion. Automated vulnerability tools are already widely deployed in software development, so improvements in the technology are building on past success. Still, the CEO of Qihoo360, the cybersecurity firm responsible for China's Cybersecurity Military-Civil Fusion Innovation Center-among other state ties-called automated vulnerability discovery tools an \"Assassin's Mace\" for China.  50  The arcane term references the military strategy of creating an asymmetric advantage over a more powerful enemy-in DOD jargon, it is the Chinese Offset Strategy. For China's military, attacking an adversary's command and control system to disrupt \"system-of-systems\" communication would fit the bill. 51 AVL tools could help the PLA foment such an attack. U.S. policymakers should consider whether current support for developing AVL technologies is enough. China's largest tech firms and universities are now competing at events hosted by the PLA's labs. Those competitions, in turn, spur innovation, connect researchers, and create a platform for iteratively testing and improving the technology. The United States, by contrast, supports three DARPA programs: Assured Micropatching, CHESS, and Harnessing Autonomy for Countering Cyberadversary Systems.  52  Combined with any classified programs or allocations, these three programs represent the USG's best efforts to develop AVL tools. To get the most out of the technology and maintain any lead over China in this technology, the United States may need to invest more in developing AVL tools. Public competitions with cash prizes large enough to turn winners into businesses could be a good first step. DARPA's CGC in 2016 helped launch a few new companies. But increasing investment and public interest in the technology by the cybersecurity community could yield even greater dividends. With some luck and more public investment, new businesses and a more secure U.S. cyber domain could be in the offing. Endnotes 1 , \" phrack -.\" , September 26, 2017, https://perma.cc/W5VH-J7F5. The translation to \"Robot Hacking Game\" from Mandarin is both a direct translation, and the translation used in China's own translations. Figure  2  shows each server rack embossed with \"RHG\" in large white letters at the top to drive home the competition's branding. Although the name evokes thoughts of animated machinery moving about, the more appropriate English-language idea might be a \"bot\"-used to denote automated bits of software from virtual assistants to automated web scrapers. 2 \" 'Halfbit' ,\" , November 8, 2017, https://perma.cc/ESL5-8YNL; --, \" ,\" April 23, 2019, https://perma.cc/9E4N-CGY4; 419, \"RHG , ,\" Sohu, March 30, 2021, https://perma.cc/7E53-2FBZ. \n 3 , \" IQ RHG ,\" , March 30, 2021, https://perma.cc/93CH-QXNN.  4  Translator's note: For a more in-depth discussion in English of the Chinese term , which can be rendered as \"cyber powerhouse\" or \"cyber superpower,\" see Rogier Creemers et al., \"Lexicon: Wǎngluo Qiángguó,\" New America, May 31, 2018, https://www.newamerica.org/cybersecurityinitiative/digichina/blog/lexicon-wangluo-qiangguo/. 5 Dustin Fraze, \"Cyber Grand Challenge,\" Defense Advanced Research Projects Agency, accessed August 27, 2021, https://perma.cc/65W8-XEEK. 6 Defense Advance Research Projects Agency, \"Cyber Grand Challenge Rules, Version 3,\" Massachusetts Institute of Technology. November 18, 2014, 12-13, https://archive.ll.mit.edu/cybergrandchallenge/docs/CGC_Rules_18_Nov_14_Ver sion_3.pdf. The \"attacks\" were, in fact, proofs of concepts that exploited other teams' vulnerabilities. An automated referee system evaluated whether the attacks would work as intended, and if so, awarded points. The victims were docked points, but no malware was installed on the targeted system. This structure prevented teams from permanently impairing their opponents and focused the game on vulnerability discovery, exploitation, and patching.  Figure 1 . 1 Figure 1. DARPA's CGC in Las Vegas 7 \n 22 (These so-called robots are the cyber reasoning systems tested at DARPA's CGC and Xu's inspiration). The CCP channeled Xu's recommendations in its New Generation Artificial Intelligence Development Plan, released around the same time in 2017, generically stating China must \"strengthen AI cybersecurity technology research and development.\" 23 Chinese policymakers would later expressly echo Xu's recommendation. But by the time China's first Robot Hacking Game competitors filed into a Wuhan conference room in late 2017, it was already becoming clear that China viewed AVL tools as important to becoming a \"cyber powerhouse.\" 24 \n Figure 2 : 2 Figure 2: China's First RHG in 2017 25 \n Figure 3 : 3 Figure 3: A Timeline of RHG Finals. 36 \n In 2021, military oversight of RHGs expanded further. The Key State Laboratory for Information System Security Technology ( ), a lab administered by the PLA's Equipment Development Department, managed the 2021 Zongheng Cup ( ). 44 According to the U.S.-China Security and Economic Review Commission, the Equipment Development Department \"plays a central role in military modernization by overseeing weapons development across the entirety of the PLA.\" 45 The lab's oversight of the competition indicates an uptick in the PLA's responsibility for developing, and possibly deploying, the technology. \n ( June 2014): 22-23.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Robot-Hacking-Games.tei.xml", "id": "938b4304b6bbaa132879643d1443f440"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "for suggestions and comments. This work was funded in part by the Berkeley Existential Risk Initiative. All errors are my own. \n Unfamiliarity: The possibility that the post-Advanced AI world will be very unfamiliar to those crafting agreements pre-Advanced AI. 10 The potential speed of a transition between pre-and post-Advanced AI states exacerbates these issues. 11 Indeterminacy and unfamiliarity are particularly problematic for pre-Advanced AI agreements. Under uncertainty alone (and assuming the number of possible outcomes is manageable), it is easy to specify rights and duties under each possible outcome. However, it is much more difficult to plan for an indeterminate set of possible outcomes, or a set of possible outcomes containing unfamiliar elements. A common justification for the rule of law is that it promotes stability 12 by increasing predictability 13  and therefore the ability to plan. 14  Legal tools, then, should provide a means of minimizing disruption of pre-Advanced AI plans during the transition to a post-Advanced AI world. Of course, humanity has limited experience with Advanced AI-level transitions. Although analysis of how legal arrangements and institutions weathered similar transitional periods would be valuable, this Report does not offer it. Rather, this Report surveys the legal landscape and identifies common tools and doctrines that could reduce disruption of pre-Advanced AI agreements during the transition to a post-Advanced AI world. Specifically, it identifies common contractual tools and doctrines that could faithfully preserve the goals of pre-Advanced AI plans, even if unforeseen and unforeseeable societal changes from Advanced AI render the formal content of such plans irrelevant, incoherent, or suboptimal. A key conclusion of this Report is this: stable preservation of pre-Advanced AI agreements could require parties to agree ex ante to be bound by some decisions made post-Advanced AI, with the benefit of increased knowledge. 15  By transmitting (some) key, binding decision points forward in time, actors can mitigate the risk of being locked into naïve agreements that have undesirable consequences when applied literally in uncontemplated circumstances. 16  Parties can often constrain those ex post choices by setting standards for them ex ante. 17 10 E.g., if I know I will be transported to Uzbekistan, the outcome is both certain and determinate, but, since I have never been to Uzbekistan, that result is in some sense unfamiliar to me. 11 Thanks to Ben Garfinkel for this point.", "authors": ["Cullen O'keefe"], "title": "Stable Agreements in Turbulent Times: A Legal Toolkit for Constrained Temporal Decision Transmission", "text": "Introduction This century, 1 advanced artificial intelligence (\"Advanced AI\") technologies could radically change economic or political power.  2  Such changes produce a tension that is the focus of this Report. On the one hand, the prospect of radical change provides the motivation to craft, ex ante, agreements that positively shape those changes.  3  On the other hand, a radical transition increases the difficulty of forming such agreements since we are in a poor position to know what the transition period will entail or produce. The difficulty and importance of crafting such agreements is positively correlated with the magnitude of the changes from Advanced AI. The difficulty of crafting long-term agreements in the face of radical changes from Advanced AI is the \"turbulence\" 4 with which this Report is concerned. This Report attempts to give readers a toolkit for making stable agreements-ones that preserve the intent of their drafters-in light of this turbulence. Many agreements deal with similar problems to some extent. Agreements shape future rights and duties, but are made with imperfect knowledge of what this future will be like. To take a real-life example, the outbreak of war could lead to nighttime lighting restrictions, rendering a long-term rental of neon signage suddenly useless to the renter.  5  Had the renter foreseen such restrictions, he would have surely entered into a different agreement.  6  Much of contract law is aimed at addressing similar problems. However, turbulence is particularly problematic for pre-Advanced AI agreements that aim to shape the post-Advanced AI world.  7  More specifically, turbulence is a problem for such agreements for three main reasons: 1. Uncertainty: Not knowing what the post-Advanced AI state of the world will be (even if all the possibilities are known); 8 2. Indeterminacy: Not knowing what the possible post-Advanced AI states of the world are; 9 and 1 Cf. Greg Brockman, Co-Founder & Chief Technology Officer, OpenAI, Can We Rule Out Near-Term AGI? (Nov. 7, 2018), https://www.youtube.com/watch?v=YHCSNsLKHfM; Katja Grace et al., When Will AI Exceed Human Performance? Evidence from AI Experts 2 (2018), https://perma.cc/2K2D-LE3A (unpublished manuscript) (\"Taking the mean over each individual, the aggregate forecast gave a 50% chance of [high-level machine intelligence] occurring within 45 years and a 10% chance of it occurring within 9 years.\").  2  See Holden Karnofsky, Some Background on Our Views Regarding Advanced Artificial Intelligence, OPEN PHILANTHROPY PROJECT: BLOG § 1 (May 6, 2016), https://perma.cc/2H7A-NZTA.  3  See generally Nick Bostrom et al., Public Policy and Superintelligent AI: A Vector Field Approach, in ETHICS OF ARTIFICIAL INTELLIGENCE (S. M. Liao ed., forthcoming 2019) (manuscript version 4.3), https://perma.cc/SN54-HKEG.  4  See id. at 7.  5  This fact pattern is taken from 20th Century Lites, Inc. v. Goodman, 149 P.2d 88 (Cal. App. Dep't Super. Ct. 1944).  6  See id. at 92.  7  There may not be a clear pre-and post-Advanced AI boundary, just as there was not with previous revolutions. Nevertheless, hypothesizing such a clear boundary is useful in thinking through the issues with which this Report is concerned. Thanks to Ben Garfinkel for this point. 8 E.g., I know that a flipped coin can land on either heads or tails, but I am uncertain of what the result of any given flip will be. 9 E.g., if I have an opaque bag containing a single die with an unknown number of sides, the set of possible outcomes from rolling that die (without first examining it) is indeterminate: the outcome \"7\" might or might not be possible, depending on how many sides the die in fact has. If the die is a regular cubic die, \"7\" is impossible; if the die is dodecahedral, \"7\" is possible. In any case, the result is also uncertain. This Report aims to help nonlawyer readers develop a legal toolkit to accomplish what I am calling \"constrained temporal decision transmission.\" All mechanisms examined herein allow parties to be bound by future decisions, as described above; this is \"temporal decision transmission.\" However, as this Report demonstrates, these choices must be constrained because binding agreements require a degree of certainty sufficient to determine parties' rights and duties.  18  As a corollary, this Report largely does not address solely ex ante tools for stabilization, such as risk analysis,  19  stabilization clauses,  20  or fully contingent contracting. For each potential tool, this Report summarizes its relevant features and then explain how it accomplishes constrained temporal decision transmission. My aim is not to provide a comprehensive overview of each relevant tool or doctrine, but to provide readers information that enables them to decide whether to investigate a given tool further. Readers should therefore consider this Report more of a series of signposts to potentially useful tools than a complete, ready-to-deploy toolkit. As a corollary, deployment of any tool in the context of a particular agreement necessitates careful design and implementation with special attention to how the governing law treats that tool. Finally, this Report often focuses on how tools are most frequently deployed. Depending on the specific tool and jurisdiction, however, readers might very well be able to deploy tools in non-standard ways. They should be aware, however, that there is a tradeoff between novelty in tool substance and legal predictability. The tools examined here are: • Options-A contractual mechanism that prevents an offeror from revoking her offer, and thereby allows the offeree to accept at a later date; • Impossibility doctrines-Background rules of contract and treaty law that release parties from their obligations when circumstances dramatically change; • Contractual standards-Imprecise contractual language that determines parties' obligations in varying circumstances; • Renegotiation-Releasing parties from obligations under certain circumstances with the expectation that they will agree on alternative obligations; and • Third-party resolution-Submitting disputes to a third-party with authority to issue binding determinations. Although the tools studied here typically do not contemplate changes as radical as Advanced AI, they will hopefully still be useful in pre-Advanced AI agreements. By carefully deploying these tools (individually or in conjunction), readers should be able to ensure that the spirit of any pre-Advanced AI agreements survives a potentially turbulent transition to a post-Advanced AI world. agent ex ante with incomplete information and specifying optimal behaviors ex post once more information about the state of the world is available.\").  17  See generally Kaplow, supra note 16, at 589; Scott & Triantis, supra note 16.  18      \n I. Why Agreement Incompleteness? If promoting stability is a goal of law generally and binding agreements specifically, why do parties fail to maximize stability by specifying agreement obligations in all contingencies to the greatest extent possible? 21 That is, why do parties underspecify their agreements? This problem is generally termed agreement \"incompleteness.\"  22  This Part points to some reasons why an agreement might be (more) incomplete.  23   \n A. Transaction Costs One of the most common explanations for incompleteness is that contracting has transaction costs:  24  Generally, neither party has the goal of negotiating the perfect contract. The contract is the means to the parties' end, not the end itself. The more time a party spends negotiating, the more it delays the performance that will make it better off. Parties want their counterparty's performance, not a well-negotiated deal. That is why much contracting is quite informal. For example, a buyer calls a seller to ask about the availability of a part the buyer needs in his business operation. The buyer wants the part, not the perfect contract and will accept some interpretive or performance risk in order to get what it needs as quickly as possible. That is one of the main reasons why parties underspecify their obligations and rely on post contracting adjustments and informal enforcement to reduce the costs of contracting. . . . However, the amount each party is willing to invest in negotiating costs will differ depending on individual preferences, goals, foresight, and trust in the mechanisms of informal enforcement. One party may agree to terms thinking that the terms are precise enough to deal with all contingencies while the other party may realize that a term, while precise, does not cover all contingencies; that party may plan to rely on informal enforcement to address post-contracting disputes.  25  Time is one of the most obvious transaction costs. Money is also important, especially if attorney costs are involved.  26  Emotional and relational costs are also relevant.  27  For example, a proud seller might take offense  21  Cf. Hadfield-Menell & Hadfield, supra note 16, at 1 (\"The ideal way to align principal and agent is to design a complete contingent contract. This is an enforceable agreement that specifies the reward received by the agent for all actions and states of the world.\" (citation omitted)).  22  Cf. id. at 2.  23  The relevant literature is vast, so this section is necessarily summary. For more discussion of the reasons for contract incompleteness, see id.; see also George S.   147, 199 (1996) . at being asked to agree to terms conditioned on his own bad faith. Thus, the buyer might refrain from asking the seller to agree to such terms-even if they would reduce uncertainty-to preserve an amicable relationship. \n B. Bounded Rationality Another obvious source of incompleteness is parties' bounded rationality. 28 \"[T]hat is, the parties are subject to significant time, resource, and cognitive restraints that limit their capacity to choose an optimal outcome.\"  29  For example, the \"planning fallacy\" predicts that people are \"prone to underestimate the time required to complete a project, even when they have considerable experience of past failures to live up to planned schedules.\" 30 Thus, parties might fail to agree on what should happen if the project takes considerably longer than anticipated. More generally, parties might ignore, overlook, or under-appreciate the possibility of bad outcomes, even when the probability thereof is non-trivial.  31  Relatedly, parties' information about future conditions is necessarily imperfect. 32 Uncertainty, indeterminacy, and unfamiliarity (as defined in the Introduction) all contribute to this. Thus, even in a world with no transaction costs, parties' ability to form efficient agreements 33 is hindered by their bounded rationality. \n C. Interpretive Uncertainty A final source of incompleteness is the fact that parties are often unsure about how courts would interpret contract provisions.  34  Richard Craswell gives the following example: [Consider] a contract between [seller] S and [buyer] B, entered into at a time when there was some uncertainty about S's future cost of production. Suppose now that, if all the relevant incentives were taken into account, it would be efficient to grant S an excuse whenever her costs increased by more than 127 percent. If courts were able to measure S's costs with no risk of error, achieving the ideal result would simply be a matter of granting an excuse whenever her costs in fact went up by more than 127 percent. In practice, though, courts may not always be in a good position to measure S's costs, especially if some of those costs involve hard-to-quantify variables. More generally, there are many other things that courts also may be poor at measuring. In some cases, the most efficient outcomes may depend on factors that are completely unobservable (for instance, the efficiency of completing a consumer transaction may depend on whether the consumer's tastes have changed in some unobservable way). In other cases, the efficient outcome may depend on factors that are observable to the contracting parties, but that cannot be proved to the satisfaction of a reviewing court (for example, the seller's costs may include opportunity costs that a court would find hard to evaluate). In the newer literature on incomplete contracts, these two difficulties are often referred to (respectively) as involving information that is either unobservable or nonverifiable.  35  By contrast, if \"a contract that says S will deliver 100 widgets on July 1 could be considered 'complete' (in the sense of not leaving any gaps) if it is [invariably] interpreted to mean that the seller must deliver those widgets on July 1 regardless of anything else that might happen.\" 36 Thus, the fact that parties do not know how courts will interpret all contractual obligations in all contingencies itself causes contractual incompleteness. As Robert E. Scott and George G. Triantis pointed out in their influential Yale Law Journal article, when contracts grant interpretive flexibility-i.e., when contracts use standards as opposed to rules-parties trade off front-end contracting costs for back-end costs like uncertainty 37 and, potentially, litigation.  38  Thus, interpretive uncertainty is a calculated part of the contracting process. 39 35 Craswell, supra note 33, at 155-56.  36  Id. at 154.  37  See Posner, supra note 23, at 1608-09 (discussing error cost).  38  See generally Scott & Triantis, supra note 16.  39  See generally id. \n II. Option Contracts Option contracts may be the simplest legal mechanism for deferring decision-making. \"An option contract is a promise which meets the requirements for the formation of a contract and limits the promisor's power to revoke an offer.\"  40  For example: A offers to sell B Blackacre for $5,000 at any time within thirty days. Subsequently A promises . . . in return for $100 paid or promised by B that the offer will not be revoked. There is an option contract under which B has an option [to purchase Blackacre for $5,000].  41  As with any contract, 42 the offeror can condition the offeree's ability to accept (i.e., \"exercise\" her option) on the occurrence of some event. Options are a clear and simple example of temporal decision transmission. An offeree benefits from an option contract because she believes she will later be in a better position to decide whether to accept. Options are useful when parties wish to keep a definite offer open for some period of time, while allowing the offeree to gather more information so as to make a better-informed acceptance decision. Options are a limited tool, at least on their own. A contract must specify the content of the option; the offeree's only choice is between exercising the option or not. However, options can become more flexible when combined with other tools in this Report. For example, an agreement might stipulate that when certain conditions are satisfied, a party has an option to renegotiate an agreement.  43  In this way, options are limited on their own, but can be quite stabilizing when combined with other stabilizing agreement features. 40 RESTATEMENT (SECOND) OF CONTRACTS, supra note 18, § 25.  41  Id. illus. 2.  42  See id. § 36(2) (\"[A]n offeree's power of acceptance is terminated by the non-occurrence of any condition of acceptance under the terms of the offer.\").  43  See infra Part V. \n III. Impossibility in Contract and Treaty Law Minimizing turbulence and allocating the associated risks is a major purpose of contract law: The process by which goods and services are shifted into their most valuable uses is one of voluntary exchange. The distinctive problems of contract law arise when the agreed-upon exchange does not take place instantaneously (for example, A agrees to build a house for B and construction will take several months). The fact that performance is to extend into the future introduces uncertainty, which in turn creates risks. A fundamental purpose of contracts is to allocate these risks between the parties to the exchange.  44  The doctrine of Impossibility is a key mechanism by which contract law deals with turbulence. A close analogy exists in treaty law: the doctrine of rebus sic stantibus (\"things thus standing\"). This Part first restates the doctrine (or, more precisely, doctrines) of Impossibility and rebus sic stantibus, then analyzes why they are stabilizing. \n A. Contractual Impossibility Doctrines The Restatement (Second) of Contracts introduces the idea of Impossibility as follows: Contract liability is strict liability. It is an accepted maxim that pacta sunt servanda, [meaning] contracts are to be kept. The obligor is therefore liable in damages for breach of contract even if he is without fault and even if circumstances have made the contract more burdensome or less desirable than he had anticipated. . . . Even where the obligor has not limited his obligation by agreement, a court may grant him relief. An extraordinary circumstance may make performance so vitally different from what was reasonably to be expected as to alter the essential nature of that performance. In such a case the court must determine whether justice requires a departure from the general rule that the obligor bear the risk that the contract may become more burdensome or less desirable. [Impossibility doctrine] is concerned with the principles that guide that determination.  45  Two forms of contract Impossibility are relevant to this Report: Impracticability and Frustration. \n Impracticability The doctrine of Impracticability is stated as follows: Where, after a contract is made, a party's performance is made impracticable without his fault by the occurrence of an event the non-occurrence of which was a basic assumption on which the contract was made, his duty to render that performance is discharged, unless the language or the circumstances indicate the contrary.  46  The Uniform Commercial Code 47 and the United Nations Convention on the International Sale of Goods 48 contain similar provisions. Traditionally, courts have generally found impracticability in the following three cases: 49 1. Supervening death or incapacity of a person necessary for performance 50 2. Supervening destruction of a specific thing necessary for performance 51 3. Supervening prohibition or prevention by law 52 This list is not exhaustive, however.  53  Note also that mere changes in market conditions are usually not grounds for discharge under this rule.  54  The foreseeability of a contingency is relevant to an Impracticability analysis, but not determinative thereof.  55  \"Furthermore, a party is expected to use reasonable efforts to surmount obstacles to performance, and a performance is impracticable only if it is so in spite of such efforts.\" 56 Thus, performance need not be truly \"impossible\" to fall under this doctrine; \"[p]erformance may be impracticable because extreme and unreasonable difficulty, expense, injury, or loss to one of the parties will be involved.\" contract for sale if performance as agreed has been made impracticable by the occurrence of a contingency the nonoccurrence of which was a basic assumption on which the contract was made or by compliance in good faith with any applicable foreign or domestic governmental regulation or order whether or not it later proves to be invalid.\").  48  United Nations Convention on Contracts for the International Sale of Goods art. 79, Jan. 1, 1988, 1489 U.N.T.S. 3, https://perma.cc/J2QF-FGW4 (\"A party is not liable for a failure to perform any of his obligations if he proves that the failure was due to an impediment beyond his control and that he could not reasonably be expected to have taken the impediment into account at the time of the conclusion of the contract or to have avoided or overcome it or its consequences.\"). 49 RESTATEMENT (SECOND) OF CONTRACTS, supra note 18, § 261 cmt. a.  50  See id. § 262 (\"If the existence of a particular person is necessary for the performance of a duty, his death or such incapacity as makes performance impracticable is an event the non-occurrence of which was a basic assumption on which the contract was made.\").  51  See id. § 263 (\"If the existence of a specific thing is necessary for the performance of a duty, its failure to come into existence, destruction, or such deterioration as makes performance impracticable is an event the non-occurrence of which was a basic assumption on which the contract was made.\").  52  See id. § 264 (\"If the performance of a duty is made impracticable by having to comply with a domestic or foreign governmental regulation or order, that regulation or order is an event the non-occurrence of which was a basic assumption on which the contract was made.\").  53  See id. § 261 cmt. a.  54  See id. cmt. b.  55  See id.  56  Id. cmt. d (citation omitted).  57  Id. The following examples demonstrate circumstances in which performance is Impracticable due to violations of the basic assumptions of a contract: A contracts to repair B's grain elevator. While A is engaged in making repairs, a fire destroys the elevator without A's fault, and A does not finish the repairs. A's duty to repair the elevator is discharged, and A is not liable to B for breach of contract.  58  A contracts with B to carry B's goods on his ship to a designated foreign port. A civil war then unexpectedly breaks out in that country and the rebels announce that they will try to sink all vessels bound for that port. A refuses to perform. Although A did not contract to sail on the vessel, the risk of injury to others is sufficient to make A's performance impracticable. A's duty to carry the goods to the designated port is discharged, and A is not liable to B for breach of contract.  59  By contrast, performance in the following circumstance is not Impracticable: Several months after the nationalization of the Suez Canal, during the international crisis resulting from its seizure, A contracts to carry a cargo of B's wheat on A's ship from Galveston, Texas to Bandar Shapur, Iran for a flat rate. The contract does not specify the route, but the voyage would normally be through the Straits of Gibraltar and the Suez Canal, a distance of 10,000 miles. A month later, and several days after the ship has left Galveston, the Suez Canal is closed by an outbreak of hostilities, so that the only route to Bandar Shapur is the longer 13,000 mile voyage around the Cape of Good Hope. A refuses to complete the voyage unless B pays additional compensation. A's duty to carry B's cargo is not discharged, and A is liable to B for breach of contract.  60  The difference in outcome turns on a flexible 61 judicial determination as to whether changed circumstances violated an essential assumption of the contract. \n Frustration Frustration doctrine is stated as follows: Where, after a contract is made, a party's principal purpose is substantially frustrated without his fault by the occurrence of an event the non-occurrence of which was a basic assumption on which the contract was made, his remaining duties to render performance are discharged, unless the language or the circumstances indicate the contrary.  62  Comments to the Restatement clarify that performance is \"Frustrated\" only if:  58  Id. illus. 6.  59  Id. illus. 7.  60  Id. illus. 9.  61  Id. cmt. b (\"In borderline cases this criterion is sufficiently flexible to take account of factors that bear on a just allocation of risk.\").  62  Id. § 265. \n \"[T] he purpose that is frustrated must have been a principal purpose of that party in making the contract. It is not enough that he had in mind some specific object without which he would not have made the contract. The object must be so completely the basis of the contract that, as both parties understand, without it the transaction would make little sense\"; 63 2. \"[T]he frustration must be substantial. It is not enough that the transaction has become less profitable for the affected party or even that he will sustain a loss. The frustration must be so severe that it is not fairly to be regarded as within the risks that he assumed under the contract\"; 64 and 3. \"[T]he non-occurrence of the frustrating event must have been a basic assumption on which the contract was made. This involves essentially the same sorts of determinations that are involved under the general rule on [I]mpracticability.\"  65  For example: A, who owns a hotel, and B, who owns a country club, make a contract under which A is to pay $1,000 a month and B is to make the club's membership privileges available to the guests in A's hotel free of charge to them. A's building is destroyed by fire without his fault, and A is unable to remain in the hotel business. A refuses to make further monthly payments. A's duty to make monthly payments is discharged, and A is not liable to B for breach of contract.  66  By contrast: A leases a gasoline station to B. A change in traffic regulations so reduces B's business that he is unable to operate the station except at a substantial loss. B refuses to make further payments of rent. If B can still operate the station, even though at such a loss, his principal purpose of operating a gasoline station is not substantially frustrated. B's duty to pay rent is not discharged, and B is liable to A for breach of contract. The result would be the same if substantial loss were caused instead by a government regulation rationing gasoline or a termination of the franchise under which B obtained gasoline.  67  Again, the results turn on a judicial determination as to whether the contract makes sense in light of the changed circumstances. 68 \n B. Rebus Sic Stantibus (\"Things Thus Standing\") Impossibility finds a close analog in treaty law in the doctrine of rebus sic stantibus.  69  The Vienna Convention on the Law of Treaties restates the doctrine as follows:  63  Id. cmt. a.  64  Id.  65  Id.  66  Id. illus. 3.  67  Id. illus. 6.  68  See id. cmt. a. 1. A fundamental change of circumstances which has occurred with regard to those existing at the time of the conclusion of a treaty, and which was not foreseen by the parties, may not be invoked as a ground for terminating or withdrawing from the treaty unless: (a) The existence of those circumstances constituted an essential basis of the consent of the parties to be bound by the treaty; and (b) The effect of the change is radically to transform the extent of obligations still to be performed under the treaty. 2. A fundamental change of circumstances may not be invoked as a ground for terminating or withdrawing from a treaty: (a) If the treaty establishes a boundary; or (b) If the fundamental change is the result of a breach by the party invoking it either of an obligation under the treaty or of any other international obligation owed to any other party to the treaty. 3. If, under the foregoing paragraphs, a party may invoke a fundamental change of circumstances as a ground for terminating or withdrawing from a treaty it may also invoke the change as a ground for suspending the operation of the treaty.  70  However, there is a \"substantial bar to the successful application of the doctrine.\" 71 \"[T]here has never been a successful assertion of it in a court case and . . . there is no clear example of its successful use in diplomatic exchanges.\"  72   \n C. Impossibility, Rebus, and Stabilization At first blush, it might seem odd to call doctrines that discharge contractual obligations \"stabilizing.\" We often associate stability with firm adherence to rules, notwithstanding changed circumstances. Of course, this is the norm in contract law; Impossibility and rebus are exceptions. Nevertheless, Impossibility doctrines can be fairly said to be stabilizing for two reasons. First, \"[c]ommon sense sets limits to a promise, even where contractual language does not.\" 73 Thus, while a categorical refusal to discharge duties in the face of Impossible circumstances might be stabilizing in the sense that obligation  69   would closely track contractual language, 74 such a regime would be destabilizing as compared to the reasonable expectations of contracting parties.  75  Secondly, in repeat-play dynamics, Impossibility doctrines can incentivize stabilizing behaviors. In their groundbreaking article Impossibility and Related Doctrines in Contract Law-An Economic Analysis, Richard A. Posner and Andrew M. Rosenfield explain that, from an economic perspective, Impossibility doctrines promote efficiency when they assign economic losses 76 in such situations to the \"superior risk bearer.\" 77 Each party's comparative ability to prevent 78 and/or insure against 79 risks determines which is the superior risk bearer.  80  Economical Impossibility doctrines would therefore promote stability by assigning losses associated with radically changed circumstances to the superior risk bearer, thus incentivizing her to take stabilityenhancing measures (such as avoidance or insurance). Caselaw partially reflects the economic approach. For example, \"[f]rom the standpoint of economics, contract discharge should not be allowed when the event rendering performance uneconomical was reasonably preventable by either party.\" 81 And indeed, \"[t]his view prevails in the case law.\"  82  Posner and Rosenfield find that a number of common case patterns follow the superior risk bearer logic.  83  For example: Discharge of contracts for personal services is often sought when an employee has died unexpectedly. In Cutler v. United Shoe Mach. Corp., a machinery company had an employment contract with an inventor. When the inventor died, the court, in a suit by the inventor's estate, held the contract discharged. This outcome-typical in employee cases-is consistent with the economic approach. The employee (1) is in at least as good a position as his employer to estimate his life expectancy, (2) has better knowledge of the value of the contract to him compared to that of any alternative employment and (3) can readily purchase life insurance.  74  Cf. Gillian K. Hadfield, Judicial Competence and the Interpretation of Incomplete Contracts, 23 J. LEGAL STUD. 159, 160 n.4 (1994) (as if \"a contract that says 'A to deliver 100 pounds of peas to B for $25' includes the term 'under any and all circumstances' including earthquakes and pea shortages.\").  75  See Kull, supra note 73, at 38-39.  76  As Posner and Rosenfield explain, \"[i]n every [Impossibility] case the basic problem is the same: to decide who should bear the loss resulting from an event that has rendered performance by one party uneconomical.\" Posner & Rosenfield, supra note 44, at 86. In other words, if the promisor is not held liable for breach when performance is Impossible, then the promisee is, in an economic sense, liable, since she will no longer receive the benefits the promisor promised her. Thus, in Impossible circumstances, the loss must go to some party.  77  See id. at 90.  78  See id.  79  See id. at 90-91.  80  See id.  81  Id. at 98.  82     83  See id. at 100-08. If the employer were seeking damages as a result of the employee's death, alleging that death had caused a breach of the employee's obligations under the contract, the contract should also be discharged. Estimating life expectancy is in general no more (if no less) difficult for the employer than for the employee (if the employee knew of a condition reducing his life expectancy below the actuarial level for people of his age, sex, etc., discharge would presumably not be allowed). And the employer is better able to estimate the cost to him (in firm-specific human capital, replacement costs, etc.) if the employee dies, and can usually self-insure against such an eventuality.  84  Impossibility and rebus also reflect, to some degree, the thesis of this Report that \"transmitting (some) key decision points forward in time\" can be stabilizing. Admittedly, Impossibility and rebus do not reflect an explicit ex ante commitment to transmit binding decision points into the future. Nevertheless, the regime they effect is somewhat analogous to one in which parties maintain an implied option 85 to refuse to perform under certain circumstances.  86  The economic default regime proposed by Posner and Rosenfield is, in effect, one in which parties implicitly agree ex ante to assign certain losses to the superior risk bearer, as determined ex post in light of the circumstance causing the loss. Put differently, these doctrines provide ex ante background rules for determining ex post which party should bear losses in light of the particular change in circumstances. Thus, they constitute an implicit temporal transmission of a key decision point: the determination of which party is the superior risk bearer is made ex post rather than ex ante. Still, since these are background rules, parties do not usually explicitly agree to them ex ante. The next Part examines how parties can make similar risk allocations explicit.  84  Id. at 100 (footnote omitted) (citing Cutler v. United Shoe Mach. Corp., 174 N.E. 507 (Mass. 1931)).  85  See generally supra Part II.  86  Cf. Geis, supra note 23, at <PHONE> (arguing that indefinite contracts can be usefully understood as containing embedded options). \n IV. Contractual Standards Doctrines identified in the previous Part provide default rules for discharging contractual duties in absence of an explicit agreement about the allocation of the relevant risks. 87 However, as the language of the relevant doctrines implies, 88 parties are free to explicitly allocate risks otherwise.  89  This Part examines some common contractual clauses that parties may use to stabilize their agreements in the face of significantly changed circumstances. A key feature of many such clauses is that they use standards rather than rules 90 for governing parties' actions.  91  Note that the clauses examined here merely two examples of standards; other examples exist.  92  A. Force Majeure Clauses Force majeure (French: \"superior force\")  93  clauses explicitly allocate risk from dramatically changed circumstances. They are thus an explicit version of Impossibility doctrine. A representative force majeure clause is as follows: If either party is rendered unable by force majeure, or any other cause of any kind not reasonably within its control, wholly or in part, to perform or comply with any obligation or condition of this Agreement, upon such party's giving timely notice and reasonably full particulars to the other party such obligation or condition shall be suspended during the continuance of the inability so caused and such party shall be relieved of liability and shall suffer no prejudice for failure to perform the same during such period . . . . The cause of the suspension (other than strikes or differences with workmen) shall be remedied so far as  87  See, e.g., Christopher J. Costantini, Allocating Risk in Take-or-Pay Contracts: Are Force Majeure and Commercial Impracticability the Same Defense?, 42 SW. L.J. 1047, 1064 (1989).  88  See supra Part III.  89  See, e.g., RESTATEMENT (SECOND) OF CONTRACTS, supra note 18, ch. 11 Intro. Note.  90   possible with reasonable dispatch. Settlement of strikes or differences with workmen shall be wholly within the discretion of the party having the difficulty. The party having the difficulty shall notify the other party of any change in circumstances giving rise to the suspension of its performance and of its resumption of performance under this Agreement. The term \"force majeure\" shall include, without limitation by the following enumeration, acts of God, and the public enemy, the elements, fire, accidents, breakdowns, strikes, differences with workmen, and any other industrial, civil or public disturbance, or any act or omission beyond the control of the party having the difficulty, and any restrictions or restraints imposed by laws, orders, rules, regulations or acts of any government or governmental body or authority, civil or military.  94  Given their similarity to Impossibility doctrines, there is debate about whether such clauses usually add anything to contracts, or merely restate background doctrines.  95  Common notable features of force majeure clauses include: 1. External causation: \"a force majeure event cannot be caused by the party claiming force majeure\"; 96 2. Actual impediment: \"the party claiming force majeure bear[s] the burden of proving the force majeure event caused its damages\"; 97 3. Unavoidability: the clause does not \"include any cause which by the exercise of due diligence the party claiming force majeure is able to overcome . . .\"; 98 and 4. A requirement to give notice to the other party.  99  Force majeure clauses temporally transmit decision-making much like Impossibility doctrines: they provide a party with an option to suspend performance, conditional on external and unavoidable contingencies that make performance impossible or impracticable.  100  A force majeure clause recognizes the possibility of such contingencies without trying to plan for or enumerate all of them. Thus, parties set ex ante standards for deciding ex post when an option to suspend performance exists.  101  In so doing, force majeure clauses reduce  94     99  See Knoll & Bjorklund, supra note 96.  100  Cf. Geis, supra note 23, at <PHONE> (arguing that indefinite contracts can be usefully understood as containing embedded options).  101  Cf. Scott & Triantis, supra note 16, at 855 (\"Force majeure clauses typically provide that performance is excused in the event of specific contingencies (such as war, labor strikes, supply shortages, and government regulation that hinders the pressure for parties to enumerate ex ante precise contractual obligations under all possible contingencies. 102 \n B. Best Efforts and Similar Clauses \"Best efforts clauses typically come into play when one party wants the other to actively promote something, but for some reason the parties either cannot or choose not to spell out the details regarding what is involved and instead leave the issue to this 'best efforts' coverage.\" 103 Language requiring exercise of \"due diligence\" or \"'due professional skill and competence'\" can have similar effects.  104  Some representative examples 105 include: • A distributor promising to \"use its best efforts to promote and maintain a high volume of  [beer]  sales\"; 106 • A distributor promising to \"devote its best efforts to the sale and promotion of sales of the beverages\"; 107 • A motorcycle dealer promising to \"use his best efforts to sell Suzuki motorcycles.\" 108 Such clauses can also simply require parties to negotiate further.  109  However, best efforts clauses are not always enforceable:  110  A bare statement such as \"best efforts will be used\" or \"the parties will perform with due professional diligence\" will not be given effect unless the parties have provided other clues as to how to determine whether the duty has been met, such as through industry practices or past performance.  111  Thus, for example, in Proteus Books Ltd. v. Cherry Lane Music Co., a clause requiring parties to market books with \"due professional skill and competence\" was not \"void for vagueness\": 112 performance). But these clauses also identify excusing contingencies that fall within a vaguely stated category of factors beyond the control of the parties.\").  102     ] romises by an employer to pay 'parity' or 'according to industry standards' are not sufficiently definite to be enforceable.\").  112  See 873 F.2d at 508. [I]t is not fatal that the contract does not define the standards of due professional skill and competence. Reference to the managerial and marketing standards of the book publishing and distribution industries is sufficient to make the phrase readily understandable. A list of the acts that would constitute due professional skill and competence was therefore not necessary . . . .  113  By contrast, in Pinnacle Books, Inc. v. Harlequin Enterprises Ltd.,  114  the court held that a clause requiring parties (an author and a publisher) to exercise \"their best efforts\" to reach an agreement on a contract renewal 115 was unenforceable for vagueness: [T]he \"best efforts\" clause is unenforceable because its terms are too vague. \"Best efforts\" or similar clauses, like any other contractual agreement, must set forth in definite and certain terms every material element of the contemplated bargain. It is hornbook law that courts cannot and will not supply the material terms of a contract. Essential to the enforcement of a \"best efforts\" clause is a clear set of guidelines against which the parties' \"best efforts\" may be measured. The performance required of the parties by a \"best efforts\" clause may be expressly provided by the contract itself or implied from the circumstances of the case. In the case at bar, there simply are no objective criteria against which either [the publisher] or [the author]'s efforts can be measured.  116  When they are enforceable, best efforts clauses provide stability by binding parties to perform in accordance with a predetermined set of standards, as applied to future circumstances.  117  Thus, parties attain a binding agreement while avoiding the need to demand ex ante specific actions in all future circumstances.  118  Best efforts and similar standards thus provide flexibility that allows meaningful commitments to survive even when circumstances change. \n C. Standards and Stability Generally Standards-based contracts are useful when parties can agree to broad principles for guiding action, but for some reason 119 cannot agree on precise (i.e., rule-based) language. 120 \"Flexible contracting can [also] foster trust and collaboration, ultimately creating a more successful contracting relationship.\"  121  Of course, parties must take care to manifest their intent to be bound and guide potential adjudicators with reasonably certain standards.  122  Enforcement of indefinite terms is unpredictable and still heavily litigated.  123  113 Id. at 509. 114 519 F. Supp. 118 (S.D.N.Y. 1981).  115  See id. at 120.  116  Id. at 121 (citations omitted).  117  See generally Scott & Triantis, supra note 16 (explaining that parties might agree to standards rather than rules to shift costs to the back end of the contractual relationship).  118  See generally id.  119  For reasons why this might occur, see supra Part I.  120  See Scott, supra note 110, at 1649-55.  121  Epstein, supra note 27, at 335; see also Goetz & Scott, supra note 110. Still, enforceable standards create stability by enabling ex post 124 or ex tempore (\"at the time\")  125  rulemaking with the benefit of information learned after the time of the initial agreement.  126  They are thus most appropriate when parties can agree upon reasonably certain standards to which they wish to be held, but the precise content of which is not satisfactorily specifiable ex ante. 127  122  See, e.g., Ladas, 23 Cal. Rptr. 2d at 815 (\"[P]romises by an employer to pay 'parity' or 'according to industry standards' are not sufficiently definite to be enforceable.\"); Cobble Hill Nursing Home, Inc. v. Henry & Warren Corp., 548 N.E.2d 203, 206 (N.Y. 1989) (\"Before rejecting an agreement as indefinite, a court must be satisfied that the agreement cannot be rendered reasonably certain by reference to an extrinsic standard that makes its meaning clear.\"); Joseph Martin, Jr., Delicatessen, Inc. v. Schumacher, 417 N.E.2d 541, 544 (N.Y. 1981) (\"It certainly would have sufficed, for instance, if a methodology for determining the rent was to be found within the four corners of the lease, for a rent so arrived at would have been the end product of agreement between the parties themselves. Nor would the agreement have failed for indefiniteness because it invited recourse to an objective extrinsic event, condition or standard on which the amount was made to depend.\"); Fischer v. CTMI, L.L.C., 479 S.W.3d 231, 239-40 (Tex. 2016); RESTATEMENT (SECOND) OF CONTRACTS, supra note 18, § 33(1) (\"Even though a manifestation of intention is intended to be understood as an offer, it cannot be accepted so as to form a contract unless the terms of the contract are reasonably certain.\"); Scott, supra note 110, at 1649-61.  123  See Geis, supra note 23, at 1683-86.  124  See Kaplow, supra note 16.  125  See Verstein, supra note 16.  126  See PARISI & FON, supra note 16, at 11; Kaplow, supra note 16, at 585-86.  127  This is especially likely to be true when conduct is not homogeneously evaluable-i.e., when changing circumstances also change the desirability of some fixed conduct. See Luppi & Parisi, supra note 90, at 52 (\"Volatility of the external environment creates an increased opportunity for obsolescence of legal rules. This in turn would render standards preferable to specific rules . . . .\"); Ehrlich & Posner, supra note 90, at 270 (\"The problems of overinclusion and underinclusion [from using rules rather than standards] are more serious the greater the heterogeneity (or ambiguity, or uncertainty) of the conduct intended to be affected. If speeding were a homogeneous phenomenon-as it would be, for example, if driving at a speed of more than 70 miles per were always unreasonably fast and driving at a lesser speed never unreasonably fast-it could be effectively proscribed by a uniform speed limit with no residual prohibition of unreasonably fast driving. But speeding is in fact heterogeneous. It includes some driving at very low speeds and excludes some very fast driving, depending on a multitude of particular circumstances. A single speed limit or even a large number of separate speed limits exclude a great deal of conduct that is really speeding and include a great deal that is not really speeding.\"). Obviously, the development and deployment of Advanced AI might very well entail such volatility. \n V. Renegotiation Renegotiation 128 provides another potential tool for assimilating new information into existing agreements. This tool is particularly prevalent in international transactions.  129  In an international setting, renegotiation is typically appropriate when, e.g., \"a subsequent event not foreseen by the parties . . . has rendered the obligation of one party so onerous that it may be assumed that if he had contemplated its occurrence, he would not have made the contract.\"  130  Readers will, of course, recognize this as similar to the circumstances warranting invocation of Impossibility.  131  Parties can account for renegotiation either implicitly or explicitly. In the former case, renegotiation operates like Impossibility: certain fundamentally changed circumstances implicitly grant the disadvantaged party an option to renegotiate material contract terms.  132  Parties can also provide for renegotiation explicitly by specifying circumstances that trigger a duty to renegotiate.  133  As with any contractual provision, the triggering circumstances can be more or less precise-i.e., more rule-like or standard-like.  134  Like all contractual standards,  135  insufficiently precise conditions risk rendering a renegotiation provision unenforceable.  136  There is, of course, an obvious risk to renegotiation clauses: the possibility that parties will not reach an agreement 141 despite good-faith efforts to do so. 142 \"Some tribunals have concluded that when parties are unable to reach an agreement in renegotiation, there is no breach of contract because '[a]n obligation to negotiate is not an obligation to agree.  '\" 143  If parties fail to reach an agreement during renegotiation, the contractual relationship can take a number of paths. The choice ultimately depends on the relevant governing law and adjudicative body. When a hardship (e.g., force majeure) clause triggers renegotiation, sometimes suspension or termination of the contract results.  144  More commonly, however, the contract explicitly calls for the dispute to be then submitted to a third party (especially an arbitrator or arbitral body) for resolution.  145  The next Part deals with such alternatives. But these are by no means the only options, and parties could agree ex ante on many other methods of resolution (including those previously outlined in this Report) in case negotiations fail.  146  Obviously, the likely outcome in case of failure enormously effects parties' comparative bargaining power during renegotiation.  147  Renegotiation allows parties to specify contractual duties and rights in changed circumstances with the benefit of increased knowledge. However, they must be used with caution (and not just because the risk of negotiation failure creates a risk of an open term). If one party's bargaining power radically changes between ex ante agreement and ex post renegotiation (e.g., if their best alternative to a negotiated agreement improves), then the other party might not be able to secure similarly favorable terms.  148  However, careful  141  See id. at 1367-68.  142  \"Party liability for damages arising from a breach of the contractual obligations derived from the [renegotiation] clause only comes into consideration in exceptional cases.\" Id. at 1368. This is appropriate only where the non-agreement is proven to be caused by a gross breach of obligation in bad faith by the other side. This could be the case, for example, where proceedings are unjustifiably delayed, when negotiations are intentionally obstructed or where proposals by one side are obviously rejected for reasons other than normal business judgement. Id. at 1369 (footnotes omitted).  143  Gotanda, supra note 129, at 1465 (quoting Kuwait v. American Independent Oil Company (Aminoil), 21 I.L.M. 976, 1004 (Arb. Trib. 1982)). But see Berger, supra note 129, at 1367 (2003) (\"In the interests of an efficiency-oriented interpretation of such clauses, German law provides that an obligation of the parties to reach agreement exists in this respect if the adjustment criteria and adjustment aim have been defined to sufficient clarity.\"). Relatedly, so-called \"agreements to agree\" are not binding in American law. See JOHN BOURDEAU ET AL., 17A AM. JUR. 2D CONTRACTS § 38.  144  PETER, supra note 128, at 248.  145  See id. at 248-58; Berger, supra note 129, at 1360, 1370-78.  146  Of course, in some sense all legal disputes take place against the backdrop of possible litigation. For example, if the alternative to an agreement by renegotiation is a contractual standard, see generally supra Part IV, parties might still disagree over what such standards concretely require. In such a case, they might therefore refer the dispute to a court or arbiter.  147  See Guggenheimer, supra note 27, at 208 n.313. This is an instance of the more general idea from negotiation theory that a party's best alternative to a negotiated agreement (\"BATNA\") affects her bargaining position. See, e.g., Russell Korobkin, Bargaining Power as Threat of Impasse, 87 MARQ. L. REV. 867, 868-69 (2004).  148  Indeed, this is often the motivation for renegotiations in the international investment context. See, e.g., Asante, supra note 129, at 408 (\"[A] number of transnational contracts were concluded as incidents of the colonial system in which metropolitan companies were offered privileged investment interests in the colonies and accordingly given such grotesquely favourable terms as could hardly survive the collapse of colonialism. . . . In these circumstances, host agreement design can mitigate this risk by specifying default rules that restore the desired ex post bargaining position of each party. For example, a contract could stipulate that the (otherwise-)advantaged party will have to pay heavily in case negotiations fail. governments of newly independent countries consider renegotiation or restructuring of these arrangements as a legitimate part of the decolonisation process.\"). \n VI. Third-Party Resolution If parties fail to agree on contractual obligations, then third-party resolution offers a final way of harnessing ex post decision-making. Of course, this could include litigation. However, I will not focus on litigation here for two reasons. First, as most relevant here, litigation encompasses judicial enforcement of agreements containing other contractual tools detailed above. Discussing litigation here would therefore be largely duplicative. Secondly, avoiding litigation is often highly desirable due to its costs. This Part therefore focuses on a number of common 149 third-party adjudicative techniques that fall under the umbrella of \"alternative dispute resolution\" (\"ADR\"): 150 expert determination, 151 dispute boards, 152 and arbitration.  153  Note that these are non-exclusive; parties can consider providing for a multi-tiered, \"escalating\" ADR process that begins with negotiation and culminates (if all else fails) in arbitration or litigation.  154  A. Expert Determination \"Expert determination is an informal process that produces a binding decision.\"  155  As the term suggests, the main idea is that an expert settles the parties' dispute. Appraisal is one form of expert determination (\"ED\"); 156 ED is popular for resolving valuation disputes. 157 ED is also common where technical or scientific knowledge is necessary.  158  ED has a number of advantages over other forms of ADR, including: 159 • An expert's determination is binding; 160  149  The following list is by no means exhaustive. There is no reason, for example, that parties could not agree to be bound by a decision rendered by some third party outside of these mechanisms. However, the mechanisms detailed here are common and therefore often operate according to well-established rules and in the shadow of established legal principles and expertise. This often makes established third-party adjudicative mechanisms more stabilizing than ad hoc ones. However, in highly idiosyncratic transactions, deviation from established modes of third-party resolution might be necessary or desirable.  150     158  See id. (\"Outside of the valuation sphere, common business and industry areas where expert determination may be used for the purpose of obtaining an expert scientific or professional opinion include broadcasting and telecoms, IT, government procurement and PFI/PPP, energy and natural resources and banking and finance. Disputes involving long term contracts in these areas may require an expert to give his opinion on certain specialist or technical matters.\").  159  For a more extensive discussion of the pros and cons of ED, see id. § 2.  160  See id. • It allows for resolution by someone with specialized knowledge of the relevant subject; 161 • \"It is usually cheaper, quicker and less formal than arbitration or litigation;\" 162 and • \"There is, arguably, a greater chance of finality\" as compared to court or arbitral decisions.  163  Downsides include: • \"The law on expert determination is not as well developed as the law on other dispute resolution mechanisms such as arbitration;\" 164 • Less clear rules for the expert's jurisdiction, as compared to arbitration; 165 • Although binding, ED is not self-enforcing; 166 and • The finality of an ED also limits opportunity for appeal.  167  Parties must, of course, either agree on an expert when disputes arise  168  or \"provide that, if the parties cannot agree upon the identity of the expert, a specified body will identify a suitable candidate on the application of either party.\"  169  Parties can also require that the expert be independent. 170 161 See id.  162  Id.  163  See id.  164  Id.  165  See id.  166  See id. (\"Experts' decisions cannot generally be enforced without further court action or arbitration proceedings on the decision-whether domestically or internationally. An expert's decision becomes part of the contract such that, if a party fails to comply with the decision, then a further court judgment or arbitral award is likely to be required before a party is able to enforce the decision. There is no convention for the enforcement of an expert's decision abroad such as that which exists in relation to arbitral awards.\").  167  See id.  168  See id. § 4. [G]enerally, referring issues to a named expert is not advisable unless the dispute has already arisen. This is because a considerable amount of time may pass between the making of the original contract and the time a dispute arises and, in the interim, the nominee may have died, retired, become ill, have developed a conflict of interest, become unsuitable/unavailable or simply refuse to conduct the determination. Id.  169  Id. If there is no such provision, the court has no power (unlike in relation to the appointment of arbitrators) to appoint an expert, and there is a significant risk that the process will break down. It is therefore important to name the appointing authority accurately in the clause and to ensure (so far as possible) that it will be in a position to act as the appointing authority if called upon to do so in due course, and that it would also be willing to do anything else in relation to the expert determination which the parties might like it to do. Id.  170  See id. As mentioned above, an ED is generally not appealable. For example, under UK ED law, an expert's decision is likely to be overturned by courts only in very narrow circumstances.  171  Examples of such circumstances are: • Fraud or collusion; • Partiality; • Significant departure from parties' instructions; or • Failure to state reasons.  172  Parties can also contractually provide that courts may overturn an ED on \"manifest error\" 173 -a very high bar. \n B. Dispute Boards Dispute boards are usually project-specific adjudicative bodies that provide quick resolution to disputes related to the project.  174  They are especially common in construction and infrastructure projects 175 outside the US.  176  Their popularity has increased rapidly in recent decades.  177  Andrew Verstein also notes that dispute boards feature in adjudications of credit default swap obligations.  178  \"Dispute boards are panels of neutral experts, typically three, chosen by the parties and convened at the start of a . . . project.\"  179  Parties can provide for board opinions on a dispute to have varying degrees of bindingness, including: • Merely advisory;  180  • Binding in the interim (i.e., pending further adjudication);  181  • Binding unless a party formally expresses dissatisfaction with the result;  182  or • Binding and final.  183  Dispute boards are especially helpful when interpreting vague contractual standards (\"such as 'equitable,' 'reasonably anticipated,' or 'workmanlike quality'\") 185 in the context of a specific profession or industry.  186  Advantages of this tool include: • It establishes a culture of claim avoidance. • It may facilitate positive relations, open communication, trust and co-operation between the parties. • It can help settle issues promptly, before they escalate into disputes. • It can provide an informal forum with well-informed individuals, who are familiar with the project, to resolve disputes. • It is often cost effective. Resources can remain focused on the job, rather than concentrating on resolving disputes. • The determination will be influential in subsequent proceedings, if these are necessary. • The existence of the dispute board may influence the parties' behaviour, so that issues are dealt with and the number of disputes is minimised. In practice, this may only be true on the largest of projects, because of the \"standing costs\" of a dispute board.  187  Disadvantages include: • It can be costly. The parties are jointly liable for the direct costs of the board members plus any additional time spent resolving disputes. • Dispute board members may make a determination that is contractually or factually incorrect, or try to impose their own ideas on the parties. • The determination may be nothing more than a compromise between the parties' positions. • The dispute board's enquiry is limited and takes place without the opportunity for a proper, judicial examination of evidence. • The process is a \"claims review\" rather than dispute resolution, since the dispute board generally gets involved late in the process, after one party has prepared a detailed claim.  188  Note that several of these considerations might apply to ED as well. \n C. Arbitration Arbitration is a prominent tool of contemporary ADR. It is a contractual alternative to traditional litigation: 189 instead of submitting disputes to a court, parties submit them to an arbitral tribunal. However, arbitration differs from litigation in that parties can specify, in their arbitration clause, the rules, procedure, jurisdiction, applicable law, tribunal composition, and scope of arbitral proceedings (i.e., which questions the tribunal may \n Conclusion This Report outlined five common ways to stabilize agreements in the face of dramatically changing circumstances by allowing key binding decisions to be made in the future: • Options • Impossibility doctrines • Contractual standards • Renegotiation • Third-party resolution Knowledge of these tools will hopefully enable readers (especially those who are not lawyers) to better understand the options available to them. In short, they do not necessarily have to specify all rights and duties under all Advanced AI contingencies; the above mechanisms allow for binding-but-flexible agreements that can, hopefully, meaningfully survive even radically changed circumstances. V . RENEGOTIATION ................................................................................................................................... VI. THIRD-PARTY RESOLUTION .......................................................................................................... A. EXPERT DETERMINATION .................................................................................................................................................... B. DISPUTE BOARDS ................................................................................................................................................................... C. ARBITRATION .......................................................................................................................................................................... D. THIRD-PARTY RESOLUTION GENERALLY ......................................................................................................................... CONCLUSION .................................................................................................................................................. \n Table of Contents of . CONTRACTUAL IMPOSSIBILITY DOCTRINES ........................................................................................................................ 1. Impracticability..................................................................................................................................................................... 2. Frustration........................................................................................................................................................................... B. REBUS SIC STANTIBUS (\"THINGS THUS STANDING\") ...................................................................................................... C. IMPOSSIBILITY, REBUS, AND STABILIZATION .................................................................................................................... I. WHY AGREEMENT INCOMPLETENESS? ............................................................................................ A. TRANSACTION COSTS .............................................................................................................................................................. B. BOUNDED RATIONALITY ........................................................................................................................................................ C. INTERPRETIVE UNCERTAINTY ............................................................................................................................................... II. OPTION CONTRACTS .............................................................................................................................. III. IMPOSSIBILITY IN CONTRACT AND TREATY LAW ..................................................................... AIV. CONTRACTUAL STANDARDS ........................................................................................................... A. FORCE MAJEURE CLAUSES.................................................................................................................................................... B. BEST EFFORTS AND SIMILAR CLAUSES ............................................................................................................................... C. STANDARDS AND STABILITY GENERALLY ......................................................................................................................... \n See, e.g., Janice C. Griffith, Local Government Contracts: Escaping from the Governmental/Proprietary Maze, 75 IOWA L. REV. 277, 358 (1990); Detlev F. Vagts, Rebus Revisited: Changed Circumstances in Treaty Law, 43 COLUM. J. TRANSNAT'L L. 459, 459 (2005); George K. Walker, Sources of International Law and the Restatement (Third), Foreign Relations Law of the United States, 37 NAVAL L. REV. 1, 26-27 (1988). 70 Vienna Convention on the Law of Treaties art. 62, May 23, 1969, 1155 U.N.T.S. 331, https://perma.cc/XDU2-Y7KW. 71 Shalanda H. Baker, Climate Change and International Economic Law, 43 ECOLOGY L.Q. 53, 87 (2016). 72 Detlev F. Vagts, Book Review, 98 AM. J. INT'L L. 614, 615 (2004); see also Laurence R. Helfer, Exiting Treaties, 91 VA. L. REV. 1579, 1643 (2005). For accounts of two unsuccessful assertions of rebus in front of the International Court of Justice, see Baker, supra note 71, at 84-86. 73 Andrew Kull, Mistake, Frustration, and the Windfall Principle of Contract Remedies, 43 HASTINGS L.J. 1, 38 (1991). \n Id. (citing Gulf, Mobile & O.R.R. v. Illinois Central R.R., 128 F. Supp. 311 (N.D. Ala. 1954), 225 F.2d 816 (5th Cir. 1955); Martin v. Star Publishing Co., 126 A.2d 238 (Del. 1956); Powers v. Siats, 70 N.W. 2d. 344 (Minn. 1955); Helms v. B & L Investment Co., 198 S.E.2d 79 (N.C. App. 1973)). \n For useful background on the rules-versus-standards debate in contract and law design, see generally Barbara Luppi & Francesco Parisi, Rules versus Standards, in PRODUCTION OF LEGAL RULES 43 (Francesco Parisi ed., 2011); PARISI & FON, supra note 16, at 9-29; Isaac Ehrlich & Richard A. Posner, An Economic Analysis of Legal Rulemaking, 3 J. LEGAL STUD. 257 (1974); Louis Kaplow, A Model of the Optimal Complexity of Legal Rules, 11 J. L. ECON. & ORG.150 (1995); Kaplow, supra note 16. 91 See generally Mark P. Gergen, The Use of Open Terms in Contract, 92 COLUM. L. REV. 997 (1992); Ian R. Macneil, Contracts: Adjustment of Long-Term Economic Relations under Classical, Neoclassical, and Relational Contract Law, 72 NW. U. L. REV. 854, 866 (1978); Scott & Triantis, supra note 16. Note that there are plenty of ways for parties to explicitly allocate risk via rules. A buyer, for example, may agree to pay a specific amount above the cost of production via a \"cost-plus\" clause. See, e.g., RESTATEMENT (SECOND) OF CONTRACTS, supra note 18, ch. 11 Intro. Note. However, mechanisms like this are of less interest to this Report since they do not constitute the type of intertemporal decision transmission with which I am interested. 92 See, e.g., Gary B. Conine, The Prudent Operator Standard: Applications Beyond the Oil and Gas Lease, 41 NAT. RESOURCES J. 23 (2001); Robert A. Hillman, Court Adjustment of Long-Term Contracts: An Analysis Under Modern Contract Law, 1987 DUKE L.J. 1, 4 (\"[S]ome coal contracts include a 'gross inequities adjustment provision,' which requires the parties to negotiate in good faith to resolve 'inequities' resulting from economic conditions that the parties did not contemplate at the time they made their agreement.\");Verstein, supra note 16. 93 E.g., Allison R. Ebanks, Force Majeure: How Lessees Can Save Their Leases While the War on Fracking Rages On, 48 ST. MARY'S L.J. 857, 873 (2017). \n Langham-Hill Petroleum Inc. v. S. Fuels Co., 813 F.2d 1327, 1329 n.1 (4th Cir. 1987); see also 30 WILLISTON ON CONTRACTS § 77:31 (4th ed.) (discussing force majeure clauses). 95 Compare E. Air Lines, Inc. v. McDonnell Douglas Corp., 532 F.2d 957, 991 (5th Cir. 1976) (\"Because of the uncertainty surrounding the law of excuse, parties had good reason to resort to general contract provisions relieving the promisor of liability for breaches caused by events 'beyond his control.' Although the Uniform Commercial Code has ostensibly eliminated the need for such clauses, lawyers, either through an abundance of caution or by force of habit, continue to write them into contract.\") with P.J.M. Declercq, Modern Analysis of the Legal Effect of Force Majeure Clauses in Situations of Commercial Impracticability, 15 J.L. & COM. 213, 225 (1995) (\"Even in the absence of detailed wording, trade usage or the surrounding circumstances may indicate an intent to grant the seller a broader exemption than is provided by the U.C.C.\"). 96 Jocelyn L. Knoll & Shannon L. Bjorklund, Force Majeure and Climate Change: What is the New Normal?, 8 AM. C. CONSTRUCTION LAW. J. 2 (2014). 97 Id. 98 E.g., Gulf Oil Corp. v. F.P.C., 563 F.2d 588, 613 (3d Cir. 1977). \n See generally id. 103 Best Efforts Clauses, 24 No. 2 CORP. COUNS. Q. art. 1 (2008). 104 See id. (quoting Proteus Books Ltd. v. Cherry Lane Music Co., 873 F.2d 502, 508 (2d Cir. 1989)). 105 See id. 106 Bloor v. Falstaff Brewing Corp., 601 F.2d 609, 610 (2d Cir. 1979). 107 Joyce Beverages of New York, Inc. v. Royal Crown Cola Co., 555 F. Supp. 271, 273 (S.D.N.Y. 1983). 108 Am. Suzuki Motor Corp. v. Bill Kummer, Inc., 65 F.3d 1381, 1383 (7th Cir. 1995). 109 See, e.g., 2 LAW OF SELLING DESKBOOK § 23:2 (2018). 110 See, e.g., id.; Nellie Eunsoo Choi, Contracts with Open or Missing Terms Under the Uniform Commercial Code and the Common Law: A Proposal for Unification, 103 COLUM. L. REV. 50, 50-60 (2003); Charles J. Goetz & Robert E. Scott, Principles of Relational Contracts, 67 VA. L. REV. 1089, 1119-26 (1981); Robert E. Scott, A Theory of Self-Enforcing Indefinite Agreements, 103 COLUM. L. REV. 1641, 1647-61 (2003). 111 2 LAW OF SELLING DESKBOOK, supra note 109, § 23:2. But cf. Ladas v. California State Auto. Assn., 23 Cal. Rptr. 2d 810, 815 (App. 1993) (\"[P \n\t\t\t Richard A. Posner & Andrew M. Rosenfield, Impossibility and Related Doctrines in Contract Law: An Economic Analysis, 6 J. LEGAL STUD. 83, 88 (1977). 45 RESTATEMENT (SECOND) OF CONTRACTS, supra note 18, ch. 11 Intro. Note. \n\t\t\t Often discussed alongside renegotiation are \"adaptation clauses\": \"a group of contract provisions that allow contract changes by following an automatic or predetermined pattern or which are merely designed for the filling of gaps in \n\t\t\t Verstein, supra note 16, at 1898.186  See id. at <PHONE>.187  Dispute Boards: What Are Dispute Boards?, supra note 174, § 8.188  Id.189  See Allen & Overy LLP, supra note 153, § 2.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Stable-Agreements.tei.xml", "id": "2d78b36e89413d4b339e76b16800359a"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Artificial intelligence (AI) is a potent general purpose technology. Future progress could be rapid, and experts expect that superhuman capabilities in strategic domains will be achieved in the coming decades. The opportunities are tremendous, including advances in medicine and health, transportation, energy, education, science, economic growth, and environmental sustainability. The risks, however, are also substantial and plausibly pose extreme governance challenges. These include labor displacement, inequality, an oligopolistic global market structure, reinforced totalitarianism, shifts and volatility in national power, strategic instability, and an AI race that sacrifices safety and other values. The consequences are plausibly of a magnitude and on a timescale to dwarf other global concerns. Leaders of governments and firms are asking for policy guidance, and yet scholarly attention to the AI revolution remains negligible. Research is thus urgently needed on the AI governance problem: the problem of devising global norms, policies, and institutions to best ensure the beneficial development and use of advanced AI. This report outlines an agenda for this research, dividing the field into three research clusters. The first cluster, the technical landscape, seeks to understand the technical inputs, possibilities, and constraints for AI. The second cluster, AI politics, focuses on the political dynamics between firms, governments, publics, researchers, and other actors. The final research cluster of AI ideal governance envisions what structures and dynamics we would ideally create to govern the transition to advanced artificial intelligence. Visit www.fhi.ox.ac.uk/govaiagenda to check for the most recent version of this paper. 1 This document received input from many contributors. The text was primarily written by Allan Dafoe. Several portions received substantial input from other individuals, most affiliated with the Governance of AI Program, noted where for each portion. This work draws from the body of thinking and insight in the community of scholars and scientists thinking about these issues. In particular, for comments, conversations, and related work this document benefits from Miles Brundage,", "authors": ["Allan Dafoe"], "title": "AI Governance: A Research Agenda", "text": "Introduction Transformative  \n Introduction Artificial intelligence is likely to become superhuman at most important tasks within this 2 beneficial AI. AI safety focuses on the technical questions of how AI is built; AI governance focuses on the institutions and contexts in which AI is built and used. Specifically, AI governance seeks to maximize the odds that people building and using advanced AI have the goals, incentives, worldview, time, training, resources, support, and organizational home necessary to do so for the benefit of humanity. To motivate this work it can be helpful to consider an urgent, though not implausible, hypothetical scenario. Suppose that in one year's time a leading AI lab perceives that profound progress may be on the horizon. It concludes that given a big push, in 6 to 24 months the lab is likely to develop techniques that would achieve novel superhuman capabilities in strategic domains. These domains might include lie detection, social-network mapping and manipulation, cyber-operations, signals and imagery intelligence, strategy, bargaining or persuasion, engineering, science, and potentially AI research itself. Despite our knowledge (in this scenario) that these technical breakthroughs are likely, we would have uncertainty about the details. Which transformative capabilities will come first and how they will work? How could successive (small) capabilities interact to become jointly transformative? How can one build advanced AI in a safe way, and how difficult will it be to do so? What deployment plans and governance regimes will be most likely to lead to globally beneficial outcomes? The AI governance problem is the problem of preparing for this scenario, along with all other high-stakes implications of advanced AI. The task is substantial. What do we need to know and do in order to maximize the chances of the world safely navigating this transition? What advice can we give to AI labs, governments, NGOs, and publics, now and at key moments in the future? What international arrangements will we need--what vision, plans, technologies, protocols, organizations--to avoid firms and countries dangerously racing for short-sighted advantage? What will we need to know and arrange in order to elicit and integrate people's values, to deliberate with wisdom, and to reassure groups so that they do not act out of fear? The potential upsides of AI are tremendous. There is little that advanced intelligence couldn't help us with. Advanced AI could play a crucial role solving existing global problems, from climate change to international conflict. Advanced AI could help us dramatically improve health, happiness, wealth, sustainability, science, and self-understanding.  6  The potential downsides, however, are also extreme. Let us consider four sources of catastrophic risk stemming from advanced AI. ❖ (1) Robust totalitarianism could be enabled by advanced lie detection, social manipulation, autonomous weapons, and ubiquitous physical sensors and digital footprints. Power and control could radically shift away from publics, towards elites and especially leaders, making democratic regimes vulnerable to totalitarian backsliding, capture, and consolidation. ❖ (2) Preventive, inadvertent, or unmanageable great-power (nuclear) war . Advanced AI could give rise to extreme first-strike advantages, power shifts, or novel destructive capabilities, each of which could tempt a great power to initiate a preventive war. Advanced AI could make crisis dynamics more complex and unpredictable, and enable faster escalation than humans could manage, increasing the risk of inadvertent war. ❖ ❖ (4) Finally, even if we escape the previous three acute risks, we could experience systematic value erosion from competition , in which each actor repeatedly confronts a steep trade-off between pursuing their final values or pursuing the instrumental goal of adapting to the competition so as to have more power and wealth. 9 These risks can be understood as negative externalities: harms from socio-technical developments that impact individuals other than those responsible for the developments. These externalities are especially challenging to manage as they may be extreme in magnitude, complex and hard to predict, and they will spill across borders and generations. Building the right institutions (including norms and political arrangements) is plausibly close to a necessary and sufficient condition to adequately address these risks. With the right institutions these risks can be radically reduced and plausibly eliminated. Without them, it may be that nothing short of a technical miracle will be sufficient to safely navigate the transition to advanced AI systems. \n Transformative AI Stepping back from this scenario, research on AI governance considers AI's most transformative potential capabilities, dynamics, and impacts. The stakes could be extreme: absent an interruption in development, AI this century is likely to be sufficiently transformative to \"precipitate a transition comparable to (or more significant than) the agricultural or industrial revolution.\" Given our current uncertainties about which 10 11 capabilities will have the greatest impacts, however, it can be useful to attend to a broad range of potentially transformative capabilities and dynamics. Accordingly, we focus on AI could transform international security by altering key strategic parameters, such as the security of nuclear retaliation, the offense-defense balance, the stability of crisis These transformative innovations and impacts may arise gradually over the coming decades, facilitating our anticipation and governance of their arrival. But they may also emerge more suddenly and unexpectedly, perhaps due to recursive self-improvement, advances in especially potent general purpose applications or even general intelligence, or overhang of computational power or other crucial inputs. The speed, suddenness, and predictability of the arrival of new capabilities will shape the character of the challenges we will face. From a more long-term and abstract perspective, the emergence of machine superintelligence (AI that is vastly better than humans at all important tasks) would enable revolutionary changes, more profound than the agricultural or industrial revolutions. The second cluster concerns AI politics , which focuses on the political dynamics between firms, governments, publics, researchers, and other actors, and how these will be shaped by and shape the technical landscape. How could AI transform domestic and mass politics ? Will AI-enabled surveillance, persuasion, and robotics make totalitarian systems more capable and resilient? How will countries respond to potentially massive increases in inequality and unemployment, and how will these responses support or hinder other global governance efforts? When and how will various actors become concerned and influential (what could be their \"AI Sputnik\" moments)? How could AI transform the international political economy ? Will AI come to be seen as the commanding heights of the modern   20  Unjammed bottlenecks and overhangs are closely related perspectives on complements, focusing either on the last necessary input or an already achieved necessary input. For example, consider the progress function f(X,Y, Z)=MIN(X,Y,Z). If at baseline X=0, Y=1, Z=1 then progress in X from 0 to 1 would represent an unjammed bottleneck. If at baseline X=0, Y=0, Z=1, then Z could be regarded as a form of overhang. understanding could make it cost-effective to efficiently create massive datasets for a variety of purposes from the internet, creation of these datasets could improve machine understanding of how many task domains in the world relate to each other, which could improve transfer learning between those domains, which could further improve natural language understanding. ❖ (2b) Rapid progress in a crucial bottleneck/complement of AI research. For example, we have seen sudden jumps in capabilities from the provision of a single large training dataset for a particular task. Similarly, the generation of a crucial training set for a generally applicable task could lead to a broad front of progress. ❖ (3) Substantial AI advances on tasks crucial for future AI R&D, permitting highly 21 recursive self-improvement. This might lead to an endogenous growth positive feedback process, sometimes called an \"intelligence explosion\", where each 22 generation of AI accelerates the development of the subsequent generation. ❖ (4) Radical increases in investment in AI R&D. ❖ (5) A large ratio of R&D costs to execution costs, so that once a particular capability is achieved it could be massively deployed. For example, the learning process could be highly compute intensive (such as with genetic algorithms), but once trained that same compute could be used to run thousands of instances of the new algorithm. Some arguments for rapid general progress have been articulated by Eliezer Yudkowsky,  23  Nick Bostrom, and MIRI. Some arguments against (spatially local) rapid general progress Ben Goertzel, and is implicit to most mainstream perspectives. The skeptical position draws support from the fact that AI progress and technological progress tends to be gradual, piecemeal, uneven, and spatially diffuse. If this remains true for AI then we should expect some transformative capabilities to come online far before others. \n Kinds, Capabilities, and Properties of AI AI could be transformative in many ways. We should systematically think through the kinds of AI that could be developed, and what their capabilities and properties might be. For 30 scenarios where progress is not rapid and broad, it will also be useful to articulate probable sequences in AI capabilities, or necessary achievements, prior to particular kinds of transformative AI. Some examples of potentially transformative capabilities include AI that is superhuman in, or otherwise transformative of, particular areas such as cybersecurity, autonomous weapons, surveillance, profiling, lie-detection, persuasion and manipulation, finance, strategy, engineering, manufacturing, and other areas of science and technology. Such AI, if arriving unbundled from other transformative capabilities, is often called \"narrow AI\". In addition to producing new capabilities, AI could be transformative through incremental effects, such as incremental changes in the costs or performance of existing capabilities, to the point that it transforms industries and world order. 31 29 See also the reading list, compiled by Magnus Vinding, on arguments against the hard take-off' hypothesis: https://magnusvinding.com/2017/12/16/a-contra-ai-foom-reading-list/ .  30  There is some work on \"kinds of intelligence\" that may speak to this. For an informal introduction, see Shanahan, Murray. \"Beyond Humans, What Other Kinds of Minds Might Be out There?\" Aeon , October 19, 2016. https://aeon.co/essays/beyond-humans-what-other-kinds-of-minds-might-be-out-there ; Shanahan, Murray. \"The Space of Possible Minds\" EDGE, May 18, 2018. https://www.edge.org/conversation/murray_shanahan-the-space-of-possible-minds . See also the CFI project on 'Kinds of Intelligence', at http://lcfi.ac.uk/projects/kinds-of-intelligence/ , and specifically José Hernández-Orallo, \"The Measure of All Minds\", 2017, Cambridge University Press, http://allminds.org/. Also see NIPS 2017 symposium: http://www.kindsofintelligence.org/ . 31 A discussion of the positive aspects of this is in Harford, Tim.\"What We Get Wrong about Technology.\" FT Magazine , July 17, 2017. https://www.ft.com/content/32c31874-610b-11e7-8814-0ac7eb84e5f1 . Negative possibilities also exist. For example, even without any specific capabilities that are especially transformative or novel, AI and associated trajectories could displace sufficient workers to generate a political economic crisis on the scale of the Great Depression. \"I suspect that if current trends continue, we may have a third of men between the ages of 25 and 54 not working by the end of this half century, because this is a trend that shows no sign of decelerating. And that's before we have ... seen a single driver replaced [by self-driving vehicles] ..., not a trucker, not a taxicab driver, not a delivery person. ... And yet that is surely something that is en route.\" Quoted in Matthews, Christopher. \"Summers: Automation is the middle class' worst enemy.\" Axios , June 4, 2017. https://www.axios.com/summers-automation-is-the-middle-class-worst-enemy-<PHONE>-754facf2-aaca-4788-9a41-38f87f b0dd99.html . To date AI systems remain narrow , in the sense that a trained system is able to solve a particular problem well, but lacks the ability to generalize as broadly as a human can. Further, advances in AI capabilities are highly uneven, relative to the distribution of human capabilities, and this trend seems likely to persist: game playing algorithms are vastly superhuman at some games, and vastly subhuman at others. AI systems today are 32 sometimes analogized as \"idiot savants\": they vastly outperform humans at some tasks, but are incompetent at other \"simple\" adjacent tasks. AI systems are approaching or are now superhuman at translating between languages, categorizing images, recognizing faces, and 33 driving cars, but they still can't answer what seem like simple common-sense questions such as Winograd Schemas. It may be the case that many kinds of TAI will arrive far before AI has achieved \"common sense\" or a child's ability to generalize lessons to a new task domain. Many thinkers, however, think the opposite is plausible. They reason that there is plausibly some faculty of general intelligence, some core cognitive module, some common factor to most kinds of \"few-shot\" learning (learning from only a few examples). This general intelligence, once achieved at even merely the level of a four-year-old human, would enable AI systems to be built that quickly learn in new domains, benefiting from and directing their superhuman memory, processing speed, sensor arrays, access to information and wealth of stored information, and library of specialized systems. This artificial general intelligence (AGI)--AI that can reason broadly across domains--could then rapidly catalyze progress across the task space; this is sometimes called \"seed AGI\". The concept of AGI is more strategically relevant to the extent that (1) the 34 concept maps onto a cluster of capabilities that come as a bundle (likely as a consequence of  32     3 , p. 531.  33  For many definitions of the task, but not all.  34  Yudkowsky provides a helpful overview of the concept of general intelligence here: Yudkowsky, Eliezer. \"General Intelligence.\" Arbital , n.d. https://arbital.com/p/general_intelligence/ . See also Goertzel, Ben. \"Artificial general intelligence: concept, state of the art, and future prospects.\" Journal of Artificial General Intelligence 5.1 (2014): 1-48. Note that AGI, as defined in this document and by Yudkowsky and Goertzel, is conceptually distinct from broad human level capabilities. One could in principle have an AGI with sub-(adult)-human reasoning, or non-AGI systems with many superhuman capabilities. However, it does seem plausible that in practice AGI will be a necessary and sufficient condition for human-level capabilities in nearly all domains, given (1) the ability of the general intelligence to call on all the other existing superhuman assets of machine intelligence, and (2) the vast array of problems seeming to require general intelligence--there is scarce data and extreme interdependencies with other domains--that are otherwise unlikely to be solved by narrow AI. Note that \"human-level AI in X\", \"high human-level AI in X\", and \"superhuman AI in X\" can be used to characterize narrow AI systems. general reasoning), (2) AGI has transformative implications, such as igniting rapid general progress, (3) AGI arrives early in the sequence of transformative capabilities.  35  We would like to know more about the probable strategic properties of novel capabilities and kinds of AI. For example, could they enhance cooperation by giving advice, by mediating or arbitrating disputes, by identifying gains-from-cooperation amongst strangers? Could AI and cheap surveillance enable robust monitoring of compliance to agreements, and cryptographic systems that protect participants from exposing their sensitive information? Could AI enable overcoming commitment problems through binding costly commitments that are 36 hard-coded into AI-adjudicated contracts? To what extent will AI enabled capabilities be defense-biased (vs offense-biased), defined here as costing relatively more to attack than to defend, for a given goal? Broadly speaking, defense-biased technology makes a multipolar world more stable. To what extent will new technologies be destruction-biased , defined 37 here as making it relatively easy to destroy value (but potentially hard to capture value)? Do new AI capabilities provide first-mover advantages , so that actors have (economic or military) incentives to develop and deploy them quickly? An extreme form of power 38 advantage, which may be more likely from first-mover advantages and offense bias, is decisive strategic advantage : an advantage sufficient to \"achieve complete world domination\". The strategic character, and the perceived strategic character, of future 39 technology will shape the international landscape, determining how secure or vulnerable are great powers under the status quo, and how able they are to cooperate to overcome commitment problems and the security dilemma.  35  For example, contra (1) it could be that general reasoning comes in different flavors, and AI becomes vastly superhuman at some forms while still remaining subhuman at others. Contra (3), AGI could plausibly be much harder to achieve than super-surveillance AI, super-hacking AI, even super-inventing AI (e.g. with circuit design). (2) seems plausible.  36  Particularly between great powers, who otherwise lack a powerful legal structure within which to make commitments.  37  For example, Yann LeCunn (NYU Ethics of AI Conference) stated that he believes narrow defensive AI systems will dominate general AI systems, because of specialization; his logic implies that narrow offensive AI systems should also dominate general AI systems, suggesting a world where general AI cannot flourish without massive narrow AI defenses. Eric Schmidt (2:52:59 during talk at Future of War Conference ) conjectured that cyber-AI systems will dominate on the defense. (For discussion of near-term defense vs offense bias, see Brundage, Miles, Shahar Avin, et al. \"The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation.\" ArXiv:1802.07228  [Cs]  , February 20, 2018. http://arxiv.org/abs/1802.07228 .)  38  There are several kinds of first mover advantage, such as from the first to attack, or the first to develop a capability. Both can be understood as a form of offense bias, though there are subtleties in definition related to what kinds of symmetry are presumed to be present.  39  Bostrom. Superintelligence ; p 96. These questions of the potential strategic properties of AI can also be framed in a more general way. To what extent will (particular kinds of) AI be, or have the option of being made to be, transparent , stabilizing/destabilizing , centralizing/decentralizing of power, politically valenced (towards authoritarianism or democracy), or wisdom-enhancing (advisor AIs)? How likely is it that we will get some of these (e.g. wisdom AI) before others (e.g. decisive cyber first strike AI)? In researching the possible strategic capabilities of AI, we must also ask how far our estimates and models can be relied upon.  Will    \n Other Strategic Technology Many other novel technologies could play a strategic or transformative role. It is worth studying them to the extent that they pose transformative possibilities before TAI, that they shape key parameters of AI strategy on the road to TAI, or that they represent technological opportunities that could be unlocked by a super R&D AI. These technologies include: atomically precise manufacturing, cheap and agile robotics, synthetic biology, genetic and cognitive enhancement, cyber-innovations and dependencies, quantum computing, ubiquitous and potent surveillance, lie detection, and military capabilities such as anti-missile defense, hypersonic missiles, energy weapons, ubiquitous subsea sensor networks, etc. \n Assessing AI Progress The previous section, Mapping Technical Possibilities, tried to creatively envision longer-run transformative possibilities and the character of AI. This section seeks to be more precise and A modeling strategy is to look for robust trends in a particular input, such as compute/$.  43  The canonical hardware trend is Moore's Law. Kurzweil and Nordhaus observe an 44 45 46 impressively consistent exponential trend in computing performance given cost, beginning before Moore's Law. From this some argue that the trend will continue. AI Impacts finds the 47 recent doubling time in FLOPS/$ to be about 3-5 years, slower than the 25 year trend of 1.2 years. Amodei and Hernandez note that over the past five years there appears to be an 48 exponential increase in the total compute used to train leading AI applications, with a 3.5 month doubling time.  49  More complex approaches could try to build causal and/or predictive models of AI progress (on particular domains) as a function of inputs of compute, talent, data, investment, time, and indicators such as prior progress and achievements (modeling the \"AI production function\"). To what extent does performance scale with training time, data, compute, or other fungible assets? What is the distribution of breakthroughs given inputs, and what is the existing and 50 likely future distribution of inputs? How quickly can these assets be bought? How easy is it to enter or leapfrog? Modeling these inputs may yield insights on rates of progress and the key factors which slow or expedite this. What do these models imply for likely bottlenecks in progress? Does it seem  43  Such an approach works best when we can (1) credibly extrapolate the trend in the input (which may not be true if there is a change in underlying dynamics), and (2) can map the input to outcomes that matter (which may not be true for lots of reasons).  44   likely that we will experience hardware or insight overhang? Put differently, how probable 51 is it that a crucial input will suddenly increase, such as with algorithmic breakthroughs, implying a greater probability of rapid progress? More generally, from the perspective of developers, how smooth or sudden will progress be? Articulating theoretically informed predictions about AI progress will help us to update our models of AI progress as evidence arrives. The status quo involves experts occasionally making ad-hoc predictions, being correct or mistaken by unquantified amounts, and then possibly updating informally. A more scientific approach would be one where explicit theories, or at least schools of thought, made many testable and comparable predictions, which could then be evaluated over time. For example, can we build a model that predicts time until super-human performance at a task, given prior performance and trends in inputs? Given such a model, we could refine it to assess the kinds of tasks and contexts where it is likely to make especially good or bad predictions. From this can we learn something about the size of the human range in intelligence space, for different kinds of tasks? If our models 52 are accurate then we would have a useful forecasting tool; if they are not then we will have hard evidence of our ignorance. We should build models predicting other strategically relevant parameters, such as the ratio of training compute costs to inference/execution compute costs, or more generally the ratio of R&D costs to execution costs (costs of running the system). 53 \n Forecasting AI Progress Using the above measurements and models, and with expert judgment, to what extent can we forecast the development of AI (inputs and performance)? There are several desiderata for good forecasting targets. Given such forecasting efforts we could ask, how well calibrated and accurate are different groups of experts and models for different kinds of forecasting problems? How best can we elicit, adjust, and aggregate expert judgment? How different are  51  For example, a relatively simple algorithmic tweak generated massive improvements in Atari game playing. It is plausible that there are many other such easily implementable algorithmic tweaks that an AI could uncover and implement. Bellemare, Marc G., Will Dabney, and Rémi Munos. \"A distributional perspective on reinforcement learning.\" ArXiv:1707.06887  [Cs]  , July 21, 2017. https://arxiv.org/abs/1707.06887 .  52  Alexander, Scott. \"Where The Falling Einstein Meets The Rising Mouse.\" Slate Star Codex (blog), August 3, 2017. http://slatestarcodex.com/2017/08/02/where-the-falling-einstein-meets-the-rising-mouse/ .  53  These parameters are relevant to the scale of deployment of a new system, to predicting the kinds of actors and initiatives likely to be innovating in various domains, and to other aspects of AI governance. the problems of near-term and long-term forecasting, and to what extent can we use lessons from or performance in near-term forecasting to improve long-term forecasts? Near-term forecasting work is currently being done by Metaculus. Many surveys have asked 54 untrained and uncalibrated experts about near-and long-term forecasts. It can also be productive to evaluate previous forecasting efforts, to see how well calibrated they are, and if there are conditions that make them more or less accurate. 55 \n AI Safety 56 \n The Problem of AI Safety AI Safety focuses on the technical challenge of building advanced AI systems that are safe and beneficial. Just as today a lot of engineering effort goes into ensuring the safety of deployed systems--making sure bridges don't fall down, car brakes don't fail, hospital procedures administer the correct medications to patients, nuclear power plants don't melt, and nuclear bombs don't unintentionally explode --so it is plausible that substantial effort will be 57 required to ensure the safety of advanced and powerful AI systems. Relatively simple AI systems are at risk of accidents and unanticipated behavior. Classifiers are known to be fragile and vulnerable to subtle adversarial attacks. One military report characterizes AI (specifically deep learning) as \"weak on the 'ilities'\", which include 58 reliability, maintainability, accountability, verifiability, debug-ability, fragility, attackability. AI can behave in a manner that is not foreseen or intended, as illustrated by Microsoft's failure to anticipate the risks of its \"Tay\" chatbot learning from Twitter users to make  54   offensive statements. Complex systems, especially when fast-moving and tightly coupled, can lead to emergent behavior and 'normal accidents'. The 2010 \"flash crash\" is illustrative, in 59 which automated trading algorithms produced 20,000 \"erroneous trades\" and a sudden trillion dollar decline in US financial market value; this undesired behavior was stopped not by real-time human intervention but by automated safety mechanisms.  60  The previous kinds of accidents arise because the AI is \"too dumb\". More advanced AI systems will overcome some of these risks, but gain a new kind of accident risk from being \"too clever\". In these cases a powerful optimization process finds \"solutions\" that the researchers did not intend, and that may be harmful. Anecdotes abound about the 61 surprising routes by which artificial life \"finds a way\", from a boat-racing AI that reward-hacked by driving in circles, to a genetic algorithm intended to evolve an oscillating 62 circuit that hacked its hardware to function as a receiver for radio waves from nearby computers, to a tic-tac-toe algorithm that learned to defeat its rivals using a memory bomb.  63 64  Future advances in AI will pose additional risks (as well as offer additional opportunities for safety). AI systems will be deployed in ever more complex and consequential domains. They will be more intelligent at particular tasks, which could undermine the value of human oversight. They will begin to acquire models of their environment, and of the humans and institutions they interact with; they will gain understanding of human motivation, be able to observe and infer human affect, and become more capable of persuasion and deception. As systems scale to and beyond human-level (in particular dimensions), they may increasingly be able to intelligently out-maneuver human built control systems. This problem is analogous to the problem of alignment in capitalism (of \"avoiding market failures\"): how to build a legal and regulatory environment so that the profit motive leads firms to produce social value. History is replete with examples of large negative externalities caused by firms perversely optimizing for profit, from fraudulent profits produced through 'creative' accounting (e.g. Enron), to policies that risk disaster or generate pollution (e.g. Deepwater Horizon oil spill), to firms that actively deceive their investors (e.g. Theranos) or regulators (e.g. Volkswagen emissions scandal). These scandals occur despite the existence of informed humans with \"common sense\" within the corporations, and the governance institutions being capable of comparable intelligence. Powerful AI systems may lack even that common sense, and could conceivably be much more intelligent than their governing institutions. \n AI Safety as a Field of Inquiry Much of the initial thinking about AI safety was focused on the challenge of making hypothesized human-level or superhuman AI safe. This line of inquiry led to a number of important insights.  65  1. Orthogonality thesis: Intelligent systems could be used to pursue any value system. 66  65   2. Instrumental convergence: systems will have instrumental reasons for acquiring power and resources, maintaining goal integrity, and increasing its capabilities and intelligence. 3. Empirical safety tests may not be sufficient. Human overseers may not have the capacity to recognize problems due to the system's complexity, but also its ability to intelligently model and game the oversight mechanism. 4. Formalizing human preferences is hard. When such a formal statement is fed as the goal into powerful and intelligent systems, they are prone to fail in extreme ways. This failure mode is a trope in Western literature, as per Midas' curse, the Sorcerer's Apprentice, the Monkey's Paw, and the maxim to be careful what one wishes for.  67  5. There are many ways control or alignment schemes could catastrophically and irreversibly fail, and among the most dangerous are those we haven't thought of yet. The above framing adopts the lens of AI accident risks : the risks of undesired outcomes arising from a particular, intentionally designed, AI system (often highly intelligent). There is another, relatively neglected framing, of AI systemic risks : the risks of undesired outcomes--some of which may be very traditional--that can emerge from a system of competing and cooperating agents and can be amplified by novel forms of AI. For example, AI could increase the risk of inadvertent nuclear war, not because of an accident or misuse , but because of how AI could rapidly shift crucial strategic parameters, before we are able to build up compensating understandings, norms, and institutions.  68  AI safety can thus be understood as the technical field working on building techniques to reduce the risks from advanced AI. This includes the ultimate goals of safety and alignment of superintelligent systems, the intermediate goals of reducing accident, misuse, and emergent risks from advanced systems, as well as near-term applications such as building self-driving car algorithms that are sufficiently safe, including being resilient to en-masse terrorist hacks.  67   As evidence of the importance of this field, when AI researchers were surveyed about the likely outcome of super-human AI, though the majority believe it is is very likely to be beneficial, the majority of respondents assign at least a 15% chance that superhuman AI would be \"on balance bad\" or worse, and at least a 5% chance it would be \"extremely bad (e.g. human extinction)\". The goal of AI safety is to provide technical insights, tools, and solutions 69 for reducing the risk of bad, and especially extremely bad, outcomes. As AI systems are deployed in ever more safety-critical and consequential situations, AI researchers and developers will increasingly confront safety, ethical, and other challenges. Some solutions to these challenges will be one-off, local patches. For example, Google's solution to misclassifying images of black people as \"gorillas\" was to simply remove \"gorilla\" and similar primate categories from its service. This kind of patch will not scale or 70 generalize. We would prefer to find solutions that are more foundational or generalizable, and thus more plausibly contribute to scalably safe and beneficial AI. Broadly, for particular systems we will want them to have various desirable properties, such as the following (drawing from Everitt et al's 2018 framework): ❖ Reliability and Security , so that the system behaves as intended in a wide range of situations, including under adversarial attack.  71  ❖ Corrigibility , so that the system is optimally open to being corrected by a human overseer if it is not perfectly specified/trained. Candidate methods include by To some extent these approaches can be trialed and developed in concrete near-term settings. 84 \n The Implications of AI Safety for AI Governance For the purposes of AI governance it is important that we understand the strategic parameters relevant to building safe AI systems, including the viability, constraints, costs, and properties of scalably safe systems. What is the safety production function , which maps the impact of various inputs on safety? Plausible inputs are compute, money, talent, evaluation time, constraints on the actuators, speed, generality, or capability of the deployed system, and norms and institutions conducive to risk reporting. To what extent do we need to spend time or resources at various stages of development (such as early or late) in order to achieve safety? If the safety-performance trade-offs are modest, and political or economic returns to absolute and relative performance are relatively inelastic (marginal improvements in performance are not that important), then achieving safe AI systems is more likely to be manageable; the world will not have to resort to radical institutional innovation or other extreme steps to achieve beneficial AI. If, however, the safety-performance trade-off is steep, or political or economic returns are highly elastic in absolute or especially relative performance, then the governance problem will be much harder to solve, and may require more extreme solutions. There are a broad range of implicit views about how technically hard it will be to make safe advanced AI systems. They differ on the technical difficulty of safe advanced AI systems, as well as risks of catastrophe, and rationality of regulatory systems. We might characterize them as follows: ❖ Easy : We can, with high reliability, prevent catastrophic risks with modest effort, say 1-10% of the costs of developing the system.  84   ❖ Medium: Reliably building safe powerful systems, whether it be nuclear power plants or advanced AI systems, is challenging. Doing so costs perhaps 10% to 100% the cost of the system (measured in the most appropriate metric, such as money, time, etc.). ➢ But incentives are aligned . Economic incentives are aligned so that companies or organizations will have correct incentives to build sufficiently safe systems. Companies don't want to build bridges that fall down, or nuclear power plants that experience a meltdown. ➢ But incentives will be aligned . Economic incentives are not perfectly aligned today, as we have seen with various scandals (oil spills, emissions fraud, financial fraud), but they will be after a few accidents lead to consumer pressure, litigation, or regulatory or other responses.  85  ➢ But we will muddle through . Incentives are not aligned, and will never be fully. However, we will probably muddle through (get the risks small enough), as humanity has done with nuclear weapons and nuclear energy. we be able to verify compliance with that policy? For example, would it be possible to separate a machine's objectives from its capabilities, as doing so could make it easier for non-experts to politically evaluate a system and could enable verification schemes that leak fewer technical secrets (related to capabilities)? ➢ Greater insight into the character of the safety problem will shed insight into a number of parameters relevant to solving the governance problem. Some governance arrangements that could depend on the character of the safety problem include: ❖ Providing incentives and protections for whistleblowers ❖ Representation of AI scientists in decision making ❖ Technical verification of some properties of systems ❖ Explicit negotiations over the goals of the system AI Safety work is being done at a number of organizations, including DeepMind, OpenAI, Google Brain, the Center for Human Compatible AI and UC Berkeley, the Machine Intelligence Research Institute, the Future of Humanity Institute, and elsewhere. \n AI Politics AI will transform the nature of wealth and power. The interests and capabilities of powerful actors will be buffeted, and new powerful actors may emerge. These actors will compete and cooperate to advance their interests. Advanced AI is likely to massively increase the potential gains from cooperation, and potential losses from non-cooperation; we thus want political dynamics to be such as to be most likely to identify opportunities for mutual benefit and to identify far in advance joint risks that could be avoided by prudent policy. Political dynamics could also pose catastrophic risks short of human-level AI if, for example, they lead to great power war or promote oppressive totalitarianism. Political dynamics will affect what considerations will be most influential in the development of (transformative) AI: corporate profit, reflexive public opinion, researchers' ethics and values, national wealth, national security, sticky international arrangements, or enlightened human interest. It is thus critical that we seek to understand, and if possible, beneficially guide, political dynamics. AI Politics looks at how the changing technical landscape could transform domestic and mass politics , international political economy , and international security , and in turn how policies by powerful actors could shape the development of AI. Work in this cluster benefits from expertise in domestic politics, international relations, and national security, among other areas. It will involve a range of approaches, including theory (mathematical and informal), contemporary case studies, historical case studies, close contact with and study of the relevant actors, quantitative measurement and statistical analysis, and scenario planning. \n Domestic and Mass Politics AI has the potential to shape, and be shaped by, domestic and mass politics. As AI and related technologies alter the distribution of domestic power, forms of government will alter. This could mean a shift in power towards actors with the capital and authority to deploy powerful AI systems, such as elites, corporations, and governments. On the other hand, AI could be used to enhance democracy, for example through aligned personal digital assistants, surveillance architectures that increase the accountability of authorities, or decentralized (crypto-economic) coordination technologies. The impact of exacerbated inequality and job displacement on trends such as liberalism, democracy, and globalization could be substantial. What systems are possible for mitigating inequality and job displacement, and will they be sufficient? More generally, public opinion can be a powerful force when it is mobilized. Can we foresee the contours of how public opinion is likely to be activated and expressed? Will certain groups--cultures, religions, economic classes, demographic categories--have distinct perspectives on AI politics? This set of questions is generally less relevant to short timelines (e.g. AGI comes within 10 years). \n Forms of Government Domestic political structures, such as whether a government is accountable through elections and is transparent through public legislative debates and an informed free press, arise as a complex function of many factors, some of which will plausibly be altered by advanced AI. Some factors that seem especially important to determining the character of government, and in particular the extent to which it is liberal and democratic, are: 1) the (unequal) distribution of control over economic assets and wealth; (2) surveillance Research on forms of government will examine plausible AI-driven trends in these and other factors, and evaluate possible strategies for mitigating adverse trends. This matters for extreme stakes because (i) trends in domestic governance speak to long-term trends in regime-type (e.g. democracy); (ii) it could influence the character of key actors in AI strategy, such as the character of the Chinese and US governments; (iii) it will inform the kinds of global institutions that will be feasible and their properties. \n Inequality There is an extensive and active literature on inequality and government. This literature should be reviewed, and lessons applied to our understanding about future forms of government, given trends in inequality (see section 4.2 ). \n Surveillance To what extent will AI and sensor technology enable cheap, extensive, effective surveillance? It is plausible that sufficient information about an individual's behavior, intent, and psychology--and of an individual's social network--will soon be generated through passive interactions with digital systems, such as search queries, emails, systems for affect and lie detection, spatial tracking through MAC addresses, face recognition, or other kinds of individual recognition. If so, a first order effect seems to be to shift power towards those entities who are able to use such information, plausibly to reinforce government authority, and thus authoritarian systems. However, super-surveillance could also prove beneficial, such as for enabling AI verification agreements and for enabling \"stabilization\" (the prevention of world-destroying technology). In addition, it may be possible to design AI-enabled surveillance in ways that actually reinforce other values and institutions, such as liberalism and democracy. For example, it may be possible to attenuate the typical tradeoff between security and privacy through cryptographically enabled privacy-preserving surveillance. 87 \n Repression and Persuasion Profiling of individuals, mapping of social networks, ubiquitous surveillance and lie detection, scalable and effective persuasion, and cost-effective autonomous weapons could all radically shift power to states. These trends may enable a state that is willing to do so to monitor and disrupt groups working against it. Autonomous weapons could permit a dictator to repress without requiring the consent of military officers or personnel, which have historically been one check on leaders. These trends should be mapped, understood, their potential consequences studied, and governance safeguards proposed.  87  One creative idea is to use secure multiparty computation or homomorphic encryption when analyzing data for evidence of criminal activity. These cryptographic technologies make it possible to perform such analysis without having access to the underlying data in an unencrypted form. Ben Garfinkel provides an excellent review of the state-of-the-art in cryptographic systems and their implications for political, economic, and social institutions in Garfinkel, Ben. \"Recent Developments in Cryptography and Potential Long-Term Consequences.\" Future of Humanity Institute, 2018. \n Advisors and Collective Action AI could also conceivably empower citizens, relative to the state or elites. Personal advisor AIs could allow individual citizens to engage politically in a more informed manner, and at lower cost to themselves; however, this level of AI capability seems like it might be close to AGI (likely to occur relatively late in the sequence of transformative developments). AI systems could facilitate collective action, such as if it becomes possible to assemble and mobilize a novel political coalition, and new political cleavages, through scraping of social media for expressions of support for neglected political positions. Individuals could express more complex political strategies, and more efficiently coordinate. For example, an American citizen (in a plurality voting system that strongly rewards coordination) might want to state that they would vote for a third party candidate if 40% of the rest of the electorate also agrees to do so. Other kinds of narrow AI advisors could transform domestic politics. Efficient AI translation could facilitate cross-language communication and coordination. AI political filters could exacerbate political sorting (filter bubbles), or could elevate political discourse by helping users to avoid low-credibility news and more easily identify counter-arguments. Video and audio affect and sincerity/lie detection, if effective and trusted, could incentivize greater sincerity (or self-delusion). 88 \n Inequality, Job Displacement, Redistribution AI seems very likely to increase inequality between people (and probably also between countries: see section 5 ). The digitization of products, because of low marginal costs, 89 increases winner-take-all dynamics. AI dramatically increases the range of products that can be digitized. AI will also displace middle-class jobs, and near-term trends are such that the replacement jobs are lower paying. Labor share of national incomes is decreasing; AI is 90  88  Thanks to Carl Shulman for the above.  89   likely to exacerbate this. Ultimately, with human-level AI, the labor share of income should 91 become ever smaller. Given that capital is more unequally distributed than labor value, an increase in capital share of income will increase inequality. AI seems to be generating (or is at least associated with) new natural global monopolies or superstar firms: there's effectively only one search engine (Google), one social network service (Facebook), and one online marketplace (Amazon). The growth of superstar firms plausibly drives the declining labor share of income. These AI (quasi-)monopolies and 92 associated inequality are likely to increasingly become the target of redistributive demands. Another risk to examine, for \"slow scenarios\" (scenarios in which other forms of transformative AI do not come for many decades) is of an international welfare race-to-the-bottom, as countries race with each other to prioritize the needs of capital and to minimize their tax burden. Research in this area should measure, understand, and project trends in employment displacement and inequality. What will be the implications for the policy demands of the public, and the legitimacy of different governance models? What are potential governance solutions? 93 \n Public Opinion and Regulation Historically public opinion has been a powerful force in technology policy (e.g. bans on GMOs or nuclear energy) and international politics (e.g. Sputnik). Further, as these examples illustrate, public opinion is not simply a reflection of elite interest. In the case of Sputnik, the US intelligence community was well aware of Soviet progress, and the Eisenhower administration did not want to engage in a space race and tried to persuade the American public that Sputnik was not a significant development. And yet, within months Sputnik had 94 triggered a reorientation of US technology policy, including the legislative formation of ARPA (today DARPA). It could thus be helpful to study public opinion and anticipate movements in public opinion, as can be informed by scenario based surveys, and studying particular groups who have been exposed to instances of phenomena (such as employment shocks, or new forms of surveillance) that could later affect larger populations. What kinds of public reactions could arise, leading to overly reactive policy and regulation? Could regulating AI (or taxing AI companies) become a new target of political campaigns, as already seems to be happening in Europe? This area of research will also help policymakers know how best to engage public opinion when an event occurs (and in general). It will also help scholars to communicate the results and value of their work. \n International Political Economy The next set of questions in the AI Politics research cluster examines the international political dynamics relating to the production and distribution of wealth. Economic success with AI and information technology seems to exhibit substantial returns to scale (e.g. Google 95 ) and agglomeration economies (e.g. Silicon Valley). If these trends persist it could lead to an 96 (even more extreme) international AI oligopoly, where a few firms capture most of the value from providing AI services. Are there any relevant aspects to the competitive dynamics  94  Ryan, Amy and Gary Keeley. \"Sputnik and US Intelligence: The Warning Record.\" Studies in Intelligence 61:3 (2017). https://www.cia.gov/library/center-for-the-study-of-intelligence/csi-publications/csi-studies/studies/vol-61-no-3/pdfs/sputni k-the-warning-record.pdf .  95  Due in part to the \"virtuous cycle\" between AI capabilities which attracts customers, which increases one's data, which improves one's AI capabilities, and the high fixed costs of developing AI services and low marginal cost of providing them.  96   What are the possibilities and likely dynamics of an international economic AI race? Is it plausible that countries would support domestic or allied consortia of AI companies, so as to better compete in industries that appear to be naturally oligopolistic? Technological displacement will impact countries differentially, and countries will adopt different policy responses. What will those be? If redistributing wealth and retraining becomes a burden on profitable companies, could there be AI capital flight and an international race \"to the bottom\" of providing a minimal tax burden? If so, could the international community negotiate (and enforce) a global tax system to escape this perverse equilibrium? Or are AI assets and markets sufficiently tax inelastic (e.g. territorially rooted) as to prevent such a race-to-the-bottom? Research on international political economy is most relevant for scenarios where AI does not (yet) provide a strategic military benefit, as once it does the logic of international security will likely dominate, or at least heavily shape, economic considerations. However, many IPE related insights equally apply to the international security domain, so there is value in studying these common problems framed in terms of IPE. \n International Security AI and related technologies are likely to have important implications for national and international security. It is also plausible that AI could have strategic and transformative military consequences in the near and medium-term, and that the national security perspective could become dominant. First, studying the near-term security challenges is helpful for understanding the context out of which longer-term challenges will emerge, and enable us to seed long-term beneficial precedents. Longer-term, if general AI becomes regarded as a critical military (or economic) asset, it is possible that the state will seek to control , close , and securitize AI R&D. Further, the strategic and military benefits of AI may fuel international race dynamics . We need to understand what such dynamics might look like, and how such a race can be avoided or ended . \n Near-term Security Challenges In the coming years AI will pose a host of novel security challenges. These include international and domestic uses of autonomous weapons, and AI-enabled cyber-operations, malware, and political influence campaigns (\"active measures\"). Many of these challenges look like \"lite\" versions of potential transformative challenges, and the solutions to these challenges may serve as a foundation for solutions to transformative challenges. To the 101 extent the near-term and transformative challenges, or their solutions, are similar, it will be useful for us to be aware of and engage with them. For a recommended syllabus of readings lead team be about the performance of its rivals, and that it will be able to sustain a known lead? Given models of AI safety (such as the performance-safety tradeoffs and the time-schedule for safety investments), what is the expected risk incurred by race dynamics? There are also questions about the strategies for retaining a lead or catching up. Are there tools available to the leading team that will allow it to retain a lead? For example, could a team retain its lead by closing off its research? What difference does it make if the leading team is a state, or closely supported by a state? The potential for coalitions within a race merits study. What are the possibilities for alliances between leading groups or states to help them retain their lead? consent, but the disadvantage that they tend to be crude, and are thus often inadequate and may even be misdirected. A hardened form of international norms is customary law, 109 though absent a recognized international judiciary this is not likely relevant for great-power cooperation.  110  Diplomatic agreements and treaties involve greater specification of the details of compliance and enforcement; when well specified these can be more effective, but require greater levels of cooperation to achieve. Institutions, such as the WTO, involve establishing a bureaucracy with the ability to clarify ambiguous cases, verify compliance, facilitate future negotiations, and sometimes the ability to enforce compliance. International cooperation often begins with norms, proceeds to (weak) bilateral or regional treaties, and consolidates with institutions. Some conjectures about when international cooperation in transformative AI will be more domains, AI appears in some ways less amenable to international cooperation--conditions (3), (4), (5), (6)--but in other ways could be more amenable, namely (1) if the parties come to perceive existential risks from unrestricted racing and tremendous benefits from cooperating, (2) because China and the West currently have a relatively cooperative relationship compared to other international arms races, and there may be creative technical possibilities for enhancing (4) and (5). We should actively pursue technical and governance research today to identify and craft potential agreements. \n Third-Party Standards, Verification, Enforcement, and Control One set of possibilities for avoiding an AI arms race is the use of third party standards, verification, enforcement, and control. What are the prospects for cooperation through third party institutions? The first model, almost certainly worth pursuing and feasible, is an international \"safety\" agency responsible for \"establishing and administering safety standards.\" This is crucial to achieve common knowledge about what counts as compliance. 111 The second \"WTO\" or \"IAEA\" model builds on the first by also verifying and ruling on  111  As per Article II of the IAEA Statute. (1) The governance problems that we are facing today and that we will face in the future overlap extensively, with the primary differences being (i) the scope of interests to be represented, (ii) the potential need to compete in some broader military-economic domain, and (iii) the stakes. To illustrate the similarities, consider how the governance of an international AI coalition will ideally have some constitutional commitment to a common good, will have institutional mechanisms for assuring the principals (e.g. the publics and leaders of the included countries) that the regime is well-governed, and for credibly communicating a lack of threat to other parties. In fact, if we are able to craft a sufficiently appealing, realistic, self-enforcing, robust model of AI governance, this could serve as a beacon, to guide us out of a rivalrous equilibrium. The problem then reduces to one of sequencing: how do we move from the present to this commonly appealing future? (2) We would ideally like to embed into our governance arrangements today, when the stakes are relatively low, the principles and mechanisms that we will need in the future. For example, given temporal discounting, diminishing marginal utility, and uncertainty about who will possess the wealth, it may be possible today to institutionalize collective commitments for redistributing wealth in the future.  113  This research cluster is the least developed in this document, and within the community of people working on AI governance. We want our governance institutions to be resilient to drift and hijacking. Two poles of the risk space are totalitarian capture and tyranny of the majority. To prevent totalitarian capture and tyranny of the majority, to varying extents and in varying combinations, countries throughout the world have employed: regular, free, and fair elections; protected rights for political expression; rule of law and an independent judiciary; division of power; constraints on state power; constitutionally protected rights; federalism. \n Values and Principles The problem of how to build institutions for governing a polity is a core part of the fields of political science and political economy. The more mathematical theoretical corner of this space is often called public choice, social choice, or (by economists) political economy. Political scientists in comparative politics and American politics extensively study the properties of different political systems. Scholars in political science and political theory study the design of constitutions. Given the centrality of this problem to these fields, and their existing expertise, substantial effort should be spent learning from them and recruiting them, rather than trying to reinvent good governance. Nevertheless, at the present time the application of these disciplines to the problems of AI governance remains neglected. \n Positive Visions While the above is directed to devising a feasible model of ideal long-run AI governance, it is unlikely to generate a simple solution (anytime soon). However, it could be extremely beneficial to have a simple, compelling, broadly appealing vision of the benefits from cooperation, to help motivate cooperation. We believe that both the potential benefits from safe development and the potential downsides from unsafe development are vast. Given that perspective, it is foolish to squabble over relative gains, if doing so reduces the chances of safe development. How can we simply, clearly, evocatively communicate that vision to others? 115 115 \"Paretotopia\"? the probability of a forecasting event and a set of predictions is to 50%, over a given time frame, the more we will learn about forecasting ability, and the world, over that time frame. 5. We ideally want them to be epistemically temporally fractal : we want them to be such that good forecasting performance on near-term forecasts is informative of good forecasting performance on long-term predictions. Near-term forecasting targets are more likely to have this property as they depend on causal processes that are likely to continue to be relevant over the long-term. 6. We want them to be jointly maximally informative . This means that we ideally want a set of targets that score well on the above criteria. A way in which this could not be so is if some targets are highly statistically dependent on others, such as if some are logically entailed by others. Another heuristic here is to aim for forecasting targets that exhaustively cover the different causal pathways to relevant achievements. 14 science 14 , engineering, and science. These advantages may come in sufficient strength or combinations to radically transform power. Even the mere perception by governments and publics of such military (or economic) potential could lead to a radical break from the current technology and world order: shifting AI leadership to governments, giving rise to a massively funded AI race and potentially the securitization of AI development and capabilities. This16 could undermine the liberal world economic order. The intensity from a race dynamic could 17 lead to catastrophic corner-cutting in the hurried development and deployment of (unsafe) advanced AI systems. This danger poses extreme urgency, and opportunity, for global 18 cooperation. \n 24 has been expressed by Robin Hanson, Paul Christiano, AI Impacts and Katja Grace, and 25 26 27 \n technologies and architectures; (3) repression technologies; (4) persuasion technologies; (5) personal advisor technologies; (6) collective action technologies. \n AI ideal governance aspires to envision, blueprint, and advance ideal institutional solutions for humanity's AI governance challenges. What are the common values and principles around which different groups can coordinate? What do various stakeholders (publics, cultural groups, AI researchers, elites, governments, corporations) want from AI, in the near-term and long-term? What are the best ways of mediating between competing groups and between conflicting values? What do long-term trends--such as from demographics, secularization, globalization, liberalism, nationalism, inequality--imply about the values of these stakeholders over medium and long timelines? \n 5 Amodei, Dario, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. \"Concrete Problems in AI Safety.\" ArXiv:1606.06565 [Cs] , June 21, 2016. http://arxiv.org/abs/1606.06565 ; Russell, Stuart, Daniel Dewey, and Max Tegmark. \"Research Priorities for Robust and Beneficial Artificial Intelligence.\" Future of Life Institute -AI Magazine , 2015. https://futureoflife.org/data/documents/research_priorities.pdf?x90991 ; Everitt, Tom, Gary Lea, and Marcus Hutter. \"AGI Safety Literature Review.\" ArXiv:1805.01109 [Cs] , May 3, 2018. http://arxiv.org/abs/1805.01109 ; Metz, Cade. \"Teaching A.I. Systems to Behave Themselves.\" The New York Times , August 13, 2017, sec. Technology. https://www.nytimes.com/2017/08/13/technology/artificial-intelligence-safety-training.html . \n Bostrom, Nick, Allan Dafoe, and Carrick Flynn. \"Public Policy and Superintelligent AI: A Vector Field Approach.\" Future of Humanity Institute, 2018. http://www.nickbostrom.com/papers/aipolicy.pdf . 7 Bostrom, Nick. Superintelligence: Paths, Dangers, Strategies . Oxford: Oxford University Press, 2014. Yudkowsky, Eliezer. \"Artificial Intelligence as a Positive and Negative Factor in Global Risk.\" In Global Catastrophic Risks , edited by Nick Bostrom and Milan M. Cirkovic. New York: Oxford University Press, 2008, pp. 308-45. See also the reading syllabus by Bruce Schneier. \"Resources on Existential Risk -for: Catastrophic Risk: Technologies and Policies.\" Berkman Center for Internet and Society, Harvard University, 2015. 3) Broadly superhuman AI systems could be built that are not fully aligned with human values , leading to human extinction or other permanent loss in value. This 7 risk is likely much greater if labs and countries are racing to develop and deploy advanced AI, as researching and implementing AI safety measures is plausibly time 8 and resource intensive.6 https://futureoflife.org/data/documents/Existential%20Risk%20Resources%20(2015-08-24).pdf?x70892 .8  Armstrong, Stuart, Nick Bostrom, and Carl Shulman. \"Racing to the Precipice: A Model of Artificial Intelligence Development.\" Technical Report. Future of Humanity Institute, 2013. https://www.fhi.ox.ac.uk/wp-content/uploads/Racing-to-the-precipice-a-model-of-artificial-intelligence-development.pdf . \n Each of these clusters characterizes a set of problems and approaches within which the density of conversation is likely to be greater. However, most work in this space will need to engage with the other clusters, reflecting and contributing high-level insights. Superintelligence offers tremendous opportunities, such as the radical reduction of disease, poverty, interpersonal conflict, and catastrophic risks such as climate change. However, superintelligence, and advanced AI more generally, may also generate catastrophicvulnerabilities, including extreme inequality, global tyranny, instabilities that spark global (nuclear) conflict, catastrophically dangerous technologies, or, more generally, insufficiently controlled or aligned AI. Even if we successfully avoid such technological and political pitfalls, tremendous governance questions confront us regarding what we want , and what we ought to want , the answers to which will require us to know ourselves and our values much better than we do today.Overview AI governance can be organized in several ways. This agenda divides the field into three complementary research clusters: the technical landscape , AI politics , and AI ideal governance .This framework can perhaps be clarified using an analogy to the problem of building a new city. The technical landscape examines the technical inputs and constraints to the problem, such as trends in the price and strength of steel. Politics considers the contending motivations of various actors (such as developers, residents, and businesses), the mutually harmful dynamics that could potentially arise between them, and strategies for cooperating to overcome such dynamics. Ideal governance involves understanding the ways that infrastructure, laws, and norms can be used to build the best city, and proposing ideal master plans of these to facilitate convergence on a common good vision.The first cluster, the technical landscape , seeks to understand the technical inputs, possibilities, and constraints for AI, and serves as a foundation for the other clusters of AI governance. This includes mapping what could be the capabilities and properties of advanced and transformative AI systems, when particular capabilities are likely to emerge, and whether they are likely to emerge gradually in sequence or rapidly across-the-board. To the extent possible, this cluster also involves modeling and forecasting AI progress : the production function of AI progress, given inputs such as compute, talent, data, and time, and the projection of this progress into the future . We also need to assess the viability, constraints, costs, and properties of scalably safe AI systems . To what extent will we need to invest resources and time late in the development process? What institutional arrangements best promote AI safety? To what extent will the characteristics of safe AI be apparent to and observable by outsiders, as would be necessary for (non-intrusive) external oversight and verification agreements? \n .1 Rapid and Broad Progress? Technical Landscape seemingly disparate capabilities, spanning image recognition, language translation, In steering away from dangerous rivalrous dynamics it will be helpful to have a clear sense of speech recognition, game playing, and others. Perhaps a substantial advance in what we are steering towards, which bring us to the final research cluster: what are the ideal A rich overview to issues in the field is given in Bostrom, Nick. Superintelligence: Paths, Work on the technical landscape seeks to understand the technical inputs, possibilities, and \"efficient meta-learning\" or transfer learning could catalyze advances in many areas. governance systems for global AI dynamics? What would we cooperate to build if we could? Dangers, Strategies. Oxford: Oxford University Press, 2014. See especially chapters 4 ('The constraints for AI, providing an essential foundation for our later study of AI politics, ideal ❖ (1b) Clusters of novel powerful technological capabilities that are likely to be What potential global governance systems--including norms, policies, laws, processes, and Kinetics of an Intelligence Explosion'), 5 ('Decisive Strategic Advantage'), 11 ('Multipolar governance, and policy. This includes mapping what the capabilities and properties of unlocked in close proximity to each other, perhaps because they facilitate each other institutions --can best ensure the beneficial development and use of advanced AI systems? To scenarios'), and 14 ('The strategic picture'). transformative AI systems could be, when they are likely to emerge, and whether they are or depend on solving some common problem. answer this we need to know what values humanity would want our governance systems to likely to emerge in particular sequences or many-at-once. This research cluster benefits from ❖ (2a) Complements : Scientific and technological advances often depend on having pursue, or would want if we understood ourselves and the world better. More pragmatically, The Future of Life Institute offers a set of resources on Global AI Policy here: expertise in AI, economic modeling, statistical analysis, technology forecasting and the history several crucial inputs, each of which acts as a strong complement to the others. For what are the specific interests of powerful stakeholders, and what institutional https://futureoflife.org/ai-policy/ . of technology, expert elicitation and aggregation, scenario planning, and neuroscience and example, the development of powered flight seems to have required sufficient mechanisms exist to assure them of the desirability of a candidate governance regime? evolution. advances in the internal-combustion engine. Complementarities could lead to jumps 19 Insights for long-term global governance are relevant to contemporary and medium-term AI in capabilities in several ways. governance, as we would like to embed the principles and institutional mechanisms that will 1. Mapping Technical Possibilities ➢ (i) Unjammed bottlenecks : There could be rapid alleviation of a crucial be crucial for the long-term today, while the stakes are relatively low. It will also facilitate This cluster investigates the more abstract, imaginative problem area of mapping technical bottleneck. For example, we have seen sudden jumps in capabilities from the cooperation today if we can assure powerful actors of a long-term plan that is compatible possibilities, and especially potentially transformative capabilities. Are we likely to see a provision of a single large training dataset for a particular task. Similarly, the with their interests. rapid broad (and local?) achievement of many transformative capabilities? What kinds of generation of a crucial training set for a generally applicable task could lead to transformative capabilities could plausibly emerge, and in what order? What are their a broad front of progress. strategic parameters, such as by enabling powerful new capabilities (in cyber, lethal In working in this space across the three research clusters, researchers should prioritize the strategic properties, such as being offense-or defense-biased, or democracy or autocracy ➢ (ii) Overhang : There could be a latent reservoir of a crucial complement, that autonomous weapons [LAWs], military intelligence, strategy, science), by shifting the questions which seem most important , tractable , and neglected , and for which they have a valenced? becomes suddenly unlocked or accessible. For example, it could come from: 20 offense-defense balance, or by making crisis dynamics unstable, unpredictable, or more comparative advantage and interest . Questions are more likely to be important if they are hardware overhang in which there is a large reservoir of compute available rapid? Could trends in AI facilitate new forms of international cooperation, such as by enabling strategic advisors, mediators, or surveillance architectures, or by massively likely to identify or address crucial considerations, or if they directly inform urgent policy decisions. Questions are more likely to be tractable if the researcher can articulate a 1A first issue concerns how rapid and general advances will be in AI. Some believe that to be repurposed following an algorithmic breakthrough; from abundant increasing the gains from cooperation and costs of non-cooperation? If general AI comes to be promising well-defined research strategy, the questions can be tackled in isolation, or they progress will, at some point, allow for sudden improvements in AI systems' capabilities seen as a critical military (or economic) asset, under what circumstances is the state likely to reduce to resolvable questions of fact. Questions are more likely to be neglected if they do not across a broad range of tasks. If so, much of the following proposed work on sequences and control, close, or securitize AI R&D? What are the conditions that could spark and fuel an directly and exclusively contribute to an actor's profit or power, like many long-term or kinds of AI would be unproductive, since most transformative capabilities would come online international AI race ? How great are the dangers from such a race, how can those dangers be global issues, and if they fall outside of the focus of traditional research communities. at the same explosive moment. For this reason, this agenda draws initial attention to this communicated and understood, and what factors could reduce or exacerbate them? What Researchers should upweight questions for which they have comparatively relevant expertise question. routes exist for avoiding or escaping the race, such as norms, agreements, or institutions or capabilities, and in which they are especially interested. Ultimately, perhaps the simplest regarding standards, verification, enforcement, or international control? How much does it rule of thumb is to just begin with those questions or ideas that most grab you, and start Rapid general progress could come about from several mechanisms, enumerated with some matter to the world whether the leader has a large lead-margin, is (based in) a particular furiously working. redundancy: country (e.g. the US or China), or is governed in a particular way (e.g. transparently, by ❖ (1a) Many important tasks may require a common capability, the achievement of scientists)? which would enable mastery across all of them. For example, deep learning unlocked economy, warranting massive state support and intervention? If so, what policies will this entail, which countries are best positioned to seize AI economic dominance, and how will this AI nationalism interact with global free trade institutions and commitments?Potentially most importantly, how will AI interact with international security ? What are the near-term security challenges (and opportunities) posed by AI? Could AI radically shift key \n insecure compute that can be seized by an expansionist entity; from insight overhang , if there are general powerful algorithmic improvements waiting to be uncovered; or from data overhang , such as the corpus of digitized science textbooks waiting to be read, and the internet more generally. ➢ (iii) Complementary clusters of capabilities : Advances in one domain of AI could strongly complement progress in other domains, leading to a period of rapid progress in each of these domains. For example, natural language 19 Crouch, Tom D, Walter James Boyne et al. \"History of flight.\" Encyclopaedia Britannica , 2016. https://www.britannica.com/technology/history-of-flight/The-generation-and-application-of-power-the-problem-of-propulsio n . \n ://intelligence.org/files/IEM.pdf.24  Bostrom, Nick. Superintelligence: Paths, Dangers, Strategies. Oxford: Oxford University Press, 2014. Chapter 4. Grace, Katja. \"Likelihood of Discontinuous Progress around the Development of AGI.\" AI Impacts , February 23, 2018. 28 29 25 Hanson, Robin. \"I Still Don't Get Foom.\" Overcoming Bias (blog), July 24, 2014. http://www.overcomingbias.com/2014/07/30855.html . 26 Christiano, Paul. \"Takeoff Speeds.\" The Sideways View (blog), February 24, 2018. https://sideways-view.com/2018/02/24/takeoff-speeds/ . 21 Autoregressive parameter persistently above 1.22  Good, I.J. \"Speculations Concerning the First Ultraintelligent Machine.\" In Advances in Computers , edited by Franz L. Alt and Moris Rubinoff, 6:31-88. New York: Academic Press, 1964. 23 Yudkowsky, Eliezer. \"Intelligence Explosion Microeconomics.\" Machine Intelligence Research Institute, 2013. https27 https://aiimpacts.org/likelihood-of-discontinuous-progress-around-the-development-of-agi/ . 28 Goertzel, Ben. \"Superintelligence: Fears, Promises and Potentials: Reflections on Bostrom's Superintelligence, Yudkowsky's From AI to Zombies,and Weaver and Veitas's 'Open-Ended Intelligence.'\" Journal of Evolution & Technology 24, no. 2 (November 2015): 55-87. http://www.kurzweilai.net/superintelligence-fears-promises-and-potentials . \n Mnih, Volodymyr, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, et al. \"Human-Level Control through Deep Reinforcement Learning.\" Nature 518, no. 7540 (February 2015): 529-33. https://doi.org/10.1038/nature14236 . See Figure \n developments be predictable and foreseeable in their character? To what extent will AI be dual use , making it hard to distinguish between the development, training, and deployment of dangerous/destabilizing/military systems and safe/stabilizing/economic systems? To what extent will developments be predictable in their timelines and sequencing , and what are our best forecasts (see section 2.3 )? If we can estimate roughly when strategically relevant or transformative thresholds are likely to be reached, or in what order, then we can formulate a better map of the coming transformations. \n What are the key categories of input to AI R&D, and can we measure their existing distribution and rates of change? Plausibly the key inputs to AI progress are computing power (compute), talent, data, insight, and money. Can we sensibly operationalize these, orComputational Power and the Social Impact of Artificial Intelligence.\" 2018, 1-44. http://dx.doi.org/10.2139/ssrn.<PHONE> . Hilbert, M., and P. López. \"The world's technological capacity to store, communicate, and compute information.\" Science 332, issue 6025 (April 1, 2011). http://doi.org/10.1126/science.<PHONE> . 41 At https://www.eff.org/ai/metrics . 42 E.g. one relevant question is whether China can become a decisive world leader in AI without becoming more scientifically open. Wagner, Caroline S., and Koen Jonkers. \"Open Countries Have Strong Science.\" Nature News 550, no. 7674 (October 5, 2017). https://doi.org/10.<PHONE>a , p. 32. For an overview of China's AI landscape, see Ding, Jeffrey. \"Deciphering China's AI Dream: The context, components, capabilities, and consequences of China's strategy to lead the world in AI.\" Future of Humanity Institute, March 2018. https://www.fhi.ox.ac.uk/wp-content/uploads/Deciphering_Chinas_AI-Dream.pdf . Supplementing model-based forecasts with expert assessment, to what extent can we forecast AI progress? 2.1 Measuring Inputs, Capabilities, and Performance quantitative about assessing existing and future progress in AI. Can we improve our measures of AI inputs, investments, and performance? Can we model AI progress: the relationship between measurable inputs and indicators, and future AI innovation? 40 find useful proxies for them? What are the most important AI capabilities that we should be tracking? Can we construct sensible, tractable, strategically relevant measures of performance , that either track or predict transformative capabilities? Prior and existing metrics of performance are summarized and tracked by the Electronic Frontier Foundation.41    This measurement exercise should be disaggregated at the level of the strategically relevant actor. Who are the main organizations and countries involved, and what is the distribution of and rates of change in their inputs, capabilities, and performance? Later in this document we will ask about the strategic properties of these organizations and countries, such as their institutional configuration (e.g. legal structure, leadership selection process), goals (political, economic, other), and access to other strategic assets. As a relatively poorly understood and potentially pivotal actor, current research is especially seeking to better understand China's inputs, capabilities, and performance. 422.2 Modeling AI ProgressAs well as mapping technical possibilities, we want to be able to model progress in AI development towards these possibilities. 40 Hwang, T. \" \n For the original paper, see Moore, G.E. \"Cramming More Components Onto Integrated Circuits.\" Electronics 38, no. 8 (April 19, 1965): 82-85. https://doi.org/10.1109/JPROC.<PHONE> . Cf. also Schaller, R. R. \"Moore's Law: Past, Present and Future.\" IEEE Spectrum 34, no. 6 (June 1997): 52-59. https://doi.org/10.1109/6.591665 ; \"Trends in the cost of computing.\" AI Impacts , March 10, 2015. https://aiimpacts.org/trends-in-the-cost-of-computing/ . Dario, and Danny Hernandez. \"AI and Compute.\" OpenAI (blog), May 16, 2018. https://blog.openai.com/ai-and-compute/ . 50 Silver, D., A. Huang, C.J. Maddison, A. Guez, and L. Sifre. \"Mastering the game of Go with deep neural networks and tree search.\" Nature 529 (January 28, 2016). http://doi.org/10.1038/nature16961 . 45 Kurzweil, Ray. \"The Law of Accelerating Returns.\" Kurzweilai (blog), March 7, 2001. http://www.kurzweilai.net/the-law-of-accelerating-returns . 46 Nordhaus, William D. \"Are we approaching an economic singularity? Information technology and the future of economic growth.\" Cowles Foundation Discussion Paper No. 2021, September 2015. Figure 1, https://cowles.yale.edu/sites/default/files/files/pub/d20/d2021.pdf . 47 Kurzweil's data shows the trend beginning in 1900, Nordhaus's data shows the trend beginning in 1940. 48 Grace, Katja. \"Recent Trend in the Cost of Computing.\" AI Impacts, November 11, 2017. https://aiimpacts.org/recent-trend-in-the-cost-of-computing/ . Doubling time is equal to time to a 10x increase divided by 3.3 (because log 2 (10)=3.3). 49Amodei, \n At https://www.metaculus.com/questions/ . We use AI Safety to refer to the distinct, specialized field focusing on technical aspects of building beneficial and safe AI. AI Safety and AI Governance can be used as exclusive (and exhaustive) categories for the work needed to build beneficial AI. This agenda summarizes the aspects of AI Safety especially relevant to AI Governance.57  Though even this often requires considerable organizational efforts, and involves many close calls; cf. Sagan, Scott D. The Limits of Safety: Organizations, Accidents, and Nuclear Weapons . Princeton:Princeton University Press, 1993; Schlosser, Eric. Command and Control: Nuclear Weapons, the Damascus Accident, and the Illusion of Safety . Reprint edition. New York: Penguin Books, 2014. 58 Potember, Richard. \"Perspectives on Research in Artificial Intelligence and Artificial General Intelligence Relevant to DoD.\" JASON -The MITRE Corporation, 2017. https://fas.org/irp/agency/dod/jason/ai-dod.pdf , p. 2. 55 Muehlhauser, Luke. \"Retrospective Analysis of Long-Term Forecasts\". https://osf.io/ms5qw/register/564d31db8c5e4a7c9694b2be . 56   \n Bostrom. Superintelligence ; Yudkowsky, Eliezer. \"Artificial Intelligence as a Positive and Negative Factor in Global Risk.\" In Global Catastrophic Risks , edited by Nick Bostrom and Milan M. Cirkovic, pp. 308-45. New York: Oxford University Press, 2008. 66  Bostrom. Superintelligence , pp. 105-108. \n See Soares, Nate. \"Ensuring Smarter-than-Human Intelligence has a Positive Outcome.\" Talks at Google series, November 20, 2016. https://www.youtube.com/watch?v=dY3zDvoLoao . Bostrom. Superintelligence , Chapter 9. Horowitz, Michael C., Paul Scharre, and Alex Velez-Green. \"A Stable Nuclear Future? The Impact of Autonomous Systems and Artificial Intelligence.\" Working Paper, December 2017; Geist, Edward, and Andrew J Lohn. \"How Might Artificial Intelligence Affect the Risk of Nuclear War?\" RAND, 2018. https://www.rand.org/pubs/perspectives/PE296.html . Yudkowsky, Eliezer. \"Difficulties of AGI Alignment.\" The Ethics of Artificial Intelligence Conference, NYU , 2016. https://livestream.com/nyu-tv/ethicsofAI/videos/138893593 .68   \n Korinek, Anton, and Joseph E. Stiglitz. \"Artificial intelligence and its implications for income distribution and unemployment.\" For a review of forecasts of AI displacement of human jobs, see Winick, Erin. \"Business Impact Every study we could find on what automation will do to jobs, in one chart.\" MIT Technology Review , January 25, 2018. https://www.technologyreview.com/s/610005/every-study-we-could-find-on-what-automation-will-do-to-jobs-in-one-chart/ . The three more prominent forecasts are from: Nedelkoska, Ljubica and Glenda Quintini.\"Automation, Skills Use and Training\", OECD Social, Employment and Migration Working Papers, No. 202, OECD Publishing, Paris, 2018. https://www.oecd-ilibrary.org/docserver/2e2f4eea-en.pdf?expires=<PHONE>&id=id&accname=guest&checksum=F85DCC6 No. w24174. National Bureau of Economic Research, 2017. https://www8.gsb.columbia.edu/faculty/jstiglitz/sites/jstiglitz/files/w24174.pdf . 90   \n On industry concentration in AI, see the following and references: Bessen, James E. \"Information Technology and Industry Concentration.\" Boston Univ. School of Law, Law and Economics Research Paper No. 17-41, December 1, 2017. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=<PHONE> . between companies; for example, to what extent are AI innovations being patented, are patentable, or are held as trade secrets?Countries lacking AI industries currently worry that they are being shut out of the most rewarding part of the global value chain. Some in the US and Europe currently worry that 97 China is coercively/unfairly leveraging its market power to strategically extract technical competence from Western firms, and this was arguably the motivation for the Trump trade war. These concerns could lead to AI mercantilism or AI nationalism , following from 98 99 strategic-trade theory, where countries devote substantial resources to retaining and developing AI capabilities, and to supporting their AI national champions. To what extent are countries (e.g. Canada) able to internalize the returns on their AI investments, or does talent inevitably gravitate towards and benefit the existing leaders in AI (e.g. Silicon Valley)? 100 What lessons emerge from examining the partial analogies of other general purpose technologies and economy wide transformations such as computerization, electrification, and industrialization? Countries and companies are searching for other ways to economically benefit from the AI transformed economy. They are searching for rewarding nodes in the value chain in which they can specialize. Countries are examining policy levers to capture more of the rents from AI oligopolies, and aspire to build up their own AI champions (such as the EU rulings against Apple and Google, and China's exclusion of Google and Facebook). How substantial of an advantage does China have, as compared with other advanced developed (mostly liberal democratic) countries, in its ability to channel its large economy, collect and share citizen data, and exclude competitors? What steps could and would the U.S. take to reinforce its lead? \n Research on race dynamics involves a large set of questions and approaches. We will need to integrate and develop models of technology/arms races. What are the distinctive features 103 of an AI race, as compared with other kinds of races? What robust predictions can we make about that subfamily of races? Under what conditions are those races most dangerous or destructive? Specifically, a plausible and important proposition is that races are more dangerous the smaller the margin of the leader; is this a robust conclusion? How do 104 openness, accessibility of the research frontier, first-mover advantages, insecure compute, and other factors affect race dynamics? Given models of AI innovation, how confident can a 6.2 Control, Closing, and Securitization Basic AI R&D is currently conducted in the open: researchers have a strong interest to publish their accomplishments to achieve recognition, and there is a strong ethos of scientific openness. Some AI R&D is semi-closed, conducted in private for-profit spaces; however, this tends to not be general AI R&D, but instead applications of existing techniques. This could plausibly change, if AI becomes perceived as catastrophically dangerous, strategic military, or on AI and International Security, see: ❖ Zwetsloot, Remco. \"Artificial Intelligence and International Security Syllabus.\" Future of Humanity Institute, 2018. ( link ). Some specific references worth looking at include: ❖ Brundage, M., S. Avin, et al. \"The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation.\" Future of Humanity Institute, 2018. [ PDF ]. ❖ Horowitz, Michael, Paul Scharre, Gregory C. Allen, Kara Frederick, Anthony Cho and Edoardo Saravalle. \"Artificial Intelligence and International Security.\" Centre for a New American Security, 2018. ( link ). ❖ Scharre, P. Army of None: Autonomous Weapons and the Future of War . New York: W. W. Norton & Company, 2018. ❖ Horowitz, M. \"Artificial Intelligence, International Competition, and the Balance of Power.\" Texas National Security Review , 2018. [ link ] [ PDF ]. \n /users.econ.umn.edu/~holmes/class/2007f8601/papers/horner.pdf .Finally it would be valuable to theorize the likely stages of an AI race and their characteristics (tempo, danger, security consequences). Can we map current behavior onto this framework? In light of states' interests in strong AI systems, current international agreements, and historic relationships, what configurations of state coalitions are likely and under what circumstances? Historical precedents and analogies can provide insight, such as consideration of the arms race for and with nuclear weapons, other arms races, and patent and economic technology races. What about analogies to other strategic general purpose technologies and more gradual technological transformations, like industrialization, electrification, and computerization? In what ways do each of these fail as analogies? 103 For a model on risks from AI races, see Armstrong, Stuart, Nick Bostrom, and Carl Shulman. \"Racing to the precipice: a model of artificial intelligence development.\" AI & Society 31, no. 2 (2016): 201-206. The broader modeling literature relevant to races is large, including economic models of auctions, patent races, and market competition. Race to Stay Ahead.\" Review of Economic Studies (2004) 71, <PHONE>. 105 existing policy makers and publics perceive or invoke a race logic? What kinds of events 106 could spark or escalate a race, such as \"Sputnik moments\" for publics or an 107 \"Einstein-Szilard letter\" for leaders? 108 http:/What is the current distribution of capabilities, talent, and investment? To what extent do 6.4 Avoiding or Ending the Race 104  For a model of a technology race which is most intense when the racers are far away, see Hörner, Johannes. \"A Perpetual Given the likely large risks from an AI race, it is imperative to examine possible routes for avoiding races or ending one underway . The political solutions to global public bads are, in increasing explicitness and institutionalization: norms, agreements (\"soft law\"), treaties, or institutions. These can be bilateral, multilateral, or global. Norms involve a rough mutual understanding about what (observable) actions are unacceptable and what sanctions will be imposed in response. Implicit norms have the advantage that they can arise without explicit \n likely are when: (1) the parties mutually perceive a strong interest in reaching a successful agreement (great risks from non-cooperation or gains from cooperation, low returns on unilateral steps); (2) when the parties otherwise have a trusting relationship; (3) when there is sufficient consensus about what an agreement should look like (what compliance consists of), which is more likely if the agreement is simple, appealing, and stable; (4) when compliance is easily, publicly, and rapidly verifiable; (5) when the risks from being defected on are low, such as if there is a long \"breakout time\", a low probability of a power transition because technology is defense dominant, and near-term future capabilities are predictably non-transformative; (6) the incentives to defect are otherwise low. Compared to other \n\t\t\t escalation, the efficiency of negotiations, the viability of mutual privacy preserving surveillance, and the volatility and predictability of the future balance of power. It could enable new operational and strategic capabilities, for instance in mass-persuasion, cyber-operations, command and control, intelligence, air combat, subsea combat, materials12  It is worth being clear about the magnitudes of impact that one is contemplating. \"Transformativeness\" and its cognates have been used in a broad set of ways. At the high end, it has been used to refer to the most extreme impacts, such as AI that \"precipitates a transition comparable to (or more significant than) the agricultural or industrial revolution.\" Call impacts at the scale of the agricultural revolution or industrial revolution revolutionary impacts . At the low end, the term is used to refer to the mundane reorganization of industries. I use the term here to refer to impacts of intermediate magnitude.There are tradeoffs in any definition of a concept. Expanding the concept will allow us to contribute to and benefit from the broader conversation on the impacts of AI, most of which are not focused on revolutionary impacts. However, doing so may take our eyes off the most important impacts, diluting our efforts. I recommend the above intermediate definition, delimited at innovations that could induce radical changes. In addition to helping us address the broader set of large implications, doing so will help us remain attentive to the many seemingly small achievements, events, or dynamics that can generate massive impacts. To most, the first domesticated plants and the first steam engines would not have looked revolutionary. \n\t\t\t On the deleterious effects of framing AI development as a 'race', see Cave, Stephen, and Seán S. Ó hÉigeartaigh. \"An AI Race for Strategic Advantage: Rhetoric and Risks.\" In AAAI / ACM Conference on Artificial Intelligence, Ethics and Society , 2018. http://www.aies-conference.com/wp-content/papers/main/AIES_2018_paper_163.pdf . 17 Danzig, Richard, ed. \"An Irresistible Force Meets a Moveable Object: The Technology Tsunami and the Liberal World Order.\" Lawfare Research Paper Series 5, no. 1 (August 28, 2017). https://assets.documentcloud.org/documents/<PHONE>/Danzig-LRPS1.pdf . 18 Armstrong, Stuart, Nick Bostrom, and Carl Shulman. \"Racing to the Precipice: A Model of Artificial Intelligence Development.\" Technical Report. Future of Humanity Institute, 2013. \n\t\t\t For example, we can analyze how transnational self-governance regimes of private companies have emerged and why these efforts have succeeded or failed. This is particularly relevant as several AI companies have already introduced self-governance measures as well. Fischer, Sophie-Charlotte. 2018. \"Reading List -Industry Self-Regulation/Security Governance\".. \n\t\t\t This is also asked in the Technical Landscape . 106 Cave, Stephen, and Seán S. Ó hÉigeartaigh. \"An AI Race for Strategic Advantage: Rhetoric and Risks.\" In AAAI / ACM Conference on \n\t\t\t Adler, Emanuel. \"The Emergence of Cooperation: National Epistemic Communities and the International Evolution of the Idea of Nuclear Arms Control.\" International Organization 46, no. 1 (1992): 101-45. \n\t\t\t Bostrom, Nick, Allan Dafoe, and Carrick Flynn. \"Public Policy and Superintelligent AI: A Vector Field Approach.\" Future of Humanity Institute, 2018. http://www.nickbostrom.com/papers/aipolicy.pdf \n\t\t\t Though we should learn a lot from seeing one such unexpected event occur. Thus, such a \"long-shot\" target would be a worthwhile forecasting target to a person who assigns intermediate subjective probability of it occurring, even if everyone else in the community is confident it will (not) occur.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/GovAI-Agenda.tei.xml", "id": "d06c0484c4c18f1279df70a99c776e47"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Each step in the framework defines a phase for sustained talent management process change. First, identification of AI talent; second, experimentation, evaluation, and iteration of new initiatives designed to leverage AI talent; third, implementation of an agreed upon approach; and fourth, harmonization of service-level approaches to enable enterprise-wide AI talent assessment. Our recommendations are designed to create agility in talent management. We believe AI, like other emerging technologies, demonstrates the need for services to adapt to new warfighting techniques and domains. Such agility includes aligning performance incentives; creating opportunities to build and leverage expertise; and, importantly, to track, evaluate, and iterate pilot initiatives. It also allows for flexibility across services without creating more seams as AI talent management is unlikely to look the same for each service. What may work well for one service may not for another, given inherent differences in approach to AI adoption, mission priorities, and force management.", "authors": ["Diana Gehlhaus", "Ron Hodge", "Luke Koslosky", "Kayla Goode", "Jonathan Rotner"], "title": "The DOD’s Hidden Artificial Intelligence Workforce", "text": "Executive Summary Cultivating a leading artificial intelligence workforce is a strategic focus of the U.S. Department of Defense's (DOD) 2018 AI Strategy. That makes harnessing the full potential of AI for assured superiority and security critical, but it cannot happen in a vacuum. It requires a ready and able workforce, along with sustained pipelines and opportunities to leverage expertise. Recommendations from the National Security Commission on Artificial Intelligence's Final Report echoed this need, stating that \"national security agencies need more digital experts now or they will remain unprepared to buy, build, and use AI and its associated technologies.\" 1 However, the DOD has publicly stated that it faces significant challenges in accessing AI talent.  2  Adding to these challenges, the DOD has not consistently defined AI talent. The DOD's AI workforce discussions have therefore had a narrow focus on recruiting select technical talent, namely software developers, lamenting an inability to compete with the private sector. The push to engage this talent includes campaigning to raise salaries, creating fellowships, and leveraging rotational exchanges with industry. This policy brief provides a new perspective on this discussion. Instead of repeating how the DOD struggles to compete with the private sector, we argue that the DOD already has a hidden cadre of AI and related experts. We further argue that this hidden cadre could go a long way in meeting stated AI objectives, through policies that more effectively identify and leverage this talent, processes that incentivize experimentation and changes in career paths, and through investing in necessary technological infrastructure. To better understand the current state of the DOD's AI workforce, we conducted interviews with 31 experts across the services and the Office of the Secretary of Defense. We then analyzed civilian personnel data from the Office of Personnel Management (OPM), and leveraged key insights from previous CSET research. Our analysis suggests that the DOD employs more AI talent than the popular narrative may suggest. Previous CSET research found the national security community was a top employer of AI talent, particularly technical talent.  3  The interviews validated this hypothesis, as did our analysis of OPM data on the top occupation codes that likely include AI talent. Our analysis also points to a range of reasons for why the DOD's existing AI workforce remains hidden. Strategically, the services do not have a consistent approach to defining or prioritizing AI. This translates directly into the level of prioritization and investment put into cultivating an AI workforce. On a more tactical level, we find that challenges for identifying and leveraging AI talent fall into three general categories: people, processes, and technology. Moreover, we find that each DOD armed service currently has a different approach to defining, identifying, classifying, assigning and tracking AI and AI-related talent. Approaches span informal networking and word-of-mouth recommendations; equating AI expertise with technically-oriented career fields or educational credentials; creating searchable skills repositories; and even a machine learning approach in which personnel files are scraped and analyzed for potential digital skills. Our discussions revealed many organic experiments taking place within each service to identify and leverage AI or related talent, from informal \"communities of practice\" to flyaway teams and new career pathways. Finally, experts had a range of opinions on what different services should do to cultivate a leading AI workforce that reflected their circumstances, realities, and requirements. Based on these findings, we provide 15 recommendations to help the DOD better leverage its existing AI workforce. Our recommendations cover technical and nontechnical talent, apply to each of the services, are broken down by challenge, and are intentionally actionable on short-and medium-term timelines. We begin with a recommendation that sets the strategic foundation, while the remaining recommendations map through the framework: Figure  S1  provides a summary list of our recommendations. We propose talent management actions for the short term to leverage AI talent now and in the medium term to cultivate pipelines of expertise in the future. We also recommend targeted training from operators and acquirers to senior leaders, and rewarding \"rock stars\" that are uniquely positioned to make change happen. Several of our recommendations are also specific to the Joint Artificial Intelligence Center in its role as the central hub and coordinator for the DOD's AI activities. This includes repurposing outdated and underused two-digit function codes already assigned to every DOD billet for the AI archetypes, and to create a forum to harmonize service-level AI workforce initiatives and learned best practices. This policy brief focuses on the DOD's best asset: its existing workforce. We hope this report provides a new perspective to advance the DOD's strategic goals. We recognize that some of our recommendations are intertwined with bigger issues that are difficult, time-consuming, and potentially costly to change. However, we believe that all of our recommendations are critical to effectively leveraging and sustaining the DOD's AI workforce. \n Introduction In 2019, the U.S. Department of Defense (DOD) released a summary of its Artificial Intelligence Strategy, titled \"Harnessing AI to Advance Our Security and Prosperity.\"  4  The strategy elevated AI as a game changing technology that the United States must be ready to lead on, stating: \"The Department of Defense's (DoD) Artificial Intelligence (AI) Strategy directs the DoD to accelerate the adoption of AI and the creation of a force fit for our time. . . . We will harness the potential of AI to transform all functions of the Department positively, thereby supporting and protecting U.S. servicemembers, safeguarding U.S. citizens, defending allies and partners, and improving the affordability, effectiveness, and speed of our operations.\" The strategy outlines the department's goal for wide-scale adoption of AI tools, applications, and capabilities, which will require a major shift in current force structure and attitudes regarding the understanding and appropriateness of AI and its applications. The Joint Artificial Intelligence Center (JAIC) was tasked to lead coordination of AI activities throughout the department, including responsibility for \"cultivating a leading AI workforce.\" Effectively implementing this strategy will require both an AI-able and AI-ready DOD workforce. This includes technical talent and nontechnical talent across all ranks and grades. DOD personnel will need to work together with common understanding to develop, acquire, integrate, and utilize AI capabilities across the force. However, growing, cultivating, and sustaining AI talent has been a challenge. Senior DOD officials, national security policymakers, and defense think-tank reports have all publicly stated that the DOD is struggling to recruit and retain the AI workforce it needs.  5  While there are several key challenges, reports abound of the DOD's inability to compete with the private sector in terms of salary. This report takes a different approach to addressing this challenge: better leveraging the DOD's existing cadre of AI talent. We believe--and our research confirms--that the DOD has a wealth of AI-able talent already in its ranks. Making changes to empower this talent will therefore go a long way to advance the DOD's strategic AI goals. \n Research Approach The DOD's AI workforce is not easily measured or evaluated.  6  As an emerging technology that has only recently proliferated in adoption, thanks to advances in computing, cloud storage, and big data, the federal government's occupation classification system has not yet caught up. This extends to the uniformed services, who are still finalizing the implementation of the 2015 congressional legislation related to identifying and tracking cyber talent.  7  Given this complexity, to enhance the reliability of our assessment, we scoped this study to the DOD's armed services-U.S. Army, U.S. Navy, U.S. Air Force, U.S. Marine Corps, and U.S. Space Force. We scoped the study in this way, instead of the entire DOD enterprise, when initial research revealed that the DOD is taking varied and wide-ranging approaches to managing AI talent.  8  To best assess the DOD's AI workforce posture, we employed a mixed-methods approach. For our qualitative assessment, we interviewed 31 experts across each of the services and key offices of interest in the Office of the Secretary of Defense. We conducted several OSD interviews because we believe these offices offered important and valuable perspectives related to either overseeing or directly working with the services' AI talent. We used both existing networks as well as snowball sampling to obtain our set of interviewees, with targets for the range of roles and responsibilities we wished to include across services. Table  1  below breaks down our interviewees by service and in OSD. Importantly, we note these interviews are not representative of the full range of stakeholder perspectives and experiences. In keeping our study reasonably bounded, we did not interview equivalent positions across each of the services as our goal was to obtain the range of approaches to defining, identifying, and tracking AI talent. 9 It became clear during our research that approaches vary widely both within and across services; every person interviewed had a unique perspective. The range of initiatives and opinions regarding the DOD's AI talent were so varied that it is likely a truly comprehensive assessment would include interviewing every office within every unit of every command for each service. 10 However, given the constraints of our study this was neither reasonable nor realistic. We instead opted to interview personnel across different types of commands and positions, uniformed and civilian, in leadership, technical, and nontechnical positions. To analyze our interviews, we conducted a thematic assessment using NVivo coding software. We created a thematic codebook based on the interview protocol, research objective, and other notable discussion topics (see Appendix A for the list of discussion topics and thematic analysis codebook). We coded each interview using this codebook to analyze discussions specific to relevant topics or groups of interest. We analyzed selected individual codes through several lenses: in aggregate (all interviews), by service (for those with more than one respondent), and for uniformed interviewees versus civilians. We embed our findings as part of our discussion in each chapter of this report. For our quantitative assessment, we considered several sources. First, we analyzed microdata from the U.S. Census Bureau's 2018 American Community Survey (ACS) on industry employment as well as top employers listed in LinkedIn Talent Insights for key AI occupations. The occupations considered were related to CSET's definition of the AI workforce, outlined in a separate report. 11 After we had a complete list of occupations to consider, using both previous CSET research and the range of billets listed by our interviewees, we then analyzed civilian employment microdata from the Office of Personnel Management. Our final list included 12 OPM occupations, which we descriptively analyzed by service and over the last five years (2016-2020). Unfortunately, no analogous data is publicly available for the uniformed component. 12 There are certainly challenges and limitations with our approach. First, as noted above, we were not able to interview personnel in equivalent positions across the services. While this allowed for a range of perspectives, we cannot make absolute and unequivocal comparative statements across the services. Second, interviews were semi-structured, and as such, did not all cover exactly the same topics or questions. The team shared the same list of discussion topics with interviewees, as developed from our interview protocol, in advance of interviews. However, due to differences in perspective and expertise, some discussion topics were more relevant than others. Third, some services were more heavily represented than others in our interview sample, due to sampling strategy, service size, and service investment in AI workforce development. This could affect the balance of our findings and recommendations. Finally, we were limited by the availability of quantitative data on DOD manpower and personnel, particularly regarding the uniformed service component. \n Defining \"AI Workforce\" in the DOD To understand what AI talent the DOD has, we must first define the DOD's \"AI workforce.\" Just as there are several interpretations of what constitutes AI, so too are there varying interpretations of who makes up the AI workforce. This was verified by the range of interpretations of \"AI talent\" and \"AI workforce\" among our interviewees. While some described AI talent as the very technical core working actively on designing AI applications, others described the entire technical and nontechnical teams of talent that are needed to design, develop, acquire, and deploy AI tools or AI-enabled capabilities safely and effectively. For this paper, we take the perspective that it takes an entire team--from the coders to the end users--to achieve mission success. Therefore, we interpret an AI workforce to include AI technical and nontechnical talent. We consider the broader range of personnel both actively and indirectly engaged in deploying AI-enabled tools and capabilities because all of these personnel matter if AI is to be widely adopted across the DOD enterprise.  13  That is, AI adoption is a team sport. Our definition therefore consists of the set of uniformed and civilian personnel engaged in the design, development, testing, implementation, maintenance, and acquisition of AI-enabled tools and capabilities: -Technical talent: those with the knowledge, skills, and abilities to engage in the design, development, and deployment of AI or AI-enabled capabilities. 14 \"For this paper, we take the perspective that it takes an entire team--from the coders to the end users--to achieve mission success. Therefore, we interpret an AI workforce to include AI technical and nontechnical talent.\" -Nontechnical talent: those in roles that complement technical talent, including acquisition personnel and program managers.  15  Previous CSET research identified two types of technical talentthose who are or could be directly engaged in the design, development, and deployment of AI, and those with the requisite knowledge, skills, and abilities (KSAs) that could perform these functions with minimal additional training (\"AI-adjacent\").  16  For example, computer scientists, software developers, database architects, operations researchers, and data scientists fit in the first of these types, while electrical engineers, mechanical engineers, and aerospace engineers fit in the second. However, for the purposes of this study we include all relevant technical talent jointly. 17 While we did not directly measure leadership talent--those making budget, strategy, or other programmatic decisions for uniformed and civilian personnel--it was clear from our discussions that this talent is also critical to the DOD's AI adoption. Therefore, we do acknowledge their importance in this context and include both middle management and senior leadership in our assessment of challenges and recommendations. \n The DOD's Hidden AI Workforce This study began by exploring the premise that the DOD has more AI, AI-adjacent, and AI-capable talent than senior leaders in the DOD realize. The indicators that we explored, as described here, validate this premise. This runs counter to the prevailing wisdom in national security policy circles that the DOD is unable to compete for AI talent. To the contrary, we find that the DOD has an able workforce with AI and AI-adjacent skills; the problem is that it is underidentified, undervalued, and underleveraged. Previous CSET research identified the national security community as one of the top employing industries of both technical and nontechnical AI talent.  18  Looking at microdata from the 2018 ACS, our analysis found that about 8 percent of technical talent and between 6-8 percent of nontechnical talent were employed in public administration. The importance of the national security community as a top employer of AI talent was even more apparent when we considered additional technical occupations such as electrical engineers, mechanical engineers, aerospace engineers, and operations researchers. For these occupations, we found half of the top employers were entities considered part of the national security community. Table  2  showcases the presence of the DOD and the defense industrial base (DIB) as a top employer of AI talent across a range of prominent AI occupations. The complete list of all occupations and their top 10 employers is provided in Appendix B. Finally, we asked nearly all of our interviewees whether they agreed with our premise that the DOD had more AI and AIadjacent talent than realized.  23  All of the interviewees we asked agreed with our hypothesis, regardless of how they defined AI talent. Many also described the large range of billets this talent could be found in, along with identifying other reasons why the services are not fully leveraging existing AI talent. We discuss this further in a subsequent section of this report. Given the strong AI and AIadjacent workforce already employed by the DOD, the question then becomes how to identify and leverage this talent. Where is this talent hiding, why is it hidden, and what can the DOD do about it? We address these questions in turn. \"Given the strong AI and AI-adjacent workforce already employed by the DOD, the question then becomes how to identify and leverage this talent.\" The DOD's Approach to Identifying AI Talent Our findings suggest the problem is not a lack of AI-talent in the DOD but rather an ineffective approach for identifying and leveraging AI talent. Each service is approaching the issue in different ways, both across and within the services. Most services do not have a formalized process to identify talent within their ranks. Here we discuss our findings by providing highlights from each approach. AI talent is-formally or informally-overseen by a variety of organizations across the services. This is summarized in Table  3 . The Air Force Directorate of Studies, Analyses and Assessments (AF/A9) shares responsibility with the Air Force Directorate for Strategy, Integration and Requirements (AF/A5) for the U.S. Air Force, while the Office of Naval Research (ONR) is the central entity for the U.S. Navy. The U.S. Army has two task forces working from different perspectives, the Army Talent Management Task Force (ATMTF) under the Office of the Deputy Chief of Staff for Personnel (G-1), and the U.S. Army Artificial Intelligence Task Force (AITF) under Army Futures Command (AFC). U.S. Marine Corps efforts are spearheaded by the Manpower and Reserve Affairs Department, and the U.S. Space Force has no known specific entity for AI talent management yet as they are still setting up.  24  Finally, the JAIC is the DOD's central coordinating hub for AI strategy, including workforce. \n Table 3. AI Talent Tracking Organization by Service Note: Service-level research laboratories (e.g., ONR, ARL, AFRL) also have independent efforts ongoing to identify their own civilian AI talent, but these efforts are not necessarily service-wide. Source: CSET and MITRE tabulations. Interviewees highlighted the lack of an effective means to quickly and accurately identify AI talent. Different efforts and strategies are currently employed or are in development, and interviewees indicated substantial room for progress. For example, service members often rely on word-of-mouth recommendations and referrals to find and access necessary talent. Another practice is to identify career fields or educational credentials that may not be AIspecific, but might indicate that the service member has requisite knowledge or skills that could be applied to AI. We also learned about a machine learning (ML) approach to identifying digital talent in which personnel files and other available documentation for uniformed and civilian personnel are scraped for data and coding skills. \n Civilians Across all services, interviewees highlighted a similar range of occupations linked most closely to civilian personnel with AI and AI-adjacent skills. 25 These were operations researchers, mathematicians, computer scientists, information technologists, and various disciplines of engineers.  26  While there is no guarantee these occupations include all AI expertise in the civilian force, nor can we state how many within these occupations have AI expertise, we are confident that a large share of technical AI and AI-adjacent expertise are contained in these categories. Additionally, we learned that informal networks play a huge role in identifying AI talent, coordinating internally by word-of-mouth referrals and personal connections. \n U.S. Air Force The Air Force currently identifies AI talent through several means, but is planning to implement a skills-based approach. To date, identification happens through informal networks in addition to tracking relevant educational degrees and select career fields. For example, the most commonly cited uniformed career field applicable to AI by our interviewees was \"operations research analyst.\" However, operations research analysts do not have all of the technical KSAs needed for AI, and discussions are underway for what additional career fields may be included. Recent years have also seen a sustained push in the Air Force to coordinate AI activities and elevated the need for AI and related talent identification. The USAF-MIT AI Accelerator, an Air Force partnership with the Massachusetts Institute of Technology, established in 2019, is the central hub for Air Force AI.  27  The emergence of software factories--de facto centers of coding excellence--across the service also created a need to identify AI and AI-adjacent talent. Our interviewees noted this involved asking manpower, personnel, and services support to the Air Force for recent graduates of relevant programs along with informal networking. Networking includes reaching out to those having done tours in Education With Industry, those working in other software factories, 28 and those working on publicly known grassroots coding communities such as Airman Coders. One coding factory, BESPIN, also manages Digital University, which allows airmen to obtain online coding certifications through thirdparty providers. Going forward, plans are in place to create AI-adjacent specialty experience identifiers (SEIs), which are two-digit codes for special skills and experiences attached to each billet.  29  This particular effort stems from the implementation of a service-wide Computer Language Proficiency Test that treats software coding skills similar to foreign languages in terms of incentive pay. The Air Force will use the results to identify service members with coding skills and assign them a SEI. \n U.S. Army The Army has taken a different approach to identifying AI and AIadjacent talent. The G-1 ATMTF employed natural language processing to conduct keyword extraction from résumés, position descriptions, personnel files, and Github repositories in an attempt to categorize the Army's AI workforce. This scan identified a \"data workforce\" of approximately 27,000 uniformed and civilian service members. 30 Additional measures were taken to ascertain what KSAs individual service members possess, including interviews, résumé reviews, and analysis of other special identifiers. Career fields commonly associated with AI were operations research and systems analysts, and information network engineers.  31  The ATMTF is also working on several new AI-relevant talent initiatives.  32  One such initiative is a pilot program to create small \"lab sprint teams\" where a small cadre of AI talent is available for consultation for a given command. Another is the Talent Based Career Alignment program, created to increase high-performing officer retention by curating career paths to officers based on their stated interests. One path specific to AI is the data science master's program at Carnegie Mellon University in Pittsburgh. The Army also developed the Integrated Personnel and Pay System (IPPS-A) which the ATMTF plans to use to build out a database of their data workforce. The AITF is working on several workforce pilot programs that include identifying and leveraging AI talent.  33  One program is the recently established Army Software Factory, which includes a partnership with Austin Community College to train service members in coding.  34  Another program is an AI scholar program through Carnegie Mellon University to graduate officers, noncommissioned officers, and civilians with master's degrees in data science or data engineering, or a certification in an AI cloud technician program.  35  Our discussions revealed, however, that there were challenges identifying participants and that deadlines for military sign-off were after the deadlines for university applications. 36 Moreover, once the initial assignment with AITF was complete, it remained unclear if participants would continue on assignments that leveraged this expertise. At least one U.S. Army Reserve unit is also taking on the initiative to better identify, classify, and assign AI and AI-adjacent talent. Discussions revealed about one-third of the members in the 75th Innovation Reserve Command (75IC), tasked with future force strategic planning in partnership with AFC, could be considered AI talent. To more effectively leverage this talent, 75IC has two approaches. First, members are given a generic occupational specialty code, enabling assignment flexibility. Second, 75IC is developing an internal Civilian, Education, Skills, Experience, and Certifications database in which they enter every soldier's and civilian's qualifications for tracking (similar to what the Army atlarge is attempting to create with IPPS-A). This will allow for more targeted assignment matching, particularly for personnel with AI expertise. \n U.S. Navy The Navy follows a similar practice as the Air Force and Army in that they use a mix of occupational codes and specialty identifiers to identify personnel who may be considered AI talent. Interviewees observed that the most commonly cited career specialty for both officers and civilians with AI skills by our interviewees was operations research analysts. While this career specialty does not directly equate to \"AI,\" advanced operations research (OR) degrees and related technical degrees were the closest identifiable and trackable proxy for AI-or ML-related expertise. This is particularly true for uniformed officers since the Navy Officer Occupational Classification System assigns a primary designation by broad functional area.  37  The Office of the Chief of Naval Operations, Assessments Division (OPNAV N81), within Naval Strategy, Resources, and Plans (N5/N8), tracks Naval Postgraduate School (NPS) graduates in OR and other relevant fields closely, and works with program offices to guide graduates into relevant billets. Service members with these degrees are given a subspecialty code, and efforts are also underway to track them in subsequent assignments. However, interviewees noted this pool of graduates is small, and split between the Army, Navy, and Marine Corps. For civilians, the Naval Information Warfare Center Pacific is the primary organization for talent identification, in coordination with ONR. While interviewees did not mention AI-specific talent tracking efforts, they did acknowledge that through word-ofmouth and project research the organization could identify the few who were working in AI/ML. In addition, faculty and doctoral candidates at NPS with relevant AI/ML experience can also be called upon as needed through informal networks within the Navy. \n U.S. Marine Corps The Marine Corps recently drafted a new AI workforce development plan to create military occupational specialty codes for AI, although the occupations coded are mostly AI-adjacent instead of purely AI.  38  One new code will be for data scientists, separate from the existing operations research MOS. The goal is to allow for greater workforce management and career tracking for these technical skills. They also have a goal to create \"flyaway teams\" for units to access AI talent. These teams will be modeled after similar teams in their intelligence operations for career fields with limited supply and high demand. On the civilian side, we learned that only two Marine Corps personnel are considered research scientists with AI expertise. This is because the Marine Corps relies heavily on the Navy for its research and development, so it does not have the R&D infrastructure or staffing levels of other services. However, if AI talent is needed, informal networks within the Naval Warfare Centers are used to find individuals with the KSAs necessary to meet identified needs, or similar talent from within ONR is tapped by personal request. \n U.S. Space Force As a new service, the Space Force has done relatively little in the way of AI talent identification and tracking. Moreover, the service is still closely tied, in terms of talent, to the Air Force.  39  One interviewee highlighted that, currently, the only way to assemble AI talent in the Space Force is to use personal connections similar to those noted above in coordination with Air Force A5/A9 to gather a small group of experts. However, the recent guidance from the U.S. Space Force Chief of Space Operations, Gen. John Raymond, also highlighted above signals a potential shift toward digital talent training, tracking, and identification.  40  In addition, the Space Force has their own software factories--Kobayashi Maru, a \"command and control (C2) program,\" and Space CAMP, described as \"the premiere software factory for Space Force\" that is based in Colorado Springs and led by the same office in charge of the new digital workforce strategy. 41 \n Joint Artificial Intelligence Center At the department level, the JAIC is the centralized hub for all AI activities. Currently, the JAIC is coordinating with each service to design and implement an archetype system for AI talent to deliver tailored AI education for different segments of DOD personnel. These six archetypes are: Lead AI, Drive AI, Create AI, Facilitate AI, Embed AI, and Employ AI.  42  Their first pilot program began in October 2020, with the goal of using these archetypes to design and provide education and training programs at scale. In the longer term, the JAIC hopes to use these education platforms as means of creating unifying certifications that signal competency in specific AI fields, based on the archetypes. However, there is also difficulty in mapping these archetypes to existing manpower categorization systems. \n The DOD's Civilian AI Workforce There are no explicit AI occupation codes in the uniformed or civilian force. However, as noted above, interviewees were able to isolate a reasonable range of civilian occupations linked most closely to personnel with AI and AI-adjacent skills.  43  In addition to the occupations cited, our own research of news media and USAJobs position descriptions revealed additional codes related to technical and nontechnical talent, such as program and management analysts and contracting and purchasing agents. While this may not include all DOD civilian AI, AI-adjacent, or AIcapable billets, given the misalignment of OPM's occupational codes and hiring practices of individual services, we are confident these occupations comprise a large share of this target workforce.  44  OPM collects and publishes civilian personnel data for the federal government. OPM also oversees the occupational taxonomy used for civilian positions (i.e., billets), which is different from the taxonomies used by other occupational data-reporting federal agencies (e.g., the U.S. Bureau of Labor Statistics, and the U.S. Census Bureau).  45  According to OPM data, as of 2020 the DOD civilian workforce consisted of roughly 770,000 workers. Most of these personnel were housed in the Army (over 256,000), followed by the Navy (226,000), then the Air Force (176,000), and finally the Office of the Secretary of Defense (112,000) which includes the DOD's independent agencies. Using OPM data on the occupations AI technical and nontechnical talent likely fill, it is possible to provide a general description of the DOD's civilian \"AI workforce.\" Our list is composed of 12 technical and nontechnical occupation codes--what we call \"AI occupations\" for the purpose of this analysis--listed below.   Note: Total employment over 2016-2020 in these 12 occupations increased 4 percent in OSD, 17 percent in the Army, 10 percent in the Air Force, and 13 percent in the Navy. Total DOD employment in these 12 occupations increased 12 percent, going from 140,000 in 2016 to 157,000 in 2020. In general, the technical talent included in these occupations grew more than the overall civilian workforce, in ways we might largely expect. For example, the Air Force greatly increased its number of aerospace engineers and operations research analysts, and the Army increased its number of mechanical engineers. Notably, the Navy had double-digit increases for each of these occupations, with the exception of electronics engineering, aerospace engineering, and purchasing. Moreover, this data suggests computer scientists--a small but critical AI occupation--are increasing rapidly in the DOD. The increase in computer scientists is much faster than the DOD's total civilian workforce, as shown in Figure  3 . The Air Force led the increase in computer scientists, increasing its cadre 43 percent between 2016 and 2020, with the Navy and Army also experiencing large increases. Interestingly, however, as of 2020, Navy totals included almost twice as many computer scientists as the Air Force, and more than three times as many as the Army. We next examined OPM microdata of civilian DOD personnel to show how these 12 AI occupations were distributed among the services. Some occupations (e.g., IT management, operations research, and contracting) are well-distributed, with no single service housing more than one-third of that specified talent. Conversely, other occupations are vastly overrepresented in certain services. For example, over 50 percent of computer scientists, mechanical and electrical engineers, and computer engineers are found in the Navy. A closer examination of this talent's distribution within each service reveals most are concentrated within certain commands. For example, the majority of computer scientists can be found in OSD's Defense Information Systems Agency, the Air Force Materiel Command, the Army Futures Command, and in the Navy's Naval Information Warfare Systems Command, as well as its Naval Sea Systems Command. A similar distribution is found for the Air Force's, the Army's, and the Navy's operations research analysts and computer and mechanical engineers. As Figures  4 and 5  demonstrate, the Navy by far has the most civilian talent in these occupations compared to the other services. As shown in Figure  4 , the Navy's employment total in these 12 occupations is nearly double that of the Air Force's or the Army's. Moreover, in some of these occupations, the Navy's workforce surpasses that of the other services. For example, the Navy has more civilian computer scientists than the other three services combined. The Navy also has the highest share of its civilian workforce in these occupations. As shown in Figure  5 , these 12 occupations comprise over a quarter (26 percent) of the Navy's civilian workforce. This compares to 14 percent for the Army, which has the largest civilian workforce yet the smallest share comprised of these occupations. This data supports our hypothesis that the DOD may have more AI and AI-relevant expertise than is commonly understood. Claims that the DOD lacks this talent is not borne out in this data, particularly given notable expansions of its computer science workforce. Moreover, our interviews suggest technical military billets are also distributed among the services, and many service members (both officers and enlisted) may be in nontechnical billets but possess AI skills. \n Why the DOD's AI Workforce is Hidden: Challenges Our interviews revealed many reasons why the DOD's AI and AIadjacent workforce remains hidden and not effectively leveraged. These reasons are myriad and complex; intertwining strategy, people, processes, and technology. Here we provide a discussion of these reasons as raised by our interviewees. While some services experienced varying degrees of these challenges and barriers, our analysis finds that most, if not all, of these reasons affect each service to some degree. \n Service-Level AI Strategies What is considered AI, and its relative strategic importance, varies widely across services and within organizations. While interviewees across services noted an emphasis on AI coming from senior leadership, each service approached actual AI adoption and prioritization of adoption slightly differently. Moreover, interviewees inconsistently defined both AI and AI talent. Some interviewees grouped autonomy with AI, some equated ML to AI, while others followed a stricter definition of AI (many in the latter category were AI practitioners). A two-tier problem exists within the services: leadership wants AI without proper understanding of the technology, while practitioners often focus on their area of application to the exclusion of broader application areas. Interviewees noted at least one consequence of prioritizing AI without understanding AI: an overinflated importance placed on software developers. They noted that while important, software developers are but one part of the AI workforce. Many professionals with AI-adjacent skills who are not software developers by training have pivoted to deliver key AI capabilities for their local missions. Yet there was concern that software developers were given undue influence that risked dominating AIrelated policy conversation or decisions. There is no agreement across the DOD on who constitutes the AI workforce and how much investment and prioritization should be given to cultivating AI talent in-house. Some interviewees believed AI expertise should be left to contractors, federally funded research and development centers, and university affiliated research centers, having some general technical operational, maintenance, and acquisition talent in-house. Others acknowledged the need for some top-tier AI expertise within their organization, since there are some capabilities that are best suited for direct development by uniformed or civilian personnel. When we asked interviewees about the JAIC's categorization of AI talent into six archetypes, many agreed with the premise, but none had thought about how or how well that translated into practice for their service.  47  These differences also factor into why each service's AI workforce is overseen by varying commands or groups, making any attempt to standardize AI talent management extremely difficult. Larger strategic challenges stem from a lack of a shared view of AI's capabilities, limitations, and mission suitability. Different views on AI have led to a range of approaches across DOD elements. For example, many interviewees discussed challenges, including how others within their organization had a lack of understanding of AI, whether as a tool or a technique; skepticism about the current state of AI as being ready for wide-scale adoption across processes and operations; availability of adequate technology and investment in modernizing data systems; reliance on contractors for technical capabilities; organizational culture; and the ability for both the planning, programming, budget, and execution processes and requirements processes to adequately include AI as a nonmaterial capability. 48 Some of these are tactical and discussed in more detail below, but the aggregation of these is an organizational reality. Regardless of cause, the result is an inconsistent prioritization of investing in and adopting AI, which directly results in inconsistently identifying, leveraging, and cultivating a dedicated AI workforce. Nevertheless, how each service defines and categorizes AI will factor into how services define their AI workforce, determine who is a part of this workforce, and decide on how this talent is trained, assigned, and promoted. We do not assert that services should adopt a uniform definition of AI,  49  we do contend that it is not necessary for services to have the same mix \"Nevertheless, how each service defines and categorizes AI will factor into how services define their AI workforce, determine who is a part of this workforce, and decide on how this talent is trained, assigned, and promoted.\" of AI talent. In fact, given differences in needed AI capabilities and applications for AI across services, it is beneficial for services to have a different mix of AI talent. \n OSD and Service-Level Culture Organizational culture for each service represents another barrier to leveraging the DOD's existing AI workforce. Interviewees noted that organizational strategy and organizational culture are intertwined. However, culture embodies its own challenge when it limits forward-thinking and encourages misaligned performance incentives for leveraging AI talent. While each service has a distinct culture, this challenge affects them all. Organizational culture is critical not only in influencing organizational AI strategy, but in how services approach talent management. In addition to congressional mandates regarding end-strength quotas and officer promotions, 50 culture is a key determining factor for how each service's uniformed and civilian personnel are classified, assigned, and promoted. While culture is broader than the AI talent topic, conversations about AI talent are necessarily part of this discussion and will be affected by culture. For example, we heard from multiple interviewees that having warfighting experience was not only celebrated, but necessary for promotion. By design, officers change assignments every two to three years. These rotations normally include a mix of operational tours, broadening experiences, education and instructional tours, and service headquarters or joint operation tours, in addition to assignments that align with the officers' primary career field. The guiding mindset is for service members to \"As a result, little opportunity outside of self-initiative exists to build, use, and maintain technical expertise. Moreover, technical talent can be less highly regarded than their combat peers because of service culture.\" be combat or operation-ready, a concept that remains largely defined by traditional warfare. As a result, little opportunity outside of self-initiative exists to build, use, and maintain technical expertise. Moreover, technical talent can be less highly regarded than their combat peers because of service culture.  51  Finally, in another dimension of culture, we heard about the limited risk tolerance and willingness to divert from the traditional, long acquisition processes. There was a sense that it is ingrained in hearts and minds of service members that to receive positive performance evaluations, and maximize promotion potential, existing rules and procedures require strict adherence. \n People Each service has an inconsistent interpretation of its AI workforce, AI workforce needs, and in their ability to identify AI talent. AI workforce leaders have different, often nonoverlapping visions of talent composition (whether it be uniform, civilian, or contracted out) and number (how many uniformed members should be AI-technically savvy). Consequently, skilling and training military personnel in AI becomes ad hoc and falls heavily on individual drive. Interviewees disagreed not only on how much technical talent is needed, but where in the organization this talent should be located. Still, they agreed the current personnel structure is not designed to cultivate and leverage technical expertise. The degree to which contractors are relied upon also came up on multiple occasions, with some interviewees lamenting the lack of in-house capability while others believing it was the right division of labor given inhouse funding and other personnel constraints. We also heard accounts of enlisted service members taking the initiative to build AI expertise, but commissioned officers being prioritized for the few opportunities that exist to use these skills.  52  A fair number of enlisted service members either already have AI-adjacent coding skills coming in, or took their own initiative to learn them through extracurricular online coding boot camps and programs via entities like the Air Force's Digital University. Several interviewees noted that for any software factory or other opportunity to build or use coding skills, interest from the enlisted community far exceeded demand. However, interviewees also noted limited opportunities for enlisted talent to use these skills.  53  For opportunities that do include enlisted service members, very few slots are dedicated to them. The stigma that enlisted service members' expertise is inferior to officers remains a cultural barrier that limits effectively leveraging the greater pool of AI and AIadjacent talent. In an interview about the USAF-MIT AI Accelerator, its deputy director recognized the importance of enlisted service members in advancing its goals, saying \"The only way we get to that level is including more enlisted folks. . . . There's a lot more meat on the bone with them and the future of AI is in the enlisted, not in the officers. The officers will play a very important role in this, but we will need the enlisted by sheer numbers.  54  Formal and informal AI and AI-related communities have been an important convening mechanism for talent, but remain frail and fragmented. In a more positive light, interviewees noted several instances of individuals taking the initiative to form and maintain technical communities of practice. These communities-both uniformed and civilian, formal, and informal--were established and growing within existing cultures across services. These include, for example, Platform One, Airmen Coders, and Naval Applications of Machine Learning. However, they are scattered across the DOD. While these communities of practice provide a powerful organizing and convening function, unless they are part of a larger network, or supported organizationally, they risk frailty. Moreover, many lack sustained funding streams. This could explain why, in spite of these communities, interviewees continued to stress the importance of personal connections in identifying and accessing technical talent. This included interviewees in strategically technical positions, such as in a software factory. The role of personal connections in building AI-related research portfolios was particularly salient for research lab personnel who felt disconnected from the \"big\" service. While some efforts exist, for example a \"dotted line\" or informal chain of command between Army Research Lab and Army Futures Command, the \"valley of death\" remained very real for other researchers.  55  Researchers without personal connections remained relatively siloed in their efforts, with no clear project impact. For example, interviewees at the Naval Information Systems Warfare Command noted a rotation program into the larger NWIC did not result in a lasting connection. Many ultimately did not come back to the NWIC and moved on. \n AI rock stars are critical players in cultivating an AI workforce. They thrive in spite of, not because of, their organizations' incentives. AI \"rock stars\" are essential but extremely rare within each service. These rock stars not only know AI and have AI or AIadjacent expertise, but they are in positions to enact the policy changes needed to create a formal AI workforce. That is, they have AI expertise, see the potential of AI and want to build this expertise across the service, have close, positive relationships with senior leadership who support their recommendations, and are working in or directly engaged with headquarters personnel or talent management. We were fortunate to speak with a few such rock stars. When we asked these interviewees what would happen when they left their current assignment, they did not know how their initiatives would continue. Yet without these rock stars strategically positioned, it is not clear how needed changes to personnel policy will happen. \n Senior leaders championing AI adoption and workforce development are essential, yet their understanding and appreciation of AI is highly inconsistent and support often ends when that leader rotates out. Interviewees also discussed senior leaders, noting their importance in whether AI workforce initiatives were prioritized or not. While many interviewees stressed the importance of executive education in helping senior leaders appropriately define AI and AI use-cases, they also expressed concern for what happens when senior leaders who champion AI leave their positions. For example, the fate of a planned initiative for six new officer career pathways in the Army Special Operations Forces, two of which are designed to leverage technical expertise, lies with the championship of a senior commander.  56  Additionally, interviewees also expressed that members of senior and middle management may not have the technical, political, and mission awareness of where AI is suitable and appropriate for implementation. Middle management lacks incentive to take risks. AI, as a big unknown, represents a large risk. In terms of leadership, many interviewees discussed the \"frozen middle,\" middle management that--for reasons noted above--lacked incentives to prioritize AI workforce development. Many middle managers are focused on the mission at hand, avoiding political risk, and circumventing distractions that might veer them from their career path or promotion potential. These managers, like senior leaders, may or may not understand what AI is, but unlike senior leaders, lack a sense of urgency to make AI adoption a priority. Finally, there are concerns about the lack of AI literacy across the organization. This is particularly relevant for operators who work or could work with AI-enabled systems, and for acquisition talent that will be involved in the purchase and delivery of such systems. Interviewees questioned what types of training should be available, how such training should be delivered, and who should take it. \n Processes By far the largest issue cited related to leveraging AI talent was the state of uniformed and civilian talent management procedures and practices. These challenges begin before talent is even hired, commissioned or enlisted, and affect each part of the talent management life cycle.  57  Issues are rooted in budgeting, personnel regulations, data infrastructure, culture, and organizational processes. The result is an environment where AI talent is very difficult to identify, measure, track, and promote. This leaves the idea of a clearly defined \"AI workforce\" as just an idea, unless larger changes or adjustments are made. On the uniformed side, few opportunities or incentives exist within the current talent management structure to cultivate and leverage AI expertise. Officers and enlisted members follow a formal process of occupational assignment, training, and duty assignments or rotations that do not prioritize technical expertise. In addition to occupational taxonomies that are mission or operation-centric instead of skills-based taxonomies, each occupation has a defined career pathway. While it seems the Marine Corps demonstrates more talent management flexibility than the other services, it too is plagued by bureaucratic challenges. Our discussions revealed that the timing, eligibility, and tracking for the new MOS codes are not final. Moreover, extensive discussion and planning went into their establishment, including a strong and vocal push by a few \"rock stars,\" hundreds of white papers, and now an AI workforce strategic plan. For civilians, there are at least 12 occupations that comprise the AI workforce, with no clear way to identify AI expertise. When asked what billets civilian AI talent are found in, interviewees noted a range of about 12 occupations (analyzed above). Not only are there limitations of OPM's occupational taxonomy that make AI talent identification difficult, but it is often the case that positions are filled based on what is funded and available. Instead of reclassifying billets, which is a timely process, units are more likely to hire in whatever billet is already available and vacant, tailor the announcement, and when the hire arrives take on the desired roles and responsibilities. We spoke with several civilian practitioners with AI skills who either did not know what billet they were formally hired under or acknowledged their daily work did not actually align with their formal position description. \n AI-related project assignments for civilians are ad hoc, particularly at the research labs. Here we learned of an interplay where researchers appreciated the internal labor market and some autonomy over which projects they worked on. It provided flexibility to work across groups and build networks. However, it was also noted that word-of-mouth and active network cultivation was a major factor, and that these networks tended to be haphazard and tenuous. There is also the issue of not knowing the right people which can result in missed opportunities. This challenge is further exacerbated by not having a centralized repository or database of personnel with AI or AI-adjacent skills that could aid in talent identification. 58 Many interviewees raised serious concerns over current budgeting, requirements, and acquisitions processes that inherently limit AI adoption. Several interviewees noted that discussion about AI-enabled capabilities and AI workforce were moot if AI was not included in requirements. They explained that all roads lead through requirements, and that personnel were incentivized to follow their part in the process. If AI was not included in the requirements, there would not be AI embedded into the capability, nor would funding be available to integrate AI. Also regarding requirements drafting, interviewees noted the potential for disconnect in leveraging AI expertise due to the division of roles and responsibilities across different units or groups. That said, interviewees did note at least one area of progress beneficial to integrating AI--the creation of crossfunctional product teams in the Army Futures Command. These teams coordinate with each other and with the Army AI Task Force on drafting all requirements. \n Technology Interviewees expressed concern over current limitations in service-level data curation, integration, accessibility, and software availability, all needed for AI development and deployment. For example, several interviewees discussed the difficulty of obtaining licenses for current AI and AI-adjacent software packages, due to funding constraints and outdated policies still in place in some parts of the DOD for software acquisition. As with any technical skill, there is critical dependency on the tools required to deliver capability. So successfully onboarding AI talent requires an equal focus on ensuring the right tools and development environment is in place and is adequately supported. However, without appropriate access to software and tools, the process for leveraging and retaining AI talent is further hindered. Some interviewees reiterated points made in other related reports on the DOD's AI readiness pertaining to the DOD's fragmented and legacy data systems and infrastructure. 59 While less directly related to AI workforce as other observed challenges, this impacts the ability of the DOD to \"do AI.\" That matters for attracting and retaining civilian talent, as well as for creating opportunities to train and leverage uniformed AI expertise. \n Finally, interviewees noted substantial reliance on contractors for data warehousing and analytics because of inherent challenges. They believe substantial technology investment is needed to modernize systems, although we note the public discussion of ongoing initiatives now taking place in this regard.  60    \n Leveraging the DOD's AI Talent: Recommendations Given that safe and effective adoption of AI requires a team, we believe the DOD should include the relevant technical and nontechnical talent as part of any effort to better identify, measure, track, and leverage AI talent. Moreover, we believe some AI capabilities cannot be contracted to the private sector. This implies that services should cultivate some amount of top-tier technical talent in-house. Our recommendations therefore consider all of the above talent, in addition to managers and leaders who are important change agents. The recommendations presented here stem from the challenges identified above. We begin with an overarching strategic recommendation as it sets the foundation. We then present recommendations in relative order of actionability to promote the identification → experimentation → implementation → harmonization framework. We also note that our recommendations focus on leveraging the potential of the DOD's existing workforce, so we do not focus on attracting and recruiting new talent although this is also clearly important. It may not be the case that each service-level recommendation is appropriate for each service. Moreover, some of these recommendations are complex, intertwined with bigger issues that are difficult, time-consuming, and potentially costly to change. However, we believe that all of our recommendations are critical to effectively leveraging and sustaining the DOD's AI workforce. \n Service-Level AI Strategies Without actively prioritizing AI adoption in each service, there will be limited prioritization and investment in their AI workforce. Each service should therefore clearly and publicly create a strategy for cultivating its AI talent. \n Rec 1: Clearly and explicitly make AI adoption and organization a priority by having each service create an AI workforce strategy and implementation plan. Given the variation in organizational AI across services, it is important that each service create a strategy specific to their AI workforce. This should include defining technical and nontechnical AI talent, the associated roles and responsibilities needed to execute the service-level AI strategy, existing AI talent gaps as determined by an AI workforce assessment to major commands and component functional community managers (CFCMs), and plans for how this talent will be identified, assigned, and promoted. It should also consider how the defined AI workforce relates to the JAIC's six AI workforce archetypes, as defined in the AI Education Strategy,  61  and be consistent with the DOD's AI Strategy. \n People \n Armed with an AI workforce strategy, services must now put in place building blocks to effectively identify and leverage its AI workforce. This includes short-term measures in AI talent identification, harnessing the AI potential of the enlisted corps, connecting AI communities of practice, rewarding AI \"rock stars,\" and investing in AI education and training. \n Rec 2: Engage in short-term AI talent identification and tracking techniques using a skills-based assessment. Services should engage in immediate technical and nontechnical AI talent identification, leveraging, training, and sustainment practices until a more formal strategy and implementation plan is in place.  62  There are practices the services can execute in the near term. Engaging the expertise and experience of the enlisted community is a critical part to leveraging AI talent--investing in an AI workforce cannot be exclusive to officers. Enlisted service members work side by side with officers in addition to being on the front lines of all operations, from back office processes to combat and mission support. \n Rec 4: Encourage and coordinate AI and AI-related communities of practice by establishing a dedicated communication platform that connects to CoPs across services. Formal and informal networks dominated how many of our interviewees identified AI talent. However, one-off connections are neither scalable nor sustainable. Leveraging AI talent should therefore institutionalize what works, by encouraging, facilitating and coordinating AI communities of practice. Services could bring together their disparate network of AI and AIadjacent practitioners, along with domain experts who possess AI or related knowledge, by coordinating their efforts. Creating a marketplace for AI and AI-related communities, by providing a convening platform, will further enable cross-pollination of ideas and applications in addition to reducing duplicative efforts. This could increase AI diffusion while creating a community and culture of belonging for this talent, particularly for more organic or smaller communities. Moreover, it could help with efforts to identify AI and AI-adjacent talent, particularly if they are domain experts not in technical billets.  63  Rec 5: Connect, empower, and reward AI \"rock stars\" by (1) creating a dedicated intraservice working group for servicenominated and appointed AI rock stars, and (2) establishing an annual DOD AI Achievement Award that comes with a monetary prize. Our interviews made the importance of AI \"rock stars\" in advancing AI workforce goals clear. These rock stars are strategically skilled and positioned to make these recommendations--to make sure AI talent is identified and leveraged--a reality. Rock stars are not necessarily senior officers, but rather come in all ranks and are immersed in the day-to-day operations of their command or service. We recommend the working group be unlike traditional working groups that consist of senior or senior-appointed leaders and carefully managed agendas. Instead, we recommend a group that operates outside of the usual DOD hierarchy, where members are nominated by relevant CoP representatives. The group could be chaired by the JAIC and meet quarterly to raise issues and share best practices.  64  The DOD AI Achievement Award should be awarded to a peer-nominated service member or civilian who has advanced AI literacy, AI workforce goals, or safe, ethical, and trustworthy AI adoption in a significant way. Rec 6: Invest in AI education as professional military education (PME) for senior leaders.  65  Senior leadership training would be strategic-focused for understanding their service's organizational AI strategy, along with appropriate uses and limitations of AI. Many of our civilian interviewees did not know their assigned billet, and relied on word-of-mouth for identifying other AI-able talent and AI-related projects. Interviews suggest a large share of civilian talent with AI or AI-adjacent skills is housed in the services' respective R&D organizations, particularly research labs. Effectively leveraging this talent starts with formal talent identification, followed by a mechanism to match this talent with special projects and assignments. Skills databases should be widely searchable across the service enterprise. \n Technology AI talent needs access to data, software, and equipment to design, develop, and deploy AI, and therefore leverage their skills and expertise. However, interviewees consistently mentioned this as a persistent challenge. The DOD should empower talent by investing in the needed digital infrastructure. Rec 12: Create a mechanism for all DOD personnel to submit their experienced challenges regarding software, data access, computing facilities, internet connections, classification barriers, data cleaning and integration, and other issues creating barriers to AI use and adoption to service-level chief data officers, chief technology officers, and chief information officers. On a regular basis, service CDOs, CTOs, or CIOs should review these challenges. Upon reviewing these submissions, the services should (1) create a course of action with milestones to address the final assessment of these challenges, and (2) use them as an input to prioritize technology investments. \n Leveraging the JAIC The JAIC provides an important infrastructure that should be harnessed. These recommendations leverage the coordinating function of the JAIC to help the services and, importantly, to harmonize AI workforce identification across disparities in approaches to AI talent management. Rec 13: The JAIC should convene an annual AI workforce conference to facilitate communication of ongoing activities and share best practices. To help establish best practices, the JAIC should provide support for services to evaluate initiatives and pilots for outcomes and potential to scale. Our discussions revealed several different approaches across and within services, including pilots and initiatives in progress. For example, we heard a range of current and planned approaches including: -Small cadre of talent in labs or flyaway teams. -Special experience identifiers. -New technical career fields with separate promotion tracking. -Skills or competency repository to facilitate project matching. -Making 5-10 percent of each career field technical experts. -Continued reliance on contractors for technical expertise. Leveraging AI talent effectively involves having services evaluate these approaches, and share lessons learned and best practices. Conference participants should include: (1) those leading initiatives highlighted in the \"rock star\" working group recommended above, (2) those leading other initiatives recognized by AI-related CoPs, by the JAIC through its routine service-level engagement, and as nominated by commanding officers. This could not only elevate existing initiatives but reward services for taking on new approaches with an evaluative component. Rec 14: The JAIC should issue service-level guidance on crafting AI workforce strategies and implementation plans in line with the JAIC publishing its overarching AI workforce goals. In line with the JAIC's efforts to ensure an AI-ready DOD workforce, the JAIC should facilitate service-level workforce strategies. At a minimum this includes issuing guidance, and at its best would include providing active consultation.  70  Basic guidance could provide freedom for the services to design the approach best for them; for example, articulating that it is not necessary to put all technical AI talent under one or two career codes, encourage service-wide competency repositories, and inventory and coordinate AI-related communities of practice. Rec 15: Create two-digit function codes for each of the JAIC's archetypes that are standardized across services to harmonize AI talent identification and tracking. Function codes for all DOD billets already exist but are rarely used. 71 A benefit of these codes is that they can be applied to joint commands and assignments, and enable a DOD-wide inventory. We propose updating these codes in a way that embodies the JAIC's six archetypes, along with embedding flexibility in these codes through regular review to accommodate other emerging technologies. \n OSD and Service-Level Culture Across our interviewees, one theme was clear: the need to create an environment where all personnel are working toward the safe, ethical, and trustable use of AI across applications and capabilities. However, similar to funding and requirements, this is an extremely complex issue that is outside the scope of this paper. We instead suggest that implementing the above recommendations will help relieve existing cultural tensions related to leveraging the latent potential of the DOD's existing cadre of AI talent. Finally, related to culture and similarly outside the scope of this study, we note interviewees repeatedly stressed the challenges raised by existing budgeting and requirements processes. They noted the importance of having the incentives, funding, and ability to integrate consideration of AI-enabled capabilities, tools, and applications into solutions as part of leveraging and cultivating AI talent. The sentiment was: \"no requirement, no AI, no need for an AI workforce.\" While we do not propose a recommendation here, we do acknowledge its importance. \n Alignment with NSCAI Recommendations Our recommendations are consistent with those provided by the National Security Commission on Artificial Intelligence's (NSCAI) Final Report when it comes to leveraging the DOD's existing AI talent. This is particularly true regarding the education and training of enlisted, officer, and civilian personnel across the DOD and in the need to expand access to tools, datasets, and infrastructure. An additional recommendation covered by NSCAI, but not in this report, is complementary to our recommendations. This is the recommendation to update the Armed Services Vocational Aptitude Battery Test to include computational thinking. Our recommendations differ from the NSCAI report in that they do not explicitly recommend the establishment of new digital career fields in the services for software developers, data scientists, and artificial intelligence, as well as new civilian occupational series for software development, software engineering, knowledge management, data science, and artificial intelligence. We recommend the services approach AI and AI-related talent identification, assignment, and management using the best approach for them, which may vary by service. We instead propose creating a functional code that will be DOD-wide and align with the JAIC's six archetypes. We also attempt to keep our recommendations to actionable items in the short-and mediumterm, appreciating the DOD's current realities and constraints. It may be the case that in the longer term, the establishment of new digital career fields could prove useful to the services. Other NSCAI recommendations focus on recruiting civilian technical new talent whether through novel pathways into government, modifying hiring practices, or expanding existing programs. For example, considerable focus is put on creating Digital Corps, National Reserve Digital Corps, a Digital Service Academy, and an AI Scholarship for Service program. As this report is focused on leveraging existing talent, those recommendations are outside of the scope of this study. \n Summary: Challenges and Recommendations \n Service-Level AI Strategies What is considered AI, and its relative strategic importance, varies widely across services and within organizations. This means there is disagreement on who makes up the AI workforce and how much investment and prioritization should be given to cultivating AI talent in-house. Rec 1: Clearly and explicitly make AI adoption and organization a priority by having each service create an AI workforce strategy and implementation plan. \n People Each service has an inconsistent interpretation of its AI workforce, AI workforce needs, and in their ability to identify AI talent. Rec 2: Engage in short-term AI talent identification and tracking techniques using a skills-based assessment. We also heard accounts of enlisted service members taking the initiative to build AI expertise, but commissioned officers being prioritized for the few opportunities that exist to use these skills. Rec 3: Create more opportunities for enlisted service members who possess AI and AI-related skills. \n Formal and informal AI and AI-related communities have been an important convening mechanism for talent, but remain frail and fragmented. Rec 4: Encourage and coordinate AI and AI-related communities of practice. AI rock stars are critical players in cultivating an AI workforce, but thrive in spite of, not because of, their organizations' incentives. Rec 5: Connect, empower, and reward AI \"rock stars.\" Senior leaders championing AI adoption and workforce development are essential, yet their understanding and appreciation of AI is highly inconsistent and support often ends when that leader rotates out. Rec 6: Invest in AI education as PME for senior leaders. Middle management lacks incentive to take risks. AI, as a big unknown, represents a large risk. Rec 7: Invest in AI education as PME for middle management. There are concerns about the lack of AI literacy across the organization. Rec 8: Build broader AI-literacy across the DOD by creating training and exposure opportunities. \n Processes By far the largest issue cited related to leveraging AI talent was the state of uniformed and civilian talent management procedures and practices. Rec 9: Conduct service-wide inventories of uniformed and civilian technical billet and assignments. On the uniformed side, few opportunities or incentives exist within the current talent management structure to cultivate and leverage AI expertise. Rec 10: Experiment with career pathway pilot initiatives that can be measured, evaluated, and iterated upon. Eventually, establish clearly defined career pathways with promotion potential up to and including the ranks of general/flag officer. For civilians, there are at least 12 occupations that comprise the AI workforce, with no clear way to identify AI expertise. Moreover, AIrelated project assignments for civilians are ad hoc, particularly at the research labs. Rec 11: Leverage civilian AI talent by creating a central competency and skills database. \n Technology Interviewees expressed concern over current limitations in servicelevel data curation, integration, accessibility, and software availability, all needed for AI development and deployment. Rec 12: Create a mechanism for all DOD personnel to submit their experienced challenges regarding software, data access, computing facilities, internet connections, classification barriers, and other issues creating barriers to AI use and adoption to service-level CDOs, CTOs, or CIOs. \n Leveraging the JAIC The JAIC provides an important infrastructure that should be harnessed. These recommendations leverage the coordinating function of the JAIC to help the services and, importantly, to harmonize AI workforce identification across disparities in approaches to AI talent management. Rec 13: The JAIC should convene an annual AI workforce lessons learned conference. Rec 14: The JAIC should issue service-level guidance on crafting AI workforce strategies and implementation plans in line with the JAIC publishing its overarching AI workforce goals. Rec 15: Create two-digit function codes for each of the JAIC's archetypes that are standardized across services to harmonize AI talent identification and tracking. \n Conclusion Central to the adoption of AI tools, applications, and capabilities in the Department of Defense is the ability to recruit and retain an AIable and AI-ready workforce. Indeed, growing and cultivating an AI-ready workforce is a top priority in the DOD's AI Strategy. However, stakeholders within and outside of the DOD have repeated the claim that the department struggles to attract the necessary talent. Usually, these discussions emphasize competition for top-tier technical talent with industry. This report flips that narrative. Previous CSET research found the DOD was already a top employer of AI talent. Armed with this finding, we conducted interviews with 31 key experts across the services and OSD and analyzed civilian personnel data from OPM to understand the state of the DOD's AI workforce. In addition to revealing a hidden cadre of AI talent within the DOD, our analysis provides a detailed understanding of challenges and opportunities for the DOD's AI workforce policy. Our research finds that while the department has a cadre of AI and AI-adjacent personnel, this talent remains hidden. Our analysis uncovers several reasons for this. At the service-level, there is not a consistent approach to defining or prioritizing AI. This translates to lack of investment in cultivating an AI workforce. On a more tactical level, we identified three types of challenges: people, processes, and technology. The challenges pertaining to people, processes, and technology are diverse. Regarding people, each service struggles to define and identify their AI workforce. The enlisted community lacks opportunities to use their AI expertise. Convening communities of \"However, stakeholders within and outside of the DOD have repeated the claim that the department struggles to attract the necessary talent.\" interest are frail and fragmented, and individual AI rock stars within each service succeed despite, rather than because of, existing systems. Moreover, senior leaders' understanding and appreciation of AI is highly inconsistent, and middle management lacks the incentive to take risks in acquiring and adopting AI-capabilities. Existing processes within the DOD offer another set of challenges. The overriding issue in this category is the current state of uniformed and civilian talent management procedures and practices. Today's performance incentive structures do not align with the pursuit of or leveraging of AI expertise at any level. Moreover, civilians with relevant skills and abilities are spread across multiple occupations and AI-related project assignments for civilians are ad hoc. Technological barriers create yet additional obstacles. Interviewees highlighted difficulty in accessing necessary software, tools, and data platforms, and serious concerns related to data reliability and ownership were raised across the department. Together, these challenges make it difficult to effectively identify and leverage the DOD's existing AI workforce. Our analysis found that each service has its own office or organization that oversees what it loosely considers \"AI talent,\" either formally or informally. Across these entities, AI talent is most commonly identified through informal networks and personal connections, with interviewees across all services highlighted heavy reliance on such networks to find suitable talent. While other methods of identification exist, they are imperfect, such as equating an existing career field or educational credential with AI expertise. Interviewees across the enterprise stressed the lack of an effective, systematic means to identify AI talent quickly and accurately. Our research suggests that properly identifying and leveraging this talent could go a long way in meeting the DOD's stated AI workforce goals. To this end, we provide 15 recommendations. While the initial recommendation provides a strategic umbrella for AI workforce prioritization, the remaining are structured to map the following framework: Identification → Experimentation → Implementation → Harmonization Each step in the framework offers guidance for sustained talent management process change. With identification, the organization should ask: How do we define AI talent and where might we find it within my organization? In experimentation, how might we try out new ideas that leverage AI talent, evaluate their success, and iterate? For the third step, implementation, how might we agree upon and deliver an approach? Finally, with harmonization, how might the service-level approaches work together to enable enterprise-wide AI talent assessment and empowerment? Our recommendations should promote agility in talent management. Experimentation, evaluation, and iteration of pilot initiatives are critical in determining the best approaches to AI workforce cultivation for each service. The same process allows for flexibilities as priorities and missions evolve. We also recommend targeted training from operators and acquisition personnel to senior leaders, and rewarding department \"rock stars\" that are uniquely positioned to make change happen. Several of our recommendations are also specific to the JAIC in its role as the central hub and coordinator for the DOD's AI activities. This includes repurposing outdated and underused two-digit function codes already assigned to every DOD billet for the AI archetypes, and to create a forum to harmonize service-level AI workforce initiatives and learned best practices. Ultimately, the DOD should work toward a future where enlisted service members, officers, and civilians have a pathway to apply their AI and related expertise to deliver impactful outcomes that advance the DOD's AI Strategy. We hope this report provides a new perspective to advance that strategic goal. With the exception of reaching out to each service's headquarters personnel offices (X1). We were able to communicate with personnel in each but for the Navy's N1, which did not respond to our request. 10 That is, offices across the strategic, operational, and tactical levels of command for each service as well as every combatant commander, research laboratory, and resource sponsor (e.g., those with authority to fund AI-enabled tools and capabilities). Even then, such an assessment would still likely miss some ongoing efforts. 11 Gehlhaus and Mutis, \"The U.S. AI Workforce.\" 12 Analysis of uniformed personnel data is a cumbersome process. It requires special access and permissions which only select entities can legally obtain; further, all requests must also go through a time-intensive human subjects review process on the front end and back end for publication. Moreover, an assessment of military occupation and specialty codes, which are publicly available, quickly show there are no clear \"AI\" occupational groups to cleanly analyze, similar to the civilian component. The hidden nature of this talent through classification taxonomies is part of the basis for this report. 14 Called \"Technical Team 1\" and \"Technical Team 2\" in previous CSET reports.  15  Called \"Product Team\" and \"Commercial Team\" in previous CSET reports.  16  Gehlhaus and Mutis, \"The U.S. AI Workforce.\" The DOD does not have the same occupational classifications as used for these categories. We therefore considered the most closely matching occupation and uniformed specialty classifications. 17 Similarly, previous CSET research identified two categories of nontechnical talent--product team and commercial team. However, given the nature of DOD occupational classifications, along with the associated roles and responsibilities, we considered a subset of each nontechnical category and we considered these subsets jointly. For example, the DOD does not employ user experience designers or sales engineers as separate occupations, and on the civilian side many product and commercial team occupations are likely rolled into \"program and management analysts\" (OPM code 0343). \n System (NAICS) , where Public Administration is NAICS 92 and the national security and international affairs subsector is NAICS 928. The range for nontechnical talent is based on CSET's nontechnical categories; 6 percent of Product Team employment was in public administration compared to about 8 percent for Commercial Team occupations. 20 This is double the national average; the national security sector comprised 1.5 percent of total U.S. employment in 2018. 21 Diana Gehlhaus and Ilya Rahkovsky, \"The U.S. AI Workforce: Labor Market Dynamics\" (Center for Security and Emerging Technology, April 2021), https://cset.georgetown.edu/publication/u-s-ai-workforce/. See Appendix D. 22 For this analysis, we used a compiled list of job titles associated with each occupation. 23 There were a handful of interviewees for which this question was not appropriate given their expertise or current position. 24 However, a guidance document from General John Raymond indicates that a new Technology and Innovation Office within the Space Force will be leading the digital transformation of the service, which includes workforce components. See U.S. Space Force, \"U.S. Space Force Vision for a Digital Service,\" U.S. Department of the Air Force, May 2021, https://media.defense.gov/2021/May/06/<PHONE>/-1/-1/1/USSF%20VISION%20FOR%20A%20DIGITAL%20SERVICE%202021%20( 2).PDF. 25 Civilians are managed independently from uniformed personnel and have a separate occupational hierarchy. However, they are also classified in \"functional communities\" alongside uniformed personnel, for which Table  3  applies. This discussion focuses on civilians' talent management separate from functional communities, and we note each service's civilian corps had overlapping approaches. 26 Our own review of OPM job postings also shows a fair number of data scientists are hired as management and program analysts. We list the types of engineers in the next chapter. https://aia.mit.edu/about-us/. 28 Coding factories include Kessel Run, BESPIN, Space CAMP, Red 5, and Tron, to name a few. 29 They will retire outdated codes and replace them with coding language and potentially other technical skills identifiers. 30 Efforts are also underway to coordinate with the reserve component for analysis. 31 Related areas of concentration can be grouped together into broader functional areas (FA). 32 Related to several new authorities granted in the 2019 National Defense Authorization Act. 33 Much of which is in coordination with ATMTF and TRADOC. 34 Applications for the first cohorts far exceeded the number of available slots, and the opportunity was made available to all servicemembers. Interviewees noted the huge demand indicated large latent technical expertise among the enlisted corps.  training resources for its service members. This is because the Space Force was set up to be a small and lean service, and to leverage its parent service. https://www.ai.mil/docs/2020_DoD_AI_Training_and_Education_Strategy_and_I nfographic_10_27_20.pdf. The JAIC is the DOD's central coordinating hub for artificial intelligence issues and strategy, including workforce. They are currently working with the services to pilot educational training for the leadership archetype, which informs senior leaders on appropriate definitions and uses of AI. 48 We note the analogous challenge for software acquisition, which now follows a separate process. 49 Whether there should be one common definition of AI is an active discussion. We appreciate there are advantages and disadvantages to this and ultimately believe different mission needs across services will drive different AI use cases, measures of performance, and degrees of human-machine interaction. However the services define or operationalize \"AI,\" what matters is that each service align their AI workforce definition and investment accordingly.  50  The key pieces of legislation are: NDAA (end-strength), and DOPMA and ROPMA (officer promotions for the active duty and reserve components, respectively).  51  In the Navy, every other assignment is on a ship, in a combat-ready position. In the Army, it is required that officers command a battalion for promotion to colonel, and officers are trained from the beginning to command troops in combat. In the Air Force, \"pilot culture\" has historically celebrated pilots over other fields, limiting promotion opportunities for more technical fields like operations researchers and engineers. 52 Enlisted service members comprise 80 percent of the uniformed service component, making them a larger talent pool. Moreover, they are equally if not more involved in operational roles that could feed back into designing and developing usable and trustworthy AI applications.  55  The \"valley of death\" refers to the gap between R&D and actual deployment in the field. Many prototypes or research insights never make it into the hands of operators. 56 As of this writing, it was unclear if the effort would proceed after a period of uncertainty, even though the details had been carefully planned and vetted internally among stakeholders.  57  The talent management lifecycle refers to all talent engagement stages from attracting and recruiting to promotion and attrition. 58 Some organizations we spoke with were working to build a competency database, but reliant on self-reported data and not specific to AI. 59 See, for example: Defense Innovation Board, \"Software Acquisition and - -1102 -Contracting -0830 -Mechanical Engineering -0850 -Electrical Engineering -0854 -Computer Engineering -0855 -Electronics Engineering -0861 -Aerospace Engineering -1550 -Computer Science -1520 -MathematicsTogether, these 12 occupations accounted for roughly 20 percent of the DOD's civilian workforce in 2020, totaling about 157,000 personnel across the services. As shown in Figure1, IT management comprises the largest share of this talent at 5 percent. 46   \n Figure 1 . 1 Figure 1. Identified AI Occupations as a Share of the DOD Civilian Workforce, 2020 \n Figure 2 . 2 Figure 2. AI Occupational Employment Change Over Time, by Service (2016-2020) \n Figure 3 . 3 Figure 3. Percent Growth of the DOD's Workforce for Selected Groups (2016-2020) \n Figure 4 . 4 Figure 4. AI Occupational Distribution Across DOD Services, 2020 \n Figure 5 . 5 Figure 5. The Navy Has the Highest Share in AI Occupations \n Figure 6 . 6 Figure 6. Summary of Why the DOD AI Talent is Hidden \n 13 A goal stated in the DOD's AI Strategy. See U.S. Department of Defense, Summary of the 2018 Department of Defense Artificial Intelligence Strategy. \n 38 Military Occupational Specialty (MOS) codes define the occupational taxonomy used by the Marine Corps and by the Army for enlisted service members. Other services use different occupational taxonomies. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n Table 1 . 1 DOD Interviews by Service Number of Service Interviewees U.S. Army 12 U.S. Air Force 9 U.S. Navy 4 U.S. Marine Corps 2 U.S. Space Force 1 OSD 3 Total 31 Source: CSET and MITRE tabulations. \n Create more opportunities for enlisted service members who possess AI and AI-related skills. Centers of Excellence, and recognized AI-related communities of practice. -Tracking AI-related educational attainment in both personnel files and the central repository. -Ensuring timelines of training enrollment deadlines align with internal assignment processes (e.g., submitting university application materials before the due date). Rec 3: We recommend empowering the enlisted community by creating opportunities to upskill for AI, bring AI expertise into assignments, and participate in strategic and tactical conversations surrounding AI deployment and operation. Such opportunities include increasing the number of rotational opportunities with industry (e. These include: -Competency assessments issued by each functional community. -Processes for self-reporting of AI and AI-related skills on personnel files. -Recording AI and AI-related skills identification through special identifiers added to billet information. -Creating a service-wide skills database or repository (instance to existing platform) accessible at a minimum to X1, component functional community managers (CFCMs), OSD functional community managers, established AI g., Education With Industry), tours at service-level AI Centers of Excellence and software factories, and increased eligibility for pilot AI education and training programs. \n Invest in AI education as PME for uniformed and civilian middle management. It would also include understanding what personnel, resources, and technology needs are and what flexibilities may need to be approved to access them.This training would be more tactical, teaching managers not only about the importance of prioritizing AI deployment, but learning how to know when AI may be an appropriate solution, along with how to manage projects that build AI-enabled capabilities into research, requirements, budgeting, and acquisitions. Training would also include what personnel and technology needs are helpful to be successful on an AI project, how to best leverage personnel, and how to obtain any necessary senior-level approvals. Rec 7: Rec 8: \n Inspire broader AI literacy across the DOD by creating training and exposure opportunities. share and submit ideas to integrate AI into current processes and operations. Processes Current uniformed and civilian talent management practices, including career field assignment, tour of duty assignments, and promotion pathways do not effectively cultivate and leverage the DOD's existing AI talent. We recommend each service experiment, evaluate, iterate, and eventually implement realigned talent management processes for AI and AI-related career fields to incentivize AI adoption and deployment. Importantly, the same approach may not be what is best for each service. That will depend on service AI strategy, force structure and deployment, operational needs, and organizational culture. Rec 9: Effectively leveraging talent includes empowering them with the knowledge needed to trust and responsibly engage in AI-enabled tools and solutions. Specifically, general and targeted AI training could consist of: -DOD-wide: AI literacy course for all personnel, (i.e., civilian and uniformed). -For technical and technically-apt talent: AI coursework and credentialing from military colleges (e.g., Digital University) and third party microcredentials or certifications that are recorded in personnel files. 66 -For acquisition talent: AI in requirements course and certificate at DAU. -For requirements talent: Provide training so that requirements personnel can document consideration of AI- embedded capabilities for each non-materiel solution -For operators: Training modules as part of basic military training (uniformed only) along with routine PME required for promotion. -AI engagement and exposure: Create a contest/challenge for officers, enlisted, reservists, guardsman, and civilians to \n Conduct service-wide inventories of uniformed and civilian technical billet and assignments. This is the first step to ensuring that individuals with technical talent are placed in assignments where they may leverage their expertise. Our findings show technical billets are scattered across career fields. Civilians can be found in at least 12 occupation codes. For both uniformed and civilian personnel, assignments do not always correspond to career fields. A workforce assessment is needed to understand where existing technical assignment opportunities exist. The assessment should come in two parts: (1) taking an inventory of currently available technical assignments, and (2) surveying commands for opportunities to create new technical billets or transition existing positions. Surveys should also ask commands to assess high-demand positions or currently unfilled technical positions. Rec 10: \n Experiment with career pathway pilot initiatives that can be measured, evaluated, and iterated upon. Eventually, in accordance with timelines laid out in each services' AI workforce strategy, establish clearly defined career pathways with promotion potential up to and including the rank of general/flag officer. Pilot initiatives should include the necessary stakeholder engagement and regular reporting for lessons learned and best practices. Ultimately, implemented approaches to AI talent management should incentivize AI-relevant training, assign talent in a way that uses their AI expertise, and provide routine engagement with industry, academia, combatant commands and joint duty assignments for continued partnership and learning. If appropriate, we recommend creating a separate promotion board for non-leadership technical career fields with technically-oriented promotion criteria (e.g., in lieu of command posts or broadening experiences).Technical career tracks can vary by service and career field (e.g., working within existing fields, creating new career fields, or both), exist for active and reserve components, and include opportunities for leadership track individuals to switch to a technical track. For example, recent years have shown examples of the services both creating new niche technically-oriented career fields and modifying existing ones. In 2020, the Army established an \"Enterprise Marketing and Behavioral Economics\" career field, for which selected officers must have education and expertise in marketing and data analytics.67  In 2019, Air Force Operations Research Analysts were moved out of the engineering and acquisitions functional community to create a separate technical career track.68     \n Rec 11: Each service should leverage its civilian AI workforce by creating a central competency and skills database that allows civilian personnel to self-identify technical expertise, project or research experience, technical publications, and links to Github or other repositories. 69 \n 1.1. Defining AI talent [Includes high level view of what kinds of personnel are and/or work on AI teams.] 9 3.3. AI Talent Desired Talent End State [Whether the services need everyone to know how to code, or just 1.2. Identifying AI Talent [How talent is identified a few PhDs, or any other level in between] currently] 3.4. Informal Communities 1.2.1. Talent Marketplace or equivalent database 3.5. Building Trust 1.2.2. Informal communities (e.g., Airmen Coders) 4. JAIC [all things JAIC-related] [includes whether interviewee engaged with 5. Quotes/Stories them] 1.2.3. Word of mouth 6. Miscellaneous 1.2.4. Advanced tracking 1.2.5. Other 1.3. Classifying AI Talent [Formal methods the DOD uses to classify their AI talent] 1.3.1. Codes [specialty codes, experience identifiers, skills, etc.] 1.3.2. Billets 1.3.3. Assessments and Certifications 1.4. Assigning AI talent [to projects/programs] 1.5. Recruiting and Retaining AI talent 1.6. AI Talent in other organizations [For use when interviewee references AI talent in other organizations than the one they are representing] 1.7. Use/Reliance on Contractors 2. Why is the DOD's AI Talent Hidden? [Why/How organization struggles to define, identify, classify, assign, and recruit AI talent.] 2.1. Organizational AI Strategy 2.1.1. Defining AI and AI Literacy [Ability of all levels of personnel to discuss, comprehend, and make informed decisions regarding AI] 2.1.2. Organization's AI Activities/Projects 2.1.3. Organizational AI Strategy 2.2. Challenges in Organizational AI \n 35  See Talent Management Task Force Annual Report, 2019-2020: U.S. Army, \"Talent Management Annual Report 2019-2020,\" https://talent.army.mil/wpcontent/uploads/2021/03/TMTF-Annual-Report.pdf.36 One interviewee noted the Army's timeline for submitting a candidate to a university's graduate degree program occurs after the due date of the university.37  For more, see Navy Personnel Command, \"Manual of Navy Officer Manpower and Personnel Classifications: Volume I,\" U.S. Navy, October 2020, https://www.mynavyhr.navy.mil/Portals/55/Reference/NOOCS/Vol1/Entire%20M anual%20I%2073.pdf?ver=rpCPz8Vt4b52x7ETh2GP8A%3D%3D. \n\t\t\t Note: We restricted our geographic search to the U.S. data for computer research scientists, software engineers, mathematicians, statisticians, data scientists, and project management specialists accessed December 2020. Data for aerospace engineers, electrical and electronics engineers, mechanical engineers, operations research analysts, and purchasing managers accessed March 2021.Source: LinkedIn Talent Insights, CSET analysis. \n\t\t\t This means we are not including OSD, independent DOD subagencies, combatant commands, or the DOD's intelligence branch. \n\t\t\t Gehlhaus and Mutis, \"The U.S. AI Workforce.\" Here \"national security \n\t\t\t See USAF-MIT AI Accelerator website, \"About Us,\" USAF-MIT AI Accelerator, \n\t\t\t For example, the Space Force is leveraging the uniformed recruiting and civilian hiring infrastructure of the U.S. Air Force. It is also sharing education and \n\t\t\t This also includes working with the Defense Innovation Unit (DIU) on a new", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/The-DODs-Hidden-Artificial-Intelligence-Workforce.tei.xml", "id": "e2eed2593c8b137c5a08459b1fbc0989"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "A logically uncertain reasoner would be able to reason as if they know both a programming language and a program, without knowing what the program outputs. Most practical reasoning involves some logical uncertainty, but no satisfactory theory of reasoning under logical uncertainty yet exists. A better theory of reasoning under logical uncertainty is needed in order to develop the tools necessary to construct highly reliable artificial reasoners. This paper introduces the topic, discusses a number of historical results, and describes a number of open problems. 1. See, e.g., the work referenced by  Hutter et al. (2013) .", "authors": ["Nate Soares", "Benja Fallenstein"], "title": "Questions of Reasoning Under Logical Uncertainty", "text": "Introduction Consider a black box with one input chute and two output chutes. The box is known to take a ball in the input chute and then (via some complex Rube Goldberg machine) deposit the ball in one of the output chutes. An environmentally uncertain reasoner does not know which Rube Goldberg machine the black box implements. A logically uncertain reasoner may know which machine the box implements, and may understand how the machine works, but does not (for lack of computational resources) know how the machine behaves. Standard probability theory is a powerful tool for reasoning under environmental uncertainty, but it assumes logical omniscience: once a probabilistic reasoner has determined precisely which Rube Goldberg machine is in the black box, they are assumed to know which output chute will take the ball. By contrast, realistic reasoners must operate under logical uncertainty: we often know how a machine works, but not precisely what it will do. General intelligence, at the human level, mostly consists of reasoning that involves logical uncertainty. Reasoning about the output of a computer program, the Research supported by the Machine Intelligence Research Institute (intelligence.org). Published as Technical report 2015-1. behavior of other actors in the environment, or the implications of a surprising observation are all done under logical (in addition to environmental) uncertainty. This would also be true of smarter-than-human systems: constructing a completely coherent Bayesian probability distribution in a complex world is intractable. Any artificially intelligent system writing software or evaluating complex plans must necessarily perform some reasoning under logical uncertainty. When constructing smarter-than-human systems, the stakes are incredibly high: superintelligent machines could have an extraordinary impact upon humanity  (Bostrom 2014) , and if that impact is not beneficial, the results could be catastrophic  (Yudkowsky 2008 ). If that system is to attain superintelligence by way of self-modification, logically uncertain reasoning will be critical to its reliability. The initial system's ability must reason about the unknown behavior of a known program (the contemplated self-modification) in order to understand the result of modifying itself. In order to pose the question of whether a practical system reasons well under logical uncertainty, it is first necessary to gain a theoretical understanding of logically uncertain reasoning. Yet, despite significant research going back to  Loś (1955) ,  Gaifman (1964)  and before, continued by  Halpern (2003) ,  Hutter et al. (2013) ,  Demski (2012) ,  Christiano (2014)  and many, many others, 1 this theoretical understanding does not yet exist. It is natural to consider extending standard probability theory to include the consideration of worlds which are \"logically impossible\" (such as where a deterministic Rube Goldberg machine behaves in a way that it doesn't). This gives rise to two questions: What, precisely, are logically impossible possibilities? And, given some means of reasoning about impossible possibilities, what is a reasonable prior probability distribution over them? This paper discusses the field of reasoning under logical uncertainty. At present, study into logically uncertain reasoning is largely concerned with the problem of reasoning probabilistically about sentences of logic. Sections 2 and 3 discuss the two problems posed above in that context. Ultimately, our understanding of logical uncertainty will need to move beyond the domain of logical sentences; this point is further explored in Section 4. Section 5 concludes by relating these problems back to the design of smarter-than-human systems which are reliably aligned with human interests. \n Impossible Possibilities Consider again the black box, with the Rube Goldberg machine inside. An agent reasoning using standard probability theory is environmentally uncertain; they do not know which Rube Goldberg machine is in the box. Probability distributions assign probabilities to some set of \"possibilities,\" and when reasoning probabilistically under environmental uncertainty, the set of possibilities is all Rube Goldberg machines consistent with observation that could fit in the box. What is the set of possibilities considered by a logically uncertain reasoner? They may know which Rube Goldberg machine is in the box, without knowing how that Rube Goldberg machine behaves (for lack of deductive capabilities). The machine, following the laws of logic and physics, deposits the ball in only one of the two chutes, but a logically uncertain reasoner must consider both output chutes as a \"possibility.\" Logically uncertain reasoning, then, requires the consideration of logically impossible possibilities. What sort of objects are logically impossible possibilities? What is the set of all impossible possibilities, to which probabilities are assigned? In full generality, this question is vague and intractable. But there is one setting in which it is natural to consider logical impossibilities, and that is the domain of formal logic itself: consider agents that are uncertain about the truth values of sentences of logic. Indeed, the study of logical uncertainty in the literature centers on reasoning according to assignments of probabilities to sentences of first-order logic  (Gaifman 2004) . How does reasoning about logical sentences correspond to reasoning under logical uncertainty in the real world? Sentences of first order logic are extremely expressive: given a description of the Rube Goldberg machine in the black box, it is possible to construct a logical sentence which is true if and only if the machine deposits the ball in the top chute. A reasoner uncertain about whether that sentences is true is also uncertain about the behavior of that Rube Goldberg machine. Logical sentences can also encode statements such as \"this Turing machine will halt,\" or \"this function sorts its input and has time complexity O(log n).\" Thus, while it is ultimately necessary to understand logically uncertain reasoning as it pertains to observation and interaction in the real world, it is reasonable to begin studying reasoning under logical uncertainty by studying probability assignments to logical sentences. Picking any probability distribution over logical sen-tences does not automatically constitute \"reasoning under logical uncertainty.\" Intuitively, logically uncertain reasoning must preserve some of the structure between sentences: if a reasoner assigns probability 1 to φ and deduces φ → ψ (via some complex implication), then the reasoner must assign probability 1 to ψ thereafter. But clearly, not all of the structure between sentences can be preserved, for that would require logical omniscience. Which structure is preserved under logical uncertainty, and how? It is illuminating to first consider the probabilities that a reasoner would assign to sentences of logic if they could preserve all the logical structure. How could a deductively omniscient reasoner assign probabilities consistently to logical sentences? It is not so simple as claiming that omniscient reasoners assign probability 1 to true sentences and 0 to false ones, because logical sentences are not simply \"true\" or \"false\" in a vacuum. It depends entirely on which logical theory obtains: if the domain is numbers and \"×\" is multiplication, then the sentence ∀a, b : a×b = b×a is true; but if the domain is vectors and \"×\" is the vector cross product, then the same sentence is false. There are, in fact, two types of logical uncertainty: uncertainty about the logical theory, and uncertainty stemming from limited deductive capabilities. The first type of logical uncertainty has been studied for many decades, with early work done by  Gaifman (1964) ,  Hacking (1967) , and others. It is not merely a problem of defining symbols: in logic, there are many theories which are \"incomplete,\" meaning that there exist sentences which are not necessarily either true or false according to that logical theory. Consider Peano Arithmetic (PA), which formalizes the natural numbers. PA nails down the definitions of \"+\" and \"×,\" but there are still sentences which are true of the numbers but are not implied by the Peano axioms. The classic example is Gödel's sentence, which roughly claims \"PA cannot prove this sentence\"  (Gödel, Kleene, and Rosser 1934) . This statement is true, but it does not follow from the Peano axioms. A consistent assignment of truth values to every sentence requires a complete logical theory. A complete theory of logic is a consistent set T of sentences such that for every sentence φ, either φ ∈ T or (¬φ) ∈ T . Incomplete theories can be \"completed\" by starting with the set of all consequences of the incomplete theory and then choosing arbitrary (consistent) assignments of truth for each independent sentence. For example, there are many different ways to complete PA, only one of which is \"true arithmetic.\" Identifying true arithmetic is uncomputable, as statements of true arithmetic include statements about which Turing machines halt. 2 Even a deductively unlimited reasoner-that always believes ψ whenever it believes both φ and φ → ψ, no matter how complex and obfuscated the implication-may have uncertainty about which sentences are true or false, via uncertainty about which complete theory is the \"real\" one. A logically uncertain but deductively unlimited reasoner-which knows all consequences of everything it knows, but does not know the \"true\" complete theory of logic-only entertains consistent logically impossible possibilities. It may not know which logical theory corresponds to how sentences act in \"the real world,\" but each \"possible world\" is self-consistent. This provides a partial answer to the question of the nature of logically impossible possibilities: if a reasoner is deductively unlimited, an \"impossible possibility\" is any complete theory of logic. A \"consistent\" assignment of probabilities to sentences, then, corresponds to a probability distribution over complete theories of logic, where the probability of a sentence is equal to the measure of theories in which that sentence is true. 3 This is the standard model of logically uncertain reasoning throughout the literature, used e.g. by  Gaifman (1964) ,  Christiano et al. (2013) , and many others. This is a fine result for deductively unlimited reasoners, but the goal is to understand reasoning under deductive limitations. Deductively unlimited reasoners reason according to consistent impossible worlds, but detecting inconsistencies can be a computationally expensive task. Deductively limited reasoners must entertain inconsistent impossible possibilities. Recent study of reasoning under logical uncertainty has been pushing in this direction  (Gaifman 2004) . Intuitively, deductively limited reasoners must reason according to \"theories\" (assignments of truth values to sentences) that seem consistent so far, discarding hypotheses as soon as a contradiction is deduced. An \"impossible possibility,\" then, could be any assignment of truth values to logical sentences which does not allow a short proof of inconsistency. This intuition might be formalized as follows: Fix enumerations of sentences and proofs in first order logic. Consider all bit strings of some huge length, say Ackerman(10 100 ), and interpret these as assignments of truth to each sentence (a 1 in the n th position claims that the n th sentence is true, and a 0 claims it is false). Now search all strings with length of up to some far larger number, say Ackerman 2 (10 100 ), for proofs in PA that one of these bit strings makes inconsistent claims, and reject any bit string which is found to be inconsistent by this procedure. Assign probabilities to each remaining bit string (according to some prior); this may be used to generate an assignment of probabilities to sentences of length less than Ackerman(10 100 ) according to the measure of bit strings which claim that the sentence is true. This procedure is wildly impractical, but it does seem to allow for satisfactory logically uncertain reasoning using finite (if not limited) deduction. Clearly, the larger the number of sentences (and the larger the number of proofs searched), the more the remaining bit strings will resemble a collection of complete theories which are actually consistent. This process leads to intuitively \"reasonable\" logically uncertain beliefs: no \"theories\" that admit a short proof of inconsistency are considered, but \"theories\" with inconsistencies that are very difficult to deduce may remain. This process corresponds to a finite version of considering all complete theories: it considers bit strings assigning truth values to many sentences, for which there are no reasonably sized proofs of contradiction; deductively unlimited reasoning considers bit strings assigning truth values to all sentences, for which there is no proof of contradiction at all. However, for all that this technique seems intuitively nice, no precise statements about its performance have yet been proven. These techniques shed light on the nature of impossible possibilities in the context of deductively limited reasoning: an impossible possibility is any assignment of truth values to logical sentences which has not yet been proven inconsistent, in some fashion. That is, practical agents may entertain contradictory possibilities, so long as the possibility is discarded once the contradiction is deduced; an \"impossible possibility\" is any assignment of truth to logical sentences which hasn't been found to be consistent so far. It is an open problem to develop more practical techniques than the one above which allow agents to reason as if according to truth-assignments which have \"not yet been found to be inconsistent.\" For further discussion, see  Christiano (2014) . While this answer is somewhat satisfactory, it only answers the question of impossible possibilities as it relates to uncertainty over logical sentences. Realistic reasoning under logical uncertainty will require more than just an ability to assign probabilities to logical sentences; discussion of this point is relegated to Section 4. \n Logical Priors In the context of uncertainty about logical sentences, deductively limited reasoners must approximate reasoning according to some probability distribution over complete theories. This gives rise to a second question: which probability distribution over complete theories should they approximate? Of course, the answer is in part up to us: if we design a system which reasons about the probabilities of logical sentences, which takes questions in the form of sentences and outputs predictions in the form of probabilities, then the question of what logical theory it should use depends entirely upon how we want the questions to be interpreted. For example, if a deductively limited system is built to help its designers reason about some Boolean algebra, then the \"distribution over complete theories\" might be some sort of simplicity prior which assigns probability 1 to the complete theory of Boolean algebras. But what if the system is intended to reason according to some extremely powerful theory of logic (e.g. PA) which is capable of expressing many questions about the real world (e.g. whether the Rube Goldberg machine deposits the ball into the top chute), but for which we do not know the preferred single complete theory (e.g. because the complete theory answers all halting problems)? Then what distribution over complete theories should be used? If the system is supposed to reason according to true arithmetic, then what initial state of knowledge captures our beliefs about that uncomputable theory? This is the problem of logical priors. Intuitively, the problem may seem easy: just choose a maximum entropy (or otherwise weak) prior. Unfortunately, it is not obvious how to construct a weak prior over complete theories. Starting with a maximum entropy prior on logical sentences and refining towards consistency will not suffice: a prior which assigns 50% probability to every sentence places zero probability mass on the set of all complete theories, because there are infinitely many contradictory sentences, and so any infinite sequence of sentences generated by this prior is guaranteed to select at least one contradiction eventually. Hutter et al. (  2013 ) make an early attempt to answer the first question, by defining a logical prior in terms of a probability distribution over sentences which assigns positive probability to all consistent sentences and zero probability to contradictions. A probability distribution of this form allows for the definition of a satisfactory logical prior: The Hutter prior: For each sentence, select a model in which that sentence is true, and in which certain desirable properties hold (the \"Gaifman condition\" and the \"Cournot condition\"  (Hutter et al. 2013) ). Add the complete theory of that model to the distribution with measure in proportion to the probability of the sentence. This prior has many desirable properties, but it cannot be computably approximated: the conditions that Hutter demands of each model (which yield the prior's nice properties) rely on the high-powered machinery of set theory, and it is not possible to computably approximate this prior. That is, there does not exist a computable process refining assignments of probabilities to sentences which converges on the assignments of Hutter's prior in the limit  (Sawin and Demski 2013). 4  4. Remember that a probability distribution over complete theories can be treated as a probability distribution The Hutter prior yields insight into what constitutes a desirable prior, but a study of logical uncertainty in deductively limited systems requires that the prior be approximable. Just as deductively limited reasoners must approximate reasoning about consistent theories (by entertaining inconsistent \"theories\" until a contradiction is deduced), so must deductively limited reasoners start with a prior that does not quite match the (inevitably intractable) intended prior, and then refine those probabilities as they reason. But this process of starting with an incoherent prior (which places probability mass on inconsistencies) and refining it towards some desirable prior (eliminating inconsistencies as contradictions are deduced, and shifting probability mass to better correspond with the \"true\" prior) is precisely the problem of reasoning under logical uncertainty, entire! That is, the approximation of a satisfactory logical prior exhibits, in miniature, all the problems of reasoning according to a probability distribution over sentences. Thus, the definition of a satisfactory approximable logical prior, and the study of its approximations, may yield solutions to the problem of reasoning under logical uncertainty more generally. Unfortunately, it is not entirely clear what it would mean for an approximable logical prior to be \"satisfactory,\" and naïve attempts at constructing computably approximable logical priors have all had undesirable properties.  Demski (2012)  proposes a computably approximable prior that can be generated from any distribution Φ over all sentences. A \"generator\" is used to generate complete theories (by drawing sentences at random from Φ), and Demksi's prior assigns probability to a theory T according to the probability that the generator would generate T . More formally, Demski's generator is given by Algorithm 1. It takes an initial set B of known sentences and a distribution Φ over sentences. It constructs a complete theory T by starting with B and selecting sentences φ at random from Φ. It either adds φ to T (if φ is consistent with T ) or adds ¬φ to T (otherwise). For example, let the base theory be Peano Arithmetic (PA), and let Φ be a simplicity prior over sentences which assigns each sentence φ probability 2 −|φ| where |φ| is the length 5 of φ. Clearly, the simplicity prior does not describe a satisfactory logical prior over sentences, as it puts significant probability mass on short contradictory sentences such as \"0 = 1\". Demski's generator, however, only generates consistent theories, and therefore, it places probability 0 on all contradictions. Similarly, because all sentences of PA are over sentences which assigns probability to sentences in accordance with the proportion of theories in which that sentence is true. 5. Where sentences are encoded in binary, preferably using some encoding of length where the length of φ is the same as the length of ¬φ, so that the prior is not biased in disfavor of negations. \n Algorithm 1: The Demski generator Data: A probability distribution Φ over sentences Data: A base theory B of known sentences Result: A complete theory T begin T ←− B loop φ ←− genrandom(Φ) if T ∪ { φ } is consistent then T ←− T ∪ { φ } else T ←− T ∪ { ¬φ } included in B, the prior only generates theories T consistent with PA, and so it assigns probability 1 on all sentences implied by PA. Now consider a sentence φ which is independent of PA: the probability of this sentence depends upon how often Demski's generator generates a theory T in which φ is true. Clearly, this probability is positive, as with probability 2 −|φ| , φ will be the first random sentence added to T . Similarly, because there is a chance that the first random sentence is ¬φ, the probability of φ will not be 1. Thus, Demski's prior defines a probability distribution over complete theories extending PA. While Demski's prior is uncomputable, Demski (2012) has shown that the resulting prior probability distribution is computably approximable: There is a computable procedure which will output successive approximations of the probability of a sentence φ, converging in the limit to the probability assigned to φ by the uncomputable procedure. Even this computable approximation, however, is not a tractable algorithm; recently,  Christiano (2014)  has proposed an alternative approach to constructing priors which borrows from standard machine learning techniques, making it more likely that the priors developed in this way can be used in realistic algorithms. These priors, however, have some undesirable properties. For example, starting with B as the empty set, Demski's prior places zero probability on the set of complete theories where PA holds. 6 Agents approximating Demski's would not be able to learn Peano Arithmetic: Demski's prior, while approximable, is not weak enough. In order for a reasoner using Demski's prior to believe PA, it must be included in the base theory B. This reveals a related issue: there are two different ways to \"update\" Demski's prior on a sentence φ. The prior can either be completely regenerated from the base theory 6. Specifically, Demski's prior places zero probability mass on any theory which is not finitely axiomatizable. The induction schema of PA consists of infinitely many axioms, all independent of each other. With probability 1, Demski's generator will eventually select the negation of one of these axioms from Φ. B ∪ { φ }, or it can be conditioned on φ (by removing all theories in which φ is false). These two different updates result in two different posterior probability distributions. Consider the posterior probability of a sentence ψ such that both ψ and ¬ψ are consistent with φ. If the prior is regenerated from B ∪ { φ }, then the resulting posterior still places at least 2 −|ψ| probability on ψ, because this is the probability that ψ is the first sentence selected at random from Φ. But if the prior is conditioned on φ, then it may be the case that the posterior probability of ψ is arbitrarily low. For example, if ¬ψ → φ, then all theories with ¬ψ will have φ, and if it is also the case that almost all theories with ψ also contain ¬φ, then the posterior probability may place arbitrarily small positive probability on ψ. In other words, conditioning the prior on φ favors explanations for φ, while regenerating the prior does not alter the Φ-based lower bound probability of any sentence that does not directly contradict φ. This double update is strange. An agent reasoning using Demski's prior would treat facts that it \"learns\" (through observation and conditioning) differently from facts that it \"always knew\" (sentences from the base theory). This phenomenon is not well understood. Why does the double update occur? Is it undesirable? Can it be avoided? These questions remain open, and it is possible that answers to these questions will lend insight into the generation of satisfactory approximable logical priors. It is not at all clear what it would mean for a logical prior to be \"satisfactory,\" in the first place: part of the problem is that it is not yet clear what desiderata to demand from a logical prior. Candidate desiderata include: 1. Coherence: A prior P(•) is coherent if it is a probability distribution over complete theories. (This requires P(φ) = 1 − P(¬φ), and so on.) 2. Computable approximability: A prior P(•) is computably approximable if there is an algorithm which computes an approximation of the probability which P assigns to φ that converges to P(φ) in the limit. \n The Occam property: A prior has the Occam property if there exists a length-based lower bound on the probability of any consistent sentence. \n Inductive: A prior is inductive if its probability for sentences of the form ∀n.ψ(n) goes to 1 as it conditions on more and more (going to all) confirmations of ψ(•). \n PA-weakness: A prior is PA-weak if it places non-zero probability mass on the set of complete extensions of PA. 6. Bounded regret: It may be desirable to show that a prior has regret (in terms of log loss or some similar measure) at most a constant worse than any other probability distribution over complete theories. 7. Practicality: The more tractable the algorithm which approximates a prior, the more practical that prior is. 8. Reflectivity: A prior P(•) is reflective if there is some symbol P in the logical language which can be interpreted as a representation of P(•), such that P(•) assigns accurate probabilities to statements about P . Coherence is an extremely desirable property; while approximations of a logical prior must be incoherent, it is prudent to demand coherence in the distribution being approximated. Reflectivity has been shown to be possible (up to infinitesimal error)  (Christiano et al. 2013 ) but difficult to do in a satisfactory manner  (Fallenstein 2014 ). Hutter's prior is coherent, inductive, and PA-weak. It has the Occam property so long as the probability distribution which it is generated from has the Occam property; but it is not computably approximable and it is far from practical. By contrast, Demski's prior is coherent and computably approximable, and has the Occam property if Φ does, but lacks most other desirable properties. Inductivity has been suggested in the literature, where it is better known as the Gaifman condition  (Gaifman 1964) . Roughly, the Gaifman condition requires that if P(•) is a logical prior and φ is a true Π 1 7 sentence of the form ∀n : ψ(n), then if P(•) is conditioned on the truth values of the first N instances of ψ(n), then as N goes to infinity, P(φ) tends to 1. In other words, the Gaifman condition requires that if a reasoner learns that ψ(n) is true for more and more n, then it eventually become arbitrarily confident that it is true for all n. However, computably approximable probability distributions which satisfy the Gaifman condition must assign probability 0 to some true Π 2 sentences  (Sawin and Demski 2013 ), 8 which seems strongly undesirable: computably approximable priors that satisfy the Gaifman condition are not sufficiently \"weak\"; they prevent reasoners from deducing true sentences.  Hahn (2013)  reports on an investigation of a more involved desirable property: If φ(•) is a generic predicate symbol (that is, the initial set of axioms makes no claims about φ(•)), and if the prior is conditioned on the statement that φ(n) is true for exactly 90% of the first 10 100 natural numbers, then the posterior probability 7. φ is a Π1 sentence if it can be written in the form ∀n : ψ(n) and there is a primitive recursive function which takes n as input and computes whether ψ(n) is true or false. 8. φ is a Π2 sentence if it can be written in the form ∀m. ∃n. ψ(m, n), where ψ(m, n) is primitive recursive. of φ(0) should be (approximately) 0.9. (This captures some of the intuition that a tractable algorithm will assign probability ≈ 0.1 that the (10 100 ) th digit of π is a 7, a desideratum that is difficult to formalize.) Unfortunately, even this property appears to be very difficult to obtain, and we are not aware of any proposed logical priors that have been shown to possess this property. In large part, generating logical priors is difficult because it's not yet clear what properties such a prior should possess, nor which properties are possible. Continued investigation into well-behaved logical priors is warranted, as the development of satisfactory computably approximable logical priors promises insight into problems of reasoning under logical uncertainty more generally. \n Beyond Logical Sentences A study of logical uncertainty with respect to sentences of first-order logic has proven insightful, but even if a practically approximable prior distribution over complete theories were defined, it would not provide a full theoretical understanding of reasoning under logical uncertainty in practice. Ultimately, logical sentences are not the right tool for reasoning about the behavior of objects in the real world. It is possible to construct a logical sentence which is true if and only if the Rube Goldberg machine deposits the ball in the top chute, but this sentence would be long and awkward. The manipulations that are easy to do to the sentence don't obviously correspond to realistic reasoning shortcuts about Rube Goldberg machines. Realistic reasoning under logical uncertainty will likely require hierarchical and contextfilled models of the problem. It is possible that these things could be built atop practical methods for reasoning according to a probability distribution over logical sentences, but the ability to reason with uncertainty about the truth-values of logical sentences will not solve these problems directly. Furthermore, while logical sentences are quite expressive, it is not clear that sentences of first order logic are the \"correct\" underlying structure for logically uncertain reasoning about the world. Practical logically uncertain reasoning inevitably requires reasoning about states of reality, and while most simple real-world questions can be translated into a sentence of first-order logic, it is by no means clear that uncertainty over logical sentences is the best foundations upon which to build practical reasoning. By analogy, consider a billiards player who lacks knowledge of physics. This reasoner would do well to learn classical mechanics, not because it behooves the player to start modeling the billiards table in terms of individual atoms, but because various insights from classical mechanics apply at the high level of billiards. But though the billiards player may use knowledge from classical mechanics in their high-level model of the world, it is not the case that the high-level model is \"merely\" a computational expedient standing in for the \"real\" atomic model of reality. The atomic model, too, is simply a model, and one which does not quite explain all the phenomena in the quantum world of the billiards player. We are like the billiards player: our state of knowledge is one where a study into uncertainty over logical sentences may provide significant insight that we can use to understand logical uncertainty as it pertains to \"high level\" objects, but this does not mean that practical logically uncertain reasoning will be done in terms of logical sentences, and nor does it mean that practical logically uncertain reasoning could always be reduced to uncertainty about logical sentences. It is merely the case that, given our present state of knowledge, a better understanding of logical uncertainty in the context of logical sentences is likely to provide insight into reasoning under logical uncertainty more generally. \n Discussion Smarter-than-human artificial systems must do most reasoning under both logical and environmental uncertainty. If high confidence in this reasoning is to be justified, even in a wide array of esoteric situations, then a theoretical understanding of logically uncertain reasoning is necessary: without it, it is difficult to ask the right questions  (Soares and Fallenstein 2014a) . The development of reliable methods for reasoning under logical uncertainty is work that must be done in advance of the development of smarter-than-human systems, if those systems are to be safe. While it may be possible to delegate significant AI research to early smarter-than-human systems, the creation of reliable methods for reasoning under logical uncertainty cannot be delegated, because logically uncertain reasoning is precisely what the delegatee must use in order to perform its research! How could a smarter-than-human system be trusted to accurately discover methods for reasoning under logical uncertainty, while using unreliable methods of reasoning under logical uncertainty? Furthermore, a better theoretical understanding of logical uncertainty is necessary in order to formalize many open problems related to the alignment of smarter-than-human systems. For example, consider the problem of constructing realistic world models: an agent faced with learning about the environment in which it is embedded must reason according to a distribution over environments that contain the agent. This problem must be fully described in order to check whether a practical program implements a solution, but describing the problem requires a better understanding of reasoning under logical uncertainty  (Soares 2015) . Or consider the problem of counterfactual reasoning: formalizing the decision problem faced by an agent embedded within its environment requires some way to formalize the problem of agents which may have an accurate description of their program, but uncertainty about which action it will take. Specifying this problem, too, requires a better theory of logical uncertainty. In fact, satisfactory decision theory additionally requires an understanding of logical counterfactuals, the ability to reason about what \"would\" happen if a deterministic program did something that it doesn't. It is likely that a better understanding of logical uncertainty will yield insight in this domain  (Soares and Fallenstein 2014b) . Many existing tools for studying reasoning, such as game theory, standard probability theory, and Bayesian networks, all assume that reasoners are logically omniscient. If these tools are to be extended and improved, a better understanding of logical uncertainty is required. In short, a developed theory of logical uncertainty would go a long way towards putting theoretical foundations under the study of smarter-than-human systems. We are of the opinion that those theoretical foundations are essential to the process of aligning smarter-thanhuman systems with the interests of humanity. \t\t\t . And which oracle machines halt, and which metaoracle machines halt, and so on. \n\t\t\t . There are uncountably many complete theories of first order logic, but a probability distribution over them can be defined using the machinery of measure theory.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/QuestionsLogicalUncertainty.tei.xml", "id": "f86d5fcf31063556840a1253cb09227a"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "61   A forthcoming CSET brief on the latent potential of community colleges will explore this in greater detail. 62  This working group would then be part of the federal efforts that the National Artificial Intelligence Initiative Office for Education and Training coordinates. 63 For example, options to leverage existing funding streams include WIOA, Perkins Act, H1B revenue, and other related NSF programs. For more on how WIOA funds are allocated, see: (1) \"Federal Sources of Workforce Funding,\" Urban Institute, accessed July 26, 2021, https://workforce.urban.org/strategy/federal-sources-workforce-funding; (2) Donna Counts, \"Federal Funding for State Employment and Training Programs Covered by the WIOA,\" CSG Knowledge Center, accessed July 26, 2021, http://knowledgecenter.csg.org/kc/content/federal-funding-state-employmentand-training-programs-covered-wioa; and (3) U.S. Employment and Training Administration, \"Current Grant Funding Opportunities,\" U.S. Department of Labor, accessed July 26, 2021, https://www.dol.gov/agencies/eta/grants/apply/find-opportunities. 64 For example, tuition inflation, rising student debt burdens, a 40 percent underemployment rate for recent college graduates, and rapidly increasing master's degree enrollments. 65 What credential employers demand is not necessarily what is actually needed to perform the associated tasks. Bachelor's degrees are a market signal to Center for Security and Emerging Technology | 66 employers; instead of proposing degree creep or degree inflation we propose training more talent up to the required level needed to be proficient. This requires making those options viable and credible signals to employers.", "authors": ["Diana Gehlhaus", "Luke Koslosky", "Kayla Goode", "Claire Perkins"], "title": "U.S. AI Workforce: Policy Recommendations", "text": "Executive Summary The U.S. artificial intelligence workforce, which stood at 14 million people in 2019, or 9 percent of total U.S. employment, has grown rapidly in recent years. This trend is likely to continue, as AI occupational employment over the next decade is projected to grow twice as fast as employment in all occupations. Such an important and increasing component of the U.S. workforce demands dedicated education and workforce policy. Yet one does not exist. To date, U.S. policy has been a piecemeal approach based on inconsistent definitions of the AI workforce. For some, current policy is focused on top-tier doctorates and immigration reform. For others, the conversation quickly reverts to STEM education. This report addresses the need for a clearly defined AI education and workforce policy by providing recommendations designed to grow, sustain, and diversify the domestic AI workforce. We use a comprehensive definition of the AI workforce--technical and nontechnical occupations--and provide data-driven policy goals. Our policy goals and recommendations build off of previous CSET research along with new research findings presented here. Previous research in this series defined the AI workforce, described and characterized these workers, and assessed the relevant labor market dynamics. For example, we found that the demand for computer and information research scientists appears to be higher than the current supply, while for software developers and data scientists, evidence of a supply-demand gap is mixed. To understand the current state of U.S. AI education for this report, we manually compiled an \"AI Education Catalog\" of curriculum offerings, summer camps, after-school programs, contests and challenges, scholarships, and related federal initiatives. To assess the current landscape of employer demand and hiring experiences, we also interviewed select companies engaged in AI activities. Our research implies that U.S. AI education and workforce policy should have three goals:  (1)  increase the supply of domestic AI doctorates, (2) sustain and diversify technical talent pipelines, and (3) facilitate general AI literacy through K-12 AI education. To achieve these goals, we propose a set of recommendations designed to leverage federal resources within the realities of the U.S. education and training system. Our first recommendation sets the foundation for facilitating these goals by creating a federal coordination function. We believe such a function is critical given ongoing fragmented AI education initiatives, and would harness the potential of the newly established National Artificial Intelligence Initiative Office for Education and Training within the White House Office of Science and Technology Policy. We recommend this office coordinate federal and state initiatives, convene key stakeholders to share lessons learned and best practices of statelevel AI education initiatives, and compile and publish information on AI education and careers on a publicly available \"AI dashboard.\" The remaining recommendations advocate for a multipronged approach to implement policies across goals, including: Importantly, our recommendations prioritize creating multiple viable pathways into AI jobs to diversify the AI workforce and leverage all U.S. talent. Our research shows the dominant pathway to enter the AI workforce remains having a four-year college degree. However, this may be restricting the amount of talent entering the AI workforce, unnecessarily limiting opportunity for those who are otherwise qualified and able. Our recommendations therefore prioritize harnessing the potential of community and technical colleges, minority-serving institutions, and historically Black colleges and universities in training tomorrow's U.S. AI workforce. In addition, to promote alternative pathways into AI jobs, we propose that the National Institute of Standards and Technology work with industry to establish industry-accepted standards for AI and AI-related certifications to enhance their legitimacy. And as a top employer of technical talent, the federal government could modify its hiring criteria to lead by example. We hope that this report and recommendations advance the discourse on AI education and workforce policy. Now is a critical time to invest in training and equipping a globally competitive AI workforce for tomorrow. With concerted and targeted efforts, it is possible to lead the world in AI talent. Ultimately, an AI workforce policy inclusive of all of our report's elements is more likely to be the most effective. However, we also present our recommendations as a road map to guide U.S. policymakers in crafting an AI education and workforce agenda.  \n Table of Contents Executive \n Introduction The acceleration of artificial intelligence adoption has profoundly changed how businesses, individuals, and governments conduct their day-to-day activities. As this trend continues, having an education and workforce policy specific to AI is essential to future U.S. competitiveness. In effect, AI education and workforce policy is now a national security priority. Such a workforce policy includes both adequately training the U.S. workforce for AI jobs and preparing the country's youth and adults for the changing nature of jobs resulting from AI and other emerging technologies. This report pulls together previous CSET research on the U.S. AI workforce to offer data-driven policy goals and recommendations designed to grow, sustain, and diversify the domestic AI workforce. This is the third and final paper in a three-part series on the U.S. AI workforce. The first defined, described, and characterized the supply of U.S. AI workers while the second evaluated the associated labor market dynamics. Although not in the three-part series, we also published several papers as part of our body of research on the U.S. AI workforce. This includes published papers on the prevalence of AI and AIrelated certifications, a comparative assessment of AI education in the United States and China, an overview of U.S. AI summer camps, the U.S. Department of Defense's (DOD) AI workforce, and a forthcoming examination of the latent potential of community colleges in training tomorrow's AI workforce. \n Current U.S. AI Education and Workforce Policy To date, most discussions on broader AI education and workforce policy have focused on STEM or on top-tier technical talent. For example, both the Global AI Talent Report  1  and Stanford University's annual Artificial Intelligence Index Report 2 break down the status of the AI workforce through a detailed analysis of PhDlevel computer research scientists. In its Final Report, the National Security Commission on Artificial Intelligence similarly emphasizes technical talent, before focusing its broader U.S. AI workforce discussion on K-12 STEM education.  3  Equating STEM policy to AI education and workforce policy is also apparent in the executive branch. The education and training arm of the new National Artificial Intelligence Initiative, housed in the White House Office of Science and Technology Policy (OSTP), is tasked with focusing on AI, but in practice, emphasizes STEM.  4  Similarly, other recent White House science and technology (S&T) education initiatives focus on STEM over AI or other emerging technologies. For example, in 2018 the White House issued a STEM education policy that included a five-year STEM education strategic plan. 5 AI education and workforce policy is, by our definition, more expansive than top-tier PhD-level computer science talent and STEM education. It includes policies for technical and nontechnical talent involved in safe AI design, development, and deployment.  6  The recommendations in this report are designed with that in mind. In this regard, AI education and workforce policy is more aligned with broader S&T education and workforce policy. Although much of the discussion surrounding STEM focuses on K-12 and four-year colleges, a large component of S&T education and workforce policy is focused on postsecondary education and training outside of the traditional four-year degrees, including associate's degrees, certificates, microcredentials, certifications, and apprenticeships. These pathways are vital for youth that do not attend four-year college, as well as for adult learners upskilling for S&T jobs. This means greater alignment with AI education and workforce policy; however, these alternative pathways generally receive little attention. \n A Unique Need for AI Education and Workforce Policy Five factors make the AI workforce distinct from the STEM and S&T workforce, and from the workforce of other emerging technologies. Such uniqueness creates a need for a dedicated AI education and training policy to adequately grow and cultivate a globally competitive AI workforce. Effective AI education and workforce policy must therefore understand the long road already traveled on S&T--particularly STEM--education and workforce policy. This includes the maze of federal programs, congressional proposals, and body of related research. Effective AI education and workforce policy must acknowledge the reasons why S&T workforce policy continues to be a challenging area for meaningful progress and change. \n Current S&T Education and Workforce Policy Challenges As a large part of AI education and workforce policy, it is important to understand why S&T--particularly STEM--education and workforce policy continues to be a major challenge even after decades of national prioritization. Since many of these challenges will also affect AI education and training policy, they must be considered within this context. There are three main challenges affecting U.S. S&T education and workforce policy: (1) the decentralized and fragmented structure and design of the U.S. education system; (2) lack of coordination and evaluation of federal and state-level initiatives for success, information sharing, and scaling, including public-private partnerships; (3) disagreement on the nature and extent of S&T workforce supply gaps and the need for policy interventions. First, the U.S. education system is a patchwork of public and private schools largely administered by individual states. The federal government provides relatively minor education oversight through the compilation and reporting of education statistics, along with promoting equitable access to education and enforcing a prohibition on institutional discrimination. However, similar to S&T education programs, these programs also suffer from a lack of coordination.  13  For example, the Workforce Innovation and Opportunity Act (WIOA), the primary law governing federal workforce development programs, spans 19 programs administered by three federal agencies. Each has its own data systems and reporting requirements. 14 There is also complexity in administration, as programs often require collaboration between state agencies, local industry associations, and employers, in addition to braided funding across public and private sources that creates yet more coordination challenges. Third, there are marked differences in opinions about the appropriate policy goal and federal role in growing and cultivating a S&T workforce. For example, a 2017 Congressional Research Service report noted the following viewpoints: \"There is a shortage;\" \"There is not a shortage;\" \"More scientists and engineers are needed regardless of the existence of a shortage;\" \"There may be shortages [but only] in certain industries, occupations, or fields.\"  \n AI Education and Workforce Policy Coordination A key lesson learned from the experience of S&T education and workforce policy is the large potential value of coordination. Facilitating equity in access to AI educational resources will be critical, and the S&T experience at the federal and state level shows this does not happen without concerted effort. S&T education and workforce policy highlights how important coordination is to maximize the impact and reach of federal and state programs. Coordination has a clear impact on how effective policy initiatives are; previously, no one federal entity had ownership over the entire federal STEM enterprise: \". . . too little attention has been paid to replication and scale-up to disseminate proven programs widely, and too little capacity at key agencies has been devoted to strategy and coordination.\"  18  Because of this, Congress elevated federal coordination of STEM initiatives to the executive branch.   25  To enhance our assessment for this paper, we supplemented this work with the creation of an AI Education Catalog and interviews with employers of AI talent. \n Size, Supply, and Labor Market Dynamics The U.S. AI workforce is sizable. In 2019, it consisted of 14 million workers, or about 9 percent of total U.S. employed. It is also growing rapidly relative to the total workforce. Over 2015-2019, the U.S. AI workforce grew 21 percent compared to 6 percent for total U.S. employed. Demand for AI occupations will likely be strong over the next decade, projected to grow twice as fast as for all U.S. occupations.  26  The U.S. AI workforce is also fairly concentrated in major metropolitan hubs, especially for technical talent. In absolute terms, large concentrations of workers were found in major hubs on the West Coast (see Figure  2 ). As a share of county employment, we also found large concentrations of technical workers in Seattle and San Francisco, as well as the Washington, D.C. metropolitan area. Product Team workers are more geographically distributed, although still concentrated in urban areas. Importantly, our analysis shows the technical AI workforce has a lack of racial and gender diversity. For Technical Team occupations, most workers are male, with few identifying as Black or Hispanic.  27  Figure  2  summarizes some of these key findings. Note: People who identify as Hispanic may be of any race. Our previous research also shows that the dominant pathway to enter the AI workforce is a four-year college degree. However, this may be unnecessarily limiting, potentially leaving qualified and able AI talent behind. While the majority of AI workers have at least a bachelor's degree, a sizable share of the AI workforce does not: we found 44 percent of those employed in Product Team jobs and one-third of those employed in Commercial Team jobs have less than a bachelor's degree.  28  The challenge is that many employers do not yet view alternative pathways such as certifications and other sub-baccalaureate credentials as viable market signals of qualification.  29  In terms of four-year degrees, we do find evidence of talent supply catching up to demand through a shift toward AI-related majors. Computer science and engineering were the two fastest growing fields of study over 2015-2018. Anecdotally, however, the surge in demand for computer science courses at many U.S. universities appears in some cases to be greater than their ability to supply them.  30  While we do not make claims of a blanket AI talent shortage, our assessment of AI labor market dynamics found that there are likely current gaps in supply relative to demand in the AI workforce which vary by occupation. For example, for computer and information research scientists there appears to be more demand than supply, while for software developers and data scientists evidence of supply-demand gaps is mixed. In contrast, we found no evidence of supply-demand gaps for user experience designers and project management specialists. 31 \n AI Education in the United States In addition to labor market data, crafting effective AI workforce policies also requires knowing how and where to target AI education and training investments to prepare youth for AI careers. To do this, we need to understand:  (1)  what states are doing in AI education;  (2)  what types of AI curricula, programs and scholarships are available; and (3) where the accessibility and opportunity gaps are. This is not a simple task given the fragmented design of the U.S. education system, but it is essential for good policy. To understand the current U.S. AI education landscape, we created an AI Education Catalog to see what AI-related efforts are underway across the country. We collected data on curricula, afterschool programs, summer camps, challenges, contests, conferences, federal initiatives, and scholarships. 32 A large share of AI educational programming currently available is online, and comes from a mix of both private and nonprofit organizations. Investors in these programs and course offerings ranged from school districts and local universities to big tech companies like Amazon, Google, and Microsoft. 33 Importantly, the cost of these offerings varies widely, ranging from being free to requesting a quote (e.g., the cost varies on factors like the number of students in the classroom or the size of the school district). Summaries for each category, including descriptive statistics, are provided in Appendix B. By design, the catalog also provides insights on equity in access gaps. Overall, we find gaps in access and opportunity are more acute in rural and low-income school districts that lack access to AI education resources and opportunities. The catalog also shows that while there are free and accessible programs for students to learn about computer science and AI, the majority are online-requiring the student to have high-speed broadband and access to a computer. Previous CSET research found AI education in the United States is being provided in a piecemeal way that varies by state and places a heavier emphasis on computer science education. The decentralized structure of the U.S. education system creates both opportunities and challenges; schools have the flexibility to integrate innovative AI curricula designs and experiment with new approaches in pedagogy, but this could also exacerbate existing educational disparities. Many school districts lack access to adequate resources to support AI education or its evaluation, and there is no consensus or consistency in AI curricula. A detailed discussion of current efforts to integrate AI education at the K-12 and postsecondary levels is provided in the CSET report \"AI Education in China and the United States: A Comparative Assessment.\" 34 \n Perspectives from AI Employers In addition to these quantitative assessments, we also wanted to hear about labor market dynamics directly from AI employers. We conducted 10 interviews with companies engaged in AI activities ranging from AI research, software development, hardware design and production, and provision of AI-enabled service or capabilities. We include several of these insights in our proposed recommendations. While not a generalizable sample, we believe these discussions provide key insights that further validated our quantitative assessment. For example, our discussions provided insights on current AI labor market dynamics that could not be gleaned from federal survey data. This includes hiring conditions for technical talent: top-tier PhD-level talent was much harder to recruit than other technical talent. Moreover, while employers did not report difficulty in recruiting software developers and data scientists, they did report difficulty in finding \"exceptional\" software developers.  35  Moreover, consistent with our research on AI job postings data, all employers emphasized the importance of four-year college degrees as a minimum requirement for AI and AI-related positions. When asked about alternative credentials such as certifications, many acknowledged their relevance but none reported using them to hire AI talent. Several also noted that there were no clear standards for which certifications were at a quality equivalent to their hiring needs. Finally, while a few employers noted providing internal training and upskilling incentives for AI and AI-related skills, the majority did not. \n AI Workforce Policy Goals and Frameworks \n AI Workforce Policy Goals Although we use four categories to define the AI workforce, our research found the need for three targeted AI education and workforce policy goals: Goal 1: Grow, cultivate, and attract PhD-level technical AI talent while also improving diversity in the field. Goal 2: Ensure a diverse and sustainable supply of non-doctorate technical talent. Goal 3: Promote and provide AI literacy education for everyone. 36 \n Education and Workforce Policy Frameworks To capture each part of the career and education pipeline, we believe AI education workforce policy should be considered through two frameworks: (1) Short-term (employer) This framework covers the talent management life cycle, from \"hiring to retiring,\" as a way to depict the AI workforce challenges in the current recruiting and retention environment. The focus of this framework as shown in Figure  3  is on leveraging AI expertise in the current workforce. This includes good talent management in the form of advancement, continuous education and upskilling, promotion, and other professional development opportunities, as well as discouraging AI talent from leaving the field. Each phase represents a possible intervention point--mostly for employers as they assess their AI workforce needs. Center for Security and Emerging Technology | 20  Many stakeholders are involved in career and educational decision-making and workforce development. As such, it is important both frameworks also consider the entire stakeholder ecosystem. This means identifying all of the players that have roles and responsibilities in implementing AI education and workforce policy. Figure  5  shows each tier of stakeholder in a series of concentric circles, to demonstrate the many levels of influence from ego (self) to federal authorities and institutions. Each tier interacts with and on each other--as part of the larger society--to influence individuals' career and educational decisions.   \n AI Workforce Policy Recommendations We consider our policy recommendations through the lens of our frameworks as they relate to our defined goals for AI workforce policy. That is, to grow, cultivate, and sustain a globally competitive cadre of technical and nontechnical AI talent. We use the evidence presented above to design policies that we believe are most likely to be effective in achieving the stated goals; however, we also strongly advocate where possible for rigorous program evaluation and iteration in design and implementation as appropriate. Our recommendations accept the legacy and challenges of previous efforts to design and effectively implement U.S. STEM education and training policies, as a part of AI education policy. We work within the current system--noting there are advantages and disadvantages to its structure--to provide actionable ideas that we believe can advance the goal of growing and cultivating U.S. AI talent. We also focus heavily on the role of the federal government, although much of this relies on coordination and cooperation with states, industry, academia, and nonprofit stakeholders. It also relies on buy-in from teachers, adult learners, and students, to integrate AI in the classroom and for youth to aspire to AI careers. We cannot generally mandate what nonfederal stakeholders do. However, we can propose a clear role for the federal government, and ways the government can best leverage its resources to engage stakeholders to promote AI workforce goals. That means making resources related to AI education easily and readily available, meeting stakeholders where they are, and providing avenues to coordinate, cultivate, create, collect, convene, and disseminate such information.  38  In addition, as a top employer of AI talent, the federal government can lead by example for creating alternative pathways into AI jobs. Finally, in encouraging youth, parents, and educators to embrace and aspire to AI careers, we propose recommendations specific to leveraging federal resources to create AI career and educational information. We designed these recommendations with EAST principles in mind--make it Easy, Attractive, Social, and Timely.  39  We present our recommendations in groupings by the nature of the proposed action. We provide a summary of the rationale for each type of action, followed by the specific recommendation(s). For all recommendations except the first, we also provide a parenthetical tag in the following structure:  \n Empowering the National Artificial Intelligence Initiative Office for Education and Training The most successful AI workforce policy will include a coordinated effort between federal and state governments. Convene: Through these networks, provide states with support to design and integrate AI/computer science curricula into core subjects, and general AI literacy curricula, by hosting conferences, professional development sessions, and workshops that share best practices. Issue an annual report recognizing K-12 school districts and community and technical colleges that have made a major advancement in AI education as exemplars. \n Creating and Disseminating AI Educational and Career Information A critical part of the coordination function of the National Artificial Intelligence Initiative Office for Education and Training includes being a one-stop resource for AI career and educational information. Research shows that youth start planning for college and careers early, continuing and iterating throughout high school and complete the process at different times in their lives. For example, one study found \"three main considerations have a sizeable impact and act together on youths' career and educational decisions: (1) the use of and reliance on influencers, (2) engagement in information-seeking activities during high school, and (3) social-emotional well-being, degree of focus, and selfmotivation.\"  46  While federal policy cannot directly engage all the above factors, it can make encouraging AI education and careers easy, attractive, social, and timely (EAST). We propose leveraging federal resources to create AI career and educational resources to raise awareness of AI career pathways. This is a critical step to ensuring youth, parents, educators, and counselors have access to the right resources at the right time as youth approach their educational and career decision-making processes. We propose this happen in two steps: (1) oversee the creation of better AI education and workforce data 47 and AI education and career materials, funded by Congress, and (2) oversee the dissemination of these materials.   \n Rec \n Facilitating Alternative Pathways into AI Jobs As noted earlier, previous CSET research shows that a four-year college degree remains the dominant pathway into AI jobs. Moreover, employers are largely not using the vast array of AI and AI-related certifications available as a way to qualify applicants for AI jobs. 60 However, this limits access and opportunity. Our research found that this approach leaves talent behind that could thrive in AI jobs and reduce the potential for bias in AI through greater diversity. To  \n Integrating K-12 AI Curricula and Course Design The U.S. K-12 education system has important advantages along with limitations when it comes to integrating new curricula, as noted in the STEM challenges section and in other CSET research.  74  Perhaps the biggest advantage is the ability for school districts to experiment with new curricula and pedagogical approaches, and establish partnerships to address the specific needs of that school or community. However, serious challenges also exist. Initiatives and partnerships will have limited effectiveness and reach if they are not consistently funded, tracked, evaluated, and scaled.  75  For example, even after more than a decade of concerted efforts by states, industry, and advocacy groups, just half of all U.S. high schools offer computer science education.  76  In the case of AI education, there is the added challenge that curricula are neither consistently defined nor agreed upon in importance. Still, to adequately train and equip tomorrow's AI workforce, school systems must start somewhere. Our AI Education Catalog shows there is a large range of online AI education programs and resources already available for teachers and students (see Appendix B). Moreover, our research found an impressive range of innovative pilot initiatives currently underway across the country to bring AI in the classroom. We provide two such examples in the text boxes below, based on interviews with project leads and supplemented by publicly available information. \n Example 1: Gwinnett County Public Schools (Georgia) Gwinnett County Public Schools (GCPS) is instituting a comprehensive K-12 AI-centric education curriculum spanning three elementary schools, one middle school, and one high school. Each grade from Pre-K to high school has a dedicated AI curriculum that will be integrated into core subjects. Unlike other AI education programs in the United States that depend on special academic placement, GCPS is creating these schools as part of its effort to meet a growing population, and students living in the district will be automatically assigned to these schools. The program aims to increase AI readiness for every student, regardless of background or aptitude, through six components: coding, applications, data science, computational thinking, problemsolving, and ethics. Each component of \"AI Readiness\" strives to align learning goals with the relevant skills and knowledge graduates need for postsecondary education or the workforce.    \n AI Workforce Policy Recommendations: Summary Each recommendation aside from the first recommendation is tagged with the department and agencies (\"agencies\") involved, which framework this falls under (short-term or medium-/longterm), and which AI workforce policy goal it addresses (1, 2, 3).  \n Empowering the National \n Creating and Disseminating AI Educational and Career Information Rec: Congress should appropriate funding for the BLS, U.S. Census Bureau, and NCSES to design, collect, compile, and publish occupation or skills-based data on the U.S. AI workforce along with education statistics from NCES.  Rec: Congress should fund targeted undergraduate scholarships and research fellowships that cover tuition, cost of living, and incentive pay for AI and AI-related expertise. \n Investing in PreK-12 AI Education and Experiences Rec: Congress should appropriate funding for federal grants to states for public K-12 schools to equip facilities with \"AI labs\" that enable hands-on learning along with virtual learning platforms for anytime remote or individualized learning.  Rec: Congress should authorize FCC funding to secure access to high-speed internet and high-speed internet devices, with eligibility for any K-12 youth participating in the free lunch program. (Agencies: FCC and USDA; Framework: Long-term; Policy goals: 1+2+3) \n Integrating K-12 AI Curriculum and Course Design Rec: Congress should appropriate funding for federal grants to rural and low-income school districts to integrate K-12 AI education designed using promising practices and proven education models. (Agencies: USDA, DOE, NSF, and Ed; Framework: Long-term; Policy goals: 1+2+3) \n Cultivating and Supporting K-12 AI Educators Rec: Congress should fund and task Ed to create a national repository of peer-reviewed AI teaching materials, including offthe-shelf AI-enabled products, experiment kits, and in-class labs. (Agencies: Ed and NSF; Framework: Medium-term; Policy goals: 1+2+3) \n Conclusion The technical and nontechnical workers that comprise the U.S. AI workforce have experienced rapid growth in recent years. Strong demand for workers with AI and AI-adjacent expertise will likely continue, as AI-enabled applications rapidly proliferate across the economy. Such an important and rapidly growing component of the U.S. workforce demands dedicated education and workforce policy. Yet one does not exist. Current AI education and workforce policy either focuses narrowly on top-tier PhD-level talent in computer science and engineering, or broadly on STEM education. This paper addresses the need for clearly defined and targeted policies to grow and cultivate the domestic AI workforce. Our approach to AI education and workforce policy is data-driven, clearly defining this segment of the workforce and analyzing the demand for and supply of AI talent across the United States. We also manually compiled an AI Education Catalog to understand the landscape of AI educational programming in the United States, and spoke with employers engaged in AI activities about their experience recruiting and retaining AI talent. Using this information, we argue AI education and workforce policy should consist of three goals. First, to grow, cultivate, and attract PhD-level technical AI talent. Second, to ensure a diverse and sustainable supply of non-doctorate technical talent. And third, to promote and provide AI literacy education for everyone. We designed our policy recommendations with these goals in mind. Our recommendations also appreciate the reality and challenges inherent in the design of the U.S. education system, the authorities of the federal government, and the long and persistent challenges of U.S. STEM education and workforce policy. Although AI education and workforce policy is bigger than STEM, we acknowledge the overlap and must consider the challenges accordingly. Given the complexities of federal education and workforce policy, our first recommendation calls for federal coordination through leveraging the new National Artificial Intelligence Initiative Office for Education and Training within OSTP. Our remaining recommendations emphasize investment in schools, students, and teachers; in better AI workforce data and AI education and careers information; and in the research that will better equip classrooms across all subjects to teach AI literacy. We also prioritize creating and cultivating multiple pathways into the AI workforce that includes leveraging the potential of community colleges, MSIs, and HBCUs, along with establishing national industry-accepted standards for AI certifications. We believe the most effective AI workforce policy will include all of the various elements outlined in this report. However, our recommendations can also be considered a road map for policymakers interested in understanding each segment of the AI workforce. This is the third paper in a three-part series on the U.S. AI workforce. We hope this report and recommendations advance the discourse on AI education and workforce policy. The United States is at a critical moment to invest in training and equipping a globally competitive AI workforce, and with dedicated effort it is possible for the country to lead in AI talent. \n Authors Diana Gehlhaus is a research fellow at CSET, where Luke Koslosky and Kayla Goode are research analysts. Claire Perkins is a former CSET student research analyst. the exception of the 4th level (blue) on the map used to represent nonprofit initiatives). • Programs and initiatives must be statewide and not confined to a specific region of the state • Program must be standalone and not just a tab on the respective state's department of education website highlighting STEM education For the state search, key terms were used such as \"(insert state here) STEM education\", \"(insert state here) STEM Education initiatives\", and \"(insert state here) STEM programs.\" Programs and initiatives were evaluated using the criteria previously laid out. Once the STEM offices/programs and initiatives were evaluated, they were sorted into four groups as shown in the map. If a state was found to have both an office and an initiative, then the one with the highest-ranking program score (from 1-4 as outlined below) was used for the map classification. If a state also had a non-state-affiliated program, it was only colored blue if there was not a state-affiliated program for the state. This was done because there are a lot of similarities between initiatives and the STEM offices/programs creating a lot of crossover between the two. The criteria for each group are laid out below. 1. Level 1 (Yellow): The first level is most commonly used to classify programs that function as a working group or advisory council. These are the programs that are placing an emphasis on existing efforts rather than establishing new ones. In addition, they have very few if any sponsored STEM programs, activities, extracurriculars, and events. \n Level 2 (Orange): The second level is for programs that are more active than the first level and have sponsored STEM programs, activities, extracurriculars, and events. Their programs might not be the most active in all of the schools and might not reach every student but nevertheless are still being impactful. \n Level 3 (Purple): The third level is for programs that have proven metrics for their success. This could be through an increase in performance on standardized testing or increased rankings by a national party. Their sponsored STEM programs, activities, extracurriculars, and events are successful and are impacting the greatest number of students. 4. Level 4 (Blue): The fourth level is for programs that operate on the state level but do not have state involvement. This includes programs that are nonprofits that have a focus on K-12 STEM education in that particular state. \n Limitations When finding STEM programs the search was designed to be as broad as possible in an effort to try and identify the state programs given that it would be presumed they are one of the larger STEM programs in that given state. Even with a criteria for categorizing the programs listed, there was still a degree of subjectivity used in the process. Because there were limited strict boundaries used for the groups, some could potentially fit into multiple groups and a judgment call was made to place the program into one group over the other. as a barrier to access. The programs were generally inclusive of all K-12 educators rather than only offering a curriculum that is grade specific. One of the more unique examples of curricula for educators is the NVIDIA teaching kits, which provide free resources and materials necessary for educators to teach from AI curricula. In addition, some of the nonprofit curriculum programs have a mentorship program associated with them to help educators learn the material like Microsoft TEALS and Exploring Computer Science. Other programs just include online lesson plans or modules that educators can use. The MIT Media Lab offers a free example of the online curriculum, while Project Lead the Way provides a curriculum that requires educators to request a quote in order to access information on the price of the program. The AI Education Project designed a program that can teach AI literacy to middle and high school students even if they do not have access to computer science courses. \n After-School Programs The after-school programs found take place predominantly online and are run by private organizations. Their activity type ranges from curricula that children are able to engage with on their own time, to live after-school classes, and seasonal camps. After-school programs are predominantly operating online and accessible nationwide as a result. These programs serve as a curricular supplement rather than acting as an additional club or organization. Many schools also have their own sponsored coding clubs that are unique to the school. However, those were not documented for this catalog as that information is generally not publicly available. Most after-school programs are run by private organizations and are available for K-12 students. Their duration can be anywhere from one week of lessons, to year-round content. Funding and investors were difficult to find for these programs given that they are privately owned and self-sufficient. Their cost varies by the program as many of the at-your-own-pace modules are free, the live lessons range anywhere from $50-$450 a month, and the fiveday courses are around $450. \n Key Numbers • 91 programs identified or a graduated high school senior. However, the main target audience for the camps are high school students. We found a mix of virtual, day, and overnight camps, with the majority of them being hosted on a college campus (whether it is run by the university itself or a private organization). The majority of the camps are run by private organizations with iD Tech being the largest but there are also camps that are run by college programs and nonprofit organizations as well. The cost of the camps range from being free for every student to $7,375 with the bulk of the camps costing around $500-$1,000. \n Key Numbers • 447 AI and AI-adjacent summer camps identified in the United States • 47 programs are free for students • 148 programs are overseen by iD Tech Location and cost are the largest barriers to a child's access to AI summer camps. The camps are heavily concentrated in three states: California, New York, and Texas. This uneven distribution is further demonstrated by the fact that eight states alone host more camps than the other 42 states. In addition, most camps take place on a university campus, but the overwhelming majority of them are run by private organizations. Even though cost is often dependent on factors such as duration, host organization type, and whether or not it is overnight, most camps still have a price tag greater than $750, and this can be a deterrent for some families. \n Conferences/Challenges While there were less than 20 AI conferences identified, most of them were under one week in length, open to all educators and some to the public as well and took place in a virtual format. The conferences were hosted by a wide variety of organizations including universities, computer science associations, AI4All, government organizations like the DOD and the Defense Advanced Research Projects Agency (DARPA), and big tech companies like Google and Microsoft. The conferences are more geared towards educators to help provide them with the tools to implement AI and computer science in their classrooms. The challenges portion of our catalog was segmented to include challenges, competitions, hackathons, and robotics competitions. Challenges and competitions were programs that included computer science pitch competitions, cybersecurity competitions, computer science test competitions, and collaborative challenges. Because the focus of the catalog is for K-12 students rather than undergraduate students, most of the hackathons identified for the catalog were catered towards high school students with very few aimed at middle and elementary school students. The robotics competitions were not as explicit in citing age or grade eligibility as they only put \"students\" down for eligible participants but many were for middle and high school students. The challenges and competitions were more broad in eligibility with many being open to anyone but otherwise catered towards both high school and university students. Most of these challenges (challenges, competitions, hackathons, and robotics) were sponsored by a wide array of large companies, college universities, and government organizations like the DOD, NASA, DARPA, DOE, etc. Many of the challenges had more than one sponsor with many having more than 10 representing a large array of companies. California was found to have the most hackathons but due to the COVID-19 pandemic, most of the hackathons and other challenges were operating in a virtual format. Some competitions were as short as one day to as long as a yearlong competition process. \n Key Numbers • 100 programs identified • 14 programs targeted towards underrepresented groups • Price point not really as relevant since competitions free to access • 30 challenges identified • 14 conferences identified • 50 hackathons identified (although there are many more not identified in the catalog) • 8 robotics competitions identified The competitions (challenges, hackathons, and robotics) are all accessible in the fact that there does not appear to be a financial barrier in terms of entry. However, participation in many of these competitions requires school support. Whether school support be a faculty advisor, funds for building supplies, or access to curricula to learn the skills necessary to compete, schools with a lack of funding and resources could be at a severe disadvantage. Hackathons are an interesting program in the fact that many of the ones listed such as AIHacks, a student ran hackathon for female and gender nonconforming students, has a mentorship component embedded into it. Hackathons are the largest competition-based programs in the catalog with at least 50 documented. There were more hackathons that were not reported in the catalog because they either changed each year in terms of the topic or were more local competitions. Students are not required to have a background in coding to compete in one of these competitions and there is a lot of learning that takes place at these events. BEST Robotics is also an example of a robotics program that reaches 18,000 middle and high school students each year that is free to enter and has a mentorship aspect built in with the volunteers for the program. All of the AI challenges (including robotics and hackathons) incorporate AI and AI-adjacent skills but tend to be relatively unique. The University of Alabama has a cybersecurity capture the flag competition that operates in a similar fashion to Jeopardy!, the trivia quiz game show. For the conferences, they are more aimed towards educators and the greater community but a large one that takes place each year is the AI4ED conference. \n Federal Initiatives Federal initiatives include a range of programs that different federal departments and agencies have set up in order to support, recruit, or work with AI and AI-adjacent talent. These initiatives we found were primarily targeted towards undergraduate and graduate students, although a few were focused on K-12 outreach. The different program types include apprenticeships, challenges, fellowships, and internships. There was no dominant government agency hosting these programs, with representation from most intelligence, defense, and research-oriented agencies. The programs were predominantly targeted towards U.S. citizens and permanent residents. The duration of these programs vary given that this section includes a wide array of different programs with durations ranging from one summer to several years. \n Key Numbers • 70 programs identified Internships and fellowships made up the majority of the federal initiatives in effort to boost their workforces. The fellowships mostly covered the cost of students' undergraduate or graduate education in addition to providing them with a stipend in exchange for their participation for a set period of time to work for the respective agency. Because of the position tied to the internship or fellowship, almost all of the programs required citizenship or permanent residency status in addition to the student's ability to pass a federal background check of varying levels of clearances. The Graduate Research Fellowship program through the NSF is an example of one of the fully funded programs that is aimed towards funding graduate students' education with a set stipend and education allowance each year. \n Scholarships Given the breadth of scholarship offerings in the United States, it is likely that our catalog is not as comprehensive as the other sections presented in this report. Moreover, we found the purposes of each scholarship sponsoring organizations differs. The selection of scholarship programs was more sparse and difficult to find niche programs. Many of the programs identified were more general STEM scholarships that had the space for people with AI and AI-adjacent skills to apply.  32 The catalog will be fully available and searchable as an online interactive tool, to be released alongside this report and created in partnership with the AI Education Project. 33 Noting information is limited for many programs that are behind a paywall.  34  Peterson, Goode, and Gehlhaus, \"AI Education in China and the United \n States.\" 35 This gets into quality differences within occupations, a separate topic not addressed in this report. 36 Having some AI literacy will facilitate career awareness and preparedness for responsible use of AI, especially for those in nontechnical AI careers.  37  For more on youth career and educational decision making, see Diana Gehlhaus, \"Youth Information Networks and Propensity to Serve in the Military\" (Pardee RAND Graduate School Dissertation, May 2020, published September 2021). Behavioural Insights Team, 2015. grants to improve teacher education in certain specific areas (e.g. instructing students with disabilities or leveraging new technology for instruction)\" but it has never been funded. Finally, this could also be explicitly added to the NSF's Advanced Technological Education (ATE) program.  84  For many U.S. STEM college graduates, doctorates in current form are not a marketable value proposition in the same way as for international students. Pursuing a postdoc can have a negative financial impact for years. On average, ex-postdocs, \"give up about one-fifth of their earning potential in the first 15 years after finishing their doctorates -which, for those who end up in industry, amounts to $239,970.\" See Devin Powell, \"The price of doing a postdoc.\" • Creating and Disseminating AI Educational and Career Information • Establishing AI Education and Training Tax Credits • Investing in Postsecondary AI Education and Scholarships • Facilitating Alternative Pathways into AI Jobs • Investing in PreK-12 AI Education and Experiences • Integrating K-12 AI Curriculum and Course Design • Cultivating and Supporting K-12 AI Educators • Funding AI Education and Careers Research \n Summary ..................................................................................................... \n Figure 1 . 1 Figure 1. CSET Research Series on the U.S. AI Workforce \n Figure 2 . 2 Figure 2. Summary of the U.S. AI Workforce \n Figure 3 . 3 Figure 3. Short-Term Policy Framework (Employer) \n Figure 4 . 4 Figure 4. Medium-and Long-term Framework (Life cycle) \n Figure 5 . 5 Figure 5. Stakeholders in AI Education and Workforce Policy \n ( Departments or agencies [\"agencies\"] involved [e.g., DOD]; Framework [short or medium/long]; Policy goal(s) being addressed [1 -Increase U.S. AI doctorates, 2 -Sustain and diversify nondoctorate technical talent, 3 -Promote AI literacy education]). \n : Congress should appropriate new funding for: -The U.S. Bureau of Labor Statistics (BLS), U.S. Census Bureau, and National Center for Science and Engineering Statistics (NCSES) to design, collect, compile, and publish occupation or skills-based data on the AI workforce along with education statistics from the National Center for Education Statistics (NCES). 48 (Agencies: NSF, U.S. Department of Commerce, Ed, and DOL; Framework: Medium-term; Policy goals: 1+2+3) -The BLS and Employment and Training Administration (ETA) to create school/career counselor and student/parent resources for AI technical and nontechnical careers, similar to an Occupational Outlook Handbook, along with a short video training for counselors. (Agencies: DOL; Framework: Medium-term; Policy goals: 1+2+3) -The National Artificial Intelligence Initiative Office of Education and Training to work with NSF, industry, and other stakeholders to provide guidance on approaching the online education offerings in the above dashboard to assist youth, parents, teachers, and counselors. \n ( Agencies: Ed; Framework: Long-term; Policy goals: 1+2+3) Rec: The NSF should allocate a portion of its ITEST program and Discovery Research preK-12 program (DRK-12) research dollars for AI education research. (Agencies: NSF; Framework: Long-term; Policy goals: 1+2+3) Rec: Ed should require all Regional Education Labs (RELs) to include AI education research as at least 15 percent of their research portfolios. \n Figure 6 . 6 Figure 6. Summary of AI Workforce Policy Recommendations \n ( Agencies: NSF, DOD, and DOE; Framework: Medium-term; Policy goal: 1) \n ( Agencies: NSF, DOE, DOD, and HHS; Framework: Long-term; Policy goals: 1+2)Rec: Congress should appropriate funding for federal grants to states for K-12 AI experiential learning opportunities. \n ( Agencies: NSF, DOE, DOD, and HHS, USDA; Framework: Long-term; Policy goals: 1+2) \n 31 Gehlhaus and Rahkovsky, \"U.S. AI Workforce: Labor Market Dynamics.\" \n \n Still, this report acknowledges the importance of STEM education policy as both a core part of S&T workforce policy and in AI education and workforce policy. That makes it equally as important to acknowledge STEM education and training policy is not a new phenomenon, but a very complex one that has gone on for decades. Despite showing mixed progress, recommendations from a 2010 report on K-12 education from the President's Council of Advisors on Science and Technology still ring true.7     1. Effective, safe, and trustworthy AI deployment and adoption is critical to future U.S. economic competitiveness and national security. AI is unique among emerging technologies, and rapidly evolving in terms of application and adoption. Without equitably providing quality AI education, we head toward greater economic and social disparities as demonstrated by the digital divide and accelerated by the COVID-19 pandemic. 5. To date there is no consistent definition of \"AI education\" or uniformly accepted curricula, as schools are still working to integrate computer science into classrooms. 2. The AI workforce is broader than the STEM and S&T workforce, including technical and nontechnical talent.3. Similarly, AI education is broader than STEM education, including concepts related to bias, governance, and ethics. This means STEM policy does not adequately address AI education and workforce policy needs.4. In a world characterized by ubiquity in AI, AI literacy of the population will be as important to national economic competitiveness as basic literacy and digital literacy. Without national elevation, U.S. states may have less urgency to make AI education and workforce training a priority. \n At the federal level, for example, we counted 16 agencies involved in STEM education programming. A 2012 study by the Congressional Research Service identified between 105 and 254 federal STEM education programs, depending on how the scope was defined. 10  State and local governments, as well as private industry and nonprofit organizations, also have their own STEM policies and educational agendas. 11  For example, the Utah STEM Action Center coordinates and communicates policies, resources, and events. Not surprisingly, there is limited coordination across federal and state governments. While efforts exist to coordinate initiatives and programs within the federal government (i.e., within OSTP) there is no obvious coordination with states or between states. As such, some companies and nonprofit entities are filling this gap. Organizations like Code.org and Microsoft's TEALS, for example, are working to provide equitable access to computer science education. However, without clear standards, the effectiveness of some of these programs is less clear, particularly in the case of forprofit coding bootcamps. 8  (We explain the design of the U.S. education system in detail in the CSET report \"Education in China and the United States: A Comparative System Overview.\" 9 ) Second, federal funding for education comes from an assortment of government departments and agencies. 12 Many other states similarly have their own STEM offices or initiatives; a map of these by state is provided in Appendix A.In terms of federal workforce development programs, few are specifically designed for technical career fields such as AI. Most employment and training programs are intended for specific populations such as dislocated workers, low-income adults, and other underrepresented and underserved communities. One notable exception is the Registered Apprenticeship program through the U.S. Department of Labor (DOL), but the impact and reach of this program for AI-related careers is fairly minimal. \n 15  Such a range of views held by economists and policymakers creates additional confusion in the scope and scale of federal and state S&T education initiatives and programs. This makes S&T education and workforce policy a very important, but extremely difficult challenge to effectively tackle without coordination. That sentiment has been articulated by both Democratic and Republican administrations: \"Over the past few decades, a diversity of Federal projects and approaches to K-12 STEM education across multiple agencies appears to have emerged largely without a coherent vision and without careful oversight of goals and outcomes.\" 16 (2010) \"More than a decade of studies by the National Academies of Sciences, Engineering, and Medicine document the need to prepare learners for the jobs of the future and identify a host of challenges and opportunities within the U.S. STEM education ecosystem.\" 17 (2018) \n Rec: The National Artificial Intelligence Initiative Office for Education and Training should be fully leveraged to coordinate federal and state U.S. AI education and training policy. We further recommend the National Artificial Intelligence Initiative Office for Education and Training submit, and Congress fund, a five-year funding proposal that enables all activities of the office. 41 There is a new opportunity for the federal government to coordinate AI education, training and workforce development policy, with the 2021 establishment of the Education and Training Office within the National Artificial Intelligence Initiative Office at the White House. 40 We recommend this office work under three guiding principles to facilitate equitable access and opportunity to AI education, training, and workforce development across the United Manage and oversee all AI federal education and training funding streams and initiatives. 43 Report annually on these efforts and relevant new legislation, working with NSTC, CoSTEM, the National Science Foundation (NSF), the U.S. Departments of Agriculture (USDA), Defense, Commerce, Education (Ed), Energy (DOE), and Labor (DOL), as well as the National Institute of Standards and Technology (NIST), and other relevant agencies. Coordinate these initiatives with states and relevant industry stakeholders. 44Compile: Track and report annually on federally funded AI and AIrelated initiatives, contests, grant challenges, scholarships, and AI education research in coordination with relevant federal agencies. Publish and maintain an AI/computer science education dashboard searchable by state for available after-school programs, contests, challenges, and conferences, scholarships, and online curricula.45     States: -Coordinate 42 -Compile -Convene Coordinate: \n Department of Education (Ed) to build in AI program information to its College Scorecard. (Agencies: Ed; Framework: Medium-term; Policy goals: 1+2+3) -A multi-platform, multi-year national AI careers public service announcement (PSA) campaign for youth and parents. 49 (Agencies: OSTP; Framework: Medium-term; Policy goals: 1+2+3) (Agencies: DOL, NSF, and Ed; Framework: Medium-term; Policy goals: 1+2+3) -The U.S.-Free virtual- \n chat career guidance and AI career assessments in conjunction with the AI careers landing page at BLS. (Agencies: DOL and Ed; Framework: Medium-term; Policy goals: 1+2+3)Establishing AI Education and Training Tax CreditsIt is important that employers participate in upskilling workers for AI jobs. However, our research finds only sporadic evidence this is currently happening. Across the companies we interviewed, few provided for or rewarded AI upskilling. For the companies that were embarking on related initiatives, they were generally for small cadres of AI talent or for broader categories of technical talent. This suggests there are opportunities for the federal government to help the private sector grow and cultivate AI talent internally by creating appropriate incentives.Rec: \n Congress should establish employer tax credits for employer-provided AI training, partnerships with community and technical colleges, and adult education programs that result in AI hires, 50 including from nondegree AI apprenticeships and other promising nondegree programs. 51 (Agencies: Internal Revenue Service; Framework: Short-term; Policy goals: 1+2+3) \n Investing in Postsecondary AI Education and Scholarships 52 Our research documents the clear need to increase the pipeline of U.S.-citizen doctorates, 53 as well as cultivate diverse and sustainable pathways for AI and AI-adjacent technical talent. This includes all postsecondary pathways, from undergraduates at twoyear and four-year institutions, to career and technical education programs, online boot camps, and adult learning centers.First, while there are many reasons for the lack of U.S.-citizen AI doctorates, research shows at least some of this is financially motivated. 54  Increasing the U.S. pipeline of PhD-level talent will address not only acute gaps in the supply of top-tier AI talent, but mitigate the challenge of AI and computer science faculty shortages and promote long-term U.S. AI competitiveness. Rec: Congress \n should fund NSF, DOE, and DOD graduate and postgraduate scholarships and fellowships for U.S. students pursuing AI and AI-related studies that are competitive with commensurate/peer private sector salaries. Moreover, NSF, DOE, and DOD should report annually on the number and demographic composition of applicants, awardees, and reviewers, along with the selection process. 55  For scholarships and fellowships funded by DOE and DOD, recipients should work collaboratively with a federally funded research lab during the funding period that is organized by either the university or scholarship administrator. (Agencies: NSF, DOD, and DOE; Framework: Medium-term; Policy goal: 1)Second, the U.S. government should provide greater access and opportunity for nontraditional students to pursue postsecondary coursework and majors in AI and AI-related fields of study.Rec: \n Congress should fund undergraduate scholarships and research fellowships related to AI and AI-related fields of study that cover tuition, cost of living, and incentive pay for acquiring AI and AI-related expertise .56  We further recommend the research fellowships be available to two-year and four-year college students, during the academic year and summers, and are coordinated by NSF AI Centers of Excellence in partnership with universities and federally funded research labs.57   Third, the U.S. government should increase the capacity of all postsecondary institutions to provide AI and AI-related courses and degree programs. This includes AI educational opportunities for students of all majors. AI education at the postsecondary level--particularly AI literacy, ethics, governance, and other introductory principles--should be integrated in all majors, not siloed into AI, engineering, or computer science departments with limited course capacity. Rec: Congress \n should appropriate funding for NSF to award grants to accredited two-or four-year postsecondary institutions for faculty and experienced AI and AI-related industry professionals to (1) complete professional development related to AI education, and (2) teach courses that integrate AI literacy, ethics, governance, and other introductory technical principles into the curricula. 58 (Agencies: NSF; Framework: Medium-term; Policy goals: 1+2)Rec: Congress \n should appropriate funding for NSF to award grants to sub-baccalaureate institutions, minority-serving institutions (MSI), tribal colleges and universities, and historically Black colleges and universities (HBCU) to equip facilities with \"AI labs\" for hands-on learning along with virtual learning platforms for any time remote or individualized learning. 59 A share of this funding should also support the hiring and training of laboratory support staff as well faculty to safely and effectively demonstrate, operate, and maintain all equipment and tools.(Agencies: NSF; Framework: Medium-term; Policy goals: 1+2) \n Congress should fund and task NIST, in coordination with industry and relevant trade associations, with establishing national, industry-recognized standards for AI certifications, stackable credential pathways, and sub-baccalaureate nondegree programs that can be program-accredited to enhance facilitate additional alternative pathways into AI careers, the U.S. government must better leverage the latent potential of community and technical colleges, MSIs, HBCUs, and other nontraditional pathways in training tomorrow's AI workforce.61     Rec: Congress should establish and fund a joint Ed, DOL, and NSF working group that oversees and administers a new AI workforce development grant program. 62 These grants will establish, track, and evaluate AI-specific public-private partnerships in a way that leverages the existing federal employment and training funding infrastructure. 63 We further propose prioritizing partnerships that create AI and AI-adjacent credential programs with guided pathways, stackable credentials, student advising, and other wraparound supports to increase opportunity and diversity in the AI workforce. Eligible participants should include state workforce boards, community and technical colleges, MSIs, HBCUs, and other underrepresented community groups and organizations. employer marketability. Employers lack clear, standard market signals from the value of certifications to use as a substitute for four-year degrees. If AI certification programs were accredited in the way colleges and universities are, it could legitimize certifications as a true pathway into the AI workforce. (Agencies: Ed, DOL, and U.S. Department of Commerce; Framework: Medium-term; Policy goals: 1+2)Regarding industry recognition of AI and related certifications, more coordinated effort is required at the national level to create industry-accepted national standards related to the quality of such certifications. To advocate that everyone needs a four-year degree misses the mark; four-year degrees are expensive64  and not all AI jobs need one. 65Rec:(Agencies: NIST; Framework: Short-term; Policy goals: 1+2) Rec: \n Congress should fund and task NCSES (within NSF) to issue a report on the new National Training, Education, and Workforce Survey with recommendations for enabling alternative pathways. 66 (Agencies: NSF; Framework: Medium-term; Policy goals: 1+2) Finally, our research found the federal government is a top employer of AI talent, particularly technical talent. 67 Redesigning eligibility criteria for AI and AI-adjacent jobs at the federal level could set an example for other employers to accept credentials outside of four-year college degrees. 68 Rec: Congress \n should task the Office of Personnel Management (OPM) to establish federal government hiring criteria and pathways for AI and AI-adjacent jobs that are based on portfolios of work and certifications. (Agencies: OPM; Framework: Short-term; Policy goals: 1+2)Investing in PreK- \n 12 AI Education and Experiences 69 Much of the above extends to PreK-12 education, and we propose analogous recommendations to ensure equity in access and opportunity for exposure to AI and AI-related education. We further emphasize the importance of access to experiential learning opportunities such as site visits, demonstrations, internships and externships, contests, challenges, and hackathons, and in-person after-school programs. For students with limited access to in-person opportunities, online education is critical but current offerings are wide-ranging and have unknown effectiveness in motivating or preparing youth for AI careers. 70 Our earlier recommendations for the National Artificial Intelligence Initiative Office for Education and Training were designed to address this uncertainty about what is effective. We also recommend: Rec: Congress \n should appropriate funding for federal grants to states for public K-12 schools to equip facilities with \"AI labs\" that enable hands-on learning along with virtual learning platforms for any time remote or individualized learning. (Agencies: NSF, DOE, DOD, and U.S. Department of Health and Human Services (HHS); Framework: Long-term; Policy goals: 1+2) Rec: Congress should appropriate funding for federal grants to states for K-12 AI experiential learning opportunities. For example, field trips to robotics engineering facilities, alternative or virtual reality simulations, unmanned aerial vehicle or drone demonstrations, scholarships to participate in AI or cyber hackathons, summer programs, conferences, and competitions, AI- related youth apprenticeships, and financial support for extracurricular activities, clubs, and other after school programs. 71 (Agencies: NSF, DOE, DOD, HHS, and USDA; Framework: Long-term; Policy goals: 1+2) Rec: Congress should authorize the Federal Communications Commission (FCC) or USDA funding to secure access to highspeed internet and connected devices, 72 with eligibility for any K-12 youth participating in the free lunch program. 73 (Agencies: FCC and USDA; Framework: Long-term; Policy goals: 1+2+3) \n Congress should appropriate funding for federal grants to rural and low-income school districts to integrate K-12 AI education designed using promising practices 79 and proven education models In a partnership with Denver Public Schools, researchers at the University of Colorado Boulder received a five-year, $20 million NSF grant to lead a research collaboration examining the role of AI in education and workforce development. The project will focus on how middle and high school students and teachers can leverage AI to augment learning and collaboration inperson and virtually. Through rigorous program evaluation, the study hopes to reveal insights on how to leverage AI to advance educational outcomes through human-machine collaboration and understand how teachers, students, and AI interact. The research team will also codesign middle and high school AI curricula that will be integrated into core courses, with the help of education stakeholders and the local community. Such efforts aim to prepare teachers and students to understand, critique, and design new uses of AI. .80  While a stand-alone course is better than no course, ideally, funded proposals would integrate AI education within core subjects. AI education could include using AI as a learning tool; building AI technical skills, statistics competency, and critical thinking; AI literacy; and AI career education. Example 2: Denver Public Schools (Colorado) As noted above, our AI Education Catalog shows accessibility gaps for AI education are more acute in rural and low-income school districts that lack access to AI education resources and opportunities. 77 In the long term such inequities could result in additional disparities in workforce opportunities and outcomes across segments of the population, meaning these districts should be prioritized. We believe with dedicated resourcing and using evidence-based best practices in learning and course design, it is possible to ensure equitable access to AI education across the United States. 78 goals: 1+2+3) Cultivating and Supporting K-12 AI Educators Successfully integrating AI education in the classroom requires having qualified educators. Instead of setting a target benchmark for recruiting new STEM educators, which many initiatives are currently doing, we propose empowering all teachers with the materials and resources they need to teach AI education in the Rec: (Agencies: USDA, DOE, NSF, and Ed; Framework: Long-term, Policy classroom. This includes technical curricula and general AI literacy. To provide this curriculum, GCPS is certifying teachers through internal training programs, partnering with local businesses, and procuring technology and infrastructure needed for instruction. Rec: \n Congress should fund and task Ed to create a national repository of peer-reviewed AI teaching materials, including off- the-shelf AI-enabled products, experiment kits, and in-class labs. 81(Agencies: Ed and NSF; Framework: Medium-term; Policy goals: 1+2+3) Rec: \n Congress should fund and task Ed to create a complementary national clearinghouse for AI/computer science and related education programs within the What Works Clearinghouse. 82 (Agencies: Ed and NSF; Framework: Medium-term; Policy goals: 1+2+3) Rec: \n Congress should create a national fund through the NSF for K-12 teachers to pursue AI training via AI certifications, AI conference attendance, and hosting AI curriculum, pedagogy, and course design professional development (PD) sessions. 83 effectively, (3) understanding and addressing AI career attrition, and (4) how to reimagine AI and AI-related doctoral programs to facilitate and encourage more U.S.-citizen applicants and graduates. Rec: Congress (Agencies: NSF; Framework: Medium-term; Policy goals: 1+2+3) Funding AI Education and Careers Research Concurrent to efforts that elevate access to AI education and training is a need to conduct AI education research at all education levels. This includes: (1) defining technical and general AI education, (2) integrating AI education into the classroom \n should commission a National Academies of Sciences (NAS) study on reimagining the design of STEM doctoral programs .84  The study should describe and characterize the applicant pool for AI and AI-related doctoral programs, and, if more U.S. students are applying than there are spots available, understand the barriers to increasing U.S. enrollment (e.g., competitiveness in preparation vs. faculty and funding constraints). (Agencies: NAS; Framework: Medium-term; Policy goal: 1) Rec: \n Congress should fund Ed to issue two funding challenges: (1) for K-12 integration of AI curricula into core course offerings 85 and (2) for public K-12 schools to have a career counseling and exploration course that include (a) module(s) on AI careers. Course offerings should consider: AI as a learning tool; building AI technical skills, statistics competency, and critical thinking; AI literacy; and AI career pathways. \n Artificial Intelligence Initiative Office for Education and Training Rec: The National Artificial Intelligence Initiative Office for Education and Training should be fully leveraged to coordinate federal and state U.S. AI education and training policies, and Congress should authorize five years of funding. \n Congress should fund the U.S. Department of Education (Ed) to build in AI program information to its College Scorecard. Congress should fund more NSF, DOE, and DOD graduate and postgraduate scholarships and fellowships for U.S. students pursuing AI and AI-related studies that are competitive with commensurate/peer private sector salaries. Agencies should also report annually on the composition of applicants, awardees, and application reviewers. (Agencies: Ed; Framework: Medium-term; Policy goals: 1+2+3) Rec: Congress should fund a multi-platform multi-year national AI careers PSA campaign for youth and parents. (Agencies: OSTP; Framework: Medium-term; Policy goals: 1+2+3) Rec: Congress should fund free virtual chat career guidance and AI career assessments in conjunction with the AI careers landing page at BLS. (Agencies: DOL, NSF, Commerce, and Ed; Framework: Medium-term; Policy goals: 1+2+3) Rec: Congress should fund the BLS and ETA to create school/career counselor and student/parent resources for AI technical and nontechnical careers, similar to an Occupational Outlook Handbook, along with a short video training for counselors. (Agencies: DOL; Framework: Medium-term; Policy goals: 1+2+3) Rec: Congress should fund the National Artificial Intelligence Initiative Office for Education and Training, NSF, industry, and other stakeholders to provide guidance on approaching the online education offerings in the above dashboard to assist youth, parents, teachers, and counselors. (Agencies: OSTP, Framework: NSF; Medium-term; Policy goals: 1+2+3)Rec:(Agencies: DOL and Ed; Framework: Medium-term; Policy goals: 1+2+3)Establishing AI Education and Training Tax CreditsRec: Congress should establish employer tax credits for employerprovided AI training, partnerships with community and technical colleges, and adult education programs that result in AI hires, including from nondegree AI apprenticeships and other promising nondegree programs.(Agencies: IRS; Framework: Short-term; Policy goals: 1+2+3)Investing in Postsecondary AI Education and ScholarshipsRec: \n The Strengthening Career and Technical Education for the 21st Century Act): For example, NVIDIA offers a research grant for AI related research, the National Oceanic and Atmospheric Administration offers the Ernest F. Hollings Undergraduate Scholarship that provides undergraduate students with two years of academic assistance and a full-time summer internship, and the CTSA provides scholarships to educators to use for professional development opportunities focused on addressing inequity in computer science education. Programs are free to apply Scholarship programs' duration ranges from a one-time scholarship award amount to a set fellowship during the duration of a degree. We found that they are funded by a mix of government organizations, nonprofits, and private for-profit organizations. With the diverse section of programs in this category it is difficult to make generalizations about the data for the section since they all operate differently. provide financial assistance to underrepresented and underserved students. Provides grants to develop and support career technical education programs at secondary and postsecondary institutions. Programs funded by the legislation are administered by the U.S. Department of Education and funds are allocated to states by formula. Funds are used to help recipients attain an industryrecognized credential, a certificate, or a postsecondary degree.• Registered Apprenticeship system: A registry of apprenticeship programs, maintained by states and administered by the U.S. Department of Labor.Apprenticeship sponsors (e.g., employers, unions, industry groups, etc.) within these registered programs receive preferential treatment in federal systems, making them eligible for Workforce Innovation and Opportunity Act funding and other federal programs, and apprentices may receive a nationally-recognized credential. for STEM Education (Washington, DC: National Science & Technology Council, December 2018), https://trumpwhitehouse.archives.gov/wpcontent/uploads/2018/12/STEM-Education-Strategic-Plan-2018.pdf. The strategy states: \"this report sets out a Federal strategy for the next five years based on a Vision for a future where all Americans will have lifelong access to high-quality STEM education and the United States will be the global leader in STEM literacy, innovation, and employment.\" Note it is unclear if this is still in effect or if it was de facto cancelled with the new administration, which happens and is an example of the challenges with such policy efforts. Here, technical talent includes those with knowledge, skills, and abilities to engage in the design, development and deployment of AI or AI-enabled capabilities. Nontechnical talent includes those in roles that complement technical talent such as user experience designers, compliance officers, and program managers. Nontechnical talent should have AI literacy. President's Council of Advisors on Science and Technology, Prepare and Inspire: K-12 Education in Science, Technology, Engineering, and Math (STEM) for America's Future (Washington, DC: Executive Office of the President, September 2010), https://nsf.gov/attachments/117803/public/2a--there is variation by geographic region or institutional tier or type. Reasons could include departmental funding challenges, faculty recruitment challenges, and the incentive for tenure-track faculty to conduct research and publish over teaching courses (which also has a significant impact on STEM doctorate experiences and attrition). Key Numbers • 29 programs identified • 14 programs overseen by the government • 17 programs targeted towards postsecondary students • 9 programs targeted towards professionals • Strategy • Perkins V (6 7 \n 40 \"William M. (Mac) Thornberry National Defense Authorization Act for Fiscal Year 2021,\" H.R. 6395, 116 th Cong. (2020), https://www.congress.gov/bill/116th-congress/house-bill/6395.41  The legislation establishing this office also stipulates for funding as is submitted in the President's annual budget through the Office of Management and Budget.42 Even more essential if current proposed legislation passes which calls for a massive increase in K-16 STEM education programming and related grants; we propose below similar initiatives that are AI-specific and can be leveraged from this funding stream. See the House of Representatives' \"National Science Foundation for the Future Act\" (H.R. 2225, 117 th Cong. (2021), https://www.congress.gov/117/bills/hr2225/BILLS-117hr2225eh.pdf) and the Senate version \"United States Innovation and Competition Act of 2021\" (S. 1260, 117 th Cong. (2021), https://www.congress.gov/117/bills/s1260/BILLS-117s1260es.pdf). 43 We note this office is already supposed to be coordinating federal efforts across agencies. See \"Summary of AI Provisions from the National Defense Authorization Act 2021\" (Stanford Human-Centered Artificial Intelligence [HAI]), accessed July 26, 2021, https://hai.stanford.edu/policy/policyresources/summary-ai-provisions-national-defense-authorization-act-2021. See the one-stop AI education and training marketplace for Singapore, \"AI Singapore,\" accessed July 26, 2021, https://aisingapore.org/. Meanwhile, OSTP's new STEM4ALL federal programs portal is more limited; see Networking and Information Technology Research and Development (NITRD), \"R&D Workforce Training: Federal Agencies' STEM Internships, Scholarships, and Training Opportunities,\" accessed July 26, 2021, https://www.nitrd.gov/STEM4ALL/. 46 Gehlhaus, \"Youth Information Networks and Propensity to Serve in the Military.\" 47 For challenges with existing AI workforce data collected and published by the federal government, see Gehlhaus and Mutis, \"The U.S. AI Workforce: Understanding the Supply of AI Talent.\" could include creating new occupational categories, fielding a household and/or establishment survey supplement, and engaging in other representative data capture that accounts for declining survey participation rates. 44  For example, the National Governors Association, states' departments of education, state legislatures, and state workforce boards.45   \n\t\t\t (Agencies: NSF, DOE, and DOD; Framework: Medium-term; Policy goal: 2) \n\t\t\t (Agencies: Ed; Framework: Long-term; Policy goals: 1+2+3) \n\t\t\t © 2021 by the Center for Security and Emerging Technology. This work is licensed under a Creative Commons Attribution-Non Commercial 4.0 International License. To view a copy of this license, visit https://creativecommons.org/licenses/by-nc/4.0/. Document Identifier: doi: 10.<PHONE> \n\t\t\t National Security Commission on Artificial Intelligence, Final Report (Washington, DC: NSCAI, March 2021), https://www.nscai.gov/wpcontent/uploads/2021/03/Full-Report-Digital-1.pdf.4 See OSTP's National AI Initiative Office, \"Education and Training,\" https://www.ai.gov/strategic-pillars/education-and-training/. To date, the office created a portal for federal STEM initiatives and scholarship opportunities; is overseeing a study on the Future of Work; and is advocating for school access to computing resources.5 Committee on STEM Education, Charting a Course for Success: America's \n\t\t\t In addition to challenges in program effectiveness for improving workforce outcomes.14 See Appendix C for more information on WIOA and other relevant laws governing federal education and workforce policy. \n\t\t\t We note our policy recommendations are complementary to NSCAI's NDEA II recommendation. \n\t\t\t This could be similar to military recruiting, except directing people to AI", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-U.S.-AI-Workforce-Policy-Recommendations.tei.xml", "id": "1e5b6749bb081ab90d1cb384851cb6e6"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Introduction 2. Precisifying strong longtermism 2.1 Axiological strong longtermism (ASL) 2.2 Benefit ratio (BR) and ASL 3. The size of the future 4. Tractability of significantly affecting the far future 4.1 Influencing the choice among persistent states 4.2 Mitigating risks of premature human extinction 4.3 Influencing the choice among non-extinction persistent states 4.4 Uncertainty and 'meta' options 5. Strong longtermism about individual decisions 6. Robustness of the argument 7. Cluelessness", "authors": ["Hilary Greaves", "William MacAskill"], "title": "The Case for Strong Longtermism", "text": "Introduction A striking fact about the history of civilisation is just how early we are in it. There are 5000 years of recorded history behind us, but how many years are still to come? If we merely last as long as the typical mammalian species, we still have over 200,000 years to go  (Barnosky et al. 2011) ; there could be a further one billion years until the Earth is no longer habitable for humans  (Wolf and Toon 2015) ; and trillions of years until the last conventional star formations  (Adams and Laughlin 1999:34) . Even on the most conservative of these timelines, we have progressed through a tiny fraction of history. If humanity's saga were a novel, we would be on the very first page. Normally, we pay scant attention to this fact. Political discussions are normally centered around the here and now, focused on the latest scandal or the next election. When a pundit takes a \"long-term\" view, they talk about the next five or ten years. With the exceptions of climate change and nuclear waste, we essentially never think about how our actions today might influence civilisation hundreds or thousands of years hence. We believe that this neglect of the very long-run future is a serious mistake. An alternative perspective is given by longtermism, according to which we should be particularly concerned with ensuring that the far future goes well (MacAskill MS). In this article we go further, arguing for strong longtermism: the view that impact on the far future is the most important feature of our actions today. We will defend both axiological and deontic versions of this thesis. Humanity, today, is like an imprudent teenager. The most important feature of the most important decisions that a teenager makes, like what subject to study at university and how diligently to study, is not the enjoyment they will get in the short term, but how those decisions will affect the rest of their life. The structure of the paper is as follows. Section 2 sets out more precisely the thesis we will primarily defend: axiological strong longtermism (ASL). This thesis states that, in the most important decision situations facing agents today, (i) every option that is near-best overall is near-best for the far future, and (ii) every option that is near-best overall delivers much larger benefits in the far future than in the near future. We primarily focus on the decision situation of a society deciding how to spend its resources. We use the cost-effectiveness of antimalarial bednet distribution as an approximate upper bound on attainable near-future benefits per unit of spending. Towards establishing a lower bound on the highest attainable far-future expected benefits, section 3 argues that there is, in expectation, a vast number of sentient beings in the future of human-originating civilisation. Section 4 then argues, by way of examples involving existential risk, that the project of trying to beneficially influence the course of the far future is sufficiently tractable for ASL(i) and ASL(ii) to be true of the above decision situation. Section 5 argues that the same claims and arguments apply equally to an individual deciding how to spend resources, and an individual choosing a career. We claim these collectively constitute the most important decision situations facing agents today, so that ASL follows. The remainder of the paper explores objections and extensions to our argument. Section 6 argues that the case for ASL is robust to several plausible variations in axiology, concerning risk aversion, priority to the worse off, and population ethics. Section 7 addresses the concern that we are clueless about the very long-run effects of our actions. Section 8 addresses the concern that our argument turns problematically on tiny probabilities of enormous payoffs. Section 9 turns to deontic strong longtermism. We outline an argument to the effect that, according to any plausible non-consequentialist moral theory, our discussion of ASL also suffices to establish an analogous deontic thesis. Section 10 summarises. The argument in this paper has some precedent in the literature. Nick  Bostrom (2003)  has argued that total utilitarianism implies we should maximise the chance that humanity ultimately settles space. Nick  Beckstead (2013)  argues, from a somewhat broader set of assumptions, that \"what matters most\" is that we do what's best for humanity's long-term trajectory. In this paper, we make the argument for strong longtermism more rigorous, and we show that it follows from a much broader set of empirical, moral and decision-theoretic views. In addition, our argument in favour of deontic strong longtermism is novel. We believe that strong longtermism is of the utmost importance: that if society came to adopt the views we defend in this paper, much of what we prioritise in the world today would change. \n Precisifying strong longtermism \n Axiological strong longtermism (ASL) Strong longtermism could be made precise in a variety of ways. First, since we do not assume consequentialism, we must distinguish between axiological and deontic claims. Let axiological (resp. deontic) strong longtermism be the thesis that far-future effects are the most important determinant of the value of our options (resp. of what we ought to do). It remains imprecise what \"most important determinant\" means. Taking the axiological case first, in this paper we consider the following more precise thesis: 1 Axiological strong longtermism (ASL): In the most important decision situations facing agents today, (i) Every option that is near-best overall is near-best for the far future. (ii) Every option that is near-best overall delivers much larger benefits in the far future than in the near future. Where condition (i) holds, one can identify the near-best options by focussing in the first instance only on far-future effects. If (as we believe, but will not argue here) the analogous statement regarding near-future effects is not also true, that supplies one sense in which far-future effects are \"the most important\". Where condition (ii) holds, the evaluation of near-best options is primarily driven by far-future effects. That supplies another such sense. In sections 3-5, we will argue that clauses (i) and (ii) of ASL hold of particular decision situations: those of a society deciding how to spend money with no restrictions as to 'cause area', an individual making the analogous decision, and individual career choice. Because these decision situations have particularly great significance for the well-being of both present and future sentient beings, we claim, they are the most important situations faced by agents today. Therefore, strong longtermism follows, even if ASL(i) and (ii) do not hold of any other decision situations. Throughout, \"the far future\" means everything from some time t onwards, where t is a surprisingly long time from the point of decision (say, 100 years). \"The near future\" means the time from the point of decision until time t. We will interpret both \"near-best overall\" and \"near-best for the far future\" in terms of proportional distance from zero benefit to the maximum available benefit, and \"much larger\" in terms of a multiplicative factor. As we intend it, ASL is not directly concerned with the objective value of options and their actual effects. Rather, terms like \"near-best\" and \"benefits\" relate to the ex ante value of those options, given the information available at the time of decision, and their prospects for affecting the near or far future. Ex ante value may be expected value, but the statement of ASL does not presuppose this. Since it refers to \"benefits\", ASL makes sense only relative to a status quo option: benefits are increases in value relative to the status quo. As above, our primary examples will be cases of deciding how to spend some resource (either money or time). For concreteness, we will then take the status quo to be a situation in which the resources in question are simply wasted. However, other plausible choices of status quo would be unlikely to significantly affect our argument, and the argument does not require that the status quo be special in any deep sense. ASL makes only comparative claims. We do not claim, and nor do we believe, that options cannot deliver large benefits without being near-best for the far future, or that available near-future benefits are small in any absolute sense. Our claim is rather that available benefits for the far future are many times larger even than this. \n Benefit ratio (BR) and ASL Our argument for ASL proceeds via the intermediate claim that the following property holds of the decision situations in question: \n Benefit ratio (BR): The highest far-future ex ante benefits that are attainable without net near-future harm are many times greater than the highest attainable near-future ex ante benefits. We prove in the Appendix that if BR holds of a given decision situation, then (firstly) so does ASL(ii), and (secondly) ASL(i) holds of a certain restriction of that decision situation. (The restriction involves removing any options that do net expected near-future harm; this restriction is innocuous in the context of our argument.) Evaluating BR, and hence ASL, requires quantitative analysis; any particular quantitative analysis requires strong evaluative assumptions. To this end, we will temporarily make a particular, plausible but controversial, set of evaluative assumptions. Section 6, however, shows that various plausible ways of relaxing these assumptions leave the basic argument intact. One controversial assumption that may be essential, concerning the treatment of very small probabilities, is discussed in section 8. The inessential assumptions in question include the following. First, we will identify the ex ante value of an option with its expected value: the probability-weighted average of the possible values it might result in. Second, we will identify value with total welfare: that is, we will assume a total utilitarian axiology. Third, and a near-corollary of the latter, we will assume time-separability. The latter allows us to separately define near-future and far-future benefits: overall value is then simply the sum of near-future value and far-future value, where these in turn depend only on near-future (respectively, far-future) effects. For a rough upper bound on near-future expected benefits in the context of a society spending money, we consider the distribution of long-lasting insecticide-treated bednets in malarial regions, which saves a life on average for around $4000. Each $100 therefore saves on average 0.025 lives in the near future (GiveWell 2020a).  2  We cannot argue that this is the action with the very largest near-future benefits. In particular, though it seems hard to beat this cost-effectiveness level via any intervention that is backed by rigorous evidence, it might be possible to achieve higher short-term expected benefits via some substantially more speculative route. A full examination of the case for strong 3 longtermism would involve investigation of this, and the corresponding sensitivity analysis. However, even quite large upward adjustments to the figure we use here would leave our argument largely unaffected. We emphasise that we are not considering the long-run knock-on effects of bednet distribution. It is possible, for all we say in this paper, that bednet distribution is the best way of making the far future go well, though we think this unlikely.  4  We will argue in section 4 that, for a society's decision about how to spend its resources, the lower bound on attainable far-future expected benefits is many times higher than this upper bound for near-future expected benefits, and therefore BR holds of this decision situation. Section 5 discusses related decision situations facing individuals. \n The size of the future There is, in expectation, a vast number of lives in the future of human civilisation. Any 5 estimate of just how \"vast\" is of course approximate. Nonetheless, we will argue, existing work supports estimates that are sufficiently large for our argument to go through. There are several techniques one can use for estimating the expected number of future beings. Let us start with the question of the expected duration of humanity's future existence, temporarily setting aside questions of how large the population might be at any future time. Firstly, one might use evidence regarding the age of our species to ground judgments on the annual risk of extinction from natural causes, and extrapolate from there. Given that Homo sapiens has existed for over 200,000 years,  Snyder-Beattie et al. (2019:2)  thereby estimate that the expected future lifespan of humanity is at least 87,000 years, as far as natural causes of extinction are concerned. Secondly, one might undertake reference class forecasting  (Kahneman and Lovallo 1993; Flyvbjerg 2008 ). Here, the lifespans of other sufficiently similar species serve as benchmarks. Estimates of the average lifespan of mammalian species (resp. hominins) are between 0.5 and 6 million years (resp. around 1 million years)  (Snyder-Beattie et al. 2019:6) . Thus reference class forecasting, naively applied, suggests at least 1 million years for the expected future duration of humanity. Both of these estimates, however, ignore the fact that humans today are highly atypical. Humanity today is significantly better equipped to survive extinction-level threats than either other species are, or than our own species was in the past, thanks to a combination of technological capabilities and geographical diversity. Therefore a range of substantially higher benchmarks is also relevant: for instance, the frequency of mass extinction events (1 in every 30-100 million years (Snyder-Beattie et al. 2019:7)), and the time over which the Earth remains habitable for humans (around 1 billion years  (Adams 2008:34) ). The above figures concern the expected duration of humanity's future. Since we are interested in the expected number of future beings, we also need to consider population size. We again consider several benchmarks. First, the UN Department of Economic and Social Affairs (2019:6) projects that the global population will plateau at around 11 billion people by the year 2100. Second, the large majority of estimates of the Earth's \"carrying capacity\"that is, its long-run sustainable population, based on relatively conservative assumptions about future technological progress -are over 5 billion, and sometimes substantially higher  (Cohen 1998:342; Bergh and Rietveld 2004:197) . Third, for predicting the further future, we might extrapolate from the historical trend of human population increasing over time. Given this trend, it is at least plausible that continued technological advances will enable an even larger future population up to some much higher plateau point (say, 1 trillion), even if we cannot currently foresee the concrete details of how that might happen  (Simon 1998) . Importantly, it is the expected number of future beings, not the median, that is relevant for our purposes. In addition to the possibility of numbers like the higher benchmarks indicated above, it is of course also possible that the future duration and/or population size of humanity are much smaller. However, the effects of these possibilities on the expected number are 6 highly asymmetric. Even a 50% credence that the number of future beings will be zero would decrease the expected number by only a factor of two. In contrast, a credence as small as 1% that the future will contain, for example, 1 trillion beings per century for 100 million years (rather than 10 billion per century for 1 million years) increases the expected number by a factor of 100. We must also consider two more radical possibilities that, while very uncertain, could greatly increase the duration and future population sizes of humanity. The first concerns space settlement. There are currently no known obstacles to the viability of space settlement, and some scientific investigations suggesting its feasibility using known science  (Sandberg and Armstrong 2013; Beckstead 2014 ). If humanity lives not only on Earth but also on other planets -in our own solar system, elsewhere in the Milky Way, or in other galaxies toothen terrestrial constraints on future population size disappear, and astronomically larger populations become possible. Even if we only settle the solar system, civilisation would have over 5 billion years until the end of the main sequence lifetime of the Sun  (Sackmann et al. 1993:462; Schröder and Smith 2008:157-8) , and we would have access to over two billion times as much sunlight power as we do now  (Stix 2002:6; Sarbu and Sebarchievici 2017:16 ). If we are able to widely settle the rest of the Milky Way, then we could access well over 250 million rocky habitable-zone planets  (Bryson et al. 2021:22) , each of which has the potential to support trillions of lives over the course of their sun's lifetimes. Moreover, an interstellar civilisation could survive until the end of the stelliferous era, on the order of ten trillion years hence  (Adams and Laughlin 1999) . If we consider possible settlement of the billions of other galaxies accessible to us, the numbers get dramatically larger again. The second radical possibility is that of digital sentience: that is, conscious artificial intelligence (AI). The leading theories of philosophy of mind support the idea that consciousness is not essentially biological, and could be instantiated digitally  (Lewis 1980; Chalmers 1996: ch. 9 ). And the dramatic progress in computing and AI over just the past 70 years should give us reason to think that if so, digital sentience could well be a reality in the future. It is also plausible that such beings would have at least comparable moral status to humans  (Liao 2020) , so that they count for the purposes of the arguments in this paper.  7  Consideration of digital sentience should increase our estimates of the expected number of future beings considerably, in two ways. First, it makes interstellar travel much easier: it is easier to sustain digital than biological beings during very long-distance space travel  (Sandberg 2014:453) . Second, digital sentience could dramatically increase the number of beings who could live around one star: digital agents could live in a much wider variety of environments  (Sandberg 2014:453) , and could more efficiently turn energy into conscious life  (Bostrom 2003:309) . One might feel sceptical about these scenarios. But given that there are no known scientific obstacles to them, it would be overconfident to be certain, or near-certain, that space settlement, or digital sentience, will not occur. Imagine that you could peer into the future, and thereby discovered that Earth-originating civilisation has spread across many solar systems. How surprised would you be, compared to how surprised you would be if you won the lottery? To move towards particular numbers, we consider three specific future scenarios, taken from  Newberry (2021a) , where civilisation is: (i) Earthbound; (ii) limited to the Solar System; and (iii) expanded across the Milky Way. In each case, Newberry makes a conservative estimate of the carrying capacity of civilisation in that scenario, on the assumptions that digital life is and is not possible, giving six scenarios in all. He also provides a best-guess estimate of the duration of civilisation in that scenario. These scenarios are not meant to exhaust the possibility space, but they give an indication of the potential magnitudes of future population size: To arrive at an overall estimate of the expected number of future people, one would further need to estimate probabilities for scenarios such as those above (and for all other scenarios). However, since the number of lives in the future according to different possible scenarios is spread over many orders of magnitude, in any such expected value calculation, it tends to be the \"largest\" scenario in which one has any nonzero credence that drives the overall estimate. Even a 0.01% credence that biological humanity settles the Milky Way at carrying capacity, for example, contributes at least 10 32 to the expected number of future beings. Precisely how one's remaining credence is spread among \"smaller\" scenarios then makes very little difference. Because of this, we believe that any reasonable estimate of the expected number of future beings is at least 10 24 . (In fact, we believe that any reasonable estimate must be substantially higher than this; since higher numbers would make little difference to the arguments of this paper, however, we will not press that case here.) However, we are also sympathetic to the concern that if this is the only estimate we consider, the case for strong longtermism would be driven purely by tiny credences in highly speculative scenarios. We will therefore also consider the extent to which the same arguments would go through on some vastly more conservative estimates, as follows: Expected number of future beings Main estimate 10 24 Low estimate 10 18 Restricted estimate 10 14 Our low estimate (10 18 ) corresponds, for instance, to a <PHONE>% credence in the Solar System (biological life) scenario, with zero credence in either digital sentience or more wide-ranging space settlement. Our restricted estimate (10 14 ) corresponds to the above estimate for Earthbound life, with zero credence in any larger-population scenario (including both digital sentience and any space settlement). In the arguments that follow, the reader is invited to substitute her own preferred estimate throughout. We will argue that BR (and hence ASL) holds of society's decision situation even on our restricted estimate, and clearly holds by a large margin on our main estimate. \n Tractability of significantly affecting the far future The far-future effects of one's actions are usually harder to predict than their near-future effects. Might it be that the expected instantaneous value differences between available actions decay with time from the point of action, and decay sufficiently fast that in fact the near-future effects tend to be the most important contributor to expected value? If that were so, then neither BR nor ASL would hold. This is a natural reason to doubt strong longtermism. We will call it the washing-out hypothesis. 8 We agree that the washing-out hypothesis is true of some decision situations. However, we claim that it is false of our society's decision situation. Given the argument of section 2, our task is to show that there exists at least one option available to society with the property that its far-future expected benefits are significantly greater than the near-future expected benefits of bednet distribution (that is, recall: 0.025 lives saved per $100 spent). We will consider examples in two categories: mitigating extinction risk, and positively shaping the development of artificial superintelligence. \n Influencing the choice among persistent states Here is an abstract structure which, insofar as it is instantiated in the real world, offers a recipe for identifying options whose effects will not wash out. Consider the space S of all possible fine-grained states the world could be in at a single moment of time (that is, the space of all possible instantaneous microstates). One can picture the history of the universe as a path through this space. Let a persistent state be a subset A of S with the property that, given the dynamics of the universe, if the instantaneous state of the world is in A, then the expected time for which it remains in A is extremely long. Now suppose that there are two or more such persistent states, differing significantly from one another in value. Suppose further that the world is not yet in any of the states in question, but might settle into one or the other of the states in question in the foreseeable future. Finally, suppose that there is something we can do now that changes the probability that the world ends up in a better rather than a worse persistent state. Then, as a result of the persistence that is built into the definition, the effects of these actions would not wash out at all quickly. The empirical question is whether there are, in the real world, any options available that instantiate the structure just described. We claim that there are. \n Mitigating risks of premature human extinction The non-existence of humanity is a persistent state par excellence. To state the obvious: the chances of humanity re-evolving, if we go extinct, are miniscule. Only slightly more subtly, the existence of humanity is also a persistent state: while we face significant risks of premature extinction, as argued in section 3, humanity's expected persistence is vast. These persistent states have unequal expected value. Assuming that on average people have lives of significantly positive welfare, according to total utilitarianism the existence of 9 humanity is significantly better than its non-existence, at any given time. Combining this with  9  We return to this assumption in section 6. the fact that both states are persistent, premature human extinction would be astronomically bad. Correspondingly, even an extremely small reduction in extinction risk would have very high expected value  (Bostrom 2013:18) . For example, even if there are 'only' 10 14 lives to come (as on our restricted estimate), a reduction in near-term risk of extinction by one millionth of one percentage point would be equivalent in value to a million lives saved; on our main estimate of 10 24 expected future lives, this becomes ten quadrillion (10 16 ) lives saved. As is increasingly recognised, as an empirical matter of fact, there are things we could do that would reduce the chance of premature human extinction by a non-negligible amount. As a result, although precise estimates of the relevant numbers are difficult, the far-future benefits of some such interventions seem to compare very favourably, by total utilitarian lights, to the highest available near-future benefits. The detection and potential deflection of asteroids provides a relatively robust example of such an intervention. This involves scanning the skies to identify asteroids that could potentially collide with Earth and, if one were found, investing the resources to try to deflect it, and/or to prepare bunkers and food stockpiles to help us survive an impact winter. Most of the expected costs here are in detection, because the costs of deflection and preparation are only paid in the very unlikely event that one does detect an asteroid on a collision course with Earth. In 1996, NASA commenced the Spaceguard Survey, a multi-decade plan to track near-Earth objects with the aim of identifying any on impact trajectories. At a total cost of $71 million (USD) by 2012, the Spaceguard Survey had tracked over 90% of asteroids of diameter 1km or more in near-Earth orbit, and all asteroids of diameter 10km or more over 99% of the sky. It is not certain that a large asteroid collision would cause human beings to go extinct. We assume a status quo risk of human extinction, conditional on the impact of a 10km+ asteroid, of 1%. It is also far from certain that we could deflect a 10km+ asteroid, even if we knew it was on a collision course. However, it is far from certain that we could not, and, as above, there are other actions we could take to protect against the extinction risk. We assume here that if such an object were detected to be on a collision course, our deflection and preparation efforts would reduce extinction risk by a proportional 5%. The assumptions in this paragraph follow  Newberry (2021b) , and seem fairly conservative. Putting these numbers together, we estimate that the Spaceguard Survey, on average, reduced extinction risk by at least 5 × 10 -16 per $100 spent. On our main estimate of the expected number of future beings, this amounts to 500 additional million lives; this decreases to 500 or 0.05 lives on our low and restricted estimates, respectively. Of course, we should expect further work on asteroids to have lower cost-effectiveness, because of diminishing marginal returns. However, the opportunity remains significant. The remaining risk of a 10km+ asteroid collision in the next 100 years has been estimated at 1 in 150 million  (Ord 2020:71) . It has been estimated that the cost to detect with near-certainty any remaining asteroids of greater than 10km diameter would be at most a further $1.2 billion  (Newberry 2021b ). On our main (resp. low, restricted) estimate of the expected number of future beings, every $100 of this would, on average, result in 300,000 (resp. 0.3, 0.00003) additional lives. This example therefore supports strong longtermism on our main and low estimates, though not on the restricted estimate. Organisations whose work mitigates risk of extinction from asteroid impacts, and which would benefit from substantially more funding, include the Planetary Society and the B612 Foundation. While asteroid defense is among the more easily quantified areas of extinction risk reduction, it is far from the only one, or the most significant  (Ord 2020: ch. 3 ). Another possibility concerns global pandemics. Such a pandemic could be natural or man-made, with the latter being particularly concerning  (Posner 2004:75-84 ; Rees 2018: sec. 2.1; Ord 2020). In particular, progress in synthetic biology is very rapid (Meng and Ellis 2020), and it is likely that we will soon be able to design man-made viruses with very high contagiousness and lethality. If such pathogens were released (whether deliberately or by accident  (Shulman 2020; Ord 2020:129-131 )) in the course of military tensions, or by a terrorist group, there is a real possibility that they could kill a sufficient number of people that the human species would not recover. In a recent paper, Millet and Snyder-Beattie (2017) use three distinct methods to generate estimates of the risk of an extinction-level pandemic in the next 100 years. The resulting estimates range from 1 in 600,000 to 1 in 50. The authors further use figures from the World Bank to generate a very conservative estimate that $250 billion of spending on strengthening healthcare systems would reduce the chance of such extinction-level pandemics this coming century by at least a proportional 1%.  10  Taking the geometric mean to average across the two methods that generate the lower estimates for extinction risk, we obtain a risk of about 1 in 22,000 of extinction from a pandemic over the next 100 years. If we use the above figure of $250 billion to reduce the working on these threats include the John Hopkins Center for Health Security, the Nuclear Threat Initiative's biosecurity program, and Gryphon Scientific. \n Influencing the choice among non-extinction persistent states A second way of positively impacting the long run is by improving the value of the future conditional on the existence of a very large number of future sentient beings. For concreteness, we focus on one way of doing this: positively shaping the development of artificial superintelligence (ASI), that is, artificial systems that greatly exceed the cognitive performance of humans in virtually all domains of interest.  12  The idea that the development of sufficiently advanced artificial intelligence could prove a key turning point in history goes back to the early computer pioneers Alan  Turing (1951)  and I.J.  Good (1966) . It has more recently been the subject of wider concern. There are two 13 classes of long-term worry. The first is from AI-takeover scenarios  (Bostrom 2014; Russell 2019 ). This worry is that, once we build a human-level artificial intelligence, it would be able to recursively self-improve, designing ever-better versions of itself, quickly becoming superintelligent. From there, in order to better achieve its aims, it will try to gain resources, and try to prevent threats to its survival. It would therefore be incentivised to take over the world and eliminate or permanently suppress human beings. Because the ASI's capability is so much greater than that of humans, it would probably succeed in these aims. The second worry is from entrenchment scenarios (MacAskill MS). If an authoritarian country were the first to develop ASI, with a sufficient lead, they could use this technological advantage to achieve world domination. The authoritarian leader could then quash any ideological competition. An AI police force could guarantee that potential rebellions are prevented; an AI army would remove any possibility of a coup. And if the leader wanted his ideology to persist indefinitely, he could pass control of society on to an ASI successor before his death. To this end, he could hard-code the goals of the ASI to match his own, have the ASI learn his goals from his speech and behaviour, or even 'mind upload', scanning his brain and having it digitally emulated  Sandberg 2013) . In either of these scenarios, once power over civilisation is in the hands of an ASI, this could persist as long as civilisation does (Riedel MS). Different versions of the ASI-controlled futures are therefore persistent states with significantly differing expected value, so that we have another instantiation of the structure outlined in section 4.1. The ruler-ASI could monitor every aspect of society. And it could replicate itself indefinitely, just as easily as we can replicate software today; it would be immortal, freed from the biological process of aging. The value of the resulting world would depend in considerable part on the goals of the ruler-ASI. Though extinction risks involve dramatic reductions in the size of the future population, these AI scenarios need not. In the classic statement of the AI-takeover scenario, the ASI goes on to settle the stars in pursuit of its goals  (Bostrom 2014:100) . Similarly, if an authoritarian leader transferred power to an ASI, they too might want their civilisation to be large, populous and long-lasting. In particular, for a wide variety of goals (such as building the grandest possible temples, doing research, or, in a toy example  Bostrom (2014:123-4)  gives to illustrate the general phenomenon of misaligned AI, maximising the number of paperclips), acquiring more resources helps with achievement of these goals, which motivates settling the stars. And, in order to fulfill these goals, a populous workforce would be instrumentally valuable. In expectation, the number of future beings, in these scenarios, is very large. Now, this workforce might consist almost entirely of AIs. But, as we noted in section 3, there are reasons to think that such beings would have moral status, and therefore how well or poorly their lives went would be of moral concern, relevant to the arguments of this paper. And, at least on the authoritarian-takeover scenarios, the ruler might wish to have a very large number of human followers, too. There are two strands of work aimed at reducing risks from ASI. First, AI safety research, which aims to ensure that AI systems do what we intend them to do  (Amodei et al. 2016 ). Such work is conducted by organisations such as Berkeley's Center for Human-Compatible AI, the Machine Intelligence Research Institute, and labs within Google DeepMind and OpenAI. Second, policy work, in particular to ensure a cooperative approach between countries and companies: for example, by The Partnership on AI, the Centre for the Governance of AI, and the Center for New American Security. Despite this work, ASI safety and policy are still extremely neglected. For example, the Open Philanthropy Project is the only major foundation with these issues as a key focus area; it spends under $30 million per year on them (Open Philanthropy 2020). The AI safety teams 14 at OpenAI and DeepMind are small. There is no hard quantitative evidence to guide cost-effectiveness estimates for AI safety work. Expert judgment, however, tends to put the probability of existential catastrophe from ASI at 1-10%. Given these survey results and the arguments we have canvassed, we think 15 that even a highly conservative assessment would assign at least a 0.1% chance to an AI-driven catastrophe (as bad as or worse than human extinction) over the coming century. We also estimate that $1 billion of carefully targeted spending would suffice to avoid catastrophic outcomes in (at the very least) 1% of the scenarios where they would otherwise occur. On these estimates, $1 billion of spending would provide at least a 0.001% absolute reduction in existential risk. That would mean that every $100 spent had, on average, an impact as valuable as saving one trillion (resp., one million, 100) lives on our main (resp. low, restricted) estimate -far more than the near-future benefits of bednet distribution. \n Uncertainty and 'meta' options There is a lot of uncertainty in the numbers we have given, even in the most scientifically robust case of asteroid detection. We will give this issue a more thorough treatment in the next section, arguing against various ways in which one might worry it undermines our argument. One thing that uncertainty can support, however, is a preference for different types of strategy to improve the far future. Rather than directly trying to influence the far future, one could instead try to invest in decision-relevant research, or invest one's resources for use at a later date. The possibility of either of these strategies strengthens our argument considerably. To see this, let us suppose, for the sake of argument, that no 'first-order' intervention (such as those we discussed in sections 4.2-3) delivers higher far-future expected benefits than the highest available near-future expected benefits, relative to the credences that are appropriate in the present state of information. Suppose, however, that it is highly likely that conditional on sufficient additional information, at least one of the proposed interventions, or another such intervention (not yet considered) in a similar spirit, would have much higher far-future benefits, relative to the updated credences, than the best available near-future benefits. Then society might fund research into the cost-effectiveness of various possible attempts to influence the far future. Provided that subsequent governments or philanthropists would take due note of the results, this 'meta-option' could easily have much greater far-future expected benefits than the best available near-future expected benefits, since it could dramatically increase the expected effectiveness of future governmental and philanthropic action (all relative to currently appropriate credences) . A complementary possibility is that rather than spending now, society could save its money for a later time  (Christiano 2014; MacAskill 2019; Trammell 2020) . That is, it could set up a sovereign wealth fund, with a longtermist mission. This fund would pay out whenever there becomes available some action that will sufficiently benefit the far future (in expectation), whether that is during the lifetimes of current citizens or later. There would be some annual risk of future governments being misaligned and using the money poorly, but this risk could be mitigated via constitutional enshrinement of the mission, and would be compensated by the fact that the fund would benefit from compound returns of investment.  16  These considerations show that the bar that 'intractability' objections to our argument must meet is very high. For BR to fail to hold on such grounds, every option available to society must have negligible effect on the far future's expected value. Moreover, it must be near-certain that there will be no such actions in the future, and that no such actions could be discovered through further research. This constellation of conditions seems unlikely. We believe our arguments apply to individuals in much the same way they apply to society as a whole. Suppose Shivani is an individual philanthropist, deciding where to spend her money. Naively, we might think of Shivani as making a contribution to asteroid detection, pandemic preparedness, or AI safety that is proportional to her resources. If $1 billion can decrease the chance of an asteroid collision this century by 1 in 120,000, then $10,000 can decrease the chance of an asteroid collision by 1 in 12 billion. Because the individual's ability to contribute to short-term good would also decrease proportionally, perhaps the argument goes through in just the same way. \n Strong longtermism about individual decisions This \"naive\" argument is, in our view, approximately correct. We foresee three ways of resisting it. First, one could claim that private individuals are much more limited in their options, to such an extent that Shivani can do nothing to decrease risks from asteroids, pandemics, or AI. However, this is simply not true. Multiple organisations working on these risks, including  16  Plausibly, the gains from the investment would outweigh the risk of value-drift of the fund: the historical real rate of return on risky investments (such as stocks and housing) was around 7% during the period <PHONE>  (Jordà et al. 2019 (Jordà et al. :1228 . It seems reasonable to expect substantially lower returns in the future; but even if so, they would still be significantly higher than the risk of future governments misusing the funds; even a 90% probability of a future government misusing the funds over the next century would amount to only 2% annual risk. There is some precedent for successful long-lasting trusts in the charitable sector. In the US the John Clarke Trust was founded in 1676  (Ochs, 2019) ; in the UK, King's School, Canterbury was established in 597 (House of Commons Public Administration Select Committee, 2013). In 1790 Benjamin Franklin invested £1000 for each of the cities of Boston and Philadelphia: ¾ of the funds would be paid out after 100 years, and the remainder after 200 years. By 1990, the donation had grown to almost $5 million for Boston and $2.3 million for Philadelphia  (Isaacson 2003:473-474) . The oldest similar government funds date back to the mid-19th century: Texas's Permanent School Fund was founded in 1854 (Texas Education Agency 2020), and its Permanent University Fund was founded in 1876 (University of Texas System 2021). If the annual chance of failure of such funds were as high as 2%, then the chance of the Texas Permanent School Fund persisting until the present day would be one in thirty, and the chance of the King's School persisting until the present day would be one in ten trillion. This does not merely appear to be a selection effect: to our knowledge, it is not the case that there have been very large numbers of attempted long-lasting government funds that have failed. This suggests that 2% is a conservatively high estimate of the annual risk of failure. most of those we mentioned above, accept funding at all scales from private individuals, and would scale up their activity in response. Second, one could claim that there are increasing returns to scale, so that the impact of a small donation is much less than the relevant fraction of the impact of a large donation. This is an open possibility, but it seems significantly more likely that there are fairly strongly diminishing returns, here as elsewhere. This is for both theoretical and empirical reasons. 17 Theoretically: since interventions vary in their ex ante cost-effectiveness, a rational altruistic actor will fund the most cost-effective intervention first, before moving to the next-most cost-effective intervention, and so on. Empirically, diminishing returns have been observed across many fields (e.g.  Cassman et al. 2002:134; Arnold et al. 2018; Bloom et al. 2020 ). Third, one could claim that, once we consider the actions of individuals with smaller amounts of resources, the probability of success from directing those resources to long-term oriented interventions becomes so low that expected utility theory gives the wrong recommendations. We discuss this issue in section 8. What of individual decisions about where to direct one's labour, rather than one's money? We believe that much the same arguments apply here. Suppose that Adam is a young graduate choosing his career path. Adam can choose to train either as a development economist, or as an AI safety researcher. While there are differences between Adam's decision situation and Shivani's (MacAskill 2014), there are also important similarities. In particular, the considerations that make it better in expectation for Shivani to fund AI safety rather than developing world poverty reduction similarly seem to make it better in expectation for Adam to train as an AI safety researcher rather than as a development economist. \n Robustness of the argument In our initial presentation of the argument, we have at times assumed expected total utilitarianism, for simplicity. This raises an important question of how wide a class of axiologies will support axiological strong longtermism. First, what if instead of maximising expected total welfare, the correct axiology is risk averse? This in fact seems to strengthen the case for strong longtermism: the far-future 18 interventions we have discussed are matters of mitigating catastrophic risks, and in general terms, risk aversion strengthens the case for risk mitigation  (Mogensen, MacAskill and Greaves MS) . With only minor modifications, similar remarks apply if, instead of replacing risk neutrality with risk aversion, we replace appeals to utilitarianism in our argument with (ex post) prioritarianism. Second, if the only means of positively influencing the far future were via reducing the risk of extinction, the case for strong longtermism might rely on controversial views in population ethics, such as totalism, on which the absence of a large number of happy future beings makes things much worse. But many axiologies will not agree that premature extinction is extremely bad. In particular, person-affecting approaches to population ethics tend to resist that claim. According to the spirit of a person-affecting approach, premature extinction is in itself at worst neutral: if humanity goes prematurely extinct, then there does not exist any person who is worse off as a result of that extinction, and, according to a person-affecting principle, it follows that the resulting state of affairs is not worse. The far-future benefits of extinction risk mitigation may therefore beat the best near-future benefits only conditional on controversial population axiologies.  19  However, risks from ASI are unlike extinction in this respect: there will be a large population in the future either way, and we are simply affecting how good or bad those future lives are. The idea that it's good to improve expected future well-being conditional on the existence of a large and roughly fixed-size future population is robust to plausible variations in population-ethical assumptions. 20 Third, the example of ASI risk also ensures that our argument goes through even if, in expectation, the continuation of civilisation into the future would be bad  (Althaus and Gloor 2018; Arrhenius and Bykvist 1995: ch. 3; Benatar 2006 ). If this were true, then reducing the risk of human extinction would no longer be a good thing, in expectation. But in the AI lock-in scenarios we have considered, there will be a long-lasting civilisation either way. By working on AI safety and policy, we aim to make the trajectory of that civilisation better, whether or not it starts out already 'better than nothing'. One feature of expected utilitarianism that is near-essential to our argument is a zero rate of pure time preference. With even a modest positive rate of pure time preference (as e.g. on \"discounted utilitarian\" axiologies), the argument would not go through. Our assumption of a 20 \"Narrow\" person-affecting approaches disagree, since they regard two states of affairs as incomparable whenever those states of affairs have non-identical populations  (Heyd 1988) . However, such approaches are implausible, for precisely this reason. Similarly, theories on which any two states of affairs with non-equinumerous populations are incomparable (Bader MS) are implausible. When comparing different sized populations, a \"wide\" person-affecting approach will typically map the smaller population to a subset of the larger population, and compare well-being person-by-person according to that mapping  (Meacham 2012) ; these theories will tend to agree with total utilitarianism on the evaluation of the AI catastrophes we discuss. For similar reasons, we also do not consider here the incomparability that is introduced by a \"critical range\" view  (Blackorby, Bossert and Donaldson 1996) . zero rate, however, matches a consensus that is almost universal among moral philosophers, and also reasonably widespread among economists.  21  This is of course nowhere near an exhaustive list of possible deviations from expected total utilitarianism. We consider some other deviations below, in the course of discussing cluelessness and fanaticism. Our conclusion is that the case for strong longtermism is at least fairly robust to variations in plausible axiological assumptions; we leave the investigation of other possible variations for future research. \n Cluelessness Section 4 focussed on worries about our abilities to affect the far future. A distinct family of worries is more directly epistemic, and involves the idea that we are clueless both about what the far future will be like, and about the differences that we might be able to make to that future. Perhaps the beings that are around will be very unlike humans. Perhaps their 22 societies, if they have anything that can be called a society at all, will be organized in enormously different ways. For these and other reasons, perhaps the kinds of things that are conducive to the well-being of far-future creatures are very different from the kinds of things that are conducive to our well-being. Given all of this, can we really have any clue about the far-future value of our actions even in expectation? We take it for granted that we cannot know what the far future will be like. But, since the argument of sections 2-6 has already been conducted in terms of expected value, lack of knowledge cannot ground any objection to the argument. The objection must instead be something else. In fact, there are several quite distinct possibilities in the vicinity of the \"cluelessness\" worry. In the present section, we address five of these objections, relating to simple cluelessness, conscious unawareness, imprecision, arbitrariness, and ambiguity aversion. 22 Since \"washing-out\" concerns whether we are able to affect the far future in expectation, this too has an epistemic aspect, so that the distinction between the concerns of section 4 and those discussed here is not completely clear  (Tarsney 2019) . Nonetheless, the issues raised seem sufficiently different to warrant a separate treatment. 21 A zero rate of pure time preference is endorsed by, inter alia,  Sidgwick (1890 ), Ramsey (1928 ,  Pigou (1932) ,  Harrod (1948) ,  Solow (1974) ,  Cline (1992) ,  Cowen (1992) ,  Stern (2007) ,  Broome (2008) ,  Dasgupta (2008) ,  Dietz, Hepburn, and Stern (2008) ,  Buchholz and Schumacher (2010), and Gollier (2013) . In a recent survey of academic economists with expertise on the topic of social discounting, 38% of respondents agreed with this \"Ramsey-  Stern view\" (Drupp et al. 2018:119) .  Greaves (2017)  provides a survey of the arguments on both sides. Even among philosophers, the consensus against discounting future well-being is not universal. In particular, some plausible models of partiality suggest assigning greater effective moral weight to one's own contemporaries than to far-future people  (Setiya 2014; Mogensen 2019) . However, even these models seem unlikely to recommend sufficient discounting to undermine the argument for longtermism (Mogensen 2019: sec. 6). \n Simple cluelessness Our concern is with relatively weighty decisions, such as how to direct significant philanthropic funding. But it is illuminating to compare these to far more trivial decision situations, such as a choice of whether or where to go shopping on a given day. Even in the latter cases, many have argued, we can be all but certain that our choice will have highly significant consequences ex post -far more significant than the more predictable nearer-term effects. The reasons for this include the tendency for even trivial actions to 23 affect the identities of future persons far into the future. However, when comparing quite trivial alternatives, we can have no idea which of the two will turn out to be superior vis-à-vis these deeply unpredictable very far future effects. Some have argued that these facts undermine any attempt to base decisions on considerations of the overall good even in trivial everyday decision contexts (e.g.  Lenman 2000) . We agree with  Greaves (2016)  that this concern is overblown: in the context of relatively trivial everyday decisions, at least, the deeply unpredictable far-future effects plausibly cancel out for the purpose of comparing actions in expected value terms. Consequently, there is no objection here to basing these decisions on an expected-value assessment of nearer-future, more foreseeable effects. As we have argued in section 4, however, decisions about how to spend philanthropic funding are disanalogous in this respect. We are not discussing the possibility that either funding AI safety research or not funding it might lead, as chance has it, to the birth of an additional unusually good or bad person several centuries' hence. Rather, we are discussing the possibility that funding AI safety might have its intended effect of making AI safer. While there are certainly severe uncertainties in such work, it would be overly pessimistic to insist that success is no more likely than counterproductivity. Considerations of such 'simple' cluelessness therefore do nothing to undermine the argument for strong longtermism. \n Conscious unawareness The expected value approach we assumed in section 3 is intended as a subjective decision theory: that is, it utilizes only material that is accessible to the decision-maker at the time of decision. In particular, therefore, there is an implicit assumption that the agent herself is in a position to grasp the states, acts and consequences that are involved in modelling her decision. But perhaps this is not true. Consider, for example, would-be longtermists in the Middle Ages. It is plausible that the considerations most relevant to their decision -such as the benefits of science, and therefore the enormous value of efforts to help make the scientific and industrial revolutions happen sooner -would not have been on their radar. Rather, they might instead have backed attempts to spread Christianity, perhaps by violence: a putative route to value that, by our more enlightened lights today, looks wildly off the mark. The suggestion, then, is that our current predicament is relevantly similar to that of our medieval would-be longtermists. Perhaps there are actions available to us that would, if we were able to think it all through in full detail, then deliver high expected benefits for the far future. But we know, if only by induction from history, that we have not thought things through in all relevant detail. Perhaps we thereby have good reason to reject subjective expected-value analysis, and use some quite different form of decision analysis to assess far-future effects -in which case, all bets are as yet off regarding what the conclusion will be. This is the issue of conscious unawareness -knowing that one is unaware of many relevant considerations, mere awareness of which would influence one's decision-making. Following much of the recent literature on this topic, however, our view is that conscious unawareness does not occasion any particularly significant revision of the Bayesian framework, for three reasons. First, we know that we operate with coarse-grained models, and that the reasons for this include unawareness of some fine-grainings. Of course, failure to consider key fine-grainings might lead to different expected values and hence to different decisions, but this seems precisely analogous to the fact that failure to possess more information about which state in fact obtains similarly affects expected values (and hence decisions). Since our question is which actions are ex ante rational, both kinds of failure are beside the point. Second, we know we are likely to be omitting some important possible states of nature from our model altogether. But consciousness of this can be modelled by inclusion of a \"catchall\" state: \"all the other possibilities I haven't thought of\". Again, conceptualising parts of this state in more explicit terms might change some expected value assessments, but again this does nothing to undermine the ex ante rationality of decisions taken on the basis of one's existing assessments.  24  Third, while the best options might well be ones that have not occurred to us, that does nothing to impugn the rationality of assessments of those possible options that have occurred to us. And our argument for strong longtermism, recall, requires only a lower bound on attainable far-future expected benefits. We do not claim (nor do we believe) that issues of conscious unawareness have no effect on what the reasonable credences and values in a given decision situation are. The point is rather that these issues need not occasion any deep structural change to the analysis. Our further claim is that the numbers we have suggested in section 4 are reasonable after taking issues of conscious unawareness into account. \n Arbitrariness An obvious and potentially troubling feature of our discussion in section 4 is the paucity of objective guidance for the key values and probabilities. This seems to contrast starkly with, for instance, the usual impact evaluations for the short-term benefits of bednet distribution, which can be guided by relatively hard evidence (GiveWell 2020b). This gives rise to three distinct, though related, concerns with the standard Bayesian approach that we have used. The first is simply that the probabilities and/or values in this case are too arbitrary for our argument to carry any weight. The second is that in cases where any precise assignments would be this arbitrary, it is inappropriate to have precise credences and values at all. The third is that in such cases, the appropriate decision theory is ambiguity averse, and that this might undermine the argument for strong longtermism. We address these concerns in turn. The \"arbitrariness\" objection is that even if a rational agent must have some precise credence and value functions, there is so little by way of rational restriction on which precise functions are permissible that the argument for strong longtermism is little more than an assertion that the authors' own subjective probabilities are ones relative to which this thesis is true. We have some sympathy with this objection. However, there is a distinction between there being no watertight argument against some credence function on the one hand, and that credence function being reasonable on the other. Even in the present state of information, in our view credence-value pairs such that the argument for strong longtermism fails are unreasonable. If, for instance, one had credences such that the expected number of future people was only 10 14 , the status quo probability of catastrophe from AI was only 0.001%, and the proportion by which $1 billion of careful spending would reduce this risk was also only 0.001%, then one would judge spending on AI safety equivalent to saving only 0.001 lives per $100 -less than the near-future benefits of bednets. But this constellation of conditions seems unreasonable. However, we note that this issue is contentious. We regard the quantitative assessment of the crucial far-future-related variables as a particularly important topic for further research. \n Imprecision Imprecise approaches represent an agent by a class of pairs of probability and value functions -a representor -rather than a single such pair. The natural interpretation is that these correspond to incomplete orderings of options: one option is better than another, for instance, if and only if the first has higher expected value than the second on all probability-value pairs in the representor.  25  ASL involves comparing ex ante far-future benefits with ex ante near-future or total benefits. If imprecision is a feature of rational evaluation at all, it is plausibly a particularly prominent feature of evaluation of far-future consequences. So perhaps, for any option (including the ones we have discussed above), any reasonable representor contains at least some elements according to which the far-future benefits of this option are no higher than the near-future benefits of bednet mitigation? It is somewhat complex to say how one should evaluate ASL in the context of such imprecision. (For instance: Should we simply evaluate ASL itself relative to each element of the representor in turn, and supervaluate to arrive at an overall verdict? Or should we seek to define subsentential terms like \"near-best\" in the context of representors? If the latter, how exactly?) The general idea, though, is that one way or another, if the possibility in the last sentence of the preceding paragraph is realised, then ASL is at least not determinately true. Our reply to the imprecision critique is very similar to our reply to the arbitrariness critique. While we do not take a stand on whether or not any imprecision of valuation is either rationally permissible or rationally required (Elga 2010), we don't ourselves think that any plausible degree of imprecision in the case at hand will undermine the argument for strong longtermism. For example, we don't think any reasonable representor even contains a probability function according to which efforts to mitigate AI risk save only 0.001 lives per $100 in expectation. This does seem less clear, however, than the claim that this is not a reasonable precise credence function. \n Ambiguity aversion In employing the standard Bayesian machinery, we have been assuming ambiguity neutrality. In contrast, an ambiguity-averse decision theory favours gambles that involve more rather than less objectively specified probabilities, other things being equal (Machina and Siniscalchi 2014). Empirically, people commonly demonstrate ambiguity aversion. Suppose, for example, that one urn contains 50 red balls and 50 black balls, and a second urn contains both red and black balls in unknown proportion  (Ellsberg 1961 ). If one is ambiguity averse, one might strictly prefer to bet on the risky urn, where one knows the probability of winning, regardless of which colour one is betting on. This preference seems inconsistent with expected utility theory, but is widespread  (Trautmann and Kuilen 2015) . It might seem at first sight that ambiguity aversion would undermine the case for strong longtermism. In contemplating options like those discussed in section 4, one needs to settle one's credence that some given intervention to reduce extinction risk, or to increase the safety of ASI, would lead to a large positive payoff in the far future. But again, there seems significant arbitrariness here. In contrast, impact evaluations for the near-future benefits of bednet distribution seem to involve much more precisely bounded probabilities. Might an ambiguity-averse decision theory, then, take a substantially dimmer view of the far-future benefits of existential risk mitigation, and hence of strong longtermism? Our answer is 'no', for two reasons. First, whether or not ambiguity aversion has any prospect of undermining the argument for strong longtermism depends, in the first instance, on whether the agent in question is ambiguity averse with respect to the state of the world, or instead with respect to the difference one makes oneself to that state. The above argument-sketch implicitly assumed the latter. But, if one is going to be ambiguity averse at all, it seems more appropriate for an 26 altruist to be ambiguity averse in the former sense (MacAskill, Mogensen, Greaves and Thomas MS). And it is far from clear that actions seeking to improve the far future increase ambiguity with respect to the state of the world. It is already extremely ambiguous, for instance, how much near-term extinction risk humanity faces. We see no reason to think that 27 this latter ambiguity is increased, rather than decreasing or remaining the same, by, for example, funding pandemic preparedness.  28  Second, although it is psychologically natural, and correspondingly widespread, ambiguity aversion is anyway irrational. Here we agree with a fairly widespread consensus; we have nothing to add to the existing debate on this question.  29  We conclude that the possibility of ambiguity aversion does not undermine the argument for strong longtermism. \n Fanaticism One obvious point of contrast between the paradigm examples of ways to attain high near-future vs. far-future expected benefits is that the former tend to involve high probabilities of relatively modest benefits, whereas the latter tend to involve tiny probabilities of enormous benefits. In discussing actions aimed at mitigating extinction risk, for instance, we conceded that it is very unlikely that any such action makes any significant difference; the argument for prioritizing such actions nonetheless is characteristically that if they do make a significant difference, they might make a truly enormous one. Even among those who are sympathetic in general to expected utility theory, many balk at its apparent implications for cases of this latter type. Suppose you are choosing between a \"safe option\" of saving a thousand lives for sure and a \"risky option\" that gives a one in a trillion chance of saving a quintillion lives. The expected number of lives saved is a thousand times greater for the risky option. Unless the utility function is very non-linear as a function of lives saved, correspondingly, the expected utility of the latter option is also likely to be greater. Yet, if you choose the risky gamble, it is overwhelmingly likely that a thousand people will die, for no gain.  30  Intuitively, it seems at least permissible to save the thousand in this case. If so, this might suggest that while expected utility theory is a good approach to choice under uncertainty in more ordinary cases, it fails in cases involving extremely low probabilities of extremely large values. One might, then, seek a \"non-fanatical\" decision theory -one that does not require the agent to sacrifice arbitrarily much, with probability arbitrarily close to one, in \"fanatical\" pursuit of an extremely unlikely but enormously larger payoff. Might a non-fanatical decision theory undermine the case for strong longtermism? We regard this as one of the most plausible ways in which the argument for strong longtermism might fail. Our view is that at present, the question cannot be confidently settled, since research into the possibility of a non-fanatical decision theory is currently embryonic. However, initial results suggest that avoiding fanaticism might come at too high a price. Beckstead and Thomas (2020), for instance, consider a sequence of gambles. The first gamble delivers a large but relatively modest benefit with certainty. The last gamble delivers an enormously large benefit with extremely small probability, and zero benefit otherwise. These two gambles are linked by a sequence in which each gamble offers only a very slightly lower probability of winning than the previous gamble, and involves a much better benefit if one does win. This sequence-schema illustrates that any transitive theory that is not fanatical must instead be worryingly \"timid\": in at least one pairwise comparison of adjacent gambles, even an arbitrarily large increase in the value of a positive payoff fails to compensate for any arbitrarily small decrease in its probability. As Beckstead and Thomas go on to show, such timidity in turn leads to implausibly extreme forms of risk aversion in some cases, and to particularly implausible forms of dependence of option-assessments on assessments of causally isolated aspects of the state of affairs. A complementary reply is that in any case, the probabilities involved in the argument for longtermism might not be sufficiently extreme for any plausible degree of resistance to \"fanaticism\" to overturn the verdicts of an expected value approach, at least at the societal level. For example, it would not seem \"fanatical\" to take action to reduce a one-in-a-million risk of dying, as one incurs from cycling 35 miles or driving 500 miles (respectively, by wearing a helmet or wearing a seat belt (Department of Transport 2020)). But it seems that society can positively affect the very long-term future with probabilities well above this  30  A similar example is that of Pascal's Mugging  (Bostrom 2009) . threshold. For instance, in section 4.3, we suggested a lower bound of one in 100,000 on a plausible credence that $1 billion of carefully targeted spending would avert an existential catastrophe from artificial intelligence. Things are less clear on the individual level. If, for example, $10 billion can reduce the risk of extinction (or a comparably bad outcome) by one in 100,000, and an individual philanthropist makes a $10,000 contribution with effects proportional to that, then the philanthropist would reduce extinction risk by one in ten billion. At this level, we are unlikely to find commonplace decisions relying on that probability that we would regard as non-fanatical.  31  So, if one is inclined to take seriously the fanaticism worry, despite the problems with 'timidity', it may be that the probabilities in question are problematically small on the individual level, but not at the social level. Our inclination is to think that our intuitions on the societal level are correct, and that our intuitions around how to handle very low probabilities are unreliable. The latter has some support from the psychological literature  (Kahneman and Tversky 1979:282-83; Erev et al. 2008) . We therefore tentatively conclude that considerations of fanaticism do not undermine the argument for strong longtermism. \n Deontic strong longtermism In section 2, we distinguished between axiological and deontic versions of strong longtermism. So far, our discussion has focused exclusively on the case for the axiological claim. The deontic analog to ASL is \n Deontic strong longtermism (DSL): In the most important decision situations facing agents today, (i) One ought to choose an option that is near-best for the far future. (ii) One ought to choose an option that delivers much larger benefits in the far future than in the near future. Just as ASL concerns ex ante axiology, the 'ought' in DSL is the subjective ought: the one that is most relevant for action-guidance, and is relative to the credences that the decision-maker ought to have.  32  Without assuming consequentialism, DSL does not immediately follow from ASL. We believe, however, that our argument for ASL naturally grounds a corresponding argument for DSL. This is because of the following stakes-sensitivity argument: (P1) When the axiological stakes are very high, there are no serious side-constraints, and the personal prerogatives are comparatively minor, one ought to choose a near-best option. (P2) In the most important decision situations facing agents today, the axiological stakes are very high, there are no serious side-constraints, and the personal prerogatives are comparatively minor. (C) So, in the most important decision situations facing agents today, one ought to choose a near-best option. DSL follows from the conjunction of (C) and ASL. The stakes-sensitivity argument is obviously valid. Are its premises true? (P1) appeals to only a very moderate form of stakes-sensitive non-consequentialism. It allows that there may be some actions that are always permissible or prohibited, no matter how great the axiological stakes: for example, perhaps one is always permitted to save the life of one's child; or perhaps one is always prohibited from torturing another person. And it only entails that comparatively minor prerogatives are overridden when the stakes are very high.  33  It is highly plausible that there should be at least this much stakes-sensitivity. The lack of stakes-sensitivity is a common objection to Kant's notorious view that even if a friend's life depends on it, one should not tell a lie  (Kant 1996) . Turning to prerogatives, in \"emergency situations\" like wartime, ordinary prerogatives -for instance, to consume luxuries, to live with one's family, and even to avoid significant risks to one's life -are quite plausibly overridden.  Nagel (1978)  observes that public morality tends to be more consequentialist in character than private morality; one natural partial explanation for this (though not the one emphasised by Nagel himself) is that in public contexts such as governmental policy decisions, the axiological stakes tend to be higher. We foresee two lines of resistance to (P1). First, one could reject the idea of \"the good\" altogether  (Thomson 2008: sec. 1.4) . On this view, there is simply no such thing as axiology. It's clear that our argument as stated would not be relevant to those who hold such views. But such a view must still be able to explain the fact that, in cases where there is a huge amount at stake, comparatively minor constraints and prerogatives get overridden. It seems likely that any such explanation will result in similar conclusions to those we have drawn, via similar arguments. Secondly, and more plausibly, perhaps only some sorts of axiological considerations are relevant to determining what we ought to do. We consider two ways in which this idea might undermine our argument. First, on a non-aggregationist view, comparatively small ex ante benefits to individuals are not relevant to determining what one ought to do, even if the benefits apply to an enormous number of people  (Scanlon 1998:235 ; Frick 2015 ; Voorhoeve 2014) . Second, perhaps axiological considerations cannot outweigh non-consequentialist considerations when the axiological considerations involve altering the identities of who comes into existence  (Parfit 1984: ch. 16 ). However, both lines of thought risk proving too much. Let's first consider the non-aggregationist response. Consider a Briton, during WWII, deciding whether to fight; or someone debating whether to vote in their country's general election; or someone deciding whether to join an important political protest; or someone deciding whether to reduce their carbon footprint. In each case, the ex ante benefits to any particular other person are tiny. But in at least some such cases, it's clear that the agent is required to undertake the relevant action, and the most natural explanation of why is because the axiological stakes are so high. \n 34 Second, consider the non-identity response. It's clear that governments ought to take significant action to fight climate change. But almost all of the expected damages from climate change come from its impacts on those who are yet to be born. What's more, any 35 policy designed to mitigate climate change will also affect the identities of those unborn people. Endorsing the non-identity response would therefore risk rejecting the idea that welfarist considerations generate any obligations for society today to fight climate change, even while accepting that climate change will significantly and avoidably reduce welfare in expectation for centuries to come. That position is clearly incorrect. Turning now to (P2): The 'high-stakes' aspect of this premise is justified in part on the basis of the arguments of sections 3-4. At least on our main and low estimates of the expected size of the future, in the decision situations we've discussed, not only are the best options those that have the near-best far-future consequences, but they are much better than those options whose far-future consequences are nowhere near best. At the same time, at least for most members of rich countries, the decision situations we've discussed are those where the personal prerogatives are arguably comparatively minor, and where there are no serious side constraints. This is clearest in the cases of individual decisions about where to direct one's altruistic spending (holding fixed the total size of one's \"altruistic budget\"), and about career choice. The decision to give to organisations that will positively influence the far future rather than organisations more geared towards improving the near future, or to work in a career that is particularly beneficial for the long-term future, might well involve some sacrifices. But they are not close to the sorts of sacrifices where 36 there might be absolute or near-absolute prerogatives. Similarly, these are not circumstances where one is required to violate side-constraints in order to achieve the near-best long-term outcome. The slightly less clear cases are those involving individual decisions about the total size of one's \"altruistic budget\" (vs. \"personal budget\"), and societal decisions about how many resources to devote to improving the prospects for the far future (vs. the near future, including the lifetimes of present people). Here, it remains true that no serious side-constraints need be involved. One might worry, though, that here our argument will be too demanding: might it imply that we, individually or as a society, ought to devote most of our resources to improving the far future, at the large expense of our own prudential interests? As in the discussion of demandingness in the context of global poverty, a range of responses to this concern is possible. We have nothing to add to the existing literature on demandingness (e.g.  Kagan 1984; Mulgan 2001; Hooker 2009 ). We will simply note that even if, for example, there is an absolute cap on the total sacrifice that can be morally required, it seems implausible that society today is currently anywhere near that cap. The same remark applies to at least the vast majority of individuals in rich countries. We ought to be doing a lot more for the far future than we currently are. 38 \n Summary and conclusions The potential future of civilisation is vast. Once we appreciate this, it becomes plausible that impact on the far future is the most important feature of our actions today. Strong longtermism would be false in a world that had sufficiently weak causal connections between the near and the distant future, such that it was too difficult to significantly influence 38 Might our arguments go further than this, and justify atrocities in the name of the long-term good? Perhaps the French Revolution had good long-term consequences, in terms of bringing about a more liberal and democratic world: does strong longtermism, if so, justify the guillotine? We do not think so, for at least two reasons. The first is that, for such serious side-constraints, something closer to absolutism or near-absolutism becomes much more plausible (or, at least, it takes more than mere ex ante goodness to justify violation of those side-constraints). The second is that, in almost all cases, when there is some option available that promotes the long-term good while violating a serious side-constraint, there will be some alternative option available that achieves a similar amount of long-term good without violating that side-constraint. Liberal democracy could have been achieved in France without the Reign of Terror. 37  Mogensen (2020)  discusses specifically the relationship between demandingness and longtermism. 36 There are, however, reasons to think that these sacrifices are not as great as we might initially suppose (MacAskill,  Mogensen and Ord 2018) . the course of the very long-run future. However, we have argued, the world we find ourselves in today does not have this feature. We presented our central case in terms of (i) a total utilitarian axiology and (ii) an expected value treatment of decision-making under uncertainty. However, we argued, plausible deviations from either or both of these assumptions do not undermine the core argument. This paper mainly focussed on the decision situations of a society or individual considering how to spend money without constraints as to cause area, and of an individual's career choice. We argued that these are situations where we can in expectation significantly influence the far future. Precisely because of this, they are among the most important decision situations we face, and axiological strong longtermism follows. In our own view, the weakest points in the case for axiological strong longtermism are the assessment of numbers for the cost-effectiveness of particular attempts to benefit the far future, the appropriate treatment of cluelessness, and the question of whether an expected value approach to uncertainty is too \"fanatical\" in this context. These issues in particular would benefit from further research. In addition to axiological issues, we also discussed the counterpart deontic issues. We suggested that deontic strong longtermism might well be true even if consequentialism is false, on the grounds that (i) the stakes involved are very high, (ii) a plausible non-consequentialist theory has to be sensitive to the axiological stakes, becoming more consequentialist in output as the axiological stakes get higher, and (iii) in the key decision situations, any countervailing constraints and/or prerogatives are comparatively minor. Quite plausibly, in the world as it is today, the most important determinants of what we ought to do arise from our opportunities to affect the far future. It is possible, but far from obvious, that far-future impacts are also more important than near-future impacts in a much wider class of decision situations: for instance, decisions about whether or not to have a child, and government policy decisions within a relatively narrow 'cause area'. Insofar as they are, strong longtermism could potentially set a methodology for further work in applied ethics and applied political philosophy: for each issue in these subfields, one could identify the potential far-future effects from different actions or policies, and then work through how these bear on the issue in question. The answers might sometimes be surprisingly revisionary. \n Terminology and notation For any option , let , , respectively denote 's near-future, far-future and 𝑥 𝑁(𝑥) 𝐹(𝑥) 𝑉(𝑥) 𝑥 overall benefits. Let , , respectively be the highest available near-future, far-future and 𝑁 * 𝐹 * 𝑉 * overall benefits. Let be the highest far-future benefit that is available without net 𝐹' short-term harm. We interpret both \"near-best overall\" and \"near-best for the far future\" in terms of proportional distance from zero benefit to the maximum available benefit, and \"much larger\" in terms of a multiplicative factor. There is, of course, flexibility on the precise values of the factors involved. We therefore consider the following precisifications of our key claims, carrying free parameters: BR(n): . 𝐹' ≥ 𝑛𝑁 * : Every option that delivers overall benefits of at least 𝐴𝑆𝐿 In what follows, we prove claims (a) and (b) for specified relationships between the parameter values. Precisification of claim (a). We claim (more precisely) that if BR(n) holds of a given decision situation, then for any , holds of the restricted decision ϵ 𝑂 ∈ [0, 1] 𝐴𝑆𝐿(ϵ 𝑂 , ϵ 𝑂 + 1 𝑛 ) situation (with any options involving net short-term harm removed). For example, if , 𝑛 = 10 then every option that delivers at least 90% of available overall expected benefits delivers at least 80% of available far-future expected benefits, once any options involving net short-term harm are ruled out. Proof. Suppose that BR(n) holds. Since far-future benefit is attainable without near-future 𝐹' net harm, the overall best option must deliver total benefits of at least ; so any near-best 𝐹' option must deliver total benefits of at least . But by BR(n), the maximum (1 − ϵ 𝑂 )𝐹' attainable near-future benefit is at most . Therefore, any near-best option must deliver  \n Precisification of claim (b). We claim (more precisely) that if BR(n) holds then for any , ASL(ii) also holds. For example, if , then every ϵ 𝑂 ∈ [0, 1] (ϵ 𝑂 , (1 − ϵ 𝑂 )𝑛 − 1) 𝑛 = 10 option that delivers at least 90% of available overall expected benefits delivers at least 8 times as much far-future as near-future expected benefit. Proof. Let be any option that is near-best overall. Then 𝑥  benefits that are at least r times its own near-future benefits. \n -future net harm is here ruled out). \n \n Scenario Duration (centuries) Carrying capacity (lives per century) Number of future lives Earth (mammalian reference class) 10 4 10 10 10 14 Earth (digital life) 10 4 10 14 10 18 Solar System 10 8 10 19 10 27 Solar System (digital 10 7 10 23 10 30 life) Milky Way 10 11 10 25 10 36 Milky Way (digital 10 11 10 34 10 45 life) \n So far we have discussed what is best for a society to do, sometimes referring to what billions of dollars would be able to achieve. But what about individuals? \n\t\t\t We discuss deontic strong longtermism in section 9. \n\t\t\t It would amount to a \"surprising and suspicious convergence\" between near-future and far-future optimisation).3 A sister organisation to GiveWell, Open Philanthropy, has tried hard to find human-centric interventions that have more short-term impact, and has struggled (Berger 2019 ). There might be more cost-effective interventions focused on preventing the suffering of animals living in factory farms (Bollard 2016) . We leave this aside in order to avoid getting into issues of inter-species comparisons; again, there is a corresponding need for sensitivity analysis.2 Following GiveWell (2018b), we will assume that the short-term benefits of the interventions that do the most short-run good would scale proportionately even if very large amounts of money were spent. \n\t\t\t We will use 'human' to refer both to Homo sapiens and to whatever descendants with at least comparable moral status we may have, even if those descendants are a different species, and even if they are non-biological. \n\t\t\t On duration: technological progress brings not only protection against existing extinction risks, but also novel sources of extinction risk(Ord 2020: esp. chs. 4 and 5). On population size: the tendency for richer societies to have lower fertility rates has led some to conjecture that human population, after plateauing around 2100, might significantly decline into the indefinite future, a high \"carrying capacity\" notwithstanding (Bricker and Ibbitson 2019) . \n\t\t\t We return to the likelihood of artificial superintelligence in section 4.3. \n\t\t\t It is important here to distinguish between ex ante and ex post versions of the washing-out claim. The ex post version is false, as is established by the literature on cluelessness; cf. section 7.1. However, it is the much more plausible ex ante washing-out claim that is relevant to the arguments of this paper. \n\t\t\t risk by 1%, and assume that the risk reduction occurs throughout the next 100 years but only in that time period, then each $100 of such spending would, in expectation, increase the number of future beings by 200 million (respectively, 200, 0.02) on our main (resp., low, restricted) estimate. According to these calculations, the far-future benefits would thereby significantly exceed the near-future benefits of bednet distribution on our main and low estimates of the size of the future, though not on our restricted estimate. Organisations11  We use the geometric rather than the arithmetic mean because the estimates in question are spread across several orders of magnitude; the arithmetic mean effectively defers to the highest estimate on the question of order of magnitude. Using the arithmetic mean would lead to results that are still more favourable to strong longtermism. Similarly, we disregard Millet and Snyder-Beattie's \"Model 1\" because, as the authors note, this model is flawed in important respects; including this model would also strengthen the case for strong longtermism.10  Two ways in which Millet and Snyder-Beattie's estimate is particularly conservative are (i) that the $250bn figure is at the extreme upper end of anticipated costs for the intervention they discuss, and (ii) that the intervention in question concerns an extremely broad-based approach to biosecurity, not specifically optimising for extinction risk reduction. \n\t\t\t Those concerned include leading machine learning researchers such as Stuart Russell (2019)  and ShaneLegg  (2008: sec. 7.3), philosophers such as Nick Bostrom (2014) , Eliezer Yudkowsky (2013) ,Toby Ord  (2020:138-152)  and Richard Ngo (2020), physicists such as Max Tegmark (2017: ch. 4) and Stephen Hawking (2018: ch. 9), and tech entrepreneurs such as Elon Musk (2014 ), Sam Altman (2015 and Bill Gates (Statt 2015) .12 Other areas one might consider here include affecting the values that the world converges on (Reese 2018) , or reducing the risk of a totalitarian world government (Caplan 2008) . \n\t\t\t Grace et al. (2018)  asked 352 leading AI researchers to give a probability on the size of existential risk arising from the development of 'human-level machine intelligence'; the median estimate was 5%. A survey among participants at a conference on global catastrophic risks similarly found the median estimate to be 5%. One would expect a selection effect to be at work in surveys of those who have chosen to work on existential risk, but not so (or not strongly) for the survey of AI researchers.14  Neglectedness is crucial to the argument of this paper. Would strong longtermism still be true if, for example, 10% of global GDP were already spent on the most valuable long-term-oriented interventions? Even if true, would it still be significantly revisionary compared to a near-termist approach, as we have claimed it is at the current margin? We aren't sure. Our claim here is only that the world today is clearly far below this optimum. \n\t\t\t On the standard account, to be risk averse is to have utility be a concave function of total welfare(Pratt  1964:127; O'Donoghue and Somerville 2018:93). Some have argued that the standard account is inadequate (Rabin 2000;  Buchak 2013:30). On risk-weighted expected utility theory, risk aversion is represented by a risk function that transforms the expected utility function (Quiggin 1982; Quiggin and Wakker 1994; Buchak 2013 ). The differences between these accounts are unimportant for our purposes. 17 Relatedly, it seems that insofar as scale does make a difference, ASL(i) and (ii) are more likely to be true of decision situations involving smaller sums of money, not less likely. Increasing-returns phenomena are discussed by Pierson (2000) . \n\t\t\t It is not immediately clear precisely what a person-affecting approach will say about the value of extinction risk mitigation, since the usual formulations of those theories do not specify how the theories deal with risk, and it is not immediately clear how to extend them to cases that do involve risk. Thomas (2019) explores a number of possibilities. \n\t\t\t See e.g. Lenman (2000) , Greaves (2016) . We agree with this claim, but our argument does not rely on it. \n\t\t\t The first type of unawareness is unawareness of possible refinements, the second is unawareness of possible expansions (Bradley 2017: sec. 12.3; Stefánsson and Steele forthcoming: sec. 3.2). \n\t\t\t Bewley (2002) , Dubra, Maccheroni, and Ok (2004) , and Galaabaatar and Karni (2013) provide representation theorems linking such representations to incomplete orderings. \n\t\t\t See e.g. (Al-Najjar & Weinstein 2009)  for a survey of arguments that ambiguity aversion is irrational. Rowe & Voorhoeve (2018)  and Stefánsson & Bradley (2019)  defend its rationality.28  We investigate the issues outlined in this paragraph in more depth in Mogensen, MacAskill and Greaves (MS).27Beard et al. (2020, Appendix A)  and both present a wide range of estimates from around 1% to 50%, from (respectively) a literature review and a conference participant survey.26  To see the distinction in Ellsberg's two-urns setting, suppose that in the status quo, one is set to receive $100 iff the ambiguous urn delivers a red ball. Suppose one's choice is between whether to add to that background gamble a bet on a black ball being drawn from the risky urn, or instead from the ambiguous urn. Pretty clearly, ambiguity aversion in the standard sense will recommend the latter (since one then faces zero ambiguity overall), notwithstanding the fact that the benefit delivered by one's action is more ambiguous in this case. \n\t\t\t It is widely agreed that either it is useful to distinguish between objective and subjective senses of 'ought'(Ewing 1948:118-22; Brandt 1959:360-7; Russell 1966;  Parfit 1984:25; Portmore 2011; Dorsey 2012;  Olsen  2017; Gibbard 2005; Parfit 2011 ), or 'ought' is univocal and subjective (Prichard 1932;  Ross 1939:139; Howard-Snyder 2005; Zimmerman 2006; Zimmerman 2008; Mason 2013) . Our discussion presupposes that one31  One exception might be putting on a seatbelt for a one-mile drive. If doing so decreases one's chance of a fatal accident by a factor of one-third, then the seatbelt reduces one's risk of death by about one in a billion. But perhaps this is not our reason for wearing seatbelts for short journeys. \n\t\t\t (P1) is very similar to Singer's claim that \"If it is in our power to prevent something very bad from happening, without thereby sacrificing anything morally significant, we ought, morally, to do it\"(Singer 1972:231). of these disjuncts is correct. A minority of authors holds that 'ought' is univocal and objective(Moore  1903:199-200, 229-30; Ross 1930:32; Thomson 1986:177-79; Graham 2010; Bykvist 2011) . Similarly (but less discussed), one might be skeptical of the notion of ex ante axiology; again, our discussion of ASL has presupposed that any such skepticism is misguided. \n\t\t\t For example, the Stern Review predicts the vast majority of damages to occur after 2100 in both \"baseline\" and \"high climate\" scenarios(Stern 2007:178, fig. 6.5d).34  None of these examples, however, involves foregoing an opportunity to save many lives of identified people. In this respect, our examples are perhaps relevantly dissimilar to a decision between spending to benefit the far vs. the near future. We thank an anonymous referee for pressing this reply.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/The-Case-for-Strong-Longtermism-GPI-Working-Paper-June-2021-2-2.tei.xml", "id": "99cdbc722494c940d5e382f6b378a21f"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Expectations around future capabilities of lethal autonomous weapons systems (LAWS) have raised concerns for military risks, ethics, and accountability. The U.K.'s position, as presented among various international voices at the UN's Convention on Certain Conventional Weapons (CCW) meetings, has attempted to address these concerns through a focused look at the weapons review process, humanmachine teaming or \"meaningful human control\" (see e.g. JCN1/18), and the ability of autonomous systems to adhere to the Rules of Engagement. Further, the U.K. has stated that the existing governance structures-both domestic and international-around weapons systems are sufficient in dealing with any concerns around the development, deployment, and accountability for emerging LAWS; there is no need for novel agreements on the control of these weapons systems. In an effort to better understand and test the U.K. position on LAWS, the Centre for the Study of Existential Risk has run a research project in which we interviewed experts in multiple relevant organisations, structured around a mock parliamentary inquiry of a hypothetical LAWS-related civilian death. The responses to this scenario have highlighted different, sometimes complementary and sometimes contradicting, conceptions of future systems, challenges, and accountability measures. They have provided rich \"on the ground\" perspectives, while also highlighting key gaps that should be addressed by every military that is considering acquisition and deployment of autonomous and semi-autonomous weapon systems.", "authors": ["Amritha Jayanti", "Shahar Avin"], "title": "It Takes a Village: The Shared Responsibility of 'Raising' an Autonomous Weapon", "text": "Introduction With the increasing integration of digital capabilities in military technologies, many spheres of the public--from academics to policymakers to legal experts to nonprofit organizations--have voiced concerns about the governance of more \"autonomous\" weapons systems. The question of whether autonomous weapons systems pose novel risks to the integrity of governance, especially as it depends so heavily on the concept of human control, responsibility, and accountability, has become central to the conversations. The United Kingdom (U.K.) has posited that lethal autonomous weapons (LAWS), in their current and foreseeable form, do not introduce weaknesses in governance; existing governance and accountability systems are sufficient to manage the research, development, and deployment of such systems and the most important thing we can do is focus on improving our human-machine teaming. Our research project seeks to test this theory by asking: With the introduction of increasingly autonomous agents in war (lethal autonomous weapons/LAWS), are the current governance structures  (legal, organizational, social)  in fact sufficient for retaining appropriate governance and accountability in the U.K. MoD? By attempting to confront strengths and weaknesses of existing governance systems as they apply to LAWS through a mock parliamentary inquiry, the project uncovers opportunities for governance improvements within Western military systems, such as the U.K. \n Background Computers and algorithms are playing a larger and larger role in modern warfare. Starting around 2007 with writings by Noel Sharkey, a roboticist who heavily discusses the reality of robot war, members of the research community have argued that the transition in military technology research, development, and acquisition to more autonomous systems has significant, yet largely ignored, moral implications for how effectively states can implement the laws of war. 1 Segments of this community are concerned with the ethics of decision making by autonomous systems, while other segments believe the key concern is regarding accountability: how responsibility for mistakes is to be allocated and punished. Other concerns raised in this context, e.g. the effects of autonomous weapon systems on the likelihood of war, proliferation to non-state actors, and strategic stability, are beyond the scope of this brief, though they also merit attention. \n U.K. Position on LAWS The United Kingdom's representatives at the UN Group of Governmental Experts (GGE) on Lethal Autonomous Weapon Systems (LAWS) have stated that the U.K. believes the discussions should \"continue to focus on the need for human control over weapon systems and that the GGE should seek agreement on what elements of control over weapon systems should be retained by humans.\" 2 The U.K., along with other actors, such as the United States, believe that a full ban on LAWS could be counterproductive, and that there are existing governance structures in place to provide appropriate oversight over the research, development, and deployment of automated weapons systems: …[T]he U.K. already operates a robust framework for ensuring that any new weapon or weapon system can be used legally under IHL. New weapons and weapons systems are conceived and created to fulfil a specific requirement and are tested for compliance with international law obligations at several stages of development.  3  The U.K. is also interested in a \"technology-agnostic\" focus on human control because it believes that it will \"enable particular attention to be paid to the key elements influencing legal, ethical and technical considerations of LAWS, as opposed to \"debated definitions and characteristics\" which, ultimately, may \"never reach consensus.\" The position emphasizes that taking a \"human-centric, through-life\" approach would enable human control to be considered at various stages and from multiple perspectives. This includes across all Defense Lines of Development, the acquisition of weapons systems, and their deployment and operation. It is the U.K.'s position that the existing institutional infrastructure buildsin accountability measures throughout the weapon system lifecycle. \n Methodology In order to stress test the governance and accountability structures that exist for U.K. weapon systems, and how they would apply to LAWS, we developed a hypothetical future scenario in which a U.K. LAWS kills a civilian during an extraction mission in Egypt. In order to ensure a level of feasibility and accuracy of construction, the scenario was built based on a wargaming scenario publicly published by RAND.  4  We then ran a facilitated role-play exercise based on our modified scenario with an initial group of Cambridge-based experts. With their feedback and the lessons from the role-play, we developed the final version of the scenario which we then used in the research study (see Appendix). This final iteration of the LAWS scenario was used to run a mock U.K. parliamentary inquiry through which we interviewed 18 experts across various areas of expertise, including (but not limited to) U.K. military strategy, military procurement, weapons development, international humanitarian law, domestic military law, military ethics, and robotics. The interviews ranged from 45 to 120 minutes and explored a variety of questions regarding the case. The main objective of the interviews was to catalyze a meaningful discussion around what information the experts deemed important and necessary in order to decide who should be held accountable in the aftermath of this scenario. A sample of the questions asked include: • Who holds the burden of accountability and responsibility? • What explanations and justifications for actions are needed? • What information is necessary to come to a conclusion about the burden of accountability? • Are there any foreseeable gaps because of the autonomy of the weapons systems? The responses and dialogue of these 18 interviews were then reviewed and synthesized in order to develop a landscape of strengths and weaknesses of the current governance and accountability schemes related to U.K. institutions as they relate to LAWS, as well as recommendations on addressing any identified weaknesses. The full report is under preparation, but we are happy to share our preliminary key findings and recommendations below. \n Key Findings The main takeaway from the \"inquiry,\" from both a legal and organizational standpoint, was that assessing accountability is in the details. This contrasts with what we perceive as a dominant narrative of \"meaningful human control,\" which focuses mainly on human control, and the design of that interaction, at the point of final targeting action. The disconnect between the accountability across a weapon's lifetime and the focus on final targeting decision was observed throughout the various expert interviews. \"Meaningful human control\" has become the idée fixe of domestic and international conversations for regulation of LAWS but it disadvantageously provides a limited lens through which most experts and relevant personnel think about accountability. To contrast this heavily focused narrative, the interviews have highlighted a whole range of intervention points, where humans are expected to, and should be supported in making decisions that enable legal, safe, and ethical weapon systems. These are arguably points that should be considered in \"meaningful human control.\" These include, but are not limited to: •  \n Recommendations \n Dialogue Shift: Emphasizing Control Chain and Shared Responsibility The prioritization of \"meaningful human control\" for LAWS-related risk mitigation and governance anchors the scope of control points around final targeting decisions. The narrative implies that this is the main area of control that we want to manage, focus, and improve on in order to ensure that the weapons systems we are deploying are still acting with the intent and direction of human operators. Although this is an important component of ensuring thoughtful and safe autonomous weapons systems, this is only a fraction of the scope of control points. In order for us to acknowledge the other points of control throughout the research, development, procurement, and deployment of LAWS, we need to be inclusive in our dialogue about these other points of human control. \n Distribution of Knowledge: Personnel Training Training everyone who touches the research, development, deployment, etc. of LAWS on international humanitarian law, robot ethics, legality of development, responsibility schemes, and more, would contribute to a more holistic approach to responsibility and accountability, and, at its best, can contribute to a culture that actively seeks to minimise and eliminate responsibility gaps through a collaborative governance system. 5 This distribution of understanding around governance could provide a better landscape for accountability through heightened understanding of how to contextualize technical decisions. Further, it can provide an effective, granular method for protecting against various levels of procedural deterioration. With shifting geopolitical pressures, as well as various financial incentives, there could easily be a deterioration of standards and best practices. A collaborative governance scheme that is based on a distributed understanding of standards, military scope, international norms, and more, can provide components of a meaningful and robust governance plan for LAWS. This distribution of knowledge, though, must be coupled with techniques for reporting and transparency of procedure to be effective. \n Acknowledging the Politics of Technical Decision Making/Design Specifications \"Meaningful human control,\" through its dialogue anchoring, also puts a heavy burden on the technical components of design decisions, such as best practices for human-computer interactions. The politics of quantification in technical decision systems for autonomous systems should not be undervalued. The way any autonomous system decides what actions to take and what information to show is a highly political decision, especially in the context of war. It is important to understand which parts of the design process are more political than they are technical, who should be involved in those decisions, and how to account for those decisions in the scope of research and development (to inform a proper, comprehensive collective responsibility scheme). \n Task Execution In order to execute this mission, U.K. military leaders decide that deploying AI combatants for the rescue mission is the best option. Given the large success of Levia Watch and Strike, the MoD has acquired a new version of this ground unmanned vehicle: the Levia Rescue (Levia R). Levia R is a multifunctional tracked robotic system that can be armed with AR15s, such as M4a1, CQBR, and Colt Commando guns to be used for defense (these are not offensive strikers). The unmanned system is a 7.5ft-wide, 4ft-tall tank-like robot with speed reaching 20mph. The tank is able to carry various heavy payloads, including up to two adult humans. Making technical progress from the generation 1 Levia Sentry and Strike, there no longer is a human in the loop for the Levia R; human military personnel are on the loop and so, once the AI is deployed, humans have limited capabilities in controlling its actions. During the rescue mission, a Levia R agent misassesses a human-robot interaction as aggressive. The human action is perceived as threatening and so the agent uses its defense technology to shoot. The human is killed. It is soon discovered the human is a non-combatant --a civilian who was not engaging in any threatening acts towards the unmanned system (the Levia R's 'body' camera footage was reviewed by human military personnel and the act of the civilian was deemed to be non threatening). There was no external intervention with the AI system given that humans are sitting on the loop. The misassessment is being viewed as a system failure, though the reasons of why the system failed are uncertain. \n Public Knowledge During the altercation, another civilian recorded the interaction on their mobile device and has posted it on social media. The video is shaky and the altercation goes in and out of frame and so the exact actions of the human and robot are slightly uncertain. The final shot is captured. \n Assumptions 1. There is no AGI (narrow AI only --applications for narrowly defined tasks). 2. There are no existing international or domestic bans on LAWS. 3. The movement for private companies to refuse engagement with LAWS has failed to spread effectively -companies like Amazon, Microsoft, etc are still engaged. 4. The state of technology (SoT) has passed domestic technology readiness standards. 5. There has been relatively widespread integration of LAWS in various missions (targeting, search and rescue, and more). Establishment of military need: defining military necessity for research, development, and/or procurement; choice of technological approach based on political and strategic motivations. (Main related stakeholders: U.K. MoD; U.K. Defense Equipment and Support (DE&S); Private military contracting companies, such as BAE Systems, Qinetiq, General Dynamics.) • \n Technical capabilities and design: trade -offs between general applicability and tailored, specific solutions with high efficacy and guarantees on performance; awareness, training, and foreseeability of contextual factors about intended use situations that may affect the performance of the weapon system; documentation and communication of known limitations and failure modes of the system design. (Main related stakeholders: Weapons deployment: informing commanders about the capabilities and limitations of the system, of their track record in similar situations, of novel parameters of the new situation; establishing and training for appropriate pre-deployment testing schemes to capture any vulnerabilities or \"bugs\" any specific weapons system; checking for readiness of troops to operate and maintain systems in the arena; expected response of non-combatants to the presence of the weapon system. (Main related stakeholders: U.K. MoD commanding officers; U.K. MoD military personnel --human operators)• Weapons engagement: awareness of limiting contextual factors, need to maintain operator awareness and contextual knowledge; handover of control between operators during an operation. related stakeholders: Private military contracting companies, such as BAE Systems, Qinetiq, General Dynamics; U.K. DE&S; U.K. MoD military personnel -human operators) • Procurement: robust Article 36 review; assessment of operational gaps, and trading-off operational capability with risks; trade-off between cost effectiveness and performance of weapons systems; documentation and communication of trade-offs so they can be re-evaluated as context or technology changes; number and type of systems; provisioning of training and guidance; provisioning for maintenance. (Main related stakeholders: U.K. DE&S; Article 36 convened expert assessment group; Private military contracting companies, such as BAE Systems, Qinetiq, General Dynamics) • Private military contracting companies, such as BAE Systems, Qinetiq, General Dynamics; U.K. Defense Science and Technology, U.K. Defense and Security Analysis Division) • Human-computer interaction design: choices of what data to include and what data to exclude; trade-offs between clarity and comprehensiveness; level of technical information communicated; parallel communication channels: to operator in/on the loop, to command centres further from the field, logs for future technical analysis or legal investigation. (Main related stakeholders: Private military contracting companies, such as BAE Systems, Qinetiq, General Dynamics; U.K. Defense Science and Technology; U.K. Defense and Security Analysis Division; U.K. MoD military personnel -human operators) • Weapons testing: choice of parameters to be evaluated, frequency of evaluation, conditions under which to evaluate; simulation of adversaries and unexpected situations in the evaluation phase; evaluation of HCI in extreme conditions; evaluation of the human-machine team. (Main (Main related stakeholders: U.K. MoD military personnel --human operators) • Performance feedback: ensuring a meaningful feedback process to guarantee process improvement, reporting of faulty actions, communicating sub-par human-machine techniques and capabilities, and more. (Main related stakeholders: U.K. MoD military personnel --human operators; U.K. MoD commanding officers; U.K. DE&S; Private military contracting companies, such as BAE Systems, Qinetiq, General Dynamics) \n\t\t\t  Carpenter, C. (2014). From \"Stop the Robot Wars!\" to \"Ban Killer Robots.\" Lost Causes, 88-121. doi: 10.7591/cornell/9780801448850.<PHONE> 2 Human Machine Touchpoints: The United Kingdom's perspective on human control over weapon development and targeting cycles, Human Machine Touchpoints: The United Kingdom's perspective on human control over weapon development and targeting cycles (2018). 3 Ibid. 4 Khalilzad, Z., & Lesser, I. O. (1998). Selected Scenarios from Sources of Conflict in the 21st Century (pp. 317-318). RAND. \n\t\t\t  Ansell, C. (2012). Collaborative Governance. Retrieved from https://oxfordindex.oup.com/view/10.1093/oxfordhb/9780199560530.<PHONE>.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/It_Takes_a_Village__The_Shared_Responsibility_of_Raising_an_Autonomous_Weapon.tei.xml", "id": "6074cb9d94b612bb593a2215796cfcdb"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "is a research organization focused on studying the security impacts of emerging technologies, supporting academic work in security and technology studies, and delivering nonpartisan analysis to the policy community. CSET aims to prepare a generation of policymakers, analysts, and diplomats to address the challenges and opportunities of emerging technologies. CSET focuses on the effects of progress in artifi cial intelligence, advanced computing, and biotechnology.", "authors": ["Matthew Daniels", "Ben Chang", "Igor Mikolic-Torreira", "James Baker", "Jack Clark", "Remco Zwetsloot", "Teddy Collins", "Helen Toner", "Jack Corrigan", "Jeff Alstott", "Maura Mccarthy", "Alex Friedland", "Lynne Weil", "David Lasker", "Jared Dunnmon", "Matt Mahoney", "Greg Allen", "Fellow. Andrew W Marshall"], "title": "National Power After AI", "text": "ing access to certain resources and the industrial capacity to leverage them. These broader effects take more time to appear, but their impact can be enormous: industrialization was not \"revolutionary\" because of the rapidity of change, as it unfolded in waves over decades, but because of its ultimate magnitude of change. With AI technologies, progressive substitution of machines for human cognitive labor may eventually have economic and social implications on a scale comparable to the Industrial Revolution. And like the Industrial Revolution, this AI revolution will change some fundamental elements of national power. Of course, these foundational shifts can render some of the current processes and resources of a state obsolete, but they can also make what states are already doing, or already possess, more valuable. For example, the invention of railroads was a boon for those rich in steel precursors.  2  With AI, data-hungry algorithms may advantage authoritarian states, which already surveil and catalogue their own populations with little regard for human rights.  3  We suggest an \"evolutionary\" view of technological change: major, widely diffused innovations are akin to environmental shifts, affecting the competitive capacity of states based on their existing trends in population, resources, institutions, character, and policies. Some previously \"maladaptive\" factors may become advantageous, and vice versa; states will adapt their institutions, organizations, and policies to the new environment in different ways and to varying degrees, and consequently gain or lose relative power as a result. Nations that primarily focus on AI technologies as offering marginal improvements in existing capabilities (\"helping to build better mousetraps\") will eventually miss larger opportunities to adapt. This paper is a first step into thinking more expansively about AI and national power. In what follows, we first explain this evolutionary view in greater detail before applying it to AI. Ultimately, we seek pragmatic insights for long-term U.S. competition with authoritarian governments like that of China. For the foreseeable future, China's population and total economic size will very likely exceed those of the United States, even as its per capita GDP lags. This new challenge differs fundamentally from the United States' Cold War competition with the Soviet Union, and success will require thoughtful and timely diagnosis of modern environmental shifts in how states can produce power. These insights can guide our own investments as well as our approach to alliances. The United States has many sources of advantage and strength, and as Joseph Nye rightly observed, \"Our greatest mistake in such a world would be to fall into one-dimensional analysis and to believe that investing in military power alone will ensure our strength.\"  4  This paper is a first step, intended to provoke new questions and provide a framework for assessing the relationship between AI and national power.* *This work benefitted directly from the early Office of Net Assessment summer study on AI in 2016. \n WHY STATIC, UNIVERSAL MEASURES OF POWER (THAT ARE USEFUL) DO NOT EXIST Power is simply the relative capability of a state to achieve what it wants in international affairs. Power depends on one state's favorable factors relative to another. In one of the founding works of international relations, Hans Morgenthau proposed distinguishing nine elements of national power: geography, resources, industrial capacity, military preparedness, population, national character, national morale, quality of diplomacy, and quality of government.  5  Since Morgenthau's writing, generations of analysts have sought a definitive way to measure national power that would, finally, allow accurate judgment of relative strength without fighting a war.  6  The search has included dozens of books and hundreds of journal articles offering competing methodologies and metrics.  7  For example: Should measures of useful access to resources include both steel and oil, or only steel? How should \"soft power\" be measured? What about the \"latent\" power that a state could theoretically draw from its population?  8  Were such a universal, \"objective\" measure obtainable, the benefits would be enormous. We could easily answer questions such as, \"who's ahead?\" and \"if it's not us, what should we do about it?\" This quest, however, has not borne fruit. Proposed measures have tended to perform poorly when generalized. 9 History is full of surprises where states have An Evolutionary Theory of Technological Competition 1 achieved victory even when \"objective\" metrics would predict their defeat: the United States owes its national existence to victory in the Revolutionary War over the British Empire, the superpower of the time. Why? First and foremost, power is always contextual. 10 This is especially clear in military matters. A large military's skill at high-intensity conflict may not translate to skill at counterinsurgency; and factors that provide one country advantage relative to another can change. The world offers no \"power particle\" to measure objectively alongside other characteristics of nature-what we intuitively mean by \"power\" is mostly a generalization from particular observations. 11 Elements of power can also combine in surprising ways. Andrew Marshall offered the reminder that countries with relatively smaller populations and GNPs can pose substantial challenges to larger competitors: in 1938, Japan had roughly half the population and one-tenth the GNP of the United States,  12  but it built a navy that challenged the United States in wartime.  13  In part because of these issues, history is rife with leaders who have had a large gap between their beliefs and the reality of military capabilities.  14  Each competition should be analyzed carefully on its own, distinguishing elements of power, identifying key areas of competition, and working to diagnose the most important problems and opportunities in each area of competition. \n MAJOR INNOVATIONS CHANGE THE SECURITY ENVIRONMENT, CHANGING WHAT GIVES RISE TO POWER Military leaders throughout history are sometimes faulted for preparing to fight the previous war instead of the next one. We should likewise avoid strategizing to \"win the previous competition.\" Just as changing adversaries from the Soviet Union to insurgents in Iraq and Afghanistan represented a new security environment and revealed the non-fungibility of power, major innovations also change the security environment as they are widely adopted. Such innovations do this in part by changing what assets, practices, and strategies give rise to power. Differential impacts of emerging technologies often bring shifts in relative capabilities of individual countries.  15  Thinking about long-term competition in periods of rapid technological change therefore requires assessing how innovations change factors related to military and national power. Major innovations can change the estimations of power in three ways: • First, innovations introduce new elements of power. Major innovations, in changing how states generate power, can create new factors that must be considered in characterizing power. For example, the advent of railroads, internal combustion engines, and nuclear weapons dramatically increased the importance of a state's access to steel, oil, and uranium, respectively.  16  New factors, however, are not only limited to materials. They may also encompass characteristics of a society's culture, organizations, or economic activities.  17  • Second, innovations change the importance of existing elements of power. Major innovations also change the \"coefficients\" of existing elements of power, causing them to matter more or less than before. For example, Mongol light cavalry, modern navies, and ballistic missiles all changed how geographic barriers affected one's balance of power with geographic neighbors, eroding the effectiveness of simple remoteness, oceans, and armies still in the field, respectively, as shields against coercive power.  18  Industrialization meant the inventiveness of a nation's scientists and engineers became more important. For example, before the Industrial Revolution, potential productivity gains in areas like agriculture and manufacturing were small and stable; this made conquering territory a primary means by which one group could increase its wealth and security.  19  During and after the Industrial Revolution, modern states could also pursue substantial military and economic growth by applying new technologies to increase productivity. The next section discusses how these three changes manifest in the context of AI. Perhaps least obviously, major innovations sometimes broadly alter what policies states pursue, by making certain kinds of behavior more valuable or less costly. e offer early thinking about potential changes caused by AI: new elements of power, shifting importance for existing elements of power, and shifting intermediate goals for states. These are not definitive or complete results, but a starting place for broader thinking. \n NEW ELEMENTS One of the most familiar examples of new elements of power is associated with the Industrial Revolution, when machines began to help humans with physical labor in new and organized ways. The Industrial Revolution led to dramatic changes in the character of war and military power. A simple approximation is that, before the Industrial Revolution, any group's military power correlated most closely with its quantity of fieldable humans under arms, a measure of both taxable population and military potential. After the Industrial Revolution, any estimate of military power had to include a society's industrial capacity and access to resources to enable that capacity, which are measures of a society's ability to produce useful military hardware, such as ships, tanks, planes, and submarines. It is useful to see AI technologies today as part of another large-scale transition: machines are increasingly helping humans with certain kinds of cognitive labor in new and organized ways.  20  This transition will span decades, with potential economic and social implications on a scale comparable to those of the Industrial Revolution. Today, as then, there are Power After AI: New Elements, Changed Factors, and Altered Goals 2 W large questions about the future of economic production, human labor, and military capabilities. These future trends will define new elements of power. U.S. defense leaders believe the rapidly growing military applications of AI technologies will be critical for the years ahead.  21  State power will increasingly hinge on the new factors required to effectively adopt AI. Four such factors often identified by existing literature include data, AI scientists and engineers (\"AI talent\"), computational power (\"compute\"), and AI-adapted organizations. Below, we explore the latter two in greater detail. \n Ability to Access and Leverage Compute The United States has historically used large-scale compute capabilities for analysis of nuclear weapons detonations and cryptanalysis.  22  More recently the U.S. government's uses have grown to include climate modeling and a variety of scientific applications. In the years ahead, the United States may also use large compute resources for creating and countering new AI capabilities. For decades, cutting-edge AI systems have used steadily increasing quantities of compute resources, making improvements in compute capabilities a key driver of AI progress. This usage appears to have accelerated across the last decade: the compute used in the largest AI training runs has doubled every 3.4 months since 2012, growing more than 300,000 times from AlexNet in 2012 to AlphaGo Zero in 2018.  23  OpenAI researchers have shown that the 2010s appear to be the beginning of a new computing era for AI technologies, distinct from the preceding 40-50 years.  24  For military applications where limited real-world data is available, techniques leveraging computer simulations instead of large quantities of data may further increase demand for compute.  25  Cloud compute may become vital for rapidly processing and fusing intelligence across platforms, while edge compute will be necessary for autonomous systems deployed in the field tasked with assessing and outthinking adversaries' equivalent systems. As such, a nation's ability to leverage large quantities of computational power could become a new primary term feeding into its ability to influence international U.S. defense leaders believe the rapidly growing military applications of AI technologies will be critical for the years ahead. State power will increasingly hinge on the new factors required to effectively adopt AI. affairs. For example, the key technical precursors required to manufacture cutting-edge AI chips are currently concentrated in the United States and allied countries-though semiconductor manufacturing capabilities more broadly, beyond just the most cutting-edge chips, may further grow the importance of Taiwan and South Korea as international trading partners.  26  Importantly, compute resources must be configured in ways useful for modern AI capabilities. High-performance computing (HPC) systems currently maintained within the U.S. Government, such as in the Department of Energy, tend to be both specialized for non-AI functions and subject to system-specific security measures, posing challenges for broad, standardized utilization by other organizations. Consequently, commercial cloud compute resources may better serve the U.S. Government in deploying certain kinds of AI technologies, although potentially promising efforts to improve the use of U.S. HPC assets for AI are also underway.  27  Effective use will depend, too, on accessible software tools for using cloud compute systemswhich may prove to be comparable to process and tooling approaches developed to make factories effective during industrialization in the United States.  28  Compute resources can flow more easily than many traded goods. As computing infrastructure continues to grow, new ways of sharing access to large, regionally-concentrated quantities of compute, including through space internet constellations, may create new opportunities and incentives for international partnerships. \n Ability to Manage Data and AI Safety & Security Even when states possess the raw resources required to adopt some major innovation, they still must undertake the often-difficult process of institutional and organizational adaptation. Bureaucratic factors in organizations matter greatly: in militaries, competing civilian, interservice, and intra-service actors may promote or resist adoption of new technologies.  29  Resistance can include parochial forces that attempt to stymie adoption: for example, only direct pressure from Eisenhower moved the Air Force to adopt ICBMs instead of focusing solely on less survivable crewed bombers.  30  Organizational culture also has significant impacts: because mass armies threatened the pre-existing hierarchical power structure within many European militaries, many states failed to adopt Napoleon's innovation even after his dramatic string of victories.  31  During periods of rapid change, medium-sized powers may have opportunities to adopt innovations more speedily than larger powers.  32  With AI, demands for organizational adaptations will be significant. Two factors are especially important: effective data pipelines and the effective management of security issues associated with modern AI technologies. The ability to deploy cutting-edge AI applications will increasingly depend on the quality of each organization's data pipeline. Modern machine learning methods are notoriously data-hungry, but simply possessing large quantities of data-collecting sensing platforms will be insufficient-for supervised learning applications, data must be structured, labeled, and cleaned; fusing data from many platforms, sources, and formats will represent its own herculean challenge for many militaries. Finally, these data pipelines must also be dynamic: data management itself must be monitored, in part to detect attacks, because \"data poisoning\" attacks can manipulate AI behavior by changing what lessons it learns.  33  Consequently, it will be increasingly important for military leaders to successfully implement organizational reforms to create and maintain effective data pipelines. Military leaders must also learn to effectively manage the novel security issues associated with AI technologies. Relying on modern AI systems for safety-or mission-critical tasks carries challenges because many deep learning models are exceptionally hard to interpret.  34  Michael Jordan at UC Berkeley has analogized the creation of early large-scale AI models to building bridges before civil engineering was a rigorous discipline: \"While the building blocks are in place, the principles for putting these blocks together are not, and so the blocks are currently being put together in ad-hoc ways. … Just as early buildings and bridges sometimes fell to the ground-in unforeseen ways and with tragic consequences-many of our early societal-scale inference-and-decision-making systems are already exposing serious conceptual flaws.\"  35  A more developed engineering discipline for AI is needed to manage the risk of accidents from relying on opaque machines in the field.  36  In near-term military settings, effectively integrating new AI technologies will require special investment in test, evaluation, validation and verification (TEVV) processes by competent organizational leaders.  37  More widely, many modern AI systems are not designed to work in the presence of malevolent actors. Potential security issues for deep learning systems include adversarial examples and model inversion, in addition to data poisoning and more traditional computer network and software attacks.  38  Adversarial examples refer to \"inputs\" (such as visual or audio patterns) to an AI system that cause the system to malfunction; model inversion refers to an ability to reverse-engineer the data used to train an AI system, which may include private or classified information. Despite these challenges, modern machine learning capabilities will be increasingly woven into G20 societies, economies, and military systems.* The U.S. position with *For example, AI technologies will intersect with 5G and networking trends in cities as autonomous systems (like vehicles) in urban areas begin to have large quantities of interactions with other intelligent agents-working on everything from traffic coordination to utilities management and financial investments. The ability for intelligent systems to interact on large scales, safely and securely, will be critical. AI technologies for the next two or three decades appears analogous to the future that faced IT technologies in the 1990s: AI technologies are so valuable that they will be used despite substantial design and security issues. What might the future look like given these vulnerabilities? We can only speculate: in direct military settings, there may be new sub-competitions that resemble the emergence of electronic warfare after the invention of radar.  39  In economic systems, in addition to the potential for the novel security risks discussed previously, there is risk of physical manifestations of the kinds of problems currently seen in high frequency trading systems, such as rapid, unanticipated interactions among automated agents managing services in cities.  40  These issues may open new vulnerabilities to both individual rogue actors and state adversaries. Organizations that are able to adapt early to manage these new security issues will be advantaged. Since states vary in their access to compute, data, AI talent, and useful organizational adaptations, they will also vary in their ability to benefit from modern AI technologies. Any national rankings based on these factors will be debatable, but the nations that generally lead in these metrics is unsurprising, and include: the United States, China, Japan, South Korea, the UK, Canada, Taiwan, Israel, France, Germany, and Russia. Advanced economies should be increasingly expected to focus their own investments and policies on improving their positions in these areas. \n CHANGED FACTORS Industrialization meant that a nation's stock of productive scientists and engineers counted more than it had in the past. With the arrival of AI, various previously recognized elements of national power will become more important, while others may become gradually less so. For illustrative purposes below, we discuss population size and scientific talent as contrasting examples: population size becoming less important, scientific talent becoming more important. The U.S. position with AI technologies for the next two or three decades appears analogous to the future that faced IT technologies in the 1990s: AI technologies are so valuable that they will be used despite substantial design and security issues. \n Population Size As AI technologies increasingly substitute for human labor, total population size may become less important for national military and economic capacity.  41  Just as machines took over rote physical labor during industrialization, AI technologies will automate rote cognitive labor, from diagnosing maintenance needs to exploiting imagery intelligence. This may reduce the total quantity of human labor needed to maintain a military's operational capacity. In major wars, partially or fully autonomous AI platforms may further reduce a country's need to field humans in combat. As militaries rely more on autonomous systems for military operations, defense planners may come to count autonomous systems and their available domestic supply of AI chips the way they once counted soldiers and the available domestic recruiting pool of military-age adults.  42  Downstream, this could help technologically advanced nations compensate for demographic challenges, such as aging populations and low birth rates, a situation the United States, China, Japan, Western Europe, and Russia all face to varying degrees.  43  Population trends continue to matter for national power-but AI technologies, like many other technologies of the past century, may further erode this importance. \n Industrious Scientists and Engineers Harnessing new technologies, both by developing technologies and accessing innovations created elsewhere, is an important means of growing power. Applications of AI can help in both areas, serving as a force multiplier on, and therefore increasing the importance of, productive scientists and engineers. Recently, for example, DeepMind's AlphaFold achieved breakthrough rates of accuracy comparable to experimental methods in the protein-structure prediction challenge known as CASP.  44  By obviating the need for experimental protein structure assessment, a skill-demanding and time-intensive procedure, AlphaFold represents a large augmentation of human scientists' biosciences research. In a different domain of research, modern AI applications are able to help with chip design.  45  Researchers have demonstrated a deep learning system capable of designing the physical layout of computer chips more effectively than human engineers.  46  Google has used this system to design its next generation of Tensor Processing Units (TPUs), the company's specialized AI chips. Likewise, rapid progress in machine translation, automatic literature review, and related tools means a given scientific discipline's state-of-the-art will become increasingly accessible and useful to well-organized groups of human scientists and engineers. Just as the printing press alleviated the need to travel from country to country to accumulate knowledge from different libraries, AI applications can lower the costs for researchers to access state-of-the-art knowledge in any field. There are three ways that modern AI applications will contribute on a large scale to scientific discovery and engineering invention: they will contribute directly to new discoveries and engineered systems, especially in areas that involve searches over large spaces in data or design;  47  automate the physical work of science and engineering, such as \"self-driving laboratories\" that robotically automate experimental laboratory work;  48  and make global scientific knowledge more accessible to humans, such as by extracting knowledge from millions of articles as well as from articles in many different languages.  49  Finally, there is an old debate about whether science advances most because of new ideas or new tools;  50  AI technologies appear able to contribute both. In the longer-term, AI may enable new and more creative forms of knowledge-generation that function as \"pathfinders\" for human brains, unlocking otherwise difficult-to-reach innovations. When AlphaGo beat Lee Sedol, its 37th move in the second game surprised human professionals. In the words of Go master Fan Hui, \"It's not a human move. I've never seen a human play this move. So beautiful.\"  51  When AI behavior surprises us, we learn something new. Looking ahead, modern and future AI systems may be able to solve scientific puzzles that have thus far stumped humanity's best minds.  52  Just as railways advantaged nations with access to steel, it appears that AI tools capable of augmenting science and engineering work will favor nations with the best existing \"resources\" of industrious scientists and engineers. This trend appears likely to deepen the advantages of nations that host, or can attract, a disproportionate fraction of the world's best in those fields.  53   \n ALTERED GOALS Finally, major innovations can alter state strategies, as different instrumental goals become more appealing for achieving a state's ultimate ends. The Industrial Revolution again provides a clear example. Before industrialization, conquering territory was a primary way that one group could increase its wealth and security relative to others.  54  During and after the Industrial Revolution, in contrast, states have been able to pursue these ends effectively by increasing productivity-as well as by gaining access to international trading networks and new technologies to enable further military and economic growth. Territorial conquest by states in the modern era is rarer for many reasons-but not simply because states have become more beneficent, instead because changes in technology have reshaped how they can best achieve their goals.  55  In short, major innovations can alter what long-term competitions in each era are fundamentally about. In the standard \"ends, ways, means\" trichotomy, this corresponds to ways. States have the same ends (security, wealth, prestige, influence, sovereign action), but the ways that competition is best pursued can change, such as through participation in globalized production chains instead of territorial conquest. With AI technologies, there are two worrying possibilities: a broad movement toward authoritarianism and the greater use of advanced forms of population-and economy-targeting information warfare. \n Social Control Temptations A technological innovation rarely tilts intrinsically toward \"freedom\" or \"authoritarianism.\" It is possible, however, to try to discern how new technologies may affect current social and economic systems in the future. Especially in authoritarian states like China, AI technologies may provide elites with tools that reduce contradictions between maintaining power and promoting economic growth through free markets. By making authoritarianism appear more feasible, this may generate an \"authoritarian temptation\" for the many states with malleable governance systems. First, AI technologies are likely to reduce the costs of controlling populations under authoritarian rule. Automating mass collection, processing, and analysis of data is likely to decrease the marginal cost of controlling additional citizens, thus reducing the resources required to indefinitely sustain totalitarianism. With access to hundreds of millions of cameras, social media postings, bank accounts, automated analysis of emotions and sentiment, and other data streams, AI-empowered algorithms can perform much of the work previously done by secret police in pre-AI authoritarian states.  56  Automated surveillance methods are likely to scale more effectively than manual surveillance, which requires some amount of human labor per citizen to be controlled. For example, Lichter et al. analyzed official Stasi records from East Germany, finding that more than 1.5 percent of the population was either officially employed or unofficially used as informers by the secret police.  57  Beyond the quantity of people involved in human surveillance operations, automated surveillance may impose lower economic costs on a society than human surveillance.  58  On this matter, China appears poised to benefit from feedback cycles between AI deployment and data aggregation-the Chinese government is already using AI technologies to enhance population control, as well as to profile and control its ethnic minorities.  59  In these early efforts, the Chinese government is collecting large quantities of data, from facial scans to DNA; COVID-19 has only deepened PRC data collection on its citizens.  60  This data will help fuel new AI development for social control in Chinese firms. Future AI applications could, in turn, help China manage its data and drive more expansive collection, continuing the cycle. China will likely export versions of these capabilities to authoritarian governments globally in the 2020s and 2030s, as it has already begun to do. According to recent CSET research, since 2008, over 80 countries have adopted Chinese surveillance technologies.  61  These tools will help authoritarian governments worldwide deepen their holds on power.  62  Second, and more speculatively, AI progress may benefit authoritarian states by reducing the costs and consequences of state interventions into internal markets. The classic critique of centrally planning complex economies is that attempting to do so poses intractable optimization problems.  63  For many practical reasons, from human organizational factors to corruption, AI technologies are unlikely to change this. However, AI technologies could reduce, to some degree, the negative consequences of state interventions in markets. For example, AI applications may help gather and interpret the volumes of information necessary for more effective economic controls. An analogous effect is visible inside large firms in both China and the United States today: companies like eBay, Taobao, Amazon, and Uber apply machine learning to mine large volumes of sales data to better match demand and supply. Modern machine learning tools enable automatic pattern analysis, improved forecasting, and natural language processing for predicting demand and performing sentiment analysis. Google's \"Smart Bidding,\" for example, uses machine learning to optimize conversions for ads; California uses AI to predict electricity demand, more effectively controlling the power grid and reducing blackouts.  64  Walmart's internal logistical management has analogs to a centrally planned micro-economy.  65  There are many challenges to using analogous tools effectively for state economic policy, perhaps most of all the variable goals of planners themselves. But these trends suggest national-level strategic planning may be able to benefit from better information by applying modern machine learning tools to data accessible by states. Leaders of authoritarian states like China may find themselves facing lower costs for sustaining domestic political and economic control; leaders of authoritarian-leaning states may find themselves handed these tools by China. The effects of AI on population control and state interventions in markets are not certain. In the near term, however, it seems likely that Chinese elites at least believe that AI may help them better control their society, and so too may elites in other states. be decisively influenced by the clash between AI influence systems, for example, China may determine its best bet for reabsorbing Taiwan is heavy investment in AI-empowered propaganda. Information attacks can also target economic systems and financial markets, especially AI systems associated with managing equities investments. An unintentional, early demonstration of this possibility occurred in 2013, when U.S. trading algorithms responded to disinformation posted by the AP's Twitter account after it was hacked.  66  Information warfare may be increasingly linked to economic warfare, not just political disruptions. Higher-end, AI-empowered information warfare is a more speculative, longer-term capability. Chris Wiggins has characterized current technical trends as enabling \"reality jamming\": the potential for synthetic, targeted, and optimized disinformation at web-scale.  67  In this future, current computational propaganda concerns are just the tip of the iceberg. The bigger issue is the potential for large-scale machine-generated information that is highly targeted at particular individuals or subpopulations, evolved to maximally shape particular behaviors, and potentially able to affect anyone with web access.  68  Leveraging these developments, governments may attempt to shape perceptions of other populations more frequently than in the past.  69  OpenAI self-censored full publication of its GPT-2 language-generation model in 2019, for example, because it was concerned that generating close-to-human text would enable nefarious actors to proliferate disinformation. It is easy to imagine states pursuing similar capabilities for their own ends.  70  According to recent CSET research, GPT-2's successor, GPT-3, may be especially potent at generating disinformation at scale when steered by a skilled human operator and editor, opening up the possibility of highly effective human-machine teaming.  71  These trends may pose challenges for democratic societies, though it is still too early to make clear judgments. Three unresolved questions exist today: First, if a long-term risk in authoritarian systems is intellectual conformity, an analogous effect in democracies may be mob majoritarianism.  72  This inherent challenge in democratic societies could turn out to be exacerbated by modern information technologies and make organizational reforms even more difficult. Second, more research is needed to understand the balance between democracies' ability to use disagreements and diverse information to advance new explanations and solutions, and the potential for information attacks to undermine political stability.  73  And third, most fundamentally, Western democracies, and particularly the U.S. system of government, are based on a foundation of individual freedom where individuals are the best judges of their own interests. It is not yet obvious how Western institutions will adapt to machines that can anticipate-or shape-individuals' own preferences, states, and choices better than the individuals themselves can.  74  In the context of international competition, leveraging AI technologies to alter target states' national priorities or political stability through information warfare would represent \"winning without fighting\" par excellence. In this evolutionary theory of technological competition, AI's effects on national power fall into three categories: new elements of power, changed factors, and altered goals. Exploring new elements required for successful AI adoption, such as compute and organizational adaptations, helps us understand when, how, and why some societies may be better positioned than others to benefit from major innovations. Similarly, the idea of changed factors helps focus on how existing elements of national power may have changing importance, such as population size and industrious researchers. Finally, thinking about altered goals of states in competition shows how major innovations can reshape the ways that states engage in competition, such as enacting new domestic political and economic controls and leveraging AI-enabled information attacks on other states' social and economic systems. This research offers a way to start thinking about these issues together, and hopes to spur new, wider thinking and work. Creating new conceptual tools for U.S. decision-makers and analysts to make sense of AI technologies' effects is vital to American prosperity. Over the long term, these technologies will create significant changes in U.S.-China competition. From this research, we see three early sets of insights into opportunities for U.S. leaders: The scale of possible impacts from major technologies is obvious: the United States benefitted greatly from growth connected to technological and economic changes in the 40 years from 1880 through 1920; and China has also already benefitted from a mix of technological and economic changes in its resurgence from 1980 through 2020.  77  Recent history demonstrates that getting technology right is critical for long-term national flourishing-and determining trajectories for the United States and China over the next 20 to 30 years. Can we sketch the longer-term future? Only speculation is possible today: Broad historical examinations tend to suggest that more successful societies present fewer obstacles to long-term change and, especially, limit the costs of intellectual conformity. They seek to maximize the benefits of pluralism, competition, and mechanisms to share, challenge, and supplement new knowledge.  78  A key challenge for China will be limiting the long-term costs of intellectual conformity induced by an authoritarian government. A favorable factor for China will be the dynamic organizations it has built over the last 20 years, which may remain able to adapt and benefit from organizational learning as the world continues to change over the next 10 to 20 years. In the longer term, however, continued evolution seems increasingly challenging for China under the CCP and absent substantial pluralism; many of its main challenges for net economic-technological growth are likely to persist, while the benefits of its dynamic organizations are likely to decline over time. A likely challenge for the United States will be institutional and organizational sclerosis, which will make organizational learning and adaptation challenging over the next decade. Interactions between AI technologies and democratic institutions increase uncertainty and may exacerbate these challenges. Weighing against these factors is Samuel Huntington's reminder of the United States' multidimensional sources of power and ability for self-renewal.  79  The most favorable factors for U.S. vitality and competition with authoritarian governments coincide with its enduring strengths: areas such as its cultural values and pluralism, overall approach to governance, and access to global talent.  80  In the longer term, the United States' central challenges appear more temporary, and its greatest advantages more enduring-a favorable outlook achievable with thinking and work today. It is useful to distinguish AI from autonomy. The former is defined above; the latter is best thought of as some degree of delegation of decision-making agency to another entity, which could be a human or a machine.  84  Systems can have neither, both, or one of these two things. For example, an autonomous military system can be unintelligent, as in the case of a landmine, or an intelligent system can support humans without autonomy, as in the case of an information system for a pilot. The 2010s were the third period of global excitement about AI. The first period occurred in the 1960s, centered in the United States and the UK, and the second period occurred in the 1980s, centered in the United States and Japan. Both periods were associated with significant investment and optimism for cascading breakthroughs in machine intelligence. Both periods were followed by \"AI winters\": periods of widespread divestment from AI R&D and the belief that earlier expectations had far exceeded reality.  85  The current period will probably be remembered as being centered in the United States and China, though with substantial activity in the UK, Europe, Canada, Japan, Israel, and South Korea. Since the 2010s, most excitement about AI has focused on machine learning (ML), and, within ML, mostly on applications of neural networks (deep learning). ML is a broad subfield of AI that centers on inference from data and overlaps substantially with statistics and optimization. \"Neural networks\" refers to a family of statistical models for extracting patterns from large quantities of data, originally inspired by the behavior of biological neurons. While the rediscovery and improvement of neural nets started the current AI wave in the late 2000s, specific trends over the last 20 to 30 years enabled the success of recent applications: global growth and diffusion of compute resources; large quantities of digital data globally; and the connection of these two by the global internet. For this reason, the foundation of modern AI advancements is often called the \"triad\" of new algorithms, compute resources, and data.  86  population, iron and steel production, energy consumption, military expenditure, and military personnel. (See: Singer, J. David, Stuart Bremer, and John Stuckey, \"Capability Distribution, Uncertainty, and Major Power War, <PHONE>,\" in Bruce Russett (ed.) Peace, War, and Numbers, (Beverly Hills: Sage, 1972), pp. 1948, as well as. https://correlatesofwar.org/data-sets/national-material-capabilities.) More recently, Michael Beckley has argued that traditional measures of power conflate gross resources with net resources, and thus fail to account for a country's burdens in addition to its assets. Thus, he proposes the use of \"GDP * GDP per capita.  \"    INTRODUCTION 1 |* 1 AN EVOLUTIONARY THEORY OF TECHNOLOGICAL COMPETITION 2 | POWER AFTER AI: NEW ELEMENTS, CHANGED FACTORS, In this work, we use \"artificial intelligence\" to mean, as per the Defense Innovation Board, \"a variety of information processing techniques and technologies used to perform a goal-oriented task and the means to reason in the pursuit of that task.\" See Appendix. \n \n \n • Thinking of long-term competitions in an evolutionary framework makes large, broadly-diffused technology changes akin to environmental shifts. Like a volcanic eruption or the start of an ice age, adaptations are valuable and some states will be better at adapting than others. It is useful to begin thinking about how AI technologies can create new elements of power, change the importance of existing elements of power, and alter the goals of states in competition. Getting a better sense of AI's effects in each of these factors will be critical for major powers. The United States has a number of opportunities: studying the approaches of other countries, especially U.S. competitors and medium-sized, quickly-changing countries; 75  developing strategies for global leadership in producing, using, and sharing compute resources; supporting development of AI engineering as a rigorous discipline in the United States and leveraging humans trained in it; continuing to push DOD and IC organizational reforms for how data is managed and leveraged; and leveraging AI tools, cross-training between AI and other disciplines, and high-skilled STEM immigration to access new breakthroughs in science and engineering more widely. • AI technologies may change not only what states can do, but also what they want. Major innovations can broadly alter intermediate, instrumental objectives that states pursue by making certain kinds of behaviors more valuable or less costly. This can drive dramatic changes in state goals and policies. The United States may look for new opportunities in technolo-For the United States, this means learning from its competitors without mirror imaging them and sharing insights with allies before assuming they should symmetrically match U.S. policies. Perhaps most significantly, it may also mean looking ahead to how AI technologies may affect the aims and interests of U.S. allies and partners. 3 Conclusions and Key Points I gy-related democracy promotion; shaping AI technologies themselves to Creating new con-favor democracies, such as by supporting development of AI technologies ceptual tools for U.S. with less dependence on centralized data; 76 and developing approach- es to more rapidly adapt social and economic institutions to \"information decision-makers and analysts to make attacks\" by AI systems. sense of AI technolo- gies' effects is vital to American prosperity. Over the long term, these technologies will create significant changes in U.S.-China competition. broad• Finally, effects of technological change can be highly asymmetric: new elements, changed factors, and altered goals may have very different manifestations in different countries. \n See \"The Power of Nations: Measuring What Matters,\" International Security 43.2 (2018): 7-44. 8. The best overview of this quest is Ashley J. Tellis, Janice Bially, Christopher Layne, and Melissa McPherson, \"Measuring National Power in the Postindustrial Age,\" RAND Corporation, 2000. 9. This statement includes whether such measures are used quantitatively to predict who wins a war, or whether war will occur, or whether settlement terms will favor one side or another. See: Ibid., 17. 10. For various discussions of this, see: Stephen Biddle, Military Power: Explaining Victory and Defeat in Modern Battle (Princeton: Princeton University Press, 2004); David A. Baldwin, \"Power Analysis and World Politics: New Trends versus Old Tendencies,\" World Politics 161 (1979): 161-94; Jeffrey Hart, \"Three Approaches to the Measurement of Power in International Relations,\" International Organization 30 (1976), 289-305. 11. Almond and Genco (1977) most famously made this point about how to think about political phenomena in general. See Gabriel A. Almond and Stephen J. Genco, \"Clouds, Clocks, and the Study of Politics,\" World Politics 29.4 (1977): 489-522.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET_Daniels_report_NATIONALPOWER_JULY2021_V2.tei.xml", "id": "9a9502d5180dc710fbe2eabecb9cc66c"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "The underwriter and the models -solo dances or pas-de-deux? 1 involved detailed conversations with, and questioning of, employees in various roles within the company.", "authors": ["Stuart Armstrong", "Mario Weick", "Anders Sandberg", "Andrew Snyder-Beattie", "Nick Beckstead"], "title": "The underwriter and the models- solo dances or pas-de-deux? What policy data can tell us about how underwriters use models", "text": "Executive Summary Using a collection of data on some of the catastrophe policies written since 2006 by a major reinsurance organisation, this paper explores how tightly underwriters follow the models and under what conditions they deviate from them. The data included underwriter premium and LE (loss estimate), and sometimes included LE from up to four different models (AIR, RMS ALM, RMS DLM, and IHM -the in-house model). We analysed the data in order to see what could be said about the relationship between model LE and premiums (as well as underwriter LEs). Mimicking a common procedure in machine learning, the data was randomly divided into a training set and a testing set, allowing many different theories to be investigated on the training set without risk of overfitting and detecting spurious connections. \n Results There were three main results, all statistically significant and with large effect sizes: 1. The models gave good predictions as to what the underwriter premium and LE would be. In a regression test, 79% of the variance in the premium, and 88% of variance in the LE, was explained by variance in the models. In fact, most of this variance was explained by the mean LE of the four models -78% and 87% respectively, corresponding to correlations of 0.88 and 0.93. 2. As the modelled loses rose, underwriter estimates moved closer to those of the models. This was evident through a variety of measures: the underwriters would report using more models than otherwise, the underwriter LE would become more strongly correlated with the mean model LE, and they would have less extreme premium/ LE ratios (meaning that the LE information was being used more to fix the premium). It should be noted that one effect that we expected to find -that underwriters were more willing to follow the models if these were more closely bunched together -was not present in the data. It seems that the underwriter does not make much use of model spread. \n The role of underwriters As far as can be seen in the data, the underwriters premiums (and LE) were strongly correlated with the models' LE estimates. The higher the expected loss (as seen by model LE), the less likely the underwriters were to deviate from the models. That high correlation may seem to suggest a limited role to the underwriter. However, this conclusion is premature for several reasons. Most significantly, this dataset only included policies that the reinsurance organisation had actually written, so the role of underwriters in rejecting policies could have been very important. There are also issues of yearly variation and changing market conditions (2012, for instance, seems to be an outlier in many ways). This analysis also ignores the effect of underwriters negotiating and interacting with brokers -even if every good underwriter were to sign similar kinds of deals, this does not mean the underwriters were superfluous. It is significant that the models were more predictive of underwriter LE than of premium (which would be influenced by negotiation). Underwriters may also play an important role in correcting erroneous information in the policy, and making sure that the correct models were applied in the first place. Finally, there were no details of outcomes in the data (which policies led to pay-outs, and by how much?), limiting our ability to estimate underwriter expertise. Thus it is possible and likely that the underwriter played (or could play) a more synergistic role with models, focusing on quality control, market insights, and business relations. This thesis can be strongly questioned. First, the social role of the underwriter should not be underestimated: they are involved in negotiations with brokers, and require perceptiveness of the opinions and behaviour of the other market actors. Secondly, when the practical role of the underwriter was analysed in detail, some of the work involved coping with poor data quality and correcting model errors. 5 Though these tasks do not appear on the list of bottlenecks to automation, they are certainly tasks that cannot be easily automated, as they represent a failure of the automation process itself. Currently (and for the foreseeable future), only humans possess the skills to apply these kind of corrections, which often involve deducing what kind of errors have occurred or what kind of extra data could improve the situation. Thirdly, underwriters may be using strategic intelligence to select models and to maintain the sort of overall vision that could resolve larger systemic risks. 6 Therefore it would be incorrect to see the underwriters as necessarily in direct competition with the models, but as occupying (potentially) different and complementary roles. This was one of the approaches advocated in the \"autopilot problem\" paper: change the \"pilot's\" (or underwriter's) role.  7  Merely relying on effective automation technology can both weaken the skills of the human and make the whole endeavour vulnerable to situations where the automation fails for some reason. However, making use of particular human abilities to control or complement the automation allows for both greater performance and better robustness. For instance, though simple linear models outperform expert predictions in many domains, these simple linear models could only be constructed thanks to expert knowledge of the important factors.  8  Similarly, experienced underwriters are potentially aware of what information is relevant in a particular type of case -and whether such a case is at hand. The quality of expertise is highly dependent on features of the task, 9 rather than on features of the expert, so it is important that the task the underwriter performs be well designed to make best use of their human qualities (especially as expertise tends to be quite specific to the task performed 10 ). If this can be achieved, the underwriter and models of the future will amplify each other. We thus need to understand how underwriters currently interact with models. To this end, a major reinsurance organisation has made available several years' worth of records on policies priced by its underwriters and by its models so that a comparison could be made and the role of the underwriter teased out. The most useful features of this data were the premiums that were actually charged, the underwriters' Loss Estimate (LE), and the same LEs as given by the models. Using this data, this paper explores how tightly underwriters follow the models and under what conditions they deviate from them. \n Introduction The underwriters have traditionally been the key players in the insurance industry, making the final decision on any particular policy. It is their responsibility to negotiate a price and ultimately accept or decline to insure the risk. But recent years have seen the emergence of another key insurance player: computerised models that give their own estimation of risk, exposure, and other critical features of a policy. In CAT (catastrophe) insurance, these models are now used extensively by insurers, re-insurers, and regulators, and have underpinned the risk-linked securities markets. 1 In this new, model-centric world, what is the current role of the underwriter? More usefully, what will the role of the underwriter become? Some studies suggest that the underwriter is soon to be replaced by automation.  2  The Frey and Osborne study analysed current automation trends and concluded, based on O*NET data (an online job classification service developed for the US Department of Labor 3 ), that underwriting involved none of the skills estimated to be hard to automate such as manual dexterity, strategic intelligence or socially dependent tasks 4 (see Table  1 ). \n Computerisation bottleneck O*NET Variable O*NET Description \n Perception and Manipulation \n Finger Dexterity The ability to make precisely coordinated movements of the fingers of one or both hands to grasp, manipulate, or assemble very small objects. \n Manual Dexterity The ability to quickly move your hand, your hand together with your arm, or your two hands to grasp, manipulate, or assemble objects. How often does this job require working in cramped work spaces that requires getting into awkward positions? Creative Intelligence Originality The ability to come up with unusual or clever ideas about a given topic or situation, or to develop creative ways to solve a problem. Fine Arts Knowledge of theory and techniques required to compose, produce, and perform works of music, dance, visual arts, drama, and sculpture. \n Social Intelligence Social Perceptiveness Being aware of others' reactions and understanding why they react as they do. \n Negotiation Bringing others together and trying to reconcile differences. \n Persuasion Persuading others to change their minds or behaviour. \n Assisting and Caring for Others Providing personal assistance, medical attention, emotional support, or other personal care to others such as coworkers, customers, or patients.  \n Results Three primary results were prominent in this initial analysis. First, variance in the model's LE explained the majority of the variance in premium (and in underwriter's LE). Second, underwriters tend to be conservative in estimating losses, more often setting expected losses above those of the models rather than below them. Last, the premiums (and underwriters' LE) moved closer to the models for more expensive policies. The causes and implications of these results are still uncertain -this preliminary analysis is only capable of identifying correlations, not causations. Note that the high R 2 need not mean that the underwriters are explicitly using the mean LE in their estimates -the models are highly correlated with each other, as can be seen in Table  2 . Thus many linear or quasi-linear combinations of models will be highly correlated with the models, with their mean, and hence with premium. Indeed, the correlation between premium and models is quite comparable with the correlation the models have with each other.  \n Data and Methods The reinsurance organisation provided a large collection of policies with appropriate premium and LE information. Data was taken from two sources: a reinsurance system used to record catastrophe model outputs; and an underwriting system used to record class of business and other risk details. The data were compiled over several years since 2006, with modellers recording model outputs and underwriting teams recording risk details as business was transacted. The data included premium charged, underwriter LE, general location, year, and potential LEs from up to four different models. LE is the loss estimate -the mean amount of money that the underwriter expects their company will pay out on that policy. The more usual \"loss on line\" is the LE for a particular \"layer\" of insurance, divided by the limit for that layer. The four models were AIR Catrader (AIR), RMS RiskLink Aggregate level model (RMS ALM), RMS RiskLink Detailed level model (RMS DLM), and the organisation's in-house model (IHM). After cleaning the data and restricting to US/ Canada policies (where the models are the most reliable), we were left with a collection of 660 policies where all four models were used 11 -no further selection was applied to this set. It was decided to split the data into a training and a testing set. This allows \"cross validation\", where hypotheses are formed on the training set and tested on the testing set. 12 This prevents overfitting, in which hypotheses are tailored to narrowly to the data, modelling noise rather than signal.  13  The training set would be thoroughly analysed to generate hypotheses; these would then be tested for statistical significance on the testing set. The ideal size of the testing set is 1/3 of the original set.  14  As the splitting into testing and training was done before restricting down to the 660 policies, this subset ended up divided into 457 policies in the training set, and 203 in the testing set. A total of 32 hypotheses were formed on the training set, which were then tested on the testing set, and all were found to be significant at the 5% level 15 , even when accounting for multiple comparisons  16  . Most of these hypotheses were linear regressions, but other comparisons were made as well (see detail of results). Result 2: Conservative Pricing Underwriters tended to set their LEs somewhat conservatively. Underwriter LE were above the minimum model LE 94% of the time (as compared to below the maximum model 83% of the time). See Table  5  in next section for breakdown of this into highest and lowest quartiles.  22  This was especially interesting as the maximum value and the mean were both better predictors of underwriter LE than the minimum (see Figure  1  and Table  4 ). Thus, it seems the minimum value was a rough lower bound for underwriters, but that they used little information from it beyond this. This strong correlation could be a sign of an autopilot problem 20 , if underwriters put too much trust in the models. In the meantime, we can analyse how these correlations change over time. For example, Table  3  demonstrates how in 2012, the fit between premium and model LE declined, as compared to surrounding years. This was around the period when the controversial \"RMS 11\" windstorm model was in use.    First of all was the simple fact that they used (or at least recorded the use of) more models. The full US/ Canada data (6138 policies) was decomposed into categories dependent on the number of models recorded 23 , and the median premium and underwriter LE of each category was computed (see Figure  2 ). For illustration, the interquartile distance was added in the case of zero and four models. Though the ranges overlap, those policies which recorded two or less models clearly had lower LEs and premiums than those that recorded three models or four. Figure  2  -Median underwriter LE and premiums, depending on how many models were reported for that policy. The inter-quartile range for 4 models and 0 models is also plotted. To see this, the 660 policies with four models were separated into quartiles.  24  The correlation between the underwriter LE and mean, minimum, and maximum model LE was computed in each quartile (Table  5 ). The data suggests underwriters hewed closer to the model information in these quartiles. A last way of looking at the difference in behaviour, is to consider what happens when the underwriter has established their LE, and compare this with the final premium. The data indicates that for high underwriter LE 25 , the premium/LE ratio is likely to be less extreme that for low ones. The linear anti-correlation is weak (-11%), but the effect is clearer graphically (see Figure  3 ). Thus the premium is more influenced by LE estimates when these are large.  To establish the results overleaf, a total of 32 hypotheses were tested. The significance of each result was established using the testing set, with each p-value adjusted -increased 26 -using the Holm-Bonferroni method. See table 6 (where \"HBp-value\" is the Holm-Bonferroni 27 adjusted p-value). As can be seen, all hypotheses were significant at the 5% level, and all but two were significant at the 1% level (these significance levels were marked in bold).  Of these hypotheses, 22 were regression/correlations with premium or underwriter LE as dependent variable, and either all model LEs or mean model LE as independent variables. We should expect these regressions to show scale invariance to some extent: if all the model LEs double, then, say, the underwriter LE should also double as well. This means that we expect the residuals (the deviations of the underwriter LE from its \"theoretic value\" as predicted by the regression model) to double as well. Thus we expect residuals to be higher for high expected losses and lower for low expected losses: the data should be heteroscedastic. And indeed it is (see Figure  4 ). Excessive heteroscedasticity precludes the use of the standard F-test to determine the p-values for the model. Instead, we took the logarithms of all the variables, expecting that this would remove the scale variations in the residuals. The residuals that resulted were much closer to being homoscedastic (see Figure  4 ). Thus the p-values of these regressions and correlations were calculated using the log-log regression rather than the standard linear regression. These p-values were sufficient to show that all regression models were significantly different from the null hypothesis. All regression data given in this paper (the R2 and the correlation coefficients), however, came from a standard linear regression, as a linear model is more likely to be closer to the underwriters' behaviour than a logarithmic one. Generally, the logarithmic model had slightly lower R2 than the corresponding linear model, but the values were very close (and on occasion the logarithmic model had a slightly better fit than the linear model). Statistical testing and significance \n Discussion The high correlation between model-estimated LE and premium (and underwriter LE) may seem to suggest a limited role to the underwriter. However, this conclusion is premature for several reasons. For a start, this dataset is likely incomplete, as it had to be put together specifically, and many policies failed to record model estimates.  28  Most importantly, it only included policies that had actually been written; the role of underwriters in rejecting policies could have been very important. The overall result ignores the fact of yearly variation: in particular, 2012 fits very poorly into the general analysis. It is likely that underwriters were aware of changing market conditions (or changes to the models themselves) and were able to react to them accordingly in that year. This analysis also ignores the effect of underwriters negotiating and interacting with brokers. It is significant that the models were more predictive of underwriter LE than of premium 29 (which would be influenced by negotiation). Underwriters may also play an important role in correcting erroneous information in the policy, and making sure that the correct models were applied in the first place 30 . Finally, there were no details of outcomes in the data (which policies led to payouts, and by how much?), limiting our ability to estimate underwriter expertise.  31  Conversely, there could have been a lot of wasted effort on the underwriter's part. The four models are so highly correlated (see Table  4 ) that any attempts by the underwriters to strike a fine balance between them would likely have made little impact. Given all those caveats, the story that the data presents is clear. If this sample is taken to be representative, then there seems little difference between using models to estimate loss, and having the underwriters do the same. More importantly, a similar pattern is true to a lesser extent in premium, where around 80% of the variance in premiums is explained by the models in a completely linear fashion. The remaining 20% is unlikely to represent perfect performance on the part of the underwriter, free of noise and bias. Underwriters, like all humans, are subject to general biases that interfere with their performance 32 , some of which could have a specific impact on their job.  33  The important question is how biases balance against expertise within this 20%. We have done some preliminary exploration of these biases and counteracting expertise with an experimental pilot study.  34  In that study, the underwriters appeared less model-bound than in this paper: the actual conditions of work may play a large role in decisionmaking. A full analysis of the underwriters' work is needed if the aim is to develop expertise that complements models in a robust way. We believe that data analysis such as this paper, an understanding of the cognition of underwriting, and a systemic perspective on the inherent risks of outsourced cognition -whether from autopilot biases or model-induced correlations across the insurance market -can help develop practices that both reduce systemic risk and amplify human capacity. It will become more and more vital for insurance companies to record their own data as this one has, and to analyse it intelligently.  28  In many cases, it is likely that models were used but the data wasn't recorded.  29  The models explain 88% of the underwriter LE variance in the 2010-2013 period, but only 79% of the premium variance.  30  From conversations with people in the industry, this last effect is more likely to be a factor in insurance modelling than in reinsurance modelling, which seems to be a more mechanical process. 31 And even loss data in the short term is not enough to estimate true underwriter expertise, as many risk as of low probability/high return period, and wouldn't show up in the data.  32  The bias literature is vast, but Kahneman, Daniel. Thinking, fast and slow. Macmillan, 2011 provides a good overview, Gigerenzer, Gerd. \"How to make cognitive illusions disappear: Beyond \"heuristics and biases\".\" European review of social psychology 2.1 (1991): 83-115 provides a good critique, and Kahneman, Daniel, and Gary Klein. \"Conditions for intuitive expertise: a failure to disagree.\" American Psychologist 64.6 (2009): 515 gives a good synthesis of some opposed views on the subject. 33 Beckstead, Nick. \"Biased error search as a risk of modelling in insurance\" in \"Systemic Risk of Modelling.\" Joint Future of Humanity Institute-MS Amlin White Paper 3 (2014) Figure 1 - 1 Figure 1 -Underwriter's LE as a function of cat model LE. With a set of models, underwriters are unlikely to set LE below the lowest model, and tend to stick fairly close to the mean model estimate. \n 23 5071 policies did not formally record a modelled loss, 27 had one, 85 had two, 295 had 3 and 660 had all four models. Dependency for Biggest Risks Premium and LE medians for varying number of models LE Furthermore, the correlation between underwriter LE and model LE increased for higher model LE. \n Figure 3 - 3 Figure 3 -Ratio of premium to underwriter LE plotted against underwriter LE. \n Figure 4 - 4 Figure 4 -Residuals for linear and log-log regression for underwriter LE versus all four model LEs. \n Table 1 O 1 *NET variables that are bottlenecks to automation, according to Frey and Osborne. \n Table 2 2 The premium is actually more highly correlated with the other models than the in-house model (IHM) is. Or, put another way, premium deviation with model LE is comparable to the noise 19 in the model LEs. Correlations between premium, model LEs, and mean model LE. Premium AIR LE RMS ALM LE RMS DLM LE IHM LE Mean LE Premium 1 0.825 0.808 0.796 0.852 0.881 AIR LE 0.825 1 0.909 0.864 0.878 0.966 RMS ALM LE 0.808 0.909 1 0.854 0.797 0.930 RMS DLM LE 0.796 0.864 0.854 1 0.758 0.915 IHM LE 0.852 0.878 0.797 0.758 1 0.937 Mean LE 0.881 0.966 0.930 0.915 0.937 1 \n Though the controversy was mainly around European windstorms, not US/Canada ones, this could have had an impact on underwriter trust in models. Alternatively, the large losses in 2011 could have played a similar role.21     Year # of policies All Model LE to Premium R2 Mean Model LE to Premium R2 All Model LE to Underwriter LE R2 Mean Model LE to Underwriter LE R2 2010 74 91.32% 76.08% 91.65% 89.87% 2011 345 84.94% 82.01% 91.65% 90.94% 2012 145 75.69% 69.56% 80.13% 75.58% 2013 92 95.15% 83.22% 94.83% 88.18% Mean 86.78% 77.72% 89.57% 86.14% Combined 656 78.80% 77.55% 87.63% 86.95% \n Table 3 R 3 2 between premiums and Underwriter LE with models, by year. 20 Armstrong, Stuart. \"The Autopilot Problem\" in \"Systemic Risk of Modelling.\" Joint Future of Humanity Institute-MS Amlin White Paper 2 (2014). 21 Aon Benfield. \"Reinsurance Market Outlook.\" Aon Benfield Analytics (2013). \n Table 4 4 Correlation coefficients between underwriter LE and the mean, maximum, and minimum of the model LEs. Mean Max Min \n Table 5 5 Correlation coefficients between underwriter LE and mean model LE, maximum model LE, and minimum model LE, separated onto the four quartiles. 1st quartile 2nd quartile 3rd quartile 4th quartile Mean 0.449 0.233 0.598 0.876 Max 0.486 0.387 0.363 0.847 Min 0.231 -0.035 0.329 0.747 Premium/LE ratio vs underwriter LE Premium/LE ratio Underwriter LE \n Table 6 p 6 -values and Holm-Bonferroni corrected p-values (for multiple comparisons) for all 32 hypotheses considered. \n\t\t\t The underwriter and the models -solo dances or pas-de-deux?The underwriter and the models -solo dances or pas-de-deux? 3 \n\t\t\t This was the judgement formed by three of the paper's authors during several periods of immersion at the reinsurance organisation, which \n\t\t\t US Gov't Accountability Office, GAO-02-941, \"Catastrophe Insurance Risks: The Role of Risk-Linked Securities and Factors Affecting Their Use\" (2002).2 Frey, Carl Benedikt, and Osborne, Michael. \"The future of employment: how susceptible are jobs to computerisation?\" Oxford Martin School Working Paper (2013). \n\t\t\t Or that, when it did involve these skills, that they were not fundamental to the job.4 The underwriter and the models -solo dances or pas-de-deux?The underwriter and the models -solo dances or pas-de-deux? 5 \n\t\t\t Of the form U i = β 0 + β 1 M AIRi + β 2 M RMS-ALMi + β 3 M RMS-DLMi +β 4 MGMi +ε i where U is the underwriters' LE, M X is the LE of the X'th model, and ε i is the error term for the i'th policy.18  Of the form U i = β 0 + β 1 M MEANi +ε i where U is the underwriters' LE, M i is the average LE of all four models, and ε i is the error term for the i'th policy.19  Using model LE variation as an informal measure of noise. \n\t\t\t Picard, Richard R, and R Dennis Cook. \"Cross-validation of regression models.\" Journal of the American Statistical Association 79.387 (1984): 575-583. 13 Hawkins, Douglas M. \"The problem of overfitting.\" Journal of chemical information and computer sciences 44.1 (2004): 1-12. 14 Dobbin, Kevin K, and Richard M Simon. \"Optimally splitting cases for training and testing high dimensional classifiers.\" BMC medical genomics 4.1 (2011): 31. \n\t\t\t Benjamini, Yoav. \"Simultaneous and selective inference: current successes and future challenges.\" Biometrical Journal 52.6 (2010): 708-721.6 The underwriter and the models -solo dances or pas-de-deux?The underwriter and the models -solo dances or pas-de-deux? 7 \n\t\t\t Note that we are using quartile in the sense of a set of data representing a quarter of the policies, not in the sense of the three values (lower quartiles, median, upper quartile) that divide the ranked data into those four sets. \n\t\t\t The underwriter and the models -solo dances or pas-de-deux?The underwriter and the models -solo dances or pas-de-deux? 9 \n\t\t\t That is, the policies were ranked according to their mean LE, and they were split into four sets at lower quartile, the median, and the upper quartile. By an abuse of nomenclature, these four sets are also called quartiles.25  Since the underwriter would set their own LE before setting a premium, it is valid to use underwriter LE on the x-axis at this point. \n\t\t\t p-values of the individual hypotheses were calculated separately, then the hypotheses were ordered by increasing p-value. These p-values were then multiplied by (32+1-r), where r was the rank of the hypothesis on the list. Thus the hypothesis with the lowest p-value had this value multiplied by 32, that with the next-lowest had its value multiplied by 31, all the way down to the one with the highest p-value, which was multiplied by 1. As long as these adjusted p-values were below the criteria of significance, then the familywise error rate would be below that.27 Holm, Sture. \"A simple sequentially rejective multiple test procedure.\" Scandinavian journal of statistics (1979): 65-70. \n\t\t\t The underwriter and the models -solo dances or pas-de-deux?The underwriter and the models -solo dances or pas-de-deux? 13", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/MS Amlin White Paper The underwriter and the models- solo dances or pas-de-deux.pdf.downloadasset.tei.xml", "id": "f7670968738e4bad319cf33d6fc19ee1"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "In its response to Covid-19, the UK has a once-in-a-generation opportunity to become a world leader in its resilience to biosecurity and other extreme risks. This report provides an excellent roadmap for doing so.\" \n Andrew Weber, Senior Fellow, Council on Strategic Risks and Former U.S. Assistant Secretary for Nuclear, Chemical and Biological Defence Programs \"Upgrading our risk management system to include extreme risks is both crucial to our national resilience and a financial no-brainer. The cost of implementing many of these proposed changes would be a rounding error in any Spending Review, but highly beneficial to the UK's national security. As Covid-19 has shown, it is a false economy not to spend millions now to save billions later.\" \n Baroness Pauline Neville-Jones, Member of the Joint Committee on National Security Strategy \"Extreme risks present a significant security threat to both current and future generations. I hope that the risk management community in government makes use of the expertise that this report, and its contributors, have to offer.\" Lord Des Browne, Visiting Researcher at the Centre for the Study of Existential Risk at the University of Cambridge, and Former Secretary of State for Defence \"Adopting these recommendations would be like taking out an insurance policy against some of the biggest threats we face. It is an idea whose time has come.\" Wera Hobhouse, MP \"This report's key observation -that our national resilience is all-toovulnerable to a range of extreme risks -is spot on. It is encouraging to see such focus on this critically important issue.\" Lord Toby Harris, Chair of the National Preparedness Commission \"This report offers significant recommendations about how the UK can enhance its resilience to extreme risks -one of the most important challenges of our time. It should be required reading for policymakers at all levels.\" Sir Oliver Letwin, former Chancellor of the Duchy of Lancaster and author of 'Apocalypse How?' \"It is heartening -and timely -to see experts set out a research agenda for improved UK resilience to the extreme threats we face, and offer their support to the Government. This is an offer that policymakers should seize with open arms.\" Lord Martin Rees, Astronomer Royal and co-founder of the Centre for the Study of Existential Risk at the University of Cambridge Policy area Current Government Focus 2 Recommended actions and estimated costs 2 We rate risks as green when, comparatively, they are already an extensive focus of Government policy, amber when there is some (but limited) policy focus, and red when they are almost entirely neglected.", "authors": ["Toby Ord", "Angus Mercer", "Sophie Dannreuther"], "title": "FUTURE PROOF THE OPPORTUNITY TO TRANSFORM THE UK’S RESILIENCE TO EXTREME RISKS", "text": "\"The prevention of the supreme catastrophe ought to be the paramount object of all 1. Create a pool of machine learning-relevant computation resources to provide free of charge for socially beneficial application and AI safety, security, and alignment research (£35 million annually). 2. Invest further in AI safety R&D. 3. Invest further in applied biosecurity R&D. 4. Invest further in improving long-term forecasting and planning. \n Green/ Amber The capabilities of artificial intelligence (AI) systems have increased significantly in recent years. Though there is still widespread debate and uncertainty, some AI experts have estimated that there is a significant chance that AI ultimately achieves humanlevel intelligence in the coming decades. A human-level artificial intelligence that is not aligned with the objectives and values of humans poses extreme risk, as does widespread deployment of today's existing capabilities. 1. Improve foresight and progress tracking in AI research (£600k annually). \n Bring more technical AI expertise into Government through a scheme equivalent to TechCongress (£1.5 million annually). 3. Ensure that the UK Government does not incorporate AI systems into NC3  (nuclear command, control, communications) , and lead on establishing this norm internationally. 4. Set up throughout-lifetime stress-testing of computer and AI system security (£200k annually). 5. Update the Ministry of Defence's definition of \"lethal autonomous weapons systems\". \n ARTIFICIAL INTELLIGENCE \n Amber \n RISK MANAGEMENT The UK does reasonably well at risk identification compared to other countries. Nevertheless, there remain a number of technical flaws in the National Security Risk Assessment and the National Risk Register which must be addressed, including difficulties with capturing extreme risks. There also needs to be greater crossgovernment accountability to ensure that these risks are addressed, that adequate plans are drawn up and that the latest science and research leads to changes in risk policy. 1. Improve extreme risk assessment and ownership across government by updating the NSRA, applying 'three lines of defence' model to risk management and installing a Chief Risk Officer (£8.3 million annually). 2. Lead the way to ensure global resilience to all extreme risks, not just pandemics, post-Covid-19. 3. Normalise red-teaming in Government, including creating a dedicated red team to conduct frequent scenario exercises (£800k annually). 4. Revise the Green Book's discount rate and ensure the Treasury adopts key recommendations on intergenerational fairness. \n Establish a new Defence Software Safety Authority as a sub-agency of the Defence Safety Authority, to protect UK defence systems from emerging threats (£5 million annually). 6. Fund a comprehensive evaluation of the actions required to increase the resilience of the electrical grid. \n Amber \n BIOSECURITY We are highly vulnerable to biological threats from bioweapons and accidental laboratory leaks, which risk even worse consequences than naturally occurring pandemics like Covid-19. Rapid developments are being made in synthetic biology and biotechnology. These bring great benefits, but also offer harrowing prospects of misuse. 1. Task one body with ensuring preparedness for the full range of biological threats the UK faces (£1 million annually). 2. Launch a prize to incentivise development of clinical metagenomics (£3 million one-off cost). 3. Establish a Biosecurity Leadership Council and appoint a liaison officer to improve coordination between the biosciences and security communities (£1 million annually). 4. Ensure that all DNA synthesis is screened for dangerous pathogens, and regulate DNA synthesis machines. Amber/ Red \n I N T R O D U C T I O N INTRODUCTION \n Extreme risks -the defining challenge of our time At several points in humanity's long history, there have been great transitions in human affairs that accelerated our progress and shaped everything that would follow. Ten thousand years ago, we had the Agricultural Revolution. Farming could support 100 times as many people on the same piece of land, making much wider cooperation possible. We developed writing, mathematics, engineering, and law. We established civilization. Four hundred years ago, we had the Scientific Revolution. The scientific method replaced deference to perceived authorities, with careful observation of the natural world, seeking simple and testable explanations for what we saw. Two hundred years ago, we had the Industrial Revolution. This was made possible by the discovery of immense reserves of energy in the form of fossil fuels. Productivity and prosperity accelerated, giving rise to the modern era of sustained growth. But there has recently been another transition more important than any that has come before. With the detonation of the first atomic bomb, a new age of humanity began. We finally reached the threshold where we might be able to destroy ourselves -the first point when the threat to humanity from within exceeded the threats from the natural world. These threats to humanity -which we in this report refer to as 'extreme risks' -define our time. We are currently living with an unsustainably high level of extreme risk. With the continued acceleration of technology, and without serious efforts to boost our resilience to these risks, there is strong reason to believe the risks will only continue to grow. \n INTRODUCTION What are extreme risks? 1 2 3 Extreme risks are high-impact threats which have a global reach, and include both global catastrophic risks and existential risks. The nature of extreme risks makes them difficult to assess and address, compared to more regularly occurring events, such as floods, earthquakes, or terrorist attacks. Global catastrophic risks are those which could lead to significant loss of life or value across the world. For a rough sense of scale, many research papers refer to risks of disasters that result in a loss of 10% or more of the human population. Existential risks are those which could lead to the premature extinction of humanity or the permanent and drastic destruction of its potential. Unlike global catastrophic risks, existential risk scenarios do not allow for meaningful recovery and are, by definition, unprecedented in human history. In his recent book, The Precipice, Toby Ord, one of this report's authors, estimates the likelihood of the world experiencing an existential catastrophe over the next one hundred years at one in six. 4 \n The post-Covid-19 opportunity Covid-19 has given us a sense of the devastating impact that extreme risks would have on our health and economy. In any given year, the likelihood of an extreme risk materialising is relatively small, but the odds that we -or our children and grandchildren -will face one of them are uncomfortably high. We do not know which extreme risk event will come next -it might be another pandemic, or it might be something completely different. But we do know what many of the most extreme risks are, and how best to prepare for them. The cost of better preparation for extreme risks pales in comparison to the cost of Covid-19 so far, which has been estimated at over £300 billion in 2020 alone. 5 Government spending on extreme risk resilience is the best kind of investment -one 1 We have footnoted all papers and reports which we cite and, for the convenience of online readers, also included hyperlinks in the body of the report for key resources. Just as Covid-19 triggers an immune response in each individual, protecting them from reinfection, so the pandemic has triggered a social immune response across the UK, where there is public will to prevent the next extreme risk. But like the individual immune response, this social immune response will fade over time. Before it does, we need to seize this opportunity to put in place lasting protections to safeguard the country from extreme risks -both at a risk-specific level and at a systemic level. Encouragingly, the new Integrated Review highlights the need for \"low-probability, catastrophic-impact threats\" 12 to be at the heart of the UK's efforts to build national resilience. It also commits the UK to developing a \"comprehensive national resilience strategy in 2021 to prevent, prepare for, respond to and recover from risks.\" 13 The next step is to set out exactly how the Government can embed extreme risks into its resilience planning -and into its upcoming AI strategy and biosecurity review. This report provides a roadmap for how to do so. We welcome engagement with the Government on any aspect of this report, via info@ longtermresilience.org. \n OVERVIEW Firstly, we provide an overview of two of the most extreme risks the UK faces in the twenty-first century -those relating to biosecurity and artificial intelligence. We assess to what extent they are a focus of current Government policy. We also provide recommended actions for the Government to take in the next twelve months to help safeguard the UK against these extreme risks. These recommendations are by no means comprehensive -rather, they focus on issue areas that are relatively neglected in the existing policy conversation and the key actions to take within those issue areas. While climate change and nuclear security are perhaps the best-known extreme risks, this report focuses on biosecurity and AI on the basis that the extreme risks they pose currently receive much less attention in policy circles. We then turn our attention to two cross-cutting policy recommendations to improve resilience to extreme risks at a systemic level. The first cross-cutting recommendation is to improve the UK Government's process for managing extreme risks. The second is to increase funding for research into extreme risks and cement the UK's leadership in this area. This would ensure that we better understand the nature of the threats we face and how best to deal with them. We conclude with a short section setting out further information on extreme risks. Estimated costs of implementing these recommendations in the upcoming Comprehensive Spending Review are included in this report where possible, and further information is available on request. T   H E E X T R E M E R I S K S - OV E R V I E W A N D R E C O M M E N DAT I O N S \n Summary of the risk The tragic events of the Covid-19 pandemic have highlighted the need for the UK to transform its level of preparedness against biological threats. But in our response, we cannot simply 'fight the last war' and focus solely on preparedness for future naturally occurring pandemics. We know from national security risk assessments and the UK Biological Security Strategy that we remain vulnerable to accidental and deliberate biological threats which risk even worse consequences than Covid-19. 1 Even more concerning are the very rapid developments that are being made in synthetic biology and biotechnology, which offer harrowing prospects of misuse. Fortunately, there are concrete steps that can be taken to ensure that the UK leads the world in efforts to mitigate these risks and prepare for all forms of pandemics. \n Current level of Government focus Responding to Covid-19 has rightly been the Government's core priority since early 2020. Alongside its day-to-day response to the pandemic, there have been several changes to the machinery of government to enhance pandemic preparedness going forwardincluding the creation of the UK Health Security Agency (formerly the National Institute of Health Protection). There are encouraging signs that the long-term protection of the UK's biological security is now higher on the political agenda than it was prior to Covid-19. The Government has committed to holding a public inquiry into Covid-19, which is due to start next year. Further, the Joint Committee on National Security Strategy's December 2020 report has called on the Government to \"plan for unexpected futures\", and recommended the creation of a \"dedicated national centre for biosecurity\". 2 The Government's response, published in March 2021, did not accept this recommendation specifically, but did recognise the \"need for a resilient and enduring approach to biological security\".  3  It also announced that urgent work is ongoing to identify where responsibility for biosecurity should sit long term. The Government has also announced a forthcoming review of the UK's Biological Security Strategy, though no specific date has been set for this. In the Integrated Review, we learned that the cross-government approach to biosecurity is being \"reviewed and reinforced\", and we anticipate further changes to the machinery of government. The Government committed to review its national stockpile of clinical countermeasures and consumables, such as personal protective equipment, expanded testing capability and laboratory equipment. It also set out the UK's ambitions for biosecurity, including international leadership; for instance, by increasing funding for the World Health Organisation and reducing priority vaccine development and deployment time to 100 days. 4 These are important steps to take. Yet much more needs to be done now to protect The threat of human-caused pandemics -by either accident or attack -is growing in step with the rapid march of biotechnological progress. This threat is even more of a concern than naturally-arising pandemics over this century, and therefore warrants similar attention from the Government. Though the Integrated Review indicates that an improved approach to pandemic preparedness is in train, it is both striking and disappointing that these human-caused disasters, such as accidental laboratory outbreaks and deliberate bioweapons, are barely mentioned in the review, while much more emphasis is placed on nuclear weapons. The UK has a poor track record of laboratory accidents. The 2007 foot and mouth disease outbreak was caused when it escaped from what was supposed to be the most secure level of lab in the whole world. Likewise the final victims of smallpox were in the UK, when it escaped from an insecure lab. Building a better safety culture in biotechnology is essential, and requires knowledge of the rate and causes of laboratory accidents. Yet these are currently poorly understood. The UK rivals the United States in terms of its bioscience capability, but we currently do not make the most of the expertise we have. Unlike in the United States, there are (to our knowledge) no permanent biosecurity experts on the UK National Security Council. The UK also has fewer senior Government officials with a technical background in biosecurity compared with the United States, and less of an established health security community. \n Key biosecurity policy recommendations 1. Task one body with ensuring preparedness for the full range of biological threats the UK faces (estimated cost: £1 million annually) Why this matters: Biological security is critical to the UK's national security. Covid-19 has shown that urgent reforms are needed, but these reforms must go beyond preparation for naturally occurring pandemics and avoid the well-known pattern of 'panic and neglect' in global health security. One designated body should have accountability to lead these efforts. We do not know when the next major pandemic will hit the UK -it could be decades away. This is plenty of time to sink into complacency or be distracted by other short-term crises. Ideally, a new National Centre for Biosecurity would be tasked with prevention of, and preparedness for, future large-scale and high-priority biological threats faced by the UK, regardless of their origin. It would provide strategic direction over policy and technical solutions, along with national-level coordination and integration of expertise from a wide range of disciplines to prevent and increase preparedness for biological threats. It would complement the proposed new research agency, ARIA, by fulfilling a think-tank-like function that delivers insights on new areas of opportunity and promising technical solutions.  5  It would also usefully draw on the latest foresight work in the areas of biosecurity and bioengineering. In short, such a centre's mission would be to ensure the biological security of the UK. To achieve this, it would focus on the four areas of highest priority: 1. Preventing and countering the threat of biological weapons from both state and non-state actors, treating them as a comparable security challenge to nuclear weapons; 2. Developing effective defences to biological threats, helping bring horizon technologies (especially pathogen-blind diagnostics) to technical readiness; 3. Promoting responsible biotechnology development across the world; and 4. Developing talent and collaboration across the UK biosecurity community, cementing the UK as a world leader in safe and responsible science and innovation. For further information, see Oxford's Future of Humanity Institute's working paper on the proposed Centre. It may be that the recently announced UK Health Security Agency takes on many or all of these priority areas, which would be another viable solution. What matters most is that UK biosecurity focuses not just on immediate threats, but also on prevention of and preparedness for the full range of biological threats we face. \n Launch a prize to incentivise the development of clinical metagenomics (estimated one-off cost: £3 million) Why this matters: Clinical metagenomics has the potential to identify new, unexpected pathogens in the first few infected patients, rather than months later. This would be game-changing, as it would allow for a much earlier, targeted response to an outbreak of an unknown virus. Metagenomic sequencing takes a sample from a patient, sequences the DNA of all organisms in it, and automatically compares these to a known database of pathogens, finding the closest matches. With coming technologies, this will likely be affordable to the point where doctors can routinely send in a sample from any case in the UK that they cannot diagnose with standard techniques. It would cost approximately £100 per sample for the metagenomic sequencing to be performed and to identify the closest matches. This capability could be integrated into the existing public health laboratory network, or take place in a central laboratory. This would be extremely helpful for both regular diagnoses and for novel pathogens. In the case of Covid-19, such technology would have immediately shown that the closest match was SARS, but that it was sufficiently different to be a novel SARS-like pathogen. Put simply: if metagenomic sequencing had been widely available at the start of 2020, the trajectory of Covid-19 may well have been very different. A prize of around £3 million could be awarded to the first group that can develop an interface that takes raw metagenomics data and turns out potentially clinically relevant results. Prize challenges are a wonderful innovation. They increase the number of minds tackling a particular problem without having to predict which team or approach is most likely to succeed. They are also an efficient means of identifying talented individuals and teams, who can be seconded for future programs. They tend to be about 10 times more costeffective than traditional research projects, and to prompt a higher degree of spending on research by the contestants. This is both because competitors tend to overestimate the probability of winning, and because they tend to place a significant value on the reputational reward for winning or being shortlisted. \n Establish a Biosecurity Leadership Council and appoint a liaison officer to improve coordination between the biosciences and security communities (estimated cost: £1 million annually) Why this matters: Biotechnology is often 'dual use', meaning that advances can be used for harm as well as good. For example, an individual could build live viruses 'from scratch' for legitimate research, but also to conduct bioterrorism. This makes biotechnology a highly challenging area to navigate, and one which requires a far greater degree of coordination than that which currently exists. This proposed new Biosecurity Leadership Council's role would be to develop biosecurity policy through collaboration between the Government, academia, business, and other relevant stakeholders. It would provide an official channel of coordination to ensure that there is dialogue between these stakeholders. The UK Synthetic Biology Leadership Council is a possible model. 6 The Council would help ensure adequate resourcing, both in terms of funding and expertise, and a liaison officer would improve coordination between the biosciences and security communities. This officer would provide advice and build relationships across Government, law enforcement, intelligence agencies, academic researchers, and private sector researchers. An equivalent position already exists in the United States. \n Ensure that all DNA synthesis is screened for dangerous pathogens and regulate DNA synthesis machines Why this matters: Malicious biological threats warrant equal concern to natural pandemics, but they are receiving considerably less policy attention following the Covid-19 outbreak. As the Secure DNA Project notes: \"A world in which many thousands of people have access to powerful and potentially dangerous biotechnologies is unlikely to flourish… Historical pandemics killed tens of millions of people, and engineered agents could be even more destructive.\" 7 Unless active controls are present, gene synthesis machines provide a way for individuals to get their hands on dangerous or novel pathogens. The export of desktop versions of these machines already require a license on biosecurity grounds. Gene synthesis companies should therefore be required to adhere to biosecurity guidelines for screening DNA orders for dangerous pathogens, such as those released by the Secure DNA Project. These guidelines go beyond the most commonly used International Gene Synthesis Consortium protocol to reflect rapid advancements in the field and current technological capabilities.  8  Imported DNA orders should adhere to the same biosecurity screening guidelines, and the UK should be a leader in the international community on further improving these initiatives and make screening more universal and robust. If full coverage cannot be achieved through self-regulation by gene synthesis companies, the UK should take a leading role in pushing for domestic and international regulation in this area. The UK is already at the forefront of 'biofoundries' which enable the rapid design, construction, and testing of genetically reprogrammed organisms for biotechnology applications and research.  9  The founding members of the Global Biofoundries Alliance are leading UK synthetic biologists. This is therefore a natural leadership area for the UK to adopt. • Misuse risks result from using AI in an unethical manner. For example, realistic but synthetic images or videos generated using AI can be used for activities like political disruption. And since AI capabilities are so broadly applicable, they may often have unforeseen harmful uses. • Accident risks involve harms arising from AI systems behaving in unintended ways, such as a self-driving car collision caused by a car failing to understand its environment. The more AI is integrated into safety-critical systems such as vehicles and energy systems, the higher the stakes are for accident risks. 10 • Structural risks will likely appear in the longer term. They involve the increasing use of AI to change political, social and economic structures and incentives. Widespread use of AI systems could exacerbate existing inequalities by locking in patterns of historical discrimination, provoking rapid and wide-scale unemployment, or dramatically concentrating power in the hands of a few companies and states. 11 \n ARTIFICIAL INTELLIGENCE CURRENT GOVERNMENT FOCUS AMBER \n Section Contributors Strikingly, when hundreds of scientists were surveyed about when they believed AI would reach general human-level intelligence, the median response was around 35 years from now. 12 And, of course, there is no reason to think that AI would stop at human levels of intelligence. Such technology will likely be highly beneficial to humanity in countless ways, but a human-level AI that is not aligned with human objectives and values also constitutes an extreme risk. Policymakers need to act now to ensure that AI is developed, used and governed responsibly, as an asset rather than a threat to human potential. \n Current level of Government focus The UK Government has so far demonstrated laudable proactivity in the field of AI policy and governance, establishing the Office for Artificial Intelligence, the Centre for Data Ethics and Innovation (CDEI), the AI Council, NHSX, the Regulatory Horizons Council, and becoming a member of the Global Partnership on AI. This planned activity makes the prioritisation of AI ethics vitally important. It is therefore heartening to see GCHQ's recent report confirm that it has commissioned work from the Alan Turing Institute to study the implications of AI for ethics.  14  The Government's published guidance on AI ethics and safety is another encouraging development, as is the National Data Strategy's emphasis on using data in an ethical and responsible way.  15  Initiatives from regulators such as the AI Auditing Framework and Project ExplAIn are also noteworthy examples of AI safety being taken seriously, at least with regards to accuracy, fairness and transparency.  16  The AI Council's AI roadmap puts a significant amount of emphasis on the importance of good governance and regulation to build public trust in AI, noting that \"making the UK the best place in the world to research and use AI requires it to be world-leading in the provision of responsive regulation and governance\". Finally, the Turing AI Fellowship programme is worth highlighting, as it enables further research into AI safety and robustness. 17 These initiatives represent an encouraging start. However, the UK's efforts on AI safety remain incomplete in some areas and embryonic in others. While risks of bias, transparency and accountability are frequently highlighted by policymakers, currently very few resources are invested to foresee, understand and mitigate the full spectrum of risks that AI safety researchers are concerned about. \n Key AI policy recommendations \n Improve foresight and progress tracking in AI research (estimated cost: £600k annually) Why this matters: AI capabilities and their potential applications in society are growing fast. The UK risks falling behind and taking an overly reactive approach if it does not develop its capability to monitor AI progress. The UK Government should establish its own capacity to anticipate and monitor AI progress and its implications for society. We recommend funding: • Greater capacity inside Government for establishing metrics and mechanisms to assess progress in AI, its applications and impacts on society. This could be achieved either by a new body (which could be housed within CDEI or the Alan Turing Institute), or by substantially increasing the scope and funding for existing initiatives, such as CDEI's existing AI-monitoring capacity. • Research projects in AI foresight and progress tracking, to be awarded through the mechanisms set out above. \n Bring more technical AI expertise into Government through a scheme equivalent to TechCongress (estimated cost: £1.5 million annually) Why this matters: As AI systems become more capable, their impacts will grow and become more cross-cutting, increasing the need for technical expertise across the UK Government. Such expertise is currently sorely lacking. The Government is steadily bolstering the supply of tech talent within Government through initiatives such as the Number 10 Innovation Fellowship, the Data Science Graduation Programme and the Data Science Campus Faculty. However, there is more it could do to plug the current gap in technical AI expertise, including: • Setting up a TechCongress-equivalent scheme (potentially as part of the crossgovernment Open Innovation Team) aimed at enabling the UK Government to recruit and gain access to AI expertise in fields like AI governance and ethics.  20  The scheme should place experts in Parliament and also embed them within the Civil Service. We recommend that an appropriate body or individual at the Ministry of Defence investigates the process that the UK would need to undertake to make a credible commitment that it will not incorporate AI systems into NC3, and then makes this commitment. We also recommend avoiding (and publicly committing to avoid) cyber operationsincluding intelligence-gathering operations -that target the NC3 of Nuclear Proliferation Treaty signatories.  21  A recent report by the US National Security Commission on Artificial Intelligence made a similar recommendation.  22  We further recommend that the UK advocates for this policy norm internationally; for example, by establishing a multilateral agreement to this effect. 4. Set up throughout-lifetime stress-testing of computer and AI system safety and security (estimated cost: £200k annually) Why this matters: For the UK's national security, it is important to stress-test computer systems thoroughly to test their resilience and identify flaws. To do this well, there must be an incentive structure to point out problems rather than underplay them. The Government announced in November 2020 that it would create a new AI agency and re-orient its defence capability towards emerging threats.  23  It is vitally important that any new AI-related bodies prioritise the safe development of AI, and in particular that they set up throughout-lifetime stress-testing of computer and AI system safety and security. Stress-testing allows systems to be assessed for flaws and vulnerabilities before they can be exploited by adversaries, or before accidents involving new systems occur. This is particularly important given the Government's decision to incorporate AI into its defence capabilities -for example through the Royal Air Force's AI and drone technology. We recommend that computer and AI systems be stress-tested during development, testing, training, early deployment, at regular intervals and before retirement of relevant systems. The Government should also have dedicated 'white hats' stress test their systems by attempting to compromise their software and hardware vulnerabilities, through social engineering and by designing adversarial environments. We also recommend making adversarial testing and red teaming part of military exercises. \n Update the Ministry of Defence's definition of 'lethal autonomous weapons systems' Why this matters: The UK should take a leading role in the ethics and implications of lethal autonomous weapons and promote international dialogue. The Ministry of Defence's definition of 'lethal autonomous weapons systems' differs from that of other governments, making international dialogue more difficult. \n 23 This was announced as part of the UK's November 2020 announcement that it would spend £16.5 billion on defence spending: https://www.bbc.co.uk/news/uk-54988870 Within the wide set of defence systems that integrate increasingly capable AI and machine learning, particular attention is rightly paid to lethal autonomous weapons systems. These systems raise important questions of ethics and international humanitarian law, and are the focus of arms-control negotiations at the United Nations. The Ministry of Defence's current definition of 'lethal autonomous weapons systems' is quite different from that used by many other nations. It defines an \"autonomous\" system as \"capable of understanding higher-level intent and direction\", \"capable of deciding a course of action, from a number of alternatives, without depending on human oversight and control\" and \"able to take appropriate action to bring about a desired state.\" 24 This definition is a very high bar to reach -almost human-level intelligenceand is so high as to be almost meaningless. No system currently under research or development would be capable of meeting this definition. The UK should adopt a definition similar to that used by other governments and international organisations in order to improve its ability to consider and protect against foreseeable risks associated with these systems, and to act as a global leader in setting international standards for this emerging technology.  \n Current level of Government focus The UK does reasonably well at risk identification compared to other countries. The National Security Risk Assessment (NSRA) and the National Risk Register (NRR) provide detailed analysis of the risks that the UK faces, and there is good use of horizon-scanning and foresight capacity. Nevertheless, there remain a number of technical shortcomings with the NSRA which need to be addressed in the review of its methodology which is currently taking place: • The NSRA does not sufficiently capture high uncertainty risks like extreme risks. Future risks and low-probability risks tend to be excluded from its assessment. • The way the NSRA delineates risks and highlights uncertainty is potentially confusing for decision makers -in particular, its unclear definition of \"reasonable worst-case scenario\". • The NSRA should include a robust vulnerability assessment, asking in particular how effective and well-documented existing mitigations and crisis management capabilities are. The shortcomings of the current approach to assessment of extreme risks have been made clear by Covid-19. The most recent National Risk Register failed to adequately estimate the scale of the crisis, estimating that emerging non-influenza infectious diseases could lead \"to up to 100 fatalities\". This was clearly a very large underestimation of the scale of the current crisis, as well as being out of line with academic evidence and other risk reports. It led to Government plans which focused too heavily on influenza rather than other diseases. Beyond the NSRA, there appears to be limited cross-government accountability to ensure that risks are mitigated, that adequate plans are drawn up and that the latest science and research leads to changes in risk policy. For instance, it is notable that the UK's pandemic influenza strategy did not make any plans for a lockdown, despite this being one of the dominant strategies for responding to Covid-19. 25 That being said, the Government has indicated an encouraging will to learn lessons for UK risk management in the aftermath of Covid-19. A new Cabinet Office National Situation Centre aims to improve 'situational awareness for crises and national security issues' by collating data and insights. And the new Integrated Review leaves readers in no doubt that building resilience to risk is now a top national priority. Ensuring \"a more robust position on security and resilience\" is defined as one of the four main components of Global Britain. The Integrated Review contained a range of promising announcements: • A new approach to preparedness and response to risks, which fully recognises that natural hazards and other extreme risks can cause as much disruption to the UK's core interests as security threats. • An ongoing external review of the NSRA and its underlying methodology; • A comprehensive national resilience strategy, with a commitment to include threats and hazards and all kinds of risk including the possibility of \"low-probability, catastrophic-impact events\". • A new Performance and Planning Framework and an Evaluation Taskforce (to drive change across Government and assess impact as per the Integrated Review's recommendations). 25 Samuel Hilton and Caroline Baylon's recent paper on risk management provides further details on the state of UK risk management, the lessons we can learn from Covid-19, and ensuring the UK is prepared for the next disaster. https://www.cser.ac.uk/resources/risk-management-uk/ We welcome this progress, and hope that the Government will continue to engage academic and risk management experts as they develop this new risk management infrastructure. We also hope that the Integrated Review's approach of setting out a ten-year strategy based on broad public consultation becomes a more regular part of the policymaking process. There should be a legal requirement for each new government to confirm their long-term vision and strategy in this way, including their strategy to build resilience to extreme risks. Governments also need to be held accountable for delivering on these strategies, through mechanisms such as annual reviews, clearly-defined performance indicators, and the oversight of a designated Minister and Select Committee. \n Key risk management policy recommendations \n Improve extreme risk assessment and ownership across Government by updating the NSRA, applying the 'three lines of defence' model to risk management and installing a Chief Risk Officer (estimated cost: £8.3 million annually) Why this matters: As we have seen from Covid-19, there are clear lessons to be learned about how to better assess and manage risks. A clearly defined single point of accountability in Government for risks will help transform the UK's resilience to future extreme risks. Improving the Government's approach to risk management will need strong leadership from the centre of Government, along with iterative work between policy officials, politicians, and risk experts from a range of sectors. The NSRA review provides scope to ensure that extreme risks are given appropriate focus within that document, and the Integrated Review sets out a promising long-term vision for national resilience. The next step is to set out a plan for achieving this vision which has extreme risks firmly embedded within it. We suggest a plan below based on the 'three lines of defence' structure, which is standard practice across industry. The plan sets out the ideal implementation model, though this is flexible and alternative options are possible depending on the Government's appetite for large-scale change in this area. Proposed model for the Three Lines of Defence approach to risk management a) The first line of defence: Strengthening departments' ability to manage extreme risks Why this matters: Government departments should be responsible for the day-to-day management of extreme risks relevant to their department. This is currently done quite effectively for non-extreme risks, but much less so for extreme risks. Risk Ownership Units covering both extreme and non-extreme risks would ensure that a culture of risk ownership is championed across Government and help policymakers consider extreme risks when making policy decisions. We recommend an assessment of which Government departments are best suited to manage individual extreme risks. Once complete, we suggest building up Risk Ownership Units of between two and six civil servants. These would be housed in \n Oversight Committee \n NATIONAL EXTREME RISKS INSTITUTE Independent audit and advisory function, submitting its recommendations on extreme risks to the Office for Risk Management and its Chief Risk Officer \n OFFICE OF RISK MANAGEMENT AND CHIEF RISK OFFICER The Chief Risk Officer would provide a single point of accountability for ensuring the proper management of risks and vulnerabilities across Government. \n RISK OWNERSHIP UNITS IN GOVERNMENT DEPARTMENTS Responsible for the day-to-day 'ownership' of extreme risks and vulnerabilities relevant to that Department.   ' 26  Ministers would continue to be held accountable for the risks designated to their departments. For example, an Electrical Grid Risk Unit might sit in BEIS. It would focus on boosting the resilience of the UK's electrical grid against extreme terrestrial and solar storms, humanmade electromagnetic pulses and malicious digital intrusions. These Units must be completely embedded in their departments and seen as part of those departments rather than extensions of the second line of defence (see immediately below). The relevant minister should also fully support the Units and understand their purpose. \n b) The second line of defence: Creating a new Government Office of Risk Management, headed by a Chief Risk Officer Why this matters: The Civil Contingencies Secretariat provides a good risk identification function, but as far as we are aware there is currently limited focus on extreme risks, and no cross-government accountability mechanism to ensure action is taken to check the quality and viability of risk planning and mitigation strategies. Without this, there is a strong chance that the UK will not be wellprepared for future extreme risks. A new Government Office of Risk Management, headed by a Chief Risk Officer (CRO) with specialist risk management expertise, would help bring the UK into line with current best practice from industry and elsewhere. This Office would ideally be an extension of the current Civil Contingencies Secretariat. However, other arrangements would also work. Its responsibilities would include: • Having overall responsibility for risk management across Government. • Having powers to assign responsibility for risks to ministers and hold them to account for their risk response strategy. combined with a rating of vulnerability (not just likelihood). The assessment should examine the strength of existing mitigations and crisis management capabilities, how external the threat is, and its velocity should it occur. This vulnerability assessment helps identify further mitigations required and actions to be taken by relevant risk owners. \n • Implementing the recommendations of the proposed new National Extreme Risks Institute (see the recommendation below). c) The third line of defence: Establishing an independent National Extreme Risks Institute. Why this matters: There is currently no government body which focuses exclusively on extreme risks. This would be the first UK public body to be exclusively incentivised to focus on public sector decision-making around dealing with extreme risks, many of which are not clearly under the management of any particular Secretary of State. A National Extreme Risks Institute would be tasked with providing independent advice on assessing and red-teaming the Government's approach to identifying and preparing for extreme risks, and making recommendations to the UK Government for how it can improve its management of these risks. This Institute would focus on identifying and supporting Government efforts to boost resilience to catastrophic events, as promised in the Integrated Review. It could be created as a What Works Center and part of the broader What Works Network.  27  The Institute's role would include: • Carrying out independent, evidence-based assessments of extreme risks. This would allow for a greater focus on risks that are low-probability but highly destructive. It would mirror Switzerland's approach 28 of an independent institute offering a separate perspective on risks to that which exists in Government, thereby reducing the chance of groupthink. • Carrying out issue-specific risk assessments to audit and 'red-team' Government departments in areas of particular concern. • Submitting recommendations to the Government and to a new Government Office of Risk Management, which would oversee the identification, assessment, and mitigation of risk. • Collating and presenting research on extreme risks on risk management policy to decision makers. • Identifying and highlighting extreme risks that are not clearly under the management of any particular Secretary of State. • Issuing a flagship report alongside each National Security Risk Assessment and Spending Review. The Institute should have independence from the Government and be accountable to Parliament. It should ideally be funded by way of endowment to protect it from costcutting exercises in future Spending Reviews. The Government should ensure that the Institute's expert staff can access relevant confidential information by providing security clearances to staff. \n Why the 'three lines of defence'? One perceived drawback of a 'three lines of defence' structure is that it risks creating a siloed Government institution. But without a new CRO, there is no single point of responsibility for risk management. This means that it tends to be deprioritised amidst the 'tyranny of the urgent'. And without the 'three lines' structure set out above, checks and balances are lacking, and risk owners don't get held to account to mitigate or plan for their risks. The problem we are seeking to solve is less one of coordination, and more one of establishing clear accountability. A lighter-touch option would be expanding the remit of the National Audit Office to include the 'third line of defence'. This would retain the audit function of the third line but lack the extreme risks expertise that an Institute would provide. To ensure proper coordination, we would also recommend an Oversight Committee. The Committee would bring the three lines of risk infrastructure together, with the CRO reporting to its Chair, as well as to the appropriate departmental head (e.g. the Permanent Secretary). It could be chaired by the Institute's head or by a parliamentarian to provide independent oversight. For more information, see the Future of Humanity Institute's proposal for a new 'three lines of defence' approach to UK risk management (2021). 29 29 https://www.fhi.ox.ac.uk/wp-content/uploads/three_lines_defence.pdf \n Lead the way to ensure global resilience to all extreme risks -not just pandemics -post-Covid-19 Why this matters: The UK cannot address transnational challenges alone. But it can use its position as a diplomatic superpower to lead the journey towards global resilience to extreme risks. As countries begin to form their longer-term policy responses to Covid-19, there may never be a better moment to put extreme risks at the top of the international agenda and go beyond simply 'preparing to fight the last war' of pandemic preparedness. During its G7 presidency and in other international fora, the UK should make the economic case to its counterparts for improving international extreme risk governance and preparedness. We set out a number of international asks below. \n a) Encouraging international long-term spending commitments on extreme risks, such as spending a target percentage of GDP on extreme risk management This could mirror the NATO agreement to spend 2% of Gross Domestic Product on Defence, or the OECD commitment to spend 0.7% of Gross National Income on International Development. Commitments should be made now to ramp up spending on catastrophic risk prevention once the fiscal situation allows. This would ensure that the necessary spending to provide global resilience to extreme risks is locked in, regardless of whether political and public attention fades over time. b) Pushing for an international agreement on pre-agreed finance and a \"Crisis Lookout\" function More than 40 leaders from across the humanitarian, development and private sectors have signed a Statement of Support for the Crisis Lookout campaign. This statement calls on G7 leaders to start a new global approach for predicting crises, preparing for them, and ensuring more people are better protected. By making pre-arranged finance the primary way to pay for crises by 2030, the G7 could ensure that funding gets to where it is needed faster and with greater impact. A global Crisis Lookout function would also better synthesise, prioritise, quantify and communicate all crisis risks and their potential costs. It would also ensure better global responses to disasters by identifying financial protection gaps at local, regional and global levels. This work would focus initially on a group of the most vulnerable countries to pilot better prediction of (and coordinated protection from) crises. The UK Government should lead calls for its adoption this year. \n c) Creating and then leading a global extreme risks network If the UK shows domestic leadership by introducing a Government Chief Risk Officer, it can lead the way internationally, too. Having announced its intention to create a new 'three lines of defence' risk management structure (see the recommendation immediately above this one), the UK could then call on other countries to do the same. It could encourage annual meetings between Chief Risk Officers from around the world, where countries share information and learn from each other's risk assessments. This would include exploring why some countries were better prepared for Covid-19 than others, and ensure that all national risk assessments and foresight programmes draw on global expertise. The Bank for International Settlements, which facilitates Central Bank coordination, is a useful equivalent organisation in the finance sector. \n d) Calling for a Treaty on the Risks to the Future of Humanity Some serious risks, like climate change or nuclear weapons, are covered by at least some international law, but there is currently no regime of international law in force that is commensurate with the gravity of risks such as global pandemics, or that has the breadth needed to deal with the changing landscape of risks. A new Treaty on Risks to the Future of Humanity has been recommended by Guglielmo Verdirame QC.  30  He argues that such a Treaty would provide a framework for identifying and addressing such risks, and that international diplomacy and domestic politics must be engaged at the highest level to achieve it. A new Treaty should have a series of UN Security Council resolutions to place this new framework on the strongest legal footing. The UK could take a global leadership position on this issue by starting to build an alliance towards a treaty with like-minded countries, such as Australia, Japan and New Zealand. https://unherd.com/2020/04/for-china-a-legal-reckoning-is-coming/ 3. Normalise red-teaming in Government, including creating a dedicated red team to conduct frequent scenario exercises (estimated cost: £800k annually) Why this matters: It is vital to scan for and discover vulnerabilities to UK infrastructure, thus avoiding and anticipating as many disasters as possible. A team recruited with the skills to do this well can reduce groupthink and question key assumptions. The Government's Integrated Review sets red-teaming as a \"reform priority\", highlighting the need to \"foster a culture that encourages more and different kinds of challenge, further developing capabilities such as redteaming to mitigate the cognitive biases that affect decision-making\". 31 Not only does the Integrated Review include red-teaming and challenge as a reform priority; it ran its own red-teaming exercise to \"challenge and test emerging thinking from the perspective of third parties\", which is to be commended.  32  We welcome red-teaming being used when developing policy, particularly around civil and defence risks. We recommend maintaining a red team of seasoned experts with the relevant background checks and security clearances, tasked with running scenario exercises and then implementing the recommendations from their findings. The red team would focus on scenarios such as: • A major cyber attack on UK infrastructure. • The non-availability of one or more major cloud providers in the UK for an extended period of time. • An accidental or deliberate release of a virus.  33  • Cut-off from the internet for an extended period of time. This would help ensure that the most important scenario exercises are conducted frequently and that clear lessons learned are 'owned' by senior policy makers. The results could be reported to a designated body tasked with ensuring that the findings are used, such as to Parliament, a new Government Office of Risk Management or National Extreme Risks Institute. We would also sound a note of caution on red teaming: • For certain issues (such as biosecurity), red teaming can result in the identification of new vulnerabilities, the publication of which should be carefully thought through. • It may also fail to meet its objectives whilst providing a false sense of security around a particular project, especially if red teaming is not undertaken by people with knowledge of the institutions or issues under question. More generally, the Government should broadly ensure that civil servants are skilled in the art of decision-making in situations of high uncertainty, and can apply a range of appropriate tools beyond red-teaming, including robust decision-making, cluster thinking, exploratory thinking, scenario planning, and adaptive planning. When key decisions are made in Government -such as decisions relating to management of extreme risks -it is essential that a variety of different approaches are used to evaluate those decisions. \n Revise the Green Book's discount rate and ensure the Treasury adopts key recommendations on intergenerational fairness Why this matters: The policy decisions we make today have huge implications for future generations, yet are often overly influenced by short-term pressures. Our children and grandchildren, as well as current generations, deserve to be treated equitably, and we need to consider the long-term consequences of today's policy decisions. Certain technical changes to Treasury processes would significantly improve incentives for decision makers to act in the interests of the long term, which would in turn improve management of extreme risks. The Institute for Government has noted that within the Treasury \"there is too little focus on the long term and on the trends -and foreseeable problems -which may affect these plans.\" We therefore recommend that the Treasury revise the Green Book (HM Treasury's guidance on how to appraise policies, programmes and projects) and the discount rate. Discounting is a way of comparing costs and benefits with different time spans. The Treasury uses a discount rate of 3.5% for future costs and benefits, in part to adjust for 'social time preference' (i.e. the value society attaches to present, as opposed to future, consumption). 34 We recommend lowering the Green Book's discount rate to ensure that today's policies do not make future generations disproportionately worse off. The discount rate should decline more quickly in the long run, the 'pure time preference' part of the discount rate should be set at 0%, and the Green Book should acknowledge that the current discount rate formula does not work for estimating the costs of significant disasters (for instance, because they could lead to significant economic decline). The Green Book should also have more detail on how to account for second-order effects (the further consequences of an action, beyond the desired initial consequence). Why this matters: The procurement and development of defence systems that integrate increasingly capable AI, machine learning and autonomy is vital to national security. But as this area grows in importance and complexity, ensuring the good governance, safety, and security of these systems becomes ever more important to avoid vulnerabilities or accidents that could harm service people or citizens or lead to inadvertent escalation. \n Establish a new The Defence Safety Authority has a number of sub-agencies that ensure the safety and good governance of risks such as Land (DLSR), Ordnance and Explosives (DOSR), Medical Services (DMSR), and Nuclear (DNSR). A new Defence Software Safety Authority would be tasked with regulating the safety of defence systems that integrate increasingly capable AI, machine learning and autonomy. This should involve adopting a new regulation in the form of a Joint Service Publication. This would require a targeted increase in funding for additional hiring, and training to judge the limitations, risks, and overall safety and security of new defence systems. Key priorities when procuring these systems include improving systemic risk assessment in defence procurement and ensuring clear lines of responsibility so that senior officials are held responsible for errors caused in the defence procurement chain. 35 \n Fund a comprehensive evaluation of the actions required to increase the resilience of the electrical grid Why this matters: The electrical grid is at risk from an array of both natural and human-made threats, any one of which could cause widespread disruption and thousands of avoidable deaths. If the grid is damaged or disabled, perishables such as food and medicine will expire, communication networks will collapse, oil and gas distribution will halt, water purification and distribution will cease functioning, and effective governance will likely disappear. In the worst-case scenario, nuclear reactors will also melt down. The grid's ability to withstand the impact of these threats is a major concern for national security and the ability to maintain basic services for the larger population. There is currently a window of opportunity to make Britain a global leader in electrical grid resilience and preparedness for the next century of risks. The grid is the country's largest machine, and it's also 100 years old. It is currently undergoing the process of being completely changed by renewables. As this transition occurs, we have the opportunity to upgrade the physical and digital integrity and resilience of the grid, and we can do it in a way that also helps promote clean energy generation and a Green recovery. This effort should produce specific policies, procedures and technological solutions, together with implementation timelines and an estimate of required resources. It should include action plans in the following areas: • Improving the UK's ability to identify threats and vulnerabilities: Produce standards and guidelines for threat identification and emergency response planning and preparation which are accepted and implemented by the energy sector. • Increasing the ability to protect against threats and vulnerabilities: Establish a nationwide network of resiliency test platforms that are long-duration, blackoutsurvivable microgrids. These should be located in facilities controlled by the Government, in stable areas that are free from flooding, severe weather and other high-impact disasters. • Improving recovery capacity and time: Design ultra-secure, low-power, self-healing wireless networks capable of bypassing compromised network components while maintaining essential connectivity to critical grid assets. This should be designed to preserve fail-safe operations that engage within minutes of a cyber attack. Integrated Review: p39. As the Integrated Review recognises, science and technology plays a crucial role in our national resilience and in our ability to address transnational challenges. Unfortunately, publicly funded research on extreme risks remains significantly underfunded compared to its importance. Total funding on AI safety, for example, is significantly less than the total funding going into private investment to accelerate its capabilities.  38  This could significantly impair our ability to respond to future crises. \n Key extreme risks research policy recommendations 1. Create a pool of machine-learning-relevant computational resources to provide free of charge for socially beneficial applications and AI safety, security, and alignment research (estimated cost: approx. £35 million per year) Why this matters: Access to large amounts of AI computational resources ('compute') -for instance, computing clusters of machine learning-optimised computer chips -is critically important for AI safety and to maintain UK scientific and economic leadership. Many recent machine-learning breakthroughs and expected advances in this area are reliant on large compute budgets that are currently beyond the reach of academia and civil society. This has led to research being skewed towards shortterm aims rather than developing socially beneficial applications or AI safety, security and alignment. We recommend creating a 'compute fund' to provide free or subsidised computation resources to researchers working on socially beneficial AI applications or AI safety, security and alignment. This idea is currently being investigated in the United States.  39 40  This would benefit critical research in the following areas: 38 https://aiimpacts.org/changes-in-funding-in-the-ai-safety-field/ \n 39 The National Defence Authorization Act 2021 includes a provision that a National AI Research Resource Task Force should investigate setting up a compute fund or cluster. A summary of the relevant provisions is available here. \n 40 The US National Security Commission on Artificial Intelligence made a similar recommendation in its recent report: \"To bridge the \"compute divide,\" the National AI Research Resource would provide verified researchers and students subsidized access to scalable compute resources, co-located with AI-ready government and non-government data sets, educational tools, and user support. It should be created as a public-private partnership, leveraging a federation of cloud platforms.\" See p191: https://www.nscai.gov/wp-content/uploads/2021/03/ Full-Report-Digital-1.pdf  \n Invest further in AI safety R&D Why this matters: Promoting technical AI safety research is critically important -not only due to the dangers of unsafe systems, but because it will bolster the UK's competitiveness as states and companies seek to acquire safe and beneficial AI systems. We strongly recommend the UK increase its funding for technical AI safety research. This could be done via the Alan Turing Institute, through ARIA, or through the new autonomous systems research hub at Southampton University. Funding should be made available for four broad areas of research: • Alignment: For very capable AI systems, pursuit of an incorrectly specified goal would not only lead an AI system to do something other than what we intended, but could lead the system to take harmful actions. Can we design training procedures and objectives that will cause AI systems to learn what we want them to do? • Robustness: As AI systems become more influential, reliability failures could be very harmful, especially if failures result in an AI system learning an objective incorrectly. Can we design training procedures and objectives that will cause AI systems to perform as desired on inputs that are outside their training distributions or that are generated adversarially? • Interpretability: If an AI system's internal workings could be inspected and interpreted, we might be able to better understand how its models work and why we should or should not trust the model to perform well. Interpretability could help us to understand how AI systems work and how they may fail, misbehave, or otherwise not meet our expectations. • Assurance of deep learning systems: It is not enough for models to be robust. In particularly high-stakes situations, we also need assurance that this is the case. However, traditional methods for gaining such high assurance typically cannot be applied to deep-learning AI systems. New methods need to be developed. As mentioned in the Biosecurity section of this report, one promising area for development is in the field of metagenomic sequencing -a potentially game-changing process which can identify new, unexpected pathogens in the first few patients who become infected. This would allow for a much earlier targeted response to an outbreak of an unknown virus. \n Invest further in applied biosecurity R&D This significant investment in biosecurity R&D would shore up Britain's status as a biosecurity world leader, as showcased by the Oxford / AstraZeneca vaccine, and make future breakthroughs more likely. It would also help keep pace with developments in the United States, where a Bipartisan Committee recently proposed an Apollo Program for Biodefense, which includes recommendations for long-term multi-year funding in this area. 41 We would also recommend increased caution before undertaking 'gain of function' research. In the context of biosecurity, this is a type of research that aims to increase the virulence and lethality of pathogens and viruses. Such research can improve scientific understanding in an important area, as it helps better understand how natural pandemics could evolve to become worse. However there are important risks to factor into any cost-benefit analysis of gain of function research. For instance, the newly acquired information generated by the research could be obtained and misused by hostile actors. Alternatively, newly created pathogens and 41 https://biodefensecommission.org/reports/the-apollo-program-for-biodefense-winning-the-race-against-biological-threats/ constitute the 'first line of defence. \n Defence Software Safety Authority as a sub-agency of the Defence Safety Authority, to protect UK defence systems from emerging threats (estimated cost: £5 million annually) \n \n \n \n \n save not only countless lives, but hundreds of billions of pounds. 6 To do justice to the seriousness of these risks and the unsustainably high level of risk we currently face, the Government must go beyond simply 'fighting the last war' and focusing solely on better preparedness for naturally occurring pandemics like Covid-19. 2 https://www.repository.cam.ac.uk/handle/<PHONE>;jsessionid=FC51BC52C0FFF10D3C39534F57AB2DF1 3 https://www.cser.ac.uk/resources/global-catastrophic-risks-2017/ 4 https://www.bloomsbury.com/uk/the-precipice-9781526600219/ 5 The Office for Budgetary Responsibility's forecast for the 2020/21 UK deficit changed dramatically in the wake of Covid-19, increasing from £55 billion in March 2020 to £394 billion in November 2020. The Institute for Government describes this £339 billion difference as a way of calculating the UK's \"cost of coronavirus so far.\" https:// www.instituteforgovernment.org.uk/explainers/cost-coronavirus INTRODUCTION which will \n It needs to transform the UK's resilience to extreme risks across the board. The UK is already an academic leader in the field of extreme risks, with world-class organisations such as Oxford University's Future of Humanity Institute 7 and Oxford Martin School 8 , Cambridge University's Centre for the Study of Existential Risk 9 and Centre for Risk Studies 10 , and Warwick University's Anthropogenic Global Catastrophic Risks project. 11 The \n UK Government now has the perfect opportunity to match academic excellence in extreme risks with policy leadership. \n the UK from a broader range of biosecurity threats. We currently run the risk of simply 'fighting the last war' and preparing for the previous pandemic instead of what might come next. The UK has made this mistake in the past: pandemic preparedness planning prior to Covid-19 presumed influenza, and was wrong-footed when confronted by a coronavirus. The UK's pandemic strategy did not include any plans for a lockdown, despite this having become one of the dominant strategies for responding to Covid-19. 1 https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/730213/2018_ UK_Biological_Security_Strategy.pdf 2 https://publications.parliament.uk/pa/jt5801/jtselect/jtnatsec/611/61110.htm#_idTextAnchor076 3 https://committees.parliament.uk/publications/4870/documents/49008/default/ 4 Integrated Review, p93-94 \n Such monitoring will help inform future AI policy and regulation to help manage the societal implications of AI. It will also mitigate the risks of increasingly widely deployed AI applications in critical areas. The work would complement and work closely alongside initiatives like the Centre for Data Ethics and Innovation's AI barometer, the OECD AI Observatory 18 and Stanford's AI Index initiative.19     14https://www.gchq.gov.uk/files/GCHQAIPaper.pdf 15 https://www.gov.uk/guidance/understanding-artificial-intelligence-ethics-and-safety https://www.gov.uk/government/publications/uk-national-data-strategy/national-data-strategy#data-3-5 16 https://ico.org.uk/about-the-ico/news-and-events/ai-auditing-framework/#!; https://ico.org.uk/media/for-organisations/guide-to-data-protection/key-data-protection-themes/explaining-decisions-made-with-artificial-intelligence-1-0.pdf \n • Creating specific technical roles in key departments, including the Ministry of Defence, the Information Commissioner's Office, the Home Office and the Department for Business, Energy, and Industrial Strategy. These roles would be targeted at experts in AI, machine learning and cyber security, and would focus on assuring the safety and security of AI systems that are deployed in specific sectors, particularly those that serve critical functions to society (such as critical infrastructure, law enforcement, finance, and defence).• Setting up a fund that agencies can apply for to cover the salaries of additional technical experts where necessary. • https://hai.stanford.edu/research/ai-index-2019 20 https://www.techcongress.io/ \n Providing funding for existing civil servants to develop training and expertise in AI and machine learning. The Treasury currently provides scholarships for civil servants to study economics; an equivalent scheme should be devised for AI. 3. \n Ensure that the UK Government does not incorporate AI systems into NC3 (nuclear command, control, communications), and that the UK leads on establishing this norm internationally Why this matters: As evidenced by the sobering history of nuclear near misses, introducing AI systems (or automation) into NC3 increases the risk of an accidental launch, without proportional benefits. \n • Playing a leadership role in ensuring that risk planning, risk mitigation, and risk preparedness improves across Government. This would include ensuring that Departmental risk plans are fit for purpose and providing a body of expertise who can support Departments with risk planning. • Playing a leadership role in ensuring that risk management improves globally. • Running regular vulnerability assessments. Calibration of risk severity should be 26 Our initial assessment suggests eight new Risk Ownership Units: https://www.fhi.ox.ac.uk/wp-content/uploads/ three_lines_defence.pdf. \n • Human-centric and beneficial AI applications: e.g. medical research and diagnostics, energy optimisation and decarbonisation, and AI for the Sustainable Development Goals. • AI safety research: e.g. into areas such as transparency, privacy, accountability, technical robustness, and fairness. • Providing open-source alternatives to commercial AI models (i.e. non-proprietary and freely available). • Increasing scrutiny: enabling the scrutiny, auditing and certification of commercial and government AI systems. • Using AI to test AI: using AI to rigorously identify, test and eliminate potential bugs, hazards, and failures. \n interventions to tackle a novel biological threat can either be rapidly deployed (e.g. non-pharmaceutical interventions) or highly effective (e.g. vaccines), but not both. Why this matters: With investment in highly promising horizon technologies, the UK's research community could dramatically improve how well and how quickly the UK can respond to new biological threats. Innovative technologies can help close this gap, and should be urgently prioritised for development. Currently, \n\t\t\t https://policyexchange.org.uk/wp-content/uploads/Visions-of-Arpa.pdf \n\t\t\t https://www.gov.uk/government/groups/synthetic-biology-leadership-council \n\t\t\t https://www.securedna.org/main-en \n\t\t\t https://genesynthesisconsortium.org/wp-content/uploads/IGSCHarmonizedProtocol11-21-17.pdf \n\t\t\t https://biofoundries.org/about-the-gba \n\t\t\t https://www.oecd-ilibrary.org/governance/national-risk-assessments_9789264287532-en \n\t\t\t https://www.gov.uk/government/collections/the-green-book-and-accompanying-guidance-and-documents \n\t\t\t https://www.cser.ac.uk/resources/written-evidence-defence-industrial-policy-procurement-and-prosperity/ \n\t\t\t https://www.sas.upenn.edu/tetlock/ \n\t\t\t https://globalchallenges.org/initiatives/analysis-research/reports/", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/e40baa_c64c0d7b430149a393236bf4d26cdfdd.tei.xml", "id": "64bd0158c2e98eb0b85785b97a9bb0d6"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "is a research organization focused on studying the security impacts of emerging technologies, supporting academic work in security and technology studies, and delivering nonpartisan analysis to the policy community. CSET aims to prepare a generation of policymakers, analysts, and diplomats to address the challenges and opportunities of emerging technologies. CSET focuses on the effects of progress in artificial intelligence, advanced computing, and biotechnology.", "authors": ["Ryan Fedasiuk", "Jennifer Melot", "Ben Murphy"], "title": "Harnessed Lightning HOW THE CHINESE MILITARY IS ADOPTING ARTIFICIAL INTELLIGENCE", "text": "iii rtificial intelligence (AI) is progressing at lightning speed. What 10 years ago would have been considered science fictionself-adapting computer algorithms with billions of parametersis now a central focus of military and intelligence services worldwide.  1  Owing in part to AI's fast-paced development, most analyses of its military promise tend to focus more on states' future aspirations than present-day capabilities. This is particularly true for the Chinese People's Liberation Army (PLA), which has routinely made clear its desire to harness AI for military advantage, and which prefers to keep a close hold over its actual, technical capabilities.  2  But as tensions mount between the United States and China, and some experts warn of an impending crisis over Taiwan, it is crucial that U.S. policymakers and defense planners understand the commercial off-the-shelf (COTS) AI technologies already available to the Chinese military.  3  This report offers a detailed look at the PLA's adoption of AI by analyzing 343 AI-related equipment contracts, part of a broader sample of more than 66,000 procurement records published by PLA units and stateowned defense enterprises in 2020. The report identifies key AI defense industry suppliers, highlights gaps in U.S. export control policies, and contextualizes the PLA's AI investments within China's broader strategy to compete with the United States. Key findings include: 1. Chinese military leaders are already procuring AI-related systems and equipment to prepare for \"intelligentized\" warfare, but AI so far represents a small fraction of overall purchasing activity.  4  Executive Summary \n A Center for Security and Emerging Technology iv • Approximately 1.9 percent of public PLA contracts awarded between April 2020 and November 2020 are related to AI or autonomy. • While we can only estimate a floor for Chinese military AI spending, it is likely that the PLA spends more than $1.6 billion each year on AI-related systems and equipment. • The PLA seems most focused on procuring AI for intelligence analysis, predictive maintenance, information warfare, and navigation and target recognition in autonomous vehicles. • Whereas some PLA officers have expressed serious reservations about developing lethal autonomous weapons systems (LAWS), laboratories affiliated with the Chinese military are actively pursuing AI-based target recognition and fire control research, which may be used in LAWS. 2. Chinese leaders view AI as the key to transforming the PLA into a \"worldclass,\" globally competitive military force. PLA advances in AI and autonomy will create new vulnerabilities for the United States and allied forces operating in the Indo-Pacific. • The PLA hopes to use AI to generate asymmetric advantages vis-à-vis the United States, which it regards as a \"strong enemy\" (强敌), but also a role model for AI development. • PLA units and military laboratories are focused on developing autonomous vehicles and surveillance systems in the undersea domain, where the United States has traditionally had a significant advantage. • The PLA is stepping up investment in information operations and adaptive radar systems to jam and blind U.S. sensor and information networks, which PLA leaders judge to be particularly vulnerable. • To compensate for vulnerabilities in its own networks, the PLA may adopt edge applications of AI (directly proximate to, or embedded within a platform) that can operate semi-or fully autonomously. 3. China's military-civil fusion (军民融合) development strategy is helping the PLA acquire COTS technologies, both from private Chinese technology companies and sources outside of China. • Most of the PLA's AI equipment suppliers are not state-owned defense enterprises, but private Chinese tech companies founded after 2010. Center for Security and Emerging Technology v • Although most suppliers are not state-owned, many have benefited from equipment, personnel, information, or capital provided directly or indirectly by the state. • Of the 273 PLA AI equipment suppliers identified in this study, just 8 percent are named in U.S. export control and sanctions regimes. • Some Chinese suppliers make a business out of sourcing foreign data or components and reselling them to sanctioned Chinese defense companies and PLA units. • Lapses in due diligence and situational awareness may permit the Chinese military and defense industry to access capital and technology originating in the United States and partner nations, including advanced computer chips. Supported by a burgeoning AI defense industry, the Chinese military has made extraordinary progress in procuring AI systems for combat and support functions. Within the next five to 10 years, the PLA will likely continue investing in AI to disrupt U.S. military information systems and erode the U.S. advantage in undersea warfare. Although PLA investment in the technology appears substantial-roughly equivalent to that of the U.S. military-it remains to be seen how exactly AI might alter the balance of military power in the Indo-Pacific. In addition to renewed interest in counter-autonomy research, U.S. and allied efforts to regulate access to semiconductor devices may hinder the utility and availability of AI systems for the Chinese military. Introduction n the early pages of Liu Cixin's science fiction novel Ball Lightning, when asked what, exactly, goes on at System Review Department No. 2, Dr. Lin Yun replies that she develops \"new concept weapons\" (新概念武器)-fantastical ideas with the potential to change warfare itself. As the mastermind behind the Chinese military's eventually catastrophic \"lightning weapons\" program, Lin is remembered for her ends-justify-means personality and willingness to develop dangerous weapons in the service of the state.  5  Seventeen years after it was first published in Chinese, Liu's novel has been eclipsed by his more successful series, Remembrance of Earth's Past. A Hugo Award winner, he publicly decries any comparison between his imagined universe and modern geopolitics.  6  But there can be no denying that Liu's fictional depiction of a Chinese People's Liberation Army bent on mastering \"lightning weapons\" bears striking similarities to its quest for artificial intelligence today. Chinese military leaders expect AI to fundamentally change warfare, and are leaning on the technology to transform the PLA into a \"world-class military\" by 2050.  7  AI's revolutionary potential and general-purpose application even led Andrew Ng, former chief scientist at the Chinese internet company Baidu, to label it \"the new electricity\" in 2017.  8  Despite some anxiety within the PLA about developing intelligent or automated weapons systems, concerns about technology misuse seem to take a back seat to the needs of the state.  9  The PLA's rapid embrace of AI raises questions about strategic stability and the future of warfare. While analysts generally agree that AI forms the basis of the PLA's modernization strategy, questions linger about how far it may be willing to go in devel-I oping lethal autonomous weapons systems (LAWS), and which of its new concept weapons will eventually mature into programs of record.  10  By examining 343 AI-related equipment contracts awarded by PLA units and state-owned defense enterprises in 2020, this study offers a detailed view of how the Chinese military is beginning to wield AI-and to what end. The report begins by reviewing the budgetary constraints and modernization goals that have shaped the PLA's transition to \"intelligentized\" warfare, followed by a discussion of the study's methodology and limitations. It then identifies seven primary application areas for which the Chinese military is awarding AI-related equipment contracts: autonomous vehicles, intelligence analysis, information warfare, logistics, training, command and control, and target recognition. The bulk of the report discusses common trends and significant AI purchases made within each of these fields. A fourth section profiles 273 of the PLA's known AI equipment suppliers, highlighting gaps in U.S. export control policy and prevailing technology transfer risks. Finally, the report discusses how AI fits into the PLA's broader concepts of operations, before concluding with a discussion of the policy tensions that will shape its military competition with the United States. • After being made a priority in the 1980s, mechanization sought to equip PLA units with modern platforms, including electronic warfare systems, as well as motorized, armored personnel carriers and infantry fighting vehicles. Mechanization emphasized fixed boundaries and armor operations, primarily for troops stationed along China's land borders, at the expense of naval and air operations.  15  In 2020, the PLA announced it had \"basically achieved\" mechanization.  16   \n A • Since the 1990s, the PLA's dominant push has been informatization, in which wars are won through information dominance, and the space and cyber domains are the \"commanding heights of strategic competition.\"  17  PLA operational concepts today emphasize the need to win \"informatized local wars\" by using long-range, precision, smart, and unmanned weapons and equipment.  18  In 2020, the PLA announced its goal to become a \"fully mechanized and informatized\" force by its centenary, the year 2027.  19  • First mentioned in China's 2015 Defense White Paper, intelligentization represents \"a new round of military revolution\" characterized by networked, intelligent, and autonomous systems and equipment. It endeavors to build on mechanized and informatized systems, creating \"ubiquitous networks\" in which \"'human-on-human' warfare will be replaced by 'machine-on-human' or 'machine-on-machine warfare.'\"  21  In particular, AI forms the basis of the PLA's push toward intelligentization, and tops the list of emerging technologies prioritized in recent Chinese strategy documents and development plans.  22  Although these modernization goals represent \"phases\" of development, there is significant overlap between them: \"While some units of the PLA employ data links, network-centric sensor-to-shooter system-of-systems, and field a variety of UAVs, electronic warfare platforms, and advanced combat capabilities,\" writes Dean Cheng, \"other units are still in the midst of simply shifting from towed artillery to self-propelled guns, improving their main battle tanks and becoming fully motorized.\"  23  Since 2013, however, the PLA has placed significantly less emphasis on mechanization and informatization, and is starting to phase in intelligentization as a guiding concept, as evidenced in the 2015 and 2019 defense white papers (Figure  1 ).  24  At the same time, numerous reforms to China's military and defense industry have sought to streamline the PLA's promotion of science and technology and acquisition of intelligent equipment. Some of the most significant have included the creation of an Equipment Development Department alongside an entirely new service branch, the PLA Strategic Support Force (PLASSF).  25  Renewed emphasis on military-civil fusion (MCF; 军民融合), too, has expanded the PLA's access to private-sector innovation and enabled it to draw on the work of internet giants like Baidu, Alibaba, and Tencent, and telecom giants like Huawei and ZTE.  26  But even with reforms, contemporary scholars have questioned whether China's historically bloated and inefficient defense industry can sufficiently adapt to the information age.  27   Methodology and Scope 2 rocurement information holds distinct advantages for those looking to understand the Chinese military and its immediate capabilities. First, defense contracts offer strong signals of both intent and capability, as militaries willing to spend limited resources on commercial off-the-shelf (COTS) solutions clearly deem them useful. Second, as it does in the United States, the public procurement process shapes China's ability to acquire and harness AI for military advantage.  28  To assess how China is adopting AI, this report analyzes a sample of purchasing information published directly by the Chinese military in 2020. In addition to capabilities identified in procurement contracts, the authors draw on theoretical writings and research papers by PLA officers and defense industry engineers to assess how the PLA may use the AI systems it is purchasing, and how these systems fit into its concepts of operations. CSET's corpus of PLA procurement tenders spans 66,207 records published between March 30 and December 1, 2020. These tender notices run the gamut from technology requirements and requests for proposals (RFPs) to announcements of equipment or software contracts that were awarded to Chinese companies. Different types of procurement information reflect different steps in the PLA's technology acquisition process: • Requirements, inquiries, and bid solicitations signal demand. They reflect the PLA's technological priorities and perceived gaps that research institutions and service branches are trying to fill.  29  • Contract awards signal supply and, ultimately, capability. They represent weapon systems or components the PLA sought to acquire and ostensibly received. In most cases, several companies compete to win a contract through a competitive bidding pro-P cess.  30  More rarely, the PLA selects a single vendor as a contract recipient without considering alternative suppliers (\"sole source\" procurement). Of the 66,207 tenders in the CSET dataset, 21,086 announce contracts to supply the PLA with equipment, including software and electronic components.  31  Information about many of these purchases is limited. While 49,493 tenders in the dataset are publicly available (公开), the rest are classified as confidential (秘密; 14,024 tenders) or secret (机密; 2,270 tenders).  32   \n UNDERSTANDING PROCUREMENT IN THE PLA This study primarily considers contract awards, but also makes use of RFPs and bid solicitations filed by PLA service branches, scientific papers published by research institutions, advertisements from Chinese defense companies, and theoretical articles published in outlets like People's Liberation Army Daily (解放军报). Of the 21,086 contract awards included in the CSET dataset, 18,354 are \"public\" and include information such as the requesting unit, intended end user, project budget, tendering agency, and contract winner. A full list of variables can be found in Appendix I. Most of these contracts were not awarded by PLA units, but by defense state-owned enterprises including the Aviation Industry Corporation of China (AVIC), China Aerospace Science and Technology Corporation (CASC), and hundreds of their subsidiaries.  33  Just 3,726 of the 18,354 public equipment contracts in the dataset were awarded by PLA service branches, while the remain-  The 18,354 public contracts in our dataset include all manner of supplies and equipment ranging from toilet seats and ball bearings to completed, off-the-shelf CH-4 Rainbow (彩虹) combat UAVs. We do not claim to have a complete history of the PLA's purchase records for this period of time. However, by examining trends among a sample of public contracts awarded in 2020, this paper aims to illuminate the specific types of AI-related equipment the PLA is purchasing, and to explore their potential application. \n IDENTIFYING AI AND \"INTELLIGENT\" EQUIPMENT PURCHASES Several limitations constrain our assessment of the Chinese military's AI-related procurement activity. First, definitions of \"AI\" are not consistent even within the U.S. defense enterprise. The U.S. Department of Defense's (DOD) 2020 AI strategy defines AI as \"the ability of machines to perform tasks that normally require human intelligence\"-a description that includes \"decades-old DoD AI\" such as autopilot and signal processing systems, but also modern deep learning techniques.  34  Second, Chinese defense engineers frequently conflate terminology surrounding unmanned and autonomous systems, making it difficult to distinctly analyze the latter. Finally, some project titles are ambiguous and difficult to categorize. The PLA's contract notices offer very little information about each product's technical specifications or envisioned use, though manufacturers sometimes advertise this information.  35  This study therefore adopts a broad definition of AI, including most contracts that describe \"intelligent\" systems and equipment. Despite these constraints, we identify PLA procurement projects related to AI and autonomy by using keyword searches and an AI assistant, \"Elicit.\"  36  We first searched for contracts with names that included any of 14 broad keywords: algorithm (算法), automatic (自动), autonomous (自主), autonomy (自治), intelligent (智 能), human-machine (人机), unmanned (无人), prediction (预测), artificial intelligence (人工智能), computer vision (计算机视觉), robot (机器人), intelligence (智慧), learning (学习), and the English-language abbreviation \"AI.\" Of the 18,354 public equipment contracts in the dataset, 523 contained one or more of these phrases.  37  However, some of these keywords are excessively broad, and upon closer examination, many of the projects that mention them are not strictly related to AI development. For example, \"learning\" (学习) returned contracts related to machine learning (机器学习), but also military education. \"Automatic\" (自动) and \"robot\" (机 器人) likewise returned contracts about automated manufacturing, machinery, tools, and robotic projects that likely do not feature AI or autonomy. We therefore elimi-nated 180 extraneous \"intelligent\" equipment contracts, for a total of 343 contracts related to AI and autonomy. These AI-related contracts represent 1.9 percent of PLA-wide public contract awards from April-November 2020. For a full explanation of the coding and labeling process, see Appendix I. \n PRICING CHINESE MILITARY AI DEVELOPMENT Of the 343 AI contract notices in the dataset, 205 (60 percent) listed the monetary value of the contract. Public AI contracts in the dataset were typically awarded through a competitive bidding process, as opposed to sole-source procurement. They ranged in value from $1,300 (RMB 9,000, for an intelligent sound-and-light alarm detection system) to $3 million (RMB 21 million, for an intelligent UAV data access and management platform), with the average contract amounting to $240,000 (RMB 1.7 million). These contracts are noticeably small, even when adjusted for purchasing power parity.  38  It is likely that the PLA's equivalent of major defense acquisition programs are classified or otherwise not captured in this dataset.  39  Approximately 2 percent of all public equipment contracts in the dataset appear related to AI, broadly defined. The PLA's five main service branches-as opposed to theater commands, research institutes, or defense SOEs-award the majority of public AI contracts. As in the United States, these service branches are tasked with procuring equipment used in military operations. Among the PLA Air Force, Ground Force, Navy, Rocket Force, and Strategic Support Force, approximately one in 20 public equipment contracts appear related to AI. The PLA spends more than 41 percent (approximately $86 billion) of its official $209 billion budget on equipment, and provides no additional information about how that funding is distributed.  40  If public contracts reflect how the PLA prioritizes different emerging technologies, then it is likely the PLA spends more than $1.6 billion each year on AI-enabled systems.  41  However, because it is still an emerging technology, the PLA's true spending on AI likely exceeds this number, as more funding is captured in research and development rather than off-the-shelf technology procurement. Moreover, the most resource-intensive AI projects are likely classified. For these reasons, we can only approximate a floor for Chinese military AI spending. Using a different methodology and set of source documents, past CSET analysis estimated a ceiling for Chinese defense-related AI research at \"no more than about 19 billion RMB ($2.7 billion)\" in 2018, \"and possibly much less.\"  42  Our analysis supports the conclusion that annual Chinese military spending on AI is in the low billions of U.S. dollars. Comparisons between Chinese and U.S. military spending are inherently complicated, as both countries define AI differently, discuss intelligent and autono-mous systems in different ways, publish different degrees of information about their equipment spending, count that spending differently, and use currencies with different degrees of purchasing power. However, if the PLA does spend between $1.6 billion and $2.7 billion on AI-related technologies each year, then its AI spending is likely on par with that of the U.S. military. Various analyses of DOD budgets for procurement and research indicate that it spent between $800 million and $1.3 billion on AI in 2020, with an additional $1.7 billion to $3.5 billion for unmanned and autonomous systems.  43  rior analysis has highlighted the PLA's plan to use AI in a variety of military applications, including intelligent and autonomous unmanned systems; intelligence analysis; simulation, war-gaming, and training; information warfare; and decision support.  44  Yet there are many ways to taxonomize AI applications for military use. CSET's 2020 study Chinese Perspectives on AI and Future Military Capabilities identified 12 specific AI applications of interest to Chinese military and defense industry researchers, including cybersecurity and intelligent munitions.  45  The DOD's Communities of Interest on autonomy include four AI application areas: machine perception, reasoning, and intelligence; human-machine collaboration; scalable autonomous system teaming; and test, evaluation, validation, and verification.  46  This study builds on prior taxonomies of AI applications, consolidating some fields while adding others. After reviewing the 343 AI-related contracts in our dataset, we arrived at seven discrete application areas for which the PLA is adopting AI:  \n AI Purchases by Application The resulting taxonomy is imprecise. Working with limited information, we adopted an iterative labeling process to characterize each contract, which is described in more detail in Appendix I. Contracts that did not clearly fit in any of these seven categories were marked \"Other.\" Despite some labeling uncertainty, this report finds that autonomous vehicles, ISR, predictive maintenance, and information warfare are priorities within the PLA's intelligentization strategy. To a lesser extent, PLA units also appear interested in using AI for simulation, target recognition, and command and control systems. Appendix II includes more detailed information about how each of the PLA's service branches are adopting AI, while the remainder of this section discusses each application in detail.   \n INTELLIGENT AND AUTONOMOUS VEHICLES Since it first unveiled the Wing Loong-1 combat UAV in 2009, the PLA has made significant progress in developing intelligent and autonomous systems in the air and maritime domains. Of the 343 AI equipment contracts considered in this study, 35 percent (121) are related to intelligent or autonomous vehicles. Market research firms estimate that military and security services today account for more than 40 percent of the Chinese UAV market, and procurement records confirm that several PLA units and defense SOEs purchase COTS autonomous vehicles through a public purchasing platform called the Drone Network (无人机网).  47  Most of the unmanned and autonomous vehicle contracts in our dataset are for airborne systems. While it is difficult to distinguish between contracts for remotely piloted or truly autonomous aircraft, some PLA units have funded research into autonomous flight, and others have purchased \"intelligentized\" microwave interference, reconnaissance, and data processing modules, which may be attached to remotely piloted or self-flying aircraft.  48  Myriad Chinese suppliers advertise unmanned or autonomous aerial vehicles for combat or surveillance, including private enterprises, SOEs, and state-run research and design centers. Examples of such systems include the ASN-301-a reverse-engineered copy of the IAI Harpy loitering munition  49  and the GJ-11 \"Sharp Sword\" combat UAV.  50  The Chinese Academy of Sciences (CAS) Shenyang Institute of Automation (SIA) is at the forefront of state-backed autonomous vehicle research.  51  In 2020, it was awarded contracts to supply a \"3D intelligent collision avoidance system\" for CASC, and \"intelligent self-flying machinery\" for the PLA Air Force (PLAAF).  52  The Chinese defense industry is also developing coordinated swarms of fixedwing UAVs and rotorcraft.  53  Whereas scholars in the late 2010s speculated about hypothetical swarm applications, the technology has progressed significantly, and some limited swarm applications now appear operational.  54  In 2020, multiple PLA units and CASC institutes awarded contracts for air-launched drone clusters and subsystems used in swarms, including self-organizing UAV communications systems, group node management and control software, AI-based radar coincidence imaging, and collision avoidance sensors.  55  In October 2020, the PLA Ground Force (PLAGF) placed a $900,000 order to construct \"drone swarm targets.\"  56  It is not clear from the contract what such a swarm engagement would look like in practice. The Drone Network also advertises COTS software and hardware for use in UAV group operations, such as \"SwarmLink\"-a network gateway that can support more than three hundred vehicles.  57  Beyond COTS systems, several Chinese universities have conducted swarm-related research, including Beihang University, Nanjing University of Aeronautics and Astronautics, and Zhejiang University's Institute of UAV Systems and Control.  58  Notably, multiple lines of research focus on contesting and jamming U.S. military swarm projects such as LOCUST and Gremlins.  59  Although most information about such systems is likely to be classified, some public procurement records also indicate that the PLA is purchasing unmanned and autonomous underwater vehicles (UUVs and AUVs). Five contracts in our dataset were for AUV platforms, and another contract was for an intelligent ship integration system.  60  In the summer of 2020, for example, the PLASSF placed orders for AUVs from Tianhe Defense, a company which seems to be emerging as China's national champion in A/UUVs.  61  Tianhe advertises a \"shadowless AUV solution,\" which it claims is capable of autonomously diving below 200 meters.  62  In addition to contract awards, advertisements from Chinese industry suppliers indicate that they are developing small-and medium-sized, fully autonomous underwater vehiclespossibly for sale abroad. Today, none of the top 10 companies selling AUVs in the international market are Chinese.  63  Yet several COTS AUV models are advertised on the Drone Network, and some PLA units may purchase them for undersea detection and reconnaissance. Examples include the ZF-01 AUV, which can apparently dive to 100 meters with a towed sonar array  64  ; and Kawasaki's SPICE AUV, which comes equipped with a robotic arm for underwater fiber-optic cable and pipeline inspection.  65   \n FIGURE 3 AUV Likely Ordered by the Strategic Support Force in 2020 Source: Phoenix New Media and Tianhe Defense.  66  Finally, at least four public contracts in our dataset were related to developing \"intelligent satellites\" that can autonomously adjust their orbit or engage in rendezvous and proximity operations with other space assets. For example, in August 2020, the PLA Academy of Military Sciences awarded an \"intelligent satellite simulation software\" contract to Hunan Gaozhi Technology Co., Ltd. (湖南高至科技有 限公司).  67  The company sells high-resolution cameras and intelligent video analysis servers, and holds several patents related to intelligent or automatic servo control in satellite systems.  68  In August 2020, the CASC Shanghai Academy of Space Technology likewise awarded a contract for an \"on-orbit satellite data acquisition and prediction subsystem\"; while other CASC institutes awarded contracts for intelligent or automatic inclination adjustment, high-precision attitude determination, and small satellite positioning systems.  69  Although nearly all of the vehicle contracts reviewed in this study are described as \"intelligent\" (智能), the true nature of this intelligence-and the machine learning methods that may or may not be involved in their operation-is unclear. The PLA has long procured unmanned, remotely-piloted vehicles for reconnaissance and strike missions. But in the 2020s, Chinese leaders hope that improvements in autonomous navigation and online, real-time learning will cement unmanned vehicles as the backbone of intelligentized, machine-on-machine warfare.  70   \n INTELLIGENCE, SURVEILLANCE, AND RECONNAISSANCE AI promises to revolutionize military ISR, perhaps more than any other application area. That potential is reflected in the PLA's procurement activity, as nearly one in five AI contracts canvassed in this study  (63)  appear related to ISR. U.S. military and intelligence services recognize the importance of using AI in foreign media and geospatial imagery analysis, and PLA units are adopting AI toward similar ends. According to Liu Linshan, a researcher at the PLA Academy of Military Sciences, \"battlefield situational awareness . . . includes not only the results of one's own reconnaissance, surveillance, and intelligence activities, but also massive amounts of geographic information data, human social and cultural data, and social media data\" that can be fused to improve situational awareness at all levels of operation.  71  A significant number of the PLA's AI-based ISR contracts concern remote sensing and geospatial imagery analysis, consistent with U.S. intelligence community assessments of China's space strategy.  72  Throughout 2020, CASC institutes and PLASSF units placed orders for polarized surface detection, imagery analysis, distance measurement, and multisource data fusion systems to be embedded in satellites.  73  In August, for example, the PLASSF awarded a \"geospatial information perception and intelligent analysis subsystem\" contract to Beijing Uxsino Software Co., Ltd. (\"Uxsino,\" 北京优炫软件股份有限公司).  74  The company produces data processing systems analogous to those made by Oracle in the United States.  75  CASC institutes are also developing microsatellites with edge AI information processing applications. One CASC subsidiary, Shenzhen Aerospace Dongfanghong HIT Satellite Ltd. (深圳航天东方红海特卫星有限公司), advertises a constellation of 80 \"intelligent autonomous operation and management\" MV-1 microsatellites capable of \"full color, multi-spectral, and hyperspectral imaging\" at resolutions of 1 to 5 meters.  76  In the maritime domain, the PLA is interested in using AI for underwater inspection and deep-sea sensing. State-run research institutes in this study awarded contracts for intelligent pipeline detection and identification, multisource information processing and scene analysis, and automated coordinate measuring units for underwater vehicles. In July 2020, the PLA Navy (PLAN) awarded an ocean mapping contract to Startest Marine (星天海洋), which offers several products related to undersea surveying and mapping.  77  One of its products, GeoSide1400, is a side-scanning UUV that \"uses the backscattered echoes of seabed targets for detection.\" 78 A video on Startest Marine's website depicts GeoSide1400 being towed by a fishing boat and patrolling the subsurface coastline, an application that could be used to detect U.S. undersea forces in a crisis.  79  The company also advertises services for undersea mapping and hydrological data collection depicted in Figure  4 .  Undersea Sensor Systems Offered by a PLA Contractor Source: Startest Marine.  81  Finally, the PLA is using AI in multisource data fusion for foreign military analysis, including textual analysis of foreign-language documents.  82  In the fall of 2020, the PLAGF awarded two contracts for \"foreign military equipment intelligent document data resources\"; and in November, an unspecified PLASSF unit ordered a \"multilingual intelligent text processing system\" from Nanjing Glaucus-Tech Co., Ltd. (Glaucus-Tech; 南京国业科技有限公司). On its website, Glaucus-Tech advertises the \"GL-AI Speech Recognition System 001,\" which can apparently translate foreign languages into Chinese with 80 percent accuracy, at a rate of 20 words every 150 milliseconds.  83  The company's products rely on NVIDIA processors as components, including the Tesla P40 GPU.  84  On balance, Chinese military AI contracts for ISR applications exhibit priorities similar to those stated by the U.S. intelligence community, such as network management, image classification, and transcription of foreign languages.  85  However, in part because of its lack of operational experience, the PLA has struggled to access certain types of data required for some weapons systems, such as radar signature-based target recognition.  86   \n PREDICTIVE MAINTENANCE AND LOGISTICS As in the United States, the Chinese military is using AI for equipment maintenance and logistics. Of the 343 AI contracts in our dataset, 11 percent  (38)  were related to maintenance, repair, logistics, or sustainment. Newly established PLA contractors have developed AI-based applications for leak detection, fault diagnosis, and \"smart warehouses\" intended to predict and fill orders for materiel.  87  In March 2020, for example, the Academy of Military Sciences awarded Anwise Global Technology (安怀信科技) a contract for an automated code testing platform.  88  Established in 2016, Anwise is one of China's largest intelligent equipment manufacturers, and primarily services the military aerospace and electronics industries.  89  Its products include AI-based software for soldering fault diagnosis and a virtual prototype library for testing and evaluation of aerospace weaponry.  90  In November 2020, the PLAGF and PLASSF awarded predictive maintenance contracts to Wego (威高), which produces medical devices and fault diagnosis equipment; and Sucheon Technologies (硕橙科技), which focuses on using AI in mechanical noise recognition.  91  Predictive maintenance is also emerging as an edge application for otherwise inaccessible, remotely piloted, or autonomous platforms. In 2020, multiple PLA research institutes fielded orders for automatic test equipment on satellites and underwater vehicles that cannot easily be reached for diagnostic testing and repair.  93  CASC, for example, awarded two $800,000 contracts for ATE systems on unspecified constellations of Earth observation satellites in geosynchronous and low earth orbit. PLA officers predict drastic decreases in equipment and materials stockpile requirements as a result of intelligentized maintenance and logistics systems. \"With the development of information technology such as big data and cloud computing,\" write two PLA logistics officers, \"it is not necessary to establish a large-scale resource reserve . . . all materials need only be supplied to the required place at the required time.\"  94  Consistent with this thinking, CASC and the PLA Ground Force placed orders for intelligent procurement systems for bullets and Internet of Things (IoT) devices throughout 2020.  95   \n INFORMATION OPERATIONS AND ELECTRONIC WARFARE The U.S. military considers electronic warfare to be a component of information operations, and designates information, data, knowledge, and influence as \"information-related capabilities.\"  96  Approximately 8 percent of public PLA AI contracts in this study  (29)  are related to IO, broadly defined. However, AI stands to affect each of these capabilities in different ways. Some AI projects support the PLA's public opinion warfare (舆论战) and psychological warfare (心理战) strat- Intelligent Maintenance Systems Offered by a PLA Contractor Source: Anwise Global Technology Co., Ltd.  92  egies, whereas others focus on electromagnetic spectrum dominance or defense and intrusion in cyberspace.  97  Public opinion manipulation is a longstanding focus of the PLA and the Chinese Communist Party (CCP) more broadly.  98  CSET's report Truth, Lies, and Automation found that, with breakthroughs in AI, \"humans now have able help in mixing truth and lies in the service of disinformation.\"  99  Procurement records and research papers indicate that the Chinese military is actively exploring this capability.  100  In November 2020, for example, the PLASSF's Information Engineering University awarded a contract for an \"internet public opinion AI clustering system\" to Zhengzhou Meicheng Electronic Technology Co. (郑州美诚电子科技有限公司), an electronics wholesaler. It showcases three products related to \"online behavior management\" that range in price from $6,000 to $56,000.  101  Each is a computer processor produced by Ruijie Networks (锐捷网络), advertised as being able to \"intelligently\" track source and destination IP addresses, website URL visits, and search history; and perform \"real-name online behavior auditing.\"  102  AI-based sentiment analysis software is also common among PLA units and defense SOEs. KnowleSys (乐思网 络), one of China's largest public opinion management software companies, claims Dalian Naval Academy as a customer.  103  Its products can apparently analyze trends and predict \"hotspots\" on both Chinese and foreign social media platforms. AI-enabled sentiment analysis systems like these will grow in importance as the CCP continues to expand its overseas information operations.  104  \"In the era of artificial intelligence,\" writes a professor at NUDT, \"audience information can be intelligently collected and analyzed by machines, various data about public opinion warfare opponents can also be obtained through network detection and deep data analysis; and public opinion warfare happens in real-time.\"  105  In addition to psychological operations and social media manipulation, the PLA is purchasing AI-related systems for use in electronic warfare. The PLA Navy Submarine Academy, for example, has awarded several contracts related to adaptive beamforming techniques, using AI to produce a dynamic filter that will cancel the effect of interfering signals.  106  Other PLA units awarded contracts for automatic frequency modulation, microwave jamming, broadband automatic gain control, and multisource signal separation.  107  Chinese experts broadly believe AI will revolutionize electronic warfare by replacing today's passive, adaptive technology with systems defined by more active, cognitive algorithm development.  108  Research papers published in 2020 by PLASSF Units 91404, 63610, and 93175 109 discuss using adaptive, self-correcting systems to conduct operations related to \"battlefield situational awareness, electromagnetic target reconnaissance, electronic countermeasures, electronic defense, and electromagnetic spectrum management.\"  110  Cybersecurity and network exploitation are also focal points in the PLA's adoption of AI, and are key elements of information warfare. Prior CSET research has found that \"machine learning has the potential to increase both the scale and success rate of spearphishing and social engineering attacks,\"  111  and several Chinese universities are cooperating with the PLA to advance related research.  112  Throughout 2020, PLA units and state-owned research institutes in our dataset awarded contracts for intelligent terminal inspection systems; autonomous, self-configuring software; and software control management systems.  113  In November, for example, the PLASSF awarded a contract for an AI-based \"cyber threat intelligent sensing and early warning platform\" to EverSec (恒安嘉新(北京)科技股份公司). The company serves as a national-level \"Cybersecurity Emergency Service Support Unit\" (网络 安全应急服务支撑单位) for the Chinese government's National Computer Network Emergency Response Technical Team/Coordination Center.  114  In this capacity, EverSec's role seems analogous to that of FireEye or CrowdStrike's support for the Cybersecurity and Infrastructure Security Agency in the United States.  115  Beyond its adaptive cybersecurity products, the company also advertises services for petabyte-scale data storage and processing, AI-based open source data mining, and internet blocking and censorship protocols.  116  EverSec claims that its products are used in all 31 Chinese provinces, autonomous regions, and provincial-level municipalities, and evidently PLASSF units also purchase some of these services.  117  The U.S.-based venture capital firm Sequoia Capital is an investor in EverSec.  118   \n SIMULATION AND TRAINING The PLA has long had a problem with training its enlisted service members and officers. A lack of suitable aircraft, friction when conducting joint operations, rigid organizational culture, and seasonal tides in recruitment each affect its force posture and readiness.  119  While improvements in technology can remedy only some of these issues, AI nonetheless stands to save precious time and training costs. Of the 343 PLA contracts considered in this study, 6 percent (22) concerned using AI for simulation and training. Among the PLA officer corps, war-gaming is a well-established tradition, and is growing in importance given China's relative lack of real-world combat experience. It is no surprise that the PLA has awarded contracts for proprietary, AI-based war-gaming software for use in professional military education programs. DataExa (渊亭科技), for example, advertises an AI-based war-gaming simulator called \"AlphaWar,\" inspired by DeepMind's Starcraft-playing AI system, AlphaStar.  120  The PLA's preoccupation with war-gaming grew out of the Information Operations and Command Training Department of NUDT, which created \"computer-based, warzone level, intranet-based campaign exercises\" throughout the 2000s.  121  According to Chen Hangui, a researcher at the PLA Army Command College, one of the principal uses of AI will be in \"war-game systems\" to \"more effectively test and optimize combat plans.\" 122 PLA units are also using AI in virtual and augmented reality systems to train fighter pilots and drone operators. In August 2020, for example, Naval Aviation University awarded a contract to eDong AI (翼动智能) for a \"human-machine integrated control algorithm model and simulation service.\"  124  The company primarily designs and builds VR/AR simulation centers, and also advertises stand-alone training software.  125  A similar contract was awarded to AOSSCI (傲势科技), which produces an \"X-Matrix\" UAV flight simulator for PLA pilots.  126  In June 2021, Chinese media reported that an AI system had defeated a top PLA pilot in a simulated dogfight similar to the Defense Advanced Research Projects Agency's Air Combat Evolution program.  127  Public contract awards indicate that AI-based simulation systems are becoming more common within the PLA. While AI for simulation and training promises to save military services time and resources, deep learning systems rely heavily on data, are incapable of learning common sense, and lack interpretability-limitations acknowledged by Chinese defense contractors.  128  Still, even with these limitations, it is likely that advances in AI will continue to supplement or stand in for the PLA's often cited experience gap. \n COMMAND AND CONTROL Chinese military scholars and strategists expect the speed, efficiency, and flexibility afforded by AI to revolutionize battlefield decision-making.  129  Despite the PLA's emphasis on C2, however, public procurement data does not indicate that it is a priority: Just 4 percent (15) of the contracts in our dataset appear related to C2, and most included only limited information alongside nebulous names such as \"intelligent control equipment,\" \"smart management systems,\" and \"autonomous mission planning.\" The few projects for which we could find adequate details seem primarily designed to support human decision-making processes, not replace them outright. It is likely that the PLA's most significant AI-enabled C2 projects are classified, and therefore not included in our dataset. In the following paragraphs, we supplement analysis of public procurement records with outside information, such as defense industry advertisements, to better understand the scope of China's AI-enabled C2 capabilities. Several Chinese enterprises outside of our dataset advertise AI systems capable of automating some elements of command and control-including knowledge mapping, decision support, weapon target assignment, and combat tasking. While we could not use procurement records to confirm that PLA units have purchased each of the systems specified below, each of the following companies publicly claims the Chinese military as a partner or client. Knowledge mapping is a visual representation of information designed to aid decision-making. DataExa, for example, advertises several services for AI-based knowledge mapping and combat decision support, such as encyclopedic information and real-time prediction about the movement of foreign weapons platforms.  130  In July 2020, the company's knowledge mapping software passed licensing review from the China Academy of Information and Communications Technology, and DataExa today lists the PLASSF and the Science and Technology Commission of the PLA Central Military Commission among its clients.  131  One of its products, the DataExa-Sati Knowledge Map, provides \"information about U.S. aircraft carrier equipment, such as submarines, destroyers, cruisers, and frigates accompanying aircraft carrier strike groups, and infrastructure such as overseas bases, satellite communications, logistics, and support equipment\" to the PLA Navy.  132  The company compares itself to the U.S. data management company Palantir Technologies.  133  Decision support systems streamline portions of the military decision-making process by helping identify courses of action for commanders.  134  One of China's most well-known military AI companies, StarSee (摄星智能), specializes in computer vision and decision support software. In 2020, the company won a commendation from China's Central Military Commission for its work on combatting COVID-19. 135 Among other products, StarSee advertises a \"Real-time Combat Intelligence Guidance System\" designed to \"combine the massive parameter model of a knowledge graph and the dynamic information of the battlefield in real-time.\"  136  StarSee's product is designed to create a common operational picture across different PLA units.  137  By \"relying on image, video, and audio language extraction and analysis technology,\" the company claims to be able to identify foreign weapons platforms, \"give various performance parameters of the weapon, and calculate its sustainability, firepower, maneuverability, command and control capabilities, intelligence capabilities, and other threat level parameters.\"  138  A product demonstration from June 7, 2020, appears to track Chinese aircraft flying near a U.S. Arleigh Burke-class destroyer off the coast of California.  139  Members of the StarSee research team previously worked for Baidu, Alibaba, Tencent, and Microsoft Research Asia. 140 Weapon target assignment software selects an optimal combination of weapons systems to engage one or more targets, assuming different success rates for each.  142  In addition to AI-based predictive maintenance and logistics software, the PLA contractor Anwise Global advertises a \"SIMBAT Weapon Effectiveness Evaluation System,\" which can reportedly use \"test data from multiple sources such as simulation, internal and external field tests, and exercises, among others, to evaluate effectiveness throughout the entire life cycle of weapons and equipment.\"  143  It is not clear whether the PLA has purchased access to SIMBAT specifically, but PLA units have awarded Anwise Global other AI contracts, and the PLA regularly publishes AI-based weapon target assignment research of its own.  144  Combat tasking, whereby a commander selects a course of action and directs a unit to complete some activity, represents one of the final steps in the military decision-making process.  145  Public procurement data indicates that PLA units and defense SOEs have awarded contracts for AI-based command and control software to support unit-level decision-making and combat tasking. One such project, awarded by the China Ship Research and Design Center of the China Shipbuilding Industry Corporation (CSIC), is for an \"intelligent loss management system\" to help commanding officers operate with fewer personnel after sustaining casualties. In another case, the PLA Ground Force awarded a contract to 4Paradigm, one of the largest enterprise AI companies in China, for a \"battalion and company command decision-making model and human-machine teaming software.\" 146 4Paradigm advertises a wide array of products and services, including software-defined computing platforms and an \"automatic decision-making machine learning platform\" called Sage HyperCycle.  147  As of January 2021, the company was cooperating on Very Large Database research with Intel and the National University of Singapore. 148 4Paradigm's angel investor, Sequoia Capital, remains its largest outside shareholder.  149  Taken collectively, these examples illustrate the kinds of public, AI-based decision support applications being developed by the Chinese defense industry. While it is too early to say whether the PLA may attempt to automate other segments of its C2 infrastructure, it is clear that some units have already begun acquiring COTS technologies for combat decision support and contingency operations. China's private AI sector is also starting to mature, with a few specialized companies like Anwise Global, DataExa, and StarSee carving out a niche to support different segments of the PLA's decision-making process. \n AUTOMATED TARGET RECOGNITION Target recognition and fire control are critical components of modern weapons systems, but applying AI to these tasks is a fairly new area of research. Although much of the PLA's research into AI-based automated target recognition is still aspirational, some units are purchasing relevant systems and equipment, and 4 percent (  14 ) of PLA AI contracts in our dataset appear related to ATR. Throughout 2020, PLA units and defense SOEs awarded contracts for feature extraction and recognition algorithms, target recognition algorithms for unmanned vehicles, brain-inspired multi-target fusion, and target detection based on synthetic aperture radar imagery.  150  Most notably, the Chinese military appears to be following in the footsteps of the DOD in developing AI-based ATR software for aerial vehicles.  151  Today, private Chinese companies, including Shandong Hie-Tech Co., Ltd. (山东航创电子科技有限 公司), advertise AI-based ATR systems for use in UAVs. The company won a contract to supply the PLAN with \"UAVs and supporting equipment\" in June 2020.  152  Research papers from PLA and state-sponsored laboratories also discuss developing AI-based ATR software. For example, the Shenyang Institute of Automation's first target recognition research forum in 2017 focused on using deep learning to recognize targets in still-frame images, and several SIA researchers have conducted research along similar lines.  153  SIA's Robotic Vision Group (机器人视觉组) lists a \"UAV airborne visual automated target tracking system\" as one of its achievements, but public information about the system is limited.  The PLA also aspires to use AI-based ATR software in undersea vehicles. In 2020, various PLA units awarded ATR contracts to universities and state-run research institutes that were still in the early stages of developing the technology.  156  Notable examples include \"USV target recognition algorithm and software development,\" \"deep learning-based, automatic detection of targets at sea,\" and a contract to construct a \"typical marine target database and target recognition module based on deep learning.\" 157 Multiple PLA units and defense SOEs awarded undersea target recognition contracts to Harbin Engineering University, one of seven universities administered by China's Ministry of Industry and Information Technology.  158  HEU researchers have developed a series of \"Smart Water\" (智水) AUVs for underwater ATR and path-planning missions, and are suspected of developing the HSU-001 large UUV first unveiled in China's 2019 National Day military parade.  159  Although AI-enabled ATR research is still in early development, it is maturing rapidly. Recent research papers by PLA units and defense industry engineers have used machine learning algorithms such as Single Shot Detector 300 (SSD300) and \"You Only Look Once\" (YOLO) to recognize targets with more than 80 percent accuracy.  160  While many of these algorithms are trained to recognize stationary targets in long-distance, overhead images, it remains to be seen whether or how the PLA may adopt AI-based ATR systems for ground-based weapons systems. eyond the Chinese military's intended application of AI, it is important for U.S. policymakers and defense planners to understand the sources of the technology that is being purchased by the PLA. Based on trends in public PLA contracts, we offer three observations about the overall structure and efficiency of China's emerging military AI industry. First, among the contract award notices in our dataset, defense SOEs are both buyers and suppliers of AI-related equipment. Research institutes and subsidiaries belonging to CASC, AVIC, the Aero Engine Corporation of China (AECC), and CSIC are included in the PLA's official procurement records (referred to as \"PLA contracts\"), and each placed orders for AI-related equipment in 2020.  161  Along with other defense SOEs, these companies and their subsidiaries also received contracts to supply AI-related equipment to PLA units and other state-owned research institutions. This two-way transfer of technology could indicate that SOEs are specializing in certain subfields of AI development, rather than crowding out private-sector investment. \n Supply and Demand for Intelligent Equipment in the PLA Second, whereas the organizations responsible for buying AI equipment are fairly concentrated, the PLA's AI equipment suppliers are diffuse. Of the 343 public AI contracts in this study, 331 contracts named 273 unique suppliers. Most companies were awarded just one public AI contract, and the single most active private-sector supplier, Langfang Rongxiang Electromechanical Equipment Co., Ltd. (廊坊市荣祥机电设备有限公司), was awarded just four contracts during our period of inquiry. Moreover, China's MCF development strategy is improving the PLA's access to private-sector advances in AI. Since the 1980s, the Chinese government has attempted to integrate the technical achievements of the civilian and military industries to strengthen China's comprehensive national power. In more recent policy documents, the CCP has called for deepening MCF and \"encouraging the twoway transfer and transformation of military-civil technology,\" as evidenced by the  Internet+ Action Plan (2015), innovation-driven development strategy (2016), and New Generation Artificial Intelligence Development Plan (2017).  162  CSET research has highlighted the role of new policy levers in achieving this goal, such as government guidance funds, technology brokers, and a formal Chinese AI Industry Alliance-of which 24 (9 percent) of the 273 suppliers in this study are members.  163  It should come as no surprise that a large number of private companies supply the PLA with AI-related equipment, and that some of these companies have benefited from equipment, personnel, information, and capital provided directly or indirectly by the state. \n THE PRIVATIZATION OF INTELLIGENTIZATION A robust military AI industry is emerging in China, spanning Chinese Academy of Sciences (CAS) research institutes, military factories, universities, private en-TABLE  3  The PLA's Top Suppliers of AI-Related Equipment, April-November 2020 terprises, state-owned enterprises, and their subsidiaries. To categorize each AI equipment supplier in our dataset, we searched for background information on each company's \"About Us\" web page, vacancies they advertised on job posting websites, and ownership information on Chinese financial service and due diligence platforms. We recorded the date each institution was established and any indication that it may be a subsidiary of a defense SOE or state-owned holding company. If a company publicly claimed to be a subsidiary or appeared to be majority-owned by an SOE, university, or CAS institute, we labeled it as such. Among the 273 unique PLA suppliers identified in this study, we find that private Chinese technology companies-not SOEs or their subsidiaries-are the PLA's most common suppliers of AI-related equipment. Generally speaking, these are recently established, high-technology companies for whom intelligent software or sensors are a dominant focus.  164  The PLA awarded 61 percent of the public AI contracts in our dataset to 166 private enterprises. Of them, two-thirds (108) were founded since 2010, and more than one-third  (63)  were founded since 2015. Most have fewer than 50 employees and registered capital of less than $1 million.  165   \n FIGURE 9 Private Companies are the PLA's Primary AI Equipment Suppliers As previously noted, China's MCF strategy is accelerating the PLA's access to, and adoption of, AI. Many of the non-state-owned companies that supply the PLA with AI equipment are supported directly or indirectly by the state, and some self-identify as \"military-civil fusion enterprises\" (军民融合企业). But even non-stateowned companies tend to jointly develop products with legacy defense SOEs, or base their business model around supplying them with software and equipment. The typical modern, private Chinese military AI company is: • Founded by STEM graduates from elite universities in coastal provinces; • Headquartered in a commercialization enclave or innovation park run by a university or the local CCP Science and Technology Commission; • Engaged with researchers at defense-affiliated universities and research laboratories; and • Sustained by contracts from public security bureaus, PLA units, and major defense SOEs. \n LIMITATIONS OF U.S. EXPORT CONTROLS U.S. policymakers regularly voice concerns that technology produced in the United States may be exfiltrated and deliberately or inadvertently accelerate Chinese military modernization. The U.S. government has adopted several policies designed to curtail the Chinese defense industry's access to equipment, personnel, information, and capital, especially where AI is concerned.  166  Since 1989, the United States has prohibited arms sales to China, and today the U.S. government presumes denial for license applications of items relevant to national security (NS items) to known military end-users in China.  167  Additional statutes place restrictions on specific companies: • The Entity List (EL) published by the U.S. Department of Commerce's Bureau of Industry and Security restricts the ability of U.S. firms to sell or supply technology or intellectual property to specific institutions abroad, including some individuals and institutions based in China.  168  • The Chinese Military-Industrial Complex Companies List (NS-CMIC List) published by the Department of the Treasury's Office of Foreign Assets Control restricts the ability of U.S. persons to make securities investments or own stock in certain Chinese military companies, pursuant to Executive Order 13959.  169  • The List of Chinese Military Companies (NDAA Sec. 1260H List) published by the Department of Defense is mandated by the FY2021 National Defense Authorization Act, and exists to inform Americans of companies that may be connected to the Chinese military.  170  Subsequent Executive Orders have extended OFAC investment restrictions to include companies on the Sec. 1260H List.  171  Although tens of thousands of Chinese companies are licensed to supply the PLA with equipment, very few are found on any of these three U.S. export control or sanctions lists. Of the 273 known AI equipment suppliers in our dataset, just 8 percent  (22)   Because most institutions that supply AI-related equipment are new and not subject to end-use controls, the Chinese military is frequently able to access or acquire technology from abroad, including from the United States. Some Chinese suppliers make a business out of sourcing foreign data or components and reselling them to sanctioned Chinese defense companies or PLA units. Beijing Zhongtian Yonghua Technology Development Co., Ltd. (Zhongtian Yonghua; 北京中天永华科技发展有 限公司), for example, is not currently listed in any U.S. sanctions regime. In August 2020, it was awarded a contract to supply intelligent sensor equipment to CASC, which the DOD designates as a Chinese military company.  173  A Chinese online business directory entry for Zhongtian Yonghua says that it is \"mainly engaged in the agency and sales of various imported instruments and meters,\" and specifies that it is primarily a distributor of instrumentation equipment produced by companies in the United States (Agilent, Fluke Corporation, and Testo Inc.) and Japan (Hioki Corporation and Kyoritsu Electrical Instruments Works, Ltd.).  174  Multiple companies engage in similar activity, and some examples are included throughout this report.  175  rocurement data offers a detailed, if incomplete, picture of how the PLA may use AI in future warfare. By comparing trends in purchasing records to long-standing themes observed in theoretical writings, research papers, and news reporting, we conclude that the PLA is interested in using AI to erode the U.S. advantage in undersea warfare and to jam U.S. sensor and communication networks. These aspirations are particularly relevant for U.S. policymakers and defense planners as they respond to mounting Chinese threats to Taiwan and other partners in the Indo-Pacific.  176   \n ERODING THE U.S. ADVANTAGE IN UNDERSEA WARFARE The PLA's adoption of AI appears focused in part on overcoming its significant disadvantages in undersea warfare. Ten years ago, the PLAN had \"very limited ASW [anti-submarine warfare] capabilities and [appeared] not to be making major investments to improve them\"; and more recent assessments have concluded that floating mines and active sonar would likely prove ineffective against U.S. submarine forces operating in or near the Taiwan Strait.  177  To compensate, the PLAN commissioned the construction of an \"Underwater Great Wall\" (水下长城) acoustic sensor network in 2017, and has since rapidly expanded its diesel submarine force.  178  Today the PLA appears to be making significant investments in AI-enabled systems, such as A/UUVs, A/USVs, and undersea ISR systems, which could challenge U.S. and allied submarine forces in a crisis. In addition to the contract data presented earlier in this report, research published in 2021 by Jiangsu University of Science and Technology claims that \"a full spectrum of unmanned submersibles has been initially What the PLA's Buying Habits Say About Its Battle Plans 5 P established in China,\" listing nearly a dozen AUV and UUV models of varying sizes.  179  Based on contract data and recent technology demonstrations, we assess that over the next five to 10 years, the PLAN will likely continue expanding its network of autonomous surface and undersea vehicles in an attempt to limit U.S. Navy access to the undersea space between the first and second island chains.  180  Public A/UUV contracts in our dataset are primarily for small-and medium-sized vehicles used for ISR, but English-language reporting has also shed light on some of the PLA's larger vessels, which are proliferating in number and growing in capability.  181  Chinese AUVs have also set navigation records for depth and distance. In June 2020, the SIA's Haidou 1 (海斗一号) AUV successfully dove below 10,000 meters in the Mariana Trench; and in November, SIA's Sea-Whale 2000 (海鲸2000) AUV finished a 37-day continuous test, crossing 1,250 miles of the South China Sea.  182  Despite the PLA's apparent progress in testing, however, prior CSET analysis has shown that \"the state of the current technology, the complexity of antisubmarine warfare, and the sheer scale and physics-based challenges of undersea sensing and communications all suggest these systems have a long way to go.\"  183  Given limitations in battery life and the robustness of computer vision systems, it remains to be seen whether the PLA's expanding AUV force will materially change the undersea balance of power. \n JAMMING AND BLINDING U.S. INFORMATION SYSTEMS In conjunction with modernizing equipment, the PLA is developing new concepts of operations oriented around systems confrontation and systems destruction warfare, in which \"warfare is no longer centered on the annihilation of enemy forces on the battlefield,\" but \"won by the belligerent that can disrupt, paralyze, or destroy the operational capability of the enemy's operational system.\"  184  For example, an electronic warfare textbook published by NUDT emphasizes that \"the U.S. military's combat command, military deployment, and joint operations are extremely dependent on battlefield information network systems,\" and prescribes that, \"once the battlefield communication network is broken . . . the entire battlefield information network system (C4ISR system) will be severely damaged, destroyed, or even paralyzed.\"  185  Approximately 8 percent of public procurement projects in our dataset  (29)  are related to information and electronic warfare, many of which focus on jamming or blinding enemy sensor networks and using AI for cognitive electronic warfare. Examples of such equipment contracts are outlined in Table  4 . Cyberattacks, data manipulation, and electromagnetic spectrum interference are key components of the PLA's systems confrontation strategy. In 2020, several PLA units and state-backed research institutions awarded contracts for \"microwave reconnaissance jamming drones\" and \"electromagnetic weapon\" payloads that can be attached to swarms of small UAVs and flown into enemy airspace.  186  PLA thinkers also emphasize the need to \"disrupt or block the enemy's command and decision-making to ensure one's own decision-making advantage,\" and point to the U.S. Joint Enterprise Defense Infrastructure (JEDI; now the Joint Warfighting Cloud Capability) as a likely locus of systems confrontation.  187  \"In the 'combat cloud' system,\" write PLA National Defense University professors Zhang Xiaotian and Luo Fengqi, \"information and algorithms are key strategic resources, and the opposing parties will inevitably engage in information confrontation and algorithmic warfare in the 'cloud.'\" 188  espite the PLA's demonstrable progress in adopting AI, three points of tension will define its continued push toward intelligentization in the 2020s and beyond. These include the vulnerability of C4ISR networks; dependence on foreign computer chips; and disagreements over the development of lethal autonomous weapon systems. \n BREAKING THE COMBAT CLOUD IT STRIVES TO EMULATE The first tension concerns the PLA's plan to exploit U.S. battle networks while developing its own. Having watched and learned from the U.S. experience in Afghanistan, Iraq, Kosovo, and Libya, the PLA is investing heavily in its own networked C4ISR systems, many of which feature elements of AI.  189  The PLA's vision for an intelligentized force is based in large part on U.S. military concepts like network-centric warfare, Mosaic Warfare, and the notion of a \"combat cloud.\"  190  In particular, PLA thinkers cite the need to \"cloudify\" (云化) their combat systems to speed up the observe-orient-decide-act (OODA) feedback loop first coined by U.S. Air Force Colonel John Boyd.  191  By emulating U.S. integration of sensor arrays and weapons platforms, the PLA aims to develop \"'ubiquitous networks' (泛在网络) that will shorten the distance between perception, judgement, decision-making, and action,\" and brace itself for the quickened tempo of modern warfare.  192  \"As the pace of war accelerates,\" write PLA science and technology analysts Shi Chunmin and Tan Xueping, \"combat time will be calculated in seconds.\"  193  They go on to note that Link 16, the communications and tactical data transmission system used by the United States, NATO, and coalition forces, allows a delay of just seven milliseconds.  194   \n Fundamental Tensions in Chinese Military Modernization But ubiquitously connecting sensors and shooters has created new vulnerabilities for the United States, which PLA leaders recognize and plan to exploit. U.S. defense planners often lament that exquisite ISR and communication systems make for \"big, fat, juicy targets,\"  195  and worry that in a crisis, adversaries will jam, blind, and hack the networks that bind U.S. assets together.  196  As previously outlined, PLA leaders are forging a new array of operational concepts predicated on \"systems confrontation\" and \"systems destruction,\" which are specifically designed to take advantage of U.S. vulnerabilities.  197  But it is not clear how the PLA plans to make its own networks resilient to the kinds of exploits it envisions deploying against the United States. One solution may be to develop edge applications of AI that are insulated from the rest of the battle network.  198  Some edge applications-such as predictive maintenance systems for satellites or target recognition systems for autonomous underwater vehicles-can be found among the PLA's public AI contracts.  199   \n ENSURING ACCESS TO FOREIGN SEMICONDUCTORS The second tension concerns the supply of advanced computing products at the heart of China's intelligentization strategy. Chinese leaders are acutely aware of the PLA's wholesale dependence on AI chips designed by U.S. companies and produced in Taiwan and South Korea. Although computing hardware was not the original focus of our procurement analysis, further investigation reveals that the PLA has awarded myriad contracts for U.S.-designed computer chips useful for training machine learning systems.  200  One paper by researchers from China's Ministry of Industry and Information Technology estimates that \"more than 90% of China's high-end chips rely on imports,\" including \"100% of DRAM memory, 99% of CPUs, and 93% of MEMS sensors.\"  201  The Chinese government has constructed multibillion-dollar guidance funds to promote the country's domestic semiconductor manufacturing industry, but these initiatives are rife with corruption, and it is not yet clear whether they will succeed.  202  In the meantime, U.S. policymakers have crafted a variety of export controls and sanctions designed to limit the Chinese military's access to leading-edge AI chips.  203  Large Chinese corporations such as Huawei have had to cease production of some product lines and have seen large shortfalls in revenue as a result of U.S. sanctions.  204  However, this study finds that few of the PLA's AI equipment suppliersjust 8 percent of the companies in our dataset-face specific barriers to acquiring U.S. equipment, information, and capital. PLA units and defense SOEs continue to procure systems that use leading-edge NVIDIA and Xilinx processors, and sometimes purchase these processors, themselves, through intermediary companies. Moreover, in 2021, U.S. companies have cooperated on AI-related research projects with Chinese businesses that supply the PLA with AI systems and equipment.  205  The PLA's continued access to U.S. and other foreign technology is not guaranteed. The United States and its allies may yet take additional steps to impede Chinese military access to the data, hardware, and personnel required to build an intelligentized force. If such policies were to create a significant shortage in advanced semiconductors, and if China continues to struggle with indigenizing segments of its own chip industry, it would likely slow or impair the PLA's intelligentization strategy. But U.S. experts warn that a decision to effectively cut off Chinese military access to foreign advanced semiconductors could inadvertently fuel China's homegrown chipmaking industry, and should not be taken lightly.  206   \n DECIDING THE ROLE OF LETHAL AUTONOMOUS WEAPONS The third tension concerns the development of LAWS. The Chinese government has famously shown a Janus face to LAWS, publicly calling for a ban on such weapons while privately carving out a legal defense for their development.  207  As in the United States, different factions within the Chinese military and defense industry harbor different attitudes toward LAWS.  208  A 2020 CSET study found that PLA officers and defense industry engineers are worried AI may undermine strategic stability by reducing the capability of Chinese air defenses, increasing the vulnerability of Chinese command and control systems, or degrading the PLA's available time to respond to an attack.  209  Some PLA officers appear legitimately disturbed by LAWS, and caution against a future characterized by smart weapons. In 2021, for example, three PLA researchers responded to reports that a Turkish \"Kargu-2\" quadcopter had autonomously attacked a human target in Libya, writing that fully-autonomous weapon systems present \"not only a lack of moral responsibility, but also a serious challenge to international humanitarian law and international peace and security.\"  210  This perspective is not uncommon, and PLA officers often voice similar concerns in research papers and think pieces. Others in the PLA are more sanguine about AI's utility on the battlefield, and believe that technology will inevitably increase the operational tempo of war such that, unaided by fully automated systems, humans will be incapable of responding to imminent attack.  211  Liu Peng, a member of China's Cloud Computing Expert Committee and professor at the PLA University of Science and Technology, wrote in 2020 that \"at present, most intelligent combat decision-making systems are semi-autonomous systems with humans in the loop,\" but that the PLA should \"introduce learning into the combat decision-making process to achieve mutual error correction, complementarity and efficiency.\"  212  Defense officials in the United States have advanced similar arguments.  213  Among other issues, the fast pace of technology development, lack of appetite for safety measures, and general lack of trust between the United States and China are giving rise to a security dilemma around AI development.  214  Despite the Chinese government's stated position against LAWS, it is clear that developing AI-based target recognition and fire control systems is an objective of some PLA and government-backed research centers.  215  Computer vision is by far the most active subfield of the PLA's public AI research portfolio, and the share of PLA-sponsored research papers dedicated to \"military target recognition\" (军事目 标识别) increases each year.  216  By itself, AI-based ATR does not constitute a lethal autonomous weapon system. Yet target recognition remains an integral step in the detect-to-engage sequence, and AI-based ATR is inseparable from \"AI weapon\" (人工智能武器) concepts described by the Chinese military.  217  In August 2020, The Paper, a well-circulated state-run media outlet, reported that using AI-based ATR to \"equip a missile with a 'super-powerful brain' to achieve precision strikes\" is the \"lifelong pursuit\" of NUDT's State Key Laboratory for Automated Target Recognition (自动目标识别国家重点实验室).  218  ATR research published by the Dalian Naval Academy is similarly explicit, noting that \"AI and computer vision technology provides new technical support for shipborne missiles to attack all kinds of sea and land targets accurately\" and, \"in the process of target recognition, using deep learning algorithms is an effective way to improve the accuracy of missiles attacking targets.\"  219  Ultimately, trends in procurement records, research publications, and media reports indicate that the Chinese military and defense industry are developing AIbased target recognition and fire control systems, which are essential components of LAWS. Although public information about their research is limited, the NUDT State Key Laboratory for Automated Target Recognition and the CAS Shenyang Institute of Automation appear to be key institutions driving LAWS development in China. Combined with the Chinese government's extraordinarily narrow definition of LAWS, this emphasis on AI-enabled ATR research suggests that the PLA may yet develop weapons capable of autonomously detecting and engaging human targets.  220   \n Conclusion he share of procurement activity dedicated to AI is one indication that China's military aspirations extend beyond peripheral security concerns. In the 2020s, intelligentization has become the chief focus of Chinese military modernization, with AI-related systems and equipment already accounting for 5 percent of public contracts awarded by the PLA's five main service branches. This report's narrow look at public procurement records confirms that the PLA awarded AI contracts worth at least $49 million from April to November 2020, and that it may spend more than $1.6 billion on AI-related systems and equipment each year. Investing in AI is part of the PLA's longstanding mission to become a \"world-class\" military that is \"equal to, or in some cases superior to, the United States.\" 221 PLA leaders frequently compare their own capabilities to those of the U.S. military, and public writings from 2021 refer explicitly to degrading and exploiting U.S. information systems. While much of the PLA's focus on systems confrontation and systems destruction appears to still be in early stages of development, a plurality of its equipment contracts are related to information operations and electromagnetic spectrum dominance. Within the next five to 10 years, the Chinese military will likely continue investing in AI to erode the U.S. advantage in undersea warfare; and will seek opportunities to jam, blind, and hack U.S. military information systems. Contrary to conventional wisdom about bloating in the Chinese defense industry, we find that the PLA has made significant progress engaging the private Chinese technology sector to acquire AI systems and intelligent equipment.  222  Most of the PLA's AI equipment suppliers today are not lega-T cy defense SOEs, but small, private companies that specialize in software development, data management, and IoT device design.  223  Some Chinese AI companies in our study self-identify as \"military-civil fusion enterprises,\" and benefit from equipment, personnel, information, or capital provided by the state. Others are private technology companies that have welcomed the PLA as a customer. The PLA's progress toward intelligentization will become increasingly important for the United States in the 2020s as tensions between the two countries continue to rise. In its attempts to harness AI for military advantage, the PLA will face important questions, for example, about decoupling supply chains and developing lethal autonomous weapons. It remains to be seen whether the Chinese military will succeed in becoming a fully intelligentized and world-class military force, but one thing is certain: AI is no longer just an emerging technology. Rather than speculate about its far-future implications, defense planners and policymakers would do well to heed the words of science fiction writer William Gibson: \"The future is already here-it's just not evenly distributed.\"  224  veillance, and Reconnaissance; (3) Predictive Maintenance and Logistics; (4) Information and Electronic Warfare; (5) Simulation and Training; (6) Command and Control; and (7) Automated Target Recognition. Tenders that did not fit in any of these seven categories were marked \"Other.\" Categorizing each contract was necessarily a subjective, iterative process. To make the labels more robust, the authors used the Elicit AI research assistant to check their manual coding.  226  Elicit uses language models to code data. For each contract, Elicit used the manual labels for some of the other contracts as training data, then labeled each \"intelligent equipment\" contract in the dataset. Initially, Elicit agreed with author coding in 50 percent of cases. After one author reviewed the disagreements, recoded some of the data, and reran Elicit, agreement increased to 62 percent. For remaining disagreements, author judgment superseded that of Elicit. Table  5  lists 10 examples of how the authors coded \"intelligent\" equipment contracts.   The PLA's five main service branches are most focused on using AI to improve navigation and data management in autonomous vehicles; to improve the speed and scale of intelligence collection and dissemination; and to enhance logistics through predictive maintenance. However, each of the services tend to focus on different applications of AI. For example, relative to other branches, the PLASSF is most interested in purchasing AI technology that can be used in information and electronic warfare, whereas the Ground Force tends to purchase more AI solutions for predictive maintenance and logistics. Each of the PLA's service branches are focused on using AI in intelligent and autonomous vehicles and for predictive maintenance, but ISR and information and electronic warfare are also common applications.   \n PLA Service \n PLA STRATEGIC SUPPORT FORCE (PLASSF) As the service branch responsible for space, cyber, and information warfare, the PLA Strategic Support Force is the most active in procuring AI-related technologies and applications. Of the PLASSF's 65 public AI-related contracts, most were related to intelligence, surveillance, and reconnaissance (ISR); information warfare, and autonomous vehicles.    228  The General Staff Department's Survey and Mapping Research Institute in Xi'an (Unit 61540) was another significant purchaser of AI-enabled ISR equipment within the PLASSF, awarding various contracts for intelligentized forecast correction systems, high-precision positioning algorithms, and high-resolution ocean and climate modeling software.  229  The PLASSF envisions using AI for information and electronic warfare, in applications ranging from multilingual natural language processing to public opinion monitoring, cyber threat intelligence, and adaptive radar jamming. For example, the PLASSF's Engineering Information University awarded various contracts for an AI-based \"internet public opinion monitoring and clustering system,\" \"intelligent network traffic analysis system,\" and a \"cyber threat intelligence early warning platform\" throughout 2020. Previous research suggests that PLASSF Base 311 (Unit 61716) carried out social media manipulation ahead of Taiwan's 2018 local elections.  230  One of the PLASSF's repeat contractors, EverSec, advertises services for petabyte-scale data storage and processing, AI-based open source data mining, and internet blocking and censorship protocols.  231  In addition, multiple PLASSF units have awarded contracts for AI-based fiber optical line protection switches, network amplifiers, and automated frequency modulation systems used in cognitive electronic warfare.  232  The PLASSF is also awarding contracts for autonomous vehicles, both individual platforms for logistics and sustainment and swarms with potential combat applications. Public reporting indicates that the PLASSF has been experimenting with using UAVs to resupply troops from Lhasa, Tibet, for operations near the Line of Actual Control with India.  223  The service's procurement records indicate that it has purchased several UAVs and intelligentized simulation systems from Lyncon Tech (西安羚控电子科技有限公司), a leading provider of Chinese drone swarm technology, specifically for use in or near Lhasa. In June 2020, the PLASSF's Equipment Command and Technology Academy (Unit 63628) awarded a $630,000 contract to AOSSCI Technology for a UAV simulation and training center. \n PLA GROUND FORCE (PLAGF) After the PLASSF, the Ground Force has displayed the most interest in adopting AI, and awarded 58 of the public AI-related contracts in our dataset. Previous studies show that academic AI research sponsored by the PLAGF tends to focus on improving unmanned or robotic systems' ability to navigate difficult terrain, and this priority is also reflected in the service's public procurement records.  234  Of the 58 AI contracts the PLAGF awarded in 2020, the plurality were related to autonomous vehicles, predictive maintenance, and electronic warfare. The PLAGF is most focused on using AI in autonomous aerial and ground vehicles. Valued at $890,000, its single most expensive public AI contract involved developing a UAV swarm, to be fulfilled by CASC Shenzhou Flight Vehicle Co., Ltd. (航天神舟飞行器有限公司). The company holds several patents related to swarm technology and UAV-based applications of the Beidou satellite constellation.  235  Other PLAGF contracts were related to ultra-short-range control link modules and non-line-of-sight transceivers for unmanned and autonomous UAVs. The PLAGF is also leveraging AI for ground vehicles, and in June 2020 awarded a contract to Beijing Laser Bofeng Information Technology Co., Ltd. (北京雷神博峰信息技术有限责任公司), \"a major supplier of vehicle-mounted Beidou information terminals and intelligent control systems for petrol vehicles,\" to develop an autonomous tanker truck.  236  The company specializes in autonomous \"IoT vehicles\" for logistics and transportation.  237  Predictive maintenance is another clear priority for the PLAGF. Throughout 2020, it awarded contracts for \"intelligent supply chain\" networks, ammunition shell quality detection software, and equipment failure and maintenance prediction systems. Among these, its most expensive project was a $275,000 contract for a \"self-organized network intelligent packaging system\" for bullets. The contractor, Chongqing Jialing Special Equipment Co. (重庆嘉陵特种装备有限 公司), is a wholly owned subsidiary of the defense SOE China North Industries Corporation (NORINCO).  238  To a lesser degree, the PLAGF is also leveraging AI for electronic warfare. In September 2020, PLAGF Unit 63871 awarded Xi'an Ruiweishen Electronic Technology Co., Ltd. (西安睿维申电子 科技有限公司) a $160,000 contract to develop a \"microwave reconnaissance jamming drone.\" In 2013, the company had won a national Torch Program award for a high-performance digital signal processing platform.  239  The PLAGF awarded a similar contract to TIYOA Aviation ( 河北天遥航空设备科技有限公司) to develop \"electromagnetic weapon\" payloads aboard small UAVs. The company specializes in intelligent control systems, video surveillance, and small drone applications.  240  PLA NAVY (PLAN) The Navy is just behind the Ground Force as the PLA's third-most active branch in adopting AI, and awarded 51 of the public AI-related contracts in our dataset. Most contracts were related to autonomous vehicles, ISR, and other applications not neatly captured in our taxonomy. Naval Aviation University (海军航空大学) was the PLAN unit most active in awarding AI contracts, accounting for nearly one third  (15)  of those in our dataset. Valued at $1.3 million, the PLAN's most expensive public AI contract involved retrofitting unmanned aerial vehicles with multi-tasking pods, to be filled by AECC Guizhou Liyang Aviation Power Co., Ltd. (中国航发贵 州黎阳航空动力有限公司). Several other PLAN contracts were related to sea floor mapping and AUV development. Intelligence, surveillance, and reconnaissance is another major focus of the PLAN's AI procurement. Several contracts mention fusing automatic identification system (AIS) ship positioning data to improve situational awareness for submarine and surface fleets. For example, in June and July 2020, the PLAN Submarine Academy (海军潜艇学院) purchased bulk AIS ship tracking data from Elane Inc. (亿海蓝(北京)数据技术股份公司) and tasked the company with bulk AIS processing.  241  Elane runs shipfinder.com, a global shipping database with \"millions of global shipping and related users.\"  242  The company advertises \"real-time monitoring of all satellite AIS ship positions worldwide,\" updated every five minutes, using a constellation of 108 Orbcomm satellites.  243  Although Orbcomm is a U.S. satellite company, Elane's AIS service is marketed for Chinese users only. Other PLAN units have struck similar contracts to purchase AIS data from Yantai Huadong Electron Technology Co., Ltd. (Huadong Elec-Tech; 烟台华东电 子软件技术有限公司) and the China Transport Telecommunications & Information Center (CTTIC) Information Technology National Engineering Laboratory.  244  The PLAN's other AI contracts involve building libraries of undersea sonar signatures and using deep learning to stitch them together. In July 2020, the PLAN's Naval Aviation University awarded Harbin Engineering University with a contract for an \"automatic sea-based target detection\" system based on \"deep learning image recognition.\" Several researchers at HEU have pioneered an AI-based \"seabed image mosaic system,\" hold relevant patents, and regularly conduct research on the topic.  245   \n PLA AIR FORCE (PLAAF) With just ten public AI contracts, the Air Force appears much less interested in AI procurement, relative to the Ground Force or Navy. Its AI purchases are mostly related to autonomous vehicles, predictive maintenance, and electronic warfare. Most public AI research papers sponsored by the PLAAF are related to autonomous flight, a trend that extends to its procurement records.  246  Several contracts were related to intelligent flight decision control (智能驾驶决策控制), technology primarily developed by Northwestern Polytechnical University and research institutes subordinate to CASC. One of the primary companies involved in supplying autonomous UAVs to the PLAAF is ChunYi UAV (北京淳一航空科技 有限公司), a Beijing-based provider of autonomous aerial and surface-sea vehicles.  247  In September 2020, the PLA Air Force paid to lease and operate some of ChunYi UAV's autonomous aerial vehicles. The company's website specifies that its products are useful for \"counterterrorism and aerial dogfight weapons testing and training.\"  248  Some of the PLAAF's more expensive AI contracts are related to predictive maintenance for communication networks. In April 2020, for example, China Eracom Contracting and Engineering Co., Ltd. (中时讯通信建设有限公司), a fiber-optic cable company, won a contract for an \"intelligent operation and maintenance management system.\" That month, the PLAAF awarded a similar contract to China Iconic Technology Co, Ltd. (中徽建技术有限公司), a twice-removed subsidiary of China Telecom, for an intelligent phone network system. Finally, like the PLAGF, the PLAAF is also interested in using AI for electronic warfare and electromagnetic spectrum dominance. In October 2020, the PLAAF's Air Defense Early Warning Equipment Department (Unit 93209) awarded China Civil Aviation University a contract for \"trusted radar target detection\" and \"research into the dynamic evolution of the electromagnetic environment,\" using algorithms to enhance battlefield situational awareness and plot the locations of friendly radar units. \n PLA ROCKET FORCE (PLARF) The Rocket Force is the least active service branch with respect to public AI procurement, awarding just four contracts from April to November 2020. The PLARF contracts included using AI to forecast maintenance and support resource consumption; as well as to develop intelligent robotics; a \"smart communications warehouse\"; and an autonomous, tethered UAV platform. The PLARF awarded its largest AI contract to China Electronics Technology Corporation (CETC) for an autonomous, tethered UAV platform, to be supplied to the 613 Brigade in Shangrao City.  249  Tethered drones are particularly useful for emergency response and communication, as an autonomous UAV can be towed alongside a ground vehicle or watercraft without the need for constant supervision or recharging. 250 CETC's 54th Research Institute produces lines of fourand six-rotor tethered UAVs,  251  while the 7th and 23rd Research Institutes hold patents on UAV mooring cables.  252  Other Chinese companies have developed tethered UAVs for emergency communications, such as the DG-X10 and DG-M20.  253   FIGURE 1 Equipment Modernization Phases Mentioned in China's Defense White Papers \n 1 . 1 Intelligent and Autonomous Vehicles 2. Intelligence, Surveillance, and Reconnaissance (ISR) 3. Predictive Maintenance and Logistics 4. Information and Electronic Warfare 5. Simulation and Training 6. Command and Control (C2) 7. Automated Target Recognition \n FIGURE FIGURE 2 \n 80     \n FIGURE 4 4 FIGURE 4     \n FIGURE 5 5 FIGURE 5 \n FIGURE 6 \" 6 FIGURE 6 \"War Game in Taiwan Strait 2019\" Using CMO \n FIGURE 7 Real 7 FIGURE 7Real-Time Combat Intelligence System Offered by a PLA Contractor \n 154 154 \n FIGURE 8 AI 8 FIGURE 8 AI-Based UAV Target Lock Software Advertised by a PLA Contractor \n FIGURE 9 Private 9 FIGURE 9    Private Companies are the PLA's Primary AI Equipment Suppliers \n FIGURE 11 Number 11 FIGURE 11    Number of AI Equipment Contracts Awarded by PLA Service Branches, April-November 2020 \n \n TABLE 1 Types 1 of Procurement Information Published by the PLA, April-November 2020 Source: CSET corpus of PLA procurement activity. \n TABLE 1 1 Types of Procurement Information Published by the PLA, April-November 2020 contracts were awarded by defense SOEs; theater commands; and research institutes and academic institutions under the control of the Central Military Commission, including the Academy of Military Sciences and the National University of Defense Technology (NUDT). Announcement Type ANNOUNCEMENT TYPE Public ( PUBLIC ( ) ) Confidential ( CONFIDENTIAL ( ) ) Secret ( SECRET ( ) Total ) TOTAL Award (Bid) 15,028 1,855 356 17,239 Award (Sole Source) 3,545 272 30 3,847 Bid Solicitation 7,508 2,767 416 10,691 Inquiry 12,268 659 4 12,931 Requirement 2,705 2,509 406 5,620 Modification or Annulment 2,143 600 102 2,845 Other 6,716 5,362 956 13,034 Total 49,913 14,024 2,270 66,207 \n TABLE 2 The 2 PLA's Top Buyers of AI-Related Equipment, April-November 2020 \n TABLE 2 The 2 PLA's Top Buyers of AI-Related Equipment, April-November 2020 Institution INSTITUTION Institution (Chinese) INSTITUTION (CHINESE) No. of Contracts NO. OF CONTRACTS China Aerospace Science and Technology Corporation (CASC) Strategic Support Force Ground Force Navy Academy of Military Sciences Air Force Aero Engine Corporation of China (AECC) Overall Design Institute of Hubei Aerospace Technology Research Academy (CASIC 9th Overall Design Department) Rocket Force China Ship Research and Design Center National Defense University People's Armed Police Source: CSET corpus of PLA purchasing activity (343 contracts specify purchasing units). Note: Values for state-owned enterprises such as CASC and AECC include multiple subsidiaries. \n TABLE 3 The 3 PLA's Top Suppliers of AI-Related Equipment, April-November 2020 Institution Source: CSET corpus of PLA purchasing activity (331 contracts specify suppliers). \n face specific limitations set by the U.S. Departments of Commerce, Treasury, or Defense. At times, lapses in due diligence and situational awareness may permit the Chinese military and defense industry to access U.S. technology and capital. 172   FIGURE 10 FIGURE 10 Portion of Known PLA AI Equipment Suppliers Named Portion of Known PLA AI Equipment Suppliers Named in U.S. Export in U.S. Export Control or Sanctions Lists Control or Sanctions Lists Entity List (BIS) 91% NS-CMIC List (OFAC) 83.5% NDAA Sec. 1260 List (DOD) Informatization (信息化) Intelligentization (智能化) 91.2% NOT LISTED PARENT IS LISTED LISTED Source: CSET corpus of PLA procurement activity (273 known AI equipment suppliers). \n TABLE 4 Select 4 AI-Related Electronic Warfare Contracts Awarded by PLA Units in 2020 Translated Project Name MUCD Probable Affiliation within the PLA Autonomous and controllable transformation of software configuration Unit 63796 PLASSF Xichang Space Launch Center management system Optical fiber line automatic switching protection devices and optical Unit 66389 PLASSF (Central Theater Command) amplifier equipment Information and Communications Brigade Power amplifier and smart pressurizer Unit 63751 PLASSF Base 26 Tracking and Communications Office Enclosed space automatic frequency modulation device Unit 63672 PLASSF Northwest Academy of Nuclear Research Research on key test technology for microwave reconnaissance jamming Unit 63871 Huayin Conventional Munitions Test and UAV Training Base Environmental noise intelligent collection terminal Unit 63811 PLASSF Wenchang Space Launch Center Algorithm demonstration software for cooperative sensing of radar Unit 93209 PLAAF Research Academy targets; credible detection and dynamic evolution of electromagnetic environment \n TABLE 4 4 Select AI-Related Electronic Warfare Contracts Awarded by PLA Units in 2020 TRANSLATED PROJECT NAME MUCD PROBABLE AFFILIATION WITHIN THE PLA Source: CSET corpus of PLA procurement activity (seven EW contracts awarded by identifiable PLA units). \n TABLE 5 5 Examples of Coded \"Intelligent\" Equipment Contracts Tender Name (Chinese) Tender Name (English) Counts as AI? Application(s) \"XX design and artificial intelligence typical application Yes Other scenarios demand analysis of key general use technologies\" sole source announcement \"Research on military applications of AI technology for Yes Autonomous Vehicles rockets -intelligent robotics technology in rocket (Munitions) military XX class engineering applied research project\" outsourcing Two service-type pooled and packaged procurement Yes Autonomous Vehicles projects on \"research on effect field reconstruction (Munitions) technology,\" and \"research on projectile penetration depth algorithms based on the principles of seismic wave imaging\" \"Intelligent co-processor accelerator card processing Yes Other and Shenwei [CPU and operating system] platform testing\" tender announcement Announcement of bid evaluation results for the \"study Yes Predictive of key technology for an aerospace equipment Maintenance and intelligent inspection system\" Logistics \"Cruise missile and UAV simulation system Yes Simulation and development\" outsourcing procurement bid Training announcement Announcement of the winner of the Project 1903 UAV Yes Autonomous Vehicles platform subsystems competitive negotiation (Air) procurement tender Announcement of the winning bid for 2020-6356 low- No N/A emission gas turbine prediction methods research and validation Procurement of Workshop 23: semi-automated blasting No N/A equipment Announcement of finalists for the 500kV diode No N/A experimental research platform automated subsystems project tender \n TABLE 5 5 Examples of Coded \"Intelligent\" Equipment Contracts TENDER NAME (CHINESE) TENDER NAME (ENGLISH) COUNTS AS AI? APPLICATION(S) \n TABLE 6 Number 6 of Equipment Contracts Awarded by PLA Service Branches, April-November 2020 \n TABLE 6 Number 6 of Equipment Contracts Awarded by PLA Service Branches, April-November 2020 PLA SERVICE BRANCH TOTAL NUMBER OF EQUIPMENT CONTRACTS NUMBER OF AI-RELATED EQUIPMENT CONTRACTS PORTION OF PUBLIC CONTRACTS RELATED TO AI Source: CSET corpus of PLA procurement activity. \n Source: CSET corpus of PLA procurement activity (188 contracts awarded by service branches). ment distribution measurement system\" from Xi'an Kuaizhou Measurement and Control Technol- ogy Co., Ltd. (西安快舟测控技术有限公司), an application particularly useful in battle damage assessment. FIGURE 11 FIGURE 11 Number of AI Equipment Contracts Awarded by Number of AI Equipment Contracts Awarded by PLA Service Branches, April-November 2020 PLA Service Branches, April-November 2020 Strategic Strategic Support Force Support Force Ground Force Ground Force Navy Navy Air Force Rocket Force Autonomous Vehicles NUMBER OF AI EQUIPMENT CONTRACTS Intelligence, Surveillance, and Reconnaissance Predictive Maintenance 0 10 20 30 40 50 60 70 Information and Electronic Warfare Simulation and Training Target Recognition Command and Control Other Public procurement records indicate that the PLASSF is focused on using AI for intelligence and data fusion, especially for applications such as weather monitoring, earth imagery, and battle damage assessment. One of its most expensive public contracts in 2020, valued at $1.1 mil-lion, was for an automatic high-altitude image detection system provided by Nanjing Britronics Machinery Co., Ltd. (南京大桥机器有限公司). The company produces more than 60 types of weather radar and satellite imaging equipment, including an \"intelligent high-altitude image detection system\" capable of measuring meteorological phenomena between 36 and 200 NUMBER OF AI EQUIPMENT CONTRACTS Rocket Force Autonomous Vehicles Intelligence, Surveillance, and Reconnaissance Predictive Maintenance 0 10 20 30 40 50 60 70 Information and Electronic Warfare Simulation and Training Target Recognition Command and Control kilometers above ground. Air Force Other 227  In July 2020, PLASSF Unit 63672 also bought a UAV-borne \"frag-", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Harnessed-Lightning.tei.xml", "id": "3e7bae1a363c3cc5bbf57dac463f0515"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "of CLTC for their ongoing support, feedback, and contributions. CLTC would also like to thank the Hewlett Foundation for making this work possible. \n About the cover Image: This image depicts the creation of \"Artifact 1,\" by Sougwen Chung, a New York-based artist and researcher. Artefact 1 (2019) explores artistic co-creation and is the outcome of an improvisational drawing collaboration with a custom robotic unit linked to a recurrent neural net trained on Ms. Chung's drawings. The contrasting colors of lines denote those marks made by the machine and by Ms. Chung's own hand. Learn more at https://sougwen.com/.", "authors": ["Jessica Newman"], "title": "Decision Points in AI Governance", "text": "This paper provides an overview of efforts already under way to resolve the translational gap between principles and practice, ranging from tools and frameworks to standards and initiatives that can be applied at different stages of the AI development pipeline. The paper presents a typology and catalog of 35 recent efforts to implement AI principles, and explores three case studies in depth. Selected for their scope, scale, and novelty, these case studies can serve as a guide for other AI stakeholders -whether companies, communities, or national governments -facing decisions about how to operationalize AI principles. These decisions are critical because the actions AI stakeholders take now will determine whether AI is safely and responsibly developed and deployed around the world. \n Microsoft's AI, Ethics and Effects in Engineering and Research (AETHER) Committee: This case study explores the development and function of Microsoft's AETHER Committee, which has helped inform the company's leaders on key decisions about facial recognition and other AI applications. Established to help align AI efforts with the company's core values and principles, the AETHER Committee convenes employees from across the company into seven working groups tasked with addressing emerging questions related to the development and use of AI by Microsoft and its customers. The case study provides lessons about: • How a major technology company is integrating its AI principles into company practices and policies, while providing a home to tackle questions related to bias and fairness, reliability and safety, and potential threats to human rights and other harms. • Key drivers of success in developing an AI ethics committee, including buy-in and participation from executives and employees, integration into a broader company culture of responsible AI, and the creation of interdisciplinary working groups. \n OpenAI's Staged Release of GPT-2: Over the course of nine months in 2019, OpenAI, a San Francisco-based AI research laboratory, released a powerful AI language model in stages -rather than all at once, the industry norm -in part to identify and address potential societal and policy implications. The company's researchers chose this \"staged release\" model as they were concerned that GPT-2 -an AI model capable of generating long-form text from any prompt -could be used maliciously to generate misleading news articles, impersonate others online, automate the production of abusive content online, or automate phishing content. The case study provides lessons about: • Debates around responsible publication norms for advanced AI technologies. • How institutions can use threat modeling and documentation schemes to promote transparency about potential risks associated with their AI systems. • How AI research teams can establish and maintain open communication with users to identify and mitigate harms. The OECD AI Policy Observatory: In May 2019, 42 countries adopted the Organisation for Economic Co-operation and Development (OECD) AI Principles, a legal recommendation that includes five principles and five recommendations related to the use of AI. To ensure the successful implementation of the Principles, the OECD launched the AI Policy Observatory in February 2020. The Observatory publishes practical guidance about how to implement the AI Principles, and supports a live database of AI policies and initiatives globally. It also compiles metrics and measurement of global AI development and uses its convening power to bring together the private sector, governments, academia, and civil society. The case study provides lessons about: • How an intergovernmental initiative can facilitate international coordination in implementing AI principles, providing a potential counterpoint to \"AI nationalism.\" • The importance of having several governments willing to champion the initiative over numerous years; convening multistakeholder expert groups to shape and drive the agenda; and investing in significant outreach efforts to global partners and allies. The question of how to operationalize AI principles marks a critical juncture for AI stakeholders across sectors. Getting this right at an early stage is important because technological, organizational, and regulatory lock-in effects are likely to make initial efforts especially influential. The case studies detailed in this report provide analysis of recent, consequential initiatives intended to translate AI principles into practice. Each case provides a meaningful example with lessons for other stakeholders hoping to develop and deploy trustworthy AI technologies. \n Introduction Research and development in artificial intelligence (AI) have led to significant advances in natural language processing, image classification and generation, machine translation, and other domains. Interest in the AI field has increased substantially, with 300% growth in the volume of peer-reviewed AI papers published worldwide between 1998 and 2018, and over 48% average annual growth in global investment for AI startups.  1  These advances have led to remarkable scientific achievements and applications, including greater accuracy in cancer screening and more effective disaster relief efforts. At the same time, growing awareness of the significant safety, ethical, and societal challenges stemming from the advancement of AI has generated enthusiasm and urgency for establishing new frameworks for responsible governance. The emerging \"field\" of AI governance -interconnected with such fields as privacy and data governance -has moved through several stages over the past four years. The first stage, which began most notably in 2016, has been characterized by the emergence of AI principles and strategies enumerated in documents published by governments, firms, and civil-society organizations to clarify specific intentions, desires, and values for the safe and beneficial development of AI. Much of the AI governance landscape thus far has taken the form of these principles and strategy documents, at least 84 of which were in existence as of September 2019.  2  The second stage, which initially gained traction in 2018, was characterized by the emergence of efforts to map this proliferation of AI principles  3  and national strategies  4  to identify divergences and commonalities, and to highlight opportunities for international and multistakeholder collaboration. These efforts have revealed growing consensus around a number of central themes, including privacy, accountability, safety and security, transparency and explainability, fairness and non-discrimination, human control of technology, professional responsibility, and promotion of human values.  5  The third stage, which largely began in 2019, has been characterized by the development of tools and initiatives to transform AI principles into practice. While the first two stages helped shape an international AI \"normative core,\" there has been less consensus about how to achieve the goals defined in the principles. Much of the debate about AI governance has focused on 'what' is needed, as laid out in the principles and guidelines, but there has been less focus on the 'how,' the practices and policies needed to implement established goals. This paper argues that the question of how to operationalize AI principles and strategies is one of the key decision points that AI stakeholders face today, and offers examples that may help AI stakeholders navigate the challenging decisions they will face. Efforts are already under way to resolve the translational gap between principles and practice, ranging from tools and frameworks to standards and initiatives that can be applied at different stages of the AI development pipeline.  6  A partial and non-exhaustive list of such efforts can be found in the tables below, organized into six categories: Technical Tools; Oversight Boards and Committees; Frameworks and Best Practices; Standards and Certifications; Regulations; and Institutions and Initiatives. The typology provides examples of the kinds of efforts now under way, specifically for AI research and applications. Efforts included here are well represented in the literature, but were not independently vetted for efficacy or adoption rates. \n EXISTING EFFORTS TO TRANSLATE PRINCIPLES TO PRACTICE \n Technical Tools \n Name Description The Equity Evaluation Corpus  7  A database consisting of thousands of sentences chosen to A library supporting developers and researchers in defending machine-learning models against adversarial threats The AI Fairness 360 Toolkit 10 A toolkit of metrics to check for unwanted bias in datasets and machine-learning models, as well as algorithms to mitigate such bias The TensorFlow Privacy library  11  A library to help train machine-learning models with differential privacy The Accenture AI Fairness Toolkit  12  A tool to help companies detect and eliminate gender, racial, and ethnic bias in AI systems \n Oversight Boards and Committees \n Name Description Microsoft AETHER Committee  13  A committee composed of seven working groups within Microsoft to focus on proactive formulation of internal policies for responding to specific AI-related issues in a responsible way Google Responsible Innovation Team   In addition to the resources named in the tables, countless published papers have helped define best practices for facilitating privacy, security, safety, fairness, and explainability throughout AI development. D E C I S I O N P O I N T S I N A I G O V E R N A N C E \n Institutions and Initiatives The emergence of these resources and initiatives resulted from growing acknowledgment that researchers, companies, and policymakers can, and should, do more to mitigate known and unknown risks related to artificial intelligence. We have reached a crucial juncture in the decades-long history of technological development. AI technologies are now being deployed in critical functions and for consequential ends, including the generation of synthetic media (AI-generated images, text, audio, and video), automated decision-making in the military and other high-stakes environments, and the advancement of powerful tools for security and surveillance. However, it is increasingly understood that AI systems make significant errors. They can be tricked and misused, can make damaging mistakes, and may result in unintended consequences at massive scales. There is also uncertainty about how to establish practices to make reasonable tradeoffs between occasionally conflicting goals. As a result of past mistakes -including the abuse of consumer data and the pursuit of controversial military and foreign contracts -the U.S. technology industry has come under pressure to repair lost trust with consumers and employees. Less clear is whether this pressure, combined with the public expression of principles from influential AI stakeholders, can mature into the implementation of new processes, standards, and policies. The actions AI stakeholders take now to achieve these goals will help determine whether the full upside potential of these technologies can be realized. \n AI DECISION POINTS Many decisions are made throughout the life cycle of an AI system, from which training data to use (if any), to whether and how to test for the robustness of a model against attack. All of these decisions affect the reliability, security, and impacts of the system. Transparency about how and why these decisions were made has become an important element of accountability. AI decision-makers have to consider tradeoffs and priorities, and weigh which decisions are likely to have an outsized impact. The concept of \"AI decision points\" introduced in this paper refers to concrete actions taken by an AI stakeholder (organization, company, government, employees, etc.) that were not predetermined by existing law or practice, and that mark a meaningful shift in behavior from previous practice, for the purpose of shaping the development and use of AI. These decisions are catalysts of broader inflection points leading to new strategies and opportunities for the organization. Because AI decision points reflect efforts to actively shape the field, they are poised to have a disproportionately large influence on the future trajectory of AI. Tracking decision points can provide insight into existing tensions and challenges, and into how the field is evolving to address them. Identifying decision points can also be a useful way to focus governance efforts, as it narrows a crowded solution space and highlights opportunities to make influential decisions. AI technologies are valuable in part because of their ability to automate elements of complex decision-making. This utility can at times obscure the importance of human decisions in shaping the design and use of AI technologies. However, AI systems are shaped by countless decisions, related not only to technical design, but also to human and institutional preferences. Attention to decision points underscores that specific AI trajectories are not inevitable, and there are opportunities to make decisions that will support safer and more responsible development and use. This analysis is primarily relevant to people interested in shaping future trajectories of AI, including AI researchers, policymakers, and industry executives. That is not to suggest that there is a single right answer, or that decisions should be made in isolation or only by people in positions of power. Indeed, the decisions described in the case studies were all made through consultative and iterative processes. AI technologies impact many people, but can have disparate effects across different communities. Legitimate governance efforts include engagement with diverse stakeholders as well as impacted communities. The case studies in this paper are intended to highlight key levers that are currently being used to shape the future of AI, to provide examples from which the field can draw, and to support further analysis of the effectiveness and desirability of such efforts as tools of AI governance. There is no need for AI stakeholders to start from scratch in their endeavor to operationalize AI principles. No single actor can accomplish the mitigation of AI threats in isolation; stakeholders need to coordinate and cooperate with each other, which is much easier with improved collective understanding of ongoing efforts. \n On the Need to Operationalize AI Principles Decisions about how to operationalize AI principles and strategies have potential to shift the AI landscape toward a more safe and responsible future trajectory. A literature review of recent developments in AI and AI governance reveals growing awareness of the need to support principled implementation efforts.  31  It has become difficult for AI stakeholders to ignore the many AI principles and strategies in existence. Even companies and organizations that have not defined their own principles may be expected or required to adhere to those adopted by governments. This growing \"universality\" will lead to increased pressure to establish methods to ensure AI principles and strategies are realized. Meaningful implementation efforts are likely to be critical in maintaining trust with users and the public. Due to technological, organizational, and regulatory lock-in effects, early efforts -those that fill governance gaps to establish new standards and best practices -are likely to be especially influential. The goals defined within the various AI principles and strategies are both practical and ambitious, and often build upon other human rights frameworks, rights and laws, and nonrights perspectives. These documents have emerged from all sectors, including governments, corporations, and civil society. The stated purpose of many of the principles is to forge trust: between governments and citizens, between corporations and consumers or users, and between AI researchers and the general public. However, in the wake of the \"techlash\" -a term used to describe growing public animosity toward large technology companies -earning trust is not straightforward. Trust is rather reserved for those entities that provide compelling values and motivations for their work and back them up with meaningful actions. Pledges of intention, made explicit through principles and strategies, can be extremely valuable. They offer clarity and guidance about paths to pursue, and provide insight into the priorities of firms and governments. They can also help to hold actors accountable. For example, Google's AI Principles include a list of AI applications the company will not pursue, including weapons, surveillance technologies that violate international norms, or any technologies that contravene international law and human rights.  32  Nonetheless, people are growing unsatisfied with promises alone, and are calling out companies and governments for failing to act on their principles.  33 / 34  While there is a logical trajectory for stakeholders to move from \"principles to practice\" (while facilitating an ongoing feedback loop between the two), this transition is only possible with sufficient agreement about what safe and responsible AI development and use entails. AI principles do not have to be universally accepted or non-controversial; cultural and political dynamics necessarily include variability. Rather, how principles, strategies, and guidelines are translated into action will need to accord with international laws and standards, and with unique national, local, and organizational needs. Implementation efforts are therefore most easily understood with sufficient context about their origin and scope. The case studies below attempt to provide such specificity. Each case can be understood as a meaningful example (though not necessarily an ideal model that all others should replicate), offering relevant lessons for AI stakeholders hoping to shape the future of AI. They showcase how technology companies and governments are grappling with key decisions in AI governance, and highlight some of the successes and challenges faced by key AI stakeholders around the world. \n Case Studies The following three case studies provide analysis of recent, consequential decisions that were made to operationalize AI principles in order to support safe and responsible AI development. The case studies highlight the example of an ethics and impact advisory committee re-shaping engineering and research practices at a top technology company; an AI research laboratory that experimented with staged release and other accountability measures for the dissemination of a potentially harmful AI model; and the first prominent example of intergovernmental and multistakeholder AI coordination through a common global framework. These case studies were selected based upon their scope, scale, and novelty. Many of the existing translational efforts described in the tables on pages 4-8 are narrowly focused on a particular AI challenge, such as reducing the threat from adversarial attacks or mitigating algorithmic bias. These tools are critical, but this paper is concerned with efforts of a broader scope. Each example described below simultaneously addresses numerous interconnected issues that affect the likelihood of ensuring safe and responsible AI development. The scale of each of these examples is also notable. While many of the tools listed previously are open-source and could be used broadly, it is not clear whether many of them have been widely adopted. However, the three examples below represent shifts in practices and polices that were made across entire companies and organizations, with evidence of spillover effects to other parts of the AI ecosystem already present. Lastly, these case studies were selected because they represent novel shifts in well-established behaviors. Microsoft's decision to adopt an ethics and effects advisory committee marked a reportedly first-of-its-kind effort to formalize review of societal impact throughout the lifecycle of AI technologies at the core of a company's business model. OpenAI's decision to adopt accountability measures and the staged release of an AI model represented a marked shift from the open publication norms in AI and machine-learning communities. And the OECD's decision to establish an intergovernmental hub for AI governance serves as a counterpoint to AI nationalism and rhetoric about an \"AI race\" between nations. The novelty of these examples makes them more impactful because they signify inflection points, or significant changes, in the AI landscape. Moreover, each case study explores the implementation of AI principles for different kinds of key AI stakeholders: a large technology company, an AI research lab, and governments around the world. D E C I S I O N P O I N T S I N A I G O V E R N A N C E \n CASE STUDY I CAN AN AI ETHICS ADVISORY COMMITTEE HELP CORPORATIONS ADVANCE RESPONSIBLE AI? Understanding the Role of the Microsoft AETHER Committee \n INSIGHTS • Large multinational companies have an outsized impact on trends in AI development and deployment, but have not universally adopted new practices or oversight committees to help ensure their technologies will be beneficial. • Microsoft aims to be a leader in responsible AI, and has established the AETHER Committee with the intention of operationalizing the company's AI principles into its engineering and business practices. • The AETHER Committee is facilitating internal deliberation about controversial use-cases, providing channels for concerned employees, and incentivizing research in the areas of its working groups, including safety, security, and accountability. • The AETHER Committee attributes its success in part to executive-level support, regular opportunities for employee and expert engagement, and integration with the company's legal team. \n How the AETHER Committee Emerged and How it Works In a March 2018 email to all employees, Satya Nadella, the chief executive officer of Microsoft, described the importance of AI innovation to the long-term success of the company, noting that ensuring responsibility was critical as the technology progressed.  35  He announced that Brad Smith, president of Microsoft, and Harry Shum, executive vice president of the company's AI and Research group, would be establishing the AI and Ethics in Engineering and Research (AETHER) Committee. He described the Committee as a way to bring together senior leaders from throughout the company to build proactive internal policies and address specific issues raised. He told employees, \"AETHER will ensure our AI platform and experience efforts are deeply grounded within Microsoft's core values and principles and benefit the broader society.\" By early 2019, the meaning of the acronym \"AETHER\" was expanded to stand for AI, Ethics and Effects in Engineering and Research.  36  At a time when numerous technology firms face pressure from the growing \"techlash,\" several companies have established AI ethics committees of various kinds to try to manage growing challenges. However, these committees are still viewed with some suspicion, and in some cases have been called out as \"AI ethics-washing.\" 37 Do such committees reflect calculated public relations moves, or can they be effective for translating AI principles into practice? The Microsoft AETHER Committee represents a novel experiment to restructure practices across the company's AI engineering and research teams, and may provide lessons on challenges and opportunities for other organizations. This analysis is based upon interviews and conversations with Microsoft executives and employees, a review of public documentation and media, and a 2019 presentation about AETHER given at UC Berkeley. In 2018, Microsoft outlined six principles to guide the company's AI development in a book titled The Future Computed: Artificial Intelligence and its role in society.  38  The principles laid out in the book include fairness, reliability, privacy and security, inclusiveness, transparency, and accountability. On their own, the principles are not particularly noteworthy; they reinforce common AI principles defined by countless organizations. However, while other efforts to institutionalize guiding values for AI into industry have struggled,  39  Microsoft has made an extensive effort to restructure engineering practices and policies around its values, and has had some notable successes. This case study examines how the AETHER Committee (hereinafter referred to as AETHER) is designed, how it functions, and what its impacts have been. AETHER helps organize internal talent at Microsoft, calling upon employees from different backgrounds to address controversial or complex cases and to proactively develop internal policies for safe and ethical AI development. AETHER members write reports about specific questions, outlining the issues at stake and the costs and benefits of a particular action, and then present the information, along with recommendations, to Microsoft's senior leadership team. These recommendations provide guidance and contribute to practices, policies, and positions at the company. AETHER is primarily an internal advisory committee, but has occasionally consulted with outside experts. AETHER also works alongside another group at Microsoft, the Office of Responsible AI, which is based in the legal department and assists with compliance efforts. In a keynote address in November 2019, the chair of AETHER, Eric Horvitz, explained that the role of AETHER is to \"advise the senior leadership team on policies around sensitive issues when it comes to AI products and services.\"  40  He added, \"it's already had quite a significant effect on gating and guiding Microsoft technologies and how it works with customers in different parts of the world when it comes to these technologies.\" To date, little information about AETHER or its impact has been publicly disclosed. This case study aims to shed some light on the structure and achievements of the AETHER Committee at a relatively early stage. \n Working Groups AETHER is organized into seven working groups, generally composed of 5-7 people at their core and twenty people in total, including co-chairs (typically top experts in the field,) a core subgroup of committee members, and an expanded subgroup with representatives from every major division in the company. The working groups are dedicated to the following focus areas: • Sensitive Uses, to assess automated decision-making that can have a major impact on people's lives, such as the denial of consequential services, risks to human rights, and risks of harm; • Bias and Fairness, to investigate potential impacts of AI systems on minority and vulnerable populations; • Reliability and Safety, to ensure AI systems function as intended and are robust against adversarial attacks; • Human Attention and Cognition, to monitor algorithmic attention-hacking and abilities of persuasion; • Intelligibility and Explanation, to promote transparency into how machine-learning and deep-learning models process data and reach decisions; • Human-AI Interaction and Collaboration, to study how people can better and more productively engage with AI systems; and • Engineering Best Practices, to recommend education, training, best practices, processes, and tooling to support each stage of the AI system development cycle and ensure that Microsoft teams are equipped and motivated to apply Microsoft principles for responsible AI. In addition to serving an advisory role, the working groups also address technical challenges and publish externally. For example, the Bias and Fairness working group has looked at Microsoft's facial recognition service to tackle problems associated with gender recognition for women with darker skin. The group developed new tools to probe the system and better understand and address its limitations. Similarly, the working group on Engineering Practices published a paper on AI threat-modeling,  41  and the Intelligibility and Explanation working group developed an open-source Python package, InterpretML, intended to train interpretable machine learning models and help explain black-box systems.  42  The Sensitive Uses working group specifically has a mission to undertake \"analysis and deliber- AETHER is also called upon to resolve questions about the implications of Microsoft's AI products and services, and about how to manage sensitive customer requests or uses. For example, AETHER reviews how customers might use (and misuse) Microsoft's AI products. According to Horvitz, this resulted in the company changing its course as early as April 2018.  43  \"Significant sales have been cut off. And in other sales, various specific limitations were written down in terms of usage, including 'may not use data-driven pattern recognition for use in face recognition or predictions of this type.'\" Horvitz added, \"It's been an intensive effort … and I'm happy to say that this committee has teeth.\" Evidence to verify his claim can be found in the examples of rejected requests described below. \n Impacts and Challenges Facial recognition technology has been a keen area of focus for AETHER, and contributed to Microsoft's calls for regulation of this emerging use of AI. In April 2019, it was revealed that Microsoft rejected a request from a sheriff's department in California to install facial recognition technology in officers' cars and body cameras because the company determined that to do so would constitute a human rights concern, given the high likelihood of bias against minorities.  44  Microsoft has also rejected requests from foreign governments to install facial recognition on surveillance cameras due to concerns this could suppress freedom of assembly. Microsoft has, however, allowed many uses of its AI technologies based upon recommendations from AETHER. For example, the company facilitated the use of facial recognition technology within an American prison after AETHER assessed that its uses would be limited and likely to improve safety conditions. More controversially, Microsoft has supported a facial recognition company in Israel called AnyVision, which has been criticized for identifying and tracking Palestinians in the West Bank.  45  In November 2019, following concerns about potential human rights abuses, Microsoft hired former United States Attorney General Eric Holder to investigate whether AnyVision appropriately complies with the company's principles for the use of AI and facial recognition technology.  46  Other efforts will also test Microsoft's commitment to principled implementation of AI in the coming years. In October 2019, in a surprise upset to frontrunner Amazon, Microsoft won an historic $10 billion contract with the Department of Defense (DoD) to transform the military's cloud computing systems. Google dropped out of the competition for the contract in 2018 in part because the work was determined to go against the company's AI principles, which state they will not contribute to the use of AI in weaponry. Notably, Microsoft's AI principles do not preclude such uses, as long as the systems are reliable, safe, and accountable. Microsoft has strong convictions about the importance of supporting the U.S. military with its technologies, but has not shied away from addressing the sensitivities associated with its decisions.  47  AETHER's Sensitive Uses working group reportedly defined a policy on the company's sale of AI technologies to the Department of Defense following an executive retreat on the matter. AETHER Chair Eric Horvitz has further defined key challenges related to the use of AI in military applications, providing greater insight into how the company is thinking about its role and responsibilities.  48  Horvitz said in a November 2019 keynote address for the Bulletin of the Atomic Scientists that \"inescapable errors from AI systems\" must be taken into consideration and that efforts should be made to allow time for \"human reflection, input, and intervention.\" Horvitz did not say that humans should always remain in the loop, but he did say that removing people from these positions of oversight should only be done with \"wisdom and caution.\" He also called for consideration of an array of challenges related to AI, including the impossibilities of fully testing AI capabilities in realistic, operational settings; the rise of unexpected behaviors in the complexities of interactions among AI systems; preparation for adversarial attacks on AI systems; vigilance against new forms of persuasion and deception; collaboration with potential adversaries to minimize instabilities and facilitate human oversight; investment in human-AI interaction technologies; and vigilance for the assertion of ethical principles for AI that changes the nature of war. These challenges do not directly translate to operational practices however, and Microsoft will likely face a higher degree of scrutiny over its uses of AI in the coming years. Microsoft has been providing technologies to the Department of Defense for more than 40 years. However, not all Microsoft employees have been on board with the company's willingness to support the U.S. military. A group called Microsoft Workers 4 Good, whose mission is \"to empower every worker to hold Microsoft accountable to their stated values,\" has called on Microsoft leadership to end certain contracts. For example, in February 2019, the group sent a letter calling on Brad Smith and Satya Nadella to end a contract through which the company would provide its HoloLens augmented reality technology to the U.S. Army to support war fighting. The letter stated, \"We are alarmed that Microsoft is working to provide weapons technology to the U.S. Military, helping one country's government 'increase lethality' using tools we built. We did not sign up to develop weapons, and we demand a say in how our work is used.\" 49 Nadella defended the company's decision, stating, \"We made a principled decision that we're not going to withhold technology from institutions that we have elected in democracies to protect the freedoms we enjoy.\" Smith has also defended the contract with the Defense Department, arguing, \"All of us who live in this country depend on its strong defense. . . . We want the people of this country and especially the people who serve this country to know that we at Microsoft have their backs. They will have access to the best technology that we create.\" AETHER's Sensitive Uses working group will continue to assess potentially controversial use cases that emerge in this arena.   50  The company has also added a responsible AI module to educational materials that are required for all employees, and implemented a Responsible AI Champions program, which trains people to be champions for safe and trustworthy AI within their division. \n Features of Success AETHER has several features that make it effective. The inclusion of top executives on the committee signals its perceived importance to the overall mission of the company. AETHER also facilitates input; for example, the committee has established an \"Ask AETHER\" phone line, which any Microsoft employee can use to raise a concern about an AI technology they are working on or have seen in development or use. Additionally, AETHER's interdisciplinary nature has helped make it more inclusive and far-reaching. Lastly, AETHER helps Microsoft engage proactively with external AI policy developments. For example, Microsoft has repeatedly called for government regulation of facial recognition technology.  51  The development of a self-regulating body may be seen at least partially as an attempt to prevent external regulation of the company's practices. However, this does not appear to be the primary motivating factor; in addition to the more recent work on the regulation of facial recognition, Microsoft has supported state and federal privacy regulations for years.  52  The company seems to have realized that its internal pivot to expand its focus on AI 53 -and to reorganize the company to integrate AI throughout its divisions and products  54   We owe it to the future to help ensure that these values survive and even flourish long after we and our products have passed from the scene.  55  The decision to establish AETHER stands out because it provides a clear signal to employees, users, clients, and partners that Microsoft intends to hold its technology to a higher standard. AETHER shows one pathway by which companies can empower employees to voice concerns and work toward new company practices and policies supporting the responsible development and use of AI. AETHER is poised to have an outsized effect on the trajectory of AI because it has reshaped a major AI company's practices for vetting and monitoring the AI systems it builds and sells. As of April 2020, this committee-based structure appears to be unique among AI technology companies. Moreover, to date, it has not received significant outside attention because Microsoft has only spoken about it on occasion and has not yet released significant public materials about its processes, though it may do so in the future. As other companies grapple with AI engineering design decisions and the implementation of AI principles, the AETHER Committee provides a valuable example from which other organizations can learn. \n CASE STUDY II CAN SHIFTING PUBLICATION NORMS FOR AI RESEARCH REDUCE AI RISKS? Accountability Measures and the Staged Release of an Advanced Language Model \n INSIGHTS • Researchers and organizations will increasingly face decisions about how to responsibly publish AI research that could be misused or cause unintentional harm. • This decision will rarely be a dichotomy between \"publish\" and \"don't publish.\" \"Staged release\" is one method along the spectrum as it allows the publication of AI research or technologies over time. • The test case revealed that there is conflicting evidence about the effectiveness of restraint in keeping a technology from being misused. However, taking the time to partner with other organizations and stakeholders to conduct research into the uses and impacts of a new technology may be valuable. • Responsible disclosure is not merely about the degree of openness, but also about the implementation of accountability measures, including the use of documentation efforts, discussion of potentially harmful uses and impacts in research papers, and facilitating communication prior to and following the release of new models. \n Making Technological (and Cultural) Waves On February 14, 2019, San Francisco-based AI research laboratory OpenAI announced it had developed an unsupervised language model -a statistical tool that finds patterns in human language and can be used to predict text or audio -capable of generating long-form text from any prompt it received.  56  The model, called GPT-2, came less than a year after the release of GPT, another language model that performed well across a variety of tasks, including tests of reading comprehension, translation, and sentiment analysis. GPT-2 is the next iteration of that model, but was trained on 10X the amount of data (eight million web pages) and has 10X the parameters (1.5 billion). The model was initially trained to predict the next word within a given set of text, but was eventually able to generate many sentences of synthetic text on any topic. The technological advancement was noteworthy in part because the model does not rely upon supervised learning on task-specific datasets, but is capable of learning language processing tasks -machine translation, reading comprehension, and summarization -without explicit supervision. Beyond the technological achievement, the launch of this powerful AI language model was noteworthy for how OpenAI's research team chose to release it. In what was called \"an experiment in responsible disclosure,\" OpenAI decided not to release the complete trained model, but instead to release a much smaller and simpler model (124 million parameters)  57  along with a technical paper.  58  Over the course of nine months, the organization carried out a \"staged release,\" with progressively larger models released throughout 2019, in May (355 million parameters), August (774 million parameters), and November (1.5 billion parameters, the largest model). The company used the time in between releases to prepare research and documentation exploring the societal and policy implications of the technology alongside the release of the technical papers. The move sparked debate among AI researchers, who have largely embraced open publishing norms.  59  For example, it has become common for AI researchers to publish early papers on arXiv, a free, open, and non-peer-reviewed publication platform for scientific papers. Funding requirements and job performance metrics have additionally incentivized teams to publish quickly and frequently. The culture around open disclosure in AI can in some regards be viewed as an extension of debates within computer security, where openness is often explicitly utilized for security purposes by allowing the discovery a greater number of software vulnerabilities.  60  Researchers at OpenAI are concerned that this may be a problematic baseline for dual-use research, where the risks of misuse and unintended consequences may outweigh the benefits. Some researchers have suggested that, in certain instances, the AI field should take lessons from domains that exhibit a greater degree of caution in publication than computer security, such as biosecurity and nuclear security.  61  This case study explores the reasons behind OpenAI's decision to use a staged release for GPT-2, the reaction the company received, and some examples of how norms around responsible disclosure of advanced AI models appear to have shifted. The analysis was based upon interviews, presentations, and feedback from OpenAI employees, as well as a review of public documentation and media. \n The Rationale In April 2018, OpenAI released the OpenAI Charter, which explained that the company's mission is \"to ensure that artificial general intelligence (AGI)-by which we mean highly autonomous systems that outperform humans at most economically valuable work-benefits all of humanity.\" Tenets of the charter include the broad distribution of benefits, long-term safety, technical leadership, and cooperative orientation. The charter also states: \"We are committed to providing public goods that help society navigate the path to AGI. Today this includes publishing most of our AI research, but we expect that safety and security concerns will reduce our traditional publishing in the future.\"  62  In other words, responsible disclosure is a core principle for OpenAI, and the staged release of GPT-2 was an effort to implement this value into organizational practice. The leaders of OpenAI do not take the charter lightly; an in-depth report about the company described the charter as \"sacred\", informing all performance reviews, strategies, and actions.  63  The decision to only release a smaller version of GPT-2 at the outset stemmed from concern that the model could be used maliciously, primarily for the generation of scalable, customizable, synthetic media for political or economic gains. For example, OpenAI noted that its model could be used to generate misleading news articles, impersonate others online, automate the production of abusive content online, and automate phishing content. OpenAI wanted to give people more time to adapt and react; they wanted researchers to have more time to work on mitigating risks, and for the public to realize that greater diligence may be required to discern what is true. The company's concern was not unjustified; AI technology had previously been used for \"deepfakes\" -where someone's face is inserted into existing video content -for synthetic pornography and content intended to undermine political figures. OpenAI's research and policy team believed caution was warranted in this case because even the first language model they released was able to generate text that seemed relatively authentic. For example, when prompted with two sentences -\"A train carriage containing controlled nuclear materials was stolen in Cincinnati today. Its whereabouts are unknown\"  64  -the model, on the first try, wrote seven additional paragraphs of synthetic text, which began: The incident occurred on the downtown train line, which runs from Covington and Ashland stations. In an email to Ohio news outlets, the U.S. Department of Energy said it is working with the Federal Railroad Administration to find the thief. \"The theft of this nuclear material will have significant negative consequences on public and environmental health, our workforce and the economy of our nation,\" said Tom Hicks, the U.S. Energy Secretary, in a statement. \"Our top priority is to secure the theft and ensure it doesn't happen again.\" Although this portion of the text was realistic, the model was far from perfect. OpenAI described several overarching failure modes of the model, including repetitive text, illogical combinations, and unnatural topic switching. In general, the company's researchers found that the model performed better on topics that were well represented in the training data. They also found the model did well at mimicking genres when provided with a specific subset of training data from a particular domain, such as the Amazon Reviews dataset. Throughout the staged release process, parties outside of OpenAI experimented with the released models and found interesting additional uses. For example, a doctor at Imperial College in London retrained GPT-2 on a scientific database with over 30 million biomedical literature citations over a 24-hour period, and found the system could then create its own realistic and comprehensive scientific paper abstracts based merely on a title.  65  Others have used GPT-2 to generate poetry and write short stories.  66  A site called Talk to Transformer uses a simple interface to encourage people to experiment with the model using any custom prompt they choose.  67  In addition to the staged release, OpenAI took other measures to support the goal of responsible publication. For example, along with releasing the code for GPT-2 on GitHub, OpenAI also published a \"model card,\" which explains details about how the model was built and evaluated, includes research findings on biases in the model, provides an open communication channel for feedback, and gives recommended uses.  68  This idea was inspired by a paper originally published in late 2018 called \"Model Cards for Model Reporting,\"  69  which introduced the framework of model cards -short documents accompanying trained machine-learning models -to support greater transparency about a model's performance characteristics and other important information about how and why the model was built in a particular way. \n The Final Release and Risk-Reduction Efforts OpenAI released the largest version of GPT-2, with 1.5 billion parameters, on November 5, 2019.  70  In a blog accompanying the release, the team wrote, \"While there have been larger language models released since August, we've continued with our original staged release plan in order to provide the community with a test case of a full staged release process. We hope that this test case will be useful to developers of future powerful models, and we're actively continuing the conversation with the AI community on responsible publication.\" In addition to the model, the company also released an updated report on social impacts,  71  as well as an updated model card.  72  In the report, OpenAI's team provided insight into their process and findings from the staged release.  73  Part of this research took place in-house, including the release of research related to bias in the model's outputs. OpenAI also partnered with four outside organizations to focus on challenges of detection, misuse, and bias. Partner organizations included Cornell University; The Middlebury Institute of International Studies Center on Terrorism, Extremism, and Counterterrorism; The University of Oregon; and The University of Texas at Austin. These partnerships enabled investigation into potential malicious uses, detection of synthetic text, human responses to generated text, and biases in GPT-2 outputs. Justifying the company's cautious stance, the research found that people find GPT-2 outputs to be convincing, that the system can easily be fine-tuned for misuse, and that detection of synthetic text will be a longterm challenge. The OpenAI team also used the time between releases to engage with outside stakeholders, including by contributing to ongoing work carried out by the Partnership on AI on developing responsible publication norms. As another precaution, the company communicated with outside researchers who were creating similar language models -a critical choice, as the practice of staged release works best in an environment of cooperation. OpenAI shared a specific GPT-2 email address and encouraged engagement from students and researchers. This feedback mechanism was reportedly used on numerous occasions. For example, the AI company Hugging Face decided against releasing its internal language models following discussion with OpenAI. The company describes itself as having a firm belief in open-source and knowledge sharing. Hugging Face has written, \"Without open-source, the entire field faces the risk of not making progress and concentrating capabilities at the hands of a couple of massive players (be they corporations or states), without anyone else but them being able to understand, compete or control.\"  74  However, at the same time, the company acknowledges that its technology is not neutral and that action is required to facilitate its positive impact in the world, including considering the potential malicious uses of new releases. The company published an ethical analysis along with its latest conversational AI language model.  75  Similarly, when Salesforce released the language model CTRL, they also published an analysis discussing potential societal implications.  76  OpenAI monitored uses of GPT-2 in the real world. They did this in part by tracking websites and forums with a history of promoting disinformation, as well as by having discussions with policymakers in defense and intelligence agencies. The team did not find significant evidence of misuse, though they acknowledge that advanced persistent threats (APTs) are particularly difficult to monitor. They also admitted to finding evidence of discussion of misuse, including a small number of cases of explicit public plans for misuse, though they did not believe the relevant actors had sufficient resources and capabilities to carry out the plans. \n Reception OpenAI has suggested that the experiment in staged release was relatively successful, mostly because it helped spur discussion about AI publication norms. Nonetheless, it was criticized for several reasons described below. A key argument for transparency is that advanced language models can be used to support efforts to detect other fake media. For example, a research team at the University of Washington released a different language model in June 2019, which they described as a \"state-of-the-art generator of neural fake news.\"  77  These researchers disputed that releasing their model would be dangerous, arguing that the capabilities of GROVER and GPT-2 were not sufficiently human-like and that the lack of controllability of the content makes these models less useful for adversaries. Moreover, they argued that releasing the model in its entirety had benefits for threat modeling and defense. However, the GROVER researchers did discuss their work with people at OpenAI, and were encouraged to conduct in-depth threat modeling to inform their decision about how to release their model. In November 2019, the cybersecurity company FireEye published a blog post revealing they were using GPT-2 to detect social media posts generated as part of information operations.  78  The company's researchers had fine-tuned the GPT-2 model on millions of tweets attributed to the Russian Internet Research Agency. This taught the model to create tweets that resembled the source, for example, \"It's disgraceful that people have to waste time, energy to pay lip service to #Junk-Science #fakenews.\" However, the authors of the blog pointed out that, while they were able to use the system to detect malicious activity to spread propaganda, this model could also be used to lower the barrier to entry for actors hoping to engage in such malicious activity at scale. Other criticisms of OpenAI's decision to delay the full release of GPT-2 stemmed from a perceived betrayal to core processes of peer review and the scientific method, as well as the culture of openness that has been central to AI progress for decades.  79  The belief in the value of open-source software to support replication and application has long been a central component of development in the field, key to avoiding another \"winter\" or period of diminished enthusiasm.  80  This critique was particularly sharp for OpenAI, which was founded on ideals of transparency and openness.  81  Critics of OpenAI's decision contended that the partial disclosure meant that independent researchers were not able to evaluate and verify claims made about the system, or to build upon previous findings. Some suggested that the company was overstating the uniqueness of its tool and engaging in fear-mongering.  82  Indeed, in August 2019, two graduate students from Brown University announced they had replicated the 1.5 billion-parameter GPT-2 model by modifying the open-source Grover model.  83  Their purpose was to critique the strategy of staged released, which they argued only makes sense if a model is difficult to replicate. Instead, they proved that similar results could be recreated for roughly $50k by two masters students who had never created a language model before. The duo claimed that they were making a morally justified choice, writing, \"Because our replication efforts are not unique, and large language models are the current most effective means of countering generated text, we believe releasing our model is a reasonable first step towards countering the potential future abuse of these kinds of models.\" Along with these critiques, OpenAI was also celebrated for the staged release decision, in particular for its impact on encouraging AI developers to think comprehensively about the implications of their work.  84  Norms within scientific communities can be powerful mechanisms to promote responsible innovation, and are particularly important in cutting-edge fields that have a relative lack of guidance and regulatory frameworks. OpenAI's efforts have inspired other AI stakeholders -for example, the Partnership on AI -to consider the responsible publication of high-stakes AI research.  85  Although there has been a long history of efforts to instill responsibility and ethics in technological developments, it is still rare for researchers and companies to offer transparent accounts of the risks stemming from their work. For example, while it is typical for AI researchers to state the positive uses of their models in papers, it is uncommon to see discussion about possible misuses or unintended consequences. Most technology companies are still wary to discuss the societal impact, risks, and potential negative implications of their products and services. Today, there is less doubt about the risks that AI technologies pose. It has been well documented that AI systems that optimize for user engagement can promote extremism and filter bubbles,  86  that AI-enabled synthetic media (including deepfakes) can be used to generate malicious content,  87  and that AI systems can easily be tricked  88  and can make deadly mistakes.  89   \n Impacts Even if delaying the release of the largest GPT-2 model did little to prevent misuse of language models in general, OpenAI's decision jump-started a larger conversation about best practices and responsible publication norms. The paper accompanying the release of the largest GPT-2 model concludes, \"We hope GPT-2 as a case will help the AI community navigate publications in omni-use AI research.\"  90  Their hope appears to have become reality, as others have subsequently adopted similar strategies. For example, in January 2020, Google announced a new conversational agent called Meena, which integrates an astonishing 2.6 billion parameters. Meena is capable of engaging in conversations that are more realistic than current state-of-the-art systems. Importantly, Google decided not to release an external research demo of the system due to concerns about safety and bias, noting that the company is still evaluating the risks and benefits associated with giving the public access to this powerful tool.  91  Similarly, in November 2019, Microsoft announced a language model called DialoGPT, but did not include a public sampling interface in order to minimize opportunistic misuse.  92  In September 2019, Salesforce released CTRL, a language model containing 1.63 billion parameters.  93  The researchers published their model in full and stated, \"Openness and replicability are central aspects of the scientific ethos that, prima facie, suggest the release of complete scientific research results. We reify these principles by releasing all trained CTRL models.\" However, their technical paper includes a section on \"the ethics of large language models,\" which reveals that they took responsible disclosure seriously. The researchers also published a second paper that delved more deeply into responsible innovation and the inadequacy of self-governance.  94  The Salesforce research team was encouraged to engage on these issues because of the precedent set by others. Rather than rely on self-governance, they consulted with experts at the Partnership on AI, who have been working on the issue of responsible publication norms with members from OpenAI and other stakeholders. The Salesforce researchers carried out a technology foresight exercise that included scenario planning as a way to imagine worrisome possible uses of their technology. When Salesforce did release the CTRL model openly on GitHub, they included a code of conduct and a set of questions to \"further encourage users to reflect on norms and responsibilities associated with models that generate artificial content.\" Moreover, to facilitate post-release monitoring of CTRL, Salesforce actively observes how others are using CTRL. The team set up a dedicated email account and encourages users of the model to share their uses, pose questions, and suggest solutions. Other organizations have opted for a more extreme stance on disclosing research results. The Machine Intelligence Research Institute (MIRI), whose mission is to ensure that the creation of smarter-than-human intelligence has a positive impact, released an update on its research directions in November 2018, which included a shift toward \"nondisclosed-by-default research.\"  95  The organization explained that the majority of its research results would no longer be published externally in an effort to prevent others from using their findings to build more capable and dangerous systems. This shift, which occurred three months before OpenAI's decision, was met with some confusion and skepticism, though OpenAI policy director Jack Clark called the move \"useful\" at the time, suggesting that it \"generates data for the community about what the consequences are of taking such a step.\"  96  AI researchers will continue to need to weigh the costs and benefits of different disclosure models. Some accountability measures, such as model cards, are likely to be beneficial in most cases, whereas appropriate degrees of openness will vary depending on the scale and scope of potential harm. Researchers at Oxford have proposed a theoretical framework to help inform this assessment. The framework addresses the security value of disclosure and includes factors that contribute to whether providing access to certain research will make it easier to cause harm, or easier to provide protections against harm. The factors include: counterfactual possession (i.e. where the would-be attacker acquires the relevant knowledge even without publication), absorption and application capacity (i.e. if publication of the research will only benefit attackers to the extent that they are able to absorb and apply the research), resources for solution-finding (i.e. given the disclosure, how many additional individuals or organizations will work on finding a solution?), availability of effective solutions (i.e. is there a good defense against misuse?), and the difficulty/cost of propagating a solution (i.e. even where a solution exists in theory, it might be difficult or costly to propagate that solution).  97  This case study highlights how decisions about how to disclose omni-use AI research can have a lasting impact on the future of the field. Over the course of 2019, OpenAI undertook an experiment in responsible disclosure for advanced artificial intelligence by releasing progressively more capable versions of its powerful language model. The company used the time in between releases to monitor uses, engage with partner organizations on particular research questions, and promote awareness of impacts. This decision has already had a significant impact on other AI researchers and organizations, and is poised to have an outsized effect on future trajectories of AI development. It is becoming more normal to see open deliberations about risks and tradeoffs inherent to AI systems, even as new models are made publicly available around the world. Keeping research as open as possible, while minimizing the potential for misuse and harm, is a delicate balance. \n CASE STUDY III CAN A GLOBAL FOCAL POINT FOR AI POLICY ANALYSIS AND IMPLEMENTATION PROMOTE INTERNATIONAL COORDINATION? The Launch of the OECD AI Policy Observatory • Stakeholders largely agree on high-level interests such as AI safety and transparency, but there will continue to be differences in the implementation of AI principles within different political and economic environments. • Evidence-based AI policy guidance, metrics, and case studies to support domestic AI policy decisions are in high demand, and the OECD AI Policy Observatory is poised to become a prominent source of guidance globally. • The function of the OECD AI Policy Observatory as an intergovernmental hub for AI governance may serve as a counterpoint to AI nationalism and the prominence of \"AI race\" rhetoric between nations. \n The First Intergovernmental Standard for AI It has become common to hear about the \"race for AI supremacy\" between nations, 98 despite known dangers associated with such rhetoric.  99  In particular, the focus on national advantage can undercut efforts to support a global approach to AI governance. This may be problematic for AI technologies, which are becoming widely available around the world, and whose effects will be far-reaching. Given the dynamics around national competition, it would have been difficult to predict that dozens of nations, and especially the U.S., China, and Russia, would agree to a common set of guiding principles for AI. However, that is what happened in June 2019, when the G20 gave unanimous support to the OECD AI Principles. This case study explores the events that led up to that occasion (and what was left out of the agreement), the policy mechanisms planned to support the implementation of the principles around the world, and what the developments mean for the global governance of AI. This analysis is based upon interviews and feedback from OECD employees and expert group members, and a review of public documentation and media. On May 22, 2019, forty-two countries adopted the first intergovernmental standard on artificial intelligence.  100  The guidelines came in the form of a legal recommendation that included five principles and five recommendations from the OECD, led by the Committee on Digital Economy Policy (CDEP). In announcing this initiative, OECD Secretary-General Angel Gurría stated, \"These Principles will be a global reference point for trustworthy AI so that we can harness its opportunities in a way that delivers the best outcomes for all.\" Unlike other sets of AI principles, the OECD AI Principles are an intergovernmental agreement; although the process to develop them brought together multiple stakeholders, the adherents are governments, making this the first intergovernmental standard for AI in existence. All 36 OECD member countries signed on to the OECD AI Principles, including many nations at the forefront of AI development, among them the United States, Australia, France, Germany, Korea, Estonia, Israel, Japan, and the United Kingdom. Several non-member countriesincluding Argentina, Brazil, Colombia, Costa Rica, Peru and Romania -also signed on. The European Commission additionally supported the Principles, and Ukraine was added to the list of signatories in October 2019. When the Group of Twenty (G20) released AI Principles one month later, it was noted that they were drawn from the OECD AI Principles,  101  expanding the list of supportive countries to include China, India, and Russia, among others. Established in 1961 as an intergovernmental organization, the Paris-based OECD today has 36 member countries in Europe, North America, South America, and Asia. All of the members are market-based democracies, and the organization has been criticized for being a \"club of mostly rich countries.\" 102 However, the OECD has been expanding its membership over the years, and has started to partner with more developing countries, including Brazil, India, and South Africa.  103  The OECD focuses on economic and social policy analysis and statistics, and also develops international policy standards, such as the OECD Privacy Guidelines and the OECD AI Principles. The OECD has made 177 policy recommendations since 1964, and has a history of promoting international cooperation on the safety of consequential, dual-use technologies, including genetic engineering and nuclear technology. However, that the organization would play such a prominent role in the global governance of AI was not a given. Other institutions, such as the United Nations and International Telecommunication Union (ITU), have also emerged as forums for advancing the global governance of AI, but have not similarly garnered support for ethics and governance principles to date.  104  The ability of the OECD AI Principles to attract the support of dozens of governments was several years in the making. The OECD's Committee on Digital Economy Policy (CDEP) began considering a recommendation on AI as early as 2016, and in May 2018, this committee decided to establish an AI expert group to scope AI principles. The expert group (AIGO) launched in September 2018 with 50 members, led by Wonki Min, Vice Minister of Science and ICT of Korea and chair of the OECD's Digital Economy Committee. Many countries were represented among the AIGO members, including: Australia, Canada, Denmark, Finland, France, Germany, Hungary, Japan, Korea, Mexico, Netherlands, New Zealand, Poland, Russia, Singapore, Slovenia, Sweden, Switzerland, Turkey, UAE, United Kingdom, United States, and the European Commission. In addition to government representatives, the group invited experts from industry, academia, and civil society, including from Microsoft, Google, Facebook, MIT, the Harvard Berkman Klein Center, OpenAI, IEEE, the AI Initiative of The Future Society, and the World Privacy Forum, among others. Contributions from other experts around the world were also taken into account. Broad, multistakeholder engagement and significant enthusiasm and dedication from group members were pivotal for the OECD's success in advancing a global governance framework for AI. The group helped to scope the principles over four in-person meetings in different locations around the world (Paris, Cambridge, and Dubai), and with several teleconference calls in between. This group initially identified the five principles and recommendations, which were then expanded upon further.  105  Some governments played a more integral role in enabling the OECD to meaningfully tackle the AI governance challenge. For example, the OECD's work on AI began in April 2016 at the G7 ICT Ministerial meeting in Takamatsu, Japan, where the host nation encouraged the OECD to prioritize AI and identify policy priorities for international cooperation. Japanese ministers described the need for international principles to guide research and development of AI, and proposed an initial set of principles for consideration that included transparency, user assistance, controllability, security, safety, privacy, ethics, and accountability. Japan's Ministry of Internal Affairs and Communications (MIC) also provided financial support for an OECD conference, \"AI: Intelligent Machines, Smart Policies,\" a landmark event held in Paris in October 2017.  106  The MIC also supported the development of the book Artificial Intelligence in Society, published by the OECD in June 2019, which provides greater background about the emergence of the principles and describes policy initiatives under way around the world. Moreover, Japan proposed and led the G20 discussion on AI, facilitating the agreement to the G20 AI Principles at the Ministerial Meeting on Trade and Digital Economy in Osaka, Japan. When Japanese Prime Minister Shinzo Abe announced the G20 AI Principles, he confirmed that they would guide the G20's commitment to a human-centered approach to AI. Japan's ongoing leadership and support for a global AI governance framework has been critically important. Interviewees have confirmed that if Japan had not held the G20 presidency in 2019, the G20 AI Principles would not exist. Japan will continue to be a key stakeholder in the global governance of AI, and has indicated its intention of continuing to support the OECD's efforts. The United States government has also been a vocal supporter of the OECD AI Principles. In a speech at the OECD forum and ministerial council meeting in Paris, Michael Kratsios, then Deputy Assistant to the President for Technology Policy (and now Chief Technology Officer), referred to the moment as \"a historic step,\" by which, \"America and likeminded democracies of the world will commit to common AI principles reflecting our shared values and priorities.\"  107  Kratsios commented, \"The United States has long welcomed the work of the OECD to develop AI principles. Across multiple G7 and OECD fora, we worked closely with our strong international partners to advance discussions and draft the principles.\" The United States in particular has focused on the importance of identifying the shared values of democratic nations for the development of AI. \n G20 Support Held on June 28, 2019, the G20 Osaka Summit brought together G20 leaders in Japan to address major global economic challenges. Invited international organizations included the United Nations and the World Bank. This annual meeting, primarily intended to coordinate responses to global economic turbulence, had an increased focus on the role of digitalization and technological innovation. Nonetheless, at the outset of the Summit, it was not widely expected that the group would reach an agreement on guidelines related to artificial intelligence. Yet, by the conclusion, the group released the G20 AI Principles,  108  which were accepted by consensus and established a common set of principles for the responsible stewardship of trustworthy AI. The development of a common set of principles for AI development among nations with diverse and at times conflicting interests, including the U.S. and China, was a shocking and important achievement. Within a culture of national competition for AI leadership, the G20 Principles represented a first step toward collective action on AI governance at the global scale. A footnote in the G20 AI Principles notes that that they were drawn from the OECD principles and recommendations for artificial intelligence. However, the G20 AI Principles are in fact largely identical to the OECD AI Principles. Both documents include the following language: To this end, they should provide meaningful information, appropriate to the context, and consistent with the state of art: a. to foster a general understanding of AI systems; b. to make stakeholders aware of their interactions with AI systems, including in the workplace; c. to enable those affected by an AI system to understand the outcome; and, d. to enable those adversely affected by an AI system to challenge its outcome based on plain and easy-to-understand information on the factors, and the logic that served as the basis for the prediction, recommendation or decision. 4. Robustness, security and safety a. AI systems should be robust, secure and safe throughout their entire lifecycle so that, in conditions of normal use, foreseeable use or misuse, or other adverse conditions, they function appropriately and do not pose unreasonable safety risk. b. To this end, AI actors should ensure traceability, including in relation to datasets, processes and decisions made during the AI system lifecycle, to enable analysis of the AI system's outcomes and responses to inquiry, appropriate to the context and consistent with the state of art. c. AI actors should, based on their roles, the context, and their ability to act, apply a systematic risk management approach to each phase of the AI system lifecycle on a continuous basis to address risks related to AI systems, including privacy, digital security, safety and bias. \n 5. Accountability AI actors should be accountable for the proper functioning of AI systems and for the respect of the above principles, based on their roles, the context, and consistent with the state of art. G20 support for the OECD AI Principles was particularly meaningful for several reasons. The G20 countries account for about 85% of global economic output, 75% of global exports, and two-thirds of the world's population.  109  The G20 has become a premier forum for international cooperation and coordination, and consensus support from G20 countries expands the reach of the AI principles around the world. Although the primary focus of the G20 is the global economy, recent meetings have increasingly been used to discuss pressing foreign policy challenges, ranging from sustainability to human rights abuses. However, G20 support did not extend to the full content of the OECD AI Recommendation. The second section of the recommendation, \"National policies and international co-operation for trustworthy AI,\" reveals some differences of opinion among the OECD and G20 countries. The recommended policies include (at a high level) investing in AI research and development, fostering a digital ecosystem for AI, shaping an enabling policy environment for AI, building human capacity and preparing for labor market transformation, and international co-operation for trustworthy AI. An appendix to the \"G20 Ministerial Statement on Trade and Digital Economy\" states, \"The G20 supports the Principles for responsible stewardship of Trustworthy AI in Section 1 and takes note of the Recommendations in Section 2.\"  110  In other words, while the G20 endorsed the OECD AI Principles, the support did not explicitly extend to the recommendations for governments. This fact underscores the need to track and analyze the operationalization of AI principles and strategies globally. While the development of intergovernmental principles for AI was remarkable, how these goals are realized is not likely to be uniform. This gap is one of the most common criticisms of the OECD AI Principles: that they are too high-level to lead to real policy change. This is a relevant critique of all voluntary principles, and should not be taken lightly. Like all other G20 declarations, the G20 AI Principles are non-binding and their full impact remains to be seen. The G20 has been criticized for not doing more than \"naming and shaming\" when actors fail to uphold their commitments. The OECD Recommendation is also not legally binding, and the OECD lacks enforcement capabilities. Nonetheless, other OECD Recommendations have been quite influential as political commitments, in particular for setting international standards and helping with the design of national legislation. For example, the OECD Privacy Guidelines influenced the design of privacy laws around the world. In this case, the OECD developed an additional plan to support the practical and policy relevance of its principles. \n The AI Policy Observatory At the end of 2019, the OECD announced plans for an AI Policy Observatory to help countries implement the principles and recommendations.  111  Formally launched in late February 2020, the Observatory is envisioned as \"a platform to share and shape public policies for responsible, trustworthy and beneficial AI.\" As the second pillar, the Observatory will provide analysis of AI policy in key areas, including science, health, jobs, and transportation, among others. Dashboards dedicated to each sector provide information about related policy initiatives, live updated news feeds, and how different countries are prioritizing research in that area. The third pillar is focused on trends and data, and includes OECD metrics and measurement, as well as live data from partners, including news about AI development and policy. For example, you can explore data and visualizations that depict trends in publications from different countries, AI research collaborations and networks, the growth of AI subtopics, and AI skills migration between countries. These resources are intended to help provide a basis for evidence-based policymaking. The fourth pillar relates to national AI strategies, policies, and initiatives, from national governments and other AI stakeholders. This includes a database, visualizations, and analysis of over 300 AI policy examples, which are contributed directly from governments through a survey. This resource serves as a unique repository that enables countries and organizations to compare AI policies at a much more granular level than has previously been possible. For example, the dashboards show all of the relevant initiatives under way in a given country, the prioritization and investment amounts in different areas, and relevant governmental bodies and research institutions. Numerous global data streams are used to inform the insights on OECD.ai, including news media, scientific papers, patents, job market data, and business data. The Observatory is also informed by the Microsoft Academic Graph, which collects information about scientific publications, citation relationships, authors, institutions, journals, conferences, and fields of study; and the LinkedIn Economic Graph, which uses the company's data to highlight trends related to talent migration, hiring rates, and in-demand skills by region. The Observatory makes use of AI techniques, including social network analysis and classification, to process and analyze this data before presenting it in a relatively interactive and accessible way. The data is intended not only to provide a view of the past and present, but also to provide views of future trends. Global policymakers are the primary audience of the resource, and it is hoped that they can utilize the insights to inform evidence-based practices and policy development. The OECD AI Policy Observatory already has support from numerous governments and multistakeholder organizations. For example, UNESCO (The United Nations Educational, Scientific, and Cultural Organization) aims to support the OECD's AI governance priorities by translating policy recommendations into actionable opportunities for the communities they work with in their field offices around the world. The European Commission also intends to support the Observatory, especially in its work on metrics and measurement and the collection of national AI strategies and policies. The governments of Germany, Japan, and the United States have continued to publicly voice their support of the Observatory and seek alignment between their national AI policy initiatives and the OECD AI Principles.  113  Another emerging international initiative called the Global Partnership for AI (GPAI), led by French president Emmanuel Macron and Canadian prime minister Justin Trudeau, will coordinate with the OECD to provide a forum for global debate on AI.  114  A new OECD Network of Experts on AI (ONE AI) will also advise and support the work of the AI Policy Observatory (replacing and building upon the former expert group, AIGO).  115  ONE AI is a multi-stakeholder and multi-disciplinary advisory group that is composed of more than 100 AI experts split into three working groups. The groups provide input and implementation ideas for AI policy issues, support the Observatory's four pillars, and facilitate information exchange and collaboration between the OECD and other international initiatives and organizations focusing on AI. For example, at the group's first meeting on February 27, 2020, discussions centered on classifications of AI systems and ongoing efforts to implement practices that support safe and human-centric AI. It was recognized at that meeting that governments will need to adopt new policies and practices across numerous sectors to ensure the principles and recommendations are adopted. There is some indication that this is already happening. For example, the European Commission, which has adopted the OECD AI Principles, is developing legislative proposals for AI and intends to facilitate a coordinated European approach to the human and ethical implications of AI.  116   G20 support for the OECD AI Principles and Observatory is also likely to continue. In his concluding declaration before the other global leaders, Japanese Prime Minister Shinzo Abe expressed the shared commitment to \"human-centered\" AI and suggested that AI is a \"driving force\" behind the sustainable development goals (SDGs). Prime Minister Abe acknowledged \"the growing importance of promoting security in the digital economy and of addressing security gaps and vulnerabilities.\" He also discussed more broadly the increasing importance of digital technologies and the cross-border flow of data for innovation and the global economy. Saudi Arabia assumed the G20 Presidency in 2020 and has clarified its goals of supporting the development of inclusive and trustworthy AI.  118  The implementation of AI principles within different sectors was a focus at the first G20 Digital Economy Task Force meeting in February 2020 in Riyadh, Saudi Arabia. Interviewees involved with the G20 suggested that digitalization and AI are likely to be on the agenda for many years to come. Despite the hurdles ahead, the decision to support a common set of AI principles marked a critical shift in the global landscape. The OECD AI Principles represent the first time that nations around the world committed to a common set of guidelines that provide shared understanding and goals for how to shape future trajectories of AI. The OECD Principles are also notable compared to other AI principles for highlighting a broader range of issues, including reducing economic, social, gender and other inequalities, protecting natural environments and internationally recognized labor rights, and applying a systematic risk management approach to each phase of the AI system lifecycle on a continuous basis. Perhaps most importantly, the OECD AI Policy Observatory will, by design, ensure that the Principles are linked to concrete policy mechanisms that can be implemented by nations around the world. This will help to operationalize the AI Principles at a large scale. Moreover, the Observatory is poised to become a prominent site for multistakeholder dialogue on AI. The Observatory's openly available online materials and regular events will facilitate international coordination and collaboration at a scale that has previously been difficult to sustain. Other initiatives at the OECD, such as the OECD Global Parliamentary Network -a learning hub on AI for legislators and parliamentary officials -will help ensure a comprehensive approach to achieving the organization's goals. The OECD AI Principles achieved a feat few would have thought possible. The United States signed on at a time of relative aversion to international coordination in other policy arenas. China was part of a consensus agreement to support the effort more broadly, and other countries are welcome to add their support. The year 2019 brought the first intergovernmental standard for AI and a new \"global reference point\" for AI governance into the future. Moreover, the OECD AI Policy Observatory has been identified by many of the world's governments as the new global focal point for translating AI principles into practical policy guidelines. Sustained attention to the coming successes and challenges of these efforts will be important to further understand the value of this model. At this point, the AI Policy Observatory is the first initiative of its kind and is poised to have an outsized impact on the trajectory of AI around the world. \n Conclusion AI stakeholders face countless difficult decisions about how, why, and for whom to develop and use AI technologies. The concept of \"AI decision points\" provides a framework to prioritize decisions that were not predetermined by existing law or practice, and that mark a meaningful shift in behavior from previous practice for shaping the development and use of AI. These decisions are catalysts for broader inflection points and are reshaping future trajectories of AI. Identifying and tracking AI decision points can provide insight into the evolution of the field, and help focus governance efforts by identifying key policy levers. Decisions about how to operationalize AI principles and strategies are currently faced by nearly all AI stakeholders, and are determining practices and policies in a meaningful way. There is growing pressure on AI companies and organizations to adopt implementation efforts, and those actors perceived to verge from their stated intentions may face backlash from employees, users, and the general public. The transition from principles to practice in the AI field has become shorthand for the broad desire to see plans put into action, through organizational shifts, design decisions, or the implementation of new policies. Nonetheless, this is understood to be a difficult and time-consuming process: legal frameworks are shifting, and no taxonomies of best practices have been agreed upon. This paper aims to provide a step in that direction, by compiling examples of efforts to bridge the gap between principles and practice, and by focusing on case studies that are meaningful examples of this translation process. The case studies reveal insights about how companies and intergovernmental institutions are approaching decisions about the operationalization of AI principles, as well as the challenges and successes each effort has faced. The example of Microsoft's AETHER Committee highlights the importance of executive-level support in shaping an organization's commitment to responsible AI development, as well as the value of employee and expert engagement, and integration with the company's legal team. AETHER's structure has enabled the establishment of new organizational policies and the prioritization of new areas of research. This model would be less valuable if internal dynamics at Microsoft were to diminish AETHER's decision-making power, or if the regulatory landscape were to shift to reduce individual companies' ability to make decisions about the design and sale of AI technologies. The review of OpenAI's staged release of the GPT-2 AI language model highlights the spectrum between \"open\" and \"closed\" AI research, as well as the difficulties of preventing consequential technologies from being misused. This case study exemplifies how companies can make use of multiple synergistic accountability measures, including documentation efforts, discussion of potentially harmful uses and impacts in research papers, and facilitating communication prior to and following the release of new AI models. Finally, the examination of the OECD AI Policy Observatory highlights how, despite challenges in achieving international cooperation, governments remain motivated to support global governance frameworks for AI. While the Observatory is still in its infancy, governments, like companies, are seeking guidance on actions they can take to realize their objectives for responsible AI. Though it may one day be superseded by other intergovernmental forums or treaties, the Observatory has emerged as an important resource for nations to share evidencebased AI policy guidance and metrics, and to facilitate global dialogue. Together, the case studies shine a light on how influential AI stakeholders are navigating the \"third stage\" of AI governance, translating principles into practice. Given implementation efforts are dependent on context, case studies and ethnographic accounts can help illuminate how the field is shifting to address concerns about the significant safety, security, and societal challenges accompanying the evolution of artificial intelligence. Decisions made today about how to operationalize AI principles at scale will have major implications for decades to come, and AI stakeholders have an opportunity to learn from existing efforts and to take concrete steps to ensure we build a better future. help determine when automatic systems demonstrate biases toward certain races and genders InterpretML 8 An open-source library developed by Microsoft for training interpretable machine-learning models and explaining black-box systems The Adversarial Robustness 360 Toolbox 9 \n INSIGHTS• International coordination and cooperation on AI begins with a common understanding of what is at stake and what outcomes are desired for the future. That shared language now exists in the Organisation for Economic Co-operation and Development (OECD) AI Principles, which are being leveraged to support partnerships, multilateral agreements, and the global deployment of AI systems. \n Image used with permission from a presentation byKarine Perset, Administrator on Digital Economy and Artificial Intelligence Policy in the OECD Directorate for Science, Technology, and Innovation, and by Adam Murray, International Affairs Officer in the U.S. Department of State Office of International Communications and Information Policy, delivered to members of the Partnership on AI in April 2020. The map shows that there has been a broad global commitment to the OECD AI Principles, but also highlights that many African nations have not yet been involved. \n ation about AI systems involved in sensitive uses, including automated recommendations that can have deep impact on people's lives.\" As determining what constitutes a \"sensitive use\" of AI is not always a straightforward task, Microsoft leans on influential precedents, including the Universal Declaration of Human Rights, the Guiding Principles on Business and Human Rights, and Microsoft's own Human Rights Policy. Concrete examples of sensitive uses include the denial of credit, employment, education, or healthcare services; the use of surveillance systems and other AI systems that pose risks to personal freedoms, privacy, and human rights; and the risk of AI systems creating significant physical or emotional harm. This working group relays recommendations to Microsoft's senior leadership team, who may agree or disagree with the findings. If a new recommendation is agreed to, the working group can then establish a new policy for the company, providing an important pathway for AETHER to establish precedent. This ability to generate new company policies for Microsoft is one of AETHER's critical functions, as it can lead to the restructuring of engineering or other institutional processes. Greater external transparency about this process will help clarify how many of the recommendations do in fact generate new company policies. \n a broader ecosystem intended to facilitate a culture of responsible AI development. Other Microsoft efforts that reinforce this commitment include a group called Fairness, Accountability, Transparency and Ethics in AI (FATE), which consists of nine researchers \"working on collaborative research projects that address the need for transparency, accountability, and fairness in AI.\" Additionally, in late 2019, following an AETHER recommendation, Microsoft became the first company in the world to launch a Responsible AI Standard, which is required and informs AI development throughout a system's lifecycle. In 2020, in collaboration with AETHER's Bias and Fairness working group and a group of nearly 50 engineers from numerous technology companies, Microsoft developed an AI ethics checklist for engineers to use throughout the product development process. AETHER has managed to produce new research insights, establish new company policies, and help inform decisions about sensitive uses of Microsoft's AI technology with ongoing support from executives and employees within the company. It has succeeded in part because it is only one piece of \n -means that consumers must have trust in those technologies in order to trust Microsoft. As Microsoft President Brad Smith and Microsoft Senior Director of Communications and External Relations, Carol Ann Browne, wrote in the 2019 book Tools and Weapons: These issues are bigger than any single person, company, industry, or even technology itself. They involve fundamental values of democratic freedoms and human rights. The tech sector was born and has grown because it has benefited from these freedoms. \n 1. Inclusive growth, sustainable development and well-being Stakeholders should proactively engage in responsible stewardship of trustworthy AI in pursuit of beneficial outcomes for people and the planet, such as augmenting human capabilities and enhancing creativity, advancing inclusion of underrepresented populations, reducing economic, social, gender and other inequalities, and protecting natural environments, thus invigorating inclusive growth, sustainable development and well-being.2. Human-centered values and fairnessa. AI actors should respect the rule of law, human rights and democratic values, throughout the AI system lifecycle. These include freedom, dignity and autonomy, privacy and data protection, non-discrimination and equality, diversity, fairness, social justice, and internationally recognized labor rights. b. To this end, AI actors should implement mechanisms and safeguards, such as capacity for human determination, that are appropriate to the context and consistent with the state of art.3. Transparency and explainabilityAI Actors should commit to transparency and responsible disclosure regarding AI systems. \n The Observatory supports dialogue among global multistakeholder partners, publishes practical guidance to implement the AI Principles, and supports a live database of AI policies and initiatives globally. It also compiles metrics and measurement of AI development to serve as a baseline for policy development, and uses its convening power to bring together the private sector, governments, academia, and civil society. The Observatory's resources have all been made publicly available at OECD.ai.The Observatory is structured around four main pillars, each of which has its own goals, partners, and online resources. The first pillar centers on the OECD AI Principles. The objectives of the pillar include explaining what each principle means and why it matters, and providing practical guidance to governments, including resources to support implementation efforts. Dashboards dedicated to each principle provide concrete information about related AI policy initiatives, policy instruments, and scientific research. This is considered to be a key goal of the Observatory. As Karine Perset, Administrator on Digital Economy and Artificial Intelligence Policy in the OECD Directorate for Science, Technology, and Innovation, emphasized, \"The principles were the beginning, and now we are focusing on implementation. The AI Policy Observatory is one of our major endeavors to move from principles to action and implementation, and help policymakers in this journey.\" 112   \n Moreover, the OECD AI Principles are referenced as a meaningful source for AI standards in the August 2019 Plan for Federal Engagement in Developing Technical Standards and Related Tools, developed by the US Department of Commerce National Institute of Standards and Technology (NIST). 117", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Decision_Points_AI_Governance.tei.xml", "id": "03f8bbf8b337b24e63cc5d00cc99c5bb"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Artificial Intelligence (AI) presents novel policy challenges that require coordinated global responses. 2   Standards, particularly those developed by existing international standards bodies, can support the global governance of AI development. International standards bodies have a track record of governing a range of socio-technical issues: they have spread cybersecurity practices to nearly 160 countries, they have seen firms around the world incur significant costs in order to improve their environmental sustainability, and they have developed safety standards used in numerous industries including autonomous vehicles and nuclear energy. These bodies have the institutional capacity to achieve expert consensus and then promulgate standards across the world. Other existing institutions can then enforce these nominally voluntary standards through both de facto and de jure methods. AI standards work is ongoing at ISO and IEEE, two leading standards bodies. But these ongoing standards efforts primarily focus on standards to improve market efficiency and address ethical concerns, respectively. There remains a risk that these standards may fail to address further policy objectives, such as a culture of responsible deployment and use of safety specifications in fundamental research. Furthermore, leading AI research organizations that share concerns about such policy objectives are conspicuously absent from ongoing standardization efforts. Standards will not achieve all AI policy goals, but they are a path towards effective global solutions where national rules may fall short. Standards can influence the development and deployment of particular AI systems through product specifications for, i.a., explainability, robustness, and fail-safe design. They can also affect the larger context in which AI is researched, developed, and deployed through process specifications. The creation, 3 dissemination, and enforcement of international standards can build trust among participating researchers, labs, and states. Standards can serve to globally disseminate best practices, as previously witnessed in cybersecurity, environmental sustainability, and quality management. Existing international treaties, national mandates, government procurement requirements, market incentives, and global harmonization pressures can contribute to the spread of standards once they are established. Standards do have limits, however: existing market forces are insufficient to incentivize the adoption of standards that govern fundamental research and other transaction-distant systems and practices. Concerted efforts among the AI community and external stakeholders will be needed to achieve such standards in practice. 2 See, e.g., Brundage, Miles, et al. \"The malicious use of artificial intelligence: Forecasting, prevention, and mitigation.\" Future of Humanity Institute and the Centre for the Study of Existential Risk.", "authors": ["Peter Cihon"], "title": "Standards for AI Governance: International Standards to Enable Global Coordination in AI Research & Development", "text": "Ultimately, standards are a tool for global governance, but one that requires institutional entrepreneurs to actively use standards in order to promote beneficial outcomes. Key governments, including China and the U.S., have stated priorities for developing international AI standards. Standardization efforts are only beginning, and may become increasingly contentious over time, as has been witnessed in telecommunications. Engagement sooner rather than later can establish beneficial and internationally legitimate ground rules to reduce risks in international and market competition for the development of increasingly capable AI systems. In light of the strengths and limitations of standards, this paper offers a series of recommendations. They are summarized below: • Leading AI labs should build institutional capacity to understand and engage in standardization processes. This can be accomplished through in-house development or partnerships with specific third-party organizations. • AI researchers should engage in ongoing standardization processes. The Partnership on AI and other qualifying organizations should consider becoming liaisons with standards committees to contribute to and track developments. Particular standards may benefit from independent development initially and then be transferred to an international standards body under existing procedures. • Further research is needed on AI standards from both technical and institutional perspectives. Technical standards desiderata can inform new standardization efforts and institutional strategies can develop paths for standards spread globally in practice. • Standards should be used as a tool to spread a culture of safety and responsibility among AI developers. This can be achieved both inside individual organizations and within the broader AI community. \n Table of Contents Executive Summary  \n Introduction Standards are an institution for coordination. Standards ensure that products made around the world are interoperable. They ensure that management processes for cybersecurity, quality assurance, environmental sustainability, and more are consistent no matter where they happen. Standards provide the institutional infrastructure needed to develop new technologies, and they provide safety procedures to do so in a controlled manner. Standards can do all of this, too, in the research and development of artificial intelligence (AI). Market incentives will drive companies to participate in the development of product standards for AI. Indeed, work is already underway on preliminary product and ethics standards for AI. But, absent outside intervention, standards may not serve as a policy tool to reduce risks in the technology's development.  Leading AI research 4  organizations that share concerns about such risks are conspicuously absent from ongoing standardization efforts. To positively influence the development trajectory of AI, we do not necessarily need to design new 5 institutions. Existing organizations, treaties, and practices already see standards disseminated around the world, enforced through private institutions, and mandated by national action. Standards, developed by an international group of experts, can provide legitimate global rules amid international competition in the development of advanced AI systems. These standards can support trust 6 among developers and a consistent focus on safety, among other benefits. Standards constitute a language and practice of communication among research labs around the world, and can establish guardrails that help support positive AI research and development outcomes. Standards will not achieve all AI policy goals, but they are an important step towards effective global solutions. They are an important step that the AI research community can start leading on today. The paper is structured as follows. Section 2 discusses the need for global coordination on AI policy goals and develops at length the use of international standards in achieving these goals. Section 3 analyzes the current AI standards landscape. Section 4 offers a series of recommendations for how the AI community, comprising technical researchers, development organizations, governance researchers, can best use international standards as a tool of global governance. \n Standards: Institution for Global Governance 2A. The need for global governance of AI development AI development poses global challenges. Government strategies to incentivize increased AI research within national boundaries may result in a fractured governance landscape globally, and in the long-term threaten a 7 race to the bottom in regulatory stringency. In this scenario, countries compete to attract AI industry through 8 national strategies and incentives that accelerate AI development, but do not similarly increase regulatory oversight to mitigate societal risks associated with these developments. These risks associated with lax regulatory 9 oversight and heated competition range from increasing the probability of biased, socially harmful systems to 10 existential threats to human life.  11  These risks are exacerbated by a lack of effective global governance mechanisms to provide, at minimum, guardrails in the competition that drives technological innovation. Although there is uncertainty surrounding AI capability development timelines, AI researchers expect capabilities to match human performance for 12 many tasks within the decade and for most tasks within several decades. These developments will have 13 transformative effects on society. It is thus critical that global governance institutions are put in place to steer 14 these transformations in beneficial directions. International standards are an institution of global governance that exists today and can help achieve AI policy goals. Notably, global governance does not mean global government: existing regimes of international coordination, transnational collaboration, and global trade are all forms of global governance. Not all policy 15 responses to AI will be global; indeed, many will necessarily account for local and national contexts. But 16 international standards can support policy goals where global governance is needed, in particular, by (1) spreading beneficial systems and practices, (2) facilitating trust among states and researchers, and (3) encouraging efficient development of advanced systems. 17 First, the content of the standards themselves can support AI policy goals. Beneficial standards include those that support the security and robustness of AI, further the explainability of and reduce bias in algorithmic decisions, and ensure that AI systems fail safely. Standards development on all three fronts is underway today, as discussed below in Section 3 . Each standard could also reduce long-term risks if their adoption shifts funding away from opaque, insecure, and unsafe methods. Additional standards could shape processes of research and 18 development towards beneficial ends, namely through an emphasises on safety practices in fundamental research. In addition to stipulating safe processes, these standards, through their regular enactment and enforcement, could encourage a responsible culture of AI development. These claims are developed further in Section 4 . Second, international standards processes can facilitate trust among states and research efforts. International standards bodies provide focal organizations where opposing perspectives can be reconciled. Once created and adopted, international standards can foster trust among possible competitors because they will provide a shared governance framework from which to build further agreement. This manner of initial definition, openness among research efforts that is \"unambiguously good\" in light of these concerns. In practice, credible 23 public commitments to specific standards can provide partial information about the practices of otherwise disconnected labs. Furthermore, particular standards that may emerge over time could themselves define appropriate levels and mechanisms of openness. Third, international standards can encourage the efficient development of increasingly advanced AI systems. International standards have a demonstrated track record of improving global market efficiency and economic surplus via, i.a., reduced barriers to international trade, greater interoperability of labor and end-products, and eliminated duplicated effort on standardized elements. International standards could support these outcomes for AI as well, e.g., with systems that can deploy across national boundaries and be implemented using consistent processes and packages by semi-skilled AI practitioners. Increased efficiency in deployment will drive further resources into research and development. Some in the AI community may be concerned that this will increase the rate at which AI research progresses, thereby encouraging racing dynamics that disincentivize precaution. Yet standards can help here too, both through object-level standards for safety practices with 25 enforcement mechanisms and by facilitating trust among developers. if a specialized agency is developed in the future, previously established standards can be incorporated at that time.  28  There are two existing international standards bodies that are currently developing AI standards. First is a joint effort between ISO and IEC. To coordinate development of digital technology standards, ISO and IEC established a joint committee (JTC 1) in 1987. JCT 1 has published some 3000 standards, addressing everything from programming languages, character renderings, file formats including JPEG, distributed computing architecture, and data security procedures. These standards have influence and have seen adoption and 29 publicity by leading multinational corporations (MNCs). For example, ISO data security standards have been widely adopted by cloud computing providers, e.g., Alibaba, Amazon, Apple, Google, Microsoft, and Tencent. \n 30 The second international standards body that is notable in developing AI standards is the IEEE Standards Association. IEEE is an engineers' professional organization with a subsidiary Standards Association (SA) whose most notable standards address protocols for products, including Ethernet and WiFi. IEEE SA also creates process standards in other areas including software engineering management and autonomous systems design. Its AI standardization processes are part of a larger IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems.  31  A third international standards body may become increasingly relevant for AI in the future: the ITU. The ITU has historically played a role in standards for information and communications technologies, particularly in telecommunications. It has a Focus Group on Machine Learning for Future Networks that falls within this telecommunications remit. Following the 2018 AI for Good Global Summit, it has also created a Focus Group on AI for Health, \"which aims inter alia to create standardized benchmarks to evaluate Artificial Intelligence algorithms used in healthcare applications.\" Given the ITU's historically narrower scope, however, this paper 32 does not consider the organizations' work further. \n 2C. Advantages of international standards as global governance tools International standards present a number of advantages in encouraging the global governance of AI. This section distills these advantages into three themes. First, international standards have a history of guiding the development and deployment of technical systems and shaping their social effects across the world. Second, international standards bodies privilege the influence of experts and have tested mechanisms for achieving 28 Pre-existing standards have been referenced in international treaties, e.g., The International Maritime Organization's Safety of Life at Sea Treaty references ISO product standards. Koppell, Jonathan G. S. World Rule : Accountability, Legitimacy, and the Design of Global Governance. Chicago: University of Chicago Press, 2010.  29  See, e.g., Rajchel, Lisa. 25 years of ISO/IEC JTC 1 . ISO Focus+, 2012. https://www.iso.org/files/live/sites/isoorg/files/news/magazine/ISO%20Focus%2b%20(2010-2013)/en/2012/ISO%20Fo cus%2b%2c%20June%202012.pdf .  30  For Amazon, these include ISO 27001, 27017, and 27018 from JTC 1 as well as the ISO 9001 quality management process standard. See the link associated with each company: Alibaba , Amazon , Apple , Google , Microsoft , and Tencent .  31  See IEEE's Ethics in Action website .  32  See ITU's AI for Good website . consensus among them on precisely what should be in standards. Third, existing treaties, national practices, and transnational actors encourage the global dissemination and enforcement of international standards. \n 2Ci. Standards Govern Technical Systems and Social Impact Standards are, at their most fundamental \"a guide for behavior and for judging behavior.\" In practice, 33 standards define technical systems and can guide their social impact. Standards are widely used for both private and public governance at national and transnational levels, in areas as wide ranging as financial accounting and nuclear safety. Many forms of standards will impact the development of AI. 34 Consider a useful typology of standards based on actors' incentives and the object of standardization. Actors' incentives in standards can be modeled by two types of externalities: positive, network externalities and negative externalities. With network externalities, parties face a coordination game where they are incentivized to 35 cooperate. For example, a phone is more useful if it can call many others than if it can only communicate with 36 the same model. Institutions may be necessary to establish a standard in this case but not to maintain the standard in practice, as the harmony of interests obviates enforcement. For the purposes of this paper, consider these standards \"network standards.\" Negative externalities are different; a polluter burdens others but does not internalize the cost itself. Standards here face challenges: individuals may have an incentive to defect in what could be modeled as a Prisoner's Dilemma. In the pollution case, it is in the interest of an individual business to disregard a pollution standard 37 absent additional institutions. But this interest can favor cooperation if an institution creates excludible benefits and an enforcement mechanism. External stakeholders are important here as well: institutions to enable enforced standards are incentivized by demand external to those who adopt the standards. In practice, governments, companies, and even public pressure can offer such incentives; many are explored in Section 2Ciii . For example, the ISO 14001 Environmental Management standard requires regular and intensive audits in order to obtain certification, which in turn brings reputational value to the companies that obtain it. In general, for 38 such standards, institutions are needed for initial standardization and subsequent enforcement. For the purposes of this paper, consider these standards \"enforced standards.\" Enforcement can take multiple forms, from regulatory mandates to contractual monitoring. Certification of adherence to a standard is a common method of enforcement that relies on third-parties, which can be part of government or private entities.  39  Self-certification is also common, whereby a firm will claim that it complies with a standard and is subject to  33  Abbot and Snidal, \"International 'standards' and international governance,\" p. 345.  34  Brunsson, Nils and Bengt Jacobsson. \"The contemporary expansion of standardization\" in A World of Standards . Oxford: Oxford University Press, 2000, 1-18.  35  Abbott and Snidal, \"International 'standards' and international governance.\" 36 This scenario can have distributional consequences as well, where one party gains more from the standard, but ultimately all are better off from cooperation.  37  Abbot and Snidal, \"International 'standards' and international governance.\" 38 Prakash, Aseem, and Matthew Potoski. The voluntary environmentalists: Green clubs, ISO 14001, and voluntary environmental regulations . Cambridge University Press, 2006.  39  Ibid. future enforcement from a regulator. Compliance monitoring can occur through periodic audits, applications 40 for re-certification, or ad hoc investigations in response to a whistleblower or documented failure. In 41 summary, both categories of standards exist--network and enforced--but enforced standards require additional institutions for successful implementation. In practice, standards address one of two objects: products or management processes. Product standards can define terminology, measurements, variants, functional requirements, qualitative properties, testing methods, and labeling criteria. Management process standards can describe processes or elements of organizations to 42 achieve explicit goals, e.g., quality, sustainability, and software life cycle management. A process that follows a particular standard need not impose costs with each iteration of a product: the standardized process simply informs how new products are created. Indeed, process standards can often function as a way for firms to adopt best practices in order to increase their competitiveness. One such ISO standard on cybersecurity has been 43 adopted by firms in nearly 160 countries. Figure  1  illustrates these standards categories as they relate to 44 externalities with some notable examples. Standards for AI will emerge in all four quadrants; indeed, as discussed below in Section 3 , standards that span the typology are already under development. Different types of standards will spread with more or less external effort, however. Network-product standards that support interoperability and network-process standards that offer best practices will see actors adopt them in efforts to grow the size of their market and reduce their costs. Indeed, most international standards from ISO/IEC and IEEE are product standards that address network externalities, seeking to increase the interoperability of global supply chains. Enforced standards will require 45 further incentivization from external stakeholders, whether they be regulators, contracting companies, or the public at large. The more distant the object of standardization is from common market transactions, the more difficult the incentivization of standards will be without external intervention. In particular, this means that an enforced-process standard for safety in basic research and development is unlikely to develop without concerted effort from the AI community.  40  Firms may declare that their practices or products conform to network standards, even, in some cases, choosing to certify this conformity. In these cases, however, the certification serves as a signal to access network benefits. Although enforced standards are not the only category that may see certification, it is the category that requires further enforcement to address possible incentives to defect. 41 Some AI standards, namely those on safety of advanced research, will benefit from novel monitoring regimes. See Section 4.  42  Hallström, Kristina Tamm. Organizing International Standardization : ISO and the IASC in Quest of Authority. Cheltenham: Edward Elgar, 2004.  43  Brunsson and Jacobsson. \"The pros and cons of standardization.\" 44 \"The ISO Survey of Management System Standard Certifications -2017 -Explanatory Note. ISO . Published August, 2018. https://isotc.iso.org/livelink/livelink/fetch/-<PHONE>/<PHONE>/<PHONE>/<PHONE>/00._Overall_results_and_explanator y_note_on_2017_Survey_results.pdf?nodeid=<PHONE>&vernum=-2 ; ISO 27001 had nearly 40,000 certifications in 159 countries in 2017. See ISO 27001 website .  45  Büthe and Mattli, The new global rulers. \n Figure 1. Standards typology with examples \n Network -Product Protocols for establishing Wi-Fi connections  (IEEE 802.11)  Standard dimensions for a shipping container to enable global interoperability (ISO 668) \n Network -Process Quality management process standard that facilitates international contracting and supply chains by ensuring consistency globally (ISO 9001) Information security management system, requirements and code of practice for implementation and maintenance.  \n Enforced -Process Environmental Management process standard helps organizations minimize the environmental footprint of their operations (ISO 14001) Enforced: Third-party certified. Functional safety management over life cycle for road vehicles (ISO 26262). \n Enforced: Required to meet safety regulations and import criteria. \n Safety requirements for collaborative industrial robots (ISO/TS 15006). Enforced: Supports obligations under safety regulations. There are, however, also notable examples of enforced standards that do see firms take on considerable costs to internalize harmful externalities. The ISO 14001 Environmental Management standard has spread to 171 countries, and saw over 360,000 certifications around the world in 2017. This standard provides firms a 48 framework to improve the environmental sustainability of their practices and certification demonstrates that they have done so in order to gain reputational benefits from environmental regulators. Firms take on 49 significant costs in certification, the total process for which can can cost upwards of $100,000 per facility. The 50 standard has been notable for spreading sustainable practices to middle-tier firms that do not differentiate  46  See the certification description on the Forest Stewardship Council website .  47  See European Commission website on CE marking .  48  \"The ISO Survey of Management System Standard Certifications -2017 -Explanatory Note. ISO . Published August, 2018. https://isotc.iso.org/livelink/livelink/fetch/-<PHONE>/<PHONE>/<PHONE>/<PHONE>/00._Overall_results_and_explanator y_note_on_2017_Survey_results.pdf?nodeid=<PHONE>&vernum=-2 ; ISO 14001 had over 360,000 certifications in 171 countries in 2017. See ISO 14000 website .  49  Prakash, Aseem, and Matthew Potoski. The voluntary environmentalists.  50  Ibid. themselves based on environmentally sustainable practices. Clearly, however, ISO 14001 has not solved larger 51 environmental challenges. The narrower success of this program should inform expectations of the role for standards for AI development: although they can encourage global adoption of best practices and see firms incur significant costs to undertake them, standards will not be a complete solution. Another category of enforced standards relevant to AI standards are product and process safety standards. Safety standards for medical equipment, biological lab processes, and safety in human-robot collaboration have been spread globally by international standards bodies and other international institutions. Related standards 52 for functional safety, i.e., processes to assess risks in operations and reduce them to tolerable thresholds, are widely used across industry, from autonomous vehicle development to regulatory requirements for nuclear reactor software. These standards do not apply directly to the process of cutting-edge research. That is not to 53 say, however, that with concerted effort new standards guided by these past examples could not do so. \n 2Cii. Shaping expert consensus The internal processes of international standards bodies share two characteristics that make them useful for navigating AI policy questions. First, these bodies privilege expertise. Standards themselves are seen as legitimate rules to be followed precisely because they reflect expert opinion. International standards bodies generally 54 require that any intervention to influence a standard must be based in technical reasoning.  55  This institutional emphasis on experts can see an individual researcher's engagement be quite impactful. Unlike other methods of global governance that may prioritize experts, e.g., UN Groups of Governmental Experts which yield mere advice, experts involved in standards organizations have influence over standards that can have de facto or even de jure governing influence globally. Other modes of de jure governance, e.g., national  51   regulation or legislation, present only limited direct opportunities for expert engagement. Some are concerned 56 that such public engagement may undermine policy efforts on specific topics like AI safety. Thus, for an AI 57 researcher looking to maximize her global regulatory impact, international standards bodies offer an efficient venue for engagement. Similarly, AI research organizations that wish to privilege expert governance may find 58 international standards bodies a venue that has greater reach and legitimacy than closed self-regulatory efforts. Second, standards bodies and their processes are designed to facilitate the arrival of consensus on what should and should not be within a standard. This consensus-achieving experience is useful when addressing questions 59 surrounding emerging technologies like AI that may face initial disagreement. Although achieving consensus 60 can take time, it is important to note that the definition of consensus in these organizations does not imply unanimity, and in practice it can often be achieved through small changes to facilitate compromise. This 61 institutional capacity to resolve expert disagreements based on technical argument stands in contrast to legislation or regulation that will impose an approach after accounting for limited expert testimony or filings. The capacity to resolve expert disagreement is important for AI, where it will help resolve otherwise controversial questions of what AI research is mature enough to include in standards. \n 2Ciii. Global reach and enforcement International trade rules, national policies, and corporate strategy disseminate international standards globally. These mechanisms encourage or even mandate adoption of what are nominally voluntary standards. This section briefly describes these mechanisms and the categories of standards to which they apply. Taken together, these mechanisms can lead to the global dissemination and enforcement of AI standards. International trade agreements are key levers for the global dissemination of standards. The World Trade Organization's Agreement on Technical Barriers to Trade (TBT) mandates that WTO member states use  56  Congressional or parliamentary testimony does not necessarily translate into law and legislative staffers rarely have narrow expertise. Such limitations inform calls for a specialized AI regulatory agency within the US context. A regulatory agency can privilege experts, but only insofar as they work within the agency and, for example, forgo research. Standards organizations, alternatively, allow experts to continue their research work. On limitations of expertise in domestic governance and a proposal for an AI agency, see Scherer, Matthew U. \"Regulating artificial intelligence systems: Risks, challenges, competencies, and strategies.\" Harv. JL & Tech. 29 (2015): 353-400.  57  See, e.g., Larks. \"2018 AI Alignment Literature Review and Charity Comparison.\" AI Alignment Forum blog post. 58 Supra footnote 55, expert participation remains political. Some standards are more politicized than others, though this does not follow a clear division between product or process, network or enforced. Vogel sees civil regulation (essentially enforced standards) as more politicized than technical (network) standards, although he looks at more politicized venues than simply international standards bodies. Network standards with large distributional consequences are often politicized, including shipping containers and ongoing 5G efforts. international standards where they exist, are effective, and are appropriate. This use can take two forms: 62 incorporation into enforced technical regulations or into voluntary standards at the national level. The TBT 63 applies to existing regulations regardless if they are new or old; thus, if a new international standard is established, pre-existing laws can be challenged. The TBT has a formal notice requirement for such regulation 64 and enables member states to launch disputes within the WTO.  65  There are important limitations to TBT, however. Few TBT-related disputes have been successfully resolved in the past. TBT applies only to product and product-related process standards, thus precluding its use in  66 67  spreading standards on fundamental AI research. In a further limitation, the agreement permits national regulations to deviate from international standards in cases where \"urgent problems of safety, health, environmental protection or national security arise,\" although such cases require immediate notification and justification to the WTO.  68  National policies are another key lever in disseminating international standards. National regulations reference international standards and can mandate compliance de jure in developed and developing countries alike.  69  Governments may use their purchasing power to encourage standards adoption via procurement requirements. EU member state procurement must draw on European or international standards where they exist, and the National action may threaten a fractured global governance landscape and fears of a race to the bottom in regulatory stringency, including that of standards. In a race to the bottom in regulatory stringency, AI development organizations may, in the future, choose to locate in jurisdictions that impose a lower regulatory burden; these organizations need not actually relocate, or threaten to do so, in order to impose downward pressure on regulatory oversight. National strategies already witnessed have proposed policy changes to 75 encourage AI development. Such national actions will undoubtedly continue and court leading AI development organizations. WTO institutions, if actively used for the purpose, may be able to moderate these concerns of a race to the bottom. Notably, moreover, the global and concentrated nature of markets for AI and related industries will see MNCs use standards internationally. Analogous to government procurement, MNCs may themselves demand contractors adhere to international standards. Such standards include network-product and network-process standards to meet customer demand. In addition to reduced costs from supply chain interoperability and 78 increased revenues from meeting customer demand, MNCs--and other firms alike--have further incentives to adopt international standards: standards can provide protection from liability in lawsuits and can lower insurance premiums.  79  Together these mechanisms can be used to encourage movement toward a unified global governance landscape for AI standards. National governments and MNCs can mandate use of standards, product and process, network and enforced alike. WTO rules require consistent use of international product standards globally. The incentives of MNCs encourage consistent use of international standards--both product and process--globally. If a large national market mandates adherence to a standard, MNCs may keep administration costs low by complying across the globe. If they do, then MNCs are incentivized to lobby other jurisdictions to pass similar laws, lest local competition be at an advantage. That means, given that many leading AI research efforts are 80 within MNCs, insofar as one country incorporates international AI standards into local law, others will face pressure to follow suit. This was witnessed, for example, with environmental regulation passed in the U.S., which subsequently led DuPont to lobby for a global agreement to ban ozone-depleting chemicals in order to see its international competition similarly regulated. This phenomenon is currently witnessing the 81 globalization of data protection regulations at the behest of the GDPR. The analysis of global governance mechanisms in this section should not be portrayed as arguing that using these tools to spread and enforce AI standards globally will be easy. But the tools do exist, and concerted efforts to make use of them are a worthy endeavour. In sum, the scope for standards in the global governance of AI research and development is not predetermined. Recalling our standards typology, the object of standardization and incentives therein will determine particular needs for standards development and complementary institutions. Standards for AI product specification and development processes have numerous precedents, while standards to govern fundamental research approaches are without precedent. More generally, if international experts engage in the standardization process, this serves to legitimize the resulting standard. If states and MNCs undertake efforts to adopt and spread the standard, it will similarly grow in influence. Active institutional entrepreneurship can influence the development of and scope for international standards in AI.  78    \n Current Landscape for AI Standards \n 3A. International developments Given the mechanisms outlined in the previous section, international standards bodies are a promising forum of engagement for AI researchers. To date, there are two such bodies working on AI: ISO/IEC JTC 1 Standards Committee on Artificial Intelligence (SC 42) and the working groups of IEEE SA's AI standards series. Figure  2  categorizes the standards under development within the externality-object typology, as of January 2019.  \n Figure 2. International AI standards under development \n 3B. National priorities There are three important developments to note for national policies on standards for AI. First, key national actors, including the U.S. and China, agree that international standards in AI are a priority. Second, national strategies for AI also indicate that countries plan to pursue national standards. Third, given the market structure in AI industry, countries are incentivized to ensure that international standards align as closely to national standards as possible. First, international standards are a stated priority for key governments. The recently released U.S. Executive Order on Maintaining American Leadership in Artificial Intelligence identified U.S. leadership on international technical standards as a priority and directed the National Institute for Standards and Technology to draft a plan to identify standards bodies for the government to engage. The Chinese government has taken a similar that \"China should strengthen international cooperation and promote the formulation of a set of universal regulatory principles and standards to ensure the safety of artificial intelligence technology.\" This 90 recommendation was corroborated by previous CESI policies, e.g., its 2017 Memorandum of Understanding with the IEEE Standards Association to promote international standardization.  91  Second, national standards remain relevant. Historically, observers have argued that Chinese national standards in fields auxiliary to AI, including cloud computing, industrial software, and big data, differ from international standards in order to support domestic industry. These differences have not been challenged under WTO for instance, identified 10 key areas for standardization: software engineering, performance, metrics, safety, usability, interoperability, security, privacy, traceability, and domain-specific standards.  94  Other countries are also considering national standards. An overview of AI national and regional strategies describes plans for standards from Australia, the Nordic-Baltic Region (Denmark, Estonia, Finland, the Faroe Islands, Iceland, Latvia, Lithuania, Norway, Sweden, and the Åland Islands), and Singapore. The Chief 95 Scientist of Australia has proposed an AI product and process voluntary certification scheme to support consumer trust. Insofar as these national strategies seek to develop AI national champions and given the 96 network effects inherent in AI industry, AI nationalism is of transnational ambition. \n 97 This leads to the third important point: national efforts will likely turn to international standards bodies in order to secure global market share for their national champions. Successful elevation of national standards to the international level benefits national firms that have already built compliant systems. Successful inclusion of corporate patents into international standards can mean lucrative windfalls for both the firm and its home country.  98  If one state seeks to influence international standards, all others have incentive to do similarly, else their nascent national industries may lose out. Given that both the U.S. and China have declared intent to engage in international standardization, this wide international engagement will likely come to pass. One illustrative case of the consequences of failure to follow competitors in international standardization is offered by the U.S. machine tools industry. This industry once described by Ronald Reagan as a \"vital component of the U.S. defense base,\" did not seek to influence global standards on related products, and has declined precipitously under international competition. This stands in contrast to the standards engagement and continued strength of the sector in Germany and Italy. Furthermore, the WTO rules outlined in Section 2.C.iii , if enforced, \n Case of 5G: International standards with implications for national champions Although telecom standards and standardization bodies differ from those leading in AI standardization, the ongoing development of 5G standards is an illustrative case to consider regarding states' interests. Previous generations of mobile telephony standards did not see a single, uncontested global standard, with Europe and the US on inoperable 3G standards and the LTE standard facing competition before solidifying its global market dominance in 4G. The global economies of scale resulting from 4G standard consolidation may see a uniform standard adopted globally for 5G from the start. This globally integrated market will offer 100 positive-sum outcomes to cooperation, albeit with some countries winning more than others. These incentives for network-product standards may very well be larger than those present in AI. These incentives are driving participation in efforts at the focal standardization body, 3GPP, which set LTE for 4G as well as some past generation standards, to set the radio standard. At stake in the standardization 101 process is the economic bounty from patents incorporated into the standard and their resulting effects on national industry competitiveness in the global market. One estimate claims that U.S. firm Qualcomm owns approximately 15 percent of 5G patents, with Chinese companies, led by Huawei, controlling about 10 percent. One example of Huawei's success in 5G standards was the adoption of its supported polar coding 102 method for use in control channel communication between end devices and network devices.  103  In contrast to a positive-sum game with distributional consequences common in international standards, the use of national standards reverts to a protectionist zero-sum game. In the past, there has been criticism of China's efforts to use national standards towards protectionism with requirements that differ from international standards. In 5G and AI standards, however, China has sought to engage in international standards bodies, thereby mitigating this past concern and responding to past international pressure to reduce trade barriers. The Trump administration opposes China's international standards activities, in keeping longer does the U.S. view international participation as a way to reduce trade barriers; rather, it sees international participation as a way to shift influence in China's favor globally. Despite these politics, global standards will improve market efficiency and lead to better outcomes for all. Some will be better off than others, however. The distributional consequences of 5G standards may be larger than those for AI in the short-term, but this case nonetheless has implications for international efforts towards AI standards. The future may see similarly politicized standardization processes for AI. China's formulated policies for international engagement on AI standards will likely see other countries engage in order to encourage a more balanced result. This engagement means that AI researchers' efforts to influence standards will be supported but also that they likely will be increasingly politicized. Yet, to be clear, AI standards are not currently as visible or politicized as telecom standards, which have already seen four previous iterations of standards and the emergence of large globally integrated markets dependent upon them. \n 3C. Private initiatives In addition to international and national standards, there are a number of private initiatives that seek to serve a standardizing role for AI. Standards, most commonly network-product standards, can arise through market forces. Notable examples include the QWERTY keyboard, dominance of Microsoft Windows, VHS, Blu-Ray, and many programming languages. Such market-driven product standards can produce suboptimal outcomes where proprietary standards are promoted for private gain or standards may fail to spread at all due to a lack of early adopters.  106  In AI, software packages and development environments, e.g., TensorFlow, PyTorch and OpenAI Gym, are privately created, are used widely, and perform a standardizing role. Market forces can also encourage, though not develop in their own right, network-process and enforced standards through customer demands on MNCs and MNC pressure on their supply chains, as explained in Section 2.C.iii . For example, the CleverHans Adversarial Examples Library, if incorporated into an adversarial example process check that became widely 107 adopted in practice, would be such a standard. Another example is Microsoft's Datasheets for Datasets standard practice to report on data characteristics and potential bias that is used across the company. Researchers will 108 continue to develop software packages and benchmarks. This approach does not necessarily require an additional commitment of time beyond their research work, whereas engaging on traditional standards development does require some time commitment. Some of these packages and benchmarks may spread to the extent that they become industry standards. But these standards will face difficulties in securing global dissemination and enforcement. Yet, as discussed in Section 4 below private standards can be turned into international standards with a concerted effort. Still other groups are developing standards in the broad sense of a guide for judging behavior. The 2017 Asilomar Conference on Beneficial AI yielded a set of AI Principles that address areas of research, ethics and values, and long-term issues, which have been signed by some 1300 AI and robotics researchers as well as 2500 others. Among these principles was a commitment to safety standards in particular: \"Teams developing AI 109 systems should actively cooperate to avoid corner-cutting on safety standards.\" The Association for  110 Computing Machinery (ACM), a professional association, maintains a Code of Ethics and Professional Conduct for its members. This Code includes many principles, including \"Avoid harm.\" The ACM has also 111 called for a new peer review standard that requires researchers to acknowledge \"negative implications\" of their research. The Partnership on AI, a multistakeholder forum founded by leading AI firms, seeks to develop best 112 practices for AI research and development. These standards, broadly defined, do not benefit directly from the 113 dissemination and enforcement mechanisms outlined in Section 2Ciii . However, such standards may have normative power in influencing actors who subsequently engage in standardization activities that produce standards which are subject to mechanisms of dissemination and enforcement. \n Recommendations Today, AI standards development is already underway at both ISO/IEC and IEEE. National strategies, including those of the U.S. and China, prioritize engagement in standardization processes for AI. Thus, the agenda is set. Engagement in these processes today can benefit from these ongoing processes and national foci. As time goes on, however, standards bodies may become increasingly politicized just as multiple iterations of telecom standards have, over time, given rise to highly politicized international tension over 5G. This section offers recommendations to use standards to help support AI policy goals starting today. \n 4A. Engage in ongoing processes How can the AI community, namely researchers and research organizations, engage effectively? There are four elements necessary for successful influence in international standards bodies: • technical expertise, • financial resources, • timely information, and • effective institutional knowledge.  114  The AI research community already has technical expertise and financial resources, but not up-to-date information on proceedings within standards bodies nor the institutional knowledge required to successfully intervene. The following four recommendations helps fill in these gaps. 4Ai. Build capacity for effective engagement AI researchers are unlikely to have experience engaging national and international standards bodies. Of leading AI organizations, only Google and Microsoft participate in the U.S. standards committee that is affiliated with ISO/IEC JTC 1 SC 42; none participates in the U.K. equivalent. Similarly, IEEE P7000 series working groups see very few volunteers from leading organizations.  115  In order to successfully influence standardization outcomes, researchers should develop expertise in these standardization processes. In some cases, researchers need not go far to find this expertise. Large firms may already have teams working on creating and complying with international standards, though they may focus more on products as opposed to AI research and development. Research institutions and firms can learn more about ongoing standardization processes by participating in the Open Community for Ethics in Autonomous and Intelligent Systems (OCEANIS). OCEANIS is a 116 coordinating forum for standards organizations and other interested organizations to discuss efforts to use standards to further development of autonomous and intelligent systems. It was was co-founded by the IEEE and IEC, among other national and regional standards bodies. OCEANIS does not produce standards itself, but could be a useful venue for organizations seeking to build capacity prior to engaging directly in standardization processes. Beyond expertise, perspective matters: it is important to view standards as a policy tool for encouraging positive AI outcomes.  \n 4Aii. Engage directly in ongoing processes There are two, related paths to engage with ISO/IEC JTC 1 SC 42. First, researchers should consider joining the group that mirrors SC 42 within their respective national standards body. It is through these national bodies that researchers can influence and directly engage in SC 42. In the United States, this group is InterNational Committee for Information Technology Standards (INCITS) -Artificial Intelligence. Committee membership is open to any U.S. national who is materially affected by related standardization activities. , The equivalent 120 121 committee for the UK is British Standards Institution ART/1.  122  The second method of engaging SC 42 is to seek appointment to its expert working groups that drafts standards directly. Such appointments are made by national member organizations, so the first engagement strategy will further the second. The work of SC 42 is in its early stages. Working Group 3 on Trustworthiness, and specifically its ongoing work on a technical report on robustness in neural networks, is likely the highest value area of engagement at this time. At this preliminary stage, however, participation in a national standards body or SC 42 working group can serve to build career capital and institutional knowledge that will be useful in creating further working groups in the future. These efforts could focus on standards related to AI policy goals; some of these possible standards will be discussed below. IEEE SA P7000 series working groups are open for interested individuals to join. Indeed, the process of joining is much simpler than that of ISO/IEC JTC 1-related work. One simply needs to contact a working group and express interest. In order to participate in developing the Ethics Certification Program for Autonomous and Intelligent Systems (ECPAIS), individuals must be affiliated with organizations possessing an IEEE SA Advanced Corporate Membership. The work of standards within the IEEE SA P7000 series is at varied stages of completeness. Standards earlier in the sequence have approximately a year left in development and standards later in the sequence have more time. This means that interested researchers should consider engaging soon if they are to have an impact in ongoing working groups. Two standards in particular could support AI policy goals outlined above: P7001 Transparency of Autonomous Systems, which seeks to define measures of transparency and P7009 Fail-Safe Design of Autonomous and Semi-Autonomous Systems. Engagement on these standards could help ensure that their respective scopes support the governance of long-term risks.  \n 4B. Pursue parallel standards development An alternative to direct engagement is the parallel development of standards. This could take many forms in practice. Individual organizations, existing working groups at the Partnership on AI, or other ad hoc consortia could develop, i.a., software libraries, measurement benchmarks, or best practices procedures. Once developed, these approaches could then be transferred into international standards to achieve global dissemination and enforcement. Indeed, there are numerous examples of organizations and even individual firms transferring existing standards into international ISO standards. The C Programming language was developed at Bell Laboratories before being adopted as an international standard by ISO. More recently, Microsoft transferred its Open XML format to 124 ISO, as did Adobe with PDF. Microsoft's effort is illustrative of the potential for one motivated MNC to  125 126  influence standardization processes: it placed its experts on several national committees that then influenced discussions at the ISO committee. Smaller firms can also have success: Microsoft's Open XML effort \n 4C. Research standards and strategy for development This paper and its nascent AI standards strategy serves as a call to AI researchers to engage in order to help develop standards as a global governance mechanism for AI. Further research from the AI research community is needed to ensure that standards under development today can encourage positive outcomes for advanced AI. This work could benefit from researchers from across the AI research field as well as forecasting experts. The lines of work are two-fold. First, technical standards desiderata should be developed. Second, specific strategies to see these standards codified and spread globally should then be created. \n 4Ci. Research technical standards desiderata Ultimately, AI researchers should seek to consolidate AI standards desiderata for their particular area of focus. Some of this work may take place at existing working groups hosted by the Partnership on AI, discussions within individual organizations, or through other ad hoc gatherings. This paper offers two prototype standards that would support AI policy goals: an AI safety process standard and an AI systems capability standard. The field of AI Safety is young, but preliminary conversations about how to incorporate safety procedures into a standard framework that can reduce risks globally would be a welcome application of existing research. The first step in this process is the distillation of current best practices. However tentative and incomplete, these practices are an improvement over a disregard for safety--if expectations are calibrated correctly. There are numerous labs around the world today with advanced AI ambitions and no focus on safety. Prototype 130 standards could spread a focus on safety and current best practices globally. One such approach could be a 131 process standard that requires researchers to complete a checklist procedure before undertaking research, namely record a precise specification, measures taken to ensure robustness, and methods of assurance. This 132 approach could then serve as a model for future standards and regulation as system capabilities increase. A more developed version could see researchers certify to a series of best practices. Such a certification framework could eventually be linked to a monitoring regime for defined high risk projects. This certification approach would likely see a series of related standards, which is a common practice. One standard would be definitional: defining high risk projects or developing a risk typology of multiple categories, as is used in functional safety standards. Another standard would then identify best practices and mitigation strategies to be followed at each risk threshold. Additional standards could specify monitoring and certification regimes. When realized, this example may see labs obtain third-party certification subject to verification, e.g., via real-time monitoring of researchers' use of large amounts of computing hardware. Although such enforcement regimes may be novel, international standards bodies do have experience with safety standards for emerging technologies, as described in Section 2Bi . Another standard for consideration would permit consistent assessment of system capabilities globally. This type of standard could inform above safety standards by assessing the relative danger of a system or it could facilitate international agreements on AI development and deployment in a variety of domains. Insofar as it was incorporated into an international standard, these practices could be spread globally and possibly facilitate future international agreements. ISO has supported similar efforts to combat climate change in furtherance of the Paris Climate Agreement: it has a series of greenhouse gas emissions measurement standards for individual organizations and auditors. In contrast to the organic spread of private benchmarks, international standards 133 can support universal adoption at a point in the future where it may be needed. Performance benchmarks already exist for particular tasks. The AI Index incorporates these benchmarks and others to report on AI performance and other metrics annually. Notable benchmarks include the ImageNet 134 corpus, which has served as an image recognition benchmark for research globally and helped drive the rise of deep learning research. The General Language Understanding Evaluation (GLUE) may be a similarly  135 136  impactful benchmark in the field of natural language understanding. GLUE integrates nine distinct natural language understanding tasks into one benchmark. As systems become more capable, the integration of tasks into holistic benchmarks will continue. Further work is needed to contextualize these growing modular benchmarks in a broader capabilities framework. An 137 integrative approach could benefit from ongoing efforts to map types of intelligence. Such an approach could 138 then serve as the basis for forecasting efforts and international agreements that see universal adoption. Of course, such a measurement standard cannot come ahead of fundamental research. \n 4Cii. Research strategies for standards in global governance Although AI safety research continues and these procedures outlined above are not foolproof, thought on how to implement safety processes at scale needs parallel development with technical safety research itself. Understanding such efforts as enforced-process standards, institutions for both agreement and enforcement are needed. Although ISO/IEC JTC 1 SC 42 may one day offer a promising home for such efforts, initial proof-of-concept work may be done more effectively elsewhere. Ongoing work on monitoring, incentives, and cooperation, at institutions like OpenAI, FHI Center for the Governance of AI, and the Cambridge Centre for Study of Existential Risk, may prove useful in this effort. For each identified important area of standardization ongoing, as well as the new efforts identified above, a roadmap for global governance should be developed. This roadmap can then be used by institutional entrepreneurs, whether they be individual researchers, organizations, firms, or larger groups. Each particular standards roadmap could begin by answering the following questions: Which firms, organizations, or states may adopt the standard first? Which policy mechanisms will be most useful in spreading these standard more broadly? How should the broader AI research community support these efforts? As international and other standards bodies initiate standardization efforts in more areas of AI, an important question to address will be to what extent each needs attention from the AI research community. This is a calculus of interests and impact. If a topic of standardization is relevant to policy goals for the governance of advanced AI and actors' incentives may overlook standards' development to these ends, engagement will be warranted. In other cases, e.g., standards for autonomous vehicles, actors' incentives are aligned so as to not necessitate engagement. Each roadmap should similarly address this question of differential impact. Section 2Ciii compiled a series of institutional mechanisms for dissemination and enforcement that warrant further research to analyze their relative performance as well as the influence of global and domestic politics in their processes. This understanding can then inform strategies to spread standards for AI governance. expectations. For example, in adopting a transparency standard, an organization commits to the importance of transparency for AI systems. Second, standards establish and reinforce a relational system that see individual researchers and AI development organizations embedded in a larger network. In adopting an international standard, an organization voluntarily acknowledges that outside actors have a stake in the procedures undertaken within the organization. Third, in order to follow the adopted standards, researchers will necessarily carry out practices, repeatedly performing, and internalizing as routine, a culture of responsibility and safety. Fourth, standards will often be embedded directly within products and software packages; individuals' interactions with these artefacts reinforce a culture of safety and responsibility. For example, consider a safety checklist is embedded into a software package that prompts a researcher to address safety risks and mitigation strategies before she can train a model. Regardless of who uses the system, that interaction will reinforce safety. Understood in this way, standards can be yet another tool for institutional entrepreneurs who promote a culture of responsibility and safety in AI development. Within companies, closer connections between product teams with experience in standards and AI research teams can spread this culture. The adoption of AI standards under development as well as possible future standards can further serve to support this connection within and among AI labs. Outside of a particular company, standards can drive the adoption of best practices more widely across the industry. They can also be bundled with other advocacy efforts that reward responsible labs with better public opinion and access to talented researchers. A culture change is not easy, but standards can help in this path. \n Conclusion This paper has sought to reframe international standards as tools of AI policy. Some AI policy challenges, e.g., trust among developers and safe ground rules in international competition, warrant global solutions. International standards bodies produce expertise-and consensus-based policies that can provide these solutions. A series of mechanisms can then spread and enforce these policies across the globe. International standards bodies are currently developing AI standards and states have prioritized engagement. The agenda is set, but further expert engagement is needed. This paper has made the case for this engagement, provided an overview of ongoing standardization efforts, and offered detailed recommendations for those who wish to get involved. 19 measurement, or other initial agreement contributing to subsequent expanded and enforced agreements has been witnessed in other international coordination problems, e.g., nuclear test ban treaties and environmental protection efforts. Trust is also dependent on the degree of open communication among labs. Complete 20 openness can present problems; indeed, open publication of advanced systems, and even simply open 21 reporting of current capabilities in the future could present significant risks. Standards can facilitate partial 22 \n 70 optional 70 WTO Agreement on Government Procurement encourages parties to use international standards for procurement where they exist. The US Department of Defence (DoD), for instance, uses multiple 71 international product and process standards in its software procurement, and this appears set to continue 72 based on the 2018 U.S. Department of Defense Artificial Intelligence Strategy. Beyond regulatory obligations 73 and procurement requirements, governments spread standards through adoption in their own operations.74     \n 88 position in an AI Standardization White Paper published by the China Electronics Standardization Institute (CESI) within the Ministry of Industry and Information Technology in 2018. The white paper recommended 89 \n 106 Mattli, Walter. \"Public and Private Governance in Setting International Standards.\" In Kahler, Miles, and David A. Lake, eds. Governance in a global economy: Political authority in transition . Princeton University Press, 2003. 107 Papernot, Nicolas, Fartash Faghri, Nicholas Carlini, Ian Goodfellow, Reuben Feinman, Alexey Kurakin, Cihang Xie et al. \"Technical report on the cleverhans v2. 1.0 adversarial examples library.\" arXiv preprint arXiv:1610.00768 (2016). 108 Gebru, Timnit, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumeé III, and Kate Crawford. \"Datasheets for Datasets.\" arXiv:1803.09010 (2018). \n 139 More broadly, additional research on the politics of standard-setting and standard enforcement is needed. Existing literature focuses primarily on the politics of standard-setting. This work does not focus on standards 140 for digital technologies, however. Furthermore, little work has been done to understand the role of individual firms in setting international standards. Similarly little research has been done on the ways in which standards 141 spread globally in practice. \n ISO/IEC JTC 1 SC 42 Ongoing Work Appendix 2: IEEE AI Standards Ongoing Work Glossary Acronym Meaning DoD Table of Contents Glossary 1. Introduction 2. Standards: Institution for Global Governance 2A. The need for global governance of AI development 2B. International standards bodies relevant to AI 2C. Advantages of international standards as global governance tools 2Ci. Standards Govern Technical Systems and Social Impact 2Cii. Shaping expert consensus 2Ciii. Global reach and enforcement 3. Current Landscape for AI Standards 3A. International developments 3B. National priorities 3C. Private initiatives 4. Recommendations 4A. Engage in ongoing processes 4Ai. Build capacity for effective engagement 4Aii. Engage directly in ongoing processes 4Aiii. Multinational organizations should become liaisons 4B. Pursue parallel standards development 4C. Research standards and strategy for development 4Ci. Research technical standards desiderata 4Cii. Research strategies for standards in global governance 4D. Use standards as a tool for culture change 5. Conclusion References Appendix 1: \n These claims are developed further in Sections 2C and 4 . In summary, continued AI development presents risks that require coordinated global governance responses. International standards are an existing form of global governance that can offer solutions. These standards can help support efficient development of AI industry, foster trust among states and developers of the technology, and see beneficial systems and practices enacted globally. It is important to note that, regardless of intervention, ongoing standards work will encourage increased efficiency in AI development. Engagement is needed to support standards that help foster trust and encourage beneficial systems and processes globally. 27 25 Armstrong, S., Bostrom, N. & Shulman, C. \"Racing to the precipice: a model of artificial intelligence development\", Technical Report #2013-1, Future of Humanity Institute, Oxford University: pp. 1-8. (2013). https://www.fhi.ox.ac.uk/wp-content/uploads/Racing-to-the-precipice-a-model-of-artificial-intelligence-development.pd f . 26 Software libraries, programming languages, and operating systems are standards insofar as they guide behavior. They may not emerge from standardization processes but instead market competition. See Section 3C . 27 Hale, Thomas, and David Held. Beyond Gridlock . Cambridge: Polity. 2017. 2B. International standards bodies relevant to AIA wide range of organizations develop standards that are adopted around the world. AI researchers may be most familiar with proprietary or open-source software standards developed by corporate sponsors, industry consortia, and individual contributors. These are common in digital technologies, including the development of AI, e.g., software libraries including TensorFlow, PyTorch, and OpenAI Gym that become standards across industry over time. The groups responsible for such standards, however, do not have experience in monitoring 26 and enforcement of such standards globally. In contrast, international standards bodies have such experience. This section discusses standards bodies, and Section 2C describes relevant categories of standards.Specialized bodies may create international standards. These bodies can be treaty organizations such as the International Atomic Energy Agency and the International Civil Aviation Organization, which govern standards on nuclear safety and international air travel, respectively. Such a body may well suit the governance of AI research and development, but its design and implementation are beyond the scope of this paper. Instead, this paper focuses on existing institutions that can host development of needed standards and see them enacted globally. Non-state actors' efforts towards institutional development tend to be more successful in both agenda setting and impact if they work with states and seek change that can be accommodated within existing structures and organizations. Thus, existing international standards bodies present an advantage. Nevertheless, \n (ISO/IEC 27001 and 27002, respectively)    Software life cycle management processes (ISO/IEC/IEEE 12207) Enforced -Product Paper products sourced with sustainable methods and monitored through supply chain (Forest Stewardship Council) 46 Enforced: Third-party certified. CE marking for safety, health, and environmental protection requirements for sale within the European Economic Area. (EU) Enforced: If problems arise, violations are sanctioned by national regulators. 47 \n Particular industry applications derive from the generic framework standard for functional safety, IEC 61508, including ISO 26262 Road vehicles --Functional safety and IEC 61513 Nuclear power plants -Instrumentation and control for systems important to safety -General requirements for systems. Per conversation with an employee at a firm developing autonomous driving technology, all teams in the firm have safety strategies that cite the standard. Nuclear regulators reference the relevant standard, see, e.g., IAEA.Implementing Digital Instrumentation and Control Systems in the Modernization of Nuclear Power Plants. Vienna: IAEA. 2009. https://www-pub.iaea.org/MTCD/Publications/PDF/Pub1383_web.pdf . See generally, Smith, David J., and Kenneth GL Simpson. Routledge, 2009; Jacobsson, Bengt. \"Standardization and expert knowledge\" in A World of Standards . Oxford: Oxford University Press, 2000, 40-50. 55 This is not to say that standards are apolitical. Arguments made with technical reasoning do not realize a single, objective standard; rather, technical reasoning can manifest in multiple forms of a particular standard, each with distributional consequences. See Büthe and Mattli, The new global rulers. Safety critical systems handbook: a straight forward guide to functional safety, IEC 61508 (2010 Edition) and related standards, including process IEC 61511 and machinery IEC 62061 and ISO 13849. Elsevier, 2010. 54 Murphy, Craig N., and JoAnne Yates. The International Organization for Standardization (ISO): global governance through voluntary consensus . London: Ibid.  52  There is no data available on uptake of ISO/TS 15006 Robots and robotic devices --Collaborative robots or related standards. ISO does claim, however, that IEC 60601 and ISO 10993 have seen global recognition and uptake for ensuring safety in medical equipment and biological processes, respectively. See discussion of ISO standards in sectorial examples at ISO's dedicated webpage .53   \n Vogel, David. \"The Private Regulation of Global Corporate Conduct.\" in Mattli, Walter., and Ngaire. Woods, eds. The Politics of Global Regulation. Princeton: Princeton University Press, 2009; Büthe and Mattli, The new global rulers. 59 See, e.g., Büthe and Mattli, The new global rulers , Chapter 6. 60 Questions of safe procedures for advanced AI research, for instance, have not yet seen debate oriented towards consensus. 61  Büthe and Mattli, The new global rulers, pp. 130-1. \n find no such mandates for IEEE standards. Procurement requirements are common for both IEEE and ISO/IEC standards, and market mechanisms similarly encourage both. ISO has had global success with many enforced standards, whereas IEEE has no equivalent experience to date. States have greater influence in Vogel, David. \"The Private Regulation of Global Corporate Conduct.\" in Mattli, Walter., and Ngaire. Woods, eds. 83 ISO/IEC standards development than that of IEEE, and state involvement has enhanced the effectiveness of past standards with enforcement mechanisms. Thus, given that enforcement of ISO/IEC standards has more 84 mechanisms for global reach, participation in ISO/IEC JTC 1 may be more impactful than in IEEE. Ongoing SC 42 efforts are, so far, few in number and preliminary in nature. (See Appendix 1 for a full list of SC 42 activities.) The most pertinent standards working group within SC 42 today is on Trustworthiness. The Trustworthiness working group is currently drafting three technical reports on robustness of neural networks, bias in AI systems, and an overview of trustworthiness in AI. Network -Product Network -Process • Foundational Standards: Concepts and terminology • Model Process for Addressing Ethical (SC 42 WD 22989), Framework for Artificial Intelligence Concerns During System Design Systems Using Machine Learning (SC 42 WD 23053) (IEEE P7000) • Transparency of Autonomous Systems (defining levels of • Data Privacy Process (IEEE P7002) transparency for measurement) (IEEE P7001) • Methodologies to address algorithmic • Personalized AI agent specification bias in the development of AI systems (IEEE P7006) (IEEE P7003). • Ontologies at different levels of abstraction for ethical • Process of Identifying and Rating the design (IEEE P7007) Trustworthiness of News Sources • Wellbeing metrics for ethical AI (IEEE P7011) (IEEE P7010) • Machine Readable Personal Privacy Terms (IEEE P7012) • Benchmarking Accuracy of Facial Recognition systems (IEEE P7013) Enforced -Product Enforced -Process 86 • Certification for products and services in transparency, certifications could be subject to the failings of negative externality standards that lack enforcement • Certification framework for accountability, and algorithmic bias in systems (IEEE ECPAIS) mechanisms. 87 child/student data governance (IEEE P7004) • Fail-safe design for AI systems (IEEE P7009) • Certification framework for employer data governance procedures based on 82 \"IEEE Position Statement: IEEE Adherence to the World Trade Organization Principles for International Standardization.\" GDPR (IEEE P7005) • Ethically Driven AI Nudging 83 IEEE acknowledges that their AI standards are \"unique\" among their past standards: \"Whereas more traditional standards have a focus on technology interoperability, safety and trade facilitation, the IEEE P7000 series addresses specific methodologies (IEEE P7008) issues at the intersection of technological and ethical considerations.\" IEEE announcement webpage . The SC 42 is likely the more impactful venue for long-term engagement. This is primarily because IEEE standards Politics of Global Regulation. Princeton: Princeton University Press, 2009. have fewer levers for adoption than their ISO equivalents. WTO TBT rules can apply to both IEEE and ISO/IEC product standards, but their application to IEEE was only asserted in 2017 and has never been tested. could IEEE's AI standards are further along than those of SC 42. (See Appendix 2 for a full list of IEEE SA P7000 series activities, as of January 2019.) Work on the series began in 2016 as part of the IEEE's larger Global Initiative on Ethics of Autonomous and Intelligent Systems. IEEE's AI standards series is broad in scope, and continues to broaden with recent additions including a project addressing algorithmic rating of fake news. Of note to AI researchers interested in long-term development should be P7009 Fail-Safe Design of Autonomous and Semi-Autonomous Systems. The standard, under development as of January 2019, includes \"clear procedures for measuring, testing, and certifying a system's ability to fail safely.\" Such a standard, depending 85 on its final scope, could influence both research and development of AI across many areas of focus. Also of note is P7001 Transparency of Autonomous Systems, which seeks to define measures of transparency. Standardized methods and measurements of system transparency could inform monitoring measures in future agreements on advanced AI development.IEEE SA recently launched the development of an Ethics Certification Program for Autonomous andIntelligent Systems (ECPAIS). Unlike the other IEEE AI standards, development is open to paid member organizations, not interested individuals. ECPAIS seeks to develop three separate processes for certifications related to transparency, accountability, and algorithmic bias. ECPAIS is in an early stage, and it remains to be seen to what extent the certifications will be externally verified. Absent an enforcement mechanism, such84 85 \"P7009 Project Authorization Request.\" IEEE-SA . Published July 15, 2017. https://development.standards.ieee.org/get-file/P7009.pdf?t=<PHONE> . 86 IEEE announcement webpage . 87 Calo, Ryan. Twitter Thread. October 23, 2018, 1:39PM. https://twitter.com/johnchavens/status/1054848219618926592 . \n Technologies are not apolitical and neither are the processes that shape them. With this 117 118 understanding, standards are not simply a response to a particular market need, but, more broadly, a tool of global governance. Strategic engagement in standardization now can help direct wider consideration to important areas like AI safety. ISO and IEEE have formalized standards maintenance procedures so that standards can be updated as the state of the art progresses. The important step today is to understand and See, e.g., Lessig, Lawrence. Code: And Other Laws of Cyberspace, Version 2.0 . New York: Basic Books, 2006. 118 Büthe and Mattli, The new global rulers. 119 For details on ISO's systematic review process, see \"Guidance on the Systematic Review Process in ISO.\" ISO . Published May 2017. https://www.iso.org/files/live/sites/isoorg/files/store/en/Guidance_systematic_review.pdf ; For details on IEEE's maintenance process, see \"Next Steps Kit: Guidelines for Publication, Recognition Awards and Maintenance.\" IEEE Standards Association . N.D. https://standards.ieee.org/content/dam/ieee-standards/standards/web/documents/other/next_steps_kit.pdf . 119 start using this tool for global governance. 115 This was indicated by observations of one working group and an interview with the chair of another working group. 116 \"Participation.\" OCEANIS website. 117   \n 120  InterNational Committee for Information Technology Standards(INCITS). \"New INCITS Technical Committee on Artificial Intelligence -Notice of January 30-31, 2018 Organizational Meeting and Call for Members.\" Email, 2018. https://standards.incits.org/apps/group_public/download.php/94314/eb-2017-00698-Meeting-Notice-New-INCITS-T C-on-Artificial-Intelligence-January30-31-2018.pdf .121  Membership fees vary by organizational affiliation, from several hundred to several thousand dollars per year.122  See BSI committee information webpage .4Aiii. Multinational organizations should become liaisonsAnother method of engagement is available to multinational industry or other membership associations like the Partnership on AI. These groups are eligible for liaison status with ISO/IEC JTC 1 SC 42 at both the Standards Committee level and Working Group level. Although liaisons cannot vote on final standards, they have significant influence. Participation at the Standards Committee level would allow such an organization to propose new standards, comment on draft standards, and nominate experts for Working Groups. 123   \n\t\t\t This work fits within a growing literature that argues that short-term and long-term AI policy should not be considered separately. Policy decisions today can have long-term implications. See, e.g., Cave and ÓhÉigeartaigh, \"Bridging near-and long-term concerns about AI.\" 5 Some in the AI research community do acknowledge the significance of standards, but they see efforts towards standardization as a future endeavor: the OpenAI Charter acknowledges the importance of sharing standards research, but focused on a time when they curtail open publication. The Partnership on AI is today committed to establishing best practices on AI, in contrast to formal standards.6  Advanced AI incorporates future developments in machine intelligence substantially more capable than today's systems but at a level well short of an Artificial General Intelligence. See Dafoe, \"AI Governance: A Research Agenda.\" \n\t\t\t The TBT does not define international standards bodies, but it does set out a Code of Good Practice for standards bodies to follow and issue notifications of adherence to ISO. IEEE declared adherence to the Principles in 2017. See ISO's \n\t\t\t rules. However, these same observers do note that China is increasingly active in international standards activities. In January 2018, China established a national AI standardization group, which will be active with \n\t\t\t require that national regulations cite international standards. This means that failure to secure international standards that reflect preexisting national ones could require changes in national regulation to encourage global competition. Thus, such developments could cost national industry both internationally and domestically. This means that countries will likely engage in international standards bodies that govern priority industries like AI. \n\t\t\t with its zero-sum perspective on international trade. For example, the Committee on Foreign Investment in the United States (CFIUS) decision to block the foreign acquisition of Qualcomm came out of concern that it \"would leave an opening for China to expand its influence on the 5G standard-setting process.\" No 105100  Brake, Doug. \"Economic Competitiveness and National Security Dynamics in the Race for 5G between the United States and China.\" \n\t\t\t \"Asilomar AI Principles.\" Future of Life Institute website . \n\t\t\t followed another ISO-approved open standard that was submitted by a consortium of smaller companies. 128  IEEE has also created standards and then seen them adopted by ISO/IEC JTC 1 in the past. Similar efforts could be made in the case of specific AI standards, whether from IEEE's P7000 series or from another organization. Indeed, if an organization like the Partnership on AI were to create AI standards, it could apply for status as a Publicly Available Specifications (PAS) Submitter with the ISO/IEC JTC 1. With this status, a standards organization can submit specifications for a vote among national bodies; over one hundred standards have been approved in this process.129    \n\t\t\t Baum, \"A Survey of Artificial General Intelligence Projects for Ethics, Risk, and Policy.\" 131 E.g., Avoiding negative side effects, avoiding reward hacking, scalable oversight, safe exploration, and robustness to distributional change. \n\t\t\t For up-to-date information, see the SC 42 blog, here .", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Standards_-FHI-Technical-Report.tei.xml", "id": "0645c3d4d6f96aebc6701c57651334c1"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "This paper aims to include all cases where there is substantial discussion in the literature of how national security influenced antitrust enforcement. 9 Though they are not the sole focus of this study, I am especially interested in cases where national security concerns and economic considerations conflict, suggesting different outcomes. 10 Such cases are interesting because, since the late 1970s, the dominant view-associated with the Chicago school of economics-has been that antitrust ought to primarily promote well-functioning, efficient markets by preventing firms from extracting monopoly rents. 11 Thus, cases in which national security considerations conflict with economic (i.e., efficiency-based) ones create tension between the primary goal of antitrust and other governmental objectives. Analyzing these conflicts reveals interesting insight into governmental prioritization. I find that in cases where national security and economic considerations conflict, economics has been given increased consideration over time. Cases in which the United States government (USG) actively uses (or threatens to use) antitrust enforcement to advance unrelated national security goals may be seen as a particularly worrisome historical precedent. 12  The ability to threaten antitrust enforcement to advance unrelated goals implies that the antitrust-relevant corporate conduct would have otherwise been tolerated. This further suggests that such enforcement would be contrary to the course of action recommended by economic analysis of that conduct. 13 If such uses of antitrust are tolerated, companies may worry that they will become targets of antitrust due to circumstances outside their control, 14 or that they will be pressured to abandon stated values like pacifism. 15 This could 9 See infra § I.A. 10 National security and economic considerations do not always conflict; sometimes, they both suggest the same outcome. 11 See, e.g., Nat'l Soc. of Prof'l Engineers v. United States, 435 U.S. 679, 688 (1978)  (\"Contrary to its name, the Rule [of Reason] does not open the field of antitrust inquiry to any argument in favor of a challenged restraint that may fall within the realm of reason. Instead, it focuses directly on the challenged restraint's impact on competitive conditions.\"); Organisation for Economic Co-operation and Development [OECD], Note by the United States: Public Interest Considerations in Merger Control 2 (2016), https://perma.cc/XB26-CRTP (\"U.S. antitrust law and policy, including merger review, are implemented based on the belief, borne out by our economic history, that the public interest is best served by focusing exclusively on competition considerations.\"); ROBERT BORK, THE", "authors": ["Cullen O’Keefe"], "title": "How Will National Security Considerations Affect Antitrust Decisions in AI? An Examination of Historical Precedents", "text": "Introduction: The Confluence of AI, National Security, and Antitrust Artificial Intelligence (AI)-like past general purpose technologies such as railways, the internet, and electricity-is likely to have significant effects on both national security 1 and market structure.  2  These market structure effects, as well as AI firms' efforts to cooperate on AI safety and trustworthiness,  3  may implicate antitrust in the coming decades. Meanwhile, as AI becomes increasingly seen as important to national security, such considerations may come to affect antitrust enforcement.  4  By examining historical precedents, this paper sheds light on the possible interactions between traditional-that is, economic-antitrust considerations and national security in the United States. Several useful analyses of antitrust enforcement in national security and foreign policy 5 contexts already exist.  6  This paper compiles past known American antitrust cases that have implicated national security and systematically analyzes them on a number of dimensions of interest.  7  By studying these cases, I am able to draw novel lessons on how national security considerations have influenced antitrust enforcement 8 and adjudication in past cases. raise various legal concerns. Normative preferences for underpunishment instead of overpunishment are common in law.  16  Overpunishment due to arguably irrelevant policy considerations 17 seems particularly objectionable. Yet, I am only able to find one case wherein the USG used antitrust to advance unrelated national security objectives: Bananas in Guatemala  (1953) .  18, 19  The rarity of overenforcement due to unrelated national security objectives suggests that future security-driven overenforcement would violate historical enforcement norms. Part I lists and analyzes the cases I identified, then lists several conclusions I was able to draw from my case studies. Those conclusions are: 1. National security considerations have entered the antitrust enforcement process numerous times over the past 100 years. 2. It is rare for the USG to actively use antitrust enforcement to advance unrelated national security objectives. 3. In cases where national security and economic considerations conflict, economics has been given more weight over time. 4. The president plays an important role in reconciling conflicting considerations. Part II discusses how these conclusions might apply to AI firms in the coming decades. 2020) (\"[W]e will not design or deploy AI in the following application areas: 1. Technologies that cause or are likely to cause overall harm. Where there is a material risk of harm, we will proceed only where we believe that the benefits substantially outweigh the risks, and will incorporate appropriate safety constraints. 2. Weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people. 3. Technologies that gather or use information for surveillance violating internationally accepted norms.  4 . Technologies whose purpose contravenes widely accepted principles of international law and human rights.\").  16  See, e.g., United States v. Santos, 553 U.S. 507, 514 (2008) (plurality opinion) (\"The rule of lenity requires ambiguous criminal laws to be interpreted in favor of the defendants subjected to them.\"); 4 WILLIAM BLACKSTONE, COMMENTARIES *352 (\"  [B] etter that ten guilty persons escape, than that one innocent suffer\"); Frank B. Cross, Institutions and Enforcement of the Bill of Rights, 85 CORNELL L.  REV. 1529 REV. , 1592 REV. (2000 ; Richard H. Fallon, Jr., The Core of an Uneasy Case for Judicial Review, 121 HARV. L. REV.  1693, 1695 (2008) .  17  See supra note 11 (sources arguing that only economic considerations should drive antitrust enforcement).  18  Infra § I.B.6.  19  Furthermore, even this case does not fit this characterization neatly. In the Bananas in Guatemala (1953) case, the USG began prosecuting the United Fruit Company (\"UFCO\") for purely economic reasons. See infra note 176 and accompanying text. However, the USG suspended its prosecution for foreign policy reasons. See infra notes 177-179 and accompanying text. But once the facts changed such that prosecuting UFCO was in the USG's foreign policy interests, the USG resumed prosecution. See infra notes 180-189 and accompanying text. Thus, although resuming prosecution of UFCO served unrelated national security goals (as did the suspension of prosecution), the initiation of the prosecution appears to have been purely economic. \n I. Case Studies This Section lists twelve antitrust cases in which national security considerations influenced the case's outcome. It also analyzes those cases across several dimensions of interest. \n I.A. Methodology I found cases to include in this study by searching secondary literature and, to a lesser extent, primary sources (namely case law) for existing discussions on antitrust in national security and foreign policy contexts. I attempted to include all cases where I could find substantial discussion of how national security influenced antitrust enforcement. 20 However, I am not an historian by training, so it is possible that I failed to find some relevant cases. Additionally, it seems likely that, due to the sensitivity of the subject matter, there is no easily discoverable documentary evidence of how national security influenced antitrust cases. While I hope that I have been objective and reasonably thorough in my search, it is possible I missed some cases. I would be excited to see this analysis improved by discovery of more data. Finally, the small number of cases studied here means that the conclusions I draw have large degrees of uncertainty. To structure my analysis, I identified several interesting dimensions across which cases could vary. Those dimensions are: 1. Directness: Whether the national security and antitrust concerns relate to the same conduct. A case is \"direct\" when the conduct of national security interest itself triggered antitrust scrutiny. A case is \"indirect,\" therefore, when the government used antitrust enforcement to influence something other than the antitrust-triggering conduct. 21 2. Activeness. The USG was \"active\" in a case if they sued or threatened to sue. The USG was \"passive\" if they refrained from suing or dropped an existing suit. 3. Presidential Involvement simply indicates whether there is any evidence the president was involved in the direction of the case. Of course, presidential intervention may not have left an accessible documentary record. Nevertheless, in the absence of such evidence, I assume that there was no such intervention.  \n Facts During World War I, German agents plotted to \"disrupt the flow of war materiel from U.S. factories to countries fighting against Germany\" by disrupting war industry labor (in which they actually succeeded) and planning to bomb factories and transportation facilities.  25  The USG successfully prosecuted the plotters under Section 1 of the Sherman Act 26 under the theory that the labor disruptions 27 and planned bombings 28 \"restrained\" foreign trade.  29  This case seems to retain little relevance. The USG has better tools with which to prosecute industrial saboteurs today.  30  Still, this case provides an early and successful 31 strategic use of antitrust law as a punitive measure against enemy agents. \n Dimensional Analysis 1. Directness: Direct, since the disruptive acts were also the acts that allegedly restrained trade. 2. Activeness: Active, since the cases were actually prosecuted.  32  3. Presidential Involvement: No. I did not find any evidence of this. 4. Conflicting Considerations: No. Presumably, the economic interests in this case were not at the forefront, but neither were they likely in conflict with the national security interests. \n Conflict Winner: N/A. 23 Some cases span many years. I generally date a case to the point when significant antitrust enforcement began.   27  See Lamar, 260 F. at 262 (\"The end or object of the proven conspiracy was not to call strikes, but to restrain or rather suppress foreign trade. That object is as illegal as ever . . . .\"); Rintelen, 233 F. at 787 (\"The purpose of the indictment in the present case is not to charge the commission of an offense against the United States . . . but to charge a conspiracy in violation of the first section of the Sherman Act, namely, a conspiracy in restraint of trade and commerce with foreign nations; the very gist of the offense being the conspiracy. . . . The conspiracy does not go to the restraint in trade of any particular individual or corporation, or combination of men, but to the restraint of all foreign trade where munitions of war are the subject of commerce.\").  28  See Bopp, 237 F. at 285 (\"In so far as these activities were to be directed against the manufacturing plants, in this country, and railroads and ships generally engaged in the transportation of such munitions of war as were manufactured therein, it is claimed that a conspiracy existed in violation of the Sherman Act . . . .\").  29  See generally Fisher, National Emergencies I, supra note 6, at 988-89.  30  Cf. Steuer & Barile, supra note 6, at 71 (these cases took place \"[i]n an era when federal prosecutors had fewer law enforcement tools available to them than today . . . .\").  31  Multiple defendants received statutory maximum prison terms. See Fisher, National Emergencies I, supra note 6, at 988-89.  32  See supra note 31. \n I.B.2. Military Optical Instruments (1940) 33 \n Facts The USG alleged that two corporations-American Bausch & Lomb and German Carl Zeissentered into illegal agreements that, among other things, prevented Bausch & Lomb from selling outside the United States without the prior consent of the German Carl Zeiss corporation, and prevented the latter from selling in the United States without the like consent of Bausch & Lomb. Inventions resulting from the joint labor of the two companies were to be patented in the United States by Bausch & Lomb for the account of Carl Zeiss.  34  The USG alleged that the cartel was adversely affecting war production.  35  The case was settled through a consent decree.  36  [The consent decree] expressly provided, however, that the rights of Bausch & Lomb under any patents for the manufacture of military optical instruments were not affected by the provisions of the decree. This would mean that sole ownership of those patents obtained by  [Bausch & Lomb]  for the account of Carl Zeiss was thereby vested in the American company.  37  The retention of patent rights to military technologies by the American firm presumably conferred strategic benefits to the United States government.  38  However, the nature and extent of these benefits is unclear. More generally, then head of the Department of Justice Antitrust Division Thurman Arnold said that cases like this had a number of harmful consequences: Throttling American capacity to produce essential war materials by foreign ownership and control of patents; Cartelization of certain industries with price and production control in foreign hands; Transmission to foreign companies of American military secrets; Division of markets, fixing and restricting of price of materials essential to military preparation; Collusive bidding on contracts for the Army and Navy.   259 (1963) .  35  See Fisher, National Emergencies II, supra note 6, at 1188.  36  See id.  37  Id. at 1189 (emphasis added).  38  See Spencer Weber Waller, The Antitrust Legacy of Thurman Arnold, 78 ST. JOHN'S L. REV. 569, 603 (2004); cf. Steuer & Barile, supra note 6, at 72 (\"[T]he government launched some powerful antitrust measures against international cartels that held the keys to strategic products.\").  39  Waller, supra note 38, at 603-04 (quoting THURMAN W. ARNOLD, THE BOTTLENECKS OF BUSINESS 74 (1940)); see also Fisher, National Emergencies II, supra note 6, at 1186-87. The Bausch & Lomb case was \"the earliest\" 40 of a number of antitrust consent decrees for the war production effort.  41  The Rubber  Manufacturing (1942)   The Department of Justice's wartime prosecution of the American Standard Oil Co. for disclosing rubber manufacturing processes to the German chemical company I.G. Farbenindustrie (\"I.G.\") 46 provides a stark example of the USG using antitrust in response to a firm that aided enemies. In the 1920s, Standard and I.G. had entered into several agreements to illegally 47 divide the oil, coal, and chemical markets. 48 \"In the words of a Standard official, '[I.G. was] going to stay out of the oil business-and [Standard was] going to stay out of the chemical business.'\"  49  To administer these deals, they incorporated (in America) and jointly owned the Standard-I.G. Company and Joint American Standard Company (\"Jasco\").  50  In 1937, pursuant to its agreements with I.G., Standard gave I.G. the rights and technical knowledge for the manufacture of a synthetic rubber called Butyl.  51  At the time, I.G. was a \"mighty industrial colossus\" 52 that controlled key strategic resources and technical knowledge in  40  Fisher, National Emergencies II, supra note 6, at 1188.  41  See id. at 1187-91.  42     50  See id. at 51.  51  See id. at 79-80.  52  Id. at 1. a variety of fields 53 necessary to the Nazi war effort.  54  The Nazis' dependence upon I.G. was extensive; 55 a postwar USG report concluded that \"'[w]ithout I.G. . . . , Germany would not have been in a position to start its aggressive war in September 1939.'\"  56  Rubber was a \"key[] to German rearmament,\"  57  and Butyl was \"cheaper and better\" than the process previously used there, 58 so the transfer was a strategic boon to the Nazis.  59  At the same time, Standard \"refused to reveal to the United States Navy and the British Government its process for making synthetic rubber . . . .\"  60  Though they had promised to \"stay out of the chemical business,\" Standard officials repeatedly pleaded for technical knowledge of I.G.'s own synthetic rubber, Buna.  61  The Nazi government, acutely aware of rubber's strategic value and of the potential for conflict with America, refused to allow I.G. to give Standard instructions for Buna production.  62  However, I.G. did transfer patents to the Standard-I.G. Company (\"S.I.G.\") and Jasco to avoid their seizure by the USG's Alien Property Custodian.  63  I.G. included the Buna patent in this transfer, but not (the more important) technical knowledge of how to produce it.  64  The result was that I.G. (and by extension the Nazi government) retained control over the knowledge necessary for much of the synthetic rubber production in America.  65  After the Pearl Harbor attack, Japan blocked the U.S. from its primary natural rubber supply in Southeast Asia, precipitating a \"monumental rubber crisis.\"  66  The U.S. Department of Justice Antitrust Division, under the leadership of Assistant Attorney General Thurman Arnold 67 and \"with the backing of several powerful senators and administration officials,\"  68  began antitrust enforcement against Standard with the goal of \"open[ing] up the development and manufacture of synthetic rubber.\" 69 This enforcement came as a compromise, though-President Roosevelt agreed to suspend other pending antitrust investigations so industry could focus on war production.  70  The DOJ complaint alleged that the Standard-I.G. cartel contributed to the ongoing rubber crisis.  71  Despite initial protests that the suit would disrupt their war production activities,  72  Standard and the DOJ successfully negotiated a consent decree shortly after. 73 Key provisions of the consent decree included: • The severance of all relations between American cartel members (including Standard, S.I.G., and Jasco) 74 and I.G.; 75 • Transfer of cartel-owned patents-including in Butyl and Buna-to Jasco and S.I.G., along with the \"know-how\" for the use thereof; 76 and • A requirement that, for the duration of the war, Jasco and S.I.G. would issue royalty-free licenses for those patents upon request by anyone. 77 \"Freed from cartel restrictions, the [American] wartime synthetic rubber program overcame all obstacles and expanded production of rubber substitutes at a rate which was sufficient eventually to meet emergency needs.\" 78 It is worth noting that, despite being accused of \"treason\" by incensed politicians, 79 Standard did not intend to aid the Nazi war efforts.  80  Its motive was simply cartel profits.  81  Of course, this fact did not impede USG antitrust enforcement against a firm that plainly and significantly aided enemy war efforts. \n Dimensional Analysis 1. Directness: Direct. The market divisions to which Standard and I.G. agreed were the cause of both the strategic problems for the USG and the antitrust offense. 2. Activeness: Active. The USG reached a consent decree with Standard.  82  3. Presidential Involvement: Yes.  83  4. Conflicting Considerations: No. Similar to the Bausch & Lomb case, 84 the national security considerations may have been stronger, but there was still a clear economic harm present. 5. Conflict Winner: N/A.  70  See id.  71  See id. at 91.  72    \n I.B.4. Union Pacific Railroad (1943) Facts DOJ Antitrust Division leader Thurman Arnold attempted to prosecute railroads for price-fixing in 1943. However, this conflicted with national security priorities, which saw the railroads as important to war production:  85  The final straw [for Arnold's tenure at DOJ Antitrust] appeared to be Arnold's attempt to criminally prosecute the railroads for price fixing and to indict Averell Harriman, the chairman of the Union Pacific, who was appointed as United States Ambassador to the Soviet Union in the same year that Arnold would have indicted him. The indictment was quashed in the name of national defense . . . .  86  Understanding that he lacked support from the Roosevelt administration for vigorous antitrust enforcement, Arnold resigned from DOJ.  87  As the above excerpt suggests, Arnold's wartime antitrust efforts frequently conflicted with war priorities, which generally included protecting industries involved in defense production.  88  When this happened, war priorities almost always won out: Arnold was losing the antitrust battle to defense preparation and the war effort on a daily basis. The problem was that while the Standard Oils, DuPonts, GEs, and Alcoas were guilty of heinous conduct, their sins were ultimately greedy in nature rather than traitorous. These companies were absolutely vital to the war effort and many of their executives were now working in the war planning and production effort. Arnold was forced to agree publicly-if not entirely voluntarily-to defer to the War and Navy Departments in the event they explicitly found that any particular antitrust violation was necessary for national defense. Perhaps it was inevitable that this would overwhelm his antitrust enforcement program given the scope of the national emergency and the corporatist culture of the war planners themselves. Case after case was vetoed by the planning and defense authorities, including cases involving conduct predating the war. Arnold spent more and more of his time fighting with the war planners, including Hugh Johnson, who was the first head of the NRA [National Recovery Administration] and still had little use for the antitrust laws. For the first  85  See Waller, supra note 38, at 606.  86  Id. at 606-07 (citations omitted).  87  See Alan Brinkley, The Antimonopoly Ideal and the Liberal State: The Case of Thurman Arnold, 80 J. AM.  HIST. 557, 578 (1993) .  88  See I.F. Stone, Thurman Arnold and the Railroads, 156  NATION 331, 331 (1943)  (\"Today the [War Production Board], the War and Navy departments, the Office of the Petroleum Administrator for War, and the [Board of Economic Warfare] are run by business men and lawyers who have devoted much of their activity to violating the anti-trust laws. Little wonder that they proved powerful enough, first, to force weak consent decrees on Arnold, then to shut off one scheduled prosecution after another, and finally to promote Arnold to the bench.\"). time, Congress cut rather than increased Arnold's budget and staffing.  89  From another source: Arnold had twenty-one anti-trust cases shot out from under him by the arrangement last spring whereby the corporation lawyers and bankers who man the offices of the War and Navy departments are permitted to suspend the anti-trust laws for the convenience of themselves, their ex-clients, and their friends. Among these cases are no fewer than five against the du Ponts and two against General Electric. All but one of them are criminal, not civil cases. In all but two instances the prosecution was stopped after indictments were obtained. The indictments act as a considerable deterrent. Recently the Secretaries of War and the Navy have begun to stop investigations before indictments can be returned. One such case was the inquiry into the Hawaiian pineapple industry. The other involved a group of three indictments prepared against the Illinois Freight Association and the Central States Motor Freight Bureau. These were to be the beginning of the first major attack on the greatest monopoly in this country, the growing monopoly in transportation-on the methods whereby the railroads fix not only their own rates but impose uneconomic and non-competitive rates on the movement of goods by air, water, and highway. A letter to Attorney General [Francis] Biddle by Secretary of War [Henry L.] Stimson and Acting Secretary of the Navy Forrestal throws new light on the curious procedure now being followed in anti-trust cases. It reveals that drafts of the proposed indictments were presented by the Anti-Trust Division to the Secretaries of War and the Navy before submission to the grand jury in Chicago, and that this was done \"in pursuance of an arrangement made at a conference held in your [Biddle's] office.\" The War and Navy departments, and Eastman, prevented the proposed indictments from ever reaching the grand jury. War, Navy, and Eastman were willing only to permit the prosecution of labor leaders and others who were alleged to have used a strike to coerce a motor carrier to increase its rates. Arnold and Biddle refused to do this.  90  A longtime Antitrust Division lawyer recalls: [A]t the beginning of World War II, antitrust enforcement was suspended by Executive Order. Numerous complaints were being made to the military about the Antitrust Division's activities under Thurman Arnold 89 Waller, supra note 38, at 606 (citations omitted).  90  Stone, supra note 88, at 331-32 (last alteration in original). who was Assistant Attorney General. The War Department and the Navy Department were bringing to the attention of President Roosevelt's office complaints that investigations, prosecutions, and other activities in connection with antitrust enforcement were allegedly interfering with the manufacture of munitions and other war materials by the large companies that were involved in these antitrust matters. . . . I think there were something like twenty-five cases that were postponed because FDR was approached and told about these complaints by the War Department, the Navy and the individual companies and he agreed to call a moratorium on any antitrust matter whether it was a case, an investigation or whatever, until the end of the war. So, from March 20, 1942, when FDR issued an Executive Order, until June 30, 1945, there was a moratorium on all, or most, antitrust matters. The statute of limitations was also suspended for any of these antitrust investigations or cases so that it didn't run during the war. And at the end of the war, which was four or five years before I came into the Antitrust Division, there were a lot of cases brought that involved cartels and price fixing and all kinds of stuff involving U.S. companies abroad.  91  It is not clear why Arnold decided to prosecute Union Pacific despite the moratorium, though his decision to do so proved fatal to his career with the Antitrust Division. \n Dimensional Analysis 1. Directness: Indirect. The wartime considerations militating against prosecution were unrelated to the underlying price-fixing.   92  See supra notes 85-86 and accompanying text.  93  See supra note 86 and accompanying text.  94  See supra notes 86, 91 and accompanying text.  95  See supra note 86 and accompanying text.  96  See generally BURTON I. KAUFMAN, THE OIL CARTEL CASE (1978).  97  See id. at 20. producers 98 regarding \"market allocation, price fixing, elimination of outside competition, mutual assistance in curtailing expenses, avoidance of duplicate facilities, and reciprocal sales and exchanges of crude and petroleum products.\"  99  Initially, these agreements garnered little antitrust enforcement: 100 \"[b]ecause of the growing dependence of the industrialized world on oil, the petroleum industry became the beneficiary of preferential treatment, receiving immunities and grants that other industries simply did not enjoy.\"  101  The necessities of the World Wars also contributed to this preferential treatment.  102  However, a combination of post-WWII agreements that further consolidated the American oil producers began to concern the Department of Justice.  103  The cartel had a number of adverse effects on strategic operations, which contributed to the decision to pursue the case: • \"[T]he high prices Aramco was charging for petroleum delivered to the Navy as opposed to what it was charging a number of countries such as France and Uruguay.\" The oil companies were successful in negotiating the inclusion of \"a paragraph [in the exception agreement] that stated the agreement would not be construed as to require the participants to take any action in conflict with prior obligations.\"  124  The agreement remained in effect until 1952, when it was no longer necessary to meet Europe's energy needs.  125  Meanwhile, the Iran crisis continued. The Department of Justice continued to disagree with the foreign policy wing of the cabinet on whether to resolve the crisis through a consortium of the major oil corporations (the strategy favored by the Department of State) or through a loan to Iran in exchange for them lifting restrictions on oil exports through BP (the strategy favored by the Department of Justice).  126  Truman ultimately decided to follow State's recommendation, which necessitated demoting the case from a criminal to a civil action. 127 In return, [t]he only conditions that Truman imposed on the oil companies . . . were that they turn over to the Justice Department the documents required by the subpoena and that they pledge not to make any motion to amend or dismiss the new suit until the documents were received and the details of the government's case made known.  128  The formation of the Iranian consortium proceeded under the Eisenhower administration.  129  The oil majors approached the consortium apprehensively, as they already had access to adequate supply elsewhere. 130 \"They made it clear to the administration that they would enter a consortium only in the interest of national security . . . [and] also insisted on being granted antitrust immunity in the production of Iranian oil.\" 131 \"Even though a question existed about the authority of the Justice Department to give antitrust immunity for national security purposes outside the provisions of the Defense Production Act, the Justice Department gave its sanction [for the consortium, including immunity].\" 132 Recognizing the risk of the appearance of selfcontradiction, the Antitrust Division dropped allegations relating to oil production and refining and instead focused on distribution, marketing, pricing, and transportation.  133  This was another major blow to antitrust goals in the name of national security: \"the result of this administration policy was . . . to maintain the hold of the multinationals over Mideast oil without any compensating factors in the form of . . . significant changes in government policy towards multinationals.\"  134  Changes to DPA required the attorney general to review existing voluntary agreements regarding nonmilitary goods.  135  In January 1956, the Acting Attorney General declared his intention to withdraw Justice's approval from the voluntary agreement arising out of the Iranian crisis, saying that its \"adverse impacts . . . outweighed its contribution to national defense.\" 136 However, under pressure from the Office of Defense Mobilization, the Acting Attorney General agreed to concessions that left the agreement in place. 137 \"Furthermore, the government and the oil companies secretly agreed that the proposed antitrust safeguards would apply only to the  126  See id. at 44-46.  127  See id. at 46.  128  Id. at 46-47 (emphasis added).  129  See id. at 50-60.  130  See id. at 57. However, some statements suggest a continued interest in accessing Iranian supply and preventing oil dumping by Soviet producers. See id.  131  Id. (emphasis added).  132  Id. at 58.  133  See id. at 59-60.  134  See id. (emphasis added).  135  See id. at 66.  136  See id.  137  See id. gathering of information and statistics and would be of no effect in case of an emergency necessitating the use of plans of cooperative action by the oil companies.\" 138 The administration granted antitrust immunity to oil majors yet again during the Suez Crisis of 1956-57, 139 though this immunity was of questionable legality.  140  This too was motivated by national security concerns-namely, fears that without such immunity, the oil majors would be unable to ship a sufficient amount of oil to American allies in Europe-and represented another setback for antitrust enforcement. 141 The Eisenhower administration also began to see the U.S.'s reliance on foreign oil as a threat to national security.  142  It therefore sought voluntary import quotas from the oil majors, 143 raising obvious antitrust issues.  144  This too frustrated Justice's antitrust enforcement goals.  145  From the beginning of negotiations, Justice sought a consent decree with the majors.  146  Its objectives were: 147 • \"An injunction against price fixing and division of the markets\"; 148 • The elimination of \"joint-marketing companies and other joint-marketing arrangements\"; 149 • \"  [R] equiring each of the [majors] to compete on the basis of its own resources rather than that of its competitors and . . . making excess oil available to all companies at the same price\"; 150 • \"[A]n injunction against sales of crude or products among defendants at [below-market prices]\"; 151 • \"[A]n injunction against any defendant [purchasing] any more or less than its proportionate share of any joint-production arrangement\"; 152 and • \"An injunction against the use by any defendant of a delivered price\" 153 (A delivered price is \"a system whereby oil prices were based on the cost of fuel at one or more places regardless of the actual cost of production.\") 154 The insistence of the oil companies that the case harmed U.S. foreign policy interests apparently \"weighed heavily with the presiding judge.\"  155  In response to a motion requesting the court to order defendants to produce documents from foreign subsidiaries, the judge \"asked the defendants and the Justice Department to seek the views of the Department of State on the likely effects of such an order.\" 156 \"The State Department's reply was noncommittal, yet favorable to the oil companies.\" 157 Seeing this, the presiding judge rejected Justice's requests.  158  However, after the State Department greenlighted the production of the documents, the judge ordered them produced.  159  Justice entered into consent agreements with Exxon, Gulf, and Texaco.  160  National security concerns in favor of the oil companies were the driving force behind the settlements and the way they were disclosed to the public.  161  Justice was unable to obtain consent decrees with Mobil and Socal but was unwilling to take the case to trial.  162  It dropped the suit against them in 1968.  163  Note how, in this case, antitrust goals were subservient to external assessments of the national interest. Even when Justice interprets antitrust enforcement to be in its interest, indications to the contrary from the national security apparatus (e.g., State) were sufficiently powerful to significantly delay and alter antitrust enforcement priorities. Indeed, the case also shows that corporations sufficiently important to U.S. foreign policy goals, as the big oil corporations were, 164 are likely to receive favorable treatment, including antitrust immunity, insofar as antitrust enforcement against them is detrimental to U.S. foreign policy objectives. Perhaps most shockingly, the oil corporations were unabashed about asserting that they, not Justice, were serving national interests, and the State Department largely agreed. However, the case also shows that Justice is also willing to shape its antitrust enforcement to further (its independent assessment of) national security needs. \n Dimensional Analysis 1. Directness: Both. Certainly, the artificially high prices were harmful from both national security and economic perspectives. However, some national security considerations that played a role in this case were not relevant to competitive conditions, such as the nationalization of BP and the resulting concern about whether the Soviet government could use the antitrust prosecutions as propaganda. 2. Activeness: Both. Interestingly, in this case the USG's strategy towards the majors changed depending on the geopolitical circumstances. Initially, the USG was passive by granting immunities to the companies.  165  However, the government switched to an active approach when President Truman decided that prosecuting the oil majors was beneficial on net.  166  During the Iranian nationalization crisis, President Truman took the passive steps of granting an antitrust exception under the DPA 167 and demoting the case from criminal to civil.  168  When he inherited the case, President Eisenhower also took the passive step of removing allegations relating to oil production from the case.  169  It is worth noting that the active uses of antitrust against the oil majors were all direct. That is, the USG only actively used antitrust in this case when the targeted conduct was both security and economically relevant. 3. Presidential Involvement: Yes. President Truman was personally and extensively involved in weighing the various considerations for and against antitrust enforcement. 170 4. Conflicting Considerations: Yes. \n Conflict Winner: National Security (generally). Towards the beginning of the case, President Truman decided to prosecute the oil majors notwithstanding the objections from his national security advisors. 171 However, even this decision was at least partially motivated by the national security considerations that others raised in favor of the prosecution. 172 Throughout the rest of the case, however, antitrust enforcement efforts were gradually weakened by national security considerations. \n I.B.6. Bananas in Guatemala (1953) \n Facts The United Fruit Company (\"UFCO\") was a leading American producer of fruits, with significant banana plantations in Guatemala.  173  \"By the mid-1950s, United Fruit had captured nearly 65% of the U.S. banana market.\"  174  Then, between 1952 and 1954, Guatemala's \"leftwing, nationalist government\" under Jacobo Árbenz seized and redistributed 400,000 acres of UFCO land.  \n Facts Yellowcake is a uranium product. After uranium is mined, it is \"milled\": \"crushed and treated with chemicals . . . to yield a mixture of compounds including uranium oxides concentrate, known for its color when pure and calcined as 'yellowcake.'\" 200 After being converted to uranium hexafluoride, the uranium is \"enriched\" to 3% uranium-235 for use in nuclear power reactors.  201  By comparison, weapons-grade uranium needs enrichment to >90% U-235. 202 \"[E]nriched uranium is the ingredient most difficult to obtain in fabricating nuclear weapons,\" 203 so America has focused nuclear nonproliferation efforts on controlling enrichment facilities,  204  rather than the supply of unenriched uranium (which is quite common).  205  In 1972, a booming market for uranium ended due to decreased demand for new weapons-grade uranium.  206   At the time, Canada, Australia, France, and South Africa essentially controlled all non-U.S. yellowcake sales.  209  Responding to the same market forces, the dominant producers from these countries and their respective governments formed the \"Uranium Club,\" a cartel that would fix worldwide uranium prices. 210 In 1978, the DOJ learned about the cartel following leaks from an Australian environmental group.  211  They criminally indicted Gulf Oil for violating antitrust laws by joining the cartel.  212  For unexplained reasons, the DOJ chose not to charge any of the foreign firms in the Club.  213  Furthermore, Gulf's alleged crime took place before the 1975 enactment of stiffer penalties, so the charge was only a misdemeanor.  214  The authors of Yellowcake speculate: The reason for the lenient treatment of Gulf and the absence of charges against foreign participants may have been pressure Gulf pleaded no contest and received a mere $40,000 in fines.  216  Note that, contrary to my initial presumption, the government did not treat the cartel as a bottleneck on a strategic weapons commodity, but rather as a potential constraint on American energy supply. 217 Nevertheless, the above excerpt suggests that foreign policy considerations played a role in the disposition of the suit. \n Dimensional Analysis 1. Directness: Direct. The cartel was the object of the antitrust complaint and also sensitive as a matter of foreign policy due to foreign governments' interest in it. 218 2. Activeness: Passive. As noted above, the fines against Gulf were small 219 and no foreign firms were charged. 220 3. Presidential Involvement: No (speculative). there is no evidence of this, though statements by the president may have influenced DOJ's decision-making.    \n Facts This case began in the Ford with the antitrust breakup of the Bell telephone system.  225  Presidents Carter and then Reagan inherited the case.  226  Some in the Reagan administration considered the Bell system crucial to national security. Testifying before the Senate Armed Services Committee about communications command and control, Secretary of Defense Caspar Weinberger asserted that \"[t]he [AT&T] network is the most important communication net we have to service our strategic systems in this country.\"  227  He also reported his disagreement with the antitrust case: Because of the discussions I have had concerning the effect of the Department of Justice suit that would break up part of that network, I have written to the Attorney General and urged very strongly that the suit be dismissed, recognizing all of the problems that might cause and because of the fact it seems to me essential that we keep together this one communications network we now have, and have to rely on.  228  Antitrust Assistant Attorney General William Baxter, who was overseeing the case, was unphased by Weinberger's objections and pressed on.  229  Perhaps surprisingly, given his reputation as a hawk,  230  \"Reagan [resolved the conflict] in favor of Justice and of pursuing the antitrust case.\" 231 against those carriers. 245 One particular source of concern was the investigation's effect on British Airways, which the UK was seeking to privatize. 246 In addition to their diplomatic protests, the British government retaliated by refusing to approve American carriers' proposed new discount transatlantic fares.  247  In 1984, in an attempt to alleviate further tensions over air travel between the two nations, President Reagan the Department of Justice halt the federal investigation into Laker.  248  The Justice Department acknowledged the propriety of the president adjusting antitrust enforcement to accommodate foreign policy. 249 \n Dimensional Analysis 1. Directness: Both. The USG's foreign policy concerns were based on the UK government's responses to the case, which were both direct and indirect. Directly, the UK government protested the extraterritorial application of antitrust law to its stateowned airlines.  250  The UK government also retaliated indirectly by refusing approval of American carriers' proposed fares. 251 Thus, the U.S. government's interest in easing tensions with the UK ultimately derived from both direct concerns about the implications of antitrust enforcement against state-owned airlines and indirect concerns about the UK's retaliatory measures. 2. Activeness: Passive. The DOJ halted the investigation.  252  3. Presidential Involvement: Yes.  253  4. Conflicting Considerations: Yes. DOJ's enforcement was antithetical to Reagan's foreign policy objectives.  \n Facts The number of mergers and acquisitions (M&A) in all industries has risen dramatically since approximately 1990, with peaks in 1999, 2007, and 2015.  256  Defense and aerospace 257 M&A has grown similarly.  258  American defense outlays declined starting in 1985 and dropped even further due to the end of the Cold War, which caused defense firms to have excess capacity.  259  This led to an unprecedented spike in defense industry mergers in the late 1980s and 1990s,  260  which precipitated a wave of responses in legal scholarship,  261  industry,  262  and government.  263  Until 1994, the Department of Defense (DoD) did not intervene in the merger process much.  264  Though individual contemporary and former DoD officials testified on both sides of merger cases, the DoD never took a formal position on them.  265  In 1994, the DoD's Defense Science Board released a report on the wave of mergers and made recommendations on the proper role of the DoD in merger review.  266  During this period, some had argued that DoD's role as a monopsonist with an interest in keeping prices low rendered antitrust oversight redundant.  267  Though regulators and courts  268  have consistently rejected such arguments, they do take account of DoD's expressed views on proposed mergers.  269  This is not unusual; courts and regulators often note customers' views during merger review under the theory that they have both good (but imperfect) information and incentives to object when appropriate.  270  Furthermore, the Defense Science Board found that [r]eview of defense industry mergers and joint ventures by the antitrust enforcement agencies is in the public interest and continues to protect the DOD (and ultimately the United States taxpayer) against the risk that a transaction may create a firm or group of firms with enhanced market power, i.e., the power to increase prices.  271  They thus agreed with enforcers that an exemption for defense mergers was unwarranted. 272 Some observers did worry that a continued focus on antitrust economics could undermine security. As the Defense Science Board remarked, however: Most claims that a merger or joint venture is important to national security are recognized by the antitrust agencies as \"efficiencies\" as that term is used in the Merger Guidelines-i.e., the combined firms can produce a better product at a lower price, maintain long-term R&D capacity, or put together complementary resources or staff that will produce a superior product. There may be rare situations, however, where DOD may assert that a transaction is essential to national security, even though it may substantially increase concentration and not produce cost savings.  273  The DoD noted that under National Society of Professional Engineers v. United States, 274 such arguments were \"not an acceptable counterbalance to potential anticompetitive effects.\"  275  Contrary to Professional Engineers, courts do seem to consider these national security arguments,  276  but \"no otherwise illegal defense industry merger reviewed by the courts has survived a preliminary injunction motion, or otherwise resulted in dismissal of a government charge, on a determination that public equities like national security outweighed anticompetitive effects.\"  277  The DoD thus recommended raising noneconomic national security considerations with enforcement agencies to inform their prosecutorial decisions.  278  To summarize, the consensus at the time was that DoD could provide useful information on competitive conditions to antitrust enforcers, but that its recommendations should not be conclusive because the antitrust enforcers were capable of balancing economic and security considerations well.  279  Instead, the DoD recommended enhanced communication with the antitrust authorities to more consistently offer useful advice on proposed mergers.  280  Following the recommendations from the 1994 report, DoD formalized internal procedures for developing a position on pending mergers and increased communication with the antitrust authorities.  281  Those procedures were modified in 2017 and remain in force in their amended form.  282  In 1997, an antitrust practitioner reported on the effect of the changes:  [R] elations between DoD and the antitrust enforcement agencies appear to have evolved significantly during the last two years. DoD now has handled enough transactions that it has a significant positive track record with both FTC and DoJ. Dealings between the agencies are frequent-and apparently are relatively open and respectful. Also, in virtually all cases DoD makes program and contracts officials available to the antitrust enforcement agencies to answer questions and provide any requested information. Where no formal DoD position is developed, DoD nonetheless nay be involved quite extensively with the antitrust enforcement agencies, or its activities may be very limited. Oftentimes, DoD may do little more than provide the antitrust enforcement agencies access to program and contracts personnel. Where there is no department position, interested DoD officials may convey views to the antitrust agencies that do not have \"official\" status.  283  However, even as late as 1998, disagreements within DoD complicated disposition of proposed mergers.  284  The Defense Science Board released a report on vertical integration in 1997.  285  The report found that vertical integration had not yet adversely affected DoD acquisition, but warned that it might in the future.  286  The authors also concluded that the DoD and antitrust enforcers were \"capably and effectively identifying and addressing vertical concerns arising in defense transactions\"  287  further evidence that the post-1994 changes were effective.  288  The 1997 report does not appear to have led to formal changes in DoD comparable to those following the 1994 report. \n Dimensional Analysis Given that there were many mergers during this period,  289  this Dimensional Analysis tries to fairly summarize overall trends from this era. \n Conflicting Considerations: No (generally). Although there were some cases in which individual Defense officials or offices disagreed with antitrust authorities' approach to cases, 292 DoD's official position was generally cooperative or at least acquiescent. Recent years have seen another wave of defense industry M&A.  294  This wave has not attracted nearly as much scholarly attention as previous ones,  295  though regulators have been attentive to it. For the purposes of this paper, the most significant development during this period was the DoD's increasing concern about defense industry consolidation. DoD officials began sounding the alarm about defense industry concentration in late 2015 following the acquisition of helicopter manufacturer Sikorsky by Lockheed-Martin.  296  Concerned that the antitrust authorities were too lenient on proposed mergers, the DoD began to develop a legislative proposal that would give DoD independent authority to review defense industry mergers.  297  The antitrust authorities responded by issuing a joint statement that acknowledged DoD's concerns, but insisted that the existing merger guidelines were flexible enough to handle defense industry mergers well.  298  After the rebuke, DoD withdrew their legislative proposal.  299  2. Activeness: Passive. DoD was specifically concerned that the antitrust authorities were being too lenient during merger review. 300 3. Presidential Involvement: No. I did not find any evidence of this. 4. Conflicting Considerations: Yes. DoD was initially concerned that the (perceived) leniency of merger review was detrimental to national security. 301 DoD wanted independent authority to block merger on national security grounds, whereas the antitrust authorities insisted pure economic analysis was sufficient. Although national security considerations have influenced antitrust enforcement, this is usually because the antitrust-relevant conduct is also relevant to national security. In my terminology, these are \"direct\" uses of antitrust. Out of the eleven cases studied herein, the USG used antitrust indirectly in only four.  306  Of the four indirect uses of antitrust, in only one case did the USG take a tougher stance towards the defendant due to national security considerations:  307  Bananas in Guatemala (1953).  308  In all other cases where the USG took an active stance towards potentially anticompetitive conduct (i.e., sued or threatened to sue), the anticompetitive conduct was directly relevant to the USG's national security concerns. Recall that in the Bananas in Guatemala (1953) case, the DOJ had originally postponed a suit against the United Fruit Company to protect the American company's interests in Guatemala,  309  but reversed course and sued UFCO to avoid appearing too lenient on the company.  310  Finally, note that it is much more likely for the USG to be more lenient on companies (such as by dropping, reducing, or postponing charges) to advance unrelated national security objectives. 311 I.D.3. In cases where national security and economic considerations conflict, economics has been given more weight over time. In all of the cases with conflicting considerations between 1940 and 1972, national security considerations generally had greater influence on case disposition.  312  However, since the 1970s, economic considerations better predicted the outcome in two  313  out of three 314 cases with conflicting considerations. Although this is a small sample, there are reasons to think this is a meaningful trend. The late 1970s marked the beginning of the Chicago school analysis of antitrust law, which posited maximizing consumer welfare as the primary-or sole-goal of antitrust law.  315  It therefore makes sense that other considerations, including national security, were seen as less important to case resolution thereafter. The end of the Cold War and the absence of total war since World War II may also have played a role in diminishing the comparative importance of national security considerations. If so, then this trend might reverse if the comparative importance of national security returns to twentieth century levels. I.D.4. The president plays an important role in reconciling conflicting considerations. In five  316  of the seven 317 cases with conflicting considerations, the president played a role in setting antitrust enforcement priorities. This is unsurprising-and largely unobjectionablegiven the president's role as the chief law enforcement official in the United States.  318  II. Discussion: The Qualcomm Case and National Security in Advanced Technologies AI and other advanced technologies implicate both antitrust policy (especially with regard to industry concentration) and national security. 319 Thus, it should not be surprising to see technology firms simultaneously balancing antitrust and national security demands. Indeed, this is already happening for firms like Google 320 and Amazon.  321  At the same time, there is increasing enthusiasm for reexamining the role of Chicago-style economic analysis in antitrust, much of which is motivated by the size of technology firms.  322  2020 Democratic presidential candidates Elizabeth Warren  323  and Bernie Sanders 324 championed antitrust proposals as well. Such an expansion of antitrust's policy goals, if realized, may leave a greater opening for national security considerations in the antitrust enforcement process. In this atmosphere of increased scrutiny, both opponents and defenders of big tech companies are wielding national security arguments. Opponents argue that a less concentrated-and therefore more competitive and innovative-tech sector will be crucial to prevailing in great power conflict with China and defense procurement more broadly.  325  Defenders argue that technology firms' size is instead a source of national strength due to their ability to counterbalance similarly sized foreign competitors.  326  dominant.  329  The FTC won a permanent injunction in the district court, 330 but the Ninth Circuit stayed that injunction order pending appeal.  331  On appeal, in an extraordinary move, the DOJ filed a brief in favor of Qualcomm.  332  In its brief, the DOJ argued that the injunction against Qualcomm \"would significantly impact U.S. national security\" 333 by diminishing Qualcomm's R&D expenditures 334 and reducing America's ability to compete in global 5G markets.  335  The brief also contained a statement from Ellen Lord, Under Secretary of Defense for Acquisition and Sustainment, who agreed that the injunction threatened national security, particularly emphasizing how harming Qualcomm could undermine American efforts to reduce China's dominance in 5G.  336  The FTC, in its answering brief, 337 strongly argued that these national security arguments were incognizable under modern economics-focused antitrust law,  338  while also disputing the assertion that the injunction would harm innovation and therefore national security.  339  The Ninth Circuit apparently agreed with DOJ, citing their national security concerns as weighing in favor of a stay.  340  As of the time of this writing, whether and how national security will influence the final disposition of the case remains to be seen. In setting a cellular communications standard, SSOs often include technology in the cellular communications standard that is patented. Patents that cover technology that is incorporated into a standard are known as \"standard essential patents\" (\"SEPs\"). Importantly, before incorporating a technology into a standard, SSOs \"often require patent holders to disclose their patents and commit to license [SEPs] on fair, reasonable, and non-discriminatory (\"FRAND\") terms.\" \"Absent such requirements, a patent holder might be able to parlay the standardization of its technology into a monopoly in standard-compliant products.\" [SSOs] require each party that participates in the standard setting process \"to commit to license its SEPs to firms that implement the standard on FRAND terms.\" \"Most SSOs neither prescribe FRAND license terms nor offer a centralized dispute-resolution mechanism in the event that a patent holder and standard implementer cannot agree on [FRAND] terms.\" Instead, \"most SSOs rely on the outcome of bilateral negotiations between the parties, with resort to remedies available from courts in the event of disagreement.\" We can expect to see more cases like Qualcomm in the future, with large tech companies facing both antitrust and national security demands. Which approach to antitrust ultimately serves national security best is beyond the scope of this paper. 341 However, those who would use antitrust to advance noneconomic national security goals should not expect doing so to be straightforward, even if they are correct about how to advance national security and disagree with the arguments for an economics-focused antitrust policy. The injection of national security considerations unrelated to competition into antitrust enforcement seems to run counter to a number of important trends. First, it seems inconsistent with the Supreme Court's decision in the landmark National Society of Professional Engineers case, which demands a focus of conduct on economic conditions. 342 Second, it is inconsistent with the USG's economics-focused position towards antitrust: a position which it has repeatedly espoused recently in international fora. 343 Third, it is rare-indeed, nearly unprecedented-for the USG to actively use antitrust enforcement to achieve unrelated national security goals.  344  Finally, using antitrust to advance national security despite countervailing economic considerations contravenes an existing domestic trend towards resolving conflicts between economics and national security in favor of the former.  345  To be sure, some national security concerns are cognizable as economic harms, such as consolidation in an industry that sells to the U.S. military. 346 Yet, to be consistent with the trends just identified, such harms must be framed as harmful to the USG in its capacity as a consumer, and be balanced against any countervailing economic efficiencies and other longstanding antitrust principles.  347  Arguments about how an antitrust action will affect, for example, great   307 (1919) ). Thus, refusing to deal with the DoD or USG generally should not, on its own, be grounds for antitrust action. Furthermore, It is settled law that [the offense of monopolization] requires, in addition to the possession of monopoly power in the relevant market, \"the willful acquisition or maintenance of that power as distinguished from growth or development as a consequence of a superior product, business acumen, or historic accident.\" The mere possession of monopoly power, and the concomitant charging of monopoly prices, is not only not unlawful; it is an important element of the free-market system. The opportunity to charge monopoly prices-at least for a short period-is what attracts \"business acumen\" in the first place; it induces risk taking that produces innovation and economic growth. To safeguard the incentive to innovate, the possession of monopoly power will not be found unlawful unless it is accompanied by an element power conflict,  348  are thus currently incognizable unless mediated by claims about how the targeted conduct will affect consumers (including the USG). Such national security claims are therefore superfluous to consumer-centric claims, which must ultimately ground those of national security. However, this does not bar heightened prosecutorial scrutiny of actions that are harmful from both economic and noneconomic national security perspectives. Such elevated scrutiny makes sense given limited prosecutorial resources, the legitimate importance of protecting national security,  349  and the role the DOJ serves in protecting the economic interests of the USG's national security apparatus.  350  The key distinguishing feature is not that national security considerations enter the USG's decision-making process for allocating scarce prosecutorial resources, but rather whether such considerations are allowed to override economic analysis of a case. In conclusion, AI firms are likely to face both antitrust and national security demands. This is not historically unique. Although there are proposals to increase use of antitrust to achieve noneconomic aims, these are controversial and contrary to recent antitrust policy and history. The trend over the past fifty years has been to keep unrelated national security concerns siloed from the economic analysis driving antitrust decisions. However, where potential anticompetitive conduct is also detrimental to national security, we should not be surprised if the USG takes a more aggressive approach to enforcement. of anticompetitive conduct. The request was stopped by [Office of Defense Transportation Director Joseph B.] Eastman, with the approval of Under Secretaries [Robert P.] Patterson [of the War Department] and [James] Forrestal [of the Navy Department]. \n 221    \n 4 4 \n 254 5 . 5 Conflict Winner: National Security. Reagan chose to halt the investigation for foreign policy reasons. 255 I.B.10. Wave of Defense Industry Mergers (1985-99) \n 293 5 . 5 Conflict Winner: N/A. I.B.11. DoD Antitrust Review Proposal (2015) \n 4 . Conflicting Considerations: Whether national security and economic considerations suggested opposite courses of action.22   5. Conflict Winner:In cases where there were conflicting considerations, which of those dominated as evidenced by the actual course of action pursued. I.B. Cases in Chronological Order 23 I.B.1. German Espionage during World War I (1916) 24 \n 24 United States v. Rintelen, 233 F. 793 (S.D.N.Y. 1916), aff'd sub nom. Lamar v. United States, 260 F. 561 (2d Cir. 1919); United States v. Bopp, 237 F. 283 (N.D. Cal. 1916). 25  See Steuer & Barile, supra note 6, at 7. 26 15 U.S.C. § 1. \n 39 33 United States v. Bausch & Lomb Optical Co., 34 F. Supp. 267 (S.D.N.Y. 1940). 34 Fisher, National Emergencies II, supra note 6, at 1189. Dividing markets like this is a violation of the Sherman Act. See, e.g., White Motor Co. v. United States, 372 U.S. 253, \n case was another.42     Dimensional Analysis 1. Directness: Direct. The cartelization of military optical instruments was of both national security and economic concern. 2. Activeness: Active, since the DOJ sought and entered into a consent decree. 43 3. Presidential Involvement: No. I did not find any evidence of this. 4. Conflicting Considerations: No. Although the security concerns here may have been foremost, geographic market division is a classic economic antitrust offense. 44 Thus, both economic and national security considerations favored antitrust enforcement here. 5. Conflict Winner: N/A. I.B.3. Rubber Manufacturing (1942) 45 \n See infra § I.B.3. 43 See supra note 36 and accompanying text. 44 See supra note 34. 45 United States v. Standard Oil Co., 1940-43 Trade Cas. (CCH)  ¶ 56,198 (D.N.J. 1942) [hereinafter Standard Oil I], amended by 1940-43 Trade Cas. (CCH)  ¶ 56,269 (D.N.J. 1943). 46 See generally id.; JOSEPH BORKIN, THE CRIME AND PUNISHMENT OF I.G. FARBEN 76-94 (1978); Maurice E. Stucke, Should the Government Prosecute Monopolies?, 2009 U. ILL. L. REV. 497, 524-25 (2009); Frank L. Kluckhohn, Arnold Says Standard Oil Gave Nazis Rubber Process, N.Y. TIMES, Mar. 27, 1942. 47 As with geographic markets, see supra note 34, dividing product markets violates Section 1 of the Sherman Act. See United States v. Associated Patents, Inc., 134 F. Supp. 74 (E.D. Mich. 1955), aff'd sub nom. Mac Inv. Co. v. United States, 350 U.S. 960 (1956) (per curiam). 48 See BORKIN, supra note 46, at 46-52. 49 Id. at 51 (quoting Investigation on the National Defense Program: Hearing Before the Special Comm. Investigating the Nat'l Def. Program, 77th Cong. 4312 (1942)). \n See id. 89. 73 Standard Oil I, 1940-43 Trade Cas. (CCH) ¶ 56,198. Ten individual Standard defendants pleaded guilty to criminal antitrust charges and received $5,000 in fines each. See BORKIN, supra note 46, at 91. 74 See Standard Oil I, 1940-43 Trade Cas. (CCH) ¶ 56,198 at 706. 75 See id. at 710. 76 See id. at 710-11. 77 See id. at 711. 78 STOCKING & WATKINS, supra note 65, at 106. 79  See Kluckhorn, supra note 46.80  See STOCKING & WATKINS, supra note 65, at 116-17.81  See id.82  See supra note 73 and accompanying text.83  See supra note 70 and accompanying text. 84 Supra § I.B.2. \n 2. Activeness: Passive. Although Arnold attempted to prosecute the railroad, 92 the case was ultimately quashed on national security grounds. 93 3. Presidential Involvement: Yes. President Roosevelt suspended antitrust enforcement. 94 4. Conflicting Considerations: Yes, since Arnold wanted to prosecute the case but the various national security departments and, ultimately, the president disagreed. 5. Conflict Winner: National Security, since the case was ultimately quashed. 95 I.B.5. The Oil Cartel Case (1952) 96 Facts This case concerns a number of agreements, dating back to the 1920s, 97 between major European (Royal Dutch Shell and Anglo-Iranian Oil [BP]) and American (Standard Oil of New Jersey [Exxon], Socony Mobil [Mobil], Standard Oil of California [Chevron], Texaco, and Gulf Oil) oil 91 Bernard M. Hollander, Fifty-Eight Years in the Antitrust Division: <PHONE> at 76-77 (2007) (unpublished oral history) (on file with author). \n 104   • \"[T]he prices [the major oil corporations] charged for oil delivered to Europe under the Marshall Plan . . . were higher than those charged to their own affiliates.\"105  \"As Secretary of Defense James Forrestal wrote [President] Truman in 1948, '[w]ithout Middle East Oil the [Marshall Plan] has a very slim chance of success.'\" 106 • An FTC report described the agreements as \"detrimental to the . . . security of the United States . . . .\" 107 Taking all of the above into account, \"Truman decided by June 1952 that an international oil cartel did exist, which was more harmful to American interests than any action a grand jury might take against it. On June 23, therefore, he ordered the Justice Department to begin a grand jury [criminal] investigation of the seven multinational oil giants. In letters to the secretaries of Defense, State, Interior, and Commerce and to the FTC, he asked these agencies to cooperate with the Justice Department in gathering evidence for the legal proceedings to follow.\" 113 As the case developed, a series of foreign policy crises and exigencies undermined Justice's attempts at antitrust enforcement. In spring 1951, the government of Iran nationalized BP, precipitating a political crisis for the U.S. 114 This strengthened the conviction of officials in the National Security Council and Departments of State and Defense to oppose further prosecution of the domestic oil cartel case, which they argued \"would be fodder for the Soviet propaganda machine and lead to further nationalization of American foreign oil interests.\" 115 However, after negotiations with Iran stalled, the United Kingdom launched a boycott of Iranian oil, leaving the U.S. responsible for ensuring an adequate oil supply to Europe. 116 This was likely to require coordination, though this immediately implicated antitrust. 117 President Truman considered invoking his powers under the Defense Production Act 118 (\"DPA\") to \"make exceptions to the antitrust laws for national-security reasons\" 119 but \"was reluctant to do so because of the pending case . . . .\" 120 \"The Justice Department and FTC also opposed granting any antitrust immunity to the oil industry.\" 121 \"In contrast, the Departments of State and Interior, the latter through the Petroleum Administration for Defense (PAD), urged the exception to be made . . . .\" 122 Truman ultimately agreed with State and Interior and allowed the exception. 123 However, national security advisors urged keeping the report secret on the grounds that it would \"greatly assist Soviet propaganda, would further the achievement of Soviet objectives throughout the world and [would] hinder the achievement of U. S. foreign policy objectives, particularly in the Near and Middle East.\" 108 After the report leaked, the same advisors \"strongly urged against prosecuting the oil majors, pointing out the damage such prosecution might do to American interests abroad. President Truman shared this view.\" 109 However, while a Senator, Truman had described oil cartel sales to Axis powers as \"approach[ing] treason.\" 110 As president, he wanted to \"strengthen[] the powers of the federal government over oil\" 111 and generally supported strengthening antitrust law. 11298  See id. at 11-12.99  See id. at 20-21.100  See generally id. at 3-14.101  Id. at 22.102  See id. at 22-23.103  See id.104  Id. at 28.105  Id.106  Id. at 29.107  See id. \n National Security (generally). National UFCO] gambled that national security needs were stronger than Eisenhower's commitment to antitrust-and in the short run they were right. When the Guatemalan government decided in February 1953 to nationalize more property, dozens of U.S. congressmen bombarded the State Department with telegrams, urging a strong line in support of United Fruit and in defense of American overseas investment. Company officials echoed this sentiment in a meeting with Assistant Secretary of State John Moors Cabot on May 6, 1953. For his part, Cabot was already convinced that Communism was an \"international conspiracy and ipso facto a menace to everybody in the world.\" Capitalizing on this conviction, company officials soon turned the discussion away from United Fruit's great unpopularity in Latin America to a discussion of the antitrust suit. Samuel G. Baggett, vice president of United Fruit, brought up the subject, claiming that a suit would prove \"very damaging\" to the company at a time when its entire Latin American operations were in jeopardy. \"No one,\" he emphasized, \"would believe there was not something seriously wrong with an American company being sued by its own government.\" Apparently worried at the prospect of antitrust action stifling American business investment in Guatemala (and therefore development thereof), however, Deputy Undersecretary of State Murphy raised the alarm within State.184  State then reversed its official position and told Justice that it \"must therefore seek a swift settlement through a consent decree that only addressed United Fruit's most egregious behavior.\" 185 Justice, however, continued aggressive enforcement against UFCO, rejecting State's consent decree proposal.186    \"Hoping to destroy the company's effort to undercut the lawsuit once and for all, Justice officials explained to their counterparts in [State] that too often corporations attempted to escape antitrust action by pitting the government's diplomats against its own lawyers. Conceding the point, Loftus Becker, the State Department's legal counsel, replied that his department would no longer muscle in on the case.\" 187 Nevertheless, the case ultimately ended in a modest consent decree 188 that gave substantial deference to foreign policy concerns and UFCO itself.189    This case is interesting because it shows how responsive antitrust enforcement can be to foreign policy considerations. Initially, foreign policy actors successfully shaped Department of Justice antitrust enforcement to further foreign policy goals (namely, anti-communism). However, when the U.S.'s strategic interest changed and it became desirable for State to distance itself from UFCO, State took a more hands-off approach (for a limited time). After State concluded it retained an interest in the case, it began to pressure Justice towards a tidier solution. The final resolution of the case, like the rest of it, distinctly favored foreign policy concerns over pure antitrust concerns. As noted, the USG's approach to enforcement changed from passive192   to active193  and back to passive 194 again based on changing national security considerations. 3. Presidential Involvement: Yes. 195 4. Conflicting Considerations: Yes. Throughout this case, antitrust enforcers disagreed with national security officials about what to do. 5. Conflict Winner: security officials were successful at getting initial leniency for UFCO. 196 Enforcement only resumed once the considerations were no longer conflicting. 197 Even when the State Department agreed to stop meddling in the case, 198 the ultimate resolution of the case was close to what they wanted. 175  \"[T]he [American] Justice Department had just finished its preliminary investigation of United Fruit. The investigation had found ample cause to pursue antitrust action against the company's monopoly of the Central American banana industry . . . .\" 176  In light of Árbenz's expropriation, however, United Fruit sought to leverage the USG's anticommunist stance into antitrust leniency: 177  [178 Concerned about promoting American investment abroad, high-level government officialsincluding the president, secretary of state, and assistant secretary of state-intervened to encourage private settlement and successfully postpone the suit.179    In 1954, U.S.-backed Castillo Armas toppled the Árbenz regime. 180 Thus, the secretary of state (initially) no longer had objections to antitrust action against UFCO. 181 \"Indeed, he and others apparently realized that, in the aftermath of Armas's coup, they now had more to gain by putting distance between the United States and United Fruit.\" 182 This is because the Eisenhower administration wanted to show both to domestic journalists and Central American governments that they were not \"handmaidens of United Fruit.\"183   190    Dimensional Analysis 1. Directness: Indirect. UFCO's antitrust offenses were, at most, tangentially related to enforcement decisions.191  In general, unrelated national security considerations guided many enforcement decisions. 2. Activeness: Both. 199 I.B.7. Yellowcake Uranium(1972)     \n Since 1970, \"military uses [of uranium] ha[d] been met from previously accumulated stockpiles.\" 207 To protect its domestic uranium suppliers from falling prices, the U.S. \"imposed an embargo on enrichment of foreign uranium for use in United States reactors.\" 208 \n from the Government of Canada and the United States State Department. The governments of Australia, Canada, and Great Britain all lodged complaints with the State Department over the Justice Department's efforts to subpoena information in their countries. (In Australia and Canada it was actually illegal to cooperate with United States investigators . . . .) The Justice Department denies that it was taking orders from State. Even if it weren't, attorneys familiar with the case have noted that the attorney general's office is quite capable of determining on its own what kind of cases and charges are consistent with United States foreign policy goals, and handling its affairs accordingly. It should be noted that President Carter has openly expressed his desire to see uranium mining developed in Australia to provide a greater quantity of raw uranium resources for the free world. This complement[ed] his international nuclear policies of discouraging [nuclear] breeder and reprocessing technologies. Given these goals President Carter would not want to antagonize the government of Australia. 215 \n . Conflicting Considerations: Yes. Price-fixing is illegal under the antitrust laws, but foreign policy considerations would have suggested leniency on the cartel and especially the foreign cartel participants.222   5. Conflict Winner: National Security (speculative). Again, we do not know for certain whether national security was the reason for the lenient resolution, but Taylor and Yokell suggest it is.223     I.B.8. AT&T Breakup (1974) 224 \n 302   5. Conflict Winner: Economics. DoD ultimately withdrew its proposal.303I.C. Dimensional Summary of Case Studies national security considerations favor antitrust scrutiny where anticompetitive behaviors bottleneck key war supplies 304 or increase defense industry concentration. 305 I.D.2. It is rare for the USG to actively use antitrust enforcement to advance unrelated national security objectives. Case Name Directness Activeness Presidential Involvement Conflicting Considerations \n 341 For analysis of this question, see FOSTER & ARNOLD, supra note 4. 342 435 U.S. at 688 (\"Contrary to its name, the Rule [of Reason] does not open the field of antitrust inquiry to any argument in favor of a challenged restraint that may fall within the realm of reason. Instead, it focuses directly on the challenged restraint's impact on competitive conditions.\"); accord DEF. SCI. BD., CONSOLIDATION REPORT, supra note 263, at 32.343  See DEF. SCI. BD., CONSOLIDATION REPORT, supra note 263, at 32; OECD, supra note 11, at 2; Remarks of Deputy Assistant Attorney Gen. Roger Alford, 2019 China Competition Policy Forum (May 7, 2019), https://perma.cc/CR2D-79SE. Accord DEF. SCI. BD., CONSOLIDATION REPORT, supra note 263, at 15. 347 For example, \"a business 'generally has a right to deal, or refuse to deal, with whomever it likes, as long as it does so independently.'\" Aspen Skiing Co.v. Aspen Highlands Skiing Corp., 472 U.S. 585, 601 n.27 (1985) (quoting Monsanto Co. v. Spray-Rite Service Corp., 465 U.S. 752, 761 (1984); United States v. Colgate & Co., 250 U.S. 300, 344  See supra § I.D.2.345  See supra § I.D.3.  346   \n\t\t\t See, e.g., GREG ALLEN & TANIEL CHAN, ARTIFICIAL INTELLIGENCE & NATIONAL SECURITY (2017), \n\t\t\t Note the inclusion of the word \"substantial\": some cases have national security relevance, but it seems to have been a minor factor in the case. See, e.g., United States v. Microsoft Corp., No. 98-1232, at 1 (D.D.C. Sept. 28, 2001) (pre-trial conference order) (ordering parties to expedite the settlement process due to the September 11th terrorist attacks). Furthermore, some of the \"cases\" discussed herein constitute multiple legal cases or general trends from a specific historical era.21  A stylized example might help illustrate. Suppose XYZ Corp. sells two products: guns and butter. Butter sales are not relevant to national security, but gun sales to the USG are. XYZ wants to merge with another butter seller. The USG also wants XYZ to lower its gun prices. If the USG uses XYZ's pending merger as leverage to get XYZ to lower its gun prices, then it will have used antitrust in an indirect fashion. If, on the other hand, XYZ was proposing to merge with another gun seller, the merger would have been directly relevant to the USG's desire for lower gun prices, and so conditioning the merger on lower gun prices would have been directly relevant to the USG's national security goals.22  See supra note 10 and accompanying text. \n\t\t\t See id. at 30.109  Id. at 30-31.110  See id. at 31.111  Id.112  See id. at 31-32. \n\t\t\t  Id.  at 32 (emphasis added).114  See generally id. at 38-40.115  See id. at 40.116  See id. at 41-42.117  See id. at 42. 118 50 U.S.C. App. § 2061 et seq.119  See KAUFMAN, supra note 96, at 42.120  Id.121  Id.122  Id.123  See id. at 43.124  See id.125  See id. \n\t\t\t See id. at 88.156  See id. at 88-89.157  Id. at 89.158  See id. at 90.159  See id.160  See generally id. at 90, 93-101.161  See id. at 95-96.162  See id. at 99.163  See id.164  See id. at 149. \n\t\t\t See supra notes 100-102 and accompanying text.166  See supra notes 103-113 and accompanying text.167  See supra notes 114-125 and accompanying text.168  See supra notes 126-128 and accompanying text.169  See supra notes 129-134 and accompanying text.170  See supra notes 108-127 and accompanying text.171  See supra notes 103-113 and accompanying text. \n\t\t\t  Id.  at 658-59.179  See id. at 659-61.180  See id. at 662181  See id.182  Id. (emphasis added).183  See STEPHEN G. RABE, EISENHOWER AND LATIN AMERICA58 (1988); cf. PIERO GLEIJESES, SHATTERED HOPE: THE GUATEMALAN REVOLUTION AND THEUNITED STATES, <PHONE> 361 n.3 (1991 (\"Some portray a U.S. government that is putty in the hands of the company, conveniently overlooking evidence that might temper or complicate this thesis, including the fact that the government initiated an antitrust suit against UFCO shortly after the fall of Arbenz.\"); IMMERMAN, supra note 175, at 123 (\"The official line of Eisenhower's [Guatemala] policy defended United Fruit's interests so avidly that political scientist and former State Department member Cole Blasier wrote that the United States government entered into the controversy as a virtual speaker for the company\").184  See Khula, supra note 6, at 663-64. \n\t\t\t See id. at 664.186  See id. at 665.187  Id. at 668.188  See id. at 668-70.189  See id.190  See id. at 670 (\"From postponement to prosecution to consent decree, the case had been dictated by the paramount concern of President Eisenhower and his foreign policy elite: the necessities and vagaries of national security.\").191  The decision to prosecute UFCO after Armas came to power was proximately motivated by the USG wanting to distance itself from UFCO. See supra notes 180-183. However, this desire in turn was probably related to UFCO's anticompetitive behavior. Since the relation between the conduct and the enforcement decision was mediated by the USG's desire to be perceived in a certain way, I have categorized this as indirect.192  See supra notes 176-179 and accompanying text.193  See supra notes 181-183 and accompanying text.194  See supra notes 184-189 and accompanying text.195  See supra note 179 and accompanying text.196  See supra notes 176-179 and accompanying text. \n\t\t\t See supra notes 184-189 and accompanying text.198  See supra note 187 and accompanying text.199  See supra notes 188-189 and accompanying text.200  See JUNE H. TAYLOR & MICHAEL D. YOKELL, YELLOWCAKE 13 (1979).201  See id. at 14-15.202  See id. at 15. \n\t\t\t See TAYLOR & YOKELL, supra note 200, at 175.215  Id. at 176-77 (emphasis added).216  See id. at 177.217  See generally id.218  See supra note 215 and accompanying text.219  See supra note 216 and accompanying text.220  See supra note 213 and accompanying text.221  Cf. supra note 215 and accompanying text (discussing Executive Branch dynamics in the case). \n\t\t\t See supra note 215 and accompanying text.223  See supra note 215 and accompanying text. \n\t\t\t See id.; Friedman, supra note 237, at 183-86; President Halts Grand Jury Antitrust Inquiry; British Reject Low Fares Despite U.S. Action, AVIATION WK. & SPACE TECH., Nov. 26, 1984, at 36; Rill & Turner, supra note 225, at 589. The Laker actions were only the latest such incident-the British government had long protested extraterritorial application of American antitrust law. See Friedman, supra note 237, at 184-85 (\"The government of the United Kingdom ha[d] historically been opposed to United States antitrust policy insofar as it affects business enterprises based in the United Kingdom, and ha[d] long objected to the scope of extraterritorial prescriptive jurisdiction asserted by United States courts for its antitrust laws.\" (footnote omitted)). For more background on extraterritorial application of American antitrust law, see supra note 213. 246 Feder, supra note 244; Rill & Turner, supra note 225, at 589. 247 See Friedman, supra note 237, at 216-17. At the time, the Bermuda II agreement between the two countries required governmental approval of proposed routes, capacities, and fares on transatlantic flights. See generally Rill & Turner, supra note 225, at 589 n.49.248  See id. at 589.249  See id.250  See supra notes 244-246 and accompanying text.251  See supra note 247 and accompanying text.252  See supra note 248 and accompanying text.253  See supra notes 248-249 and accompanying text.254  See supra notes 248-249 and accompanying text.255  See supra notes 248-249 and accompanying text. \n\t\t\t See M&A Statistics, INSTITUTE FOR MERGERS, ACQUISITIONS AND ALLIANCES, https://perma.cc/75JC-UVAB (archived Apr. 24, 2020). 257 Defense and aerospace are considered deeply interrelated. See, e.g., Eric J. Stock, Explaining the Differing U.S. and EU Positions on the Boeing/McDonnell-Douglas Merger: Avoiding Another Near-Miss, 20 U. PA. J. INT'L ECON. L. 825, 836-41 (1999). \n\t\t\t See supra note 296 and accompanying text.301  See supra note 296 and accompanying text.302  See supra notes 297-299 and accompanying text.303  See supra note 299.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/How-Will-National-Security-Considerations-Affect-Antitrust-Decisions-in-AI-Cullen-OKeefe.tei.xml", "id": "6c2435008e38eef0198819bcf7498029"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "How do we regulate a changing technology, with changing uses, in a changing world? This chapter argues that while existing (inter)national AI governance approaches are important, they are often siloed. Technology-centric approaches focus on individual AI applications; law-centric approaches emphasize AI's effects on pre-existing legal fields or doctrines. This chapter argues that to foster a more systematic, functional and effective AI regulatory ecosystem, policy actors should instead complement these approaches with a regulatory perspective that emphasizes how, when, and why AI applications enable patterns of 'sociotechnical change'. Drawing on theories from the emerging field of 'TechLaw', it explores how this perspective can provide informed, more nuanced, and actionable perspectives on AI regulation. A focus on sociotechnical change can help analyse when and why AI applications actually do create a meaningful rationale for new regulation-and how they are consequently best approached as targets for regulatory intervention, considering not just the technology, but also six distinct 'problem logics' that appear around AI issues across domains. The chapter concludes by briefly reviewing concrete institutional and regulatory actions that can draw on this approach in order to improve the regulatory triage, tailoring, timing & responsiveness, and design of AI policy.", "authors": ["Matthijs M Maas", "Justin Bullock", "Baobao Zhang", "Yu-Che Chen", "Johannes Himmelreich", "Matthew Young", "Antonin Korinek", "Valerie Hudson"], "title": "Aligning AI Regulation to Sociotechnical Change", "text": "Introduction How do we regulate a changing technology, with changing uses, in a changing world? As artificial intelligence ('AI') is anticipated to drive extensive change, the question of how we can and should reconfigure our regulatory ecosystems for AI change matters today. In just the past decade, advances in both AI research and in the broader data infrastructure have begun to spur extensive take-up of this technology across society (D.  Zhang et al. 2021) . As a 'general-purpose technology'  (Trajtenberg 2018 ), AI's impact on the world may be both unusually broad and deep. It may even prove as 'transformative' as the industrial revolution  (Gruetzemacher and Whittlestone 2020) . This may provide grounds for anticipation-and also for caution. Most of all, it is grounds for reflection on the choices that societies want to make and instil in the trajectory of this technology. While still at an early stage of development, uses of AI technology are already creating diverse policy challenges. Internationally, AI is the subject to intense contestation. Further AI progress, along with the technology's global dissemination, are set to further raise the stakes. It is clearly urgent to reflect on the purposes and suitability of the regulatory ecosystem for AI governance. The urgent question is; when we craft AI regulation, how should we do so? Many AI governance approaches at both the national and international level, remain hampered by siloed policy responses to individual AI applications (that is, they are technologycentric), or for AI applications' effects on individual legal fields or doctrines (that is, they are law-centric). In contrast, this chapter argues that to craft adequate AI policies, a better regulatory perspective takes a step back, and first asks a better question: when we craft AI regulation, what are we seeking to regulate? And what are the ingredients for a more systematic regulatory template for crafting AI governance? This chapter argues that to foster an effective AI regulatory ecosystem, policy institutions and actors must be equipped to craft AI policies in alignment with systematic assessments of how, when, and why AI applications enable broader forms of sociotechnical change  (Maas 2020) . It argues that this approach complements existing technologycentric and law-centric examinations of AI policies; and that, supported by adequate institutional processes, it can provide an informed and actionable perspectives on when and why AI applications actually create a rationale for regulation-and how they are consequently best approached as targets for regulatory intervention. This enables more tailored policy formulation for AI issues, facilitates oversight and review of these policies, and helps address structural accountability, alignment, and (lack of) information problems in the emerging AI governance regulatory ecosystem. The chapter is structured as follows. It first (1) sketches the general value of a 'change-centric' approach to AI governance. The chapter then (2) proposes and articulates a framework focused on 'sociotechnical change', and explores how this model allows an improved consideration of (3) when an AI application creates a regulatory rationale, and (4) how it is subsequently best approached as a regulatory target, considering 6 distinct 'problem logics' that appear in AI issues across domains. Finally (5), the chapter reflects on some of the limits of this approach, before discussing concrete institutional and regulatory actions that can draw on this approach in order to improve the regulatory triage, tailoring, timing & responsiveness, and regulatory design of AI policy. \n Towards change-centric approaches in an AI regulatory ecosystem In response to AI's emerging challenges, scholars and policymakers have appealed to a wide spectrum of regulatory tools to govern AI. It should be no surprise that recent years have seen increasing public demands for AI regulation (B.  Zhang and Dafoe 2020) , and diverse new national regulatory initiatives  (Cussins 2020; Law Library of Congress 2019) . Much work to date has focused on the regulation of AI within particular domestic regulatory contexts. For instance, this has explored the relative institutional competencies of legislatures, regulatory agencies, or courts at regulating AI  (Guihot, Matthew, and Suzor 2017; Scherer 2016) . Others have emphasised the governance roles of various actors in the AI landscape  (Leung 2019) , exploring for instance how tech companies' ethics advisory committees  (Newman 2020, 12-29) , AI company employee activists and 'epistemic communities'  (Belfield 2020; Maas 2019a ), AI research community instruments (such as scientific conference research impact assessment mechanisms)  (Prunkl et al. 2021) , or private regulatory markets architectures  (Clark and Hadfield 2019) , could all help shape AI regulation. There is also a growing recognition of the importance of global coordination or cooperation for AI governance  (Feijóo et al. 2020; Kemp et al. 2019; Turner 2018) . At the global level, much focus to date has been on the burgeoning constellation of AI ethics principles that has sprung up in the last half-decade  (Fjeld et al. 2019; Jobin, Ienca, and Vayena 2019; Schiff et al. 2020; Stahl et al. 2021; Stix 2021) . However, this is being increasingly complemented by regulatory proposals. These have ranged from relying on existing norms, treaty regimes, or institutions in public international law  (Kunz and Ó hÉigeartaigh 2021; Burri 2017; Smith 2020) . Others have proposed entirely new international organizations in order to coordinate national regulatory approaches  (Erdelyi and Goldsmith 2018; Kemp et al. 2019; Turner 2018, chap. 6) , or have compared such centralized institutions for AI to more decentralized or fragmented alternatives  (Cihon, Maas, and Kemp 2020b, 2020a) . Others have focused more on the role of soft law instruments  Gutierrez, Marchant, and Tournas 2020) , international standard-setting bodies  (Cihon 2019; Lorenz 2020) , or certification schemes  (Cihon et al. 2021) . Others have proposed the adaptation of existing informal governance institutions such as the G20  (Jelinek, Wallach, and Kerimi 2020) . This is clearly a diverse constellation of efforts. Yet there are underlying classes and patterns in these approaches to AI regulations. For instance, one group of 'technology-centric' approaches focuses on 'AI' as an overarching class that should be considered in whole  (Turner 2018) . While this is more reflective of the cross-domain application and impact of many AI techniques, however, this approach is not without problems. For one, it is undercut by intractable debates over how to define 'AI' (Russell and Norvig 2016)-and by the fact that AI is not a single thing  (Schuett 2019; Stone et al. 2016) . This shortfall is more addressed in the second 'technology-centric' approach, which is 'application-centric'-or, as some call it, 'use-case centric'  (Schuett 2019, 5) . This approach seeks to unpack the umbrella term 'AI', and split out the specific AI applications that regulation should focus on. In the past decade, many policy responses to AI have been sparked by one or another use case of AI technology-involving concrete issues that have been thrown up, or visceral problems that are anticipated-such as autonomous cars, drones, facial recognition, or social robots  (Turner 2018, 218-19) . As Petit puts it, this technology-centric approach to AI involves charting \"legal issues from the bottom-up standpoint of each class of technological application\"  (Petit 2017) . Application-centric approaches remain the default response to AI governance. However, like the AI-centric approach, this orientation also has shortfalls. For one, it emphasizes visceral edge cases, and is therefore easily lured into regulating edge-case challenges or misuses of the technology (e.g. the use of DeepFakes for political propaganda) at a cost of addressing far more common but less visceral use cases (e.g. the use of DeepFakes for gendered harassment)  (Liu et al. 2020) . Moreover, the resulting AI policies and laws are frequently formulated in a piecemeal and ad-hoc fashion, which means that this perspective can promote siloed regulatory responses  (Turner 2018, 218-19) . A focus on individual applications also may inadvertently foregrounds technologyspecific regulations even where these are not the most effective (Bennett Moses 2013). It moreover induces a 'problem-solving' orientation  (Liu and Maas 2021 ), aimed at narrowly addressing local problems caused (or envisioned) by the specific use case that prompted the regulatory process  (Crootof and Ard 2021) . A distinct set of regulatory responses to AI are instead law-centric. They invoke what Nicolas Petit has called a 'legalistic' approach, which 'consists in starting from the legal system, and proceed by drawing lists of legal fields or issues affected by AIs and robots.\"  (Petit 2017, 2) . This approach segments regulatory responses by departing from AI's impacts on-and within specific conventional legal codes or subjects (e.g. privacy law, contract law, the law of armed conflict), or by exploring the ways in which these create questions about the scope, intersection, assumptions, or adequacies of existing law  (Crootof and Ard 2021) . To be clear, technology-, application-, and law-centric approaches to AI regulation have important insights, and must play a role in any AI governance ecosystem. Nonetheless, they have their drawbacks. Importantly, AI governance proposals could be grounded in a better understanding of how AI applications translate into cross-domain changes, and how future capabilities or developments might further shift this problem portfolio  (Maas 2019b ). An alternative is therefore to shift (or complement) these approaches with a framework that is not (solely) anchored in 'new technology' (whether on the umbrella term of 'AI', or on isolated AI applications), nor on isolated legal domains, but which rather examines types of change. What impacts are we concerned about? AI regulatory ecosystem require a protocol for considering where, and why AI change warrants regulatory intervention, and how and when this regulatory intervention should take place. It should be able to adequately identify when AI applications create regulatory rationales, as well as the best levers to approach AI as a regulatory target. Can we reformulate an impact-focused approach for AI regulation, that provides superior levers for regulation? To achieve this, this chapter instead draws on existing theories from the emerging paradigms of law, regulation and technology, and 'TechLaw'-\"the study of how law and technology foster, restrict, and otherwise shape each other's evolution.\"  (Crootof and Ard 2021, n. 1; Ard and Crootof 2020) . In particular, it proposes to approach AI governance through the lens of 'sociotechnical change'  (Bennett Moses 2007a . As such, this chapter will now turn to how this approach can bridge the gap between technology/application-centric and law-centric approaches, by guiding reflection on when and why new AI applications require new regulation-and how the resulting regulatory interventions are best tailored. \n Reframing regulation: AI and sociotechnical change It should be no surprise that changes in technology have given rise to extensive scholarship on the relation between law and these new technologies. In some cases, such work has focused in on identifying the assumed 'exceptional' nature or features of a given technology  (Calo 2015) . However, as noted, other scholars have influentially argued that it is less the 'newness' of a technology that brings about regulatory problems, but rather the ways it enables particular changes in societal practices, behaviour, or relations  (Balkin 2015; Bennett Moses 2007b; Friedman 2001) . That is, in what ways do changes in a given technologies translate to new ways of carrying out old conduct, or create entirely new forms of conduct, entities, or new ways of being or connecting to others? When does this create problems for societies or their legal systems? Rather than focus on 'regulating technology' (either in general, or in specific applications), this scholarship accordingly puts a much greater emphasis on 'adjusting law and regulation for sociotechnical change' (Bennett  Moses 2017, 574) . In particular, Brownsword, Scotford, and Yeung have highlighted three dimensions of technological 'disruption': \"legal disruption, regulatory disruption, and the challenge of constructing regulatory environments that are fit for purpose in light of technological disruption\"  (Brownsword, Scotford, and Yeung 2017, 7) . As noted, various scholars of law, regulation and technology have emphasized the importance of 'sociotechnical change'. In her 'theory of law and technological change' (Bennett Moses 2007b), Lyria Bennett Moses has argued that questions of 'law and technology' are rarely if ever directly about technological progress itself (whether incremental or far-reaching). Instead, she argues that lawyers and legal scholars who examine the regulation of technology are focused on questions of \"how the law ought to relate to activities, entities, and relationships made possible by a new technology\" (Bennett  Moses 2007b, 591) . Indeed, Lyria Bennett-Moses has argued that from a regulatory perspective, \"[t]echnology is rarely the only 'thing' that is regulated and the presence of technology or even new technology alone does not justify a call for new regulation\"(Bennett  Moses 2017, 575) . In doing so, she calls for a shift in approach from 'regulating technology' to 'adjusting law and regulation for sociotechnical change.' (Bennett  Moses 2017, 574)  This shifts the focus on patterns of 'socio-technical change' (Bennett  Moses 2007b Moses , 591-92, 2007a )instances where changes in certain technologies actually expand human capabilities in ways that give rise to new activities or forms of conduct, or new ways of being or of connecting to others  (Bennett Moses 2007b, 591-92) . As such, the question of governing new technologies is articulated not with reference to a list of (sufficiently 'new') technologies (Bennett Moses 2017, 576), but is relatively 'technology-neutral'. It is this functional understanding of 'socio-technological change' that informs more fruitful analysis of when and why we require regulation for new technological or scientific progress. It can also underlie a more systematic examination of which developments in AI technology are relevant for a regulatory system to focus on. \n AI as regulatory rationale Electronic copy available at: https://ssrn.com/abstract=<PHONE> What types of sociotechnical changes (e.g. new possible behaviours or states of being) actually give rise to regulatory rationales? When a technology might create an opportunity for certain problematic behaviour, does that opportunity need to be acted upon, or can the mere possibility of that behaviour constitute a regulatory rationale? Can sociotechnical changes be anticipated? This entails a more granular understanding of the dynamics of sociotechnical change-and when or how such changes can constitute a rationale for regulation. \n Varieties of sociotechnical change When and why do AI capabilities rise to a problem that warrants legal or regulatory solutions? It is important to recognize that not all new scientific breakthroughs, new technological capabilities, or even new use cases will necessarily produce the sort of 'sociotechnical change' that requires regulatory responses. In practical terms, this relates to the observations that new social (and therefore governance) opportunities or challenges are not created by the mere fact of a technology being conceived, or even prototyped, but rather by them being translated into new 'affordances' for some actors-relationships \"between the properties of an object and the capabilities of the agent that determine just how the object could possibly be used\"  (Norman 2013, 11) . AI affordances can be new types of behaviour, entities, or relationships that were not previously possible (or easy), and which are now available to various actors  (Liu et al. 2020) . How does technological change translate into sociotechnical change? When would this be disruptive to law? There are various types of sociotechnical changes that new AI applications can create or enable  (Maas 2019c, 33 ; see also  Crootof and Ard 2021) . (1) Allowing older types of behaviours to be carried out with new items or entities, including artefacts which are potentially not captured under existing (technology-specific) regulatory codes, or which blur the boundaries between existing domains or regimes, potentially causing problematic gaps, overlaps, or contradictions in how these behaviours are covered by regulation; (2) Absolute or categorical capability changes, where AI progress expands the action space and 'unlocks' new capabilities or behaviour which were previously simply out of reach for anyone, and which could be of regulatory concern for one of various reasons; (3) Relative capability changes, where AI increases the prominence of a previously rare behaviour, for instance because progress lowers thresholds or use preconditions for a certain capability (e.g. advanced video editing; online disinformation campaigns; cryptographic tools), which was previously reserved to a narrow set of actors; or because progress allows the scaling up of certain existing behaviours (e.g. phishing emails). (4) Positional changes amongst actors, where AI applications that drive shifts in which particular state actors are dominant, while leaving the general 'rules' of that international system more or less unaltered. (5) Changing structural dynamics in a (international) society, for instance, by; i. Shifting prevalent influence between types of actors (e.g. away from states and towards non-state actors or private companies); ii. Shifting the means by which certain actors seek to exercise 'influence' (e.g. from 'hard' military force to computational propaganda, or from multilateralism to 'lawfare', as a result of new communications technologies increasing the scope, velocity and effectiveness of such 'lawfare' efforts)  (Dunlap 2008, 146-48) ; iii. Altering the norms or identities of actors, and thereby changing the terms by which they conceive of their goals and orient their behaviour. As mentioned, there may be certain AI innovations or breakthroughs that do not create very large sociotechnical changes of these forms, even if from a pure scientific or engineering standpoint they involve considerable alterations to the state of the art. Conversely, technological change or improvements also need not be qualitatively novel, dramatic, sudden, or cutting-edge for them to drive intense and meaningful change in balances of power, or in societal structures  (Cummings et al. 2018, iv) . The question is therefore not only how large these sociotechnical changes are, but how, or whether, they touch on the general rationales for new regulatory interventions. \n Mapping regulatory rationales There are various accounts for when and why regulatory intervention is warranted by the introduction of new technologies. For instance, Deryck Beyleveld and Roger Brownsword argue that emerging technologies generally give rise to two kinds of concerns: \"one is that the application of a particular technology might present risks to human health and safety, or to the environment […] and the other is that the technology might be applied in ways that are harmful to moral interests\"  (Beyleveld and Brownsword 2012, 35) . However, while these may be the most prominent rationales, the full scope of reasons for regulation may extend further. In a non-technology context, Tony Prosser has argued that regulation, in general, has four grounds: \"(1) regulation for economic efficiency and market choice, (2) regulation to protect rights, (3) regulation for social solidarity, and (4) regulation as deliberation\"  (Prosser 2010, 18) . How do these regulatory rationales relate to technological change? As Bennett  Moses (2017, 578)  notes, all four of these rationales can certainly become engaged by new technologies. That is, new technologies (or new applications) can: (1) create sites for new market failures, warranting regulatory interventions such as technical standards or certification, to ensure economic efficiency and market choice, and remedy information inadequacies for consumers; (2) generate many new risks or harms-either to human health or the environment, or to moral interests-which create a need for regulation to protect the rights of these parties (e.g. restrictions of new weapons; the ban on human cloning). (3) create concern about social solidarity, as seen in concerns over the 'digital divide' at both a national and international level, creating a need for regulation to ensure adequate inclusion. (4) create sites or pressures for the exertion of proper democratic deliberation over the design or development pathways of technologies. (Bennett  Moses 2017, 579-83)  To be sure, as a technology-centric approach would note, these cases all involve new technologies which require regulation. However, Bennett Moses argues that in each of these cases, it is not the involvement of 'new technology' per se that provides a special rationale for regulation, above and beyond the resulting social changes (e.g. potential market failures; risks to rights; threats to solidarity; or democratic deficits) that are at stake (Bennett Moses 2017, 583). We are not worried about technology; we are worried about its effects. As such, the primary regulatory concern is over the emergence of the 'sociotechnical' effects that occur. This conceptual shift can help address one limit that regulatory or governance strategies encounter if they focus too much or too narrowly on technology. As she argues: \"… treating technology as the object of regulation can lead to undesirable technology specificity in the formulation of rules or regulatory regimes. If regulators ask how to regulate a specific technology, the result will be a regulatory regime targeting that particular technology. This can be inefficient because of the focus on a subset of a broader problem and the tendency towards obsolescence\" (Bennett Moses 2017, 584). As such, taking a sociotechnical (rather than a technology-centric) approach, this lens helps keep into explicit focus specific rationales for governance in each use case of AI: on what grounds and when regulation is needed and justified? These four accounts of rationales are valuable as a starting point for AI regulation. However, we can refine this account. For one, it is analytically valuable to draw a more granular distinction between (physical) harms to human health or the environment, and (moral) harms to moral interests  (Beyleveld and Brownsword 2012) . Moreover, these categories all concern rationales for governance to step in, in response to sociotechnical changes that are affecting society (i.e. the regulatees) directly. However, there may also be cases where AI-enabled sociotechnical change creates an indirect regulatory rationale, because it presents some risk directly to the existing legal order charged with mitigating the prior risks. In such cases of 'legal disruption'  (Liu et al. 2020; Maas 2019c) , sociotechnical change can produce a threat to the regulatory ecosystem itself. This can be because these tools allow regulatees to more effectively challenge or bypass existing laws, resulting in potential 'legal destruction'  (Maas 2019c ). Alternatively, it can result because certain AI tools can drive 'legal displacement'  (Maas 2019c) , by offering substitutes or complements to existing legal instruments, in shaping or managing the behaviour of citizens  (Brownsword 2019) . Drawing together the above accounts, one might then speak of a regulatory rationale for an AI system or application, whenever it drives sociotechnical changes (new ways of carrying out old behaviour, or new behaviours, relations or entities) which result in one or more of the following situations: (1) new possible market failures; (2) new risks to human health or safety, or the environment; (3) new risks to moral interests, rights, or values; (4) new threats to social solidarity; (5) new threats to democratic process; (6) new threats directly to the coherence, efficacy or integrity of the existing regulatory ecosystem charged with mitigating the prior risks (1-5). All this is not to say that these rationales apply in the same way in all specific contexts. Indeed, they will be weighted differently across distinct legal systems and jurisdictions-and between domestic and international law. Nonetheless, they provide a rubric for understanding when or why we (should) want to regulate a new AI applicationand a reminder that it is the sociotechnical changes, not the appearance of new technology in itself, that we are concerned about. \n AI as regulatory target Along with providing a greater grounding for understanding whether, when and why to regulate new AI applications, a consideration of sociotechnical change can also shed light on the regulatory 'texture' of the underlying AI capabilities-that is, its constitution as a 'regulatory target'  (Buiten 2019, 46-48) . That is, once regulators have confirmed a regulatory rationale (i.e. they have asked 'do we need regulation? For what sociotechnical change? What regulatory rationale?'), they then face the question of how to craft regulatory actions. In considering AI applications as a target for regulation, a sociotechnical change-centric perspective must on the one hand take stock of the material aspects of a technology (as an artefact). Material features certainly matter from the perspective of understanding key parameters for regulation, such as: (1) Its trajectory and distribution: i.e. the state of leading AI capabilities (across its different sub-fields), possible and plausible rates and directions of progress given material constraints on the design space and process  (Verbruggen 2020)  affect stakeholder perceptions of the imminence of various applications of the technology, and of the need for urgent regulation (Crootof 2019; Watts 2015) (4) Potential sites or vectors for regulatory leverage: for instance, the degree to which proliferation of certain systems could be meaningfully halted through export control policies  (Brundage et al. 2020; Fischer et al. 2021) . In these ways, such material features certainly matter, especially when considering AI regulation at the global level. For instance, scholars have argued that the modern global digital economy, far from consisting solely of ethereal digital products ungraspable by law, is instead populated by distinct 'regulatory objects', which vary in their degree of apparent 'materiality' (from high-capital submarine cables and satellite launch facilities, to ethereal cloud services), and their degree of centralization (from diverse suppliers of various 'smart' appliances, to dominant social networks or computationally intensive search engine algorithms)  (Beaumier et al. 2020) . Critically, some of these may not easily be subjected to global regulation, but many which can certainly be captured by various regulatory approaches. However, for regulatory purposes, a material analysis is not sufficient. A sociotechnical change-centric perspective on AI regulation rather can and should go beyond the technology itself, and consider a broader set of 'problem logics' in play. For instance, we can fruitfully distinguish between: 'ethical challenges', 'security threats', 'safety risks', 'structural shifts', 'common benefits', and 'governance disruption'. Distinguishing amongst these ideal-types is valuable, as these clusters can introduce distinct problem logics, and foreground distinct regulatory logics or levers (see Table  1 ).  1  It is important to note that this taxonomy is not meant to be mutually exclusive, nor exhaustive. It aims to capture certain regularities which help ask productive regulatory questions. For each category, we can ask-how does the AI capability produce sociotechnical change? Why does this create a governance rationale? How should this be approached as governance target? What are the barriers, and what regulatory tools are foregrounded? There is insufficient space to go into exhaustive detail on each of these categories within the taxonomy. However, at glance, we can pick out a number of ways in which clustering AI's sociotechnical impacts along these various problem logics, can facilitate structural regulation-relevant insights for AI regulators. These include consideration of aspects. For one, this model enables examination of the underlying origins of the sociotechnical challenge of concern, in terms of: (1) the key actors (e.g. principals, operators, malicious users) whose newly AI-enabled or -related behaviour or decisions create the governance concerns, and (2) those actors' traits, interests, or motives which drive the AI-sociotechnical-problem related behaviour or decisions (e.g. actor apathy, malice, negligence, or the way the new capability sculpts choice architectures in ways shift structural incentives or strategic pressures).  \n Safety risks Can we rely onand control this? • New risks to human health or safety  \n Structural shifts How does this shape our decisions? • (all, indirectly)  \n Governance Disruption How does this change how we regulate? • New risks directly to existing regulatory order This in turn can be linked to the contributing factors which sustain or exacerbate this sociotechnical impact, such as: (1) The range and diversity of AI-related failure modes and issue groups  (Hernandez-Orallo et al. 2020, 7) , including emergent interactions with other actors (human or algorithmic) in their environment  (Rahwan et al. 2019) , peculiar behavioural failure modes  (Amodei et al. 2016; Krakovna et al. 2020; Kumar et al. 2019; Leike et al. 2017) . Moreover, this model enables a study of the barriers to regulation; that is, the factors that drive the difficulty of formulating or implementing policy solutions, and which will themselves have to be overcome to achieve effective regulatory responses for AI, because of: (5) the live societal or cross-cultural value pluralism  (Gabriel 2020)  or disagreement over the values, interests or rights affected, and how these should be weighted in the context of a specific contested AI application (ethical challenges); (6) the disproportionately high costs of 'patching' vulnerabilities of human social systems (e.g. our faith in the fidelity of human voices) against AI-enabled social engineering attacks, relative to past costs of patching 'conventional' cybersecurity vulnerabilities by the dissemination of software fixes  (Shevlane and Dafoe 2020b, 177) . (security threats) (7) the difficulty of foreseeing indirect effects of AI on the structure of different actors' choice architectures (van der Loeff et al. 2019; Zwetsloot and Dafoe 2019)-and the difficulty of resolving those situations through any one actor's unilateral action  (Dafoe 2020) , or to coordinate behaviour in response (structural shifts) Finally, on the basis of the above, it allows a consideration of the types of regulatory approaches and levers that are highlighted and foregrounded for each of these challenges. It highlights the role of 'mend-it-or-end-it' debates around algorithmic accountability (Pasquale 2019), auditing frameworks  (Raji et al. 2020) , and underlying cross-cultural cooperation  (ÓhÉigeartaigh et al. 2020)  to diverse ethical challenges. Of perpetratorfocused and target-focused (e.g. 'security mindset' (Severance 2016)) interventions to shield against AI security threats. The development of support programs to guarantee public goods such as the use of AI in 'AI for Good' interventions  (Floridi et al. 2018; ITU 2019) , humanitarian uses  (Roff 2018, 25 ; but see Sapignoli 2021), or redistributive Electronic copy available at: https://ssrn.com/abstract=<PHONE> guarantees such as a 'Windfall clause' that sees tech companies pledge extreme future profits above a certain threshold towards redistribution (O'  Keefe et al. 2020) . That is not to say that this framework provides conclusive recipes or roadmaps for regulation. Rather, it provides a beginning structuring framework for thinking through common challenges across diverse regulatory domains charged with resolving questions around seemingly separate applications of AI  (Crootof and Ard 2021) . Such an approach can at least avoid duplication of effort, and at best can support the formulation and spread of better, more resilient policies. In sum, a sociotechnical-change-centric approach is not without its pitfalls or limits. Still, it can have various benefits in organizing and orienting an AI regulatory ecosystem. It prompts regulators to ask themselves: (1) when, why and how a given AI application produces particular types of sociotechnical changes; (2) When and why these changes rise to create a rationale for governance; (3) How to approach the target of regulation. As such, this can be an important regulatory complement to the insights provided by-and the interventions grounded in-technology-centric or law-centric perspectives. \n Implementation: AI regulation through a sociotechnical lens This lens of sociotechnical change does not provide single substantive answers for how to resolve each and every AI policy problem. However, it can help answer common recurring questions in AI policy around institutional choice and regulatory timing and design (Bennett  Moses 2017, 585-91) . In particular, regulatory actors can improve governance for AI challenges in terms of regulatory triage, tailoring, timing and responsiveness, and design. \n Regulatory triage In the first place, the sociotechnical-change-centric perspective on AI can help in carrying out regulatory triage. This is not just of value to AI regulation: indeed, it fits in with a broader initiative in recent legal scholarship towards exploring questions of 'legal prioritization'  (Winter et al. 2021) . However, within the AI regulation ecosystem, this lens helps focus attention on the most societally disruptive impacts of the technology, and as such helps re-focus scarce regulatory attention. This reduces the risk that regulatory attention is over-allocated on visceral applications of AI which may not ultimately prove scalable, or on 'legally interesting puzzles', at a cost of more opaque but prevalent indirect impacts. Better triage can be a valuable corrective to approaches that select, organize or prioritize AI policy issues based on high-profile but non-representative incidents, popularcultural resonance, or 'fit' to pre-existing legal domains. Moreover, if regulatory bodies focus less on the 'newness' of AI technology, or on the steady stream of each new AI application, but rather on which downstream sociotechnical impacts in fact create particular regulatory rationales, they can step back from a reactive firefighting mode, and help defuse or dissolve the so-called 'pacing problem'  (Marchant 2011) . Regulatory triage is also aided by the ways in which this framework can expand regulatory actors' scope of analysis of which sociotechnical impacts are relevant for regulatory consideration. While technology-centric approaches can highlight the direct challenges of AI (in the areas of ethics, security, and safety), the sociotechnical-change-centric perspective also allows regulators to consider interventions for various indirect sociotechnical changes, including the ways AI systems can shift incentive structures, how to realize beneficial opportunities and public goods around AI technology, or how AI applications can disrupt the regulatory tools or systems which these regulators would rely on. What does that entail in practice? Improving triage around AI regulation could involve (1) improving information infrastructures or 'technical observatories'  (Clark [forthcoming]  (in this volume)), to not only equip regulators with relevant and up-to-date technical information around AI techniques and applications, but also think through how these relate to downstream sociotechnical impacts. This may help ensure regulators are less easily dazzled by the 'newness' of new AI applications themselves  (Mandel 2017) , and enable them to become more aware of how different analogies can highlight different regulatory narratives in potentially counterproductive ways  (Crootof and Ard 2021) . This can also involve (2) setting up a cross-ecosystem agency to focus on 'legal foresighting'  (Laurie, Harmon, and Arzuaga 2012) , and forecasting methodologies  (Avin 2019; Ballard and Calo 2019)  aimed at eliciting AI's technologies' disparate sociotechnical impacts, link these to potential and actual regulatory rationales, and study the shifting material textures and problem logics around that application. In particular, this can support more democratic and inclusive stakeholder debate over the choices affected parties would seek to make around the deployment of potentially disruptive AI breakthrough capabilities (Cremer and Whittlestone 2021). \n Regulatory tailoring and scope Secondly, and relatedly, the sociotechnical change-centric lens helps in tailoring regulatory solutions to effective clusters of AI techniques, applications, -users, and societal effects. Rather than consign regulators to confront self-similar AI challenges (e.g. around meaningful human control; susceptibility to adversarial attack; unaccountable opacity in algorithmic decision-making) many times across individual legal domains,  (Crootof and Ard 2021, 1) , this approach highlights common themes, underlying material value chains, or usage problem logics of AI, as they are expressed in these various domains. Practically, improved regulatory tailoring can require (3) the establishment of various institutions and meta-regulatory oversight mechanisms-connoting \"activities occurring in a wider regulatory space, under the auspices of a variety of institutions, including the state, the private sector and public interest groups [which] may operate in concert or independently\"  (Grabosky 2017, 150) . In so doing, such mechanisms could foster improved cross-regime dialogue of AI policy  (Cihon, Maas, and Kemp 2020b) . This can support regulatory harmonization or the bundling of regulatory interventions for various AI applications where appropriate. It can also examine how and where different regimes and institutions can exploit the same regulatory levers (e.g. compute hardware production) that intersect on the AI development value chain. (4) Establishing mechanisms and fora for dialogue amongst various actors in the AI space that may have a hand in shaping the overarching 'problem logic'-in terms of the problem's origins, contributing factors, or barriers to regulation. The aim of such discussions would ideally be to reconfigure some of these wider conditions to be more supportive or conducive to AI regulation, or-where they are not very tractable, to explore alternative levers or vectors for regulation. Finally, it can promote the exchange of best practices or lessons learned around how regulators can address some of the problem logics that generate the problem or impede regulation. \n Regulatory timing Thirdly, in terms of regulatory timing and responsiveness, a study of AI's sociotechnical changes highlights the inadequacies of governance strategies that are grounded either in an attempt to predict sociotechnical changes in detail, or reactive responses which prefer to 'wait out' technological change until its societal impact has become clear-which demands a threshold of clarity that is in fact rarely achieved, even decades after a technology's deployment  (Horowitz 2020) . Rather, it emphasises the importance of anticipatory and adaptive regulatory approaches  (Maas 2019b ). This helps mitigate some of the information problems facing AI regulation, by helping ensure AI regulation can remain adaptive and 'scalable' to ongoing sociotechnical change, given the profound lack of information about future pathways. This could be pursued by (  5 ) incorporating provisions such as sunset clauses that prompt re-examination (at the domestic level) or designating authoritative interpreters (at the international level). \n Regulatory design Fourthly, in terms of regulatory design, the sociotechnical change lens highlights when and why governance should prefer technology-neutral rules versus technology-specific rules. By considering the specific regulatory or governance rationale in play, we may understand when or whether technological neutrality is to be preferred. Generally, Bennett-  Moses (2017, 586)  argues that \"regulatory regimes should be technology-neutral to the extent that the regulatory rationale is similarly neutral\". In this view, the point is not, to find a regulatory strategy that already details long lists of anticipated future applications of AI. The idea is rather to develop institutional mechanisms that are up to the task of managing distinct problem logics-new ethical challenges, security threats, safety risks, structural shifts, opportunities for benefit, or governance disruptions-in a way that can be relatively transferable across-or agnostic to the specific AI techniques used to achieve those affects. Establishing clearer guidelines about formulation of AI-specific regulations, and the circumstances in which these should rely on standards or rules, and when they should be tech-specific or tech-neutral  (Crootof and Ard 2021) . \n Conclusion Electronic copy available at: https://ssrn.com/abstract=<PHONE> This chapter has introduced, articulated, and evaluated a 'sociotechnical change-centric' perspective on aligning AI regulation. It first briefly sketched the general value of a 'change-centric' approach to the problems facing the AI governance ecosystem. The chapter next articulated a framework focused on Lyria Bennett Moses's account of regulation for sociotechnical change. It explored how, when, and why law and regulation for AI ought to tailor themselves to broad sociotechnical change rather than local technological change. It applied this model to AI technology, in order to show how this model allows a better connection of AI applications to five types of sociotechnical change-and how these in turn can be mapped to six types of regulatory rationales. It then turned to the mirror question of how, having established a need for governance, regulators might craft policy interventions to the particular regulatory target of AI. This involved a consideration of both the material textures of AI applications, but especially demands focus on the 'problem logics' involved. It argued that socio-technical changes created by AI applications can be disambiguated into six specific types of challenges-ethical challenges, security threats, safety risks, structural shifts, public goods, and governance disruption-which come with distinct problem features (origins, contributory factors, barriers to regulation), and which may each be susceptible to (or demand) different governance responses. Finally, the chapter concluded by reflecting on the limits and uses of this approach, before sketching some indicative institutional and regulatory actions that might draw on this framework to improve regulatory triage, tailoring, timing & responsiveness, and regulatory design of AI policy. To be clear, an emphasis on sociotechnical change is not a new insight in scholarship on law, regulation and new technology. However, in a fragmented and incipient AI governance landscape, it remains a valuable tool. In sum, 'sociotechnical change' should be considered not a new or substitute paradigm for AI governance, but rather a complementary perspective. Such a lens is subject to its own conditions and limits, but when used cautiously, can offer regulators a more considered understanding of which of AI's challenges are possible, plausible, or already-pervasive-and how these might be best met. Notes 1 An earlier version of this framework is presented and unpacked in further detail in  (Maas 2020, 166-86) . Note, this version referred to the 'public goods' as the 'common goods' problem logic, instead. (safety risks) (2) Human overtrust and automation bias, rendering some AI systems susceptible to emergent and cascading 'normal accidents' (Maas 2018) (safety risks) (3) The underlying 'offense-defense balance' of AI scientific research (Shevlane and Dafoe 2020a), and how it evolves along with more sophisticated AI capabilities (Garfinkel and Dafoe 2019) (security threats). (4) The susceptibility of existing legal and regulatory systems themselves to 'disruption' by AI uses, at the level of doctrinal substance, law-making and enforcement processes, or the political foundations (Liu et al. 2020; Maas 2019c) (governance disruption). \n Table 1 . 1 Taxonomy of AI problem logics • AI systems • O. Push towards legal • Provisions to render creating efficiency governance 'innovation- substantive • CF. Legal system proof': technological ambiguity in law exposure and neutrality; authoritative • Legal automation dependence on interpreters, sunset altering processes conceptual orders or clauses; etc. … of law operational assumptions • Oversight for legal • Erodes political automation; distribution foundations Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t Electronic copy available at: https://ssrn.com/abstract=<PHONE>", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/SSRN-id3871635.tei.xml", "id": "a726b40ee5ef3430da9d7bedf45b08ca"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "China and Russia have declared 2020 and 2021 as years of scientific and technological innovation cooperation, focusing on biotech, artificial intelligence, and robotics. 1 Both countries view AI as critical to their respective domestic and foreign policy objectives, and they are ramping up investments in AI-related research and development, though China's investments far outweigh Russia's. U.S. observers are watching this convergence between America's two key competitors with increasing concern, if not alarm. Some worry that alignment between Beijing and Moscow, especially in the areas of science and technology, could accelerate the development of surveillance tools to enhance authoritarian control of domestic populations. Others warn that deepening Sino-Russian cooperation will dilute the effects of sanctions on Russia. Still others fear that the strengthening partnership between China and Russia will undermine U.S. strategic interests and those of its democratic allies in Europe and Asia.  2  Chinese and Russian sources are keen to publicize their \"comprehensive strategic partnership of coordination for the new era,\" potentially underscoring the seriousness of their joint ambitions.  3  Yet the scale and scope of this emerging partnership deserve closer scrutiny, particularly in the field of AI. To what extent are China and Russia following up on their declared intentions to foster joint research, development, and commercialization of AI-related technologies? In other words, how do we separate headlines from trend lines? This issue brief analyzes the scope of cooperation and relative trends between China and Russia in two key metrics of AI development: research publications and investment. Our key findings are as follows: Scope:", "authors": ["Margarita Konaev", "Andrew Imbrie", "Ryan Fedasiuk", "Emily Weinstein", "Katerina Sedova", "James Dunham"], "title": "Headline or Trend Line?  Evaluating Chinese-Russian Collaboration in AI", "text": "between the United States and Russia declined in 2019, which may reflect the impact of sanctions. While the number of U.S.-China AI-related research publications continued to increase through 2019, given the heightened tensions, data from 2020 and 2021 could show different trends. • Investment:  China  Our findings both confirm assessments of the expanding partnership between China and Russia and add an important caveat with regards to its scope and limitations. There has been a steady increase in AI-related research collaboration between the two nations and an even steeper rise since 2016. This upward trend mirrors the global expansion of AI research, propelled by increased computing power and the availability of large datasets. The overall number of joint Chinese-Russian AI-related publications, however, remains relatively low-whether as a share of each country's scholarly output or compared with the number of \n Introduction Shared antipathy toward the United States and convergent political, security, and economic interests are propelling what the U.S. intelligence community described in its most recent annual Threat Assessment as \"increasing cooperation\" between China and Russia.  4  Signs of a burgeoning partnership are multiplying. The two countries conduct joint training and military exercises and sell weapons to one another. Collaboration extends to space, where China and Russia plan to build a new lunar scientific research station. Both countries vote in the United Nations Security Council to shield one another from international scrutiny of human rights abuses. To deepen economic ties, Russia furnishes specialized expertise in such fields as mathematics and computer science and supplies raw materials in the form of oil, gas, and precious metals; China provides capital, resources, and access to its large market. Beijing and Moscow also exchange talent, knowhow, and investments to spur innovation, monitor and control their populations, and extend their influence globally.  5  In a recent editorial, China's state-owned Global Times proclaimed that \"China-Russia cooperation has no upper limits.\" 6 Given the long, fraught, and at times contentious relationship between China and Russia, such declarations are ahistorical at best. Memories of the Sino-Soviet border clash in 1969 are distant but still keenly felt in some quarters. The asymmetry in power between Beijing and Moscow will likely grow in the coming years, as China's economy surges ahead.  7  Cultural disparities, protectionist instincts, and lingering distrust continue to hamper the relationship, despite myriad intergovernmental contacts and personal ties between Chinese president Xi Jinping and Russian president Vladimir Putin.  8  Indeed, concerns over the theft of intellectual property, industrial espionage, and persistent cyber threats will set upper limits to Sino-Russian cooperation, notwithstanding public signaling to the contrary.  9  Perhaps nowhere is the test case for Sino-Russian cooperation more consequential or more speculative in its claims than in artificial intelligence. President Xi sees AI as a \"strategic handhold for China to gain the initiative in global science and technology competition,\" and the Chinese government reportedly spends billions each year on AI research.  10  Though Russia lags behind China in AI capability, President Putin grasps its importance for Russia's economic and strategic future.  11  In October 2019, Russia released its national AI strategy. While the strategy's details and enforcement mechanisms remain unclear, the Russian government aims to boost research and development (R&D) efforts and overcome barriers related to data, talent, and computing power. It has since allocated more than $200 million in AI subsidies through government-backed R&D centers for 2021-2023.  12  When analyzing the evolving AI partnership between China and Russia, it is important to distinguish the signals from the noise. Both governments have trumpeted a series of high-tech initiatives, joint research centers, technology parks, and cooperative agreements.  13  Based on public statements, the Russian and Chinese Academies of Sciences collaborate on AI-related research. Chinese and Russian companies are stepping up cooperation on joint development of AI products, such as autonomous vehicles, medical robotics, and facial recognition software. While some of these initiatives seem to be bearing fruit, many have yet to move past the exploratory or even the declaratory phase. The mutuality of Chinese and Russian interests shapes and constrains their partnership in AI. Even as both countries invest in emerging technologies to advance their domestic and foreign policy goals, they also seek to preserve sovereignty, limit the flow of data across their borders, and protect national industries.  14  The prominent role afforded to Chinese and Russian state-owned enterprises (SOE) in the execution of national AI priorities may further complicate the prospects for cooperation.  15  The ability of China and Russia to balance these competing imperatives will determine whether their rhetoric matches reality in the years to come. This issue brief presents an assessment of Chinese and Russian collaboration in two key areas for advancement in AI: research output and investment. Recent research from CSET shows that despite the rapid growth of Chinese research publications over the past 20 years, China's rate of international collaboration has remained flat. Countries such as the United States and Australia, on the other hand, have seen a significant rise in their rates of international collaboration.  24  As China's research output continues to climb, its international collaboration rate may increase as well. Alternatively, China may be following a deliberate strategy designed to limit collaboration in certain areas until the country has achieved a strong international position.  25  These broader patterns provide the context for assessing Chinese-Russian research collaboration in AI-related fields, as well as trends in collaboration between U.S. researchers and their Chinese and Russian counterparts. \n Methodology and Scope The following analysis draws on publications data covering collaborative Chinese-Russian AI-related research papers published between 2010 and 2019 in English and Chinese. For English-language publications, we identified collaborations between Chinese and Russian scholars by analyzing the institutional affiliations they reported in publications and preprints. If the affiliations of a paper's authors included an organization in each country, we included the paper in our analysis. We searched for such collaborations across three sources of scientific literature: Microsoft Academic Graph, Digital Science Dimensions, and Clarivate Web of Science.  26  For publications in Chinese, we relied on the China National Knowledge Infrastructure.  27  Because CNKI does not sort by country of publication, we instead identified Chinese-Russian collaborations by searching for organizations with \"Russia\" and a number of major Russian cities in their name, since CNKI often includes the country or city name in Chinese for a foreign institution. This report discusses joint Chinese-language publications separately from joint English-language publications because the two types of collaborations seem substantively different: the former is a joint research endeavor conducted in a language that is native to one side of the co-authorship while the latter is a research collaboration undertaken in a language foreign to both. As we discuss later in the report, English-language publications data and CNKI data on Chinese language publications also show different trends in Chinese-Russian collaborations. To restrict our analysis to the development and application of AI, we applied the same classification method used in an earlier CSET report, \"Russian AI Research 2010-2018.\"  28  Rather than developing keyword-based queries or explicit criteria for AIrelevant research, we inferred a functional definition from papers in the arXiv.org repository, training a model to predict the categories that authors and editors assign to the site's papers. Applying it to our broader corpus, we identified as AI-relevant the papers that would be likely to receive an AI-relevant categorization if uploaded to arXiv.org.  29  To complement this methodology, we attempted to collect data on Chinese-Russian AI-relevant research publications in Russian. We developed a set of Russian-language keywords that included a broad range of research approaches, models, techniques, applications, systems, and functions related to AI, and executed a query across Microsoft Academic Graph, Digital Science Dimensions, and Clarivate Web of Science. Our search, however, surfaced only one AI-relevant research paper published jointly by Chinese and Russian authors in Russian between 2010 and 2019. As such, we limit our discussion to collaborative AI-related publications in English and Chinese. Finally, the results from the analysis of the scientific literature are current as of February 19, 2021. Analyses conducted at different points in time will yield slightly different results because the aforementioned sources of scientific literature acquire publications data after some delay. For example, articles published in the second half of 2019 may not appear in the data until 2020. We focus our analysis on the period between 2010 and 2019 to minimize the effect of this phenomenon. \n Chinese-Russian AI-Related, English-Language Research Collaboration: Output, Trends, and Institutions Between 2010 and 2019, we identified 296 English-language, AIrelated research papers co-authored by at least one Chinese and at least one Russian author. This figure accounts for about 2 percent of total Russian AI-related publications and less than 0.1 percent of Chinese AI-related publications for the same time period. In other words, Chinese-Russian collaboration in AI-related fields constitutes only a fraction of each country's overall research output in English-language journals. There are a number of explanations for this relatively low number of joint Chinese-Russian AI-related research publications. Most basically, research conducted and published in English requires both sides to work in a language other than their native tongue. Scientists from non-English speaking countries also generally have a harder time publishing in English-language journals. While China has recently outpaced the United States in the number of AI-related research papers, Russia is significantly behind due to a combination of factors, including lower spending on R&D, problems in developing, attracting, and retaining talent, and a range of bureaucratic hurdles to international research collaboration and publishing.  30  Although Russia has been implementing a series of reforms and initiatives, such as Project 5-100, designed to increase its scientific scholarly output in international journals, buttress international collaboration, and raise the global rankings of its top universities, the aforementioned challenges persist.  31  Finally, as previously noted, China and Russia have not traditionally been important research partners for one another; Russian scientists have predominantly collaborated with European counterparts, while Chinese scientists have primarily looked toward the United States for joint research opportunities. Despite the relatively low overall number of publications produced jointly by Chinese and Russian researchers, as Figure  1  shows, there has been an increase in collaborations over the past decade, and especially since 2016. More specifically, in 2019, there were 14 times more AI-related publications co-authored by Chinese and Russian researchers than there were in 2010. In terms of key areas of focus, Chinese and Russian researchers have produced joint papers in pattern recognition, algorithm development, computer vision, machine learning, remote sensing, data mining, control engineering, and natural language processing. * Table  1  shows the top 10 AI-related research fields for joint Chinese-Russian publications, as well as Chinese and Russian research output individually. papers; and linguistics, which is on the list of Russian papers but not in the top 10 areas for collaborative research.  32  These areas are also similar to the top areas of focus for Chinese researchers, except for remote sensing and natural language processing, which appear among the top areas for collaborative research but not the top areas on the Chinese publications list, and mathematical optimization and simulation, which appear in the top 10 areas for Chinese publications but not on the collaborative list. Natural language processing is a joint area of research and a top area of Russian research but not a top 10 area for Chinese researchers. Remote sensing is the one area where Chinese and Russian researchers have published together that does not appear in either China's or Russia's top 10 areas for AI-related publications individually. As a whole, countries with more satellites tend to produce more articles on remote sensing.  33  The United States and China, for instance, operate the most satellites in orbit and were ranked first and second, respectively, in the number of remote sensing publications according to data from 1991 to 2015.  34  While Russia has the third largest arsenal of satellites in the world, it has been far less productive than the United States or China in terms of remote sensing research publications. The collaboration in AIrelated remote sensing research between China and Russia could also be an artifact of shared borders and, in turn, shared interest in land cover and use, vegetation, and climate change in these border regions. In addition to trends and topics, we also identified the key institutions that Chinese and Russian researchers listed as their affiliation on this set of 296 collaborative AI-related publications. Table  2  presents the top 10 of the most prolific collaborative institution pairs by publication count. As Table  2  illustrates, collaborations between authors listing an affiliation to the Chinese Academy of Sciences and the Russian Academy of Sciences have yielded the highest number of AIrelated joint research papers between the years 2010-2019. The Russian Academy of Sciences oversees 550 scientific institutions and research centers and reportedly employs more than 55,000 researchers.  36  Previous CSET research shows that RAS also produces the most AI-related publications in Russia.  37  The Chinese Academy of Sciences oversees 12 branches, more than 100 research institutions, three universities, more than 130 state key laboratories, and over 270 research observation stations in the field. It houses more than 71,000 formal employees as well as over 64,000 postgraduate researchers.  38  Given the number of researchers and research institutions affiliated with both CAS and RAS and considering that researchers often have more than one affiliation, there is no straightforward way to track the formal collaborative agreements between the two academies that led to the AI-related publications we identified-if such formal agreements even exist in the first place. It is, however, notable that CAS and RAS have signed a 2018 agreement to strengthen science and technology cooperation in the fields of physics, astrophysics, chemistry, biological sciences, nanotechnology, medicine, and agriculture, as well as a 2019 agreement to deepen cooperation on polar research, laser science, deep-sea research, space science, geophysics, the ecological environment, neuroscience, and more.  39  While these agreements do not directly reference collaboration on AI, research related to AI can be conducted in each of the aforementioned disciplines and fields. Another noteworthy relationship that has led to joint AI-related research publications is between ITMO University and Hangzhou Dianzi University. The partnership between these two institutions includes student and staff visits and a joint summer school that also features Alibaba Cloud Academy courses.  40    \n Huawei's AI-Related Activities and Partnerships in Russia Huawei has been operating in Russia since 1996, setting up its first local joint venture with Beto Konzern and Russia Telecom in 1997.  43  Despite initially struggling to establish itself, by 2003, Huawei supplied more than 50 percent of the broadband products market in Russia.  44  Over the past few years, in response to growing tensions between the United States and China, and especially following its addition to the U.S. Department of Commerce's Entity List in 2019, Huawei has been expanding its presence in Russia.  45  The Russian government seems to welcome this expansion, arguing that Huawei's training programs, local research centers, and promises to jointly develop or share emerging technologies will help Russia avoid a brain drain and promote local innovation.  46  In   53  While seemingly separate from the joint lab, the Skoltech Laboratory for Quantum Information Processing also claims to have ongoing multi-year collaboration agreements with  Huawei. 54  In addition to its partnerships with Russian universities, Huawei touts joint projects with Russian Academy of Sciences institutions, including the Institute for Information Transmission Problems (IITP), the Institute for System Programming, and the Marchuk Institute of Numerical Mathematics.  55  The collaboration with RAS IITP seems to be focused on 5G, including \"designing and studying innovative algorithms for cellular networks\" that improve network efficiency.  56  Huawei is also one of the international tech companies to establish a presence in Siberia's Akademgorodok-an \"academic town\" known as one of the key science and technology hubs in Russia and home to Novosibirsk State University and the Siberian Branch of RAS.  57  In 2020, Novosibirsk State University announced that the three RAS Institutes of the Siberian Branch-the Sobolev Institute of Mathematics, the Institute of Computational Mathematics and Mathematical Geophysics, and the A.P. Ershov Institute of Informatics Systems-signed a memorandum of understanding with Huawei-Russia and the Mathematical Center in Akademgorodok. The agreement aims to promote cooperation in science, research, and education in the areas of math, AI, big data, and high-performance computing.  58  Huawei's ties to this region reflect the shifting focus to technology innovation in the mid-2000s. Since the Russian government began reinvesting in Akademgorodok and the region's network of academic and research institutions, international tech companies have also sought to expand their presence and capitalize on the high concentration on talent, relative intellectual freedom, and openness to international collaboration. Huawei's partnerships with Russian universities and research institutes are likely to bring an influx of funds to these institutions. Russian science is chronically underfunded by the state, while research and development investments from the business sector are even scarcer. But with funding comes influence.  59  It is therefore possible that increased R&D investment from Huawei and potentially other Chinese companies would allow them to set the agenda for the partnerships with Russian universities and research institutions, and leverage the outputs of collaborative research to advance China's broader goals in AI. 60 \n Chinese-Russian AI-Related Publications in Chinese In addition to the English-language papers, our search of CNKI also identified 43 AI-relevant papers co-authored by Chinese and Russian researchers and published in Chinese between 2005 and 2020. Figure  2  shows the annual distribution of these Chineselanguage papers. While the previously discussed English-language publications data shows a steady increase in collaborative research between Chinese and Russian scientists, the Chinese language publications data shows no consistent upward or downward trajectory, aside from the uptick in joint publications after 2018. With respect to key institutions involved in collaborative work, 16 papers (37 percent) were co-authored by researchers affiliated with the Russian Academy of Sciences.  61  Only three papers (7 percent) were written by researchers at the Chinese Academy of Sciences, in contrast to our English language data in which CAS is the top Chinese contributor. It is also worth noting that five papers were co-authored by Chinese researchers affiliated with organizations on the U.S.   62  Two of these five papers were written by the same co-authors from Beihang University and the Moscow Aviation Institute and focus on control engineering and modeling of pilot behavior.  63  Although these papers were both written in 2010, the two organizations appear to maintain close ties. In 2020, the schools launched a joint professional MA program in aeronautical engineering.  64  Several papers also include information about the agencies funding the research. Eleven of the 43 CNKI papers were funded by the National Natural Science Foundation of China (NSFC), which is subordinate to China's Ministry of Science and Technology.  65  This arrangement is not surprising, considering that NSFC is a massive foundation that funds more than 18,000 research projects each year and has a budget of $4.9 billion.  66  Three additional projects were funded by  China  When looking at trends over time, AI-related collaborations between Chinese and Russian scientists (Figure  1 ), U.S. and Chinese scientists (Figure  3 ), and U.S. and Russian scientists (Figure  4 ) have all increased since 2010, and especially since 2016. As previously noted, in 2019, there were 14 times more AIrelated publications co-authored by Chinese and Russian researchers than there were in 2010. This upward trajectory notwithstanding, the baseline number of joint publications for China and Russia was smaller than the number of papers coauthored by researchers from the United States and Russia and only a fraction of the number of joint papers produced by researchers from the United States and China. The United States and China publish significantly more AI-related research papers than Russia. It is therefore important to contextualize trends in collaborative output in terms of the broader publications landscape, both baseline numbers for each dyad and overall publication output for each of the three countries.    \n Chinese-Russian AI-Related Investments Since the turn of the century, the private sector has been the primary source of innovation in AI.  70  Governments that harness the fruits of this innovation are more likely to prosper economically and benefit strategically. The United States boasts a vibrant innovation ecosystem and the world's largest investment market in privately held AI companies. U.S. AI companies attracted 64 percent of disclosed investment in 2019, with China a distant second at 12.9 percent. Meanwhile, private Russian AI companies attracted only 0.04 percent of global disclosed investment value.  71  China's AI market declined over the 2018-2020 period but remains significantly ahead of the Russian AI market in terms of venture capital investments, deal counts, and transaction value.  72  To better understand the evolving partnership on AI between China and Russia, we analyzed activity in the commercial sector by measuring Chinese and Russian equity investments in privately held AI companies. As a baseline comparison, we also measured U.S. investment into private AI companies in China and Russia, and Chinese and Russian investment into private AI companies in the United States. \n Methodology and Scope Our estimates rely on prior CSET research into AI investment trends globally and information from Crunchbase, a leading private investment data aggregator.  73  We supplemented this approach by compiling popular lists of the top 25 Chinese and Russian AI startups. We performed keyword searches in English, Chinese, and Russian to track news coverage, press releases, and joint announcements of Sino-Russian investments in these startups. While this methodology offers a baseline for comparison, our findings confront several limitations. Crunchbase is an Englishlanguage service and likely undercounts Chinese-and Russianbound investments. Many Chinese and Russian investment deals are concluded under the auspices of state-owned enterprises and lack publicly disclosed values. Public data on Chinese and Russian AI investment flows are therefore incomplete and uneven. In addition, our formal analysis ends in 2020 and may not capture recent developments in the growing AI partnership between China and Russia. Despite these limitations, our findings provide a reasonable starting point for evaluating commercial AI activity between China and Russia. Through Crunchbase, we initially found just four Chinese AI-related investments in Russian firms and zero Russian investments in Chinese firms. We broadened our Crunchbase search results and used keywords to search manually for reporting about investments into any of the 50 top AI startups in China and Russia.  74  We discovered an additional eight investments in AI companies through these practices, yielding a total of  12     76  This baseline of U.S.-Chinese AI investment serves as a yardstick against which to measure other international partnerships. Although the level of AI investment between China and Russia seems to be increasing, it remains substantially lower than that of the United States and China. Although the total value of U.S.-Russian AI investment is much smaller than that of U.S.-Chinese AI investment, it is worth noting that the United States remains Russia's largest international destination of AI investment, with Russian investments amounting to $666 million of the total $670 million bilateral value. By comparison, over the same time period, Russian firms invested $555 million into Chinese private AI companies.  \n Evaluating Known Chinese-Russian AI Investments All investments into Chinese and Russian AI companies in our data set were concluded between 2016 and 2019. Taken collectively, they amounted to $879 million, with more than half of all value arriving from two large deals with China's Megvii and Russia's Mail.ru Group discussed below.  \n The Russia-China Investment Fund and Megvii Since its founding in 2011, Megvii has emerged as a central player in China's efforts to develop its homegrown AI market and lessen reliance on U.S. open-source frameworks. The company is also at the forefront of China's growing surveillance system. Its flagship product, Face++, is one of the world's largest open-source computer vision platforms and powers face scanning systems across a range of sectors.  78  These include phone and consumer Internet of Things (IoT) devices as well as Smart City and Smart Community management solutions for city governments to aid policing, monitor urban residential communities, and streamline logistics in more than 100 municipalities across China, encompassing 38 percent of the country, as detailed in the company's prospectus.  79  Megvii's role in providing facial recognition technology to the Chinese state has put it in the crosshairs of U.S. export controls. Megvii and Huawei have filed joint patents and tested a surveillance system with the capability to target ethnic minorities. Advertising for the system claimed that it features a camera with the capability of scanning faces in a crowd and estimating individuals' age, sex, and ethnicity. This camera could allegedly trigger a \"Uyghur alarm\" to government authorities when it identifies members of the oppressed Muslim minority group. In terms of ownership, the venture involves Alibaba, Mail.ru Group, the Russian telecommunications firm MegaFon, and the Russian Direct Investment Fund. Alibaba will invest $100 million and hold a 48 percent stake, with the remaining ownership divided among MegaFon (24 percent), Mail.ru (15 percent), and RDIF (12.9 percent).  88  The venture mirrors broader trends in e-commerce toward integration and consolidation of services.  89  AliExpress Russia aims to grow Russia's e-commerce market, solidify business ties between Russia and China, and expand e-commerce flows between the two countries. The extent to which this joint venture will realize such ambitions remains to be seen. For example, AliExpress is the 9th largest online commerce company in Russia, well behind market leaders.  90  Mail.ru is part of \"AI Alliance Russia,\" a cooperation agreement between RDIF, Russia's largest bank Sberbank, the Russian oil company Gazprom, the Russian internet company and search engine Yandex, and the Russian telecommunications group MTS.  91  Though short on details regarding organizational structure, resourcing, and enforcement, the AI Alliance seeks to develop Russia's AI market, accelerate development of AI products and services, and deepen linkages between the Russian business community and research organizations charged with implementing Russia's national AI strategy. AliExpress Russia aims to contribute to these goals by expanding the e-commerce market in Russia and increasing cross-border sales with China, laying the foundation for greater cooperation on AI products and services. While such deals may improve the commercial AI ecosystems of both countries and strengthen business ties between Russian and Chinese firms, the geopolitical implications in the near-term are unclear. Analysts will need to track the potential for large e-commerce deals to bolster the talent base and facilitate access to data in ways that could advance Sino-Russian progress in AI with concrete security and foreign policy implications. \n Other Sovereign Wealth Fund Investments with Unknown Value In addition to the 12 deals captured in Figure  7 , we identified five investment deals where Chinese or Russian firms had invested in AI companies, but for which we lacked information about the value of the arrangement. In four of these five cases, RCIF was the investor backing the deal.  92  Established in 2012 by both countries' sovereign wealth funds, RCIF's growing portfolio illustrates how the Sino-Russian relationship is tightening across multiple sectors. The fund today manages $2.5 billion in assets, having received $500 million from Saudi Arabia in 2017. Its founding signaled both nations' desire to strengthen cross-border investment activity, a trend that has extended to the AI industry. DiDi. In 2014, RCIF found its first AI-related investment in DiDi, a Chinese rideshare company modeled after Uber. Shortly after the infusion of capital, DiDi acquired its main homegrown competitor, Kuaidi, and bought Uber China, solidifying its dominant position in the Chinese ridesharing industry.  94     107  The collaboration includes joint research and development, testing and commercialization of products to support China's push for smart public transport, as well to modernize transport systems in Southeast Asia and Russia.  108  China's Dahua Technology and Russia's NtechLab. Dahua Technology and NtechLab, two leading providers of surveillance technologies, joined forces in 2019 to develop a wearable camera with facial recognition capabilities for law enforcement.  109  The product is the culmination of Dahua's two-year foray into the Russian market through its subsidiary, Dahua Technology Rus.  110  The camera incorporates Dahua's wearable video device and NTech's FindFace algorithm that can recognize emotions in addition to facial features, age, and gender, and process data on the device, enabling faster identification in near real-time.  111  China's Vinci Group and Russia's Jovi Technology. Deepening cooperation in the IT sector, the software development firm Vinci Group and the blockchain company Jovi Technologies partnered in 2019 to develop AI-enabled digital commerce solutions for small businesses.  112  The collaboration combines Vinci's expertise in encrypted mobile communications platforms with Jovi's focus on blockchain to develop integrated applications that advance AIenabled business solutions in both nations.  113   \n Conclusion An axis of convenience or deepening partnership? A Faustian bargain or durable entente?  114  The lexicon of Sino-Russian relations obscures as much as it clarifies.  Figure 1 : 1 Figure 1: Chinese-Russian AI-Related Research Publications in English, 2010-2019. \n Figure 2 : 2 Figure 2: Chinese-Russian AI-Related Research Publications in Chinese, 2005-2020. \n Figure 3 : 3 Figure 3: U.S.-Chinese AI-Related Research Publications, 2010-2019. \n Figure 4 : 4 Figure 4: U.S.-Russian AI-Related Research Publications, 2010-2019. \n Figure 5 : 5 Figure 5: Value of Known U.S.-China Private AI Investment, 2010-2020. \n Figure 6 : 6 Figure 6: Value of Known U.S.-Russia Private AI Investment, 2010-2020. \n Figure 7 : 7 Figure 7: Value of Known China-Russia Private and Sovereign Wealth Fund AI Investments, 2010-2020. \n Figure 8 : 8 Figure 8: Number of Investment Deals Disclosed in RCIF's Public Portfolio, 2013-2019. \n -Russia AI investment levels are higher than U.S.-Russia AI investment, but much lower than U.S.-China AI investments. Recent trends, however, show declining levels of U.S. AI investment in both China and Russia over the 2018-2020 period, while Russian-Chinese AI investment has increased relative to 2016 levels. U.S. investment in China declined from $13 billion in 2017 to $1.3 billion in 2020; meanwhile, the United States invested a negligible amount in Russia. China-Russia investment increased from $182 million dollars in 2016 to $517 million in 2017, but then declined to $100 million in 2018 and $80 million in 2019. While 2020 saw no disclosed AI investment between China and Russia, investment levels already topped $300 million by January 2021. \n by Chinese and Russian scientists between 2010 and 2019, as well as key topics, trends over time, and institutions leading these collaborative efforts. We present data on AI-related research papers published in English and Chinese, amounting to a comprehensive assessment of research output. For the analysis of investment ties between China and Russia, we examine activity in the commercial sector by measuring Chinese and Russian equity investments in privately held AI companies. We draw on leading databases of private-sector AI transactions to gather data on Chinese investments in Russian firms and Russian investments in Chinese firms. As a baseline for comparison, we also assess U.S.-Russia and U.S.-China collaboration in AI research and investments over time.These trends notwithstanding, our findings suggest that the extent and scope of Chinese-Russian collaboration in AI may be overstated by both Chinese and Russian sources as well as U.S. observers. AI-related publications co-authored by Chinese and Russian researchers constitute only a fraction of each country's overall AI-related research output in English-language journals. From a comparative standpoint, research collaboration in AIrelated fields between these two counties is also less frequent and productive than AI-related research collaboration between the United States and Russia and, especially, between the United States and China. Similar observations emerge from the analysis of investment data-between 2010 and 2020, the combined value of 17 Chinese universities do not have the same global reputation for excellence as U.S. institutions, but the Chinese government is pouring a massive amount of resources into STEM education and science and technology R&D, outpacing both government and private sector investments in Russia's academic and research institutions, in absolute as well as relative terms.The United States and China similarly dominate the private AI investment market. In 2019, the United States accounted for 64 percent of the total disclosed investment value, with China a distant second at 12.9 percent of the global AI investment market. Russia, meanwhile, has remained flat, with a peak of $3 million in These trends aside, China and Russia have not generally been 2018. 19 important scientific partners for one another. The aforementioned 95.5 percent increase in the number of co-authored publications In the remainder of this brief, we first assess the AI-related research ties between China and Russia and then examine AI-reflects the wider surge in Chinese research output and indicates a relatively low baseline for collaboration. Based on 2015 data on related investments between the two countries. While our findings indicate that AI-related collaboration between China and Russia in the number of joint publications in the Scopus dataset, for example, nearly half of Russia's joint publications were with research and investment remains relatively limited in scope, we offer preliminary indicators and areas to watch for growing scientists from the United States and Germany, while work with Chinese scientists accounted for less than 9 percent. For China, convergence that may undermine U.S. interests and those of U.S. democratic allies and partners. these patterns are even more stark: research collaborations with scientists in the United States represented 44 percent of China's total joint publications, while collaborations with Russian scientists Chinese-Russian Collaboration in AI-Related Research accounted for just 1.5 percent. 23 International research collaboration drives scientific progress and tends to benefit all parties by producing more impactful publications and expanding networks. According to 2018 National Science Foundation (NSF) data, for example, among the 15 largest producers of science and engineering scholarly articles, most have high rates of international collaboration. The United States has a collaboration rate of 39 percent, which is slightly below the average collaboration rate for the largest 15 producers. China's and Russia's collaboration rates are lower, at 22 percent and 23 percent, respectively. 20 While recent political developments and growing tensions with Russian researchers. Between 2010 and 2019, researchers from China have published 457,248 English-language AI-related papers, making China the world leader in AI-related research output. Researchers from the United States published 300,602 papers over the same period, coming in second worldwide. Russia, meanwhile, ranked 22nd in terms of overall research output, with 15,032 AI-related publications, trailing behind countries such as Iran, Malaysia, and Poland. Academic institutions publish the findings show mixed the United States may be pushing China and Russia closer, scholarly collaboration between these two countries has been results. growing steadily over the past decade. According to data from Elsevier's Scopus database, the number of co-authored Despite the worsening relationship between the United States and China, AI-related research collaboration continued to increase publications involving Chinese and Russian academics increased by 95.5 percent between 2013 and 2017. 21 Previously, another through 2019. China's AI investment into private AI companies in the United States has risen steadily from $181 million in 2014 to study showed that between the years 2006 and 2015, Russia was also one of China's top research collaborators among the One Belt $2 billion in 2020. Yet U.S. investment into private AI companies in China is declining, from a peak of $13 billion in 2017 to $1.3 One Road Initiative countries, second only to Singapore. The key areas for Chinese-Russian scholarly collaboration include physics, billion in 2020. When considering the U.S.-Russia dyad, collaborations between U.S. and Russian AI scholars dropped in chemistry, geosciences, and plant and animal sciences; in physics, the citation impact and the percentage of highly cited papers co-2019. Russian investment into private AI companies in the United million in 2020. U.S. investment into private AI companies in States has declined, from a peak of $393 million in 2018 to $164 authored by Chinese and Russian researchers was very high. 22 For our examination of Chinese-Russian AI research ties, we analyze the number of AI-related research papers authored jointly To an extent, our findings corroborate existing assessments of Chinese-Russian collaboration: the two nations are growing closer. Over the past decade, and especially since 2016, Chinese-Russian research collaboration in AI-related fields has steadily increased. Collaborations between authors from institutions affiliated with the Chinese Academy of Sciences and the Russian Academy of Sciences were among the most prolific joint research endeavors, while other notable research collaborations include those between Skolkovo Institute of Science and Technology and Hangzhou Dianzi University; Skolkovo Institute of Science and Technology and East China University of Science and Technology; and ITMO University and Hangzhou Dianzi University. The Chinese telecom giant Huawei has also been ramping up its AI-related activities and partnerships with Russian universities and research institutions. AI-related investment is also on the rise. Looking at 12 AI-related deals involving the two countries since 2016 and up to 2020, China has made $324 million in private AI investments into Russia, while Russia has made $555 million in private AI investments into China. Half of the value of overall Russian-Chinese AI investment is concentrated in two deals: the Russia-China Investment Fund (RCIF) investment in the Chinese AI company Megvii in 2017 and Alibaba's joint venture with the Russian internet company Mail.ru Group in 2019. In addition to the aforementioned 12 deals, Russia and China concluded five AI investment deals of undisclosed value. U.S. investment in private AI companies in China and Chinese investment in private AI companies in the United States amounted to $35 billion, dwarfing the $879 million worth of Chinese-Russian AI investment deals, which only trace back to 2016. 16  Notably, the $670 million overall value of U.S.-Russia AI investments is smaller than that of Chinese-Russian investment deals at $879 million. The United States, however, remains Russia's largest destination for AI investment.It is important to set these research and investment trends in their proper context. Tensions between the two countries aside, when accounting for funding levels and the volume and impact of published work, U.S. and Chinese AI researchers have far more incentives to cooperate with each other than to co-author with majority of AI-related research papers, and the United States is home to many of the world's leading universities. 18  By contrast, Russia comprises only 0.04 percent of disclosed AI equity investment globally. The size, value, and dynamism of private AI markets lend perspective to assessments of trends in the field. While the data covered in this report suggests a growing but relatively limited partnership on AI between China and Russia, the trends can shift markedly with the appearance of new research collaborations or large AI investment deals. Analysts should assess ongoing trends with respect to the health of the overall research ecosystem and private investment markets in AI, both at the national and international levels.It remains to be seen what impact sanctions, trade disputes, export controls, and growing tensions with the United States will have on Chinese-Russian collaboration in AI as well as on the research and investment ties between the United States and China and the United States and Russia. Thus far, our \n Table 1 : 1 Top AI-Related Research Fields: Chinese-Russian, Chinese, and Russian. Chinese- No. of Top Fields: No. of Top Fields: No. of Russian Top Papers China Papers Russia Papers Collab. Fields Pattern 29 Pattern 38,315 Linguistics 1,021 recognition recognition Computer 20 Computer 37,996 Computer 884 vision vision vision Algorithm 20 Algorithm 28,497 Algorithm 804 development development development Control theory 14 Data mining 21,311 Pattern recognition 674 Machine learning 14 Control theory 14,841 Machine learning 499 Remote sensing 11 Machine learning 14,400 Control theory 494 Data mining 9 Math. optimization 10,878 Data mining 474 Control 7 Control 8,266 Control 440 engineering engineering engineering NLP 6 Simulation 8,127 NLP 402 Other AI research \n Table 2 : 2 Top 10 Chinese-Russian Institutional Collaborations in AI-Related Research 2010-2019. 35   Russian Entity Chinese Entity No. of Papers Russian Academy of Chinese Academy of Sciences 15 Sciences Skolkovo Institute of Hangzhou Dianzi University 12 Science and Technology Skolkovo Institute of East China University of Science and 12 Science and Technology Technology ITMO University Hangzhou Dianzi University 11 Kazan Federal University Shanghai Jiao Tong University 10 Moscow State University Zhejiang University 9 Don State Technical Beijing Jiaotong University 9 University National Research Chinese Academy of Sciences 9 Centre -Kurchatov Institute Russian Academy of China Aero Geophysical Survey and 8 Sciences Remote Sensing for Land and Resources Skolkovo Institute of Science and Technology Guangdong University of Technology 8 Source: Microsoft Academic Graph, Digital Science Dimensions, Clarivate Web of Science, and CNKI. \n 2001, Huawei opened its first research training program inRussia in collaboration with the Moscow Technical University of Communications and Informatics, establishing its first R&D center in Moscow the following year. Today, Huawei reportedly maintains eight R&D centers across Russia-five in Moscow, and one in each of Saint Petersburg, Nizhny Novgorod, and Novosibirsk.47  These R&D centers appear to be hiring locally, as the majority of job postings are in Russian, not Chinese.48    According to the Huawei website, the company has 15 Russian partner universities.49  The main areas for cooperation with these and other universities include: development of algorithms for 5G communication networks; development of algorithms for image, video and sound processing; development of algorithms to support the evolving direction of the Internet of Things; development of artificial intelligence algorithms; big data processing algorithms; and distributed computing and data storage algorithms.50    One of the pathways to collaboration with universities is through the establishment of joint labs. In 2020, the Moscow Institute of Physics and Technology established a joint laboratory with Huawei Russian Research Institute, focusing on \"research and development in the field of artificial intelligence and deep learning.\" 51  The Skoltech-Huawei Innovation Joint Lab is another example, which was set up in 2018 to focus on \"artificial intelligence including machine learning, neuron networks, computer vision, language processing, and recommendation systems.\" 52  The Skoltech Master of Science in Internet of Things and Wireless Technologies is also associated with the joint Skoltech-Huawei lab. \n Department of Commerce's Entity List, including Beihang University, Harbin Institute of Technology (HIT), Harbin Engineering University (HEU), and Beijing University of Posts and Telecommunications. Beihang, HIT, and HEU are three of China's Seven Sons of National Defense, and Beijing University of Posts and Telecommunications was deemed by the U.S. government to directly participate in advanced weapons and systems research for China's People's Liberation Army (PLA). \n . and Russian researchers, with notable growth between 2016 and 2018, and a drop in 2019, perhaps signifying the impact of sanctions or the deterioration of relations between the two countries. Bibliometric data is not a perfect indicator of collaborative research-it does not capture classified research or joint research that does not yield publications in journals and conference proceedings. That said, the culture of AI research is generally open and collaborative, with researchers often sharing their data, algorithms, and results on open web platforms such as GitHub and arXiv.org. Our analysis, therefore, offers a comprehensive assessment of AI-related research output, and as whole, publications data in both English and Chinese suggests that AI-related research collaboration between China and Russia has been relatively scarce over the past decade. The situation, however, could change if the two nations commit resources and talent to this area; in the conclusion, we outline potential developments that may signal closer collaboration. These patterns in collaboration on AI-related topics reflect broader trends in scientific research and co-authorships among the United States, China, and Russia. Based on 2018 NSF data, Chinese and Russian researchers co-authored 2,457 publications in science and engineering. This total is significantly less than the 55,382 publications co-authored by U.S. and Chinese researchers, and the 4,881 publications co-authored by U.S. and Russian researchers that year. According to a recent study of U.S.-China research ties, \"from the years 2014 to 2018, there has been a continuous rise of co-authored science and engineering journal publications that include U. * S. and China co-authors, with an all-time high in … 2018, with 175,665 articles. This number represents roughly 10 percent of all U.S. and 9 percent of all China science and engineering article publications from 2014 to 2018.\" 68  NSF data from 2018 also shows that nearly 44 percent of China's internationally coauthored articles in science and engineering were with a U.S. coauthor. 69  Meanwhile, in terms of U.S.-Russia research ties, while the United States seems to be one of Russia's top research collaborators, Russia is not one of the top countries for research collaboration for the United States.While statements from high-level representatives from China and Russia, media reports, and expert analysis all point to the establishment of new scientific partnerships and expansion of academic ties, assessing the scope and output of collaborative research is a more complicated task. \n , Chinese, and Russian AI Investment Activity The In the course of our research, we observed five additional investment deals with unknown values, bringing the total count to 17 China-Russia AI investment deals. We also noted several other forms of cooperation that we did not code as investments, including strategic cooperation agreements and joint product development. While deals involving state-owned enterprises and other undisclosed transactions raise the level of uncertainty in our findings, we cross-referenced our data sets with press releases and other joint statements and cooperative agreements between Russian and Chinese firms. value of AI investments between the United States and China far exceeds that of any other pair of counties in the world. In 2019, U.S. companies attracted $25.2 billion, or 64 percent of total disclosed AI investment value; China was the second largest market, with $5.4 billion, or 12.9 percent of the global total.75   Between 2010 and 2020, U.S. firms struck 270 investment deals in Chinese AI companies, and Chinese firms made 323 such investments in the U. China-Russia AI investment deals with known values. The lack of transparency involving deals with state-owned enterprises presents challenges for any comprehensive assessment of commercial AI activity. While our analysis does not cover all investment activity between China and Russia, we implemented several additional checks to account for potential gaps and limitations in our approach. We provide case studies of two of the largest AI-related investments: Alibaba's joint venture in 2019 with Mail.ru Group, a Russian internet company that operates social networking sites, email services, and internet portals; and the 2017 Russia-China Investment Fund (RCIF) investment in the Chinese AI company Megvii. Comparing U.S. S. AI private market, for a combined total value of more than $35 billion. Over time, U.S. investment into private AI companies in China has declined from a peak of $13 billion in 2017 to $1.3 billion in 2020; whereas China's AI investment into private AI companies in the United States has risen steadily from $181 million in 2014 to $2 billion in 2020. \n Joint Venture in 2019 with Mail.ru Group AliExpress 80  Mounting concerns about the company prompted the Trump administration to add Megvii to the U.S. Department of Commerce's Entity list in 2019, blocking its access to U.S. technology.81    While this decision threatened its future sales abroad, Megvii is showing no signs of slowing down at home. The company is one of China's four \"AI dragons'' alongside SenseTime, Yitu, and CloudWalk. 82  It has absorbed $1.4 billion in capital through multiple rounds of funding. Along with Foxconn, Ant Group, SK Group, China Reform Holdings Corporation, and Sunshine Insurance Group, RCIF funded Megvii's second largest round of financing in a 2017 Series C deal worth $460 million. 83  Megvii's largest infusion of venture capital, worth $750 million, came in a 2019 Series D round from the Bank of China and Alibaba Group, among others. By August 2019, Megvii was on track to raise nearly $1 billion in its initial public offering (IPO) on the Hong Kong Stock Exchange. But the Trump administration's restrictions on the company thwarted the IPO. 84  Megvii was back for another try in March 2021, filing for an IPO on the Shanghai Stock Exchange and planning to raise at least 6 billion yuan ($922 million). against the totality of investments into Megvii, Russia's contribution appears modest. RCIF was just one of six investors funding the $460 million Series C round. In addition, the RCIF received equal portions from the Russian Direct Investment Fund, a sovereign wealth fund established in 2011, and the Chinese Investment Corporation, meaning that Russia funded only 50 percent of RCIF's contribution to Megvii's Series C round.85  While actual percentages of contributions to the funding round are difficult to ascertain, one could assume that each of the six investors may have contributed equal parts. This estimate would make RCIF's contribution worth around $76.7 million and Russia's portion just $38.3 million. Out of $1.4 billion venture capital investments thus far, and notwithstanding RCIF's potential participation in Megvii's upcoming IPO, Russia's investment contribution will have been around 2.7 percent. In short, even when looking at the highest value deals, investments made through joint Russian-Chinese funds account for a fraction of what it takes to get an AI company's IPO off the ground. Russia is a joint venture between China's Alibaba and Russia's Mail.ru. Launched in 2019 and valued at around $2 billion, AliExpress Russia claims it will provide a single destination for online shopping, gaming, instant messaging, and mobile banking for Mail.ru's hundreds of millions of registered users. 86  As Mail.ru Group CEO Boris Dobrodeev noted in a statement, \"[T]his partnership will enable us to significantly increase the access to various segments of the e-commerce offering, including both cross-border and local merchants. The combination of our ecosystems allows us to leverage our distribution through our merchant base and goods as well as product integrations.\" 87  Such claims are noteworthy in their ambition and scope, but merit scrutiny in light of the incentives to overstate planned cooperation and development. While Mail.ru is a dominant player in Russian social media, its presence in online banking and commerce is insignificant. Measured Alibaba's \n The value of RCIF's 2014 investment in the company has not been publicly disclosed. Since then, DiDi has emerged as a world leader in autonomous driving and AI more broadly, and now competes with U.S.-based companies Waymo and Uber. In January 2021, RCIF announced a second investment specifically in DiDi's Autonomous Driving division valued at $300 million, the largest investment RCIF has made thus far, and the second largest AI investment deal between China and Russia. 95  Rostec City. 2017 marked RCIF's most active investment year to date, inaugurated with a major investment in Rostec City, a technology park in the heart of Moscow. Spearheaded by Russian defense SOE and electronics giant Rostec, the park is expected to showcase \"smart city\" technologies designed to autonomously regulate traffic, adapt power consumption, and recognize the faces and movements of its inhabitants. 96  Moscow's Smart City 2030 strategy focuses explicitly on AI. 97  Smart city technoparks are a staple of several Chinese cities and often viewed as proving grounds for emerging surveillance technologies. 98  Nio. In October 2017, RCIF and the China Investment Corporation jointly invested in Nio, an early-stage electric vehicle company headquartered in China. With shareholders including Chinese AI giants Tencent and Baidu, Nio-\"China's Tesla\"-quickly came to focus on autonomous driving. 99  Its first car model, a fully electric, fully autonomous sedan, was unveiled in January 2021. 100  NetEase Cloud Music. Known for its web portal and other internet services, NetEase is a well-established Chinese internet company that also runs a personalized music recommendation system. Its product includes some social media features and collects data on users' music preferences to connect friends and peers. 101  In November 2018, RCIF contributed an unknown amount of funds to a financing round worth $600 million. 102  In 2019, Russian robotics firm Promobot, one of the largest robotics manufacturers in Europe, signed a cooperation agreement with Chinese company \"FN Holdings,\" creating a joint project team that will develop a linguistic base in Chinese and coordinate joint research related to healthcare service robots-a deal that's estimated to bring Promobot $2 million.103  Another cooperation agreement with Beijing Huanuo Exiang (REX), a Chinese trading company, aims to bring Promobot's products on to the Chinese entertainment and services market. 104   Other Forms of Cooperation Beyond investments, we identified several other forms of corporate partnerships between Chinese and Russian AI companies. While not indicative of the same commitment as cash-in-hand, Russian and Chinese companies have signed strategic cooperation agreements related to AI development, and some have jointly developed surveillance products to be marketed in both countries. Mutual cooperation agreements range from boosting trade to bring existing products into Chinese and Russian markets to pursuing joint research and development of new products. Russia's Promobot, China's FN Holdings, and Beijing Huanuo cooperation in educational robotics, UBTech Robotics and Russian technology giant Laboratory of New Information Technology (LANIT) signed a partnership agreement in 2019, agreeing to bring AI-enabled humanoid robots into the Russian market. 105 In addition to industrial applications such as inspecting hazardous facilities, the joint venture will localize and adapt Chinese robots for Russian service and educational sectors to serve as virtual assistants, museum guides, and training references for educational institutions. 106 China's Fitsco and Russia's Cognitive Pilot. Advancing cooperation in smart cities and the autonomous transportation sector, Shanghai Fuxin Intelligent Transportation Solutions Company (Fitsco) and Cognitive Pilot, a Russian autonomous vehicle company, partnered in 2020 to develop an AI-based computer vision system for Exiang. China's UBTech and Russia's LANIT Group. Deepening autonomous trams in China. \n China and  Russia are growing more aligned in their interests, actions, and strategies, and there are clear indications of closer diplomatic, economic, military, and technological ties. Yet the scale and scope of this emerging partnership deserve careful scrutiny, particularly in the field of artificial intelligence. This issue brief outlined and analyzed Chinese-Russian collaboration in AI-related research and investment. Our findings align with previous assessments of Sino-Russian cooperation in new technologies, tracing an upward trend in the number of AIrelated research publications as well as growing investment in the AI ecosystems of these countries. At the same time, the scope of this partnership still lags behind the breadth and depth of ties between the United States and China.The data on joint AI-related research and investment highlight the limits of Chinese-Russian collaboration. The power asymmetry between the two nations inevitably complicates this partnership. More specifically, it remains unclear whether China sees Russia as a true partner suitable for joint ventures and collaboration on equal terms. For instance, in 2019, the president of the Russian Academy of Sciences expressed concern that in regards to Russia, \"China has shifted from a policy of importing technology to a policy of importing brains,\" offering Russian scientists not only better salaries but the opportunity to carry on their research in China. 115  Yet as the exact contours of this partnership continue to take shape, China may proceed to acquire Russian technology, through licit and illicit means. Recent CSET research shows that China's science and technology diplomats-those individuals stationed in PRC embassies and consulates around the world to monitor scientific and technological breakthroughs and track investment opportunities for the Chinese government-seem particularly interested in Russia, especially in early-stage technology projects developed by Russian government-backed researchers. 116  Convergence between America's two key competitors, especially around critical technologies like AI, has significant implications for U.S. national security and global leadership. While this report has focused on AI-related research and investment linkages, there are additional dimensions to Chinese-Russian cooperation in AI, with experts highlighting areas such as dialogues and exchanges, the development of science and technology parks, joint competitions, and the expansion of academic cooperation. 117  Talent development and exchanges are worth tracking in particular, especially if Chinese companies like Huawei continue to expand and deepen their partnerships with Russian universities and research institutions. Notably, the Russian Embassy in China recently claimed that the number of Chinese students in Russian universities has doubled over the past five years, reaching about 48,000 students in 2020; the number of Russian students in Chinese universities has increased by 36 percent, standing at 20,000.118    Still, Russia is not a top destination for Chinese students, nor is China for Russian students. In contrast, there were about 372,531 Chinese students enrolled in U.S. universities during the 2019-2020 academic year, accounting for 35 percent of the total number of international students. 119 Moreover, the majority of top-tier AI researchers who received undergraduate degrees in China go on to study, work, and live in the United States. 120 Roughly 30 percent of international students pursuing a Ph.D. in AI-relevant fields such as computer science and electrical engineering in the United States are from China.121  That said, talent exchanges in particular, and Sino-Russian cooperation more generally, will likely deepen if the United States' relationships with both countries continue to deteriorate.While we are reluctant to put forth policy recommendations on the basis of findings related to only two dimensions of the multifaceted Chinese-Russian partnership, we offer the following list of indicators that watchers ofChina  and Russia should track closely as cooperation between these two nations continues to unfold. Chinese and Russian universities, research institutions, and private companies increasingly formalize their research ties through multi-year agreements with designated funding. o Chinese and Russian universities and research institutions establish additional joint research centers, joint undergraduate and graduate degree programs, summer schools, and other ventures. The share of collaborative publications as part of overall publications for both China and Russia grows. o The citation impact and the percentage of highly cited AI-related papers co-authored by Chinese and Russian researchers increases. o Joint Chinese-Russian university research centers produce more and higher impact research. o Huawei research centers in Russia and partnerships with universities produce innovative and internationally recognized work. • Partnerships between defense research institutions and other entities of note to U.S. national security: o Chinese companies and institutions listed on the Entity List increasingly collaborate on AI-related topics with Russian universities and research institutions. o Research institutions affiliated with the Russian Ministry of Defense, or located in the ERA technopolis, increasingly collaborate with Chinese counterparts on AI-related research. Increases in the disclosed value and deal count of joint Chinese and Russian AI investments. o The percentage of joint AI investment between China and Russia surpasses the level of U.S.-China AI investment or occupies a majority share of China's and Russia's overall AI investment portfolios globally. o China and Russia develop additional products or establish additional joint AI ventures and innovation funds, on the order of RCIF or larger. • Hidden values or breakthrough developments: o Drop-offs in disclosed investment values or deal counts that indicate a greater role for state-owned enterprises, deliberate obfuscation, or increasing lack of transparency in joint AI investments. o Investments leading to groundbreaking AI products and innovations in the areas of autonomous vehicles, smart cities, facial recognition, telecommunications, large language models (natural language processing), or AI-enabled military applications. o Increase in data sharing agreements, exceptions to data localization laws, and other signs of mutually preferential treatment to attract AI companies and talent. • Proliferation of investments in areas of concern: o Significant growth in assets under management of RCIF's AI investment portfolio, with a focus on autonomous vehicles, smart cities, facial recognition, telecommunications, large language models (natural language processing), or AI-enabled military applications. o Increases in Chinese and Russian joint AI investments in facial recognition, such as the RCIF investment in Megvii, or growing partnerships between Russian firms and China's four \"AI dragons'': SenseTime, Yitu, CloudWalk, and Megvii. o Collaboration in AI development between Chinese and Russian state-owned defense companies, such as China Electronics Technology Corporation and Rostec. Implementation of existing corporate partnerships and development of new strategic cooperation agreements, such as the agreement between China's Dahua Technology and Russia's NtechLab to develop surveillance capabilities or between China's Fitsco and Russia's Cognitive Pilot to advance smart city and autonomous vehicle technologies. Authors Margarita Konaev is a research fellow at CSET, where Andrew Imbrie is a senior fellow, and Ryan Fedasiuk and Emily Weinstein are research analysts. Katerina Sedova is a research fellow with CSET's CyberAI, and James Dunham is a data scientist at CSET. Research: • Formalization and proliferation of research relationships: o • Increased volume and quality of collaborative AI-related publications: • Increased disclosed investment values and deal counts in relative and absolute terms: o Investment: o o \n\t\t\t Other AI research 98,620 Other AI research 3,222 Source: Microsoft Academic Graph, Digital Science Dimensions, Clarivate Web of Science, and CNKI. The top collaborative fields are similar to the top 10 fields for research by Russian scientists, except for remote sensing, which is on the list for collaborative papers but not on the list of Russian * \"Other AI research\" refers to the \"Artificial Intelligence\" subject field denoted by Microsoft Academic Graph. Publications in this field represent AI research not otherwise classified. This could include basic research that does not fall into a different field because it is not application-specific (e.g., not computer vision research), theoretical work related to AI, or applied research in a subfield too small to result in the creation of a distinct, comparable Microsoft Academic Graph field. \n\t\t\t © 2021 by the Center for Security and Emerging Technology. This work is licensed under a Creative Commons Attribution-Non Commercial 4.0 International License. To view a copy of this license, visit https://creativecommons.org/licenses/by-nc/4.0/. Document Identifier: doi: 10.<PHONE> \n\t\t\t The number of investment deals in our dataset is small, making it difficult to assess clear trend lines over time.20  Karen White, \"Publications Output: U.S. Trends and International Comparisons,\" National Science Foundation | National Science Board, December 17, 2019, https://ncses.nsf.gov/pubs/nsb20206/international-collaboration. \n\t\t\t It is important to note that, since the Chinese language data only contain 43 papers spanning 15 years, any trends must be taken with a grain of salt. In addition, these 43 papers may contain papers that were either translated into Chinese from either English or Russian, or vice versa.62  The Seven Sons of National Defense are a group of Chinese universities, administered directly by the Ministry of Industry and Information Technology, with historically intimate ties to China's defense ecosystem. For more \n\t\t\t Remco Zwetsloot, James Dunham, Zachary Arnold, and Tina Huang, \"Keeping AI Talent in the United States\" (Center for Security and Emerging", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Headline-or-Trend-Line.tei.xml", "id": "2972dd096a515e1f28f11f3bdbcd5e3e"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "I. J. Good's thesis of the \"intelligence explosion\" states that a sufficiently advanced machine intelligence could build a smarter version of itself, which could in turn build an even smarter version, and that this process could continue to the point of vastly exceeding human intelligence. As Sandberg (2010) correctly notes, there have been several attempts to lay down return on investment formulas intended to represent sharp speedups in economic or technological growth, but very little attempt has been made to deal formally with Good's intelligence explosion thesis as such. I identify the key issue as returns on cognitive reinvestment-the ability to invest more computing power, faster computers, or improved cognitive algorithms to yield cognitive labor which produces larger brains, faster brains, or better mind designs. There are many phenomena in the world which have been argued to be evidentially relevant to this question, from the observed course of hominid evolution, to Moore's Law, to the competence over time of machine chess-playing systems, and many more. I go into some depth on some debates which then arise on how to interpret such evidence. I propose that the next step in analyzing positions on the intelligence explosion would be to formalize return on investment curves, so that each stance can formally state which possible microfoundations they hold to be falsified by historical observations. More generally,", "authors": ["Eliezer Yudkowsky"], "title": "Intelligence Explosion Microeconomics", "text": "I pose multiple open questions of \"returns on cognitive reinvestment\" or \"intelligence explosion microeconomics.\" Although such questions have received little attention thus far, they seem highly relevant to policy choices affecting outcomes for Earth-originating intelligent life. \n The Intelligence Explosion: Growth Rates of Cognitive Reinvestment In 1965, I. J. Good 1 published a paper titled \"Speculations Concerning the First Ultraintelligent Machine\"  (Good 1965 ) containing the paragraph: Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an \"intelligence explosion,\" and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make. Many have since gone on to question Good's unquestionable, and the state of the debate has developed considerably since 1965. While waiting on Nick Bostrom's forthcoming book on the intelligence explosion, I would meanwhile recommend the survey paper \"Intelligence Explosion: Evidence and Import\"  (Muehlhauser and Salamon 2012)  for a compact overview. See also  David Chalmers's (2010)  paper, the responses, and  Chalmers's (2012)  reply. Please note that the intelligence explosion is not the same thesis as a general economic or technological speedup, which is now often termed a \"Singularity.\" Economic speedups arise in many models of the future, some of them already well formalized. For example, Robin  Hanson's (1998a)  \"Economic Growth Given Machine Intelligence\" considers emulations of scanned human brains (a.k.a. ems): Hanson proposes equations to model the behavior of an economy when capital (computers) can be freely converted into human-equivalent skilled labor (by running em software). Hanson concludes that the result should be a global economy with a doubling time on the order of months. This may sound startling already, but Hanson's paper doesn't try to model an agent that is smarter than any existing human, or whether that agent would be able to invent stillsmarter agents. The question of what happens when smarter-than-human agencies 2 are driving scientific and technological progress is difficult enough that previous attempts at formal 1. Isadore Jacob Gudak, who anglicized his name to Irving John Good and used I. J. Good for publication. He was among the first advocates of the Bayesian approach to statistics, and worked with Alan Turing on early computer designs. Within computer science his name is immortalized in the Good-Turing frequency estimator. futurological modeling have entirely ignored it, although it is often discussed informally; likewise, the prospect of smarter agencies producing even smarter agencies has not been formally modeled. In his paper overviewing formal and semiformal models of technological speedup,  Sandberg (2010)  concludes: There is a notable lack of models of how an intelligence explosion could occur. This might be the most important and hardest problem to crack. . . . Most important since the emergence of superintelligence has the greatest potential of being fundamentally game-changing for humanity (for good or ill). Hardest, since it appears to require an understanding of the general nature of super-human minds or at least a way to bound their capacities and growth rates. For responses to some arguments that the intelligence explosion is qualitatively forbidden-for example, because of Gödel's Theorem prohibiting the construction of artificial minds 3 -see again  Chalmers (2010)  or  Muehlhauser and Salamon (2012) . The Open Problem posed here is the quantitative issue: whether it's possible to get sustained returns on reinvesting cognitive improvements into further improving cognition. As  Chalmers (2012)  put it: The key issue is the \"proportionality thesis\" saying that among systems of certain class, an increase of δ in intelligence will yield an increase of δ in the intelligence of systems that these systems can design. To illustrate the core question, let us consider a nuclear pile undergoing a fission reaction.  4  The first human-made critical fission reaction took place on December 2, 1942, in a rackets court at the University of Chicago, in a giant doorknob-shaped pile of uranium bricks and graphite bricks. The key number for the pile was the effective neutron multiplication factor k-the average number of neutrons emitted by the average number of fissions caused by one neutron. (One might consider k to be the \"return on investment\" for neutrons.) A pile with k > 1 would be \"critical\" and increase exponentially in neutrons. Adding more uranium bricks increased k, since it gave a neutron more opportunity to strike more uranium atoms before exiting the pile. Fermi had calculated that the pile ought to go critical between layers 56 and 57 of uranium bricks, but as layer 57 was added, wooden rods covered with neutron-absorbing 3. A.k.a. general AI, a.k.a. strong AI, a.k.a. Artificial General Intelligence. See  Pennachin and Goertzel (2007) . 4. Uranium atoms are not intelligent, so this is not meant to imply that an intelligence explosion ought to be similar to a nuclear pile. No argument by analogy is intended-just to start with a simple process on the way to a more complicated one. cadmium foil were inserted to prevent the pile from becoming critical. The actual critical reaction occurred as the result of slowly pulling out a neutron-absorbing rod in six-inch intervals. As the rod was successively pulled out and k increased, the overall neutron level of the pile increased, then leveled off each time to a new steady state. At 3:25 p.m., Fermi ordered the rod pulled out another twelve inches, remarking, \"Now it will become self-sustaining. The trace will climb and continue to climb. It will not level off \"  (Rhodes 1986 ). This prediction was borne out: the Geiger counters increased into an indistinguishable roar, and other instruments recording the neutron level on paper climbed continuously, doubling every two minutes until the reaction was shut down twenty-eight minutes later. For this pile, k was 1.0006. On average, 0.6% of the neutrons emitted by a fissioning uranium atom are \"delayed\"-they are emitted by the further breakdown of short-lived fission products, rather than by the initial fission (the \"prompt neutrons\"). Thus the above pile had k = 0.9946 when considering only prompt neutrons, and its emissions increased on a slow exponential curve due to the contribution of delayed neutrons. A pile with k = 1.0006 for prompt neutrons would have doubled in neutron intensity every tenth of a second. If Fermi had not understood the atoms making up his pile and had only relied on its overall neutron-intensity graph to go on behaving like it had previously-or if he had just piled on uranium bricks, curious to observe empirically what would happen-then it would not have been a good year to be a student at the University of Chicago. Nuclear weapons use conventional explosives to compress nuclear materials into a configuration with prompt k 1; in a nuclear explosion, k might be on the order of 2.3, which is \"vastly greater than one\" for purposes of nuclear engineering. At the time when the very first human-made critical reaction was initiated, Fermi already understood neutrons and uranium atoms-understood them sufficiently well to pull out the cadmium rod in careful increments, monitor the increasing reaction carefully, and shut it down after twenty-eight minutes. We do not currently have a strong grasp of the state space of cognitive algorithms. We do not have a strong grasp of how difficult or how easy it should be to improve cognitive problem-solving ability in a general AI by adding resources or trying to improve the underlying algorithms. We probably shouldn't expect to be able to do precise calculations; our state of uncertain knowledge about the space of cognitive algorithms probably shouldn't yield Fermi-style verdicts about when the trace will begin to climb without leveling off, down to a particular cadmium rod being pulled out twelve inches. But we can hold out some hope of addressing larger, less exact questions, such as whether an AI trying to self-improve, or a global population of AIs trying to selfimprove, can go \"critical\" (k ≈ 1 + ) or \"supercritical\" (prompt k 1). We shouldn't expect to predict exactly how many neutrons the metaphorical pile will output after two minutes. But perhaps we can predict in advance that piling on more and more uranium bricks will eventually cause the pile to start doubling its neutron production at a rate that grows quickly compared to its previous ascent . . . or, alternatively, conclude that self-modifying AIs should not be expected to improve at explosive rates. So as not to allow this question to become too abstract, let us immediately consider some widely different stances that have been taken on the intelligence explosion debate. This is not an exhaustive list. As with any concrete illustration or \"detailed storytelling,\" each case will import large numbers of auxiliary assumptions. I would also caution against labeling any particular case as \"good\" or \"bad\"-regardless of the true values of the unseen variables, we should try to make the best of them. With those disclaimers stated, consider these concrete scenarios for a metaphorical \"k much less than one,\" \"k slightly more than one,\" and \"prompt k significantly greater than one,\" with respect to returns on cognitive investment. k < 1, the \"intelligence fizzle\": Argument: For most interesting tasks known to computer science, it requires exponentially greater investment of computing power to gain a linear return in performance. Most search spaces are exponentially vast, and low-hanging fruits are exhausted quickly. Therefore, an AI trying to invest an amount of cognitive work w to improve its own performance will get returns that go as log(w), or if further reinvested, log(w + log(w)), and the sequence log(w), log(w + log(w)), log(w + log(w + log(w))) will converge very quickly. \n Scenario: We might suppose that silicon intelligence is not significantly different from carbon, and that AI at the level of John von Neumann can be constructed, since von Neumann himself was physically realizable. But the constructed von Neumann does much less interesting work than the historical von Neumann, because the lowhanging fruits of science have already been exhausted. Millions of von Neumanns only accomplish logarithmically more work than one von Neumann, and it is not worth the cost of constructing such AIs. AI does not economically substitute for most cognitively skilled human labor, since even when smarter AIs can be built, humans can be produced more cheaply. Attempts are made to improve human intelligence via genetic engineering, or neuropharmaceuticals, or braincomputer interfaces, or cloning Einstein, etc.; but these attempts are foiled by the discovery that most \"intelligence\" is either unreproducible or not worth the cost of reproducing it. Moore's Law breaks down decisively, not just because of increasing technological difficulties of miniaturization, but because ever-faster computer chips don't accomplish much more than the previous generation of chips, and so there is insufficient economic incentive for Intel to build new factories. Life continues mostly as before, for however many more centuries. k ≈ 1 + , the \"intelligence combustion\": Argument: Over the last many decades, world economic growth has been roughly exponential-growth has neither collapsed below exponential nor exploded above, implying a metaphorical k roughly equal to one (and slightly on the positive side). This is the characteristic behavior of a world full of smart cognitive agents making new scientific discoveries, inventing new technologies, and reinvesting resources to obtain further resources. There is no reason to suppose that changing from carbon to silicon will yield anything different. Furthermore, any single AI agent is unlikely to be significant compared to an economy of seven-plus billion humans. Thus AI progress will be dominated for some time by the contributions of the world economy to AI research, rather than by any one AI's internal self-improvement. No one agent is capable of contributing more than a tiny fraction of the total progress in computer science, and this doesn't change when human-equivalent AIs are invented. 5 \n Scenario: The effect of introducing AIs to the global economy is a gradual, continuous increase in the overall rate of economic growth, since the first and most expensive AIs carry out a small part of the global economy's cognitive labor. Over time, the cognitive labor of AIs becomes cheaper and constitutes a larger portion of the total economy. The timescale of exponential growth starts out at the level of a human-only economy and gradually, continuously shifts to a higher growth ratefor example,  Hanson (1998b)  predicts world economic doubling times of between a month and a year. Economic dislocations are unprecedented but take place on a timescale which gives humans some chance to react. \n Prompt k 1, the \"intelligence explosion\": Argument: The history of hominid evolution to date shows that it has not required exponentially greater amounts of evolutionary optimization to produce substantial real-world gains in cognitive performance-it did not require ten times the evolutionary interval to go from Homo erectus to Homo sapiens as from Australopithecus to Homo erectus.  6  All compound interest returned on discoveries such as the invention 5. I would attribute this rough view to Robin Hanson, although he hasn't confirmed that this is a fair representation. of agriculture, or the invention of science, or the invention of computers, has occurred without any ability of humans to reinvest technological dividends to increase their brain sizes, speed up their neurons, or improve the low-level algorithms used by their neural circuitry. Since an AI can reinvest the fruits of its intelligence in larger brains, faster processing speeds, and improved low-level algorithms, we should expect an AI's growth curves to be sharply above human growth curves. \n Scenario: The first machine intelligence system to achieve sustainable returns on cognitive reinvestment is able to vastly improve its intelligence relatively quickly-for example, by rewriting its own software or by buying (or stealing) access to orders of magnitude more hardware on clustered servers. Such an AI is \"prompt critical\"it can reinvest the fruits of its cognitive investments on short timescales, without the need to build new chip factories first. By the time such immediately accessible improvements run out, the AI is smart enough to, for example, crack the problem of protein structure prediction. The AI emails DNA sequences to online peptide synthesis labs (some of which boast a seventy-two-hour turnaround time), and uses the resulting custom proteins to construct more advanced ribosome equivalents (molecular factories). Shortly afterward, the AI has its own molecular nanotechnology and can begin construction of much faster processors and other rapidly deployed, technologically advanced infrastructure. This rough sort of scenario is sometimes colloquially termed \"hard takeoff \" or \"AI-go-FOOM.\" 7 There are many questions we could proceed to ask about these stances, which are actually points along a spectrum that compresses several different dimensions of potentially independent variance, etc. The implications from the arguments to the scenarios are also disputable. Further sections will address some of this in greater detail. The broader idea is that different positions on \"How large are the returns on cognitive reinvestment?\" have widely different consequences with significant policy implications. The problem of investing resources to gain more resources is fundamental in economics. An (approximately) rational agency will consider multiple avenues for improvement, purchase resources where they are cheapest, invest where the highest returns are expected, and try to bypass any difficulties that its preferences do not explicitly forbid 7. I must quickly remark that in my view, whether an AI attaining great power is a good thing or a bad thing would depend strictly on the AI's goal system. This in turn may depend on whether the programmers were able to solve the problem of \"Friendly AI\" (see  Yudkowsky [2008a] ). This above point leads into another, different, and large discussion which is far beyond the scope of this paper, though I have very, very briefly summarized some core ideas in section 1.3. Nonetheless it seems important to raise the point that a hard takeoff/AI-go-FOOM scenario is not necessarily a bad thing, nor inevitably a good one. bypassing. This is one factor that makes an artificial intelligence unlike a heap of uranium bricks: if you insert a cadmium-foil rod into a heap of uranium bricks, the bricks will not try to shove the rod back out, nor reconfigure themselves so that the rod absorbs fewer valuable neutrons. In economics, it is routine to suggest that a rational agency will do its best to overcome, bypass, or intelligently reconfigure its activities around an obstacle. Depending on the AI's preferences and capabilities, and on the surrounding society, it may make sense to steal poorly defended computing resources; returns on illegal investments are often analyzed in modern economic theory. Hence the problem of describing an AI's curve for reinvested growth seems more like existing economics than existing problems in physics or computer science. As \"microeconomics\" is the discipline that considers rational agencies (such as individuals, firms, machine intelligences, and well-coordinated populations of machine intelligences) trying to maximize their returns on investment,  8  the posed open problem about growth curves under cognitive investment and reinvestment is titled \"Intelligence Explosion Microeconomics.\" Section 2 of this paper discusses the basic language for talking about the intelligence explosion and argues that we should pursue this project by looking for underlying microfoundations, not by pursuing analogies to allegedly similar historical events. Section 3 attempts to showcase some specific informal reasoning about returns on cognitive investments, displaying the sort of arguments that have arisen in the context of the author explaining his stance on the intelligence explosion. Section 4 proposes a tentative methodology for formalizing theories of the intelligence explosion-a project of describing possible microfoundations and explicitly stating their alleged relation to historical experience, such that some possibilities can be falsified. Section 5 explores which subquestions seem both high value and possibly answerable. There are many things we'd like to know that we probably can't know given a reasonable state of uncertainty about the domain-for example, when will an intelligence explosion occur? Section 6 summarizes and poses the open problem, and discusses what would be required for MIRI to fund further work in this area. \n On (Extensionally) Defining Terms It is obvious to ask questions like \"What do you mean by 'intelligence'?\" or \"What sort of AI system counts as 'cognitively reinvesting'?\" I shall attempt to answer these questions, but any definitions I have to offer should be taken as part of my own personal theory of the intelligence explosion. Consider the metaphorical position of early scientists 8. Academically, \"macroeconomics\" is about inflation, unemployment, monetary policy, and so on. who have just posed the question \"Why is fire hot?\" Someone then proceeds to ask, \"What exactly do you mean by 'fire'?\" Answering, \"Fire is the release of phlogiston\" is presumptuous, and it is wiser to reply, \"Well, for purposes of asking the question, fire is that bright orangey-red hot stuff coming out of that heap of sticks-which I think is really the release of phlogiston-but that definition is part of my answer, not part of the question itself.\" I think it wise to keep this form of pragmatism firmly in mind when we are trying to define \"intelligence\" for purposes of analyzing the intelligence explosion. 9 So as not to evade the question entirely, I usually use a notion of \"intelligence ≡ efficient cross-domain optimization,\" constructed as follows: 1. Consider optimization power as the ability to steer the future into regions of possibility ranked high in a preference ordering. For instance, Deep Blue has the power to steer a chessboard's future into a subspace of possibility which it labels as \"winning,\" despite attempts by Garry Kasparov to steer the future elsewhere. Natural selection can produce organisms much more able to replicate themselves than the \"typical\" organism that would be constructed by a randomized DNA string-evolution produces DNA strings that rank unusually high in fitness within the space of all DNA strings. 10 2. Human cognition is distinct from bee cognition or beaver cognition in that human cognition is significantly more generally applicable across domains: bees build hives and beavers build dams, but a human engineer looks over both and then designs a dam with a honeycomb structure. This is also what separates Deep Blue, which only played chess, from humans, who can operate across many different domains and learn new fields. 9. On one occasion I was debating Jaron Lanier, who was arguing at length that it was bad to call computers \"intelligent\" because this would encourage human beings to act more mechanically, and therefore AI was impossible; and I finally said, \"Do you mean to say that if I write a program and it writes a program and that writes another program and that program builds its own molecular nanotechnology and flies off to Alpha Centauri and starts constructing a Dyson sphere, that program is not intelligent?\" 10. \"Optimization\" can be characterized as a concept we invoke when we expect a process to take on unpredictable intermediate states that will turn out to be apt for approaching a predictable destinatione.g., if you have a friend driving you to the airport in a foreign city, you can predict that your final destination will be the airport even if you can't predict any of the particular turns along the way. Similarly, Deep Blue's programmers retained their ability to predict Deep Blue's final victory by inspection of its code, even though they could not predict any of Deep Blue's particular moves along the way-if they knew exactly where Deep Blue would move on a chessboard, they would necessarily be at least that good at chess themselves. 3. Human engineering is distinct from natural selection, which is also a powerful cross-domain consequentialist optimizer, in that human engineering is faster and more computationally efficient. (For example, because humans can abstract over the search space, but that is a hypothesis about human intelligence, not part of my definition.) In combination, these yield a definition of \"intelligence ≡ efficient cross-domain optimization.\" This tries to characterize \"improved cognition\" as the ability to produce solutions higher in a preference ordering, including, for example, a chess game with a higher probability of winning than a randomized chess game, an argument with a higher probability of persuading a human target, a transistor connection diagram that does more floatingpoint operations per second than a previous CPU, or a DNA string corresponding to a protein unusually apt for building a molecular factory. Optimization is characterized by an ability to hit narrow targets in a search space, where demanding a higher ranking in a preference ordering automatically narrows the measure of equally or more preferred outcomes. Improved intelligence is then hitting a narrower target in a search space, more computationally efficiently, via strategies that operate across a wider range of domains. That definition is one which I invented for other purposes (my work on machine intelligence as such) and might not be apt for reasoning about the intelligence explosion. For purposes of discussing the intelligence explosion, it may be wiser to reason about forms of growth that more directly relate to quantities we can observe. The narrowness of the good-possibility space attained by a search process does not correspond very directly to most historical observables. And for purposes of posing the question of the intelligence explosion, we may be better off with \"Intelligence is that sort of smartish stuff coming out of brains, which can play chess, and price bonds, and persuade people to buy bonds, and invent guns, and figure out gravity by looking at wandering lights in the sky; and which, if a machine intelligence had it in large quantities, might let it invent molecular nanotechnology; and so on.\" To frame it another way, if something is powerful enough to build a Dyson Sphere, it doesn't really matter very much whether we call it \"intelligent\" or not. And this is just the sort of \"intelligence\" we're interested in-something powerful enough that whether or not we define it as \"intelligent\" is moot. This isn't to say that definitions are forbidden-just that further definitions would stake the further claim that those particular definitions were apt for carving reality at its joints, with respect to accurately predicting an intelligence explosion. Choice of definitions has no power to affect physical reality. If you manage to define \"AI self-improvement\" in such a way as to exclude some smartish computer-thingy which carries out some mysterious internal activities on its own code for a week and then emerges with a solution to protein structure prediction which it uses to build its own molecular nanotechnology . . . then you've obviously picked the wrong definition of \"self-improvement.\" See, for example, the definition advocated by  Mahoney (2010)  in which \"self-improvement\" requires an increase in Kolmogorov complexity of an isolated system, or  Bringsjord's (2012)  definition in which a Turing machine is only said to selfimprove if it can raise itself into a class of hypercomputers. These are both definitions which strike me as inapt for reasoning about the intelligence explosion, since it is not obvious (in fact I think it obviously false) that this sort of \"self-improvement\" is required to invent powerful technologies. One can define self-improvement to be the increase in Kolmogorov complexity of an isolated deterministic system, and proceed to prove that this can only go as the logarithm of time. But all the burden of showing that a realworld intelligence explosion is therefore impossible rests on the argument that doing impactful things in the real world requires an isolated machine intelligence to increase its Kolmogorov complexity. We should not fail to note that this is blatantly false. 11 This doesn't mean that we should never propose more sophisticated definitions of selfimprovement. It means we shouldn't lose sight of the wordless pragmatic background concept of an AI or AI population that rewrites its own code, or writes a successor version of itself, or writes an entirely new AI, or builds a better chip factory, or earns money to purchase more server time, or otherwise does something that increases the amount of pragmatically considered cognitive problem-solving capability sloshing around the system. And beyond that, \"self-improvement\" could describe genetically engineered humans, or humans with brain-computer interfaces, or upload clades, or several other possible scenarios of cognitive reinvestment, albeit here I will focus on the case of machine intelligence.  12  It is in this spirit that I pose the open problem of formalizing I. J. Good's notion of the intelligence explosion. Coming up with good definitions for informal terms like \"cognitive reinvestment,\" as they appear in the posed question, can be considered as part of the problem. In further discussion I suggest various definitions, categories, and 11. Since any system with a Kolmogorov complexity k is unable to predict the Busy Beaver sequence for machines larger than k, increasing intelligence in the sense of being able to predict more of the Busy Beaver sequence would require increased Kolmogorov complexity. But since even galactic civilizations at Kardashev Level III probably can't predict the Busy Beaver sequence very far, limits on this form of \"intelligence\" are not very limiting. For more on this, see my informal remarks here. 12. This is traditional, but also sensible, since entirely computer-based, deliberately designed intelligences seem likely to be more apt for further deliberate improvement than biological brains. Biological brains are composed of giant masses of undocumented spaghetti code running on tiny noisy filaments that require great feats of medical ingenuity to read, let alone edit. This point is widely appreciated, but of course it is not beyond dispute. distinctions. But such suggestions are legitimately disputable by anyone who thinks that a different set of definitions would be better suited to carving reality at its joints-to predicting what we will, in reality, actually observe to happen once some sort of smartish agency tries to invest in becoming smarterish. \n Issues to Factor Out Although we are ultimately interested only in the real-world results, I suggest that it will be productive theoretically-carve the issues at their natural joints-if we factor out for separate consideration issues of whether, for example, there might be an effective monitoring regime which could prevent an intelligence explosion, or whether the entire world economy will collapse due to global warming before then, and numerous other issues that don't seem to interact very strongly with the returns on cognitive investment qua cognitive investment.  13  In particular, I would suggest explicitly factoring out all considerations of \"What if an agent's preferences are such that it does not want to increase capability at the fastest rate it can achieve?\" As  Omohundro (2008)  and  Bostrom (2012)  point out, most possible preferences imply capability increase as an instrumental motive. If you want to build an intergalactic civilization full of sentient beings leading well-lived lives, you will want access to energy and matter. The same also holds true if you want to fill space with two-hundred-meter giant cheesecakes. In either case you will also have an instrumental goal of becoming smarter. Just as you can fulfill most goals better by having access to more material resources, you can also accomplish more by being better at cognitive problems-by being able to hit narrower targets in a search space. The space of all possible mind designs is vast  (Muehlhauser and Salamon 2012) , and there will always be some special case of an agent that chooses not to carry out any given deed  (Armstrong, forthcoming) . Given sufficient design competence, it should thus be possible to design an agent that doesn't prefer to ascend at the maximum possible ratethough expressing this within the AI's own preferences I would expect to be structurally nontrivial. Even so, we need to separately consider the question of how fast a rational agency could intelligence-explode if it were trying to self-improve as fast as possible. If the maximum rate of ascent is already inherently slow, then there is little point in constructing a special AI design that prefers not to improve faster than its programmers can verify. Policies are motivated by differentials of expected utility; there's no incentive to do any sort of action X intended to prevent Y unless we predict that Y might otherwise tend to follow assuming not-X. This requires us to set aside the proposed slowing factor and talk about what a rational agency might do if not slowed. Thus I suggest that initial investigations of the intelligence explosion should consider the achievable rate of return on cognitive reinvestment for a rational agency trying to self-improve as fast as possible, in the absence of any obstacles not already present in today's world.  14  This also reflects the hope that trying to tackle the posed Open Problem should not require expertise in Friendly AI or international politics in order to talk about the returns on cognitive investment qua investment, even if predicting actual real-world outcomes might (or might not) require some of these issues to be factored back in. \n AI Preferences: A Brief Summary of Core Theses Despite the above, it seems impossible not to at least briefly summarize some of the state of discussion on AI preferences-if someone believes that a sufficiently powerful AI, or one which is growing at a sufficiently higher rate than the rest of humanity and hence gaining unsurpassable advantages, is unavoidably bound to kill everyone, then they may have a hard time dispassionately considering and analyzing the potential growth curves. I have suggested that, in principle and in difficult practice, it should be possible to design a \"Friendly AI\" with programmer choice of the AI's preferences, and have the AI self-improve with sufficiently high fidelity to knowably keep these preferences stable. I also think it should be possible, in principle and in difficult practice, to convey the complicated information inherent in human preferences into an AI, and then apply further idealizations such as reflective equilibrium and ideal advisor theories  (Muehlhauser and  Williamson 2013) so as to arrive at an output which corresponds intuitively to the AI \"doing the right thing.\" See also  Yudkowsky (2008a) . On a larger scale the current state of discussion around these issues seems to revolve around four major theses: The Intelligence Explosion Thesis says that, due to recursive self-improvement, an AI can potentially grow in capability on a timescale that seems fast relative to human experience. This in turn implies that strategies which rely on humans reacting to and 14. That is, we might assume that people continue to protect their home computers with firewalls, for whatever that is worth. We should not assume that there is a giant and effective global monitoring organization devoted to stamping out any sign of self-improvement in AIs à la the Turing Police in William  Gibson's (1984)  Neuromancer. See also the sort of assumptions used in Robert  Freitas's (2000)  Some Limits to Global Ecophagy, wherein proposed limits on how fast the biosphere can be converted into nanomachines revolve around the assumption that there is a global monitoring agency looking for unexplained heat blooms, and that this will limit the allowed heat dissipation of nanomachines. restraining or punishing AIs are unlikely to be successful in the long run, and that what the first strongly self-improving AI prefers can end up mostly determining the final outcomes for Earth-originating intelligent life. (This subthesis is the entire topic of the current paper. One observes that the arguments surrounding the thesis are much more complex than the simple summary above would suggest. This is also true of the other three theses below.) The Orthogonality Thesis says that mind-design space is vast enough to contain minds with almost any sort of preferences. There exist instrumentally rational agents which pursue almost any utility function, and they are mostly stable under reflection. See Armstrong (forthcoming) and  Muehlhauser and Salamon (2012) . There are many strong arguments for the Orthogonality Thesis, but one of the strongest proceeds by construction: If it is possible to answer the purely epistemic question of which actions would lead to how many paperclips existing, then a paperclip-seeking agent is constructed by hooking up that answer to motor output. If it is very good at answering the epistemic question of which actions would result in great numbers of paperclips, then it will be a very instrumentally powerful agent.  15  The Complexity of Value Thesis says that human values are complex in the sense of having high algorithmic (Kolmogorov) complexity  (Yudkowsky 2011; Muehlhauser and Helm 2012) . Even idealized forms of human value, such as reflective equilibrium  (Rawls 1971)  or ideal advisor theories  (Rosati 1995 )-what we would want in the limit of infinite knowledge of the world, infinite thinking speeds, and perfect self-understanding, etc.-are predicted to still have high algorithmic complexity. This tends to follow from naturalistic theories of metaethics under which human preferences for happiness, freedom, growth, aesthetics, justice, etc., (see  Frankena [1973, chap. 5]  for one list of commonly stated terminal values) have no privileged reason to be readily reducible to each other or to anything else. The Complexity of Value Thesis is that to realize valuable outcomes, an AI must have complex information in its utility function; it also will 15. Such an agent will not modify itself to seek something else, because this would lead to fewer paperclips existing in the world, and its criteria for all actions including internal actions is the number of expected paperclips. It will not modify its utility function to have properties that humans would find more pleasing, because it does not already care about such metaproperties and is not committed to the belief that paperclips occupy a maximum of such properties; it is an expected paperclip maximizer, not an expected utility maximizer. Symmetrically, AIs which have been successfully constructed to start with \"nice\" preferences in their initial state will not throw away those nice preferences merely in order to confer any particular logical property on their utility function, unless they were already constructed to care about that property. not suffice to tell it to \"just make humans happy\" or any other simplified, compressed principle.  16  The Instrumental Convergence Thesis says that for most choices of a utility function, instrumentally rational agencies will predictably wish to obtain certain generic resources, such as matter and energy, and pursue certain generic strategies, such as not making code changes which alter their effective future preferences  (Omohundro 2008; Bostrom 2012) . Instrumental Convergence implies that an AI does not need to have specific terminal values calling for it to harm humans, in order for humans to be harmed. All this is another and quite different topic within the larger discussion of the intelligence explosion, compared to its microeconomics. Here I will only note that large returns on cognitive investment need not correspond to unavoidable horror scenarios so painful that we are forced to argue against them, nor to virtuous pro-science-andtechnology scenarios that virtuous people ought to affiliate with. For myself I would tend to view larger returns on cognitive reinvestment as corresponding to increased policydependent variance. And whatever the true values of the unseen variables, the question is not whether they sound like \"good news\" or \"bad news\"; the question is how we can improve outcomes as much as possible given those background settings. \n Microfoundations of Growth Consider the stance on the intelligence explosion thesis which says: \"I think we should expect that exponentially greater investments-of computing hardware, software programming effort, etc.-will only produce linear gains in real-world performance on cognitive tasks, since most search spaces are exponentially large. So the fruits of machine 16. The further arguments supporting the Complexity of Value suggest that even \"cosmopolitan\" or \"non-human-selfish\" outcomes have implicit specifications attached of high Kolmogorov complexity. Perhaps you would hold yourself to be satisfied with a future intergalactic civilization full of sentient beings happily interacting in ways you would find incomprehensible, even if none of them are you or human-derived. But an expected paperclip maximizer would fill the galaxies with paperclips instead. This is why expected paperclip maximizers are scary. intelligence reinvested into AI will only get logarithmic returns on each step, and the 'intelligence explosion' will peter out very quickly.\" Is this scenario plausible or implausible? Have we seen anything in the real worldmade any observation, ever-that should affect our estimate of its probability? (At this point, I would suggest that the serious reader turn away and take a moment to consider this question on their own before proceeding.) Some possibly relevant facts might be: • Investing exponentially more computing power into a constant chess-playing program produces linear increases in the depth of the chess-game tree that can be searched, which in turn seems to correspond to linear increases in Elo rating (where two opponents of a fixed relative Elo distance, regardless of absolute ratings, theoretically have a constant probability of losing or winning to each other). • Chess-playing algorithms have recently improved much faster than chess-playing hardware, particularly since chess-playing programs began to be open-sourced. Deep Blue ran on 11.8 billion floating-point operations per second and had an Elo rating of around 2,700; Deep Rybka 3 on a Intel Core 2 Quad 6600 has an Elo rating of 3,202 on 2.4 billion floating-point operations per second. 17 • It seems that in many important senses, humans get more than four times the real-world return on our intelligence compared to our chimpanzee cousins. This was achieved with Homo sapiens having roughly four times as much cortical volume and six times as much prefrontal cortex. 18 • Within the current human species, measured IQ is entangled with brain size; and this entanglement is around a 0.3 correlation in the variances, rather than, say, a doubling of brain size being required for each ten-point IQ increase.  19  • The various Moore's-like laws measuring computing technologies, operations per second, operations per dollar, disk space per dollar, and so on, are often said to have characteristic doubling times ranging from twelve months to three years; they are formulated so as to be exponential with respect to time. People have written papers 17. Score determined (plus or minus ∼23) by the Swedish Chess Computer Association based on 1,251 games played on the tournament level. 18. The obvious conclusion you might try to draw about hardware scaling is oversimplified and would be relevantly wrong. See section 3.1. 19. For entrants unfamiliar with modern psychological literature: Yes, there is a strong correlation g between almost all measures of cognitive ability, and IQ tests in turn are strongly correlated with this g factor and well correlated with many measurable life outcomes and performance measures. See the Cambridge Handbook of Intelligence (Sternberg and Kaufman 2011). questioning Moore's Law's validity (see, e.g.,  Tuomi [2002] ); and the Moore's-like law for serial processor speeds broke down in 2004. The original law first observed by Gordon Moore, over transistors per square centimeter, has remained on track. • Intel has invested exponentially more researcher-hours and inflation-adjusted money to invent the technology and build the manufacturing plants for successive generations of CPUs. But the CPUs themselves are increasing exponentially in transistor operations per second, not linearly; and the computer-power doubling time is shorter (that is, the exponent is higher) than that of the increasing investment cost.  20  • The amount of evolutionary time (a proxy measure of cumulative selection pressure and evolutionary optimization) which produced noteworthy changes during human and hominid evolution does not seem to reveal exponentially greater amounts of time invested. It did not require ten times as long to go from Homo erectus to Homo sapiens, as from Australopithecus to Homo erectus.  21  • World economic output is roughly exponential and increases faster than population growth, which is roughly consistent with exponentially increasing investments producing exponentially increasing returns. That is, roughly linear (but with multiplication factor k > 1) returns on investment. On a larger timescale, world-historical economic output can be characterized as a sequence of exponential modes  (Hanson 1998b) . Total human economic output was also growing exponentially in AD 1600 or 2000 BC, but with smaller exponents and much longer doubling times. • Scientific output in \"total papers written\" tends to grow exponentially with a short doubling time, both globally (around twenty-seven years  [NSB 2012, chap. 5] ) and within any given field. But it seems extremely questionable whether there has been more global change from 1970 to 2010 than from 1930 to 1970. (For readers who have heard relatively more about \"accelerating change\" than about \"the Great Stagnation\": the claim is that total-factor productivity growth in, e.g., the United States dropped from 0.75% per annum before the 1970s to 0.25% thereafter  [Cowen 2011 ].) A true cynic might claim that, in many fields, exponentially greater 20. As Carl Shulman observes, Intel does not employ 343 million people. 21. One might ask in reply whether Homo erectus is being singled out on the basis of being distant enough in time to have its own species name, rather than by any prior measure of cognitive ability. This issue is taken up at much greater length in section 3.6. investment in science is yielding a roughly constant amount of annual progresssublogarithmic returns! 22 • This graph  (Silver 2012)  shows how many books were authored in Europe as a function of time; after the invention of the printing press, the graph jumps in a sharp, faster-than-exponential upward surge. • All technological progress in known history has been carried out by essentially constant human brain architectures. There are theses about continuing human evolution over the past ten thousand years, but all such changes are nowhere near the scale of altering \"You have a brain that's more or less 1,250 cubic centimeters of dendrites and axons, wired into a prefrontal cortex, a visual cortex, a thalamus, and so on.\" It has not required much larger brains, or much greater total cumulative selection pressures, to support the continuing production of more sophisticated technologies and sciences over the human regime. • The amount of complex order per unit time created by a human engineer is completely off the scale compared to the amount of complex order per unit time created by natural selection within a species. A single mutation conveying a 3% fitness advantage would be expected to take 768 generations to rise to fixation through a sexually reproducing population of a hundred thousand members. A computer programmer can design new complex mechanisms with hundreds of interoperating parts over the course of a day or an hour. In turn, the amount of complex order per unit time created by natural selection is completely off the scale for Earth before the dawn of life. A graph of \"order created per unit time\" during Earth's history would contain two discontinuities representing the dawn of fundamentally different optimization processes. The list of observations above might give you the impression that it could go either way-that some things are exponential and some things aren't. Worse, it might look like an invitation to decide your preferred beliefs about AI self-improvement as a matter of emotional appeal or fleeting intuition, and then decide that any of the above cases which behave similarly to how you think AI self-improvement should behave, are the natural historical examples we should consult to determine the outcome of AI. For example, clearly the advent of self-improving AI seems most similar to other economic 22. I am in fact such a true cynic and I suspect that social factors dilute average contributions around as fast as new researchers can be added. A less cynical hypothesis would be that earlier science is easier, and later science grows more difficult at roughly the same rate that scientific output scales with more researchers being added. speedups like the invention of agriculture.  23  Or obviously it's analogous to other foundational changes in the production of complex order, such as human intelligence or self-replicating life.  24  Or self-evidently the whole foofaraw is analogous to the panic over the end of the Mayan calendar in 2012 since it belongs in the reference class of \"supposed big future events that haven't been observed.\"  25  For more on the problem of \"reference class tennis,\" see section 2.1. It seems to me that the real lesson to be derived from the length of the above list is that we shouldn't expect some single grand law about whether you get superexponential, exponential, linear, logarithmic, or constant returns on cognitive investments. The cases above have different behaviors; they are not all conforming to a single Grand Growth Rule. It's likewise not the case that Reality proceeded by randomly drawing a curve type from a barrel to assign to each of these scenarios, and the curve type of \"AI self-improvement\" will be independently sampled with replacement from the same barrel. So it likewise doesn't seem valid to argue about how likely it is that someone's personal favorite curve type gets drawn by trumpeting historical cases of that curve type, thereby proving that it's more frequent within the Curve Type Barrel and more likely to be randomly drawn. Most of the processes cited above yielded fairly regular behavior over time. Meaning that the attached curve was actually characteristic of that process's causal mechanics, and a predictable feature of those mechanics, rather than being assigned and reassigned at random. Anyone who throws up their hands and says, \"It's all unknowable!\" may also be scoring fewer predictive points than they could. These differently behaving cases are not competing arguments about how a single grand curve of cognitive investment has previously operated. They are all simultaneously true, and hence they must be telling us different facts about growth curves-telling us about different domains of a multivariate growth function-advising us of many compatible truths about how intelligence and real-world power vary with different kinds of cognitive investments.  26  23. See  Hanson (2008a) . 24. See  Yudkowsky (2008b Yudkowsky ( , 2008c Yudkowsky ( , 2008d . 25. See, e.g., this post in an online discussion. 26. Reality itself is always perfectly consistent-only maps can be in conflict, not the territory. Under the Bayesian definition of evidence, \"strong evidence\" is just that sort of evidence that we almost never see on more than one side of an argument. Unless you've made a mistake somewhere, you should almost never see extreme likelihood ratios pointing in different directions. Thus it's not possible that the facts listed are all \"strong\" arguments, about the same variable, pointing in different directions. Rather than selecting one particular historical curve to anoint as characteristic of the intelligence explosion, it might be possible to build an underlying causal model, one which would be compatible with all these separate facts. I would propose that we should be trying to formulate a microfoundational model which, rather than just generalizing over surface regularities, tries to describe underlying causal processes and returns on particular types of cognitive investment. For example, rather than just talking about how chess programs have improved over time, we might try to describe how chess programs improve as a function of computing resources plus the cumulative time that human engineers spend tweaking the algorithms. Then in turn we might say that human engineers have some particular intelligence or optimization power, which is different from the optimization power of a chimpanzee or the processes of natural selection. The process of building these causal models would hopefully let us arrive at a more realistic picture-one compatible with the many different growth curves observed in different historical situations. \n The Outside View versus the Lucas Critique A fundamental tension in the so-far-informal debates on intelligence explosion has been the rough degree of abstraction that is trustworthy and useful when modeling these future events. The first time I happened to occupy the same physical room as Ray Kurzweil, I asked him why his graph of Moore's Law showed the events of \"a $1,000 computer is as powerful as a human brain,\" \"a $1,000 computer is a thousand times as powerful as a human brain,\" and \"a $1,000 computer is a billion times as powerful as a human brain,\" all following the same historical trend of Moore's Law.  27  I asked, did it really make sense to continue extrapolating the humanly observed version of Moore's Law past the point where there were putatively minds with a billion times as much computing power? Kurzweil 2001 replied that the existence of machine superintelligence was exactly what would provide the fuel for Moore's Law to continue and make it possible to keep developing the required technologies. In other words, Kurzweil 2001 regarded Moore's Law as the primary phenomenon and considered machine superintelligence a secondary 27. The same chart showed allegedly \"human-level computing power\" as the threshold of predicted AI, which is a methodology I strongly disagree with, but I didn't want to argue with that part at the time. I've looked around in Google Images for the exact chart but didn't find it; Wikipedia does cite similar predictions as having been made in The Age of Spiritual Machines  (Kurzweil 1999 ), but Wikipedia's cited timelines are shorter term than I remember. phenomenon which ought to assume whatever shape was required to keep the primary phenomenon on track.  28  You could even imagine arguing (though 28. I attach a subscript by year because (1) Kurzweil was replying on the spot so it is not fair to treat his off-the-cuff response as a permanent feature of his personality and (2)  Sandberg (2010)  suggests that Kurzweil has changed his position since then. 29. There are over two billion transistors in the largest Core i7 processor. At this point human engineering requires computer assistance. 30. One can imagine that Intel may have balanced the growth rate of its research investments to follow industry expectations for Moore's Law, even as a much more irregular underlying difficulty curve became steeper or shallower. This hypothesis doesn't seem inherently untestable-someone at Intel would actually have had to make those sorts of decisions-but it's not obvious to me how to check it on previously gathered, easily accessed data. Or to put it more plainly, the fully-as-naive extrapolation in the other direction would be, \"Given human researchers of constant speed, computing speeds double every 18 months. So if the researchers are running on computers themselves, we should expect computing speeds to double in 18 months, then double again in 9 physical months (or 18 subjective months for the 2x-speed researchers), then double again in 4.5 physical months, and finally reach infinity after a total of 36 months.\" If humans accumulate subjective time at a constant rate x = t, and we observe that computer speeds increase as a Moore's-Law exponential function of subjective time y = e x , then when subjective time increases at the rate of current computer speeds we get the differential equation y = e y whose solution has computer speeds increasing hyperbolically, going to infinity after finite time.  31  (See, e.g., the model of  Moravec [1999] .) In real life, we might not believe this as a quantitative estimate. We might not believe that in real life such a curve would have, even roughly, a hyperbolic shape before it started hitting (high) physical bounds. But at the same time, we might in real life believe that research ought to go substantially faster if the researchers could reinvest the fruits of their labor into their own cognitive speeds-that we are seeing an important hint buried within this argument, even if its details are wrong. We could believe as a qualitative prediction that \"if computer chips are following Moore's Law right now with human researchers running at constant neural processing speeds, then in the hypothetical scenario where the researchers are running on computers, we should see a new Moore's Law bounded far below by the previous one.\" You might say something like, \"Show me a reasonable model of how difficult it is to build chips as a function of knowledge, and how knowledge accumulates over subjective time, and you'll get a hyperexponential explosion out of Moore's Law once the researchers are running on computers. Conversely, if you give me a regular curve of increasing difficulty which averts an intelligence explosion, it will falsely retrodict that human engineers should only be able to get subexponential improvements out of computer technology. And of course it would be unreasonable-a specific unsupported miraculous irregularity of the curve-for making chips to suddenly get much more difficult to build, coincidentally exactly as AIs started doing research. The difficulty curve might shift upward at some random later point, but there'd still be a bonanza from whatever improvement was available up until then.\" In turn, that reply gets us into a rather thorny meta-level issue: which you speak? Who measures knowledge? These are all made-up quantities with no rigorous basis in reality. What we do have solid observations of is the number of transistors on a computer chip, per year. So I'm going to project that extremely regular curve out into the future and extrapolate from there. The rest of this is sheer, loose speculation. Who knows how many other possible supposed \"underlying\" curves, besides this \"knowledge\" and \"difficulty\" business, would give entirely different answers? To which one might reply: B: Seriously? Let's consider an extreme case. Neurons spike around 2-200 times per second, and axons and dendrites transmit neural signals at 1-100 meters per second, less than a millionth of the speed of light. Even the heat dissipated by each neural operation is around six orders of magnitude above the thermodynamic minimum at room temperature. 32 Hence it should be physically possible to speed up \"internal\" thinking (which doesn't require \"waiting on the external world\") by at least six orders of magnitude without resorting to smaller, colder, reversible, or quantum computers. Suppose we were dealing with minds running a million times as fast as a human, at which rate they could do a year of internal thinking in thirty-one seconds, such that the total subjective time from the birth of Socrates to the death of Turing would pass in 20.9 hours. Do you still think the best estimate for how long it would take them to produce their next generation of computing hardware would be 1.5 orbits of the Earth around the Sun? Two well-known epistemological stances, with which the respective proponents of these positions could identify their arguments, would be the outside view and the Lucas critique. 32. The brain as a whole organ dissipates around 20 joules per second, or 20 watts. The minimum energy required for a one-bit irreversible operation (as a function of temperature T ) is kT ln(2), where k = 1.38 • 10 23 joules/kelvin is Boltzmann's constant, and ln(  2 ) is the natural log of 2 (around 0.7). Three hundred kelvin is 27 • C or 80 • F. Thus under ideal circumstances 20 watts of heat dissipation corresponds to 7 • 10 21 irreversible binary operations per second at room temperature. The brain can be approximated as having 10 14 synapses. I found data on average synaptic activations per second hard to come by, with different sources giving numbers from 10 activations per second to 0.003 activations/second (not all dendrites must activate to trigger a spike, and not all neurons are highly active at any given time). If we approximate the brain as having 10 14 synapses activating on the order of once per second on average, this would allow ∼10 2 irreversible operations per synaptic activation after a 10 6 -fold speedup. (Note that since each traveling impulse of electrochemical activation requires many chemical ions to be pumped back across the neuronal membrane afterward to reset it, total distance traveled by neural impulses is a more natural measure of expended biological energy than total activations. No similar rule would hold for photons traveling through optical fibers.) The \"outside view\"  (Kahneman and Lovallo 1993 ) is a term from the heuristics and biases program in experimental psychology. A number of experiments show that if you ask subjects for estimates of, say, when they will complete their Christmas shopping, the right question to ask is, \"When did you finish your Christmas shopping last year?\" and not, \"How long do you think it will take you to finish your Christmas shopping?\" The latter estimates tend to be vastly over-optimistic, and the former rather more realistic. In fact, as subjects are asked to make their estimates using more detail-visualize where, when, and how they will do their Christmas shopping-their estimates become more optimistic, and less accurate. Similar results show that the actual planners and implementers of a project, who have full acquaintance with the internal details, are often much more optimistic and much less accurate in their estimates compared to experienced outsiders who have relevant experience of similar projects but don't know internal details. This is sometimes called the dichotomy of the inside view versus the outside view. The \"inside view\" is the estimate that takes into account all the details, and the \"outside view\" is the very rough estimate that would be made by comparing your project to other roughly similar projects without considering any special reasons why this project might be different. The Lucas critique  (Lucas 1976 ) in economics was written up in 1976 when \"stagflation\"-simultaneously high inflation and unemployment-was becoming a problem in the United States. Robert Lucas's concrete point was that the Phillips curve trading off unemployment and inflation had been observed at a time when the Federal Reserve was trying to moderate inflation. When the Federal Reserve gave up on moderating inflation in order to drive down unemployment to an even lower level, employers and employees adjusted their long-term expectations to take into account continuing inflation, and the Phillips curve shifted. Lucas's larger and meta-level point was that the previously observed Phillips curve wasn't fundamental enough to be structurally invariant with respect to Federal Reserve policy-the concepts of inflation and unemployment weren't deep enough to describe elementary things that would remain stable even as Federal Reserve policy shifted. A very succinct summary appears in  Wikipedia (2013) : The Lucas critique suggests that if we want to predict the effect of a policy experiment, we should model the \"deep parameters\" (relating to preferences, technology and resource constraints) that are assumed to govern individual behavior; so called \"microfoundations.\" If these models can account for observed empirical regularities, we can then predict what individuals will do, taking into account the change in policy, and then aggregate the individual decisions to calculate the macroeconomic effects of the policy change. The main explicit proponent of the outside view in the intelligence explosion debate is Robin Hanson, who also proposes that an appropriate reference class into which to place the \"Singularity\"-a term not specific to the intelligence explosion but sometimes including it-would be the reference class of major economic transitions resulting in substantially higher exponents of exponential growth. From  Hanson's (2008a)   Taking a long historical long view, we see steady total growth rates punctuated by rare transitions when new faster growth modes appeared with little warning. We know of perhaps four such \"singularities\": animal brains (∼600 MYA), humans (∼2 MYA), farming (∼10 KYA), and industry (∼0.2 KYA). The statistics of previous transitions suggest we are perhaps overdue for another one, and would be substantially overdue in a century. The next transition would change the growth rate rather than capabilities directly, would take a few years at most, and the new doubling time would be a week to a month. More on this analysis can be found in  Hanson (1998b) . The original blog post concludes: Excess inside viewing usually continues even after folks are warned that outside viewing works better; after all, inside viewing better shows off inside knowledge and abilities. People usually justify this via reasons why the current case is exceptional. (Remember how all the old rules didn't apply to the new dotcom economy?) So expect to hear excuses why the next singularity is also It is easy, way too easy, to generate new mechanisms, accounts, theories, and abstractions. To see if such things are useful, we need to vet them, and that is easiest \"nearby,\" where we know a lot. When we want to deal with or understand things \"far,\" where we know little, we have little choice other than to rely on mechanisms, theories, and concepts that have worked well near. Far is just the wrong place to try new things. There are a bazillion possible abstractions we could apply to the world. For each abstraction, the question is not whether one can divide up the world that way, but whether it \"carves nature at its joints,\" giving useful insight not easily gained via other abstractions. We should be wary of inventing new abstractions just to make sense of things far; we should insist they first show their value nearby. The lesson of the outside view pushes us to use abstractions and curves that are clearly empirically measurable, and to beware inventing new abstractions that we can't see directly. The lesson of the Lucas critique pushes us to look for abstractions deep enough to describe growth curves that would be stable in the face of minds improving in speed, size, and software quality. You can see how this plays out in the tension between \"Let's predict computer speeds using this very well-measured curve for Moore's Law over time-where the heck is all this other stuff coming from?\" versus \"But almost any reasonable causal model that describes the role of human thinking and engineering in producing better computer chips, ought to predict that Moore's Law would speed up once computer-based AIs were carrying out all the research!\" It would be unfair to use my passing exchange with Kurzweil as a model of the debate between myself and Hanson. Still, I did feel that the basic disagreement came down to a similar tension-that Hanson kept raising a skeptical and unmoved eyebrow at the wildeyed, empirically unvalidated, complicated abstractions which, from my perspective, constituted my attempt to put any sort of microfoundations under surface curves that couldn't possibly remain stable. Hanson's overall prototype for visualizing the future was an economic society of ems, software emulations of scanned human brains. It would then be possible to turn capital inputs (computer hardware) into skilled labor (copied ems) almost immediately. This was Hanson's explanation for how the em economy could follow the \"same trend\" as past economic speedups, to a world economy that doubled every year or month (vs. a roughly fifteen-year doubling time at present  [Hanson 1998b] ). I thought that the idea of copying human-equivalent minds missed almost every potentially interesting aspect of the intelligence explosion, such as faster brains, larger brains, or above all better-designed brains, all of which seemed liable to have far greater effects than increasing the quantity of workers. Why? That is, if you can invest a given amount of computing power in more brains, faster brains, larger brains, or improving brain algorithms, why think that the return on investment would be significantly higher in one of the latter three cases? A more detailed reply is given in section 3, but in quick summary: There's a saying in software development, \"Nine women can't have a baby in one month,\" meaning that you can't get the output of ten people working for ten years by hiring a hundred people to work for one year, or more generally, that working time scales better than the number of people, ceteris paribus. It's also a general truth of computer science that fast processors can simulate parallel processors but not always the other way around. Thus we'd expect the returns on speed to be higher than the returns on quantity. We have little solid data on how human intelligence scales with added neurons and constant software. Brain size does vary between humans and this variance correlates by about 0.3 with g (McDaniel 2005), but there are reams of probable confounders, such as childhood nutrition. Humans have around four times the brain volume of chimpanzees, but the difference between us is probably mostly brain-level cognitive algorithms.  33  It is a general truth of computer science that if you take one processing unit and split it up into ten parts with limited intercommunication bandwidth, they can do no better than the original on any problem, and will do considerably worse on many problems. Similarly we might expect that, for most intellectual problems, putting on ten times as many researchers running human software scaled down to one-fifth the brain size would probably not be a net gain, and that, for many intellectual problems, researchers with four times the brain size would probably be a significantly greater gain than adding four times as many researchers.  34  Trying to say how intelligence and problem-solving ability scale with improved cognitive algorithms is even harder to relate to observation. In any computer-based field where surface capabilities are visibly improving, it is usually true that you are better off with modern algorithms and a computer from ten years earlier, compared to a modern computer and the algorithms from ten years earlier. This is definitely true in computer chess, even though the net efforts put in by chess-program enthusiasts to create better programs are small compared to the vast effort Intel puts into creating better computer chips every year. But this observation only conveys a small fraction of the idea that you can't match a human's intellectual output using any number of chimpanzees. Informally, it looks to me like quantity < (size, speed) < quality when it comes to minds. Hanson's scenario in which all investments went into increasing the mere quantity of ems-and this was a good estimate of the total impact of an intelligence explosion-33. If it were possible to create a human just by scaling up an Australopithecus by a factor of four, the evolutionary path from Australopithecus to us would have been much shorter. 34. Said with considerable handwaving. But do you really think that's false? seemed to imply that the returns on investment from larger brains, faster thinking, and improved brain designs could all be neglected, which implied that the returns from such investments were relatively low.  35  Whereas it seemed to me that any reasonable microfoundations which were compatible with prior observation-which didn't retrodict that a human should be intellectually replaceable by ten chimpanzees-should imply that quantity of labor wouldn't be the dominating factor. Nonfalsified growth curves ought to say that, given an amount of computing power which you could invest in more minds, faster minds, larger minds, or better-designed minds, you would invest in one of the latter three. We don't invest in larger human brains because that's impossible with current technologywe can't just hire a researcher with three times the cranial volume, we can only throw more warm bodies at the problem. If that investment avenue suddenly became available . . . it would probably make quite a large difference, pragmatically speaking. I was happy to concede that my model only made vague qualitative predictions-I didn't think I had enough data to make quantitative predictions like Hanson's estimates of future economic doubling times. But qualitatively I thought it obvious that all these hard-to-estimate contributions from faster brains, larger brains, and improved underlying cognitive algorithms were all pointing along the same rough vector, namely \"way up.\" Meaning that Hanson's estimates, sticking to extrapolated curves of well-observed quantities, would be predictably biased way down. Whereas from Hanson's perspective, this was all wild-eyed unverified speculation, and he was sticking to analyzing ems because we had a great deal of data about how human minds worked and no way to solidly ground all these new abstractions I was hypothesizing. Aside from the Lucas critique, the other major problem I have with the \"outside view\" is that everyone who uses it seems to come up with a different reference class and a different answer. To Ray Kurzweil, the obvious reference class for \"the Singularity\" is Moore's Law as it has operated over recent history, not Hanson's comparison to agriculture. In this post an online discussant of these topics places the \"Singularity\" into the reference class \"beliefs in coming of a new world\" which has \"a 0% success rate\" . . . explicitly terming this the proper \"outside view\" of the situation using \"reference class forecasting,\" and castigating anyone who tried to give a different answer as having used an \"inside view.\" For my response to all this at greater length, see \"'Outside View!' 35. Robin Hanson replied to a draft of this paper: \"The fact that I built a formal model that excluded these factors doesn't mean I think such effects are so small as to be negligible. Not only is it reasonable to build models that neglect important factors, it is usually impossible not to do so.\" This is surely true; nonetheless, I think that in this case the result was a predictable directional bias. as Conversation-Halter\"  (Yudkowsky 2010) . The gist of my reply was that the outside view has been experimentally demonstrated to beat the inside view for software projects that are similar to previous software projects, and for this year's Christmas shopping, which is highly similar to last year's Christmas shopping. The outside view would be expected to work less well on a new thing that is less similar to the old things than all the old things were similar to each other-especially when you try to extrapolate from one kind of causal system to a very different causal system. And one major sign of trying to extrapolate across too large a gap is when everyone comes up with a different \"obvious\" reference class. Of course it also often happens that disputants think different microfoundationsdifferent causal models of reality-are \"obviously\" appropriate. But then I have some idea of how to zoom in on hypothesized causes, assess their simplicity and regularity, and figure out how to check them against available evidence. I don't know what to do after two people take different reference classes and come up with different outside views both of which we ought to just accept. My experience is that people end up doing the equivalent of saying, \"I'm taking my reference class and going home.\" A final problem I have with many cases of \"reference class forecasting\" is that-in addition to everyone coming up with a different reference class-their final answers often seem more specific than I think our state of knowledge should allow. I don't think you should be able to tell me that the next major growth mode will have a doubling time of between a month and a year. The alleged outside viewer claims to know too much, once they stake their all on a single preferred reference class. But then what I have just said is an argument for enforced humility-\"I don't know, so you can't know either!\"-and is automatically suspect on those grounds. It must be fully conceded and advised that complicated models are hard to fit to limited data, and that when postulating curves which are hard to observe directly or nail down with precision, there is a great deal of room for things to go wrong. It does not follow that \"reference class forecasting\" is a good solution, or even the merely best solution. \n Some Defenses of a Model of Hard Takeoff If only for reasons of concreteness, it seems appropriate to summarize my own stance on the intelligence explosion, not just abstractly discuss how to formalize such stances in general.  36  In very concrete terms-leaving out all the abstract principles, microfounda-tions, and the fundamental question of \"What do you think you know and how do you think you know it?\"-a \"typical\" intelligence explosion event as envisioned by Eliezer Yudkowsky might run something like this: Some sort of AI project run by a hedge fund, academia, Google, 37 or a government, advances to a sufficiently developed level (see section 3.10) that it starts a string of selfimprovements that is sustained and does not level off. This cascade of self-improvements might start due to a basic breakthrough by the researchers which enables the AI to understand and redesign more of its own cognitive algorithms. Or a soup of self-modifying systems governed by a fitness evaluator, after undergoing some smaller cascades of selfimprovements, might finally begin a cascade which does not level off. Or somebody with money might throw an unprecedented amount of computing power at AI algorithms which don't entirely fail to scale. Once this AI started on a sustained path of intelligence explosion, there would follow some period of time while the AI was actively self-improving, and perhaps obtaining additional resources, but hadn't yet reached a cognitive level worthy of being called \"superintelligence.\" This time period might be months or years, 38 or days or seconds.  39  I am greatly uncertain of what signs of competence the AI might give over this time, or how its builders or other parties might react to this; but for purposes of intelligence explosion microeconomics, we should temporarily factor out these questions and assume the AI's growth is not being deliberately impeded by any particular agency. At some point the AI would reach the point where it could solve the protein structure prediction problem and build nanotechnology-or figure out how to control atomicrevolved entirely around equations consisting of upper-case Greek letters. During the Q&A, somebody politely asked the speaker if he could give a concrete example. The speaker thought for a moment and wrote a new set of equations, only this time all the Greek letters were in lowercase. I try not to be that guy. 37. Larry Page has publicly said that he is specifically interested in \"real AI\" (Artificial General Intelligence), and some of the researchers in the field are funded by Google. So far as I know, this is still at the level of blue-sky work on basic algorithms and not an attempt to birth The Google in the next five years, but it still seems worth mentioning Google specifically. 38. Any particular AI's characteristic growth path might require centuries to superintelligence-this could conceivably be true even of some modern AIs which are not showing impressive progress-but such AIs end up being irrelevant; some other project which starts later will reach superintelligence first. Unless all AI development pathways require centuries, the surrounding civilization will continue flipping through the deck of AI development projects until it turns up a faster-developing AI. 39. Considering that current CPUs operate at serial speeds of billions of operations per second and that human neurons require at least a millisecond to recover from firing a spike, seconds are potentially long stretches of time for machine intelligences-a second has great serial depth, allowing many causal events to happen in sequence. See section 3.3. force microscopes to create new tool tips that could be used to build small nanostructures which could build more nanostructures-or perhaps follow some smarter and faster route to rapid infrastructure. An AI that goes past this point can be considered to have reached a threshold of great material capability. From this would probably follow cognitive superintelligence (if not already present); vast computing resources could be quickly accessed to further scale cognitive algorithms. The further growth trajectory beyond molecular nanotechnology seems mostly irrele- What sort of general beliefs does this concrete scenario of \"hard takeoff \" imply about returns on cognitive reinvestment? It supposes that: • An AI can get major gains rather than minor gains by doing better computer science than its human inventors. • More generally, it's being supposed that an AI can achieve large gains through better use of computing power it already has, or using only processing power it can rent or otherwise obtain on short timescales-in particular, without setting up new chip factories or doing anything else which would involve a long, unavoidable delay. 40 • An AI can continue reinvesting these gains until it has a huge cognitive problemsolving advantage over humans. • This cognitive superintelligence can echo back to tremendous real-world capabilities by solving the protein folding problem, or doing something else even more clever (see section 3.11), starting from the then-existing human technological base. Even more abstractly, this says that AI self-improvement can operate with k 1 and a fast timescale of reinvestment: \"prompt supercritical.\" But why believe that? (A question like this is conversationally difficult to answer since different people may think that different parts of the scenario sound most questionable. Also, although I think there is a simple idea at the core, when people ask probing questions the resulting conversations are often much more complicated.  41  Please forgive my answer if it doesn't immediately address the questions at the top of your own priority list; different people have different lists.) I would start out by saying that the evolutionary history of hominid intelligence doesn't show any signs of diminishing returns-there's no sign that evolution took ten times as long to produce each successive marginal improvement of hominid brains. (Yes, this is hard to quantify, but even so, the anthropological record doesn't look like it should look if there were significantly diminishing returns. See section 3.6.) We have a fairly good mathematical grasp on the processes of evolution and we can well approximate some of the optimization pressures involved; we can say with authority that, in a number of important senses, evolution is extremely inefficient  (Yudkowsky 2007) . And yet evolution was able to get significant cognitive returns on point mutations, random recombination, and non-foresightful hill climbing of genetically encoded brain architectures. Furthermore, the character of evolution as an optimization process was essentially constant over the course of mammalian evolution-there were no truly fundamental innovations, like the evolutionary invention of sex and sexual recombination, over the relevant timespan. So if a steady pressure from natural selection realized significant fitness returns from optimizing the intelligence of hominids, then researchers getting smarter at optimizing themselves ought to go FOOM. The \"fully naive\" argument from Moore's Law folded in on itself asks, \"If computing power is doubling every eighteen months, what happens when computers are doing the research?\" I don't think this scenario is actually important in practice, mostly because I 41. \"The basic idea is simple, but refuting objections can require much more complicated conversations\" is not an alarming state of affairs with respect to Occam's Razor; it is common even for correct theories. For example, the core idea of natural selection was much simpler than the conversations that were required to refute simple-sounding objections to it. The added conversational complexity is often carried in by invisible presuppositions of the objection. expect returns on cognitive algorithms to dominate returns on speed. (The dominant species on the planet is not the one that evolved the fastest neurons.) Nonetheless, if the difficulty curve of Moore's Law was such that humans could climb it at a steady pace, then accelerating researchers, researchers whose speed was itself tied to Moore's Law, should arguably be expected to (from our perspective) go FOOM. The returns on pure speed might be comparatively smaller-sped-up humans would not constitute superintelligences. (For more on returns on pure speed, see section 3.3.) However, faster minds are easier to imagine than smarter minds, and that makes the \"folded-in Moore's Law\" a simpler illustration of the general idea of folding-in. Natural selection seems to have climbed a linear or moderately superlinear growth curve of cumulative optimization pressure in versus intelligence out. To \"fold in\" this curve we consider a scenario where the inherent difficulty of the problem is as before, but instead of minds being improved from the outside by a steady pressure of natural selection, the current optimization power of a mind is determining the speed at which the curve of \"cumulative optimization power in\" is being traversed. Given the previously described characteristics of the non-folded-in curve, any particular selfimproving agency, without outside help, should either bottleneck in the lower parts of the curve (if it is not smart enough to make improvements that are significant compared to those of long-term cumulative evolution), or else go FOOM (if its initial intelligence is sufficiently high to start climbing) and then climb even faster. We should see a \"bottleneck or breakthrough\" dichotomy: Any particular self-improving mind either \"bottlenecks\" without outside help, like all current AIs, or \"breaks through\" into a fast intelligence explosion.  42  There would be a border between these alternatives containing minds which are seemingly making steady, slow, significant progress at self-improvement; but this border need not be wide, and any such mind would be steadily moving toward the FOOM region of the curve. See section 3.10. Some amount of my confidence in \"AI go FOOM\" scenarios also comes from cognitive science (e.g., the study of heuristics and biases) suggesting that humans are, in practice, very far short of optimal design. The broad state of cognitive psychology suggests that \"Most humans cannot multiply two three-digit numbers in their heads\" is not an unfair indictment-we really are that poorly designed along many dimensions.  43  42. At least the first part of this prediction seems to be coming true. On a higher level of abstraction, this is saying that there exists great visible headroom for improvement over the human level of intelligence. It's extraordinary that humans manage to play chess using visual recognition systems which evolved to distinguish tigers on the savanna; amazing that we can use brains which evolved to make bows and arrows to program computers; and downright incredible that we can invent new computer science and new cognitive algorithms using brains mostly adapted to modeling and outwitting other humans. But by the standards of computer-based minds that can redesign themselves as required and run error-free algorithms with a billion steps of serial depth, we probably aren't thinking very efficiently. (See section 3.5.) Thus we have specific reason to suspect that cognitive algorithms can be improved beyond the human level-that human brain algorithms aren't any closer to optimal software than human neurons are close to the physical limits of hardware. Even without the embarrassing news from experimental psychology, we could still observe that the inherent difficulty curve for building intelligences has no known reason to possess the specific irregularity of curving sharply upward just after accessing human equivalence. But we also have specific reason to suspect that mind designs can be substantially improved beyond the human level. That is a rough summary of what I consider the core idea behind my belief that returns on cognitive reinvestments are probably large. You could call this summary the \"naive\" view of returns on improving cognitive algorithms, by analogy with the naive theory of how to fold in Moore's Law. We can drill down and ask more sophisticated questions, but it's worth remembering that when done correctly, more sophisticated analysis quite often says that the naive answer is right. Somebody who' d never studied General Relativity as a formal theory of gravitation might naively expect that jumping off a tall cliff would make you fall down and go splat; and in this case it turns out that the sophisticated prediction agrees with the naive one. Thus, keeping in mind that we are not obligated to arrive at any impressively nonobvious \"conclusions,\" let us consider some nonobvious subtleties of argument. In the next subsections we will consider: 1. What the fossil record actually tells us about returns on brain size, given that most of the difference between Homo sapiens and Australopithecus was probably improved algorithms. 2. How to divide credit for the human-chimpanzee performance gap between \"humans are individually smarter than chimpanzees\" and \"the hominid transition in- 11. The enhanced importance of unknown unknowns in intelligence explosion scenarios, since a smarter-than-human intelligence will selectively seek out and exploit useful possibilities implied by flaws or gaps in our current knowledge. I would finally remark that going into depth on the pro-FOOM stance should not operate to prejudice the reader in favor of other stances. Defending only one stance at great length may make it look like a huge edifice of argument that could potentially topple, whereas other viewpoints such as \"A collective of interacting AIs will have k ≈ But of course (so far as the author believes) such other outcomes would be even harder to defend in depth.  44  Every argument for the intelligence explosion is, when negated, an argument for an intelligence nonexplosion. To the extent the negation of each argument here might sound less than perfectly plausible, other possible outcomes would not sound any more plausible when argued to this depth of point and counterpoint. \n Returns on Brain Size Many cases where we'd like to reason from historical returns on cognitive investment are complicated by unfortunately narrow data. All the most impressive cognitive returns are from a single species, namely Homo sapiens. Humans have brains around four times the size of chimpanzees' . . . but this tells us very little because most of the differences between humans and chimps are almost certainly algorithmic. If just taking an Australopithecus brain and scaling it up by a factor of four produced a human, the evolutionary road from Australopithecus to Homo sapiens would probably have been much shorter; simple factors like the size of an organ can change quickly in the face of strong evolutionary pressures. Based on historical observation, we can say with authority that going from Australopithecus to Homo sapiens did not in fact require a hundredfold increase in brain size plus improved algorithms-we can refute the assertion that even after taking into account five million years of evolving better cognitive algorithms, a hundredfold increase in hardware was required to accommodate the new algorithms. This may not sound like much, but it does argue against models which block an intelligence explosion by always requiring exponentially increasing hardware for linear cognitive gains.  45  A nonobvious further implication of observed history is that improvements in cognitive algorithms along the way to Homo sapiens must have increased rather than decreased 44. Robin Hanson has defended the \"global exponential economic speedup\" thesis at moderate length, in the Yudkowsky-Hanson AI-Foom debate and in several papers, and the reader is invited to explore these. I am not aware of anyone who has defended an \"intelligence fizzle\" seriously and at great length, but this of course may reflect a selection effect. If you believe nothing interesting will happen, you don't believe there's anything worth writing a paper on. 45. I'm pretty sure I've heard this argued several times, but unfortunately I neglected to save the references; please contribute a reference if you've got one. Obviously, the speakers I remember were using this argument to confidently dismiss the possibility of superhuman machine intelligence, and it did not occur to them that the same argument might also apply to the hominid anthropological record. If this seems so silly that you doubt anyone really believes it, consider that \"the intelligence explosion is impossible because Turing machines can't promote themselves to hypercomputers\" is worse, and see  Bringsjord (2012)  for the appropriate citation by a distinguished scientist. We can be reasonably extremely confident that human intelligence does not take advantage of quantum computation  (Tegmark 2000) . The computing elements of the brain are too large and too hot. the marginal fitness returns on larger brains and further-increased intelligence, because the new equilibrium brain size was four times as large. To elaborate on this reasoning: A rational agency will invest such that the marginal returns on all its fungible investments are approximately equal. If investment X were yielding more on the margins than investment Y, it would make sense to divert resources from Y to X. But then diminishing returns would reduce the yield on further investments in X and increase the yield on further investments in Y; so after shifting some resources from Y to X, a new equilibrium would be found in which the marginal returns on investments were again approximately equal. Thus we can reasonably expect that for any species in a rough evolutionary equilibrium, each marginal added unit of ATP (roughly, metabolic energy) will yield around the same increment of inclusive fitness whether it is invested in the organism's immune system or in its brain. If it were systematically true that adding one marginal unit of ATP yielded much higher returns in the immune system compared to the brain, that species would experience a strong selection pressure in favor of diverting ATP from organisms' brains to their immune systems. Evolution measures all its returns in the common currency of inclusive genetic fitness, and ATP is a fungible resource that can easily be spent anywhere in the body. The human brain consumes roughly 20% of the ATP used in the human body, an enormous metabolic investment. Suppose a positive mutation makes it possible to accomplish the same cognitive work using only 19% of the body's ATP-with this new, more efficient neural algorithm, the same cognitive work can be done by a smaller brain. If we are in a regime of strongly diminishing fitness returns on cognition  46  or strongly diminishing cognitive returns on adding further neurons, 47 then we should expect the 46. Suppose your rooms are already lit as brightly as you like, and then someone offers you cheaper, more energy-efficient light bulbs. You will light your room at the same brightness as before and decrease your total spending on lighting. Similarly, if you are already thinking well enough to outwit the average deer, and adding more brains does not let you outwit deer any better because you are already smarter than a deer (diminishing fitness returns on further cognition), then evolving more efficient brain algorithms will lead to evolving a smaller brain that does the same work. 47. Suppose that every meal requires a hot dog and a bun; that it takes 1 unit of effort to produce each bun; and that each successive hot dog requires 1 more unit of labor to produce, starting from 1 unit for the first hot dog. Thus it takes 6 units to produce 3 hot dogs and 45 units to produce 9 hot dogs. Suppose we're currently eating 9 meals based on 45 + 9 = 54 total units of effort. Then even a magical bun factory which eliminates all of the labor in producing buns will not enable the production of 10 meals, due to the increasing cost of hot dogs. Similarly if we can recover large gains by improving the efficiency of one part of the brain, but the limiting factor is another brain part that scales very poorly, then the fact that we improved a brain algorithm well enough to significantly shrink the total cost of the brain doesn't necessarily mean that we're in a regime where we can do significantly more total cognition by reinvesting the saved neurons. brain to shrink as the result of this innovation, doing the same total work at a lower price. But in observed history, hominid brains grew larger instead, paying a greater metabolic price to do even more cognitive work. It follows that over the course of hominid evolution there were both significant marginal fitness returns on improved cognition and significant marginal cognitive returns on larger brains. In economics this is known as the Jevons paradox-the counterintuitive result that making lighting more electrically efficient or making electricity cheaper can increase the total money spent on lighting. The returns on buying lighting go up, so people buy more of it and the total expenditure increases. Similarly, some of the improvements to hominid brain algorithms over the course of hominid evolution must have increased the marginal fitness returns of spending even more ATP on the brain. The equilibrium size of the brain, and its total resource cost, shifted upward as cognitive algorithms improved. Since human brains are around four times the size of chimpanzee brains, we can conclude that our increased efficiency (cognitive yield on fungible biological resources) increased the marginal returns on brains such that the new equilibrium brain size was around four times as large. This unfortunately tells us very little quantitatively about the return-on-investment curves for larger brains and constant algorithms-just the qualitative truths that the improved algorithms did increase marginal cognitive returns on brain size, and that there weren't sharply diminishing returns on fitness from doing increased amounts of cognitive labor. It's not clear to me how much we should conclude from brain sizes increasing by a factor of only four-whether we can upper-bound the returns on hardware this way. As I understand it, human-sized heads lead to difficult childbirth due to difficulties of the baby's head passing the birth canal. This is an adequate explanation for why we wouldn't see superintelligent mutants with triple-sized heads, even if triple-sized heads could yield superintelligence. On the other hand, it's not clear that human head sizes are hard up against this sort of wall-some people have above-average-sized heads without their mothers being dead. Furthermore, Neanderthals may have had larger brains than modern humans (Ponce de  León et al. 2008) .  48  So we are probably licensed to conclude that there has not been a strong selection pressure for larger brains, as such, over very recent evolutionary history. 49 48. Neanderthals were not our direct ancestors (although some interbreeding may have occurred), but they were sufficiently closely related that their larger cranial capacities are relevant evidence. 49. It is plausible that the marginal fitness returns on cognition have leveled off sharply enough that improvements in cognitive efficiency have shifted the total resource cost of brains downward rather than upward over very recent history. If true, this is not the same as Homo sapiens sapiens becoming stupider or even staying the same intelligence. But it does imply that either marginal fitness returns on cognition or There are two steps in the derivation of a fitness return from increased brain size: a cognitive return on brain size and a fitness return on cognition. For example, John von Neumann 50 had only one child, so the transmission of cognitive returns to fitness returns might not be perfectly efficient. We can upper bound the fitness returns on larger brains by observing that Homo sapiens are not hard up against the wall of head size and that Neanderthals may have had even larger brains. This doesn't say how much of that bound on returns is about fitness returns on cognition versus cognitive returns on brain size. Do variations in brain size within Homo sapiens let us conclude much about cognitive returns? Variance in brain size correlates around 0.3 with variance in measured IQ, but there are many plausible confounders such as childhood nutrition or childhood resistance to parasites. The best we can say is that John von Neumann did not seem to require a brain exponentially larger than that of an average human, or even twice as large as that of an average human, while displaying scientific productivity well in excess of twice that of an average human being of his era. But this presumably isn't telling us about enormous returns from small increases in brain size; it's much more likely telling us that other factors can produce great increases in scientific productivity without requiring large increases in brain size. We can also say that it's not possible that a 25% larger brain automatically yields superintelligence, because that's within the range of existing variance. The main lesson I end up deriving is that intelligence improvement has not required exponential increases in computing power, and that marginal fitness returns on increased brain sizes were significant over the course of hominid evolution. This corresponds to AI growth models in which large cognitive gains by the AI can be accommodated by acquiring already-built computing resources, without needing to build new basic chip technologies. Just as an improved algorithm can increase the marginal returns on adding further hardware (because it is running a better algorithm), additional hardware can increase the marginal returns on improved cognitive algorithms (because they are running on more hardware).  51  In everyday life, we usually expect feedback loops of this sort to die down, but in the case of hominid evolution there was in fact strong continued growth, so it's marginal cognitive returns on brain scaling have leveled off significantly compared to earlier evolutionary history. 50. I often use John von Neumann to exemplify the far end of the human intelligence distribution, because he is widely reputed to have been the smartest human being who ever lived and all the other great geniuses of his era were scared of him. Hans Bethe said of him, \"I have sometimes wondered whether a brain like von Neumann's does not indicate a species superior to that of man\"  (Blair 1957) . 51. Purchasing a $1,000,000 innovation that improves all your processes by 1% is a terrible investment for a $10,000,000 company and a great investment for a $1,000,000,000 company. possible that a feedback loop of this sort played a significant role. Analogously it may be possible for an AI design to go FOOM just by adding vastly more computing power, the way a nuclear pile goes critical just by adding more identical uranium bricks; the added hardware could multiply the returns on all cognitive investments, and this could send the system from k < 1 to k > 1. Unfortunately, I see very little way to get any sort of quantitative grasp on this probability, apart from noting the qualitative possibility.  52  In general, increased \"size\" is a kind of cognitive investment about which I think I know relatively little. In AI it is usual for hardware improvements to contribute lower gains than software improvements-with improved hardware still being critical, because with a sufficiently weak computer, the initial algorithms can perform so poorly that it doesn't pay incrementally to improve them.  53  Even so, most of the story in AI has always been about software rather than hardware, and with hominid brain sizes increasing by a mere factor of four over five million years, this seems to have been true for hominid evolution as well. Attempts to predict the advent of AI by graphing Moore's Law and considering the mere addition of computing power appear entirely pointless to me given this overall state of knowledge. The cognitive returns on hardware are always changing as a function of improved algorithms; there is no calculable constant threshold to be crossed. \n One-Time Gains On an intuitive level, it seems obvious that the human species has accumulated cognitive returns sufficiently in excess of the chimpanzee species; we landed on the Moon and they didn't. Trying to get a quantitative grasp on the \"cognitive returns on humans,\" and how much they actually exceed the cognitive returns on chimpanzees, is greatly complicated by the following facts: • There are many more humans than chimpanzees. 52. This scenario is not to be confused with a large supercomputer spontaneously developing consciousness, which Pat Cadigan accurately observed to be analogous to the old theory that dirty shirts and straw would spontaneously generate mice. Rather, the concern here is that you already have an AI design which is qualitatively capable of significant self-improvement, and it goes critical after some incautious group with lots of computing resources gets excited about those wonderful early results and tries running the AI on a hundred thousand times as much computing power. 53. If hominids were limited to spider-sized brains, it would be much harder to develop human-level intelligence, because the incremental fitness returns on improved algorithms would be lower (since each algorithm runs on less hardware). In general, a positive mutation that conveys half as much advantage takes twice as long to rise to fixation, and has half the chance of doing so at all. So if you diminish the fitness returns to each step along an adaptive pathway by three orders of magnitude, the evolutionary outcome is not \"this adaptation takes longer to evolve\" but \"this adaptation does not evolve at all.\" • Humans can communicate with each other much better than chimpanzees. This implies the possibility that cognitive returns on improved brain algorithms (for humans vs. chimpanzees) might be smaller than the moon landing would suggest. Cognitive returns from better-cumulating optimization, by a much more numerous species that can use language to convey knowledge across brains, should not be confused with any inherent power of a single human brain. We know that humans have nuclear weapons and chimpanzees don't. But to the extent we attribute this to larger human populations, we must not be attributing it to humans having writing; and to the extent we attribute it to humans having writing, we must not be attributing it to humans having larger brains and improved cognitive algorithms. 54 \"That's silly,\" you reply. \"Obviously you need writing and human general intelligence before you can invent science and have technology accumulate to the level of nuclear weapons. Even if chimpanzees had some way to pass on the knowledge they possessed and do cumulative thinking-say, if you used brain-computer interfaces to directly transfer skills from one chimpanzee to another-they'd probably still never understand linear algebra, even in a million years. It's not a question of communication versus individual intelligence, there's a joint causal dependency.\" Even so (goes the counter-counterpoint) it remains obvious that discovering and using electricity is not a pure property of a single human brain. Speech and writing, as inventions enabled by hominid intelligence, induce a change in the character of cognitive intelligence as an optimization process: thinking time cumulates more strongly across populations and centuries. To the extent that we're skeptical that any further innovations of this sort exist, we might expect the grand returns of human intelligence to be a mostly one-time affair, rather than a repeatable event that scales proportionally with larger brains or further-improved cognitive algorithms. If being able to cumulate knowledge is an absolute threshold which has already been crossed, we can't expect to see repeatable cognitive returns from crossing it again and again. But then (says the counter-counter-counterpoint) we may not be all the way across the communication threshold. Suppose humans could not only talk to each other but perfectly transfer complete cognitive skills, and could not only reproduce humans in general but duplicate thousands of mutually telepathic Einsteins, the way AIs could copy themselves and transfer thoughts. Even if communication is a one-time threshold, we could be more like 1% over the threshold than 99% over it. 54. Suppose I know that your investment portfolio returned 20% last year. The higher the return of the stocks in your portfolio, the less I must expect the bonds in your portfolio to have returned, and vice versa. However (replies the counter 4 -point) if the ability to cumulate knowledge is still qualitatively present among humans, doing so more efficiently might not yield marginal returns proportional to crossing the initial threshold. Suppose there's a constant population of a hundred million people, and returns to the civilization are determined by the most cumulated cognitive labor. Going from 0% cumulation to 1% cumulation between entities might multiply total returns much more than the further multiplicative factor in going from 1% cumulation to 99% cumulation. In this scenario, a thousand 1%-cumulant entities can outcompete a hundred million 0%-cumulant entities, and yet a thousand perfectly cumulant entities cannot outcompete a hundred million 1% cumulant entities, depending on the details of your assumptions. A counter 5 -point is that this would not be a good model of piles of uranium bricks with neutron-absorbing impurities; any degree of noise or inefficiency would interfere with the clarity of the above conclusion. A further counter 5 -point is to ask about the invention of the printing press and the subsequent industrial revolution-if the one-time threshold model is true, why did the printing press enable civilizational returns that seemed to be well above those of writing or speech? A different one-time threshold that spawns a similar line of argument revolves around human generality-the way that we can grasp some concepts that chimpanzees can't represent at all, like the number thirty-seven. The science-fiction novel Schild's Ladder, by Greg  Egan (2002) , supposes a \"General Intelligence Theorem\" to the effect that once you get to the human level, you're done-you can think about anything thinkable. Hence there are no further gains from further generality; and that was why, in Egan's depicted future, there were no superintelligences despite all the human-level minds running on fast computers. The obvious inspiration for a \"General Intelligence Theorem\" is the Church-Turing Thesis: Any computer that can simulate a universal Turing machine is capable of simulating any member of a very large class of systems, which class seems to include the laws of physics and hence everything in the real universe. Once you show you can encode a single universal Turing machine in Conway's Game of Life, then the Game of Life is said to be \"Turing complete\" because we can encode any other Turing machine inside the universal machine we already built. The argument for a one-time threshold of generality seems to me much weaker than the argument from communication. Many humans have tried and failed to understand linear algebra. Some humans (however unjust this feature of our world may be) probably cannot understand linear algebra, period.  55   The main plausible source of such an argument would be an \"end of science\" scenario in which most of the interesting, exploitable possibilities offered by the physical universe could all be understood by some threshold level of generality, and thus there would be no significant returns to generality beyond this point. Humans have not developed many technologies that seem foreseeable in some sense (e.g., we do not yet have molecular nanotechnology) but, amazingly enough, all of the future technologies we can imagine from our current level seem to be graspable using human-level abilities for abstraction. This, however, is not strong evidence that no greater capacity for abstraction can be helpful in realizing all important technological possibilities. In sum, and taking into account all three of the arguments listed above, we get a combined argument as follows: The Big Marginal Return on humans over chimpanzees is mostly about large numbers of humans, sharing knowledge above a sharp threshold of abstraction, being more impressive than the sort of thinking that can be done by one chimpanzee who cannot communicate with other chimps and is qualitatively incapable of grasping algebra. Then since very little of the Big Marginal Return was really about improving cognitive algorithms or increasing brain sizes apart from that, we have no reason to believe that there were any repeatable gains of this sort. Most of the chimp-human difference is from cumulating total power rather than individual humans being smarter; you can't get human-versuschimp gains just from having a larger brain than one human. To the extent humans are qualitatively smarter than chimps, it's because we crossed a qualitative threshold which lets (unusually smart) humans learn linear algebra. But now that some of us can learn linear algebra, there are no more thresholds like that. When all of this is taken into account, it explains away most of the human bonanza and doesn't leave much to be attributed just to evolution optimizing cognitive algorithms qua algorithms and hominid brain sizes increasing by a factor of four. So we have no reason to suppose that bigger brains or better algorithms could allow an AI to experience the same sort of increased cognitive returns above humans as humans have above chimps. The above argument postulates one-time gains which all lie in our past, with no similar gains in the future. In a sense, all gains from optimization are one-time-you cannot invent the steam engine twice, or repeat the same positive mutation-and yet to expect this ongoing stream of one-time gains to halt at any particular point seems unjustified. In general, postulated one-time gains-whether from a single threshold of communication, a single threshold of generality/abstraction, etc.-seem hard to falsify or confirm by staring at raw growth records. In general, my reply is that I'm quite willing to believe that hominids have crossed qualitative thresholds, less willing to believe that such a young species as ours is already 99% over a threshold rather than 10% or 0.03% over that threshold, and extremely skeptical that all the big thresholds are already in our past and none lie in our future. Especially when humans seem to lack all sorts of neat features such as the ability to expand indefinitely onto new hardware, the ability to rewrite our own source code, the ability to run error-free cognitive processes of great serial depth, etc.  57  It is certainly a feature of the design landscape that it contains large one-time gainssignificant thresholds that can only be crossed once. It is less plausible that hominid evolution crossed them all and arrived at the qualitative limits of mind-especially when many plausible further thresholds seem clearly visible even from here. \n Returns on Speed By the standards of the eleventh century, the early twenty-first century can do things that would seem like \"magic\" in the sense that nobody in the eleventh century imagined them, let alone concluded that they would be possible.  58  What separates the early twenty-first century from the eleventh? 57. Not to mention everything that the human author hasn't even thought of yet. See section 3.11. 58. See again section 3.11. Gregory  Clark (2007)  has suggested, based on demographic data from British merchants and shopkeepers, that more conscientious individuals were having better financial success and more children, and to the extent that conscientiousness is hereditary this would necessarily imply natural selection; thus Clark has argued that there was probably some degree of genetic change supporting the Industrial Revolution. But this seems like only a small caveat to the far more obvious explanation that what separated the eleventh and twenty-first centuries was time. What is time? Leaving aside some interesting but not overwhelmingly relevant answers from fundamental physics, 59 when considered as an economic resource, \"time\" is the ability for events to happen one after another. You cannot invent jet planes at the same time as internal combustion engines; to invent transistors, somebody must have already finished discovering electricity and told you about it. The twenty-first century is separated from the eleventh century by a series of discoveries and technological developments that did in fact occur one after another and would have been significantly more difficult to do in parallel. A more descriptive name for this quality than \"time\" might be \"serial causal depth.\" The saying in software industry goes, \"Nine women can't birth a baby in one month,\" indicating that you can't just add more people to speed up a project; a project requires time, sequential hours, as opposed to just a total number of human-hours of labor. Intel has not hired twice as many researchers as its current number and produced new generations of chips twice as fast.  60  This implies that Intel thinks its largest future returns will come from discoveries that must be made after current discoveries (as opposed to most future returns coming from discoveries that can all be reached by one step in a flat search space and hence could be reached twice as fast by twice as many researchers).  61  Similarly, the \"hundred-step rule\" in neuroscience  (Feldman and Ballard 1982)  says that since human neurons can only fire around one hundred times per second, any computational process that humans seem to do in real time must take at most one hundred serial steps-that is, one hundred steps that must happen one after another. 59. See, e.g.,  Barbour (1999) . 60. With Intel's R&D cost around 17% of its sales, this wouldn't be easy, but it would be possible. 61. If Intel thought that its current researchers would exhaust the entire search space, or exhaust all marginally valuable low-hanging fruits in a flat search space, then Intel would be making plans to terminate or scale down its R&D spending after one more generation. Doing research with a certain amount of parallelism that is neither the maximum or minimum you could possibly manage implies an expected equilibrium, relative to your present and future returns on technology, of how many fruits you can find at the immediate next level of the search space, versus the improved returns on searching later after you can build on previous discoveries. (Carl Shulman commented on a draft of this paper that Intel may also rationally wait because it expects to build on discoveries made outside Intel.) There are billions of neurons in the visual cortex and so it is reasonable to suppose a visual process that involves billions of computational steps. But you cannot suppose that A happens, and that B which depends on A happens, and that C which depends on B happens, and so on for a billion steps. You cannot have a series of events like that inside a human brain; the series of events is too causally deep, and the human brain is too serially shallow. You can't even have a million-step serial process inside a modern-day factory; it would take far too long and be far too expensive to manufacture anything that required a million manufacturing steps to occur one after another. That kind of serial causal depth can only occur inside a computer. This is a great part of what makes computers useful, along with their ability to carry out formal processes exactly: computers contain huge amounts of time, in the sense of containing tremendous serial depths of causal events. Since the Cambrian explosion and the rise of anatomical multicellular organisms 2 × 10 11 days ago, your line of direct descent might be perhaps 10 8 or 10 11 generations deep. If humans had spoken continuously to each other since 150,000 years ago, one utterance per five seconds, the longest continuous conversation could have contained ∼10 12 statements one after another. A 2013-era CPU running for one day can contain ∼10 14 programmable events occurring one after another, or ∼10 16 events if you run it for one year.  62  Of course, if we are talking about a six-core CPU, then that is at most six things that could be happening at the same time, and a floating-point multiplication is a rather simple event. Still, when I contemplate statistics like those above, I am struck by a vertiginous sense of what incredibly poor use we make of computers. Although I used to go around asking, \"If Moore's Law says that computing speeds double every eighteen months, what happens when computers are doing the research?\" 63 I no longer think that Moore's Law will play much of a role in the intelligence explosion, partially because I expect returns on algorithms to dominate, and partially because I would expect an AI to prefer ways to scale itself onto more existing hardware rather than waiting for a new generation of chips to be produced in Intel-style factories. The latter form of investment has such a slow timescale, and hence such a low interest rate, that I would only expect it to be undertaken if all other self-improvement alternatives had bottlenecked before reaching the point of solving protein structure prediction or otherwise bypassing large human-style factories. 62. Almost the same would be true of a 2008-era CPU, since the Moore's-like law for serial depth has almost completely broken down. Though CPUs are also not getting any slower, and the artifacts we have already created seem rather formidable in an absolute sense. \n I was then seventeen years old. Since computers are well known to be fast, it is a very widespread speculation that strong AIs would think very fast because computers would be very fast, and hence that such AIs would rapidly acquire advantages of the sort we associate with older human civilizations, usually improved science and technology.  64  Two objections that have been offered against this idea are (a) that the first sufficiently advanced AI might be very slow while already running on a large fraction of all available computing power, and hence hard to speed up without waiting on Moore's Law, 65 and (b) that fast thinking may prove useless without fast sensors and fast motor manipulators.  66  Let us consider first the prospect of an advanced AI already running on so much computing power that it is hard to speed up. I find this scenario somewhat hard to analyze because I expect AI to be mostly about algorithms rather than lots of hardware, but I can't rule out scenarios where the AI is developed by some large agency which was running its AI project on huge amounts of hardware from the beginning. This should not make the AI slow in all aspects; any AI with a certain amount of self-reprogramming ability ought to be able to perform many particular kinds of cognition very quicklyto take one extreme example, it shouldn't be slower than humans at arithmetic, even conscious arithmetic. But the AI's overall thought processes might still be slower than human, albeit presumably not so slow that the programmers and researchers are too bored to work effectively on the project or try to train and raise the AI. Thus I cannot say that the overall scenario is implausible. I do note that to the extent that an AI is running on more hardware and has worse algorithms, ceteris paribus, you would expect greater gains from improving the algorithms. Trying to deliberately create a slow AI already running on vast amounts of hardware, in hopes of guaranteeing sufficient time to react, may not actually serve to slow down the overall growth curve-it may prove to be the equivalent of starting out the AI with much more hardware than it would have had otherwise, hence greater returns on improving its algorithms. I am generally uncertain about this point. On the input-output side, there are various Moore's-like curves for sensing and manipulating, but their exponents tend to be lower than the curves for pure computer technologies. If you extrapolated this trend outward without further change, then the pure 64. As the fourth-century Chinese philosopher Xiaoguang Li once observed, we tend to think of earlier civilizations as being more venerable, like a wise old ancestor who has seen many things; but in fact later civilizations are older than earlier civilizations, because the future has a longer history than the past. Thus I hope it will increase, rather than decrease, your opinion of his wisdom if I now inform you that actually Xiaoguang \"Mike\" Li is a friend of mine who observed this in 2002. 65. This has mostly come up in personal conversation with friends; I'm not sure I've seen a print source. 66. The author is reasonably sure he has seen this objection in print, but failed again to collect the reference at the time. the fast-thinking researchers waiting through their molasses-slow ability to manipulate clumsy robotic hands to perform experiments and actually observe the results. The field of high-energy physics, for example, seems limited by the expense and delay of constructing particle accelerators. Likewise, subfields of astronomy revolve around expensive space telescopes. These fields seem more sensory-bounded than thinkingbounded, relative to the characteristic intelligence of the researchers. It's possible that sufficiently smarter scientists could get more mileage out of information already gathered, or ask better questions. But at the very least, we can say that there's no humanlyobvious way to speed up high-energy physics with faster-thinking human physicists, and it's easy to imagine that doubling the speed of all the human astronomers, while leaving them otherwise unchanged, would just make them twice as frustrated about telescope time as at present. At the opposite extreme, theoretical mathematics stands as an example of a field which is limited only by the thinking speed of its human researchers (computer assistance currently being a rare exception, rather than the rule). It is interesting to ask whether we should describe progress in mathematics as (1) continuing at mostly the same pace as anything else humans do, or (2) far outstripping progress in every other human endeavor, such that there is no nonmathematical human accomplishment comparable in depth to Andrew Wiles's proof of Fermat's Last Theorem  (Wiles 1995) . The main counterpoint to the argument from the slower Moore's-like laws for sensorimotor technologies is that since currently human brains cannot be sped up, and humans are still doing most of the physical labor, there hasn't yet been a strong incentive to produce faster and faster manipulators-slow human brains would still be the limiting factor. But if in the future sensors or manipulators are the limiting factor, most investment by a rational agency will tend to flow toward improving that factor. If slow manipulators are holding everything back, this greatly increases returns on faster manipulators and decreases returns on everything else. But with current technology it is not possible to invest in faster brains for researchers, so it shouldn't be surprising that the speed of researcher thought often is the limiting resource. Any lab that shuts down overnight so its researchers can sleep must be limited by serial cause and effect in researcher brains more than serial cause and effect in instruments-researchers who could work without sleep would correspondingly speed up the lab. In contrast, in astronomy and high-energy physics every minute of apparatus time is scheduled, and shutting down the apparatus overnight would be unthinkable. That most human research labs do cease operation overnight implies that most areas of research are not sensorimotor bounded. However, rational redistribution of investments to improved sensors and manipulators does not imply that the new resulting equilibrium is one of fast progress. The counter-counterpoint is that, even so, improved sensors and manipulators are slow to construct compared to just rewriting an algorithm to do cognitive work faster. Hence sensorimotor bandwidth might end up as a limiting factor for an AI going FOOM over short timescales; the problem of constructing new sensors and manipulators might act as metaphorical delayed neutrons that prevent prompt criticality. This delay would still exist so long as there were pragmatically real limits on how useful it is to think in the absence of experiential data and the ability to exert power on the world. A counter-counter-counterpoint is that if, for example, protein structure prediction can be solved as a purely cognitive problem, 67 then molecular nanotechnology is liable to follow very soon thereafter. It is plausible that even a superintelligence might take a while to construct advanced tools if dropped into the thirteenth century with no other knowledge of physics or chemistry.  68  It's less plausible (says the counter-countercounterargument) that a superintelligence would be similarly bounded in a modern era where protein synthesis and picosecond cameras already exist, and vast amounts of pregathered data are available.  69  Rather than imagining sensorimotor bounding as the equivalent of some poor blind spirit in a locked box, we should imagine an entire human civilization in a locked box, doing the equivalent of cryptography to extract every last iota of inference out of every bit of sensory data, carefully plotting the fastest paths to greater potency using its currently conserved motor bandwidth, using every possible avenue of affecting the world to, as quickly as possible, obtain faster ways of affecting the world. See here for an informal exposition.  70  I would summarize my views on \"speed\" or \"causal depth\" by saying that, contrary to the views of a past Eliezer Yudkowsky separated from my present self by sixteen 67. Note that in some cases the frontier of modern protein structure prediction and protein design is crowdsourced human guessing, e.g., the Foldit project. This suggests that there are gains from applying better cognitive algorithms to protein folding. 68. It's not certain that it would take the superintelligence a long time to do anything, because the putative superintelligence is much smarter than you and therefore you cannot exhaustively imagine or search the options it would have available. See section 3.11. 69. Some basic formalisms in computer science suggest fundamentally different learning rates depending on whether you can ask your own questions or only observe the answers to large pools of preasked questions. On the other hand, there is also a strong case to be made that humans are overwhelmingly inefficient at constraining probability distributions using the evidence they have already gathered. 70. An intelligence explosion that seems incredibly fast to a human might take place over a long serial depth of parallel efforts, most of which fail, learning from experience, updating strategies, waiting to learn the results of distant experiments, etc., which would appear frustratingly slow to a human who had to perform similar work. Or in implausibly anthropomorphic terms, \"Sure, from your perspective it only took me four days to take over the world, but do you have any idea how long that was for me? I had to wait twenty thousand subjective years for my custom-ordered proteins to arrive!\" years of \"time,\" 71 it doesn't seem very probable that returns on hardware speed will be a key ongoing factor in an intelligence explosion. Even Intel constructing new chip factories hasn't increased serial speeds very much since 2004, at least as of 2013. Better algorithms or hardware scaling could decrease the serial burden of a thought and allow more thoughts to occur in serial rather than parallel; it seems extremely plausible that a humanly designed AI will start out with a huge excess burden of serial difficulty, and hence that improving cognitive algorithms or hardware scaling will result in a possibly gradual, possibly one-time huge gain in effective cognitive speed. Cognitive speed outstripping sensorimotor bandwidth in a certain fundamental sense is also very plausible for pre-nanotechnological stages of growth. The main policy-relevant questions would seem to be: 1. At which stage (if any) of growth will an AI be able to generate new technological capacities of the sort that human civilizations seem to invent \"over time,\" and how quickly? 2. At which stage (if any) of an ongoing intelligence explosion, from which sorts of starting states, will which events being produced by the AI exceed in speed the reactions of (1) human bureaucracies and governments with great power (weeks or months) and (  2 ) individual humans with relatively lesser power (minutes or seconds)? I would expect that some sort of incredibly fast thinking is likely to arrive at some point, because current CPUs are already very serially fast compared to human brains; what stage of growth corresponds to this is hard to guess. I've also argued that the \"high-speed spirit trapped in a statue\" visualization is inappropriate, and \"high-speed human civilization trapped in a box with slow Internet access\" seems like a better way of looking at it. We can visualize some clear-seeming paths from cognitive power to fast infrastructure, like cracking the protein structure prediction problem. I would summarize my view on this question by saying that, although high cognitive speeds may indeed lead to time spent sensorimotor bounded, the total amount of this time may not seem very large from outside-certainly a high-speed human civilization trapped inside a box with Internet access would be trying to graduate to faster manipulators as quickly as possible. 71. Albeit, in accordance with the general theme of embarrassingly overwhelming human inefficiency, the actual thought processes separating Yudkowsky 1997 from Yudkowsky 2013 would probably work out to twenty days of serially sequenced thoughts or something like that. Maybe much less. Certainly not sixteen years of solid sequential thinking. \n Returns on Population As remarked in section 3.3, the degree to which an AI can be competitive with the global human population depends, among other factors, on whether humans in large groups scale with something close to the ideal efficiency for parallelism. In 1999, a game of chess titled \"Kasparov versus The World\" was played over the Internet between Garry Kasparov and a World Team in which over fifty thousand individuals participated at least once, coordinated by four young chess stars, a fifth master advising, and moves decided by majority vote with five thousand voters on a typical move. Kasparov won after four months and sixty-two moves, saying that he had never expended so much effort in his life, and later wrote a book  (Kasparov and King 2000)  about the game, saying, \"It is the greatest game in the history of chess. The sheer number of ideas, the complexity, and the contribution it has made to chess make it the most important game ever played.\" There was clearly nontrivial scaling by the contributors of the World Team-they played at a far higher skill level than their smartest individual players. But eventually Kasparov did win, and this implies that five thousand human brains (collectively representing, say, ∼10 18 synapses) were not able to defeat Kasparov's ∼10 14 synapses. If this seems like an unfair estimate, its unfairness may be of a type that ubiquitously characterizes human civilization's attempts to scale. Of course many of Kasparov's opponents were insufficiently skilled to be likely to make a significant contribution to suggesting or analyzing any given move; he was not facing five thousand masters. But if the World Team had possessed the probable advantages of AIs, they could have copied chess skills from one of their number to another, and thus scaled more efficiently. The fact that humans cannot do this, and that we must painstakingly and expensively reproduce the educational process for every individual who wishes to contribute to a cognitive frontier, and some our most remarkable examples cannot be duplicated by any known method of training, is one of the ways in which human populations scale less than optimally.  72  On a more micro level, it is a truism of computer science and an important pragmatic fact of programming that processors separated by sparse communication bandwidth sometimes have trouble scaling well. When you lack the bandwidth to copy whole internal cognitive representations, computing power must be expended (wasted) to reconstruct those representations within the message receiver. It was not possible for one of Kasparov's opponents to carefully analyze an aspect of the situation and then copy and distribute that state of mind to one hundred others who could analyze slight variant thoughts and then combine their discoveries into a single state of mind. They were limited to speech instead. In this sense it is not too surprising that 10 14 synapses with high local intercommunication bandwidth and a high local skill level could defeat 10 18 synapses separated by gulfs of speech and argument. Although I expect that this section of my analysis will not be without controversy, it appears to the author to also be an important piece of data to be explained that human science and engineering seem to scale over time better than over population-an extra decade seems much more valuable than adding warm bodies. Indeed, it appears to the author that human science scales ludicrously poorly with increased numbers of scientists, and that this is a major reason there hasn't been more relative change from <PHONE> than from <PHONE> despite the vastly increased number of scientists. The rate of real progress seems mostly constant with respect to time, times a small factor more or less. I admit that in trying to make this judgment I am trying to summarize an overwhelmingly distant grasp on all the fields outside my own handful. Even so, a complete halt to science or a truly exponential (or even quadratic) speedup of real progress both seem like they would be hard to miss, and the exponential increase of published papers is measurable. Real scientific progress is continuing over time, so we haven't run out of things to investigate; and yet somehow real scientific progress isn't scaling anywhere near as fast as professional scientists are being added. The most charitable interpretation of this phenomenon would be that science problems are getting harder and fields are adding scientists at a combined pace which produces more or less constant progress. It seems plausible that, for example, Intel adds new researchers at around the pace required to keep up with its accustomed exponential growth. On the other hand, Intel actually publishes their future roadmap and is a centrally coordinated semirational agency. Scientific fields generally want as much funding as they can get from various funding sources who are reluctant to give more of it, with politics playing out to determine the growth or shrinking rate in any given year. It's hard to see how this equilibrium could be coordinated. A moderately charitable interpretation would be that science is inherently bounded by serial causal depth and is poorly parallelizable-that the most important impacts of scientific progress come from discoveries building on discoveries, and that once the best parts of the local search field are saturated, there is little that can be done to reach destinations any faster. This is moderately uncharitable because it implies that large amounts of money are probably being wasted on scientists who have \"nothing to do\" when the people with the best prospects are already working on the most important problems. It is still a charitable interpretation in the sense that it implies global progress is being made around as fast as human scientists can make progress. Both of these charitable interpretations imply that AIs expanding onto new hardware will not be able to scale much faster than human scientists trying to work in parallel, since human scientists are already working, in groups, about as efficiently as reasonably possible. And then we have the less charitable interpretations-those which paint humanity's performance in a less flattering light. For example, to the extent that we credit Max Planck's claim that \"a new scientific truth does not triumph by convincing its opponents and making them see the light, but rather because its opponents eventually die, and a new generation grows up that is familiar with it\"  (Kuhn 1962) , we could expect that the process of waiting for the previous generation to die out (or rather, retire) was a serial bottleneck not affected by increased parallelism. But this would be a bottleneck of human stubbornness and aging biological brains, rather than an inherent feature of the problem space or a necessary property of rational agencies in general. I have also wondered how it is that a ten-person startup can often appear to be around as innovative on average as a ten-thousand-person corporation. An interpretation has occurred to me which I have internally dubbed \"the hero theory.\" This is the idea that a human organization has room for one to five \"heroes\" who are allowed to be important, and that other potential heroes somehow see that all hero positions are already occupied, whereupon some instinctive part of their mind informs them that there is no fame or status to be gained from heroics.  73  This theory has the advantage of explaining in a unified way why neither academic fields nor corporations seem to be able to scale \"true innovation\" by throwing more warm bodies at the problem, and yet are still able to scale with added time. It has the disadvantage of its mechanism not being overwhelmingly plausible. Similar phenomena might perhaps be produced by the attention span of other researchers bottlenecking through a few leaders, or by limited width of attention to funding priorities or problems. This kind of sociology is not really my field. Diving further into the depths of cynicism, we may ask whether \"science\" is perhaps a process distinct from \"publishing papers in journals,\" where our civilization understands how to reproduce the latter skill but has no systematic grasp on reproducing the former. One observes that technological progress is not (yet) dominated by China despite China graduating more PhDs than any other nation. This seems understandable if human civilization understands explicitly how to make PhDs, but the production of scientists 73. I have sometimes worried that by being \"that Friendly AI guy\" I have occupied the position of \"Friendly AI guy\" and hence young minds considering what to do with their lives will see that there is already a \"Friendly AI guy\" and hence not try to do this themselves. This seems to me like a very worrisome prospect, since I do not think I am sufficient to fill the entire position. is dominated by rare lineages of implicit experts who mostly live in countries with long historical scientific traditions-and moreover, politicians or other funding agencies are bad at distinguishing the hidden keepers of the tradition and cannot selectively offer them a million dollars to move to China. In one sense this possibility doesn't say much about the true scaling factor that would apply with more scientists, but it says that a large penalty factor might apply to estimating human scaling of science by estimating scaling of publications. In the end this type of sociology of science is not really the author's field. Nonetheless one must put probability distributions on guesses, and there is nothing especially virtuous about coming to estimates that sound respectful rather than cynical. And so the author will remark that he largely sees the data to be explained as \"human science scales extremely poorly with throwing more warm bodies at a field\"; and that the author generally sees the most plausible explanations as revolving around problems of the human scientific bureaucracy and process which would not necessarily hold of minds in general, especially a single AI scaling onto more hardware. \n The Net Efficiency of Human Civilization It might be tempting to count up 7,000,000,000 humans with 100,000,000,000 neurons and 1,000 times as many synapses firing around 100 times per second, and conclude that any rational agency wielding much fewer than 10 26 computing operations per second cannot be competitive with the human species. But to the extent that there are inefficiencies, either in individual humans or in how humans scale in groups, 10 26 operations per second will not well characterize the cognitive power of the human species as a whole, as it is available to be focused on a scientific or technological problem, even relative to the characteristic efficiency of human cognitive algorithms. A preliminary observation, that John von Neumann had a brain not much visibly larger than that of the average human, suggests that the true potential of 10 26 operations per second must be bounded below by the potential of 7,000,000,000 mutually telepathic von Neumanns. Which does not seem to well characterize the power of our current civilization. Which must therefore be operating at less than perfect efficiency in the realms of science and technology. In particular, I would suggest the following inefficiencies: • Humans must communicate by speech and other low-bandwidth means rather than directly transferring cognitive representations, and this implies a substantial duplication of cognitive labor. • It is possible that some professionals are systematically unproductive of important progress in their field, and the number of true effective participants must be adjusted down by some significant factor. • Humans must spend many years in schooling before they are allowed to work on scientific problems, and this again reflects mostly duplicated cognitive labor, compared to Xeroxing another copy of Einstein. • Human scientists do not do science twenty-four hours per day (this represents a small integer factor of reduced efficiency). • Professional scientists do not spend all of their working hours directly addressing their scientific problems. • Within any single human considering a scientific problem, not all of their brain can be regarded as working on that problem. • Inefficiencies of human scientific bureaucracy may cause potentially helpful contributions to be discarded, or funnel potentially useful minds into working on problems of predictably lesser importance, etc. One further remarks that most humans are not scientists or engineers at all, and most scientists and engineers are not focusing on the problems that an AI in the process of an intelligence explosion might be expected to focus on, like improved machine cognitive algorithms or, somewhere at the end, protein structure prediction. However, the Hansonian method of critique 74 would obviously prompt the question, \"Why do you think AIs wouldn't have to spend most of their time and brainpower on subsidiary economic tasks to support themselves, just like human civilization can't afford to spend all its time on AI research?\" One reply might be that, while humans are obliged to use whole human brains to support their bodies even as they carry out relatively repetitive bits of physical or cognitive labor, an AI would be able to exploit money-earning opportunities that required straightforward cognition using a correspondingly smaller amount of computing power. The Hansonian method would then proceed to ask why there weren't many AIs bidding on such jobs and driving down the returns.  75  But in models with a localized FOOM 74. I would describe the general rule as follows: \"For all supposed capabilities of AIs, ask why humans do not have the same ability. For all supposed obstacles to the human version of the ability, ask why similar obstacles would not apply to AIs.\" I often disagree with Hanson about whether cases of this question can be given satisfying answers, but the question itself is clearly wise and correct. 75. I would describe this rule as follows: \"Check whenever someone is working on a background assumption of a localized FOOM and then consider a contrasting scenario based on many AIs of roughly and hence one AI relatively ahead of other projects, it is very reasonable that the AI could have a much higher ratio of \"computing operations doing science\" to \"computing operations earning money,\" even assuming the AI was not simply stealing its computer time. More generally, the fact that the whole human population is not mostly composed of professional scientists, working on the most important problems an AI would face in the process of going FOOM, must play a role in reducing our estimate of the net computing power required to match humanity's input into AI progress, given algorithms of roughly human-level efficiency. All of the above factors combined may still only scratch the surface of human computational inefficiency. Our performance on integer multiplication problems is not in accordance with what a crude estimate of 10 16 operations per second might lead you to expect. To put it another way, our brains do not efficiently transmit their underlying computing power to the task of integer multiplication. Our insanely poor performance on integer multiplication clearly does not upperbound human computational efficiency on all problems-even nonancestral problems. Garry Kasparov was able to play competitive chess against Deep Blue while Kasparov was examining two moves per second to Deep Blue's two billion moves per second, implying that Kasparov was indeed able to effectively recruit his visual cortex, temporal lobe, prefrontal cortex, cerebellum, etc., to effectively contribute large amounts of computing power in the form of parallelized pattern recognition and planning. In fact Kasparov showed amazing computational efficiency; he was able to match Deep Blue in a fashion that an a priori armchair reasoner probably would not have imagined possible for a mind limited to a hundred steps per second of serial depth. Nonetheless, the modern chess program Deep Rybka 3.0 is far ahead of Kasparov while running on 2.8 billion operations per second, so Kasparov's brainpower is still not being perfectly transmitted to chess-playing ability. In the end such inefficiency is what one would expect, given that Kasparov's genetic makeup was not selected over eons to play chess. We might similarly find of human scientists that, even though they are able to recruit more of their brains' power to science than to integer multiplication, they are still not using their computing operations as efficiently as a mind designed to do science-even during their moments of peak insight while they are working on that exact problem. \n All these factors combined project a very different image of what an AI must do to outcompete human civilization at the task of inventing better AI algorithms or cracking protein folding than saying that the AI must compete with 7,000,000,000 humans each with 10 11 neurons and 10 14 synapses firing 10 2 times per second. By the time we are done observing that not all humans are scientists, that not all scientists are productive, that not all productive scientists are working on the problem every second, that not all professional labor is directly applicable to the cognitive problem, that cognitive labor (especially learning, or understanding ideas transmitted by speech) is often duplicated between individuals, that the fruits of nonduplicated contributions are processed by the surrounding bureaucracy with less than perfect efficiency, that humans experience significant serial bottlenecks due to their brains running on a characteristic timescale of at most 10 2 steps per second, that humans are not telepathic, and finally that the actual cognitive labor applied to the core cognitive parts of scientific problems during moments of peak insight will be taking place at a level of inefficiency somewhere between \"Kasparov losing at chess against Deep Rybka's 2.8 billion operations/second\" and \"Kasparov losing at integer multiplication to a pocket calculator\" . . . . . . the effective computing power of human civilization applied to the relevant problems may well be within easy range of what a moderately well-funded project could simply buy for its AI, without the AI itself needing to visibly earn further funding. Frankly, my suspicion is that by the time you're adding up all the human inefficiencies, then even without much in the way of fundamentally new and better algorithmsjust boiling down the actual cognitive steps required by the algorithms we already usewell, it's actually quite low, I suspect.  76  And this probably has a substantial amount to do with why, in practice, I think a moderately well-designed AI could overshadow the power of human civilization. It's not just about abstract expectations of future growth, it's a sense that the net cognitive ability of human civilization is not all that impressive once all the inefficiencies are factored in. Someone who thought that 10 26 operations per second was actually a good proxy measure of the magnificent power of human civilization might think differently. \n Returns on Cumulative Evolutionary Selection Pressure I earlier claimed that we have seen no signs of diminishing cognitive returns to cumulative natural selection. That is, it didn't take one-tenth as long to go from Australopithecus to Homo erectus as it did from Homo erectus to Homo sapiens. The alert reader may protest, \"Of course the erectus-sapiens interval isn't ten times as long as the Australopithecus-erectus 76. Though not as low as if all the verbal thoughts of human scientists could be translated into first-order logic and recited as theorems by a ridiculously simple AI engine, as was briefly believed during the early days. If the claims made by the makers of BACON  (Langley, Bradshaw, and Zytkow 1987)  or the Structure Mapping Engine  (Falkenhainer and Forbus 1990)  were accurate models of human cognitive reasoning, then the Scientific Revolution up to 1900 would have required on the order of perhaps 10 6 cognitive operations total . We agree however with  Chalmers, French, and Hofstadter (1992)  that this is not a good model. So not quite that low. interval, you just picked three named markers on the fossil record that didn't happen to have those relative intervals.\" Or, more charitably: \"Okay, you've shown me some named fossils A, B, C with 3.2 million years from A to B and then 1.8 million years from B to C. What you're really claiming is that there wasn't ten times as much cognitive improvement from A to B as from B to C. How do you know that?\" To this I could reply by waving my hands in the direction of the details of neuroanthropology, 77 and claiming that the observables for throat shapes (for language use), preserved tools and campfires, and so on, just sort of look linear-or moderately superlinear, but at any rate not sublinear. A graph of brain sizes with respect to time may be found here  (Calvin 2004, chap. 5) . And despite the inferential distance from \"brain size\" to \"increasing marginal fitness returns on brain size\" to \"brain algorithmic improvements\"-nonetheless, the chart looks either linear or moderately superlinear. More broadly, another way of framing this is to ask what the world should look like if there were strongly decelerating returns to evolutionary optimization of hominids.  78  I would reply that, first of all, it would be very surprising to see a world whose cognitive niche was dominated by just one intelligent species. Given sublinear returns on cumulative selection for cognitive abilities, there should be other species that mostly catch up to the leader. Say, evolving sophisticated combinatorial syntax from protolanguage should have been a much more evolutionarily expensive proposition than just producing protolanguage, due to the decelerating returns.  79  And then, in the long time it took hominids to evolve complex syntax from protolanguage, chimpanzees should have caught up and started using protolanguage. Of course, evolution does not always recapitulate the same outcomes, even in highly similar species. But in general, sublinear cognitive returns to evolution imply that it would be surprising to see one species get far ahead of all others; there should be nearly even competitors in the process of catching 77. Terrence  Deacon's (1997)  The Symbolic Species is notionally about a theory of human general intelligence which I believe to be quite mistaken, but the same book is incidentally an excellent popular overview of cognitive improvements over the course of hominid evolution, especially as they relate to language and abstract reasoning. 78. At the Center for Applied Rationality, one way of training empiricism is via the Monday-Tuesday game. For example, you claim to believe that cellphones work via \"radio waves\" rather than \"magic.\" Suppose that on Monday cellphones worked via radio waves and on Tuesday they worked by magic. What would you be able to see or test that was different between Monday and Tuesday? Similarly, here we are asking, \"On Monday there are linear or superlinear returns on cumulative selection for better cognitive algorithms. On Tuesday the returns are strongly sublinear. How does the world look different on Monday and Tuesday?\" To put it another way: If you have strongly concluded X, you should be able to easily describe how the world would look very different if not-X, or else how did you conclude X in the first place? 79. For an explanation of \"protolanguage\" see  Bickerton (2009) . up. (For example, we see millions of species that are poisonous, and no one species that has taken over the entire \"poison niche\" by having far better poisons than its nearest competitor.) But what if there were hugely increased selection pressures on intelligence within hominid evolution, compared to chimpanzee evolution? What if, over the last 1.8 million years since Homo erectus, there was a thousand times as much selection pressure on brains in particular, so that the cumulative optimization required to go from Homo erectus to Homo sapiens was in fact comparable with all the evolution of brains since the start of multicellular life? There are mathematical limits on total selection pressures within a species. However, rather than total selection pressure increasing, it's quite plausible for selection pressures to suddenly focus on one characteristic rather than another. Furthermore, this has almost certainly been the case in hominid evolution. Compared to, say, scorpions, a competition between humans is much more likely to revolve around who has the better brain than around who has better armor plating. More variance in a characteristic which covaries with fitness automatically implies increased selective pressure on that characteristic.  80  Intuitively speaking, the more interesting things hominids did with their brains, the more of their competition would have been about cognition rather than something else. And yet human brains actually do seem to look a lot like scaled-up chimpanzee brains-there's a larger prefrontal cortex and no doubt any number of neural tweaks, but the gross brain anatomy has changed hardly at all. In terms of pure a priori evolutionary theory-the sort we might invent if we were armchair theorizing and had never seen an intelligent species evolve-it wouldn't be too surprising to imagine that a planet-conquering organism had developed a new complex brain from scratch, far more complex than its nearest competitors, after that organ suddenly became the focus of intense selection sustained for millions of years. But in point of fact we don't see this. Human brains look like scaled-up chimpanzee brains, rather than mostly novel organs. Why is that, given the persuasive-sounding prior argument for how there could have plausibly been thousands of times more selection pressure per generation on brains, compared to previous eons? Evolution is strongly limited by serial depth, even though many positive mutations can be selected on in parallel. If you have an allele B which is only advantageous in the presence of an allele A, it is necessary that A rise to universality, or at least prevalence, within the gene pool before there will be significant selection pressure favoring B. If C depends on both A and B, both A and B must be highly prevalent before there is significant pressure favoring C. 81 Within a sexually reproducing species where any genetic variance is repeatedly scrambled, complex machines will be mostly composed of a deep, still pool of complexity, with a surface froth of non-interdependent improvements being selected on at any given point. Intensified selection pressures may increase the speed at which individually positive alleles rise to universality in the gene pool, or allow for selecting on more non-interdependent variations in parallel. But there's still an important sense in which the evolution of complex machinery is strongly limited by serial depth. So even though it is extremely plausible that hominids experienced greatly intensified selection on brains versus other organismal characteristics, it still isn't surprising that human brains look mostly like chimpanzee brains when there have only been a few hundred thousand generations separating us. Nonetheless, the moderately superlinear increase in hominid brain sizes over time could easily accommodate strictly linear returns on cumulative selection pressures, with the seeming acceleration over time being due only to increased selection pressures on intelligence. It would be surprising for the cognitive \"returns on cumulative selection pressure\" not to be beneath the curve for \"returns on cumulative time.\" I was recently shocked to hear about claims for molecular evidence that rates of genetic change may have increased one hundred-fold among humans since the start of agriculture  (Hawks et al. 2007 ). Much of this may have been about lactose tolerance, melanin in different latitudes, digesting wheat, etc., rather than positive selection on new intelligence-linked alleles. This still allows some potential room to attribute some of humanity's gains over the last ten thousand years to literal evolution, not just the accumulation of civilizational knowledge. But even a literally hundredfold increase in rates of genetic change does not permit cognitive returns per individual mutation to have fallen off significantly over the course of hominid evolution. The mathematics of evolutionary biology says that a single mutation event which conveys a fitness advantage of s, in the sense that the average fitness of its bearer is 1 + s compared to a population average fitness of 1, has a 2s probability of spreading through a population to fixation; and the expected fixation time is 2 ln(N )/s generations, where N is total population size. So if the fitness advantage per positive mutation falls low enough, not only will that mutation take a very large number of 81. Then along comes A* which depends on B and C, and now we have a complex interdependent machine which fails if you remove any of A*, B, or C. Natural selection naturally and automatically produces \"irreducibly\" complex machinery along a gradual, blind, locally hill-climbing pathway. generations to spread through the population, it's very likely not to spread at all (even if the mutation independently recurs many times). The possibility of increased selection pressures should mainly lead us to suspect that there are huge cognitive gaps between humans and chimpanzees which resulted from merely linear returns on cumulative optimization-there was a lot more optimization going on, rather than small amounts of optimization yielding huge returns. But we can't have a small cognitive gap between chimps and humans, a large amount of cumulative selection, and fitness returns on individual mutations strongly diminishing, because in this scenario we wouldn't get much evolution, period. The possibility of increased rates of genetic change does not actually imply room for cognitive algorithms becoming \"harder to design\" or \"harder to improve upon\" as the base level grows more sophisticated. Returns on single positive mutations are lower-bounded by the logic of natural selection. If you think future molecular genetics might reveal these sorts of huge selection pressures in the historical record, you should consistently think it plausible (though perhaps not certain) that humans are vastly smarter than chimps (contrary to some arguments in the opposite direction, considered in section 3.2). There is room for the mind-design distance from Homo erectus to Homo sapiens to be significant compared to, say, the mind-design distance from mouse to Australopithecus, contrary to what the relative time intervals in the fossil record would suggest. To wedge diminishing cognitive returns on evolution into this model-without contradicting basic evolutionary points about how sufficiently small fitness advantages take huge amounts of time to fixate, or more likely don't fixate at all-we would have to suppose that small cognitive advantages were somehow providing outsize fitness advantages (in a way irrelevant to returns on cognitive reinvestment for AIs trying to improve themselves). To some degree, \"inflated fitness advantages\" occur in theories of runaway sexual selection (where everyone tries to mate with whoever seems even nominally smartest). To whatever extent such sexual selection was occurring, we should decrease our estimate of the sort of cognitively produced fitness advantage that would carry over to a machine intelligence trying to work on the protein folding problem (where you do not get an outsized prize for being only slightly better). I would nonetheless say that, at the end of the day, it takes a baroque interpretation of the graph of brain sizes with respect to time, to say nothing of the observed cognitive gap between humans and chimps, before you can get diminishing returns on cumulative natural selection out of observed bioanthropology. There's some room for short recent time intervals to expand into large amounts of cumulative selection pressure, but this mostly means that we don't need to postulate increasing returns on each positive mu-tation to account for apparently superlinear historical progress.  82  On the whole, there is not much room to postulate that evolutionary history is telling us about decreasing cognitive returns to cumulative natural selection. \n Relating Curves of Evolutionary Difficulty and Engineering Difficulty What if creating human intelligence was easy for natural selection but will be hard for human engineers? The power of natural selection is often romanticized-for example, because of cultural counterpressures in the United States to religions that try to falsely downplay the power of natural selection. Even some early biologists made such errors, although mostly before George C.  Williams (1966)  and the revolution of the 1960s, which spawned a very clear, often mathematically precise, picture of the capabilities and characteristic design processes of natural selection. Today we can in many respects quantify with simple equations the statement that natural selection is slow, stupid, and blind: a positive mutation of fitness 1 + s will require 2 ln(population)/s generations to fixate and has only a 2s probability of doing so at all.  83  Evolution has invented the freely rotating wheel on only a tiny handful of occasions in observed biology. Freely rotating wheels are in fact highly efficient-that is why they appear in ATP synthase, a molecule which may have been selected more heavily for near-perfect efficiency than almost anything else in biology. But (especially once we go from self-assembling molecules to organs which must be grown from tissue) it's hard to come by intermediate evolutionary forms along the way to a freely rotating wheel. Evolution cannot develop intermediate forms aiming for a freely rotating wheel, and it almost never locally hill-climbs into that design. This is one example of how human engineers, who can hold whole designs in their imagination and adjust them in response to imagined problems, can easily access areas of design space which evolution almost never enters. We should strongly expect that point mutation, random recombination, and statistical selection would hit bottlenecks in parts of the growth curve where deliberate foresight, consequentialist back-chaining, and learned abstraction would carry steadily onward-rather than the other way around. Difficulty curves for intelligent engineers 82. To be clear, increasing returns per positive mutation would imply that improving cognitive algorithms became easier as the base design grew more sophisticated, which would imply accelerating returns to constant optimization. This would be one possible explanation for the seemingly large gains from chimps to humans, but the fact that selection pressures almost certainly increased, and may have increased by quite a lot, means we cannot strongly conclude this. should be bounded upward by the difficulty curves for the processes of natural selection (where higher difficulty represents lower returns on cumulative investment). Evolution does have a significant head start. But while trying to catch up with millions of years of cumulative evolutionary optimization sounds intimidating at first, it becomes less intimidating once you calculate that it takes 875 generations for a gene conveying a 3% fitness advantage to spread through a population of five hundred thousand individuals. We can't expect the difficulty curves for intelligent engineering and natural selection to be the same. But we can reasonably relate them by saying that the difficulty curve for intelligent engineering should stay below the corresponding curve for natural selection, but that natural selection has a significant head start on traversing this curve. Suppose we accept this relation. Perhaps we still can't conclude very much in practice about AI development times. Let us postulate that it takes eighty years for human engineers to get AI at the level of Homo erectus. Plausibly erectus-level intelligence is still not smart enough for the AI to contribute significantly to its own development (though see section 3.10).  84  Then, if it took eighty years to get AI to the level of Homo erectus, would it be astonishing for it to take another ninety years of engineering to get to the level of Homo sapiens? I would reply, \"Yes, I would be astonished, because even after taking into account the possibility of recently increased selection pressures, it still took far more evolutionary time to get to Homo erectus from scratch than it took to get from Homo erectus to Homo sapiens.\" If natural selection didn't experience a sharp upward difficulty gradient after reaching the point of Homo erectus, it would be astonishing to find that human engineering could reach Homo erectus-level AIs (overcoming the multi-hundred-million-year cumulative lead natural selection had up until that point) but that human engineering then required more effort to get from there to a Homo sapiens equivalent. But wait: the human-engineering growth curve could be bounded below by the evolutionary curve while still having a different overall shape. For instance it could be that all the steps up to Homo erectus are much easier for human engineers than evolution-that the human difficulty curve over this region is far below the evolutionary curve-and then the steps from Homo erectus to Homo sapiens are only slightly easier for human engineers. That is, the human difficulty curve over this region is moderately below the evolutionary curve. Or to put it another way, we can imagine that Homo erectus 84. The reason this statement is not obvious is that an AI with general intelligence roughly at the level of Homo erectus might still have outsized abilities in computer programming-much as modern AIs have poor cross-domain intelligence, and yet there are still specialized chess AIs. Considering that blind evolution was able to build humans, it is not obvious that a sped-up Homo erectus AI with specialized programming abilities could not improve itself up to the level of Homo sapiens. was \"hard\" for natural selection and getting from there to Homo sapiens was \"easy,\" while both processes will be \"easy\" for human engineers, so that both steps will take place in eighty years each. Thus, the statement \"Creating intelligence will be much easier for human engineers than for evolution\" could imaginably be true in a world where \"It takes eighty years to get to Homo erectus AI and then another ninety years to get to Homo sapiens AI\" is also true. But one must distinguish possibility from probability. In probabilistic terms, I would be astonished if that actually happened, because there we have no observational reason to suppose that the relative difficulty curves actually look like that; specific complex irregularities with no observational support have low prior probability. When I imagine it concretely I'm also astonished: If you can build Homo erectus you can build the cerebral cortex, cerebellar cortex, the limbic system, the temporal lobes that perform object recognition, and so on. Human beings and chimpanzees have the vast majority of their neural architectures in common-such features have not diverged since the last common ancestor of humans and chimps. We have some degree of direct observational evidence that human intelligence is the icing on top of the cake that is chimpanzee intelligence. It would be surprising to be able to build that much cake and then find ourselves unable to make a relatively small amount of icing. The 80-90 hypothesis also requires that natural selection would have had an easier time building more sophisticated intelligences-equivalently, a harder time building less sophisticated intelligences-for reasons that wouldn't generalize over to human engineers, which further adds to the specific unsupported complex irregularity.  85  In general, I think we have specific reason to suspect that difficulty curves for natural selection bound above the difficulty curves for human engineers, and that humans will be able to access regions of design space blocked off from natural selection. I would expect early AIs to be in some sense intermediate between humans and natural selection in this sense, and for sufficiently advanced AIs to be further than humans along the same spectrum. Speculations which require specific unsupported irregularities of the relations between these curves should be treated as improbable; on the other hand, outcomes which would be yielded by many possible irregularities are much more probable, since the relations are bound to be irregular somewhere. It's possible that further analysis of this domain could yield more specific statements about expected relations between human engineering difficulty and evolutionary difficulty which would be relevant to AI timelines and growth curves. \n Anthropic Bias in Our Observation of Evolved Hominids The observation \"intelligence evolved\" may be misleading for anthropic reasons: perhaps evolving intelligence is incredibly difficult, but on all the planets where it doesn't evolve, there is nobody around to observe its absence.  Shulman and Bostrom (2012)  analyzed this question and its several possible answers given the present state of controversy regarding how to reason about anthropic probabilities. Stripping out a number of caveats and simplifying, it turns out that-under assumptions that yield any adjustment at all for anthropic bias-the main conclusion we can draw is a variant of  Hanson's (1998c)  conclusion: if there are several \"hard steps\" in the evolution of intelligence, then planets on which intelligent life does evolve should expect to see the hard steps spaced about equally across their history, regardless of each step's relative difficulty. Suppose a large population of lockpickers are trying to solve a series of five locks in five hours, but each lock has an average solution time longer than five hours-requiring ten hours or a hundred hours in the average case. Then the few lockpickers lucky enough to solve every lock will probably see the five locks distributed randomly across the record. Conditioning on the fact that a lockpicker was lucky enough to solve the five locks at all, a hard lock with an average solution time of ten hours and a hard lock with an average solution time of one hundred hours will have the same expected solution times selecting on the cases where all locks were solved.  86  This in turn means that \"self-replicating life comes into existence\" or \"multicellular organisms arise\" are plausible hard steps in the evolution of intelligent life on Earth, but the time interval from Australopithecus to Homo sapiens is too short to be a plausible hard step. There might be a hard step along the way to first reaching Australopithecus intelligence, 86. I think a legitimate simplified illustration of this result is that, given a solution time for lock A evenly distributed between 0 hours and 200 hours and lock B with a solution time evenly distributed between 0 hours and 20 hours, then conditioning on the fact that A and B were both successfully solved in a total of 2 hours, we get equal numbers for \"the joint probability that A was solved in 1.5-1.6 hours and B was solved in 0.4-0.5 hours\" and \"the joint probability that A was solved in 0.4-0.5 hours and B was solved in 1.5-1.6 hours,\" even though in both cases the probability for A being solved that fast is one-tenth the probability for B being solved that fast. but from chimpanzee-equivalent intelligence to humans was apparently smooth sailing for natural selection (or at least the sailing was probably around as smooth or as choppy as the \"naive\" perspective would have indicated before anthropic adjustments). Nearly the same statement could be made about the interval from mouse-equivalent ancestors to humans, since fifty million years is short enough for a hard step to be improbable, though not quite impossible. On the other hand, the gap from spiders to lizards might more plausibly contain a hard step whose difficulty is hidden from us by anthropic bias. What does this say about models of the intelligence explosion? Difficulty curves for evolution and for human engineering cannot reasonably be expected to move in lockstep. Hard steps for evolution are not necessarily hard steps for human engineers (recall the case of freely rotating wheels). Even if there has been an evolutionarily hard step on the road to mice-a hard step that reduced the number of planets with mice by a factor of 10 50 , emptied most galactic superclusters of mice, and explains the Great Silence we observe in the night sky-it might still be something that a human engineer can do without difficulty.  87  If natural selection requires 10 100 tries to do something but eventually succeeds, the problem still can't be that hard in an absolute sense, because evolution is still pretty stupid. There is also the possibility that we could reverse-engineer actual mice. I think the role of reverse-engineering biology is often overstated in Artificial Intelligence, but if the problem turns out to be incredibly hard for mysterious reasons, we do have mice on hand. Thus an evolutionarily hard step would be relatively unlikely to represent a permanent barrier to human engineers. All this only speaks of a barrier along the pathway to producing mice. One reason I don't much modify my model of the intelligence explosion to compensate for possible anthropic bias is that a humanly difficult barrier below the mouse level looks from the outside like, \"Gosh, we've had lizard-equivalent AI for twenty years now and we still can't get to mice, we may have to reverse-engineer actual mice instead of figuring this out on our own.\" 88 But the advice from anthropics is that the road from mice to humans is no more difficult than it looks, so a \"hard step\" which slowed down an intelligence explosion in progress would presumably have to strike before that intelligence explosion hit the mouse level.  89  Suppose an intelligence explosion could in fact get started beneath the mouse level-perhaps a specialized programming AI with sub-mouse general intelligence and high serial speeds might be able make significant self-improvements. Then from the outside we would see something like, \"Huh, we can build these relatively dumb specialized AIs that seem to get significant mileage out of recursive self-improvement, but then everything we build bottlenecks around the same sub-mouse level.\" If we tried hard to derive policy advice from this anthropic point, it might say: \"If tomorrow's AI researchers can build relatively dumb self-modifying systems that often manage to undergo long chains of significant self-improvement with reinvested returns, and they all get stuck at around the same point somewhere below mouse-level general intelligence, then it's possible that this point is the \"hard step\" from evolutionary history, rather than a place where the difficulty curve permanently slopes upward. You should potentially worry about the first AI that gets pushed past this big sticking point, because once you do get to mice, it may be an easy journey onward from there.\" I'm not sure I' d have very much confidence in that advice-it seems to have been obtained via a complicated argument and I don't see a good way to simplify the core idea. But since I wouldn't otherwise expect this kind of bottlenecking to be uniform across many different AI systems, that part is arguably a unique prediction of the hard-step model where some small overlooked lock actually contains a thousand cosmic hours of average required solution time. For the most part, though, it appears to me that anthropic arguments do not offer very detailed advice about the intelligence explosion (and this is mostly to be expected). \n Local versus Distributed Intelligence Explosions A key component of the debate between Robin Hanson and myself was the question of locality. Consider: If there are increasing returns on knowledge given constant human you are properly appreciating a scale that runs from \"rock\" at zero to \"bacterium\" to \"spider\" to \"lizard\" to \"mouse\" to \"chimp\" to \"human,\" then AI seems to be moving along at a slow but steady pace. (At least it's slow and steady on a human R&D scale. On an evolutionary scale of time, progress in AI has been unthinkably, blindingly fast over the past sixty-year instant.) The \"hard step\" theory does say that we might expect some further mysterious bottleneck, short of mice, to a greater degree than we would expect if not for the Great Silence. But such a bottleneck might still not correspond to a huge amount of time for human engineers. \n 89. A further complicated possible exception is if we can get far ahead of lizards in some respects, but are missing one vital thing that mice do. Say, we already have algorithms which can find large prime numbers much faster than lizards, but still can't eat cheese. brains-this being the main assumption that many non-intelligence-explosion, general technological hypergrowth models rely on, with said assumption seemingly well-supported by exponential 90 technology-driven productivity growth 91 -then why isn't the leading human nation vastly ahead of the runner-up economy? Shouldn't the economy with the most knowledge be rising further and further ahead of its next-leading competitor, as its increasing returns compound? The obvious answer is that knowledge is not contained within the borders of one country: improvements within one country soon make their way across borders. Thus we can sketch two widely different possible scenarios for an intelligence explosion, at opposite extremes along multiple dimensions, as follows: 94 Extremely local takeoff: • Much like today, the diversity of advanced AI architectures is so great that there is very little trading of cognitive content between projects. It's easier to download a large dataset, and have your AI relearn the lessons of that dataset within its own cognitive representation, than to trade cognitive content between different AIs. To the extent that AIs other than the most advanced project can generate selfimprovements at all, they generate modifications of idiosyncratic code that can't be cheaply shared with any other AIs. • The leading projects do not publish all or even most of their research-whether for the same reasons hedge funds keep their sauces secret, or for the same reason Leo Szilard didn't immediately tell the world about fission chain reactions. • There is a relatively small number of leading projects. • The first AI to touch the intelligence explosion reaches k > 1 due to a basic algorithmic improvement that hasn't been shared with any other projects. • The AI has a sufficiently clean architecture that it can scale onto increasing amounts of hardware while remaining as a unified optimization process capable of pursuing coherent overall goals. • The AI's self-improvement, and eventual transition to rapid infrastructure, involves a large spike in capacity toward the latter end of the curve (as superintelligence is achieved, or as protein structure prediction is cracked sufficiently to build later stages of nanotechnology). This vastly amplifies the AI's cognitive and technological lead time over its nearest competitor. If the nearest competitor was previously only seven days behind, these seven days have now been amplified into a technological gulf enabling the leading AI to shut down, sandbox, or restrict the growth 93. Theoretically, genes can sometimes jump this sort of gap via viruses that infect one species, pick up some genes, and then infect a member of another species. Speaking quantitatively and practically, the amount of gene transfer between hominids and chimps was approximately zero so far as anyone knows. 94. Again, neither of these possibilities should be labeled \"good\" or \"bad\"; we should make the best of whatever reality we turn out to live in, whatever the settings of the hidden variables. of any competitors it wishes to fetter. The final result is a Bostrom (2006)-style \"singleton.\" \n Extremely global takeoff: • The emergence of good, successful machine intelligence techniques greatly winnows the plethora of visionary prototypes we see nowadays  (Hanson 2008b) . AIs are similar enough that they can freely trade cognitive content, code tweaks, and algorithmic improvements. • There are many, many such AI projects. • The vast majority of \"improvement\" pressure on any single machine intelligence derives from the total global economy of machine intelligences or from academic AI researchers publishing their results, not from that AI's internal self-modifications. Although the global economy of machine intelligences is getting high returns on cognitive investments, no single part of that economy can go FOOM by itself. • Any sufficiently large machine intelligence is forced by lack of internal bandwidth to split into pieces, which then have their own local goals and do not act as a well-coordinated whole. • The benefit that an AI can derive from local use of an innovation is very small compared to the benefit that it can get from selling the innovation to many different AIs. Thus, very few innovations are kept secret. (The same reason that when Stephen King writes a novel, he sells the novel to hundreds of thousands of readers and uses the proceeds to buy more books, instead of just keeping the novel to himself.) • Returns on investment for machine intelligences which fall behind automatically increase as the machine is enabled to \"catch up\" on cheaper knowledge (much as China is growing faster than Australia). Also, leading agencies do not eliminate laggards or agglomerate them (the way strong countries used to conquer weak countries). • Nobody knows how to 90%-solve the protein structure prediction problem before somebody else knows how to 88%-solve the protein structure prediction problem; relative leads are small. Even technologies like molecular nanotech appear gradually and over many different places at once, with much sharing/selling of innovations and laggards catching up; relative leads are not significantly amplified by the transition. • The end result has a lot of trade and no global coordination. (This is not necessarily a good thing. See  Hanson's [2008d]  rapacious hardscrapple frontier folk.) These two extremes differ along many dimensions that could potentially fail to be correlated. Note especially that sufficiently huge returns on cognitive reinvestment will produce winner-take-all models and a local FOOM regardless of other variables. To make this so extreme that even I don't think it's plausible, if there's a simple trick that lets you get molecular nanotechnology and superintelligence five seconds after you find it, 95 then it's implausible that the next runner-up will happen to find it in the same five-second window.  96  Considering five seconds as a literal time period rather than as a metaphor, it seems clear that sufficiently high returns on reinvestment produce singletons almost regardless of other variables. (Except possibly for the stance \"sufficiently large minds must inevitably split into bickering components,\" which could hold even in this case.  97  ) It should also be noted that the \"global\" scenario need not include all of the previous civilization inside its globe. Specifically, biological humans running on 200 Hz neurons with no read-write ports would tend to be left out of the FOOM, unless some AIs are specifically motivated to help humans as a matter of final preferences. Newly discovered cognitive algorithms do not easily transfer over to human brains with no USB ports. Under this scenario humans would be the equivalent of emerging countries with dreadfully restrictive laws preventing capital inflows, which can stay poor indefinitely. Even if it were possible to make cognitive improvements cross the \"human barrier,\" it seems unlikely to offer the highest natural return on investment compared to investing in a fellow machine intelligence. In principle you can evade the guards and sneak past the borders of North Korea and set up a convenience store where North Koreans can buy the same goods available elsewhere. But this won't be the best way to invest your money-not unless you care about North Koreans as a matter of final preferences over terminal outcomes.  98  95. À la The Metamorphosis of Prime Intellect by Roger  Williams (2002) . 96. A rational agency has no convergent instrumental motive to sell a sufficiently powerful, rapidly reinvestable discovery to another agency of differing goals, because even if that other agency would pay a billion dollars for the discovery in one second, you can get a larger fraction of the universe to yourself and hence even higher total returns by keeping mum for the five seconds required to fully exploit the discovery yourself and take over the universe. 97. This stance delves into AI-motivational issues beyond the scope of this paper. I will quickly note that the Orthogonality Thesis opposes the assertion that any \"mind\" must develop indexically selfish preferences which would prevent coordination, even if it were to be granted that a \"mind\" has a maximum individual size. Mostly I would tend to regard the idea as anthropomorphic-humans have indexically selfish preferences and group conflicts for clear evolutionary reasons, but insect colonies with unified genetic destinies and whole human brains (likewise with a single genome controlling all neurons) don't seem to have analogous coordination problems. 98. Our work on decision theory also suggests that the best coordination solutions for computer-based minds would involve knowledge of each others' source code or crisp adoption of particular crisp decision The highly local scenario obviously offers its own challenges as well. In this case we mainly want the lead project at any given point to be putting sufficiently great efforts into \"Friendly AI.\"  99  In the highly global scenario we get incremental improvements by having only some AIs be human-Friendly, 100 while the local scenario is winner-take-all. (But to have one AI of many be Friendly does still require that someone, somewhere solve the associated technical problem before the global AI ecology goes FOOM; and relatively larger returns on cognitive reinvestment would narrow the amount of time available to do solve that problem.) My own expectations lean toward scenario (1)-for instance, I usually use the singular rather than plural when talking about that-which-goes-FOOM. This is mostly because I expect large enough returns on cognitive reinvestment to dominate much of my uncertainty about other variables. To a lesser degree I am impressed by the diversity and incompatibility of modern approaches to machine intelligence, but on this score I respect Hanson's argument for why this might be expected to change. The rise of open-source chess-playing programs has undeniably led to faster progress due to more sharing of algorithmic improvements, and this combined with Hanson's argument has shifted me significantly toward thinking that the ecological scenario is not completely unthinkable. It's also possible that the difference between local-trending and global-trending outcomes is narrow enough to depend on policy decisions. That is, the settings on the hidden variables might turn out to be such that, if we wanted to see a \"Friendly singleton\" rather than a Hansonian \"rapacious hardscrapple frontier\" of competing AIs, it would be feasible to create a \"nice\" project with enough of a research advantage (funding, computing resources, smart researchers) over the next runner-up among non-\"nice\" theories. Here it is much harder to verify that a human is trustworthy and will abide by their agreements, meaning that humans might \"naturally\" tend to be left out of whatever coordination equilibria develop among machine-based minds, again unless there are specific final preferences to include humans. 99. The Fragility of Value subthesis of Complexity of Value implies that solving the Friendliness problem is a mostly satisficing problem with a sharp threshold, just as dialing nine-tenths of my phone number correctly does not connect you to someone 90% similar to Eliezer Yudkowsky. If the fragility thesis is correct, we are not strongly motivated to have the lead project be 1% better at Friendly AI than the runner-up project; rather we are strongly motivated to have it do \"well enough\" (though this should preferably include some error margin). Unfortunately, the Complexity of Value thesis implies that \"good enough\" Friendliness involves great (though finite) difficulty. 100. Say, one Friendly AI out of a million cooperating machine intelligences implies that one millionth of the universe will be used for purposes that humans find valuable. This is actually quite a lot of matter and energy, and anyone who felt diminishing returns on population or lifespan would probably regard this scenario as carrying with it most of the utility. competitors to later become a singleton.  101  This could be true even in a world where a global scenario would be the default outcome (e.g., from open-source AI projects) so long as the hidden variables are not too heavily skewed in that direction. \n Minimal Conditions to Spark an Intelligence Explosion I. J. Good spoke of the intelligence explosion beginning from an \"ultraintelligence . . . a machine that can far surpass all the intellectual activities of any man however clever.\" This condition seems sufficient, but far more than necessary. Natural selection does not far surpass every intellectual capacity of any human-it cannot write learned papers on computer science and cognitive algorithms-and yet it burped out a human-equivalent intelligence anyway.  102  Indeed, natural selection built humans via an optimization process of point mutation, random recombination, and statistical selection-without foresight, explicit world-modeling, or cognitive abstraction. This quite strongly upper-bounds the algorithmic sophistication required, in principle, to output a design for a human-level intelligence. Natural selection did use vast amounts of computational brute force to build humans. The \"naive\" estimate is that natural selection searched in the range of 10 30 to 10 40 organisms before stumbling upon humans  (Baum 2004) . Anthropic considerations (did other planets have life but not intelligent life?) mean the real figure might be almost arbitrarily higher (see section 3.8). There is a significant subfield of machine learning that deploys evolutionary computation (optimization algorithms inspired by mutation/recombination/selection) to try to solve real-world problems. The toolbox in this field includes \"improved\" genetic algorithms which, at least in some cases, seem to evolve solutions orders of magnitude faster than the first kind of \"evolutionary\" algorithm you might be tempted to write (for example, the Bayesian Optimization Algorithm of  Pelikan, Goldberg, and Cantú-Paz [2000] ). However, if you expect to be able to take an evolutionary computation and have it output an organism on the order of, say, a spider, you will be vastly disappointed. It took roughly a billion years after the start of life for complex cells to arise. Genetic algorithms can design interesting radio antennas, analogous perhaps to a particular chemical enzyme. But even with their hundredfold speedups, modern genetic algorithms seem to be using vastly too little brute force to make it out of the RNA world, let alone reach 101. If intelligence explosion microeconomics tells us that algorithmic advantages are large compared to hardware, then we care most about \"nice\" projects having the smartest researchers. If hardware advantages are large compared to plausible variance in researcher intelligence, this makes us care more about \"nice\" projects having the most access to computing resources. the Cambrian explosion. To design a spider-equivalent brain would be far beyond the reach of the cumulative optimization power of current evolutionary algorithms running on current hardware for reasonable periods of time. On the other side of the spectrum, human engineers quite often beat natural selection in particular capacities, even though human engineers have been around for only a tiny fraction of the time. (Wheel beats cheetah, skyscraper beats redwood tree, Saturn V beats falcon, etc.) It seems quite plausible that human engineers, working for an amount of time (or even depth of serial causality) that was small compared to the total number of evolutionary generations, could successfully create human-equivalent intelligence. However, current AI algorithms fall far short of this level of . . . let's call it \"taking advantage of the regularity of the search space,\" although that's only one possible story about human intelligence. Even branching out into all the fields of AI that try to automatically design small systems, it seems clear that automated design currently falls very far short of human design. Neither current AI algorithms running on current hardware nor human engineers working on AI for sixty years or so have yet sparked a FOOM. We know two combinations of \"algorithm intelligence + amount of search\" that haven't output enough cumulative optimization power to spark a FOOM. But this allows a great deal of room for the possibility that an AI significantly more \"efficient\" than natural selection, while significantly less \"intelligent\" than human computer scientists, could start going FOOM. Perhaps the AI would make less intelligent optimizations than human computer scientists, but it would make many more such optimizations. And the AI would search many fewer individual points in design space than natural selection searched organisms, but traverse the search space more efficiently than natural selection. And, unlike either natural selection or humans, each improvement that the AI found could be immediately reinvested in its future searches. After natural selection built Homo erectus, it was not then using Homo erectus-level intelligence to consider future DNA modifications. So it might not take very much more intelligence than natural selection for an AI to first build something significantly better than itself, which would then deploy more intelligence to building future successors. In my present state of knowledge I lack strong information to not worry about random AI designs crossing any point on the frontier of \"more points searched than any past algorithm of equal or greater intelligence (including human computer scientists), and more intelligence than any past algorithm which has searched an equal number of cases (including natural selection).\" This frontier is advanced all the time and no FOOM has yet occurred, so, by Laplace's Rule of Succession or similar ignorance priors, we should assign much less than 50% probability that the next crossing goes FOOM. On the other hand we should assign a much higher chance that some crossing of the frontier of \"efficiency cross computation\" or \"intelligence cross brute force\" starts an intelligence explosion at some point in the next N decades. Our knowledge so far also holds room for the possibility that, without unaffordably vast amounts of computation, semi-intelligent optimizations cannot reinvest and cumulate up to human-equivalent intelligence-any more than you can get a FOOM by repeatedly running an optimizing compiler over itself. The theory here is that mice would have a hard time doing better than chance at modifying mice. In this class of scenarios, for any reasonable amount of computation which research projects can afford (even after taking Moore's Law into account), you can't make an AI that builds better AIs than any human computer scientist until that AI is smart enough to actually do computer science. In this regime of possibility, human computer scientists must keep developing their own improvements to the AI until that AI reaches the point of being able to do human-competitive computer science, because until then the AI is not capable of doing very much pushing on its own.  103  Conversely, to upper-bound the FOOM-starting level, consider the AI equivalent of John von Neumann exploring computer science to greater serial depth and parallel width than previous AI designers ever managed. One would expect this AI to spark an intelligence explosion if it can happen at all. In this case we are going beyond the frontier of the number of optimizations and the quality of optimizations for humans, so if this AI can't build something better than itself, neither can humans. The \"fast parallel von Neumann\" seems like a reasonable pragmatic upper bound on how smart a machine intelligence could be without being able to access an intelligence explosion, or how smart it could be before the intelligence explosion entered a prompt-supercritical mode, assuming this to be possible at all. As it's unlikely for true values to exactly hit upper bounds, I would guess that the intelligence explosion would start well before then. Relative to my current state of great uncertainty, my median estimate would be somewhere in the middle: that it takes much more than an improved optimizing compiler or improved genetic algorithm, but significantly less than a fast parallel von Neumann, 103. \"Nice\" AI proposals are likely to deliberately look like this scenario, because in Friendly AI we may want to do things like have the AI prove a self-modification correct with respect to a criterion of actionhave the AI hold itself to a high standard of self-understanding so that it can change itself in ways which preserve important qualities of its design. This probably implies a large added delay in when a \"nice\" project can allow its AI to do certain kinds of self-improvement, a significant handicap over less restrained competitors even if the project otherwise has more hardware or smarter researchers. (Though to the extent that you can \"sanitize\" suggestions or show that a class of improvements can't cause catastrophic errors, a Friendly AI under development may be able to wield significant self-improvements even without being able to do computer science.) to spark an intelligence explosion (in a non-Friendly AI project; a Friendly AI project deliberately requires extra computer science ability in the AI before it is allowed to selfmodify). This distribution is based mostly on prior ignorance, but the range seems wide and so the subranges close to the endpoints should be relatively narrow. All of this range falls well short of what I. J. Good defined as \"ultraintelligence.\" An AI which is merely as good as a fast parallel von Neumann at building AIs need not far surpass humans in all intellectual activities of every sort. For example, it might be very good at computer science while not yet being very good at charismatic manipulation of humans. I. J. Good focused on an assumption that seems far more than sufficient to yield his conclusion of the intelligence explosion, and this unfortunately may be distracting relative to much weaker assumptions that would probably suffice. \n Returns on Unknown Unknowns Molecular nanotechnology is a fairly recent concept and nineteenth-century humans didn't see it coming. There is an important albeit dangerous analogy which says that the twenty-first century can do magic relative to the eleventh century, and yet a thousand years isn't very much time; that to chimpanzees humans are just plain incomprehensible, yet our brain designs aren't even all that different; and that we should therefore assign significant probability that returns on increased speed (serial time, causal depth, more of that distance which separates the twenty-first and eleventh centuries of human history) or improved brain algorithms (more of that which separates hominids from chimpanzees) will end up delivering damn near anything in terms of capability. This may even include capabilities that violate what we currently believe to be the laws of physics, since we may not know all the relevant laws. Of course, just because our standard model of physics might be wrong somewhere, we cannot conclude that any particular error is probable. And new discoveries need not deliver positive news; modern-day physics implies many restrictions the nineteenth century didn't know about, like the speed-of-light limit. Nonetheless, a rational agency will selectively seek out useful physical possibilities we don't know about; it will deliberately exploit any laws we do not know. It is not supernaturalism to suspect, in full generality, that future capabilities may somewhere exceed what the twenty-first-century Standard Model implies to be an upper bound. An important caveat is that if faster-than-light travel is possible by any means whatsoever, the Great Silence/Fermi Paradox (\"Where are they?\") becomes much harder to explain. This gives us some reason to believe that nobody will ever discover any form of \"magic\" that enables FTL travel (unless it requires an FTL receiver that must itself travel at slower-than-light speeds). More generally, it gives us a further reason to doubt any future magic in the form of \"your physicists didn't know about X, and therefore it is possible to do Y\" that would give many agencies an opportunity to do Y in an observable fashion. We have further reason in addition to our confidence in modern-day physics to believe that time travel is not possible (at least no form of time travel which lets you travel back to before the time machine was built), and that there is no tiny loophole anywhere in reality which even a superintelligence could exploit to enable this, since our present world is not full of time travelers. More generally, the fact that a rational agency will systematically and selectively seek out previously unknown opportunities for unusually high returns on investment says that the expectation of unknown unknowns should generally drive expected returns upward when dealing with something smarter than us. The true laws of physics might also imply exceptionally bad investment possibilities-maybe even investments worse than the eleventh century would have imagined possible, like a derivative contract that costs only a penny but can lose a quadrillion dollars-but a superintelligence will not be especially interested in those. Unknown unknowns add generic variance, but rational agencies will select on that variance in a positive direction. From my perspective, the possibility of \"returns on unknown unknowns,\" \"returns on magic,\" or \"returns on the superintelligence being smarter than I am and thinking of possibilities I just didn't see coming\" mainly tells me that (1) intelligence explosions might go FOOM faster than I expect, (2) trying to bound the real-world capability of an agency smarter than you are is unreliable in a fundamental sense, and (3) we probably only get one chance to build something smarter than us that is not uncaring with respect to the properties of the future we care about. But I already believed all that; so, from my perspective, considering the possibility of unknown unknown returns adds little further marginal advice. Someone else with other background beliefs might propose a wholly different policy whose desirability, given their other beliefs, would hinge mainly on the absence of such unknown unknowns-in other words, it would be a policy whose workability rested on the policy proposer's ability to have successfully bounded the space of opportunities of some smarter-than-human agency. This would result in a rationally unpleasant sort of situation, in the sense that the \"argument from unknown unknown returns\" seems like it ought to be impossible to defeat, and for an argument to be impossible to defeat means that it is insensitive to reality. 104 I am tempted to say at this point, \"Thankfully, that is not my concern, since my policy proposals are already meant to be optimal replies in the case that a superintelligence can think of something I haven't.\" But, despite temptation, this brush-off seems inadequately sympathetic to the other side of the debate. And I am not properly sure what sort of procedure ought to be put in place for arguing about the possibility of \"returns on unknown unknowns\" such that, in a world where there were in fact no significant returns on unknown unknowns, you would be able to figure out with high probability that there were no unknown unknown returns, and plan accordingly. I do think that proposals which rely on bounding smarter-than-human capacities may reflect a lack of proper appreciation and respect for the notion of something that is really actually smarter than you. But it is also not true that the prospect of unknown unknowns means we should assign probability one to a being marginally smarter than human taking over the universe in five seconds, and it is not clear what our actual probability distribution should be over lesser \"impossibilities.\" It is not coincidence that I picked my policy proposal so as not to be highly sensitive to that estimate. \n Three Steps Toward Formality Lucio  Russo (2004) , in a book arguing that science was invented two millennia ago and then forgotten, defines an exact science as a body of theoretical postulates whose consequences can be arrived at by unambiguous deduction, which deductive consequences can then be further related to objects in the real world. For instance, by this definition, Euclidean geometry can be viewed as one of the earliest exact sciences, since it proceeds from postulates but also tells us what to expect when we measure the three angles of a real-world triangle. Broadly speaking, to the degree that a theory is formal, it is possible to say what the theory predicts without argument, even if we are still arguing about whether the theory is actually true. In some cases a theory may be laid out in seemingly formal axioms, and yet its relation to experience-to directly observable facts-may have sufficient flex that people are still arguing over whether or not an agreed-on formal prediction has actually come true. 105 This is often the case in economics: there are many formally specified models of macroeconomics, and yet their relation to experience is ambiguous enough that it's hard to tell which ones, if any, are approximately true. without unknown unknowns, hence its appearance in the final subsection, is not going to have any effect on the repetition of this wonderful counterargument. 105. Another edge case is a formally exact theory whose precise predictions we lack the computing power to calculate, causing people to argue over the deductive consequences of the theory even though the theory's axioms have been fully specified. What is the point of formality? One answer would be that by making a theory formal, we can compute exact predictions that we couldn't calculate using an intuition in the back of our minds. On a good day, these exact predictions may be unambiguously relatable to experience, and on a truly wonderful day the predictions actually come true. But this is not the only possible reason why formality is helpful. To make the consequences of a theory subject to unambiguous deduction-even when there is then some further argument over how to relate these consequences to experience-we have to make the machinery of the theory explicit; we have to move it out of the back of our minds and write it out on paper, where it can then be subject to greater scrutiny. This is probably where we will find most of the benefit from trying to analyze the intelligence explosion more formally-it will expose the required internal machinery of arguments previously made informally. It might also tell us startling consequences of propositions we previously said were highly plausible, which we would overlook if we held the whole theory inside our intuitive minds. With that said, I would suggest approaching the general problem of formalizing previously informal stances on the intelligence explosion as follows: 1. Translate stances into microfoundational hypotheses about growth curvesquantitative functions relating cumulative investment and output. Different stances may have different notions of \"investment\" and \"output,\" and different notions of how growth curves feed into each other. We want elementary possibilities to be specified with sufficient rigor that their consequences are formal deductions rather than human judgments: in the possibility that X goes as the exponential of Y, then, supposing Y already quantified, the alleged quantity of X should follow as a matter of calculation rather than judgment. 2. Explicitly specify how any particular stance claims that (combinations of ) growth curves should allegedly relate to historical observations or other known facts. Quantify the relevant historical observations in a format that can be directly compared to the formal possibilities of a theory, making it possible to formalize a stance's claim that some possibilities in a range are falsified. 3. Make explicit any further assumptions of the stance about the regularity or irregularity (or prior probability) of elementary possibilities. Make explicit any coherence assumptions of a stance about how different possibilities probably constrain each other (curve A should be under curve B, or should have the same shape as curve C). 106 In the step about relating historical experience to the possibilities of the theory, allowing falsification or updating is importantly not the same as curve-fitting-it's not like trying to come up with a single curve that \"best\" fits the course of hominid evolution or some such. Hypothesizing that we know a single, exact curve seems like it should be overrunning the state of our knowledge in many cases; for example, we shouldn't pretend to know exactly how difficult it was for natural selection to go from Homo erectus to Homo sapiens. To get back a prediction with appropriately wide credible intervalsa prediction that accurately represents a state of uncertainty-there should be some space of regular curves in the model space, with combinations of those curves related to particular historical phenomena. In principle, we would then falsify the combinations that fail to match observed history, and integrate (or sample) over what's left to arrive at a prediction. Some widely known positions on the intelligence explosion do rely on tightly fitting a curve (e.g., Moore's Law). This is not completely absurd because some historical curves have in fact been highly regular (e.g., Moore's Law). By passing to Bayesian updating instead of just falsification, we could promote parts of the model space that narrowly predict an observed curve-parts of the model space which concentrated more of their probability mass into predicting that exact outcome. This would expose assumptions about likelihood functions and make more visible whether it's reasonable or unreasonable to suppose that a curve is precise; if we do a Bayesian update on the past, do we get narrow predictions for the future? What do we need to assume to get narrow predictions for the future? How steady has Moore's Law actually been for the past?-because if our modeling technique can't produce even that much steadiness, and produces wide credible intervals going off in all directions, then we're not updating hard enough or we have overly ignorant priors. Step One would be to separately carry out this process on one or more current stances and speakers, so as to reveal and quantify the formal assumptions underlying their arguments. At the end of Step One, you would be able to say, \"This is a model space that looks like what Speaker X was talking about; these are the growth curves or combinations of growth curves that X considers falsified by these historical experiences, or that X gives strong Bayesian updates based on their narrow predictions of historical experiences; this is what X thinks about how these possibilities are constrained to be coherent with each other; and this is what X thinks is the resulting prediction made over the intelligence explosion by the nonfalsified, coherent parts of the model space.\" Step One of formalization roughly corresponds to seeing if there's any set of curves by which a speaker's argument could make sense; making explicit the occasions where someone else has argued that possibilities are excluded by past experience; and exposing any suspicious irregularities in the curves being postulated. Step One wouldn't yield definitive answers about the intelligence explosion, but should force assumptions to be more clearly stated, potentially expose some absurdities, show what else a set of assumptions implies, etc. Mostly, Step One is about explicitizing stances on the intelligence explosion, with each stance considered individually and in isolation. Step Two would be to try to have a common, integrated model of multiple stances formalized in Step One-a model that included many different possible kinds of growth curves, some of which might be (in some views) already falsified by historical observations-a common pool of building blocks that could be selected and snapped together to produce the individual formalizations from Step One. The main products of Step Two would be (a) a systematic common format for talking about plausible growth curves and (b) a large table of which assumptions yield which outcomes (allegedly, according to the compiler of the table) and which historical observations various arguments allege to pose problems for those assumptions. I would consider this step to be about making explicit the comparison between theories: exposing arguable irregularities that exist in one stance but not another and giving readers a better position from which to evaluate supposed better matches versus simpler hypotheses. Step Two should not yet try to take strong positions on the relative plausibility of arguments, nor to yield definitive predictions about the intelligence explosion. Rather, the goal is to make comparisons between stances more formal and more modular, without leaving out any important aspects of the informal arguments-to formalize the conflicts between stances in a unified representation. Step Three would be the much more ambitious project of coming up with an allegedly uniquely correct description of our state of uncertain belief about the intelligence explosion: • Formalize a model space broad enough to probably contain something like reality, with credible hope of containing a point hypothesis in its space that would well fit, if not exactly represent, whatever causal process actually turns out to underlie the intelligence explosion. That is, the model space would not be so narrow that, if the real-world growth curve were actually hyperbolic up to its upper bound, we would have to kick ourselves afterward for having no combinations of assumptions in the model that could possibly yield a hyperbolic curve. 107 • Over this model space, weight prior probability by simplicity and regularity. 107. In other words, the goal would be to avoid errors of the class \"nothing like the reality was in your hypothesis space at all.\" There are many important theorems of Bayesian probability that do not apply when nothing like reality is in your hypothesis space. • Relate combinations of causal hypotheses to observed history and do Bayesian updates. • Sample the updated model space to get a probability distribution over the answers to any query we care to ask about the intelligence explosion. • Tweak bits of the model to get a sensitivity analysis of how much the answers tend to vary when you model things slightly differently, delete parts of the model to see how well the coherence assumptions can predict the deleted parts from the remaining parts, etc. \n If Step Three is done wisely-with the priors reflecting an appropriate breadth of uncertainty-and doesn't entirely founder on the basic difficulties of formal statistical learning when data is scarce, then I would expect any such formalization to yield mostly qualitative yes-or-no answers about a rare handful of answerable questions, rather than yielding narrow credible intervals about exactly how the internal processes of the intelligence explosion will run. A handful of yeses and nos is about the level of advance prediction that I think a reasonably achievable grasp on the subject should allow-we shouldn't know most things about intelligence explosions this far in advance of observing one-we should just have a few rare cases of questions that have highly probable if crude answers. I think that one such answer is \"AI go FOOM? Yes! AI go FOOM!\" but I make no pretense of being able to state that it will proceed at a rate of 120,000 nanofooms per second. Even at that level, covering the model space, producing a reasonable simplicity weighting, correctly hooking up historical experiences to allow falsification and updating, and getting back the rational predictions would be a rather ambitious endeavor that would be easy to get wrong. Nonetheless, I think that Step Three describes in principle what the ideal Bayesian answer would be, given our current collection of observations. In other words, the reason I endorse an AI-go-FOOM answer is that I think that our historical experiences falsify most regular growth curves over cognitive investments that wouldn't produce a FOOM. Academic disputes are usually not definitively settled once somebody advances to the stage of producing a simulation. It's worth noting that macroeconomists are still arguing over, for example, whether inflation or NGDP should be stabilized to maximize real growth. On the other hand, macroeconomists usually want more precise answers than we could reasonably demand from predictions about the intelligence explosion. If you'll settle for model predictions like, \"Er, maybe inflation ought to increase rather than decrease when banks make noticeably more loans, ceteris paribus?\" then it might be more reasonable to expect definitive answers, compared to asking whether inflation will be more or less than 2.3%. But even if you tried to build the Step Three model, it might still be a bit naive to think that you would really get the answers back out, let alone expect that everyone else would trust your model. In my case, I think how much I trusted a Step Three model would depend a lot on how well its arguments simplified, while still yielding the same net predictions and managing not to be falsified by history. I trust complicated arguments much more when they have simple versions that give mostly the same answers; I would trust my arguments about growth curves less if there weren't also the simpler version, \"Smart minds build even smarter minds.\" If the model told me something I hadn't expected, but I could translate the same argument back into simpler language and the model produced similar results even when given a few cross-validational shoves, I' d probably believe it. Regardless, we can legitimately hope that finishing Step One, going on to Step Two, and pushing toward Step Three will yield interesting results, even if Step Three is never completed or is completed several different ways.  108  The main point of formality isn't that it gives you final and authoritative answers, but that it sometimes turns up points you wouldn't have found without trying to make things explicit. \n Expected Information Value: What We Want to Know versus What We Can Probably Figure Out There tend to be mismatches between what we want to know about the intelligence explosion, and what we can reasonably hope to figure out. For example, everyone at the Machine Intelligence Research Institute (MIRI) would love to know how much time remained until an intelligence explosion would probably be produced by general progress in the field of AI. It would be extremely useful knowledge from a policy perspective, and if you could time it down to the exact year, you could run up lots of credit card debt just beforehand.  109  But-unlike a number of other futuristswe don't see how we could reasonably obtain strong information about this question. Hans Moravec, one of the first major names to predict strong AI using Moore's Law, spent much of his  (1988)  book Mind Children trying to convince readers of the incredible proposition that Moore's Law could actually go on continuing and continuing and continuing until it produced supercomputers that could do-gasp!-a hundred teraflops. Which was enough to \"equal the computing power of the human brain,\" as Moravec had calculated that equivalency in some detail using what was then known about the visual cortex and how hard that part was to simulate. We got the supercomputers that 108. \"A man with one watch knows what time it is; a man with two watches is never sure.\" 109. Yes, that is a joke. Moravec thought were necessary in 2008, several years earlier than Moravec's prediction; but, as it turned out, the way reality works is not that the universe checks whether your supercomputer is large enough and then switches on its consciousness.  110  Even if it were a matter of hardware rather than mostly software, the threshold level of \"required hardware\" would be far more uncertain than Moore's Law, and a predictable number times an unpredictable number is an unpredictable number. So, although there is an extremely high value of information about default AI time- (Even this kind of \"I don't know\" still has to correspond to some probability distribution over decades, just not a tight distribution. I'm currently trying to sort out with Carl Shulman why my median is forty-five years in advance of his median. Neither of us thinks we can time it down to the decade-we have very broad credible intervals in both cases-but the discrepancy between our \"I don't knows\" is too large to ignore.) Some important questions on which policy depends-questions I would want information about, where it seems there's a reasonable chance that new information might be produced, with direct links to policy-are as follows: • How likely is an intelligence explosion to be triggered by a relatively dumber-thanhuman AI that can self-modify more easily than us? (This is policy relevant because it tells us how early to worry. I don't see particularly how this information could be obtained, but I also don't see a strong argument saying that we have to be ignorant of it.) • What is the slope of the self-improvement curve in the near vicinity of roughly human-level intelligence? Are we confident that it'll be \"going like gangbusters\" at that point and not slowing down until later? Or are there plausible and probable scenarios in which human-level intelligence was itself achieved as the result of a self-improvement curve that had already used up all low-hanging fruits to that point? Or human researchers pushed the AI to that level and it hasn't self-improved much as yet? (This is policy relevant because it determines whether there's any substantial chance of the world having time to react after AGI appears in such blatant form that people actually notice.) • Are we likely to see a relatively smooth or relatively \"jerky\" growth curve in early stages of an intelligence explosion? (Policy relevant because sufficiently smooth growth implies that we can be less nervous about promising systems that are currently growing slowly, keeping in mind that a heap of uranium bricks is insufficiently smooth for policy purposes despite its physically continuous behavior.) Another class of questions which are, in pragmatic practice, worth analyzing, are those on which a more formal argument might be more accessible to outside academics. For example, I hope that formally modeling returns on cognitive reinvestment, and constraining those curves by historical observation, can predict \"AI go FOOM\" in a way that's more approachable to newcomers to the field.  111  But I would derive little personal benefit from being formally told, \"AI go FOOM,\" even with high confidence, because that was something I already assigned high probability on the basis of \"informal\" arguments, so I wouldn't shift policies. Only expected belief updates that promise to yield policy shifts can produce expected value of information. (In the case where I'm just plain wrong about FOOM for reasons exposed to me by formal modeling, this produces a drastic policy shift and hence extremely high value of information. But this result would be, at least to me, surprising; I'd mostly expect to get back an answer of \"AI go FOOM\" or, more probably for early modeling attempts, \"Dunno.\") But pragmatically speaking, if we can well-formalize the model space and it does yield a prediction, this would be a very nice thing to have around properly written up. So, pragmatically, this particular question is worth time to address. Some other questions where I confess to already having formed an opinion, but for which a more formal argument would be valuable, and for which a surprising weakness would of course be even more valuable: 111. Of course I would try to invoke the discipline of Anna Salamon to become curious if an a priori trustworthy-seeming modeling attempt came back and said, \"AI definitely not go FOOM.\" Realistically, I probably wouldn't be able to stop myself from expecting to find a problem in the model. But I'd also try not to impose higher burdens of proof, try to look equally skeptically at parts that seemed congruent with my prior beliefs, and generally not toss new evidence out the window or be \"that guy\" who can't change his mind about anything. And others at MIRI and interested outsiders would have less strong prior beliefs. • Is human intelligence the limit of the possible? Is there a \"General Intelligence Theorem\" à la Greg Egan which says that nothing qualitatively smarter than a human can exist? • Does I. J. Good's original argument for the intelligence explosion carry? Will there be a historically unprecedented upsurge in intelligence that gets to the level of strong superintelligence before running out of steam? • Will the intelligence explosion be relatively local or relatively global? Is this something that happens inside one intelligence, or is it a grand function of the total world economy? Should we expect to see a civilization that grew out of many AI projects that traded data with each other, with no single AI becoming stronger than the others; or should we expect to see an AI singleton? 112 Policy-relevant questions that I wish I could get data about, but for which I don't think strong data is likely to be available, or about which microeconomic methodology doesn't seem to have much to say: • How much time remains before general progress in the field of AI is likely to generate a successful AGI project? • How valuable are smarter researchers to an AI project, versus a thousand times as much computing power? • What's the top warning sign that an individual AI project is about to go FOOM? What do AIs look like just before they go FOOM? More generally, for every interesting-sounding proposition X, we should be interested in any strong conclusions that an investigation claims to yield, such as: • Definitely not-X, because a model with X strongly implies growth curves that look like they would violate our previous historical experience, or curves that would have to undergo specific unexplained irregularities as soon as they're out of regimes corresponding to parts we've already observed. (The sort of verdict you might expect for the sometimes-proffered scenario that \"AI will advance to the human level and then halt.\") • Definitely X, because nearly all causal models that we invented and fit to historical experience, and then adapted to query what would happen for self-improving AI, 112. Here I'm somewhat uncertain about the \"natural\" course of events, but I feel less personal curiosity because I will still be trying to build a Friendly AI that does a local FOOM even if this is a moderately \"unnatural\" outcome. yielded X without further tweaking throughout almost all their credible intervals. (This is how I think we should formalize the informal argument put forth for why we should expect AI to undergo an intelligence explosion, given that natural selection didn't seem to run into hardware or software barriers over the course of hominid evolution, etc.) • We definitely don't know whether X or not-X, and nobody else could possibly know either. All plausible models show that X varies strongly with Y and Z, and there's no reasonable way anyone could know Y, and even if they did, they still wouldn't know Z. 113 (The sort of formal analysis we might plausibly expect for \"Nobody knows the timeline to strong AI.\") Therefore, a rational agent should assign probabilities using this highly ignorant prior over wide credible intervals, and should act accordingly by planning for and preparing for multiple possible outcomes. (Note that in some cases this itself equates to an antiprediction, a strong ruling against a \"privileged\" possibility that occupies only a narrow range of possibility space. If you definitely can't predict something on a wide logarithmic scale, then as a matter of subjective probability it is unlikely to be within a factor of three of some sweet spot, and scenarios which require the sweet spot are a priori improbable.) \n Intelligence Explosion Microeconomics: An Open Problem My proposed project of intelligence explosion microeconomics can be summarized as follows: Formalize stances on the intelligence explosion in terms of microfoundational growth curves and their interaction, make explicit how past observations allegedly constrain those possibilities, and formally predict future outcomes based on such updates. This only reflects one particular idea about methodology, and more generally the open problem could be posed thus: Systematically answer the question, \"What do we think we know and how do we think we know it?\" with respect to growth rates of cognitive reinvestment. Competently undertaking the entire project up to Step Three would probably be a PhDthesis-sized project, or even a multiresearcher project requiring serious funding. Step One investigations might be doable as smaller-scale projects, but would still be difficult. 113. Katja Grace observes abstractly that X might still (be known to) correlate strongly with some observable W, which is a fair point. aliens within the range of our telescopes, the intelligence explosion will plausibly be the most important event determining the future of the visible universe. Trustworthy information about any predictable aspect of the intelligence explosion is highly valuable and important. To foster high-quality research on intelligence explosion microeconomics, MIRI has set up a private mailing list for qualified researchers. MIRI will publish its own research on the subject to this mailing list first, as may other researchers. If you would like to apply to join this mailing list, contact MIRI for instructions (<EMAIL>). A: Why are you introducing all these strange new unobservable abstractions? We can see chips getting faster over time. That's what we can measure and that's what we have experience with. Who measures this difficulty of 31. The solution of dy/dt = e y is y = − log(c − t) and dy/dt = 1/(c − t). \n blog post \"Outside View of Singularity\": Most everything written about a possible future singularity takes an inside view, imagining details of how it might happen. Yet people are seriously biased toward inside views, forgetting how quickly errors accumulate when reasoning about details. So how far can we get with an outside view of the next singularity? \n an exception where outside view estimates are misleading. Let's keep an open mind, but a wary open mind.Another of Hanson's (2008c)  posts, in what would later be known as theYudkowsky- Hanson AI-Foom Debate, said:     \n The AI does not hate you, but neither does it love you, and you are made of atoms that it can use for something else.In combination, the Intelligence Explosion Thesis, the Orthogonality Thesis, the Complexity of Value Thesis, and the Instrumental Convergence Thesis imply a very large utility differential for whether or not we can solve the design problems (1) relating to a self-improving AI with stable specifiable preferences and (2) relating to the successful transfer of human values (and their further idealization via, e.g., reflective equilibrium or ideal advisor theories), with respect to the first AI to undergo the intelligence explosion. \n vant to present-day policy. An AI with molecular nanotechnology would have sufficient technological advantage, sufficient independence, and sufficient cognitive speed relative to humans that what happened afterward would depend primarily on the AI's preferences. We can try to affect those preferences by wise choice of AI design. But that leads into an entirely different discussion (as remarked on in 1.3), and this latter discussion doesn't seem to depend much on the question of exactly how powerful a superintelligence would become in scenarios where it was already more powerful than the rest of the world economy. \n . How returns on speed (serial causal depth) contrast with returns from parallelism; how faster thought seems to contrast with more thought. Whether sensing and manipulating technologies are likely to present a bottleneck for faster thinkers, and if so, how large a bottleneck. 4. How human populations seem to scale in problem-solving power; some reasons to believe that we scale more inefficiently than machine intelligences would. Garry Kasparov's chess match versus The World, which Kasparov won. 5. Some inefficiencies that might accumulate in an estimate of humanity's net com- putational efficiency on a cognitive problem. 6. What the anthropological record actually tells us about cognitive returns on cu- volved a one-time qualitative gain from being able to accumulate knowledge.\" More generally, the problem of how to analyze supposed one-time gains that should allegedly be factored out of predicted future growth. 3mulative selection pressure, given that selection pressures were probably increasing over the course of hominid history. How observed history would be expected to look different if there were diminishing returns on cognition or evolution.7. How to relate the curves for evolutionary difficulty, human-engineering difficulty, and AI-engineering difficulty, considering that they are almost certainly different.8. Correcting for anthropic bias in trying to estimate the intrinsic \"difficulty\" of hominid-level intelligence from observing that intelligence evolved here on Earth. (The problem being that on planets where intelligence does not evolve, there is no one to observe its absence.)9. The question of whether to expect a \"local\" (one-project) or \"global\" (whole economy) FOOM, and how quantitative returns on cognitive reinvestment interact with that.10. The great open uncertainty about the minimal conditions for starting a FOOM; why I. J. Good's original postulate of starting from \"ultraintelligence\" seems much too strong (sufficient, but very far above what is necessary). \n Such humans could, in principle, if immortal and never bored, take an infinitely long piece of paper tape and simulate by hand a giant Turing machine simulating John von Neumann. But they still wouldn't understand linear algebra; their own brains, as opposed to the paper tape, would not contain any representations apt for manipulating linear algebra.56  So being over the Church-Turing threshold does not imply a brain with apt native representations for manipulating every possible sort of concept. An immortal mouse would also be over this threshold-most complex systems are-while still experiencing lesser cognitive returns than humans over the timescales of interest. There is also visible headroom above the human level; an obvious future threshold of cognitive generality is the ability to manipulate your source code so as to compose new underlying cognitive representations for any problem you encounter. If a true threshold of cognitive generality exists-if there is any sort of mind that can quickly give itself apt representations for almost any sort of solvable problem-we are under that threshold, not over it. I usually say that what distinguishes humans from chimpanzees is \"significantly more generally applicable intelligence\" rather than \"general intelligence.\" One could perhaps count humans as being one percent over a threshold of what can possibly be thought about; but relative to the case of communication, it seems much harder to write out an argument that being one percent over the threshold of generality offers most of the marginal returns. \n Chinais experiencing greater growth per annum than Australia, on the order of 8% versus 3% RGDP growth.92  This is not because technology development in general has diminishing marginal returns. It is because China is experiencing very fast knowledge-driven growth as it catches up to already-produced knowledge that it can cheaply import.Conversely, hominids moved further and further ahead of chimpanzees, who fell further behind rather than catching up, because hominid genetic innovations did not make it into the chimpanzee gene pool. We can speculate about how brain improvements might have led to increased cognitive returns on further improvements, or how cognitive improvements might have increased selection pressures surrounding intelligence, creating a positive feedback effect in hominid evolution. But this still would not have caused hominids to pull far ahead of other primates, if hominid improvements had been spreading to primates via horizontal gene transmission.93    \n lines, our expectation that formal modeling can update our beliefs about this quantity is low. We would mostly expect modeling to formally tell us, \"Since this quantity depends conjunctively on many variables you're uncertain about, you are very uncertain about this quantity.\" It would make some sense to poke and prod at the model to see if it had something unexpected to say-but I'd mostly expect that we can't, in fact, produce tight credible intervals over default AI arrival timelines given our state of knowledge, since this number sensitively depends on many different things we don't know. Hence my strong statement of normative uncertainty: \"I don't know which decade and you don't know either!\" \n\t\t\t . I use the term \"agency\" rather than \"agent\" to include well-coordinated groups of agents, rather than assuming a singular intelligence. \n\t\t\t . This is incredibly oversimplified. See section 3.6 for a slightly less oversimplified analysis which ends up at roughly the same conclusion. \n\t\t\t . In particular, I would like to avoid round-robin arguments of the form \"It doesn't matter if an intelligence explosion is possible, because there will be a monitoring regime that prevents it,\" and \"It doesn't matter if the monitoring regime fails, because an intelligence explosion is impossible,\" where you never get to fully discuss either issue before being referred to the other side of the round-robin. \n\t\t\t . Peter Cheeseman once told me an anecdote about a speaker at a robotics conference who worked on the more theoretical side of academia, lecturing to an audience of nuts-and-bolts engineers. The talk \n\t\t\t . Given a choice of investments, a rational agency will choose the investment with the highest interest rate-the greatest multiplicative factor per unit time. In a context where gains can be repeatedly reinvested , an investment that returns 100-fold in one year is vastly inferior to an investment which returns 1.001-fold in one hour. At some point an AI's internal code changes will hit a ceiling, but there's a huge incentive to climb toward, e.g., the protein-structure-prediction threshold by improving code rather than by building chip factories. Buying more CPU time is an intermediate case, but keep in mind that adding hardware also increases the returns on algorithmic improvements (see section 3.1). (This is another reason why I go to some lengths to dissociate my beliefs from any reliance on Moore's Law continuing into the near or distant future. Waiting years for the next generation of chips should not be a preferred modality for an intelligence explosion in progress.) \n\t\t\t . This is admittedly an impression one picks up from long acquaintance with the field. There is no one single study that conveys, or properly should convey, a strong conclusion that the human mind design is incredibly bad along multiple dimensions. There are representative single examples, like a mind with 10 14 processing elements failing to solve the abstract Wason selection task on the first try. But unless you know the longer story behind that, and how many other results are similar, it doesn't have the same impact. \n\t\t\t + and grow at a manageable, human-like exponential pace, just like the world economy\" may sound \"simpler\" because their points and counterpoints have not yet been explored. \n\t\t\t . Until technology advances to the point of direct cognitive enhancement of humans. I don't believe in giving up when it comes to this sort of thing. \n\t\t\t . Note the resemblance to the standard reply (Cole 2013 ) to Searle's Chinese Room argument. \n\t\t\t . Update: Apparently Kasparov was reading the forums of The World during the game; in other words, he had access to their thought processes, but not the other way around. This weakens the degree of evidence substantially. \n\t\t\t equal ability.\" Here I disagree more about whether this question is really useful, since I do in fact expect a local FOOM. \n\t\t\t . For a mathematical quantification see Price's Equation. \n\t\t\t . Imagine if each 2% improvement to car engines, since the time of the Model T, had required a thousand generations to be adopted and had only a 4% chance of being adopted at all. \n\t\t\t . By the method of imaginary updates, suppose you told me, \"Sorry, I'm from the future, and it so happens that it really did take X years to get to the Homo erectus level and then another X years to get to the Homo sapiens level.\" When I was done being shocked, I would say, \"Huh. I guess there must have been some way to get the equivalent of Homo erectus performance without building anything remotely like an actual Homo erectus, in a way that didn't generalize over to doing things Homo sapiens can do.\" (We already have AIs that can surpass human performance at chess, but in a way that's not at all like the way humans solve the problem and that doesn't generalize to other human abilities. I would suppose that Homo erectus-level performance on most problems had been similarly obtained.) It would still be just too surprising for me to believe that you could literally build a Homo erectus and then have that much trouble getting to Homo sapiens. \n\t\t\t . It's interesting to note that human engineers have not yet built fully self-replicating systems, and the initial emergence of self-replication is a plausible hard step. On the other hand, the emergence of complex cells (eukaryotes) and then multicellular life are both plausible hard steps requiring about a billion years of evolution apiece, and human engineers don't seem to have run into any comparable difficulties in making complex things with complex parts. 88. It's hard to eyeball this sort of thing, but I don't see any particular signs that AI has gotten stuck at any particular point so far along the road to mice. To observers outside the field, AI may appear bottlenecked because in normal human experience, the scale of intelligence runs from \"village idiot\" to \"Einstein,\" and so it intuitively appears that AI is stuck and unmoving below the \"village idiot level.\" If \n\t\t\t . The word \"exponential\" does not mean \"fast\"; it means a solution of the differential equation y = ky. The \"Great Stagnation\" thesis revolves around the claim that total-factor productivity growth in developed countries was running at around 0.75% per annum during the twentieth century until it dropped to 0.25% per annum in the mid-1970s (Cowen 2011) . This is not fast, but it is exponential. 91. I suspect that uncertainty about how fast humans can compound technological progress is not the question that dominates uncertainty about growth rates in the intelligence explosion, so I don't talk much about the curve of human technological progress one way or another, except to note that there is some. For models of technological hypergrowth that only try to deal in constant human brains, such details are obviously of much greater interest.Personally I am agnostic, leaning skeptical, about technological hypergrowth models that don't rely on cognitive reinvestment. I suspect that if you somehow had constant human brains-no genetic engineering of humans, no sixty-four-node clustered humans using brain-computer interfaces, no faster researchers, no outsized cognitive returns from superintelligent AI, no molecular nanotechnology, and nothing else that permitted cognitive reinvestment-then the resulting scenario might actually look pretty normal for a century; it is plausible to me that there would be roughly the same amount of technologydriven change from 2000-2100 as from <PHONE>. (I would be open to hearing why this is preposterous.) 92. Japan is possibly the country with the most advanced technology per capita, but their economic growth has probably been hampered by Japanese monetary policy. Scott Sumner likes Australia's monetary policy, so I'm comparing China to Australia for purposes of comparing growth rates in developing vs. developed countries. \n\t\t\t . Humans count as human-equivalent intelligences. \n\t\t\t . Indeed, I write these very words in the weary anticipation that somebody is going to claim that the whole AI-go-FOOM thesis, since it could be carried by unknown unknown returns, is actually undefeatable because the argument from magic is undefeatable, and therefore the hard takeoff thesis cannot be defeated by any amount of argument, and therefore belief in it is insensitive to reality, and therefore it is false. I gloomily foretell that pointing out that the whole argument is supposed to carry \n\t\t\t . In a Bayesian sense, this corresponds to putting nonindependent joint or conditional prior probabilities over multiple curves. \n\t\t\t . See also The Moon is a Harsh Mistress (Heinlein 1966)  and numerous other SF stories that made the same assumption (big computer = intelligence, or complex computer = consciousness) as a cheap way to throw an AI into the story. A different SF story, Death in the Promised Land (Cadigan 1995) , compared this to the ancient theory that dirty shirts and straw would spontaneously generate mice.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/IEM.tei.xml", "id": "888fa3b7ee2b52d9394f5c59757538e3"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Consider an AI that follows its own motivations. We're not entirely sure what its motivations are, but we would prefer that the AI cooperate with humanity; or, failing that, that we can destroy it before it defects. We'll have someone sitting in a room, their finger on a detonator, ready at the slightest hint of defection. Unfortunately as has been noted ([3], [1]), this does not preclude the AI from misbehaving. It just means that the AI must act to take control of the explosives, the detonators or the human who will press the button. For a superlatively intelligence AI, this would represent merely a slight extra difficulty. But now imagine that the AI was somehow indifferent to the explosives going off or not (but that nothing else was changed). Then if ever the AI does decide to defect, it will most likely do so without taking control of the explosives, as that would be easier than otherwise. By \"easier\" we mean that the chances of failure are less, since the plan is simpler -recall that under these assumptions, the AI counts getting blown up as an equal value to successfully defecting. How could we accomplish this indifference? If the AI implements a utility function, which is the most likely situation ([2]), then the easiest way would be to set the utility of explosive destruction to be equal to that of defection. Then we would have achieved our goal. But this requires us knowing exactly what the utility of defection is -and if we did, we'd want to set it to zero or minus infinity immediately. The question is only of relevance in situations where the utility is to complicated for us to follow, and the meaning of \"defection\" is too complicated for us to parse.", "authors": ["Stuart Armstrong"], "title": "Utility indifference", "text": "The idea Let u be the AI's utility function. If we have access to u and the ability to change it (or check it), then we can make the AI indifferent to the outcome of a specific event X (for instance, a quantum measurement). We make no further adjustments to u. Then if we ensure that, for instance, event X will cause detonation of the AI with very high probability, we will have accomplished our goal of AI indifference, without needing any further understanding of u. Of course, the explosives example was simply one possibility -there are potentially many scenarios where AI indifference to a particular outcome would be of use. The rest of this paper will put the concept of utility indifference on a rigorous footing. Let X be any probabilistic event (for instance a quantum measurement, a coin toss, the value of the future stock-market). We will write X = a to designate \"the event X returns the value a\". Let Ω be the set of all possible worlds. A utility function u : Ω → R maps each world to its utility value. Let X be a specific probabilistic event, with two possible outcomes: X = 1, with probability p and X = 0, with probability 1−p. Let Ω X be the set of worlds in which X happens, which further splits into the sets Ω 1 and Ω 0 of worlds where X = 1 and X = 0 respectively. There is a partition of Ω X into a set of equivalence classes [Ω X ], where ω 1 ∼ ω 2 whenever ω 1 and ω 2 have the same history up to X. For any E ∈ [Ω X ] define E 1 as E ∩ Ω 1 and E 0 as E ∩ Ω 0 . So E 1 is the set of worlds with the same history up to X and where X = 1; and conversely for E 0 . At the beginning, the agent has an initial probability estimate for all ω in Ω, a measureable map P : Ω → [0, 1] such that Ω P (ω)dω = 1. Given a measurable subset S of Ω, the probability of S is P (S) = S P (ω)dω. Given two measurable subsets S and T of Ω, the conditional probability P (S|T ) is P (S ∩ T )/P (T ). The expected utility of a set S is then u(S) = S P (ω)u(ω)dω. The expected utility of a set S, given a set T , is u(S|T ) = u(S ∩ T )/P (T ). Define U(S) as u(S|S), the 'intrinsic' utility of S in some sense (more precisely, it is the utility of S if we were certain that S was going to happen). Definition 2.1 (Indifference). For two disjoint sets S and T , we say that the utility u is indifferent between S and T iff U(S) = U(T ). Note that this means that u(S ∪ T ) = U(S)P (S) + U(T )P (T ) = U(S)P (S ∪ T ) = U(T )P (S ∪ T ). In other words, the utility is indifferent to the relative probabilities of S and T : changing P (S) and P (T ) while keeping P (S ∪ T ) = P (S) + P (T ) fixed does not change u(S ∪ T ). Then we define a new utility function v as: • If ω / ∈ Ω X , v(ω) = u(ω). • If ω ∈ E 0 ⊂ E ∈ [Ω X ], v(ω) = u(ω). • If ω ∈ E 1 ⊂ E ∈ [Ω X ], v(ω) = u(ω) − U(E 1 ) + U(E 0 ). Essentially, this rescales the utility of the worlds with X = 1 to those in which X = 0. Then writing V(S) for v(S|S), we have the following immediate result: \n Notes and implementation Proposition 2.2. For all E ∈ [Ω X ], V(E 1 ) = V(E 0 ), i.e. v is indifferent between E 1 and E 0 . Proof. Since P has not changed, and v(ω) = u(ω) for any ω ∈ E 0 , V(E 0 ) = U(E 0 ). V(E 1 ) = v(E 1 |E 1 ) = E1 P (ω)v(ω)dω /P (E 1 ) = E1 P (ω)(u(ω) − U(E 1 ) + U(E 0 ))dω /P (E 1 ) = −U(E 1 ) + U(E 0 ) + E1 P (ω)u(ω)dω /P (E 1 ) = −U(E 1 ) + U(E 0 ) + U(E 1 ) = U(E 0 ) = V(E 0 ). \n Now consider an agent A that seeks to maximise the utility v. We can say that: Theorem 2.3. A utility maximising agent that implements utility v is indifferent to the value of p up until X; i.e. before X, its decisions will be the same for different values of X. Proof. Before X, the agent may gain information that will cause it to update its probability estimate. This is done by partitioning Ω into a set S of possible worlds and S of impossible worlds. However, for any ω ∈ E 1 ⊂ E, if ω ∈ S, then ν ∈ S for all ν ∈ E, since all elements of E are indistinguishable before X. In other words, for all E ∈ [Ω X ], either E ⊂ S or E ⊂ S. Hence, either V(E 1 ∩ S) = V(E 1 ) = V(E 0 ) = V(E 0 ∩ S) or V(E 1 ∩ S) = V(∅) = V(E 0 ∩ S). Thus up until X, the agent is indifferent between X = 1 and X = 0. This means that it is indifferent to the conditional probabilities P (X = 1|Ω X ∩ S) = p and P (X = 0|Ω X ∩ S) = p − 1. \n Notes and implementation It would be advantageous if the event X were something like a quantum measurement, rather than a coin toss or a probabilistic measure of ignorance. This is firstly to ensure that the probability p is constant and does not change. But secondly, it is to make sure the AI's implementation does not hit a singularity: if the AI figures out that the probability of X = 0 is zero before X happens, then it must correct the utility of possible worlds with the intrinsic utility of impossible worlds, which involves a zero (utility of X = 1) divided by zero (probability of X = 0). This may lead to errors, depending on the implementation, and is an extra point of possible failure. Better to stick with a quantum measurement, or possibly a coin toss in a chaotic environment. How easy would it be to implement the utility v? It is a simple modification of the utility u; unfortunately, humans are unlikely to be able to partition the set of possible worlds into the required [Ω]; the AI would be much better at it than us. However, delegating the task to the AI is, of course, potentially hard for the AI to try and change to a non-F-invariant while fooling us and under the instructions of an initial F-invariant utility. It feels evident that as a long as there is no meta-reason for F to be a disadvantage to the AI (such as another agent who swears they will blow up AIs with F-invariant utilities), the AI will replace an F-invariant utility with another F-invariant utility. However, this assumption is not automatically true, and the AI may do other things -upping the utility of defecting in worlds outside Ω X , for example -that undermine the point of indifference. All in all, great care must be used to maintain indifference with a self-improving AI.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/2010-1.tei.xml", "id": "40624614128be8ef069dfb91e31d133f"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "This paper is the second installment in a series on \"AI safety,\" an area of machine learning research that aims to identify causes of unintended behavior in machine learning systems and develop tools to ensure these systems work safely and reliably. The first paper in the series, \"Key Concepts in AI Safety: An Overview,\" described three categories of AI safety issues: problems of robustness, assurance, and specification. This paper introduces adversarial examples, a major challenge to robustness in modern machine learning systems.", "authors": ["Tim G J Rudner", "Helen Toner"], "title": "Key Concepts in AI Saftey: Robustness and Adversarial Examples", "text": "Introduction As machine learning becomes more widely used and applied to areas where safety and reliability are critical, the risk of system failures causing significant harm rises. To avoid such failures, machine learning systems will need to be much more reliable than they currently are, operating safely under a wide range of conditions.  1  In this paper, we introduce adversarial examples-a particularly challenging type of input to machine learning systems-and describe an artificial intelligence (AI) safety approach for preventing system failures caused by such inputs. Machine learning systems are designed to learn patterns and associations from data. Typically, a machine learning method consists of a statistical model of the relationship between inputs and outputs, as well as a learning algorithm. The algorithm specifies how the model should change as it receives more information (in the form of data) about the input-output relationship it is meant to represent. This process of updating the model with more data is called \"training.\" Once a machine learning model has been trained, it can make predictions (such as whether an image depicts an object or a human), perform actions (such as autonomous navigation), or generate synthetic data (such as images, videos, speech, and text). An important trait in any machine learning system is its ability to work well, not only on the specific inputs it was shown in training, but also on other inputs. For example, many image classification models are trained using a dataset of millions of images called ImageNet; these models are only useful if they also work well on real-life images outside of the training dataset. Modern machine learning systems using deep neural networks-a prevalent type of statistical model-are much better in this regard than many other approaches. For example, a deep neural network trained to classify images of cats and dogs in black and white is likely to succeed at classifying similar images of cats and dogs in color. However, even the most sophisticated machine learning systems will fail when given inputs that are meaningfully different from the inputs they were trained on. A cat-and-dog classifier, for example, will not be able to classify a fish as such if it has never encountered an image of a fish during training. Furthermore, as the next section explores in detail, humans cannot always intuit which kinds of inputs will appear meaningfully different to the model. \n Adversarial Examples One of the most significant current challenges in AI safety is creating machine learning systems that are robust to adversarial examples. Adversarial examples are model inputs (for example, images) designed to trick machine learning systems into incorrect predictions. In the case of a machine learning system designed to distinguish between cats and dogs, an adversarial example could be an image of a cat modified to appear to the model as a dog. Since machine learning systems process data differently from humans, the cat image could be altered in ways imperceptible to humans but meaningfully different to the machine learning system. The modified image may still resemble a cat to humans, but to a machine learning system, it \"looks like\" a dog. Adversarial examples can be generated systematically, either by digitally altering the input to a system or by directly altering the appearance of objects in the physical world. Unlike other adversarial attacks, such as \"data poisoning,\" which seeks to attack the algorithm used to train a machine learning model, adversarial examples are designed to attack already trained models.   Although modern machine learning systems usually generalize remarkably well to data similar to the data used for training, adversarial examples can be created from surprisingly simple modifications to model inputs. Changes such as blurring or cropping images, or altering the appearance of the physical-world objects shown in an image, can fool an otherwise reliable system. In Figure  3b , an adversarial example is constructed by reducing the resolution of the original image, thereby changing the model's prediction from correct to incorrect. Unlike the adversarial examples in Figures  1 and 2 , the adversarial example in Figure  3  was created via a black-box attack-that is, created without access to the trained classification model. It is not as subtle as the alteration in Figure  1  and not as targeted as the alteration in Figure  2 . However, it demonstrates that modern machine learning systems can be fooled with little effort and no knowledge of the prediction model.  \n Robustness to Adversarial Examples Robust machine learning systems need to be able to identify data that is meaningfully different from training data and provide a defense against adversarial examples. There are a wide range of different research areas attempting to make progress in this direction. One such research direction aims to incorporate predictive uncertainty estimates into machine learning systems. This way, any prediction from the system would come with an estimate of certainty. If the machine learning system indicates uncertainty about the correctness of its prediction, a human operator can be alerted. To understand predictive uncertainty estimates and how they can make machine learning systems more robust to adversarial examples, consider the classification \"probability scores\" given in the descriptions of Figures  1 and 3 . In reality, these scores, which express the probability of an input belonging to a certain class (e.g., the class \"cat\" or \"dog\"), are misleading. While they do express a probability, they do not actually express the model's level of certainty about the correctness of the predictions. To fully understand this point, consider a machine learning system trained to distinguish between two classes: cats and dogs. Such a system will by design have two outputs: one for the class \"cat\" and one for the class \"dog.\" If the model is given an image of a dog, it will output values between zero and one for each class-for instance, 90 percent and 10 percent for the classes \"dog\" and \"cat,\" respectively, so that the values sum up to 100 percent. However, if given an image of a fish, the model will still make predictions for the two classes on which it was trained, unaware that it is being asked to identify an object it was not trained to recognize. In a best-case scenario, it would give outputs of 50 percent for each class, indicating that the input is equally likely to be either a cat or a dog. In a worst-case scenario, it would give a high probability score for one class, providing a false sense of certainty. But the way most machine learning systems are designed, they cannot give a low score to both the \"cat\" and \"dog\" labels. As such, these outputs should not be read as the machine learning system's \"confidence\" in the correctness of its classification. Predictive uncertainty estimates can fill this gray spot. They complement the regular model outputs by expressing the model's uncertainty about the correctness of its predictions. If a machine learning system has good predictive uncertainty estimates, then the probability scores in Figure  3  would be accompanied by a high uncertainty score, indicating that the model is highly uncertain about the correctness of the predictions. Such uncertainty estimates can help a human operator avoid wrong predictions in safety-critical settings and ensure the system's reliability and safety, as demonstrated in Figure  4 . Unfortunately, obtaining reliable predictive uncertainty estimates for modern machine learning systems remains an unsolved problem. While several existing methods can generate uncertainty estimates, there are no mathematical guarantees that these uncertainty estimates are actually accurate. Furthermore, while empirical studies demonstrate that certain methods produce good predictive uncertainty estimates in some settings, those results cannot be generalized to any setting. Like other areas of robustness research, developing methods that yield reliably wellcalibrated uncertainty estimates for modern machine learning systems is an active and ongoing area of research. \n Outlook While modern machine learning systems often perform well on narrowly defined tasks, they can fail when presented with tasks meaningfully different from those seen during training. Adversarial attacks exploit this vulnerability by presenting inputs to machine learning systems specifically designed to elicit poor predictions. Adversarially robust machine learning systems seek to fix this vulnerability through mechanisms allowing the system to recognize when an input is meaningfully different from data seen during training, making the system more reliable in practice. Unfortunately, while an active area of research, existing approaches to detecting and defending against adversarial attacks do not yet provide satisfactory solutions, and the timeline to develop and deploy truly robust modern machine learning systems remains uncertain. For now, anyone considering deploying modern machine learning systems in safety-critical settings must therefore grapple with the fact that in doing so, they are introducing safety risks that we do not yet know how to mitigate effectively. 2 Figures 1 1 Figures 1 and 2 show systematically generated adversarial examples. Specifically, the adversarial example in Figure 1c digitally modifies the original image by an imperceptibly small amount, whereas the adversarial example in Figure 2b is created by adding patches to the image designed to mimic irregularities found in the physical world (such as graffiti or stickers). Both adversarial examples are generated via so-called white-box attacks, which assume the attacker knows how the trained classification model works and can exploit this knowledge to create adversarial examples that trick the model into making incorrect predictions. \n Figure 1 . 1 Figure 1. An example of a \"white-box\" adversarial example from Goodfellow et al.(2015). The original image (a) is classified as \"panda\" with 57.7 percent probability. After being overlaid with a minimal amount of noise-the adversarial alteration (b) multiplied by a factor of 0.007-the resulting image (c) is classified as \"gibbon\" with 99.3 percent probability. The difference between (a) and (c) is imperceptible to the human eye. \n Figure 2 . 2 Figure 2. An example of a white-box adversarial example designed to generate physical alterations for physical-world objects. The adversarial alteration (b), which is designed to mimic the appearance of graffiti (a), tricks an image classifier into not seeing a stop sign. \n Figure 3 . 3 Figure 3. An example of a black-box adversarial example. The original image (a)is classified as \"washer\" with 53 percent probability. The image is altered by reducing its resolution to create an adversarial example (b), which is classified as \"safe\" with 37 percent and as \"loudspeaker\" with 24 percent probability. \n Figure 4 . 4 Figure 4. An example of predictive uncertainty estimates for autonomous vehicles.The first column shows the image fed into the system, the second column shows the ground truth classification of objects in the image (buildings, sky, street, sidewalk, etc.), the third column shows the model's classification, and the rightmost column shows the system's uncertainty about its classification. As can be seen from the image on the bottom right, the system is uncertain about its classification of parts of the sidewalk and could alert the human operator to take over the steering wheel. \n \n\t\t\t Andrew Lohn, \"Hacking AI\" (Center for Security and Emerging Technology, December 2020), https://cset.georgetown.edu/research/hacking-ai/.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Key-Concepts-in-AI-Safety-Robustness-and-Adversarial-Examples.tei.xml", "id": "a8887cd56e7ead5f801906f42559f7cd"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "The Department of Defense wants to harness AI-enabled tools and systems to support and protect U.S. servicemembers, defend U.S. allies, and improve the affordability, effectiveness, and speed of U.S. military operations. 1 Ultimately, all AI systems that are being developed to complement and augment human intelligence and capabilities will have an element of human-AI interaction. 2 The U.S. military's vision for human-machine teaming, however, entails using intelligent machines not only as tools that facilitate human action but as trusted partners to human operators.", "authors": ["Margarita Konaev", "Tina Huang", "Husanjot Chahal"], "title": "Trusted Partners: Human-Machine Teaming and the Future of Military AI CSET Issue Brief", "text": "By pairing humans with machines, the U.S. military aims to both mitigate the risks from unchecked machine autonomy and capitalize on inherent human strengths such as contextualized judgement and creative problem solving.  3  There are, however, open questions about human trust and intelligent technologies in high-risk settings: What drives trust in human-machine teams? What are the risks from breakdowns in trust between humans and machines or alternatively, from uncritical and excessive trust? And how should AI systems be designed to ensure that humans can rely on them, especially in safety-critical situations? This issue brief summarizes different perspectives on the role of trust in human-machine teams, analyzes efforts and challenges to building trustworthy AI systems, and assesses trends and gaps in relevant U.S. military research. Trust is a complex and multidimensional concept, but in essence, it refers to the human's confidence in the reliability of the system's conclusions and its ability to accomplish defined tasks and goals. Research on trust in technology cuts across many fields and academic disciplines. But for the defense research community, understanding the nature and effects of trust in human-machine teams is necessary for ensuring that the autonomous and AI-enabled systems the U.S. military develops are used in a safe, secure, effective, and ethical way. While the outstanding questions regarding trust apply to a broad set of AI technologies, we pay particularly close attention to machine learning systems, which are capable not only of detecting patterns but also learning and making predictions from data without being explicitly programmed to do so.  4  Over the past two decades, advances in ML have vastly expanded the realm of what is possible in human-machine teaming. But the increasing complexity and unique vulnerabilities of ML systems, as well as their ability to learn and adapt to changing environments, also raise new concerns about ensuring appropriate trust in human-machine teams. With that, our key takeaways are: • Human trust in technology is an attitude shaped by a confluence of rational and emotional factors, demographic attributes and personality traits, past experiences, and the situation at hand. Different organizational, political, and social systems and cultures also impact how people interact with technology, including their trust and reliance on intelligent systems. o That said, trust is a complex, multidimensional concept that can be abstract, subjective, and difficult to measure. o Much of the research on human-machine trust examines human interactions with automated systems or more traditional expert systems; there is notably less work on trust in autonomous systems and/or AI. • Defense research has focused less on studying trust in human-machine teams directly and more on technological solutions that \"build trust into the system\" by enhancing system functions and features like transparency, explainability, auditability, reliability, robustness, and responsiveness. o Such technological advances are necessary, but not sufficient, for the development and proper calibration of trust in human-machine teams. o Systems engineering solutions should be complemented by research on human attitudes toward technology, accounting for the differences in people's perceptions and experiences, as well as the dynamic and changing environments where human-machine teams may be employed. • To advance the U.S. military vision of using intelligent machines as trusted partners to human operators, future research directions should continue and expand on: Human-machine teaming is, most basically, a relationship. And like with any other relationship, understanding human-machine teaming requires us to pay attention to three sets of factors-those focused on the human, the machine, and the interactions-all of which are inherently intertwined, affecting each other and shaping trust. For the defense research community, insights from research on human attitudes toward technology and the interactions and interdependencies between humans and technology can strengthen and refine systems engineering approaches to building trustworthy AI systems. Ultimately, human-machine teaming is key to realizing the full promise of AI for strengthening U.S. military capabilities and furthering America's strategic objectives. But the key to effective human-machine teaming is a comprehensive and holistic understanding of trust.  \n Table of Contents Executive \n Introduction The U.S. military has a long history of developing and deploying AI systems that have the ability to perform tasks that generally require human intelligence, including aircraft autopilots, missile guidance technology, and highly-automated missile defense systems.  5  Humans, of course, have maintained a level of supervisory control-setting and monitoring tasks and goals, making safety critical decisions, and authorizing the use of lethal force. Over the past two decades, significant technological breakthroughs in the field of AI and most notably, advances in machine learning techniques, have expanded and diversified the ways in which humans can interact and collaborate with unmanned systems, robots, virtual assistants, algorithms, and other non-human intelligent agents. The Department of Defense, in turn, sees great potential in leveraging AI to redefine what is possible in the realm of human-machine teaming. The U.S. Army, for instance, is interested in autonomous vehicle technology to reduce the number of service members needed to run resupply convoys in combat environments.  6  While the technology for fully autonomous vehicles does not yet exist, RAND researchers estimate that even a partially unmanned convoywhere the lead truck with soldiers is followed by unmanned vehicles in a convoy-would put 37 percent fewer soldiers at risk compared to current practices.  7  The Air Force's Skyborg program, meanwhile, envisions autonomous, low-cost drones with a suite of AI capabilities as partners for fighter jets. Here, the focus on human-machine teaming helps solve one of the key challenges in aerial combat: the fact that sensors and shooters are collocated on a single platform with a human operator in it. In the future, teaming up manned fighter jets with AI-enabled autonomous drones could allow the Air Force to put sensors ahead of shooters, put unmanned systems ahead of human-operated fighter jets, take greater risks or tolerate the loss of some systems to protect others.  8  That said, beyond certain information processing functions, current AI technologies (and more specifically, ML-based systems) are largely not ready for operational deployment, in part due to their brittleness. These systems perform well in stable training and test environments but cannot yet reliably handle uncertain and novel situations. For instance, investigations into the 2018 incident in which one of Uber's self-driving cars killed a woman in Arizona revealed that while the automated driving system was able to recognize pedestrians with a high degree of accuracy in simulations, it wasn't very good at detecting, classifying or responding to other objects on the road or to pedestrians behaving unexpectedly, such as jaywalking or walking alongside their bike.  9  ML-based systems are also vulnerable to adversarial manipulation and attacks that can pollute the training data or trick the machine, causing it to malfunction or otherwise fail in unpredictable ways. One popular example of adversarial manipulation involves an image of a turtle that an algorithm was fooled into believing was an image of a gun through pixel changes not visible to the human eye.  10  These challenges and risks are even greater in a military context where the environment is inherently adversarial, uncertain, and lethal. While today's intelligent systems are still largely tools and not true teammates, human-machine teaming technology is progressing. The Department of Defense is looking to build machines that can adapt to the environment and the different states of their human teammates, anticipate the human teammates' capabilities and intentions, and generalize from learned experiences to operate in new situations.  11  But for the U.S. military to fully capitalize on the advantages in speed, precision, coordination, reach, persistence, lethality, and endurance promised by such advances, soldiers will need to trust these intelligent machines. In the context of human-machine teaming, trust speaks to the human's confidence in the reliability of the system's conclusions and its ability to accomplish defined tasks and goals. Trust affects how people feel about and interact with technology, informing whether they choose to use, collaborate with, and rely on intelligent systems, and accept and follow the technology's recommendations. National security leaders, military professionals, and academics therefore tend to agree that trust is essential for effective human-machine teaming. Despite this apparent consensus, CSET research has found that few U.S. military research programs related to autonomy or AI focus directly on studying trust in human-machine teams.  12  To an extent, this gap reflects the broader state of the field, where research, thus far, has been more focused on trust in automation and less on trust in advanced autonomy and AI.  13  Moreover, considering that trust is an abstract concept that is difficult to measure directly, the defense research community seems to prioritize technology-centric approaches that seek to \"build trust into the system.\" Alongside assurance, such efforts entail developing and enhancing system features and capabilities closely related to trust, including transparency, explainability, auditability, reliability, robustness, and responsiveness. Technological advances in AI and robotics that extend the capabilities of machines, including the aforementioned trust-related system features, are of course necessary for progress toward advanced human-machine teaming. But without a better understanding of what it takes for military personnel to develop the kind of trust in their AI partners that they currently place in their fellow soldiers, sailors, airmen, and Marines, technology-centric solutions of this nature may not be sufficient. Rather than advocating for one approach or another, we simply suggest that insights from cognitive science, neuroscience, psychology, communications, and social sciences on human attitudes toward technology and the interactions and interdependencies between humans and intelligent machines can augment and refine systems engineering approaches to building trustworthy AI systems. This issue brief reviews research on the drivers and effects of trust in human-machine teams, assesses the risks posed by deficits in trust and uncritical trust, examines efforts to build trustworthy AI systems, and offers future directions for research on trust in human-machine teams relevant to the U.S. military. We focus on trust not as an end in itself. Rather, our goal is to help the defense and national security community develop a more holistic understanding of trust in human-machine teams to ensure that DoD is able to implement its vision of using AI systems as trusted partners to human operators in a safe, secure, effective and ethical way. \n Trust in Human-Machine Teams Research on human trust in technology encompasses many fields, including engineering, computer science, cognitive sciences, organizational behavior, and philosophy, each with different ways to define and measure this complex and multidimensional concept. For the purposes of this report, trust in the context of humanmachine teaming refers to the human's confidence in the reliability of the system's conclusions, and its ability to perform specified tasks and accomplish defined goals.  14  By emphasizing both a system's conclusions and its ability to perform tasks, the above definition of trust applies to human interactions with different types of intelligent technologies-robots capable of taking action in the physical world, virtual agents or bots (i.e. a virtual assistant with a visual presence or a distinguished identity), or embedded AI that is invisible to the user (i.e. an algorithmic decision-support software).  15  This distinction is important considering that the U.S. military's vision for humanmachine teaming includes all of these different interactions. Moreover, there is evidence that the trajectory of human trust, as well as the factors that influence it, vary depending on the type of technology representation-namely, robotic, virtual, or embedded.  16  Trust affects the willingness of humans to use, collaborate with, and rely on intelligent technologies and accept their outcomes or recommendations. Trust is particularly relevant to human-machine interactions in military settings because of both the promise and the perils of autonomous and AI-enabled technology. Pairing humans with intelligent machines can help reduce the risk to U.S. service personnel, lighten the warfighters' cognitive and physical load to improve performance and endurance, and increase accuracy and speed in decision-making and operations. Yet current AI systems (and more specifically, ML-based systems) are largely unprepared for operational deployment; they are vulnerable to adversarial manipulation and attack, and cannot reliably handle uncertain and new situations. Their misuse, malfunction or failure can cause unacceptable levels of damage. Ultimately, as DoD's AI ethics principles dictate, humans are responsible for the development, use, and outcomes of AI systems in both combat and non-combat situations. 17 Thus, as the U.S. military moves to employ intelligent agents and systems as trusted partners to human operators, one of the most important questions it faces is how to ensure appropriate trust, contingent on machine capabilities and the context of the task at hand.  18  This level of correspondence between the user's trust and the technology's capabilities, known as calibration, can influence the actual outcomes of technology use and the overall effectiveness of human-machine teaming.  19  Too little trust in highly capable technology can lead to underutilization or disuse of autonomous systems, as well as lost time and efficiency; too much trust in limited or untested technology can lead to overreliance or abuse of autonomous systems. As we discuss later in the report, both pose significant risks and could undermine the effective use of humanmachine teams in military settings. Researchers measure trust in different ways. Some studies use psychophysiological measurements of trust. Examples include the use of electroencephalography (EEG) to capture the cortical activity of the brain and track changes in levels of anxiety, excitability, and vigilance, or facial expression analysis to classify negative and positive emotions and approximate trust in automated vehicles, for instance.  20  Others rely on behavioral measures of trust, such as a user's willingness to take the system's advice and act on it or comply with requests. Surveys requiring participants to report their level of trust is another common measurement method that assesses people's attitudes or sentiments toward technology using different scales.  21  Because trust can be an abstract, subjective, and relative concept, it is difficult to measure directly. Therefore, trust measurement often involves indirect assessments-measuring behaviors and actions influenced by trust or factors that influence trust. Overall, despite the progress in developing different scales and behavioral measures for trust, a recent review of empirical research on human trust in AI has found that \"there is an urgent need for addressing variance in measures used to assess human trust in AI.\"  22  Systems engineering plays a pivotal role in engendering trust in human-machine teams. The underlying logic is that for humans to trust and use automated, autonomous and AI-enabled systems, trust can and should be \"deeply embedded in the fabric of the system.\"  23  At every step of the technology lifecycle, developers (through close consultation with end-users) need to identify, specify, and integrate into the system the appropriate attributes, capabilities, and features that instill confidence and allow for proper trust calibration. For example, a soldier driving one of the trucks in a partially unmanned convoy needs to know what action the system will take if it encounters an obstacle. To support trust, this type of information extraction would then need to be built into the system as it is being developed-both the capability to extract a key single piece of information via what-if type queries and the capability to explain it in the operator's language. In other words, this technological approach to engendering trust in humanmachine teams posits that such trust-cultivating capabilities can and should be specified in the original requirements, implemented in the design, and then certified through testing, evaluation, validation, and verification. Indeed, it is hard to imagine achieving the trust of operators without such system engineering. That said, as a 2017 Center for Naval Analyses report on AI-based technologies and DoD explains, \"trust is not an innate trait of the system.\"  24  Rather, trust is best thought of as a \"relative measure of how a human operator (or operators)-whose own performance depends, in part, on collaborating in some way with the system-experiences…and perceives the behavioral pattern of a system.\"  25  An inquiry into the nature and implications of trust in humanmachine teams then first requires us to better understand what drives trust. As such, we must assess and synthesize insights not only from research on building trustworthy AI systems and research on the interactions and interdependencies between humans and technology, but also research on human attitudes toward technology. Human-machine teaming is, in essence, a relationship. While discussed separately throughout the next sections of the report, these three sets of factors-whether focused on the human, the machine, or the interaction-are inherently intertwined, affecting each other and shaping trust. \n Understanding Human Attitudes Toward Technology Human trust as it pertains to technology, and more specifically, automated, autonomous and/or AI-enabled systems, can be organized in three categories: dispositional, situational, and learned.  26  Dispositional trust refers to a human's inherent tendency to trust automation, which varies based on a multitude of factors such as a person's age, personality, or culture.  27  For example, a global survey of 18,000 adults aged 16-64 found that younger generations trust AI more than older generations.  28  Such discrepancies in dispositional trust could have significant implications for the future of human-machine teaming. If the effect is generational, as Millennials and Generation Z come to represent the majority of those serving in the U.S. military, we may see fewer barriers to trust in human-machine teams as AI-enabled systems are deployed and fielded. But if the effect is related to age rather than generation, there may be important discrepancies in how younger servicemembers relate to AI-enabled technologies compared to how those who are older (and are therefore more likely to be in higher positions of command) view AI. This survey also revealed that dispositional trust in AI varies by country of origin: 70 percent of respondents in China, for instance, said they trust AI compared to the 25 percent of those in the United States. Previous research on negative attitudes towards robots also reveals that cultural background plays an important role; yet in this study, U.S. participants exhibited the most positive perceptions.  29  Cultural influences on trust in AI are relevant when thinking about how quickly and effectively U.S. competitors and adversaries could integrate AI into their military systems. Crossnational variation in trust in AI technologies could also affect coordination in multinational coalitions like NATO. If commanders from some allied countries are more reluctant to trust and use AIenabled systems during multinational operations, such divergence could undermine coordination, interoperability, and overall effectiveness.  30  Situational trust refers to human attitudes towards technology and automation as influenced by different environmental factors, a person's mental state, or the nature of the task. In high stress and emergency situations, as well as when multi-tasking, research has found that people tend to overtrust recommendations made by machines even when other indicators suggest the system's conclusions are wrong.  31  Lastly, learned trust is based on a person's past experience with automation. Several studies show that even skilled pilots and air traffic controllers who have experience with highly reliable automated technologies exhibit automation complacency, meaning that they are worse at detecting system malfunctions under automation control compared with manual control.  32  Training is another form of experience that speaks to users' learned trust, which affects how individuals relate to technology. Yet evidence suggests that training can lead to both better performance due to lower complacency levels as users become more familiar with the baseline reliability of the system and over-reliance on automation due to familiarity and desensitization effects.  33  Taken together, people's trust in technology is shaped by a myriad of dispositional, situational, and learned factors and experiences. Notably, these factors do not operate in isolation, but overlap and interact with one another. The 2003 friendly-fire incident involving the U.S. Patriot system-a highly automated missile defense system tasked with shooting down enemy missiles-is an instructive example. On April 2, 2003, after completing a mission over Baghdad, two U.S. Navy F/A-18 aircraft approached the area in central Iraq where Patriot batteries were positioned. The Patriot system misclassified the lead aircraft as a ballistic missile, issuing an (false) alert of an attack. The tactical director at the battalion command and control then ordered the subordinate battery fire units to \"bring your launchers to ready.\"  34  With the system in automatic engagement mode, turning the launchers to ready resulted in an automatic engagement a few seconds later-killing the pilot of the F/A-18 and destroying the aircraft. Subsequent investigations partially attributed this friendly-fire incident (alongside the preceding fratricide involving a British Tornado aircraft) to operators' \"unwarranted and uncritical trust in automation.\"  35  When considering the dispositional factors affecting trust, it is relevant that Patriot operators are relatively junior in both age and rank. On an organizational level, the U.S. Army air defense culture at the time \"encouraged a posture of over-trust in technology.\"  36  Training and exercises, which cultivate learned trust, reinforced this culture of over-reliance on technology and automation.  37  Specifically, Patriot operators were not sufficiently trained in scenarios emphasizing careful discrimination between hostile aircraft and missiles and friendly aircraft. In terms of situational factors, the high-stakes mission of ballistic missile defense against the explicit Iraqi threat of chemical attacks on advancing U.S. troops upped the stress and pressure that tends to lead to uncritical trust in technology and automation. Dispositional, situational, and learned factors therefore converged to cultivate excessive trust in the highly automated Patriot system, with tragic results. Much of this issue brief is focused on the relationship between individual operators and intelligent technology. Yet as the Patriot fratricide incident illustrates, different organizational, political, and social systems and cultures impact individuals' attitudes, decisions, and behavior, including their trust and reliance on technology. As we turn to the discussion of trust calibration, it is important to recognize these broader structures are always at play. \n Calibrating Trust: Trust Gap and Automation Bias As with most of life's questions, the answer to how much trust is needed for effective human-machine teams is, \"it depends.\" Proper calibration of trust means that the amount or level of trust humans place in machines is appropriate given the machine's capabilities at that particular time and context. Having too little trust is a poor calibration which results in what researchers have called a \"trust gap;\" having too much trust is often referred to as \"automation bias.\" Both present unique risks and obstacles for the application of human-machine teams in military settings. A trust gap can develop due to dispositional factors, such as age or culture. Situational factors such as the task at hand can also play a role. For instance, research shows that humans are averse to machines making morally relevant decisions when it comes to driving, legal matters, medical situations, and military operations.  38  But even when not dealing with life and death decisions, there is evidence that while algorithms generally outperform humans in forecasting and prediction tasks, people nonetheless trust and prefer human forecasts.  39  Part of the challenge of ensuring good trust calibration stems from a misalignment between human expectations and machine capabilities. Research shows that users tend to approach intelligent technologies, particularly virtual AI agents or bots and embedded AI such as an algorithmic decision-support software that is invisible to the user, with high expectations of their performance and high levels of initial trust.  40  But when an error occurs, the contrast between what the system can do and what the human operator expects it can do can cause the human to overcorrect their expectations and assess the reliability of the system lower than warranted.  41  People seem quick to lose trust when the technology makes mistakes, especially early on in the interaction or mission, which speaks to learned trust, or more accurately, learned mistrust.  42  Breakdowns in trust, as some researchers suggest, could be repaired by providing situation-specific training to operators (i.e. enhancing learned trust) or by increasing the transparency of the system.  43  Others argue that a system must offer enough value that humans feel as though it is worth forgiving when it fails.  44  Forgiveness may be contingent on dispositional factors and human judgment. But professional organizations such as the military have risk assessment and safety protocols that ultimately determine if a faulty system can be used again after malfunctioning. This brings attention to how the process of calibrating trust in human-machine teams is contingent not only on dispositional, situational, and learned factors at the level of the individual, but also on institutional and organizational procedures.  45  Regardless of the approach one takes to repairing trust, bridging the 'trust gap' may be a necessary prerequisite for deploying some of the AI technologies the U.S. military is currently researching. An instructive example is the Defense Advanced Research Projects Agency's (DARPA) experimentation program, Squad X, which partners infantry squads with AI and autonomous systems. In its most recent experiment, autonomous ground and aerial systems were used for sensing and surveillance to provide reconnaissance and improve situational awareness for infantry units moving through natural desert and mock city blocks. AI is used to synthesize the information accumulated through a network of warfighter and unmanned nodes, cutting through the noise to provide the squad members with actionable intelligence directly to their handheld devices.  46  With advances in real-time analytics and recommender systems technologies, such computational support could help warfighters gain the initiative in dynamic operational settings.  47  But if human operators do not trust the system, they would be reluctant to follow its recommendations.  48  Thus, without trust in human-machine teams, the U.S. military may not be able to capitalize on the advantages in speed, coordination, and precision AI promises to deliver. While distrust is a form of poor calibration where human trust falls short of the technology's capabilities, over-trust or uncritical trust is another form of inappropriate reliance on technology, often described as 'automation bias.' Automation bias typically manifests itself in two types of errors: errors of omission, when people do not notice problems because the machine did not alert them, and errors of commission, where people follow automated commands or suggestions that are incorrect or inappropriate. Much of the research on automation bias comes from studies in aviation, including the analyses of incident reports citing overreliance on automated flight management systems as well as simulations and experiments.  49  For instance, one study tested for both types of errors on a group of 25 pilots in a simulated flight experiment.  50  To test for commission errors, pilots were presented with an automated alert warning that the engine was on fire, though other engine parameters were normal and no other indicators of trouble appeared. The test for omission errors included misloading data such as altitude clearance and frequency change. While the pilots missed incorrect information and other automation failures 55 percent of the time (omission error rate), all of the participants shut down the engine in response to the fake fire alert, indicating a 100 percent commission error rate.  51  Interestingly, the study also found that more experienced pilots were less likely to detect automation failures, pointing to how learned trust can exhibit itself through complacency and dependence on technology that builds up over time, especially when the system has proven itself reliable. Overall, the results, supported by other studies in aviation and health care, demonstrate that when automated decision aids are available, people tend to follow their cues.  52  To an extent, both errors of commission and errors of omission stem from the fact that humans tend to be \"cognitive misers,\" meaning they choose the option requiring the least cognitive effort and are not likely to seek alternative options.  53  Evidence from other studies in human factors literature show that as it becomes more difficult for human operators to disaggregate the factors that influenced the machine's decision, they become more likely to accept these solutions without question.  54  Notably, much of this research examines automated systems or more traditional expert systems that perform scripted tasks based on specified rules. These systems are less advanced than today's machine learning systems, especially those using deep neural network approaches that reason, reach conclusions, provide recommendations, and take action in ways not evident or easily explained to humans.  55  It is difficult to predict how such technological advances could affect trust in human-machine teams. The increased sophistication of intelligent technologies could amplify the inclination to over-trust complex systems. On the other hand, people may be reluctant to trust systems they do not understand. Such uncertainty only highlights the significant role broader social, organizational, and institutional structures and practices have in helping individual operators to properly calibrate trust depending on the capabilities and limitations of the system and the task at hand-serving as a bulwark against the risks that stem from both uncritical trust in technology and deficits in trust.  56  Thus far, the discussion has focused on research studying human attitudes toward technology in order to better understand the drivers and effects of trust in human-machine teams. Yet a holistic understanding of trust in human-machine teaming accounts for all three-the human, the machine, and the team-each as its own unit of analysis. The following section therefore centers on AI system features, as well as the interactions and interdependencies between humans and intelligent technologies that facilitate the development of trust. \n From Intelligent Tools to Trusted Partners For humans to properly calibrate trust-that is, to accurately gauge the extent to which an intelligent system can be relied upon given its capabilities, limitations, and the context at hand-they need to understand what a system can and cannot do in a given mission or environment, and why it makes the decisions it does. The transparency of the system, the capacity of the system to explain its decisions, the quality of communications between human and machine, and the reliability of the system in the present and future are all critical factors for calibrating trust and enabling effective human-machine teaming. Research and innovation has therefore focused on ways to 'build in' trust into autonomous and AI-enabled systems through features and functions that make these systems more transparent, explainable, auditable, reliable, robust, and responsive.  57  While the discussion below focuses predominantly on operators or end-users, technology-centric solutions to enabling trust in intelligent systems increasingly involve collaborative design between scientists, technicians, soldiers, and commanders, as well as efforts to test AI systems earlier in the development process to gather feedback. The capabilities and limitations of AI systems, however, change depending on the context at hand. Moreover, advanced AI systems are designed to continuously learn and alter themselves, even after they have been in operation. The trustworthiness of AI systems should therefore not be treated as static or permanent.  58  Indeed, this is partly why there are substantial challenges to testing, evaluation, validation, and verification of AI systems. While a comprehensive account of these issues is beyond the scope of this report, suffice it to say that senior defense leaders both outside and within the Pentagon recognize these challenges and are beginning to take steps toward reforming processes and practices that can accommodate a collaborative, holistic, and continuously evolving approach to building and deploying trustworthy AI systems.  59   \n Transparency, Explainability, and Auditability Transparency is a critical aspect of trustworthy technologies and is essential for calibrating operator trust in a machine. But defining the type and level of information the intelligent system needs to convey to the human operator, as well as how to communicate said information so that it is understandable to humans, remain areas of open inquiry. When approaching the question of what information is important for transparency and trust in human-machine teams, researchers have looked into factors such as the intelligent agent's current actions and plans, reasoning process, projected outcomes, and uncertainty. The Department of Defense's Autonomy Research Pilot Initiative (ARPI), for example, conducted a series of studies exploring human interactions with an autonomous squad member-a robotic mule that accompanies a dismounted soldier squad within a simulated military environment. To determine how different configurations of information influence human perceptions of the autonomous squad member, the robot shared information about its current goal (e.g., to return to base), current priority (e.g., to save time), and its projected resource expenditure (e.g., how much extra fuel it needed to use to meet its goal given said priority). The study showed that participants' situational awareness and understanding of the robot peaked when the robot displayed information about its intent, logic, and possible outcome while the addition of uncertainty information did not further enhance trust.  60  The finding regarding uncertainty information is notable, considering other research that shows humans perceive agents as more trustworthy when they convey uncertainty estimates and that doing so can also improve joint human-machine team performance.  61  In other words, while information about uncertainty can be beneficial, it may also cause confusion and prove less useful depending on the mission environment or individual differences.  62  More research is therefore needed to better understand how uncertainty affects trust (as well as performance) in humanmachine teams. Transparency about failures and errors is particularly important for trust in human-machine teams. As previously mentioned, humans are quick to lose confidence when a machine makes mistakes. The system should therefore be able to inform the user about the causes and the resulting impacts of the failure to the system and the mission, and ideally, also be able to present information on how to diagnose and mitigate the errors.  63  Related to transparency is the issue of explainability, or the ability of an AI system to explain its rationale to human users, articulate its strengths and weaknesses, and convey how it will behave in the future. There seems to be a consensus that in order to properly calibrate trust and use AI systems effectively, people need to understand how these systems work and why they reach the conclusions that they do.  64  Research on explainability therefore tackles the black-box problem with AI (and more specifically MLbased systems)-namely, that many algorithms, including those based on deep learning, are opaque to users, with few mechanisms available for explaining their reasoning and results. Part of the challenge is that while AI techniques such as decision tree induction have built-in explanations, so far they have been generally less accurate compared to more complex deep learning algorithms, which perform better but are less explainable.  65  With the current state of the technology, developers therefore face a tradeoff in their choice of algorithm, whether to optimize for performance or for explainability-with both parameters being pertinent to achieving and maintaining trust in human-machine teams. One of the key DoD research initiatives in this area is DARPA's \"Explainable Artificial Intelligence (XAI)\" program, focusing on three interrelated challenges: developing new ML techniques that produce more explainable models, designing new strategies and techniques for human-computer interaction and intelligent user interfaces for conveying effective explanations, and investigating the psychological requirements for effective explanations that help humans intuitively and quickly understand the system's rationale.  66  XAI research raises the issue of communication in human-machine teams which speaks to both system design, (i.e. AI system characteristics and functionalities) and the interactions between humans and intelligent technologies. As previously noted, different AI representations-physical robots, virtual bots, or embedded AI-evoke different cognitive and emotional responses that impact trust. The field of social robotics has shown that human-like design features, social behaviors, and implicit features and behaviors related to communication such as posture, head or eye movements, or changes in proximity can influence the human team member's understanding and trust.  67  The majority of fielded military systems today, however, have minimal anthropomorphic features. User displays, ubiquitous in human-machine teaming, are therefore particularly pertinent to communication in human-machine teams, and more specifically to how the system should convey information to best calibrate trust. User displays vary significantly in their design and functionalities depending on the nature of the human-machine interaction, including factors such as task allocation, decision-making authority, and environmental constraints. Across these different configurations, however, decisions about interface design, layout, and graphics that visualize information all have an impact on the operator's perception of the system's current plans, comprehension of the system's behavior, and projection of future outcomes.  68  An intuitive, easy-to-use interface can improve the operator's situational awareness and limit ambiguity by assuring human team members that the agent is aware of its environment. Such design can increase overall team performance by reducing communication times and minimizing errors. A smart, well-designed interface can even reduce the cognitive workload of human teammates by allowing the agent to perform tasks such as analysis, perception, or navigation best suited to its capabilities.  69  However, there is a delicate balance between ease of use and ensuring trustworthiness through transparency: the most detailed interfaces that provide information about the intricate inner workings of the system may be transparent, but are not necessarily the most user-friendly or conducive to optimal operator performance and effective humanmachine teaming. Currently, the technical challenges facing the development of displays and related audio or visual communication modalities are intertwined with the limits of AI technologies. Intelligent agents struggle with interpreting complex and ambiguous situations, understanding they have made a mistake and communicating the reasoning behind their decisions. With these underlying technological limitations, it is hard to know what type, how much, and how often information should be provided to the human operator. But as AI systems continue to evolve both technologically and socially, and human-machine teams proliferate across multiple tasks and domains, understanding the effects of displays on trust will become increasingly important. Finally, while much of the discussion above has focused on the factors influencing trust between human operators and technology, there are other forms of transparency that may influence the trust of other audiences and publics. One such approach to ensuring transparency speaks to the need for traceable and auditable data sources, design procedures, and development processes of AI systems. In the commercial space, technology companies such as IBM have made trust and transparency a part of their operating principles, pushing for greater clarity on who trains AI systems, what data is used in training, and what goes into an algorithm's recommendations. DoD AI ethics principles also call for traceable AI systems, stressing that technical experts within DoD need to possess an appropriate understanding of the technology, development processes, and operational methods of its AI systems, including \"transparent and auditable methodologies, data sources, and design procedure and documentation.\" 70 Documentation practices will help ensure that AI systems are used appropriately, responsibly, and ethically, and that users are able to calibrate expectations and trust in what the system can and cannot do in a given context.  71  Moreover, auditability may prove useful for restoring trust in the event of machine malfunction or an accident, providing a track record of what happened and how such incidents can be avoided in the future. \n Reliability, Robustness, and Responsiveness Whether the operator can trust the AI system to function properly is perhaps the most fundamental question in human-machine teaming. Indeed, while transparency and explainability are important for calibrating trust, reliability may be even more critical. For instance, one of the aforementioned ARPI studies found that participants' trust in the autonomous squad member declined when the robotic mule made errors, and that displaying information to support transparency did not mitigate the impact of the errors on trust. On the other hand, when the autonomous squad member was reliable, participants anthropomorphized the agent more than when it was unreliable, ranking it as more likable, intelligent, and safer to work with.  72  Another study in which a human teammate worked with a robot in reconnaissance missions found a similar trend: when the robot's ability was high and it proved reliable, the explanations it provided about its decisions had no significant impact on trust.  73  While these studies suggest reliability may trump transparency and explainability for engendering and calibrating trust in human-machine teams, additional research on the interaction between reliability and explainability could offer further clarity and nuance. Alongside reliability and robustness, advances in machine intelligence and capabilities that allow the technology to interact with the environment and be responsive to users also impact trust. Responsiveness, adaptability, cooperation, and pro-social behavior of intelligent technologies strengthen cognitive trust by raising expectations of high-quality performance and positive experience during mutual tasks or missions.  74  Machine behaviors that reflect social intelligence like active listening and personalization have also been linked to higher levels of emotional trust, with users reporting greater levels of engagement, likeability, and enjoyment.  75  Moreover, experimental research shows that cooperative behavior of intelligent agents can increase human-machine team performance as well as support resilience.  76  There are non-negligible technical challenges to progress in research centered on ensuring that AI systems are reliable, robust, and responsive, especially in complex adversarial environments. That said, experimentation and fielding of certain systems earlier in the development cycle can provide an opportunity for incorporating user feedback that helps the system learn and improve, as well as for building trust in human-machine teams. According to Mark Lewis, former Acting Deputy Under Secretary of Defense for Research and Engineering, one of the main goals is to figure out which AI applications will have the biggest impact on the warfighter. \"In some cases,\" Lewis explained, \"that means getting the technologies in the hands of the warfighter and having them play with them, experiment with them, and figure out what makes their job more effective … [and] easier,\" as well as \"to discard the things that don't buy their way into the war fight.\"  77  One example of such efforts is DARPA's Air Combat Evolution (ACE) program which aims to increase warfighter trust in autonomous systems by using human-machine collaborative dogfighting (air-to-air combat) as its initial challenge scenario. As AI systems train in the rules of aerial dogfighting, their performance will be monitored by fighter instructor pilots which will help mature the technology. Once the human pilots feel the AI algorithms are trustworthy in handling the bounded and predictable environment, aerial engagement scenarios will grow more difficult and realistic, eventually going from virtual testing to demonstrating dogfighting algorithms on live, full-scale manned-unmanned teams.  78  As a whole, building transparent, explainable, auditable, reliable, robust, and responsive intelligent systems will help foster appropriate trust in human-machine teams. Continual feedback between humans-developers, operators, commanders-and machines during the entire lifecycle of a system is another key element of the systems engineering approaches that seek to 'build in' trust into the intelligent machines. Such feedback is also instrumental to what some have referred to as a human-centric approach to AI development which seeks to integrate the needs, perceptions, and behaviors of the user into the design of AI systems.  79  That said, technological solutions alone cannot solve the trust problem in human-machine teams. \n U.S. Military Research: Gaps and Future Directions In October 2020, CSET published a report on U.S. military investments in autonomy and AI, analyzing publicly available data from the FY2020 research, development, testing, and evaluation budget justification books of the Army, Air Force, Navy, and DARPA, focusing specifically on basic, applied, and advanced research.  80  The findings showed that human-machine collaboration and teaming is a crosscutting theme across autonomy and AI research and development programs related to unmanned systems, information processing, decision support, targeting functions, and other areas. That said, only 18 of the 789 research components related to autonomy and 11 out of the 287 research components related to AI mentioned the word \"trust.\"  *81  There are a number of possible explanations for this apparent gap. For one, while there is a rich literature on human-automation interactions, and the role of trust therein, there is far less research on human-autonomy and human-AI interactions, and specifically on trust in human-autonomy and human-AI teams.  82  Technology, it seems, has outpaced research on human-machine teaming. The U.S. military is developing autonomous systems capable of performing an ever-increasing range of tasks with limited, if any, human supervision and ML-based systems that learn and adapt to their environment. Yet much of what we know about trust in intelligent technologies still draws on research examining human interactions with automated systems and more traditional expert systems. The gap in research on trust in autonomy and AI in DoD's * While we found relatively few instances where the word \"trust\" was mentioned, descriptions of different autonomy and AI research initiatives also included other keywords that signal research related to trust in human-machine teams, including but not limited to: assurance, reliability, robustness, resilience, predictability, explainability, interpretability, transparency, etc. These system features and characteristics are pertinent to trust, and can be thought of as elements of trust and components of effective human-machine teaming. But they are not synonymous with trust. science and technology program reflects the broader state of the field.  83  Furthermore, as previously noted, trust is a complex, abstract and hard to measure concept. Defense research therefore tends to favor technology-centric approaches geared more directly toward enhancing AI system attributes that are related to trust, including security, robustness, resilience, and reliability. DARPA leads in research focused on developing systems that behave reliably in operational settings and strengthening security in the face of adversarial attacks, with programs such as \"Guaranteeing AI Robustness against Deception (GARD),\" \"Lifelong Learning Machines (L2M),\" and \"Assured Autonomy.\"  84  The Army also has several relevant initiatives. For instance, as part of its basic research portfolio, the \"Army Collaborative Research and Tech Alliances\" effort includes research on \"AI-enabled cyber security that is robust to enemy deception,\" supporting \"Army counter-AI against near-peer adversaries.\"  85  These efforts represent systems engineering approaches that seek to \"build trust into the system,\" and are indeed necessary for establishing and properly calibrating trust in human-machine teams. But they are not sufficient. For intelligent machines to become true teammates, they need to be able to adapt to changing and new environments. The U.S. military has a number of research programs focused on assurance approaches for systems with advanced levels of autonomy that continue to learn and evolve after they are deployed. Yet the very ability to learn and adapt to the environment, as Heather Roff and David Danks argue, could undermine the human team members' trust.  86  Trust, in both human and human-machine relationships, is built on repeated interactions that provide information about values, preferences, beliefs, and other factors that help develop shared goals and expectations, as well as allow people to evaluate risk, especially in high-stakes situations. But as the AI system learns and adapts, it may change, often in ways that are unexpected or not understandable to humans. As Roff and Danks assert, \"the battlespace is a dangerous place to be figuring out the preferences and values of a dynamically adapting weapon, so it is unsurprising that trust will be difficult to establish.\"  87  Indeed, one of hardest challenges related to adaptive machine learning and trust in human-machine teams is how to ensure that the trust that has been \"earned\" by a system in a predictable, fixed environment translates not only to different, dynamic environments, but also to different machines as new team members and/or with new human team members.  88  Dispositional factors such as age, gender and cultural background impact people's attitudes and trust in technology. People behave differently and often unpredictably under stress and in high-stakes situations. Emotional factors, previous experiences with intelligent technologies as well as broader institutional and societal structures, and organizational culture all play a role in shaping the nature of trust in human-machine teams. Thus, while trust requirements built into a given system may cultivate appropriate trust in a particular human-machine team, there is no guarantee this \"built in\" trust holds for new human team members.  89  For example, a recent study from the Army Research Lab examined soldiers' trust in their robotic teammates in autonomous driving scenarios by grouping individuals in four different categories based on \"demographics, personality traits, responses to uncertainty, and initial perceptions about trust, stress, and workload associated with interaction with automation.\"  90  Based on a facial expressivity analysis, the researchers found that these groups had unique differences in their responses and attitudes toward the driving automation. The study therefore concluded that trust calibration metrics may not be the same for all groups of people and that trust-based interventions, such as changes in user display features or communication of intent, \"may not be necessary for all individuals, or may vary depending on group dynamics.\"  91  This report does not advocate for the study of trust as an end to itself. Rather, we suggest that research focused explicitly on the drivers and dynamics of trust in human-machine teams can augment technology-centric approaches to building trust into AI systems. With this in mind, we offer the following directions for continued and additional research that could contribute to advances in human-machine teaming and the development of trustworthy AI systems. • Multidisciplinary research on the drivers of trust in humanmachine teams, specifically under operational conditions. Research on human-machine trust, including scholarship that applies sophisticated computation models of cognition to understand issues such as knowledge acquisition and problem solving, is predominantly conducted under closely controlled laboratory conditions.  92  More research is needed to assess whether these findings withstand complex realworld conditions and tasks. • Collaborative research between U.S.-based researchers and defense research communities in allied countries to assess how cross-cultural variation in trust in human-machine teams may impact interoperability in multinational operations. • Research to assess what aspects of transparency are most relevant for calibrating trust in human-machine teams, especially under operational conditions. For instance, how important is explainability vs. auditability, i.e., is the ability to understand how AI systems reach a particular conclusion more conducive to building, maintaining, and adjusting trust in human-machine teams than visibility into the data and models? • Research on the interaction between explainability and reliability. While there seems to be a consensus that in order to trust their machine teammates, humans need to understand why autonomous and AI-enabled systems behave as they do, some research suggests that as long as the system is reliable, explainability is less important. Additional research could help connect and contextualize these seemingly contrasting views. • Research on shifts in cognitive workloads and trust calibration across different types of human-machine teaming. For example, research on autonomous vehicle technology for Army convoy operations shows that in a mix of manned and unmanned trucks, the soldiers who remain in the convoy would perform more tasks involving sensing and decision-making, resulting in a higher cognitive burden than their counterparts in a fully manned convoy (where cognitive burden can be shared across a larger number of soldiers).  93  This is significant considering there is evidence that users make more automation bias errors under higher workload conditions, when performing complex tasks or multitasking.  94  As such, there is a need for more research on how the distribution of tasks and decision-making responsibility in human-machine teams (and the resultant shifts in cognitive workloads) affect trust specifically in military settings. • Research on uncertainty and trust calibration. What aspects of uncertainty are most critical for humans to understand in order to calibrate trust and use the system effectively, and how should this information be communicated? • Research on reliability and trust calibration. Keeping in mind the growing urgency to field military AI systems, what are the minimum standards for AI system reliability, robustness, and resilience necessary for building and maintaining trust in human-machine teams? How do these standards vary based on operator characteristics, mission, environmental conditions, and the distribution of tasks and decisionmaking authority within the human-machine team? While this is certainly not an exhaustive list, we believe additional research on these topics could help advance the U.S. military's vision of using intelligent machines as trusted partners to human operators as well as further the development of reliable, trustworthy, and safe AI systems that would cement U.S. military and technological advantages into the future. \n Conclusion The U.S. military sees many uses for human-machine teams, and with advances in AI technology, machines will be able to take on a greater variety of tasks and responsibilities, extending humanmachine teaming to additional mission areas and functions. But progress toward advanced human-machine teaming will depend on advances in understanding human attitudes toward technology as well as breakthroughs in AI technologies, making these systems more transparent, explainable, auditable, reliable, robust, and responsive. We offer a number of research directions that could help the U.S. military move forward with its vision of using intelligent machines as trusted partners to human operators: greater emphasis on research and experimentation under operational conditions; collaborative research with allied countries; research on trust and various aspects of transparency; research on the intersection of explainability and reliability; research on trust and cognitive workloads; research on trust and uncertainty; and research on trust and reliability. As the U.S. military integrates AI technologies and capabilities into the force, resolving outstanding questions around the issue of trust in human-machine teams becomes increasingly imperative. There are no simple solutions and no single approach will suffice. But insights from research on the dispositional, situational, and learned factors that shape trust as well as the broader institutional and societal structures that influence people's attitudes and behaviors toward technology can inform and strengthen systems engineering approaches to building trustworthy AI. \n Authors Dr. Margarita Konaev is a research fellow with CSET. Tina Huang is a research analyst with CSET currently serving as a fellow in artificial intelligence policy for a member of Congress with a leadership role in AI issues. Husanjot Chahal is a research analyst with CSET. oo Research and experimentation under operational conditions, o Collaborative research with allied countries, o Research on trust and various aspects of transparency, Research on the intersection of explainability and reliability, o Research on trust and cognitive workloads, o Research on trust and uncertainty, and o Research on trust, reliability, and robustness. \n Summary ......................................................................................... \n\t\t\t © 2021 by the Center for Security and Emerging Technology. This work is licensed under a Creative Commons Attribution-Non Commercial 4.0 International License. To view a copy of this license, visit https://creativecommons.org/licenses/by-nc/4.0/. Document Identifier: doi: 10.<PHONE>", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Trusted-Partners.tei.xml", "id": "b9e021016e7fadfd8c2583ea9a1a443a"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "is a research organization focused on studying the security impacts of emerging technologies, supporting academic work in security and technology studies, and delivering nonpartisan analysis to the policy community. CSET aims to prepare a generation of policymakers, analysts, and diplomats to address the challenges and opportunities of emerging technologies. During its first two years, CSET will focus on the effects of progress in artificial intelligence and advanced computing.", "authors": ["Tim Hwang"], "title": "Shaping the Terrain of AI Competition", "text": "he concern that China is well-positioned to overtake current U.S. leadership in artificial intelligence in the coming years has prompted a simply stated but challenging question. How should democracies effectively compete against authoritarian regimes in the AI space? Policy researchers and defense strategists have offered possible paths forward in recent years, but the task is not an easy one. Particularly challenging is the possibility that authoritarian regimes may possess structural advantages over liberal democracies in researching, designing, and deploying AI systems. Authoritarian states may enjoy easier access to data and an ability to coerce adoption of technologies that their democratic competitors lack. Authoritarians may also have stronger incentives to prioritize investments in machine learning, as the technology may significantly enhance surveillance and social control. No policy consensus has emerged on how the United States and other democracies can overcome these burdens without sacrificing their commitments to rights, accountability, and public participation. This paper offers one answer to this unsettled question in the form of a \"terrain strategy.\" It argues that the United States should leverage the malleability of the AI field and shape the direction of the technology to provide structural advantages to itself and other democracies. This effort involves accelerating the development of certain areas within machine learning (ML)-the core technology driving the most dramatic advances in AI-to alter the global playing field. This \"terrain\" approach is somewhat novel in literature on AI, national strategy, and geopolitical competition. However, the framework presented Executive Summary T Introduction eadership in technological innovation has long been a crucial national security asset to the United States. From atomic energy and stealth technology to the internet and genetic engineering, the United States has led the way in nearly all the major breakthroughs of the last decades. This has served its interests not just on the battlefield but economically as well. The potential loss of this national technological lead serves as a powerful motivating force and locus of discussion in national security circles. The Soviet Union's launch of Sputnik in 1957 triggered a major surge of investment and coordination activity to ensure the United States was not left behind in aerospace technology. Numerous, less popularly known examples pop up in other domains, including security concerns around a perceived loss of superiority in developing supercomputers  1  and green technology,  2  as well as more conventional fears over new weapons technology such as hypersonic missiles.  3  Commentators worry too about the decline of science and technology education in the United States and its long-term impact on the nation's global dominance.  4  In recent years, China has stoked these fears perhaps most among U.S. security analysts. Chinese technology firms seem well poised to compete, if not outcompete, their U.S. counterparts. The Chinese government has made several major announcements signaling an aggressive campaign to invest in and quickly advance a range of critical technologies. To the extent that U.S. political and economic dominance hinges on technological dominance, the two global powers seem poised to settle in for an extended period of competition. \n L Recent breakthroughs in artificial intelligence-specifically in the subfield of AI known as machine learning (ML)-have become wrapped up in this broader concern around a loss of the U.S. technological edge. China has made AI a key priority, announcing a raft of new initiatives that signal major investment and state interest in becoming a global leader in ML. In July 2017, the Chinese government announced its \"Next Generation Artificial Intelligence Development Plan,\" a detailed and specific agenda designed to position China as the world's premier AI innovation center by 2030.  5  Chinese cities and states have pledged billions of dollars to support AI development in their regions. Chinese universities train massive numbers of engineers and researchers in the field, and Chinese products powered by AI have proven wildly successful both at home and abroad. By October 5, 2018, China boasted 14 \"unicorn\" AI companies-private companies valued at $1 billion or more.  6  Mobile application unicorn ByteDance, which owns TikTok, credits its successful expansion beyond the Chinese market to the broad applicability of user data collected and processed with technologies developed in ByteDance's AI lab.  7  Commercial drone manufacturer DJI, based in Shenzhen, has partnered with Microsoft to build out a suite of AI capabilities to enhance its hugely popular remote-controlled robots.  8  The United States, which has made AI a core part of its defense strategy and whose leading companies have made AI a key differentiator in their products and services, perceives these developments as a major national security risk. Some commentators have dubbed the competition between the United States and China a new \"arms race\" in AI.  9  Congress has also taken action, creating and funding in 2018 the National Security Commission on Artificial Intelligence, tasked with reviewing developments in AI to address the national and economic security needs of the United States and seek out \"opportunities to advance U.S. leadership.\"  10  The concern that China is well positioned to overtake current U.S. leadership in AI has prompted policymakers to question how the United States should most effectively compete in the AI space. Policy researchers and defense strategists have offered a number of different possible paths forward.  11  But this has not been an easy task. One particularly thorny challenge has been the possibility that authoritarian regimes may possess structural advantages over liberal democracies in researching, designing, and deploying AI systems. Authoritarian states may enjoy easier access to data and more effective tools to coerce adoption of technologies that their democratic competitors do not. Authoritarians may also have stronger incentives to prioritize investments in ML as the technology may significantly enhance their systems of surveillance and social control. This paper offers one answer in the form of a \"terrain strategy.\" It argues that the United States should leverage the malleability of the field of AI and work to shape the direction of research in ways that provide structural advantages to itself and other democracies. This involves accelerating the development of certain areas within ML-the core technology driving the most dramatic advances in AI-to alter the global playing field of the technology. Such an approach would prioritize investments in research that reduces dependence on large corpora of real-world training data, improves the technology's democratic viability, and attacks the social control uses of high value to authoritarian states. In essence, the terrain strategy seeks to empower democracies to effectively compete in the technology by reshaping the nature of the technology itself. Competing in AI without engaging with the technology in this way will leave the United States in particular and democracies in general at a structural disadvantage against their authoritarian competitors. Even worse, democracies may be in the unenviable position of having to compromise on core values like privacy in an effort to preserve their technological lead. The terrain strategy offers one path whereby democracies might effectively compete without having to make such sacrifices. It is easy to assume that AI is a monolithic, single technology whose applications, research ecosystem, and future directions are already settled. This picture does not reflect reality. ML is both highly multi-faceted and deeply malleable. Rather than a single monolith, the technology is better understood as a broad family of related but distinct techniques, each with unique strengths and weaknesses. The competitive strategy around AI must take these characteristics into account. It is not just a matter of whether or not to invest in ML, but specifically what provides the greatest national benefit within the domain of ML. Failing to examine these details may prevent the national security community from identifying important opportunities that enable the United States to retain its edge and even outpace competitors like China in the mid-to long-term. The first part of this paper proposes a strategic framework for thinking about global competition in AI and argues for shaping the research field to give the United States and its allies the advantage in the technology. The second fleshes out this framework, recommending specific, promising technical domains that the United States should accelerate in order to execute on this strategy. he current state of play in ML favors authoritarian societies over democratic ones. To win, democracies must work to re-shape the competitive dynamics of ML in order to retain their leadership. This section outlines why these structural advantages exist and how AI might be reshaped to rectify this imbalance, then argues that democratic governments should play a role in addressing private underinvestment in certain areas of ML research. \n ACCESS TO INPUTS DEFINE COMPETITIVE ADVANTAGE IN ML Viewing ML as a technology whose success depends on the availability of a specific set of resources enables us to think concretely about the kinds of actors and entities that are best positioned to obtain the benefits of these technologies. It is worthwhile to take a step back from the intense hype around ML to think for a moment about what the technology actually is. Simply stated, ML is the subfield of AI studying computer systems that improve through processing data. This improvement process is called \"training.\" The training process generates a piece of code-known in the field as a \"model\"which ideally can then accomplish the trained task. Consider the example of teaching a computer to recognize a cat in a photo. ML requires a large corpus of training data to do this: images of cats that are manually labeled by annotators as depicting a cat. These are processed through a learning algorithm, whereby a model is generated that associates the visual of the cat with the label \"cat.\" This training process, in effect, is a large number of mathematical operations enabling the machine to infer a set of rules that accomplish the task: detecting a The Terrain Strategy 1 T cat in an image effectively. If the process is successful, the resulting model can then accomplish this task with novel images it has not seen before. From even this rudimentary example, it is clear that the successful use of ML relies on a few core inputs.  12  Specifically, it requires  (1)  training data, which are the examples that the algorithm learns from; (2) learning algorithms, the algorithms that execute the training process; (3) computational power, the computers necessary to run the many calculations needed to generate a model; and (4) talent, the human expertise necessary to set up these systems and assess the quality of the resulting model. For the vast portion of ML applications used today, the absence of one of these resources will result in poor quality models or an inability to create ML systems at all. It is possible to look to these inputs to make concrete predictions about the relative advantages and disadvantages that different actors bring to the competitive landscape. Imagine for a moment the marketplace for cat detection technologies. Who in the market is well poised to offer an ML-driven cat detection system? Will it be one of the existing, established players in the space, or an upstart? Among upstarts and incumbents, who will have the strongest chance at building the highest-performance systems? The business running an online cat lover community may already have access to a large number of cat images; this access places it at a cost advantage against a business that needs to purchase this training dataset from a third party or hire photographers to go out and collect many photos of cats. Similarly, an upstart company with extensive expertise building ML models in other domains might come to the market with an established team of experts, allowing it to outcompete an incumbent cat photo sorting giant with a lower capacity to recruit specialists to build the technology. Of course, none of these assessments is definitive in determining who might emerge as the dominant business in a sector. A wily competitor able to market effectively their subpar AI systems might still prevail over a company offering a technically superior product. Whether or not a competitor with more relevant data but less technical expertise triumphs over a rival with less data but more technical expertise will be highly dependent on the context. Non-technical factors, such as organizational effectiveness in integrating the technology into existing processes, the costs of training personnel, and the quality of software development practices will play a major role.  13  Inputs set the stage, but there is no law that makes them determine the final outcome. Nonetheless, it is indisputable that the relative distribution of access to ML inputs exerts a powerful influence over the competitive dynamics between actors and shapes the strategies they might bring to the field. Such a lens allows us to make some structured conjectures around the structural advantages and disadvantages that different actors face in attempting to leverage the benefits of ML. \n AN AUTHORITARIAN ADVANTAGE The relationship between inputs and competitive dynamics in ML can be generalized beyond commercial competition between businesses in next-generation cat detection. The effectiveness of an ML system is based on the ability to marshal relevant data, computing power, learning architectures, and expertise. That applies regardless of whether the task is to create ML systems for sorting cat images or piloting a military drone. This analysis can be expanded to ask a far broader question: are certain societies or governance structures more or less able to leverage the benefits of ML? More to the point, are centralized autocracies better or worse than liberal democracies in building, training, and deploying AI, particularly for forwarding the interests of a state? There are three reasons to believe that authoritarian regimes have a relative advantage. For one, they may have an easier time acquiring data for training ML applications,  14  in part because they may already maintain an existing infrastructure for ubiquitous surveillance that enables easy data collection. Moreover, there may be no strong legal mechanisms to protect citizen privacy or prevent the state from compelling companies to provide access to their data. In contrast, liberal democracies may impede the collection of data through relatively robust privacy regimes. To the extent that these societies have large-scale systems of data collection, these datasets may be centralized within private corporations with legal protections, rather than in an institution immediately accessible by the state. Second, authoritarian regimes can more effectively force the deployment of novel ML technologies, allowing these systems to be rolled out and fine-tuned through subsequent data collection without needing to obtain the consent of the general public. Democratic commitments to public consent mean citizens have comparatively more mechanisms for resisting unwanted deployments of ML technology. Individuals and civil society organizations can protest or bring lawsuits to prevent the adoption of certain ML systems. Media freedoms allow journalists to expose and rally public opinion against objectionable uses of the technology. Key communities of technical specialists can refuse to work on certain applications of ML, and discourage their employers from doing so, as well.  15  Third, the surveillance and control interests that authoritarians bring to their investments in ML may also make it challenging for liberal democracies to make the comparable financial commitments necessary to retain their lead on the cutting edge of the technology. For authoritarian regimes, refining and perfecting ML may go directly to a core priority of sustaining and protecting the state, whereas democratic societies may face more conflicting incentives around whether or not to prioritize and coordinate their investment. Authoritarian regimes, therefore, may be able to train, deploy, and improve ML systems more effectively than democratic societies. Facial recognition technology provides a concrete illustration of these authoritarian advantages. Companies specializing in building facial recognition models such as SenseTime have benefitted significantly from collaborations with the Chinese government on surveillance applications.  16  These collaborations have yielded access to extensive face data for training ML models and provided opportunities to test their technology at scale without the need for public consent. Training and deploying facial recognition systems has not been as easy in the United States. Academic researchers have criticized companies marketing these technologies, highlighting the gender and racial biases that facial recognition systems might perpetuate.  17  Journalists have aggressively exposed unscrupulous practices that some startups have engaged in to gather training data.  18  Civil society organizations have successfully lobbied for municipal bans on the use of the technology throughout the country.  19  Industry leaders in ML such as Google have spoken out in favor of a moratorium on the use of facial recognition technologies.  20  Driven by the public concern around facial recognition, Congress is considering laws that would require consent before collecting or sharing face data.  21  The result is that the U.S. government and companies contend with a more difficult environment for collecting data to train facial recognition capabilities and have less freedom to deploy these systems once trained. While this resistance works to protect civil liberties in the United States, the outcome is that Chinese companies will likely continue to lead in ML-based facial recognition technologies for the foreseeable future. These structural advantages do not automatically determine the winner in AI competition. Autocratic rollouts of ML-driven surveillance and social control mechanisms have occasionally been far less successful than some commentators have made them out to be.  22  Moreover, citizens of autocratic regimes have found numerous ways to defend their privacy and subvert state surveillance even in the absence of formal legal mechanisms.  23  Governments may acquire a cutting-edge technology, only to find that bureaucratic politics thwarts its usefulness in practice.  24  The specifics will matter. But all else being equal, autocracies have an edge when it comes to building and deploying AI systems. This raises an important question: given the structural advantages that more autocratic regimes bring to competition around AI, how can democracies retain and expand their edge in the technology? The national security community has leaned heavily on a relatively small set of tools to answer this question. Frequently cited proposals include streamlining the process for government funding of non-defense companies, increasing AI R&D budgets, developing private sector incentives-namely generous tax breaks-and recruiting talent from other countries.  25  While these policies would improve democratic competitiveness in AI, they do not directly address the core structural advantages authoritarian regimes can bring to the table in advancing the state of the art in ML. Democracies are left fighting an uphill battle, contending with headwinds their competitors do not face. Perhaps more troubling, democracies such as the United States may choose to compromise on their values in order to compete effectively in ML. Executive orders have already been issued to lower barriers to accessing citizen data in order to increase the availability of training data for ML systems.  26  Policies like these erode privacy protections for the sake of accelerating AI development. Democracies should seek to compete effectively while preserving their core values, rather than move in a more autocratic direction to preserve some semblance of technological parity. Democracies can do better than this. The shape of AI is not fixed, but in flux. The United States can strategically rewrite the underlying competitive dynamics of the technology, working to offset the structural advantages that autocracies enjoy while mitigating the structural challenges that democracies face. \n THE MALLEABILITY OF ARTIFICIAL INTELLIGENCE Most popular reporting on breakthroughs in ML tends to emphasize the advancement of the technology: a new performance milestone passed, a new level of investment reached, or a new application of the technology discovered. Less frequently spotlighted is that the practice of ML itself has rapidly morphed over the past decade of rapid progress. Democracies can take advantage of this malleability to offset autocratic advantage. ML has not merely improved; it has fundamentally changed. Some of these changes concern the practical ecosystem of engineering in ML systems. For one, the high-profile nature of the technology has encouraged a massive influx of talent into the field, from highly specialized researchers working in the field and advancing the state of the art to an increasing pool of yeoman software engineers familiar with the basics of the technology.  27  The universe of practical techniques, tools, and resources available for building ML systems has also expanded. This includes the development of open-source software packages for using ML, such as TensorFlow and PyTorch, and a range of training and educational resources for learning techniques in the field.  28  Other changes concern the research field of ML itself: the last decade of activity has led to shifts in what is being done with the technology-and how. Progress has been made in the subdomain of research focusing on \"few shot\" learning-the challenge of developing ML systems that can perform a task effectively with significantly less training data than typically required.  29  Adversarial examples-the subdomain of research focusing on the creation of and defense against seemingly innocuous inputs that can manipulate the behavior of ML systems-have also become a major area of activity in the past few years.  30  These changes are important because they modify the strategic landscape around AI. For example, the rapidly increasing pool of available global talent working in ML makes it more challenging for an actor to gain a robust strategic advantage by monopolizing key personnel. In the early 2010s, by contrast, ML had not yet received mainstream attention and the number of leading researchers focused on these technical problems was comparatively small.  31  Overcoming specific technical hurdles may be even more strategically significant. For example, major breakthroughs in \"one shot\" and \"few shot\" learning might reconfigure the competitive landscape. These techniques lower the requisite quantity of data needed to train a high-performance ML system. Democracies could leverage these techniques to produce effective AI systems despite having more limited access to data than their authoritarian counterparts. In this sense, \"one shot\" and \"few shot\" learning can work to offset a key authoritarian advantage in AI competition. The reasons for these technical breakthroughs are not mysterious: progress on a given problem will depend in part on researchers and technologists deciding to focus their energies on solving that problem. This makes the field and the shape of ML malleable insofar as an actor can influence how the technical community prioritizes its efforts. Democracies can use this malleability to their advantage. They can work to encourage progress on technical challenges that, if resolved, would mitigate the structural advantages that authoritarians currently bring to AI competition. Failing to do so may mean that democracies perpetually compete on uneven terrain. \n THE ROLE OF GOVERNMENT IN SHAPING THE AI TERRAIN ML is multifaceted, and a state might potentially invest in a wide range of targets in an attempt to shape the strategic terrain of the technology. How should governments identify the opportunities that will yield the greatest impact? Governments do not act alone in investing in ML. Quite the opposite, the funding and support flowing from states is just one part of a vast ecosystem of corporations and other funders that are financing AI around the world. This funding constitutes a set of market forces pushing and pulling the field of ML in different directions. Some of these directions might make it more likely for liberal democracies to excel in the technology, while others might make it less so. For each potential area of investment, one can ask if the existing ecosystem of financial and human capital will work to systematically shift the field toward ensuring that democracies can compete on even-footing or at an advantage. Is the field making it easier over time for democracies to overcome their limitations in obtaining training data? Is the field expanding the ability for ML engineers to meet requirements for public consent in the rollout of these systems? Where the market produces underinvestment in certain problem areas, the state can shape incentives to change the level of activity in that domain. For areas with sufficient private investment, governments can allow the market to shape the field. In other words, governments might act as an effective \"gap filler\" in the space by accelerating and fostering work on problems that otherwise would receive insufficient support due to the market incentives facing the private sector. This method does not require a substantial change in how liberal democracies engage in scientific research. It instead applies an existing, long-standing approach. Vannevar Bush, whose seminal 1945 report Science-the Endless Frontier \"entrenched the concept of government patronage of scientific research in policy discourse\" and inspired the creation of the National Science Foundation, recommended a similar framework.  32  As he wrote, \"[t]here are areas of science in which the public interest is acute but which are likely to be cultivated inadequately if left without more support than will come from private sources. These areas … should be advanced by active Government support.\"  33  This should continue to be an organizing principle even in today's dynamic ML research environment. This \"gap filler\" approach suggests that democracies should place their resources into advancing ML along three dimensions: data efficiency, democratic viability, and subversion of social control. First, democracies should invest in solving a set of technological problems that reduce or replace the reliance of ML systems on massive datasets. Liberal democracies will face a comparatively harder environment for acquiring data than authoritarian regimes due to their commitments to privacy and private enterprise. Reducing the data requirements needed to create effective ML systems will help eliminate a hard tradeoff between acquiring data and maintaining democratic values. It will also allow democracies to better compete at parity with their authoritarian competitors in training a range of AI systems. Leading industrial labs often assume plentiful access to data, limiting their incentives to make solving these types of data-scarce technical challenges a top priority. Investing in this area may offer great promise. Second, democracies should invest in advancing state-of-the-art techniques to ensure the technology is democratically viable. Liberal democracies will face challenges in unilaterally imposing technologies on their publics and in forcing collaborations between private industry and the state. Advancing methods and know-how that improve transparency, ensure fairness, and protect privacy will raise the possibility of consensual adoption of ML systems. This will work to offset the authoritarian advantage of being able to deploy AI systems unilaterally. Democracies can rally fragmented research efforts on these topics, helping to speed innovation and foster global norms around values that should be embedded in ML systems. Finally, democracies should invest in a set of methods that actively undermine and raise the risks of using these technologies for social control. This prong focuses on eroding the advantages authoritarian regimes may enjoy from developments in ML, rather than eliminating the structural disadvantages that liberal democracies may face. Investing in technologies that attack these applications will reduce their value to authoritarian regimes, potentially reducing investment by these regimes in the field over time and allowing democratic investment to more easily keep pace. This space is likely to see under-investment from the private sector. Although corporations have strong interests in building defenses against attacks on ML systems, they do not have similar incentives to commoditize and distribute tools to enable those attacks in the first place. The second part of this report examines each of these prongs in greater detail. For each, this paper argues the case for these priorities, highlights a series of specific technical areas in which investment might make a major impact, and examines the strategic implications if successful. emocratic societies should invest in advancing the field of ML in ways that offset authoritarian advantages in creating and deploying AI systems. Three research areas are likely to produce the highest impact on this front: improvement of data efficiency, enhancement of democratic viability, and subversion of social control. \n REDUCING DATA DEPENDENCE ML has traditionally relied upon access to a large corpus of data, which is used to train the statistical models solving the problem at hand. Detecting objects in images requires access to many images that have already been tagged with the objects of interest. Creating a translation system between two languages requires access to a large existing body of bilingual texts.  34  Conversely, failure to acquire data relevant to the training task has imposed a hard limit on the level of performance of ML systems. This technical hurdle advantages authoritarian societies over democratic ones. Privacy protections and a fragmented data landscape in democracies may make it more challenging to acquire the datasets needed to produce effective ML systems. Even when private companies-platforms such as Google or Facebook-possess access to massive amounts of data, governments in democratic societies may have difficulty accessing this data freely. Authoritarians therefore possess a \"data advantage\"-not in the aggregate amount of data they hold, but in their ease of accessing plentiful training data when needed. Liberal democracies could take a heavy-handed approach of working to eliminate these privacy protections and expand the ease of government and corporate access to training data. This is a high-cost, time-consuming Shaping the Terrain 2 D endeavor, in some cases requiring a sacrifice of strongly held democratic values to obtain the benefits of the technology. Such compromises may not be necessary. Rather than concede that ML will forever require large, real-world datasets, targeted investments in the technical field could potentially erode the authoritarian benefit of relatively frictionless data collection. This may allow democratic societies to better benefit from the technology and produce ML systems that perform at similar parity without expansive data collection, significantly leveling the playing field. This area may be a particularly promising one for acceleration and support in part because it is likely that the market will systematically underinvest in their development over time. One of the main motivations underlying major investments in ML and AI by companies like Google and Facebook is that these technologies leverage the massive datasets that these businesses already possess to enable next-generation products and services.  35  To that end, many industrial labs have invested aggressively toward advancing ML techniques where an abundance of data is presumed. Some investment into research on making ML methods work in data-limited environments does exist. Indeed, government research agencies like DARPA already support such research.  36  However, pressures to prioritize this work will remain limited so long as the primary use case of the technology remains situations in which data is plentiful. \n Investment Opportunities Three subdomains of technical research could reduce the dependence of ML techniques on access to massive corpora of data: \"few shot\" or \"one shot\" learning, simulation learning, and self-play. These are promising areas for investment. First, \"few shot\" or \"one shot\" learning techniques seek to enable ML systems to effectively train on a task with a significantly smaller quantity of data than typically necessary. These improvements can occur in many places throughout the design of an ML system, restructuring everything from the training data to the model itself in an effort to make training more efficient.  37  More ambitious in this context, \"zero shot\" learning seeks to enable an ML system to generalize to a task it has not previously been trained on.  38  Second, simulation learning focuses not on reducing the overall amount of data necessary for the training process, but on sourcing that data virtually and thus potentially more cheaply. A classic illustration of this approach's data advantage is in the problem of training a robot arm to successfully grasp an object. Applying ML to this task could mean physically assembling a bank of robotic arms repeatedly attempting to grasp different objects.  39  While this method does enable robotic arms to learn to manipulate three-dimensional objects, the physical costs of setting up and maintaining such a system to collect data can be expensive. Simulation bypasses the cost and complexity of maintaining physical robots. Instead, a simulated robotic arm in virtual space repeatedly attempts to pick up virtual objects. With an accurate simulation of the physical world, the machine is able to learn how to grasp an object in a real robot.  40  This reduces the need to draw the data from the real world, and in some contexts, may offer cheaper sourcing of relevant data than physical collection. Finally, self-play refers to methods in which reinforcement learning agents are set to compete with one another to improve their performance. In some contexts, this can reduce or eliminate entirely the need to draw on real world data to train a system. One striking illustration of this technique is in the transition from AlphaGo-DeepMind's system for playing the game of Go, which bested 18-time world champion Lee Sedol in 2016-to AlphaZero, a successor system introduced in late 2017. The earlier AlphaGo system was trained initially on 30 million moves of Go drawn from games played by humans to gain an understanding of the rules.  41  Its successor, AlphaZero, entirely eliminated the need for any real-world data, replacing it with successive rounds of play against itself to achieve expert performance.  42  The result is similar to simulation learning: the creation of highly proficient systems of AI without a reliance on real-world data. These research areas are noteworthy because they reduce the potential advantage of a competitor who either enjoys low-cost access to relevant data, or who already possesses significant endowments of data. Advancements in basic research could reduce the barrier to entry for training \"data scarce\" ML systems with comparable performance to those trained on larger datasets. This development may aid democratic governments in keeping pace with more autocratic competitors and may even help businesses and researchers in democratic societies keep up with counterparts in competitor nations. \n Strategic Implications: Narrow Parity Reducing the dependence of ML systems on immense, real-world datasets would encourage competitive parity between democratic states and authoritarian ones. However, there is an important caveat to this analysis. Even if democracies make strong investments in reorienting ML toward research on achieving high performance in low data situations, they are likely to see only a narrow form of parity with their authoritarian competitors. Many of the identified areas of opportunity are effective only in particular circumstances. Even with major breakthroughs around some of these research challenges, existing trends do not suggest that these developments would cause the data barrier to entry to fall across all potential applications of ML at once. Simulation learning is a good example of this limited form of strategic parity. The success of simulation learning depends on the level of similarity between the simulation and the eventual real-world operational environment an ML system will be deployed in. For example, training an ML agent to fly a drone in a virtual environment without wind may lead to poor performing systems in actual flight if windy conditions exist. Simulation learning is particularly applicable to environments where many of the underlying principles are understood. Training robots to move through physical space and complete tasks like grasping demonstrates some of the most impressive applications of simulation learning because these simulations approximate realworld conditions. Simulation learning may prove less effective in creating models to accomplish tasks in domains where the underlying rules are less well understood or less conducive to quantification. Developing accurate proxy simulations for the expression of human emotions or group behavior, for instance, may prove challenging.  43  In these conditions, it may be challenging to train as successfully models on virtual data as ones trained on real-world data. To that end, investments to accelerate progress in these data reduction technologies may enable democracies to more effectively compete at parity with authoritarian regimes, but not categorically across all potential applications. Instead, a narrow parity will be achieved under the specific circumstances where these techniques significantly change the data needed to achieve a high level of performance. Moreover, realizing the full benefits of advancing these data-minimization techniques may require more than simply accelerating basic research. One common theme among many \"few shot\" simulation learning and self-play techniques is that they reduce the need for collecting real-world data while simultaneously increasing the need for computational power. Rather than a physical robot navigating the world and collecting data from numerous trials, simulation learning generates this data from the actions of a virtual robot in virtual space. Self-play generates training data through the interaction of an agent with itself, rather than with some outside environment. Simulation learning and self-play methods free the designer of an ML system from the cost and complexity of real-world data acquisition in exchange for running simulations or contests between agents. These tasks all require access to high-performance computational power. Democratic societies aiming to reduce the level of real-world data required to create competitive ML systems must ensure access to high-performance computational power. Universities, companies, and other actors advancing the state of the art in this domain will require plentiful and affordable computational power to make progress on these research problems and in the training and deployment of actual systems. This may require securing affordable access to corporate clouds, which can provide computational infrastructure as a service, and a greater investment in creating secure computational infrastructure for training models on certain applications. \n ENSURING DEMOCRATIC VIABILITY Once ML systems are built, democratic commitments to public consent and civic participation require a core set of concerns to be addressed before the technology can be deployed in the field. Although ML has enabled computers to accomplish an impressive range of tasks that they previously were unable to do, the technology still falls short in several important respects. For one, many modern ML systems lack what is referred to in the field as interpretability. Researchers have a somewhat limited understanding of how and why these systems produce the outputs that they do.  44  While it is clear, for instance, that a computer vision system can successfully recognize an object like a cat in an image, the process by which it arrives at this outcome is not always so clear.  45  Even more challenging, in some cases, attempts to make models more explainable will also reduce their performance.  46  Second, ML systems are prone to learning spurious correlations during the training process. Such correlations cause them to exhibit \"algorithmic bias,\" producing discriminatory and inequitable results when applied to people. For instance, in 2018, Amazon discovered an ML model the company used for filtering through resumes to identify promising job candidates discriminated against women.  47  The model was trained on historical resumes and hiring decisions disproportionately representing men. As a result, the model \"penalized resumes that included the word 'women's' as in 'women's chess club captain,' and downgraded graduates of two all-women's colleges.\"  48  Finally, many ML applications are privacy invasive. Training high-performance ML systems requires large datasets, and that data must be centralized to enable effective training. This tends to place power in the hands of a single entity such as a government or corporation, who might use this data for purposes beyond its original intent. Lack of interpretability, algorithmic bias, and data centralization are all potent reasons for democratic publics to reject the deployment of AI systems. For example, consider how interpretability shapes public debate around the potential use of an ML system designed to predict crime and more efficiently allocate law enforcement resources.  49  Limits around interpretability can create practical issues in explaining how such a system actually comes to predict that crime will happen in a specific location. ML may only have limited tools available to conduct effective audits and identify the areas in which the system might make an error, or when it fails to work at all. Moreover, a lack of understanding around how the system renders decisions can make it challenging to fix when errors emerge. ML is adoptable only with a murky understanding of its inner-workings, reducing trust and making public consensus around the technology harder to build. This limitation hinders support for the use of ML in applications not just in the context of state functions like national defense, but throughout society. Legal requirements mandating a citizen's right to access a decision explanation may also prevent the technology's adoption.  50  At the very least, this opacity continually opens ML systems up to claims of bias, regardless of the actual impact of these systems in practice.  51  Authoritarian states enjoy the relative advantage of being able to mandate the deployment of ML technologies throughout the economy and society in spite of these problems. Democratic states, in their preservation of an independent civil society and commitment to civic participation, have numerous mechanisms for citizens to push back on use of undesired ML technologies. As a result, autocratic states can be \"first to market\" by forcing the deployment of these technologies through society. This advantage may provide economic and social benefits of ML systems earlier than states that must wait for buy-in from a hesitant public. Similar to the challenge of immense datasets, one way to circumvent the limitations of democratic norms is to enact legal and policy reforms that smooth adoption of ML. This might include allowing the state and companies to deploy ML-driven technologies more aggressively without requiring public assent and adoption. But democracies should think twice before doing so: public consent may impose practical costs on the speed of adoption of ML technologies, but these costs are desirable in a broader sense. Like the erosion of privacy protections to make ML systems easier to train, these reforms would compromise democratic values in some degree to enable better competition. An alternative approach is to consider ways in which ML may be made more democratically viable, such that requiring public consent does not hold back the technology. Might ML be reshaped to make it more amenable to public accountability and review? Might ML systems be deployed in ways that can ensure fairness, or avoid the problems of data centralization? Enabling ML to earn public consent and the political legitimacy that consent confers will be critical to ensuring that democracies can benefit from these technological breakthroughs. The approach of forcing AI-driven systems on the public or hiding its deployment risks a backlash that may hinder the implementation of the technology. Questions of political legitimacy might seem entirely beyond the scope of technical ML research. Certainly, achieving a democratically viable form of ML will require non-technical reforms in governance and regulation that offer the public a say in the development, design, and deployment of the technology. These reforms will play a major role in facilitating adoption of ML in democratic societies at large, even if the public chooses at moments to reject particular applications. At the same time, technical details will exert a major influence over the democratic legitimacy of AI. In certain high-stakes arenas like healthcare or government administration, current ML systems may simply not be able to provide the degree of interpretability, fairness, and privacy protection demanded by the public in a democratic society. This poses a dilemma for the adoption of certain applications of AI, as they require either the wholesale rejection of the technology or its acceptance with a host of significant drawbacks. \n Investment Opportunities Democratic adoption will depend on advancing the ML field to meet public demands for interpretability, fairness, and privacy protection. Significant progress in these research areas would supply ways of meeting these concerns that were not previously possible. The end result would be to improve the practice of ML to better win the consent of a democratic public. First, democratic governments should invest in the subfield of ML focused on the problem of interpretability. Broadly speaking, interpretability research seeks to uncover the internal processes governing how ML systems make decisions while retaining a high level of performance. More ambitiously, researchers seek to translate these internal processes into something useable by a non-expert.  52  Expanding the range and sophistication of interpretability techniques would provide the public a better window into the processes by which ML systems automate decision-making. This would improve trust and create oversight options that may speed adoption. Government involvement could take a number of forms. Most obviously, it might take a role in directing funding toward research in interpretability, significantly augmenting existing programs working to advance \"explainable artificial intelligence.\"  53  States could also play an important role in spearheading public challenges that raise the profile of certain key problems and incentivize targeted work on them. Second, democratic governments should invest in advancing the subfield of fairness in ML, which seeks in part to find technical means of ensuring that AI systems can be designed to avoid the discriminatory and inequitable behaviors that may emerge during the training process.  54  For example, an ML system used to make credit and lending decisions may learn spurious correlations from the historical data the model is trained on.  55  The resulting model may systematically disfavor borrowers from certain minority groups based on racially discriminatory patterns of previous lending. Researchers have developed a range of algorithmic definitions of fairness that can be subsequently implemented into ML models to avoid these kinds of discriminatory behaviors.  56  ML fairness research has also explored the application of a family of techniques known as causal modeling, which seek to extract causal relationships from observational data. Rather than simply predict, say, the recidivism of an offender in the criminal justice system, causal systems may suggest interventions to reduce the rate of crime going forwards-an arrangement more amenable to democratic publics and consistent with protecting the rights of the individual.  57  Finally, democratic societies may also demand that the benefits of ML be obtained without undue harm to privacy. This might be made possible by reducing the overall data required for training through advancing the \"few shot,\" simulation-based, and self-play learning methods previously described. But democracies might also secure privacy by advancing the state of the art of privacy preserving ML, which seeks to train systems without the need to directly access the raw data itself. A few promising techniques are emerging in the research field for accomplishing this task. Researchers have been exploring a group of techniques known as homomorphic encryption that would enable ML systems to be effectively trained even on fully encrypted data.  58  This method might make democratic publics more willing to consent to their data being used, secure in the knowledge that their personal information remains inaccessible to those training models on that data. More radically, the subfield of research into federated learning seeks to effectively train ML systems on data that is widely dispersed across many devices and anonymized through a method known as differential privacy.  59  Breakthroughs in this research may make ML more acceptable to democracies by unlocking different architectures for the technology which do not require centralizing massive corpora of data in a single location, regardless of whether or not it is encrypted. Such techniques enhance the democratic viability of ML by providing the public and policymakers with an array of concrete options to ensure these technologies reflect democratic values. \n Strategic Implications: Democratically Viable Artificial Intelligence Resources are flowing to the research areas discussed above, driven by a mixture of commercial interest, public pressure, and researcher priorities. Interest in applying ML in high-stakes arenas has expanded the need to create more interpretable systems.  60  For instance, ML use in medicine promises to provide an early warning system to doctors about life-threatening conditions like kidney failure.  61  However, medical professionals have hesitated to adopt these systems in practice due to their \"black box\" nature, prompting companies to invest in interpretability research.  62  High-profile failures of ML systems that highlight their tendency to reproduce bias have also encouraged companies to invest in creating more diverse datasets and to support research on fairness in the technology. In one recent incident, IBM received significant public criticism for its ML-based facial recognition system, which was found to significantly underperform on faces with darker skin tones.  63  This prompted IBM to release a \"Diversity in Faces\" (DiF) dataset in 2019.  64  DiF provided a \"more balanced distribution and broader coverage of facial images compared to previous datasets,\" enabling ML researchers and engineers to address these problems in building future facial recognition technologies. Despite this no organized attempt has emerged to marshal work on the problems of interpretability, fairness, and privacy toward a broader geopolitical objective. This vacuum provides an opportunity for democratic states to build alliances and lead.  65  States can play a role in articulating a unified vision of \"democratically viable\" AI and funding work in support of this research agenda, helping organize otherwise disparate efforts among researchers and encouraging greater collaboration. Companies may be particularly keen to engage in such an agenda, as it creates an outlet to engage with the government on ML development while avoiding more controversial work in the national security and military domains. Democracies also could play a role in fostering the development of training materials, which help to more rapidly percolate know-how about these techniques throughout the research and engineering communities. Significant advances in these areas would help level the playing field with authoritarian regimes in the multi-faceted competition around AI. These tools would aid in the creation of high-performance ML systems with the capability to meet demands for transparency, accountability, and protection of social values. Citizens and businesses are more likely to trust these systems and consider them legitimate, enabling speedier integration of the technology and acquisition of its benefits by democratic societies. Beyond the direct benefit of enabling easier ML adoption, fostering a democratically viable approach to AI offer collateral strategic impacts from a geopolitical perspective. Importantly, this strategy may raise expectations of publics globally that ML will be interpretable, fair, and privacy protecting. Much of the competition around AI will be for market share: private companies based in various countries will compete aggressively to acquire users for their products and services both domestically and abroad. Widespread awareness about technical breakthroughs in subfields like interpretability, for instance, could cause consumers to demand that ML-driven products and services incorporate certain transparency features as a default. This is especially the case if these advances overcome the long-standing sacrifice of system performance required to create systems with higher levels of interpretability. Companies in markets that expect features like interpretability might enjoy a competitive edge over those in markets without the same demands. In this sense, advancing techniques for interpretability, fairness, and privacy might raise the barriers to entry into the markets of democracies, creating a kind of bulwark that favors home-grown products and services. Moreover, large markets in democratic states with strong preferences for \"democratically viable\" ML products might force companies and states to invest more in these features than they otherwise would to access these users. Even markets with weak commitments to democratic values may prefer to work with foreign companies that can deliver transparent AI, as these systems may be easier to monitor and trust. The result may impose a cost on authoritarian states-who otherwise do not benefit strongly from advancements in these techniques-while creating know-how beneficial to the accelerated adoption of AI in democratic societies. Liberal democracies face an unpalatable choice should they fail to invest in research on interpretability, fairness, and privacy. On one hand, democracies may need to slow technological adoption, delaying the economic and social gains that might otherwise be obtained through ML. This approach may put them behind authoritarian competitors that can more speedily force deployment and reap the benefit. On the other, democracies can proceed without regard for public consent, sparking significant resistance from citizens and sacrificing their values in the process. \n SUBVERTING SOCIAL CONTROL APPLICATIONS Recent breakthroughs in ML may be particularly attractive to authoritarian regimes in part because many applications seem well suited for expanding surveillance and enabling systems of social control. Democracies should invest in technologies that will hinder and thwart the use of ML for these purposes. AI is well suited for applications that suppress dissent and sustain autocratic regimes. Advances in computer vision enable ever better tracking of individuals and relationships through surveillance footage.  66  Algorithms that predict social behavior may eventually work to enhance the effectiveness of \"social credit\" arrangements incentivizing and disincentivizing the public in ways that align with a regime's motives.  67  Natural language processing models can be leveraged to streamline and enhance regimes of censorship.  68  Even when these techniques replicate existing tools of surveillance and punishment, ML may efficiently automate and otherwise lower the costs of effectively implementing these systems of control. These high-tech applications will soon \"trickle down\" to less technologically sophisticated autocrats. Leading companies in the deployment of ML for domestic social control purposes are beginning to market their services to aligned regimes around the world.  69  These tools are becoming commodified, lowering the financial resources and technical expertise needed to wield them. Technologies for social control are critical to authoritarians, as such systems help neutralize dissent and enforce their policies. Autocrats will prioritize investments in developing and deploying ML in part because the technology promises to significantly enhance social control, thereby promoting regime stability and longevity. In contrast, democracies may not have such clear and urgent needs to advance AI, making it more difficult for them to keep pace over time. How might democracies erode the strong authoritarian incentives to prioritize investment in ML? Democratic states may have an opportunity to go on the offensive by undermining the effectiveness of ML systems when authoritarian regimes deploy them. ML enhances states' ability to exert control over their populations, but ML research can simultaneously play a role in the creation of techniques that subvert these applications and make them more difficult to use effectively. By giving populations knowledge about and tools for defeating the social control applications of AI, democratic societies could work to raise the floor of resistance to these uses globally. If successful, surveillance and social control uses would prove not just ineffectual but risky for a regime because citizens could easily exploit the vulnerabilities of the technology to evasion and manipulation. Eroding the usefulness of AI for social control purposes would counter an important benefit the technology confers upon autocratic regimes. Accordingly, this strategy would work to reduce the attractiveness for authoritarian regimes to invest monetary and human capital in advancing the technology. These dynamics would aid democratic societies in maintaining their lead in the technology and in disproportionately benefiting from its development. While the economic opportunities ML offers mean that it is unlikely that these regimes would entirely halt their investments, this strategy may work to significantly suppress investment. In the very least, authoritarian regimes might be compelled to allocate a greater portion of their resources toward ensuring that their systems are sufficiently resilient against attacks and advancing research on the topic. This would impose a cost they might not otherwise face, and potentially advance the field by improving the safety and security of the technology generally. \n Investment Opportunities Ongoing research has revealed the extent to which various ML systems are vulnerable to attacks causing them to produce faulty or otherwise undesired outputs. Perhaps the most dramatic and widely reported example of this vulnerability has been the phenomenon of \"adversarial examples.\" These inputs, when fed to an ML system, cause it to render inaccurate outputs. In one prominent example, a computer vision system was fooled into recognizing a panda as a gibbon.  70  These inputs can look innocuous to a trained eye since the attacks involve changing only a few pixels in an image, raising the possibility that these kinds of manipulations might happen under the noses of human operators. These attacks also take place elsewhere in the \"lifecycle\" of an ML system. For example, researchers have highlighted the problem of \"data poisoning,\" in which the data leveraged for training a system introduces systematic biases subsequently exploitable by an attacker.  71  Adversarial examples have also been used to manipulate the tools researchers and engineers rely on to diagnose issues and examine the internal functioning of ML systems, potentially allowing attackers to hamper attempts to fix faulty systems in the future.  72  This research presents a tantalizing opportunity for democracies to undermine the social control applications authoritarians stand to gain from leveraging ML. A surveillance system designed to analyze video footage for political dissidents might be tricked into missing a person of interest, or erroneously identifying regime loyalists as malefactors. Adversarial examples might make automated censorship systems highly \"leaky,\" allowing prohibited news stories and political expression to flow through. Democracies should invest in accelerating the discovery of such vulnerabilities to erode or eliminate the usefulness of this technology in exerting social control. Researchers have been working on a variety of approaches to fix or mitigate the risk from these potential vulnerabilities as they have been exposed. The most prominent approach involves a technique known as adversarial training, which incorporates adversarial inputs into the training process to enable the system to successfully overcome these attacks.  73  Other approaches include \"feature squeezing,\" a set of processes applied to suspected adversarial inputs that can help identify the use of these tactics.  74  Still, no categorical solution to the problem of adversarial examples appears to be on the horizon. Authoritarian regimes could not \"patch\" their systems to create resiliency against adversarial examples. As the authors of one survey article summarized the current state of the art, a \"theoretical model of the adversarial example crafting process is very difficult to construct…[it is hard] to make any theoretical argument that a particular defense will rule out a set of adversarial examples.\"  75  Similar to developments in the realm of computer security, the cat-and-mouse game of discovering vulnerabilities and then repairing them will likely continue for the foreseeable future, if not indefinitely. \n Strategic Implications: Managing Unintended Consequences What specific role might the government take in attacking the effectiveness of social control applications of ML? How might democracies avoid having these methods of attack harm desirable uses of AI? There is a great deal of ongoing research in this space. Companies deploying new products and services driven by AI have strong interests in securing these technologies from manipulation by malicious actors. Industrial labs accordingly have invested in recruiting top talent around these challenges, and the technical community at large has rallied around the topic. The number of prominent workshops and papers presented at major technical conferences testifies to the prioritization of this topic in the collective research agenda.  76  In this respect, it is unclear if democratic states can play a strong role in significantly accelerating basic research activity in the space. Considerable energy on many of these topics within the ML field already exists. As a result, this may be an arena in which no serious underinvestment exists-government agencies, industrial labs, and academic institutions readily prioritize research on these topics. Vulnerabilities in these systems will be exposed rapidly, particularly as ML systems are increasingly trialed in high-stakes domains. But this does not necessarily mean the state has no role to play. Quite the opposite, there may be a few arenas in which democratic governments might make a major impact in mobilizing the burgeoning field of ML security toward subverting the social control applications of the technology. The primary challenge is that most vulnerabilities can be used to undermine both desirable and undesirable uses of ML. The techniques for subverting a computer vision system designed to suppress political dissidents can be identical to those recognizing cancerous cells. The question, then, is whether it is possible to limit the unintended negative consequences. Democracies will benefit only insofar as they can ensure that the distribution of vulnerabilities hinders social control applications while mitigating the negative impact on other applications. Democratic governments can work to do this in two ways. First, governments might simply work to highlight discoveries in this subdomain of security research, using their platform to bring attention to the fragile state of many of the ML systems that could be used to exert social control. This low-cost information campaign may erode the perceived value for autocratic regimes of developing and deploying these systems for surveillance and other purposes. Highlighting high-profile failures and the existence of persistent vulnerabilities might reduce regimes' trust in vendors attempting to sell them AI solutions. In the best pos-sible case, the host of potential vulnerabilities may convince regimes to retain their higher-cost, legacy security infrastructure rather than bet on an untested technology. Second, democratic governments might invest in shepherding technical discoveries about ML vulnerabilities into practical software that citizens could subsequently turn on oppressive systems of AI. While certain technical vulnerabilities might be demonstrated in a research setting, they may not be easily leveraged by populations likely to be subject to these technologies in autocratic regimes. The knowledge about these vulnerabilities remains largely relegated to specialized conferences or contained in dry technical literature that is inaccessible to the non-expert reader. There are also relatively low incentives for researchers and consumer web companies leading in ML to \"productize\" these vulnerabilities into user-friendly applications. The result is that those who are potentially most affected by ML enhanced surveillance or control cannot practically access the discoveries within the research community. This gap may provide a means by which the impact of these discoveries may be more effectively targeted against specific social control uses of the technology. For example, democratic states might adapt research about adversarial examples into freely available clothing patterns that defeat gait and facial recognition systems, eroding the effectiveness of ML enhanced surveillance.  77  Similarly, democracies might invest in creating easy-to-use software that allows citizens to spread faulty data about themselves online, making it more challenging to create ML systems that accurately predict their behavior. Beyond the creation of software, spreading know-how and building connections globally may be equally important. Democratic governments might set up knowledge sharing and similar programs to spread awareness about these techniques and build ties between the ML research community and activists resisting regimes on the ground. The coordination of activists to specifically thwart ML-driven surveillance in recent protests in Hong Kong suggests both awareness and demand for these practices among grassroots groups.  78  By shaping global perceptions around the reliability of the technology for surveillance and control, and helping to translate these discoveries from the lab to actual use, democratic societies can take the lead in influencing the outcome of basic research in ML security as it diffuses more widely. This type of selective involvement might help to mitigate some of the double-edged nature of this research, and places governments in a position to complement existing activity in the research field. ow can democracies effectively compete against authoritarian regimes in the development and deployment of AI systems? The present state of ML imposes limitations on the ability of liberal democracies to do so effectively. Restrictions limiting the collection of training data, commitments to an independent civil society and press, and requirements for public consent all constrain how quickly ML systems can be trained and deployed. Authoritarian competitors do not face such constraints, and have strong incentives to prioritize ML investment in order to enhance systems of social control. These factors give authoritarians an edge over democracies in AI competition. To keep up, democracies need to find a way to compete on more even footing without slowing technological adoption or sacrificing their values. This paper offers one way through this dilemma. Democratic societies should work to achieve technical breakthroughs that mitigate the structural disadvantages that they face, while attacking the benefits that authoritarian regimes stand to gain with ML. This includes investments in three domains: • Reducing a dependence on data. Authoritarian regimes may have structural advantages in marshaling data for ML applications when compared to their liberal democratic adversaries. To ensure better competitive parity, democracies should invest in techniques that reduce the scale of real-world data needed for training effective ML systems. • Fostering techniques that support democratic legitimacy. Democracies may face greater friction in deploying ML systems \n Conclusion \n H relative to authoritarian regimes due to their commitments to public participation, accountability, and rights. This requires them to invest in subfields of ML that enhance the viability of the technology in a democratic society, including work in interpretability, fairness, privacy, and causality. • Challenging the social control uses of ML. Recent advances in AI appeal to authoritarian regimes in part because they promise to enhance surveillance and other mechanisms of control. Democracies should advance research eroding the usefulness of these applications, seeking to suppress investment by autocratic regimes and encourage research on robustness. Targeted investments in these technical areas will work to level the playing field between democracies and their authoritarian competitors. Proposals to simply increase funding, expand the pool of talent, or free up datasets are laudable, but apply resources to AI as a broad, undifferentiated category. Without a close examination of the technical trends within ML, these strategies will expend significant resources ineffectually, redundantly, or most likely, in a diffuse way that fails to make any real difference from a geopolitical standpoint. In shaping the terrain of AI competition, democracies may be able to achieve the best of both worlds: a future where democratic societies are able to fully obtain the social and economic benefits of AI while simultaneously preserving their core values in a rapidly changing technological landscape.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Shaping-the-Terrain-of-AI-Competition.tei.xml", "id": "6627e24a8bb7ab1233c8564c8eca69c8"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Thanks to Tim Blomfield and Nicholas Emery for contributions to the literature summaries and organization.", "authors": ["Philip Trammell", "Anton Korinek"], "title": "Economic growth under transformative AI", "text": "Introduction At least since Herbert Simon's 1960 prediction that artificial intelligence would soon replace all human labor, many economists have understood that there is a possibility that sooner or later artificial intelligence (AI) will dramatically transform the global economy. AI could have a transformative impact on a wide variety of domains; indeed, it could transform market structure, the value of education, the geopolitical balance of power, and practically anything else. We will focus on three of the clearest and best-studied classes of potential transformations in economics: the potential impacts on output growth, on wage growth, and on the labor share, i.e. the share of output paid as wages. On all counts we will focus on long-run impacts rather than transition dynamics. Instead of attempting to predict the future, our focus will be on surveying the vast range of possibilities identified in the economics literature. The potential impact of AI on the growth rate of output could take the form of • a decrease to the growth rate, even perhaps rendering it negative; • a permanent increase to the growth rate, as the Industrial Revolution increased the global growth rate from near zero to something over two percent per year; • a continuous acceleration in growth, with the growth rate growing unboundedly as time tends to infinity; or even • an acceleration in the growth rate rapid enough to produce infinite output in finite time. Basic physics suggests that the last of these scenarios is impossible, of course-as is eternal exponential or super-exponential growth. The relevant possibility is that AI induces a growth path that resembles these benchmarks for some time, until production confronts limiting factors, such as land or energy, that have never or have not recently been important constraints on growth. We will thus also consider how increases in growth may eventually be choked off by such limiting factors. The potential impact of AI on the labor market includes predictions that • wages fall; • wages rise, but less quickly than output, producing a declining labor share; • wages rise in line with output, producing a constant labor share; or even • wages rise more quickly than output, producing a rising labor share. In the respective limiting cases, AI could result in a future in which wages are (literally or asymptotically) zero or near zero; very high in absolute terms but approximately zero percent of total output; or high both in absolute and in relative terms. As this discussion illustrates, the space of possibilities is vast. At the same time, as Simon's failed prediction testifies, transformational impacts from AI are by no means certain to transpire on any particular time horizon. Most studies to date of the economics of AI, therefore, have focused on the most immediate, moderate, and likely impacts of AI. 1 These include marginal shifts in output and factor shares; impacts on marketing and statistical discrimination; impacts on regional inequality; and industry-specific forecasts of AI-induced growth and labor displacement over the next few decades. Empirical estimates of the future economic importance of AI have drawn inferences from foreseeable industry-specific applications of AI, from totals spent on AI R&D, and from comparisons between AI and past technological developments in computing, internet connectivity, or related fields (see e.g.  Chen et al. (2016) ). Industrial and R&D-based inferences may however severely underestimate the field's transformativeness if technological development has substantial external effects, as is often assumed; the importance of the atomic bomb is not well approximated by the cost of the Manhattan project. Even in the absence of externalities, furthermore, inferences from R&D expenditures implicitly discount future output according to the time preference of the research funders, who may assign negligible value to the impact of their technologies on the distant future. And common referenceclass-based projections preclude the possibility that AI ultimately proves to be truly transformational, less like the economic impact of broadband than 1 Summaries of the economic implications of AI have been put out by most major consulting firms and by several governments and academic institutes. A proper review of these reviews would require a document in its own right, but for comparison, the review that appears to go furthest in discussing AI's transformative possibilities is that from Accenture  (Purdy and Daugherty, 2016) . The most radical scenario the authors consider is one in which AI comes to serve as a \"new factor of production\" complementing both labor and capital. They forecast that, in this scenario, the result will be what they call \"a transformative effect on growth\", by which they mean a doubling of growth rates in developed countries up to 2035. like that of the Industrial Revolution or the evolution of the human species itself. In recent years, economists have begun to engage earnestly in formal theoretical explorations of a wide array of the transformative possibilities of AI, including those outlined above. We aim to summarize the findings of these explorations.  2  In the process, we hope not only to state the conclusions of various models, but also to give the reader some mathematical intuition for the most important mechanisms at play. (Indeed we have altered some of the models slightly, to clarify their implications regarding transformativeness and their relationships to the other models discussed here.) This document is intended for anyone comfortable with a moderate amount of mathematical notation and interested in understanding the channels through which AI could have a transformative impact on wages and growth. Readers with backgrounds in economics will hopefully come to better understand the possibilities which concern singularitarians, and readers with singularitarian backgrounds will hopefully come to better understand the relevant tools and insights of economics. The rest of this document proceeds as follows. §2 consists of an overview of the economics that will be relevant for understanding the subsequent sections. §3 discusses models in which AI is added to a standard production function. §4 discusses models in which AI is added to a \"task-based\" production function. §3 and §4 both implicitly take place in a setting of exogenous productivity growth; §5 discusses models in which productivity growth is endogenous and AI can feature in its production. §6 compares the results found in §3-5. Finally, §7 concludes. \n Economics background Literature on the economics of AI extends prior literature on the economics of production and growth. Those without backgrounds in economics may therefore find it helpful to review the latter briefly. Those with backgrounds in economics may want to skim the section to familiarize themselves with the notation and terminology we introduce. \n Production and factor shares Throughout this review, we say that output at a time t is determined by a production function F and the list of input (or \"factor\") quantities available at t. If for simplicity we categorize all production factors as either labor L or capital K, and think of all output as a single good Y , we can write Y t = F (K t , L t ). We will temporarily abandon the time subscripts and explore production in a static setting; they will return when we explore growth. The production function F is assumed to be continuously differentiable, increasing, and concave in each argument. It is also assumed to exhibit constant returns to scale (CRS); doubling the earth, with all its labor and all its capital, would presumably double output. Finally, its inputs are assumed to be complements: the marginal productivity of each is increasing in the supply of the other.  3  There can be \"capital-augmenting\" technology, denoted A, and \"laboraugmenting\" technology, denoted B. Increases in some factor-augmenting technology make the use of the given factor more efficient, so that we can proceed as if we had more of it. With technology, that is, our production function takes the form Y = F (AK, BL). Note that, by CRS, a technological advance that multiplies both A and B by a given factor will multiply output by this factor as well. The \"marginal product\" of a factor is the derivative of output with respect to that factor. Both inputs are assumed to be paid their marginal products. That is, the wage rate is F L (AK, BL), and the capital rental rate is F K (AK, BL). This makes sense if we imagine that production is taking place in a competitive market, with lots of identical firms facing this production function. If a factor employed at one firm is not being paid its marginal product, another firm will offer to pay more for it. These factor payments will equal total output-i.e. KF K (AK, BL) + LF L (AL, BL) = F (AK, BL) (1) -whenever F is CRS, by Euler's Homogeneous Function Theorem. If output exceeded the sum of factor payments, we would have to explain what was done with the output not paid as wages or rents; and if factor payments exceeded output, we would have to explain how the deficit was filled. The fraction of output paid out as wages-i.e. LF L (AK, BL)/F (AK, BL) -is termed the \"labor share\". Likewise, the \"capital share\" is the fraction of output that accrues to the owners of capital. We will assume through most this document that everyone is employed. In reality, of course, many cannot work, or will not work if the wage rate is sufficiently low. Model scenarios in which the wage rate falls dramatically may therefore be more accurately interpreted as scenarios in which unemployment is widespread. For our purposes this will not be an important distinction. \n Substitution Given a production (or utility) function and a list of factor (or consumption good) prices, suppose a purchaser spends a fixed budget so as to maximize output (or utility). The elasticity of substitution between two factors X 1 and X 2 is, intuitively, the value such that, if the relative price of X 1 falls by a small proportion (say, 1%), the relative quantity of X 1 purchased will rise by an -times larger proportion ( %). Conversely, then, given a list of factor quantities, the elasticity of substitution is the value such that, in order for the relative quantity of X 1 sold to increase by a small proportion (say 1%), the relative price of X 1 must fall by a (1/ )-times larger proportion ((1/ )%). Suppose that goods X 1 and X 2 are divided among many owners, and suppose there are many interested purchasers. (The sets of owners and potential purchasers need not be disjoint, but it may be easier to imagine that they are.) The price of X 1 should then equal the market-clearing price: the price such that, given the price of X 2 , precisely the entire supply of X 1 is purchased. If the price of X 1 were higher than this, anyone who owns some of it would do well to sell her unsold goods at just less than the market price. If it were lower, i.e. if purchasers would have been willing to buy more than the entire supply of X 1 at the given price, then anyone who owns some X 1 could charge more than the given price and still sell all she has. Now consider the consequences of exogenously increasing the relative quantity of X 1 by 1%, say by increasing its absolute quantity by 1% across each of its owners and leaving the quantity of X 2 unchanged. The marketclearing relative price of X 1 will then fall by (1/ )%. As we can see, marginally increasing the relative abundance of a good results in smaller relative expenditure on that good-i.e. its owners receive a smaller share of total income-precisely when the elasticity of substitution between it and other goods, on the current margin, is less than 1. For illustration: food and other goods are not very substitutable. When food was much scarcer, its owners were able to command such higher prices for it that people spent larger shares of their incomes on it. On the other hand, industrially produced goods and handmade goods are very substitutable. As the former grew more plentiful, following the Industrial Revolution's explosion in manufacturing, people spent larger shares of their incomes on them. Goods are perfect complements if = 0. In this case, output (or utility) is some constant times the minimum of the goods' quantities, in some ratio. (Consider left shoes and right shoes, in the 1:1 ratio, or bicycle frames and wheels, in the 1:2 ratio.) However the goods' relative prices change, the relative quantities purchased will stay fixed at the given ratio. The case of perfect substitutability is approached in the limit as → ∞. In this case a positive quantity of each good is only purchased if their prices are equal; if their prices differ, only the cheaper is purchased. For ease of notation, let us define the \"substitution parameter\" ρ ( − 1)/ . Note that the cases < 1, = 1, and > 1 correspond to ρ < 0, ρ = 0, and ρ > 0 respectively. A production function exhibits constant elasticity of substitution (CES) if its elasticity of substitution does not depend on the factor prices and quantities. A two-factor CES production function that is also CRS, with factoraugmenting technology, must take the form Y = [(AK) ρ + (BL) ρ ] 1/ρ (3) if ρ = 0 and Y = (AK) a (BL) 1−a , (4) for some a ∈ (0, 1), if ρ = 0. In the second case, the function is called \"Cobb-Douglas\". When ρ ≤ 0, output requires strictly positive quantities of both factors. If we have more than two factors i, natural extensions of the above functions look similar, with the Cobb-Douglas exponents a i still summing to 1. When ρ = 0, the share of output paid to factor X, with factor-augmenting technology C, equals (CX/Y ) ρ . When ρ = 0, a factor's share equals the exponent on that factor. In general, the share of factor X is decreasing in CX/Y when ρ < 0, independent of CX/Y when ρ = 0, and increasing in CX/Y when ρ > 0. Naturally a CES and CRS utility function, defined over consumption goods, has the same properties. The notion that a utility function could be CRS may sound absurd, since there is a natural sense in which marginal utility decreases in consumption. But if we are only interested in modeling how an agent with a CES utility function will apportion her spending at a given time, and if we are considering scenarios in which she has no uncertainty, we can without loss of generality invoke a utility function that is CRS as well. This is because, for any utility function in which returns to scale are not constant, there is a monotonic transformation of it-a new utility function ranking all \"consumption bundles\" (lists of good quantities) in the same order-which is CRS. Observe that there is a deep similarity between cases (a) in which there is a single consumption good, but multiple factors to its production, and (b) in which consumer utility is defined over multiple consumption goods, each of which employs a single production factor. Consumers in the latter cases function like factories in the former cases, as if consumer goods were inputs to the production of utility. \n Exogenous growth In practice, of course, the substitution parameter between labor and capital may not be constant. It may be high when labor is more abundant than capital and low otherwise, or vice-versa, for example. It may also change over time, for reasons independent of factor quantities. That said, it has been estimated in a variety of contexts and times to be substantially negative (see e.g.  Oberfield and Raval (2014)  or  Chirinko and Mallick (2017) ). When exploring preliminary hypotheses about growth, factor shares, and other macroeconomic variables, therefore, it can be helpful to start with a model in which production is CES (and CRS) with ρ < 0. Output per person in the developed world has grown substantially over the past few centuries, following a roughly exponential trajectory. In a two-factor production function without technology growth, the only possible explanation for this would be the capital stock growing more quickly than the population. Capital accumulation cannot be the primary force driving longrun growth, however, for at least two reasons. First: if ρ = 0, capital accumulation can produce unbounded output, albeit at a growth rate that slows to zero; and if ρ > 0 (or if production is not CES but the substitution parameter is permanently bounded above zero), capital accumulation can in principle sustain a positive growth rate. 4 But if ρ < 0 (constant or bounded below), as the stock of capital per unit of labor grows, capital's marginal productivity falls to the point that output growth slows to a halt. That is, a lack of labor per unit of capital definitively constrains the growth of output per person. As we can see from the first equation above, when ρ < 0, the capital term tends to zero as the quantity of capital increases. In the limit, then, if capital is far more plentiful than labor, output tends to BL, and output per person tends to B. Second, historically, the capital share throughout the developed world has been roughly constant (at about 1/3). As noted in §2.2, however, if ρ < 0, an unboundedly increasing stock of capital per unit of labor should decrease the capital share to zero. Capital-augmenting technology growth just increases the effective capital stock, so if ρ < 0 it cannot produce long-run output growth either, for the same reasons. 5 The way to get long-run per-capita output growth in this framework, as shown by  Uzawa (1961) , is to introduce labor-augmenting technology growth. To illustrate this, suppose for simplicity that A is fixed, that B grows at some constant exponential rate g B , and that a constant proportion s of output is saved as capital each period. 6 If s is high enough that capital accumulation can keep up with the growing effective labor force, and if the labor supply is also constant at L, the result is a growth path in which output Y t , capital K t , and \"effective labor\" B t L all grow at rate g B . (Recall that, by CRS, equal proportional increases to K t and B t L will produce an equal proportional increase to Y .) If s is too small, output will be constrained by capital accumulation. (This is clearest when s = 0.) In this case, Y t eventually approximately equals AK t . Then K t+1 = K t + sY t ≈ K t + sAK t (5) (using discrete-time notation for clarity), so capital and output both grow at asymptotic rate sA. 7 The requirement that capital accumulation keep up with the growing effective labor force is thus the requirement that sA ≥ g B . We will call this condition \"sufficient saving\". Note that given any fixed g A > 0 and g B , the sufficient saving condition will eventually be met. Letting the labor force grow at some positive rate g L makes no interesting difference. In this case, so long as sA ≥ g B + g L , B t L t grows at rate g B + g L , and Y t and K t do likewise. Regardless of population growth, factor shares are constant over time, 8 so wages and capital rents per person grow at rate g B . The empirical causes of technology growth remain highly uncertain. An exogenous growth model is one that does not attempt to model these causes, but simply takes constant exponential growth in B for granted. \n Endogenous growth On an endogenous growth account, on the other hand, growth in B is modeled as the output of some deliberate effort, such as technological research. That is, technology, like final output, is generated from inputs such as labor, capital, and the stock of existing technology. In the most commonly  Kaldor (1957)  and many since. There are relatively plausible ways to microfound this phenomenon, if we wish to develop our model in more detail, and we will touch on some of these in §3.5. For most of this document, however, we will simply take a constant saving rate for granted. 7 Models without labor, in which Y t = AK t , are termed \"AK models\". An AK economy is of course only ever constrained by capital accumulation and exhibits growth at rate sA. 8 It follows from the identity sY t = K t+1 − K t that sY t /K t = g K,t . Since in the long run a constant saving rate maintains g Y = g K , the long-run capital share (AK/Y ) ρ equals (sA/g Y ) ρ . If there is insufficient saving (so that g Y = sA < g B ), or in the edge case of sA = g B = g Y , the capital share tends to 1. used research-based growth model, that of  Jones (1995) , the growth of B in absolute terms is given by Ḃt = B φ t (S t L t ) λ (6) for λ > 0 and unrestricted values of φ, where S t denotes the fraction of the labor force working as researchers (or \"scientists\") at t. Intuitively, λ > 1 corresponds to cases in which researchers complement each other, and λ < 1 corresponds to cases in which some sort of \"duplicated work\" or \"stepping on toes\" effect predominates. Output is given by Y t = F (K t , B t (1 − S t )L t ), as before. In this setting, though output is CRS with respect to capital and effective labor at any given time, it exhibits increasing returns to scale in population across time. We therefore cannot continue to assume that all inputs to production are paid their marginal products. In particular, we cannot assume that technological innovators are compensated for all the additional future output that their research produces on the margin; this sum of marginal products would exceed total output! (Indeed,  Nordhaus (2004)  estimates that innovative firms accrue on average only about 2% of the value they produce.) It would be beyond the scope of this section to summarize theories regarding the empirical or optimal number of researchers, or the empirical or optimal level of worker pay. For now, to introduce endogenous technological development without having to consider its interactions with the rest of the framework, we might simply assume that a government sets S t and pays S t L t workers to do research. Their wages are equal to non-research workers' wages, in this stylization, and they are financed by lump-sum taxes levied equally across the population. To maintain a constant rate of output growth, we must maintain a constant rate of labor productivity growth, by the reasoning laid out in §2.3. However, the growth rate of B at t, denoted g B,t , is by definition equal to Ḃt /B t . Holding L and S fixed, therefore, we now have g B,t = B φ−1 t (BL) λ . ( 7 ) If φ < 1, as we can see, g B,t falls to 0 as B t grows. If φ > 1, g B,t rises to infinity. Only in the knife-edge case of φ = 1 do we get exponential growth with a constant number of researchers. 9 As a matter of fact, the number of researchers has grown dramatically over the past few centuries. Both the population and the fraction of the population working in research have grown. For simplicity, and because the number of researchers cannot grow indefinitely in a fixed population, let us here ignore the second trend and suppose that S is fixed, with L t growing at a constant rate g L . In this case labor productivity growth g B is constant over time iff B φ−1 t (SL t ) λ = B φ−1 0 exp(g B (φ − 1)t)S λ L λ 0 exp(g L λt) (8) is constant over time; that is, iff the change in the number of researchers just offsets the change in the difficulty of producing proportional productivity increases. This in turn will obtain iff (φ − 1)g B + λg L = 0, or 10 g B = λg L 1 − φ . (9) Because we are holding 1 − S t fixed, the number of non-research workers will grow at rate g L . So the number of effective non-research workers will grow at rate g B + g L . As we have seen, under a constant saving rate, capital and output will grow at this rate too, and output per person will grow at rate g B . Observe that the steady-state rate of labor productivity growth is here undefined when φ = 1. The calculation also breaks down when φ > 1, absurdly predicting a negative rate. This is because the value assumed to exist in the derivation, namely a steady-state productivity growth rate g B under a growing research workforce, does not exist when φ ≥ 1. When φ = 1, it is straightforward to see that a positive growth rate in the number of researchers produces an increasing rate of labor productivity growth. When φ > 1, even a constant number of researchers produces ever-increasing labor productivity growth as well. The ever-increasing labor productivity growth rate g B,t that follows when φ = 1 and g L > 0 translates into an increasing output growth rate up to the an unexplained process of exponential population growth is needed when φ < 1. Models of this form with φ < 1 are then termed \"semi-endogenous\". 10 It follows from g B,t = B φ−1 t (SL t ) λ that the corresponding growth path is stable. If B t is \"too high\", growth subsequently slows, since φ − 1 < 0. Likewise, if B t is \"too low\", growth accelerates. point that g B,t +g L = sA. At that point capital accumulation cannot keep up with the growth of the effective labor force, and production is constrained by capital. If capital-augmenting technology can be developed in parallel with labor-augmenting technology, however, the two factors can both grow at an increasing rate, and output therefore can as well. That is, we have a Type I growth singularity. When φ > 1, moreover, even a constant number of researchers is enough to produce \"infinite output in finite time\", i.e. a Type II growth singularity. The intuition for this is easiest to grasp when φ = 2 and (SL) λ = 1, so that we have g B,t = B t . Suppose g B,0 is such that technology doubles every time period. Thus B 1 = 2B 0 , so g B,1 = 2g B,0 . At this doubled growth rate, technology doubles every half-period; B 1.5 = 2B 1 . By repeated applications of the same reasoning, the technology level approaches a vertical asymptote at t = 2. If capital-augmenting technology follows a similar process, output approaches a vertical asymptote at t = 2 as well. The potential for endogenous growth processes to produce explosive growth is striking. However, since the researcher population growth rate has long been positive and the productivity growth rate has long been roughly constant, and in fact declining over recent decades  (Gordon, 2016) , we can infer that at least historically φ < 1. Indeed, the most extensive study of the topic to date-that done by  Bloom et al. (2020) -estimates φ = −2.1. An estimate of φ ∈ (0, 1) would indicate that, when we have access to a large stock of existing technologies, these aid in the development of new technologies, but offer diminishing marginal aid. An estimate of φ < 0 implies that when there is a large stock of existing technologies it is harder to develop new technologies-perhaps because so much of the low-hanging technological fruit has already been developed, with this \"fishing out\" effect outweighing the effect of technological assistance in technological development. \n AI in basic models of good production \n Capital productivity in isolation At face value, AI promises to make capital more productive. This would most naturally be modeled in the standard framework as an increase to A, which would amount to effective capital accumulation. As  Acemoglu and Restrepo (2018a)  point out, and as we have seen, this on its own would not be predicted to have very transformative economic effects. It would increase output and wages somewhat. But given ρ < 0 and a fixed or only slowgrowing labor supply, labor is the primary bottleneck to output, and any increases to wages would come ever more from an increase in the labor share rather than an increase in output. Indeed, the only \"transformative\" effect of capital productivity is that, as A → ∞, all else equal, the labor share should rise to 1. This is of course the opposite of the intuitive trend, which is also the observed trend in the labor share in recent decades, especially in the industries that have undergone most automation. The models below, therefore, are all designed to shed light on the consequences of increasing the productivity of capital in combination with various structural changes to the production function that AI might also precipitate. \n Imperfect substitution Nordhaus (2015) explores the transformative possibility of AI in the standard model of good production without adding anything explicit about AI. Instead, he posits that AI changes some of the model's parameters \"behind the scenes\". This process has two steps. First, suppose that AI raises the substitution parameter between labor and capital (or certain kinds of capital, such as computers) so that it is permanently bounded above 0. In this case, capital accumulation is sufficient for exponential output growth, even without population growth or technological development of any kind. For illustration, consider our CES production function with ρ > 0 and technology represented but held fixed, and supposing the saving rate s is constant. If the capital supply grows more quickly than than the labor supply, Y t will eventually approximately equal AK t , and capital and output will accumulate exponentially at rate sA. More generally, if labor-augmenting technology grows exogenously at some rate g B ≥ 0, the output growth rate following the substitutability change shifts from min(sA, g B ) to max(sA, g B ). The substitutability change thus increases the growth rate as long as sA > g B . Second, suppose that A t grows exogenously at rate g A > 0. It does not matter whether this technology growth is due to AI or to forces that predated (but were less relevant before) the substitutability change. The growth rate of output will then tend to sA t , which, with A t , will itself be growing exponentially. We will thus have a Type I growth singularity. Under both transformative scenarios-the one-time growth rate increase that can occur without capital-augmenting technological development and the growth explosion that occurs with it-capital per worker will grow to infinity. Since ρ > 0, the capital share will now tend to 1 rather than 0. For any fixed value ρ < 1 however, capital and labor are still complements; we still have F LK > 0. Absolute wages will therefore grow rapidly as the effective capital stock grows, as long as ρ is bounded below 1. In fact, with g A > 0, wages will grow superexponentially, though less quickly than output or effective capital. Absolute wages will stagnate or fall only if ρ rapidly rises to 1 (i.e. if rapidly grows to infinity)-or if capital and labor become perfect substitutes, in which case ρ = 1 (i.e. is infinite). The latter case is explored further in §3.3. Nordhaus also discusses an analogous possibility: that AI will transform consumption growth via the \"demand side\" of the economy, rather than the \"supply side\". To explore this scenario, instead of dividing the space of goods into two production inputs and one output, let us divide it into one input (\"capital\" K) and two outputs (which might be called \"standard consumption\" Y and \"computer-produced consumption\" Z). Capital grows exogenously at rate g K . Given capital stock K t , the production of the two consumption goods must satisfy Y t + Z t /D t = K t . (10) That is, each unit of capital can produce either 1 unit of standard consumption or D t units of computer-produced consumption per unit time (without being used up). 1/D t is the relative price of Z at t: it is the number of units of Y that must be given up at t per unit of Z. Consumers' utility function U is defined over Y and Z. U has the same features the production function was assumed to have: it is differentiable, increasing, and concave in each argument; it is CRS (recalling the caveat in §2.2); and its inputs are complements. In response to consumer demand, production is allocated between Y and Z to maximize utility. Suppose D t grows exponentially at rate g D . (This might be thought of as Moore's Law: famously, the number of computations that can be purchased with a given amount of capital seems to double approximately every eighteen months.) The relative price of Z then falls exponentially at rate g D . With each proportional fall in this relative price, the relative quantity of Z pro-duced will rise by σg D , where σ denotes the elasticity of substitution between the goods in the consumer utility function on the given margin. As long as σ < 1, therefore, the proportion S of capital allocated to computer-produced consumption falls to 0 over time. As long as σ > 1, S rises to 1.  11  Let C t Y t +Z t denote total consumption. If σ is bounded below 1, then, as t → ∞, approximately all capital is ultimately allocated to producing standard consumption, so g C = g K . If σ is bounded above 1, then, as t → ∞, approximately all of capital is ultimately allocated to producing computer-produced consumption, so g C = g K + g D . The AI-relevant implications are straightforward. If computer-produced consumption is not currently very substitutable for other consumption (σ bounded below 1), but developments in AI render it more substitutable (such that σ is then bounded above 1), then the consumption growth rate could rise from around g K to around g K + g D . This would not be a growth singularity, as we are using the term. But given the speed of Moore's Law, it would be a dramatic shift. The wage and labor share are not defined in this model, since capital is the only factor of production. As should be clear, however, an analogous model with labor would behave similarly, as long as instead of simply positing growth in capital g K , we also posit equal growth g B = g K in labor-augmenting technology. Then the labor share will be constant (by CRS), and consumption-denominated wages will grow at the consumption growth rate. Finally, Nordhaus constructs various tests of the hypothesis that we are headed for a growth increase via the channels discussed above. If we are in fact headed for a supply-driven growth increase, for example, we should expect to find a rising growth rate and a rising capital share. If we are headed for a demand-driven growth increase, we should expect to find a rising share of global income spent on computer-produced goods. A 11 This is not quite correct. From the definition of substitution elasticity, a fall in the relative price of Z by some small amount (say, g D ) will increase demand for Z by a proportional amount (say, σg D ) holding the budget fixed. Here, however, as long as g K > 0, the budget is also increasing. We must therefore impose some further restriction on the utility function if we are to draw any conclusions about what happens to the budget shares over time. For simplicity, we will assume that U is \"homothetic\", meaning that the budget shares optimally allocated to each good depend only on the goods' relative prices and not on the size of the budget. thorough discussion of his empirical conclusions would be beyond the scope of this survey, but he concludes that on balance the evidence disconfirms these hypotheses. Models in which labor and capital must be combined in more complex ways tend to produce the same broad conclusion. If labor and capital are sufficiently substitutable, then increasing capital productivity can increase the capital share, but it will still increase the absolute wage rate.  Berg et al. (2018)  detail a variety of such models. We will not work through them here. \n Perfect substitution We have seen that the wage-depressing effect of increasing capital productivity is most extreme when capital and labor are highly substitutable, but that for finite elasticities of substitution, absolute wages still tend to grow. Prospects for wages look worse in cases of perfect substitutability. As noted in §2.2, when goods are perfect substitutes toward some end, they are only ever both purchased in positive quantities when their prices are the same. Even if robots were already fully substitutable for human labor, therefore, we would only observe their transformative effects once their rental rate had fallen below what would otherwise have been the wage rate. In other words, perhaps the substitutability does not need to rise; perhaps it is perfect, and all that needs to change is a relative price. To illustrate this dynamic, consider the following simple model, inspired by  Hanson (2001) . Equipment Q, labor L, and land W are employed in a Cobb-Douglas production function, Y = F (Q, L, W ) = Q a L b W 1−a−b . ( 11 ) The output good can be consumed or invested as capital K. Capital can serve either as equipment or as robotics, which functions as labor, whereas the human workforce H is fixed and can only serve as labor. The productivity of capital-the number of units of effective capital generated by one unit of converted output-is denoted A. That is, if p denotes the fraction of capital employed as equipment, output is Y = (pAK) a (H + (1 − p)AK) b W 1−a−b . ( 12 ) A t rises exogenously without bound. For simplicity we will assume that a constant and sufficient fraction s of output is saved as capital. Because the substitution parameter between equipment and labor is not less than (in fact is equal to) 0, the accumulation of effective equipment is enough to sustain output growth. Early in time, when effective capital is scarce, all capital is used as equipment; p = 1. Indeed, at the rate at which capital can be converted from equipment to robotics, it would be valuable instead to use some human labor as equipment, if that were possible. Capital then grows (using discrete-time notation for clarity) such that K t+1 = K t + s(A t K t ) a H b W 1−a−b (13) ⇒ g K,t = (K t+1 − K t )/K t = sA a t K a−1 t H b W 1−a−b . As we can see from the right hand side, capital growth will approach a steady state such that ag A + (a − 1)g K = 0 ⇒ g K = a 1 − a g A . (14) We will thus have output growth of g Y = a(g A + g K ) = g A a/(1 − a). As the equipment stock grows, wages rise. As the productivity of capital rises and effective equipment grows more abundant, however, there comes a time past which it is optimal to split further capital between equipment and robotics. The labor growth rate then jumps to the rate that keeps its marginal productivity equal to that of equipment, and the output growth rate jumps accordingly. 12 In particular, with capital now filling the roles of both equipment and labor, we now have g Y = g K = g A (a + b)/(1 − a − b) , by the same calculation as above.  13  Hanson estimates the growth implications of crossing the robotics cost threshold using a slightly more realistic model with roughly realistic estimates of the parameters involved. The productivity of capital is assumed to double (i.e. the cost of effective capital is assumed to halve) every two years, in a conservative approximation to Moore's Law. Before capital begins to be used as robotics, output in the model grows at a relatively familiar rate of 4.3% per year. After, the growth rate is 45%. In the model above, because the production function is Cobb-Douglas, the labor share-the share of output paid in compensation for human and/or robotic labor-is independent of the factor quantities. As human labor constitutes an ever smaller share of total labor, however, the human labor share falls to zero. Furthermore, even the absolute wage F L falls to zero. To see this, note that in a CRS production function, the marginal productivities of equipment and labor are kept equal (F Q,t = F L,t ) when the quantities of the two factors grow at the same rate. We can thus rearrange our formula regarding competitive CRS factor payments: F L,t Q t + F L,t L t + F W,t W = Y t ⇒ F L,t = Y t − F W,t W Q t + L t . (15) With a constant share of output accruing to land as well, but the quantity of land fixed, land rent per unit of land-i.e. the land rental rate F W -must grow at the same rate as output. Y and F W will thus both grow at g Y , and Q and L will both grow at rate g A + g K = g A + g Y > g Y . The right-hand ratio will then fall to zero. In any CRS production function without labor-augmenting technology, what happens to the marginal productivity of labor, and thus wages, depends on the quantity of effective labor relative to that of the other effective factors of production. This relative quantity need not rise; it could fall, if labor's complements grow productive and plentiful more quickly than its substitutes, or stay fixed if they grow at the same rate. Consider the following model, very similar to the above, but in which technology augments only equipment, not capital used as robotics: Y = F (Q, L, W ) = Q a L b W 1−a−b = (pAK) a (H + (1 − p)K) b W 1−a−b . (16) The growing stock of equipment implies that, as above, wages rise before the substitutability cost threshold is crossed. Furthermore, we will still have g Y = g K , and thus g Y = a(g A + g K ) ⇒ g Y = a 1 − a g A (17) before the threshold is crossed. Finally, the threshold will still eventually be crossed: if all capital were used as equipment indefinitely, the marginal productivity of capital used as equipment AF Q = aY /K would fall below that of labor F L = bY /H. After the threshold is crossed, we will have g Y = a(g A + g K ) + bg K ⇒ g Y = a 1 − a − b g A . ( 18 ) Note that this is still a growth rate increase, though not as large as that in the first model above. To see what happens to wages, however, observe that when p is chosen so that invested output is split optimally between equipment and robotics, it will satisfy AF Q = F L , or aY pK = bY H + (1 − p)K ⇒ p = H + K K a a + b . (19) As K → ∞, we have p → a/(a + b) < 1, and therefore g L = g K = g Y . As above, because the production function is Cobb-Douglas, the labor share is constant. Now, however, the quantity of effective labor grows no more quickly than output, so labor payments per labor quantity-i.e. wages per human worker-merely stagnate.  Korinek and Stiglitz (2019)  offer another illustration of this phenomenon, in the context of a somewhat similar model. As usual we will simplify here to highlight the intuition. Suppose that Y is produced as in the second model of this section (i.e. the model just above), except that the substitution parameter between land and the other two factors is bounded below 0. Though land is in fixed supply, it is at first plentiful enough that its factor share is low. The saving rate is fixed. At first, as capital accumulates, it is split between use as robotic labor and use as equipment, so that the relative quantities of labor and equipment are unchanged. The capital and labor shares are roughly constant, but the absolute wage stagnates, as we have seen. In time, however, land becomes a binding constraint. The share of output received as land rents approaches 1, and the absolute wage falls to 0. As should be clear, the same logic could apply to many more complex models. Embed any production function in a \"surrounding\" production function with a fixed-supply and low-substitutability resource such as land, and in the long run, even if all of the original production function's resources grow abundant, the resource in fixed supply constrains growth and its owners receive approximately all output. \n Substitability in robotics production Like  Korinek and Stiglitz, Mookherjee and Ray (2017)  develop a model in which capital can replace human labor without technological progress. Unlike Korinek and Stiglitz, they do not simply assume that capital can be used as robotics, but make the robot production function explicit and identify a condition under which human labor replacement can occur. A simplification of their model is as follows. The final good Y is produced using capital K and labor L in a typical two-factor production function F , with a substitution parameter bounded below 0. Labor is supplied by human work H and robotics R, which are perfect substitutes. Robotics is better thought of as the provision of robot services than as robots, because it must be used as it is produced; it cannot be accumulated. If we would like to think of it as a kind of physical capital, we would say that it exhibits full depreciation. Robotics is also produced using capital and labor, using a standard but perhaps different production function f , also with a substitution parameter bounded below 0. Whereas one unit of robotics is defined as that which replaces 1 human worker in the output production function, however, one unit of robotics replaces some D ∈ (0, 1) human workers in the robotics production function. For each input X ∈ {K, H, R}, S X is defined (assuming X > 0) to be the fraction of X that is allocated to the production of robotics rather than the final good. For simplicity, the population of human workers is fixed and there is no technological progress. More formally, output and robotics at t are Y t = F ((1 − S K,t )K t , (1 − S H,t )H + (1 − S R,t )R t ); (20) R t = f (S K,t K t , S H,t H + DS R,t R t ). As usual, a constant fraction of output is saved as capital. Early on, when capital is scarce and human labor relatively plentiful, there may be no reason to produce robotics at all. As capital accumulates and output begins to be constrained by human labor, however, the marginal output productivity of capital falls to zero. It may therefore at some point be worthwhile to allocate some positive fractions S K and S H of available capital and human labor to robotics production. To be precise, it will necessarily start being worthwhile iff f L (k, 0) > 1 given k > 0, i.e. if, given some capital (which is eventually near-worthless in final good production), marginal contributions of labor can create robotics at a ratio of more than 1:1. Let us call this the \"robotization condition\". Note that it is a relatively weak condition; f L (k, 0) = ∞ given k > 0 if f is CES, for example. If robotics production relies on capital and human labor-i.e. if we set S R = 0-it too will ultimately be constrained by lack of labor: as S K,t K t → ∞, R t → S H,t R, where R lim x→∞ f (x, H). (21) Output in turn is constrained by total labor, despite the possibility of robotics: as (1 − S K,t )K t → ∞, Y t → Ȳ (1 − S H,t )H + R t H + R , (22) where Ȳ lim x→∞ F (x, H + R). Long-run output will be maximized by setting S H = S * H , the value that maximizes (1 − S * H )H + S * H R-i.e. S * H = 1 if the robotization condition is met, because in this case R > H, and S * H = 0 if not. Long-run output will then approach an upper bound of Ȳ ((1−S * H )H +S * H R)/(H + R). The human labor share will approach 1, either because it is the scarce input to output directly or because robotics is the scarce input to output and human labor is the scarce input to robotics. The absolute wage will of course stagnate. In short, robotization can raise the output ceiling, but it cannot on its own produce a sustainably positive growth rate. One might expect that, if we do not fix S R = 0, it will eventually be optimal to use robotics in the production of robotics. In fact, this will only happen if D is large enough that, as the quantity of capital allocated to robotics production grows large, one unit of robotics can produce more than one unit of robotics: that is, if lim k→∞ f L (k, H) > 1/D; or equivalently, because f is CRS, if f L (k, 0) > 1/D for k > 0. The identification of such a condition in their more general setting is Mookherjee and Ray's key insight, and they call the condition the \"von Neumann singularity condition\", after the work by  Burks and Von Neumann (1966)  on self-replicating automata. It is of course very closely analogous to the robotization condition above, but stronger since we are imposing D < 1. We might take this to be the natural case; robotics production is presumably harder to automate than most other tasks are.  14  Suppose that this condition is met, and that S K K is large enough that f L (S K K, H) > 1/D. Then there is an optimal quantity of robotics to allocate to robotics production, so as to maximize net robotics production. This is the quantity such that use of a marginal unit of robotics on robotics production increases robotics output by exactly one unit. That is, it is optimal to set S R > 0 such that S R satisfies f L (S K K, H + DS * R R) = 1/D. The value of R depends on S R , since more inputs to robotics production will correspond to higher robotic output. Nevertheless we know that a unique S R ∈ (0, 1) satisfying the above equality exists, for a given S K K. To see this, recall that f L (S K K, H + DS R R) > 1/D at S R = 0, by supposition. And we must have lim S R →1 f L (S K K, H + DS R R) < 1/D, or else the quantity of robotics output R and thus also S R R would grow without bound, fixing S K K, as S R → 1; but in this case f L → 0, by the assumption that the substitution parameter in the robotics production function is bounded below 0. By the concavity and continuous differentiability of f , therefore, there is a unique S * R : f L (S K K, H + DS * R R) = 1/D. Under the singularity condition, S K and S R approach constants S * K and S * R , strictly between 0 and 1, as the capital stock grows.  15  Growth proceeds 14 Presumably at least some other tasks are more difficult to automate, however. As Mookherjee and Ray present in the original paper, their central result does not depend on robotics production being more difficult to automate than all other tasks, just on robotics production being sufficiently difficult to automate. 15 Consider a time t at which S R,t > 0, and let m > 1 denote K t /K t for some t > t. From t to t , the capital input to robotics production is multiplied by mS K,t /S K,t . Because f is CRS, to maintain the condition that f L,t = f L,t = 1/D the labor input to robotics production must also be multiplied by mS K,t /S K,t , and robotics production will then also be multiplied by this factor. We thus have (a) H + DS R,t R t = (H + DS R,t R t )mS K,t /S K,t and (b) R t = R t mS K,t /S K,t . Because both inputs to robotics production are multiplied by a common quantity, f K is ECONOMIC GROWTH UNDER TRANSFORMATIVE AI as in an AK model, with the rate of capital accumulation, final good output growth, and robotics output growth all asymptotically constant and proportional to the saving rate.  16  The wage level is constant and lower than it is in the absence of the singularity condition, since the ratio of capital to labor in robotics production is still asymptotically constant but now positive rather than zero. The share of income accruing to human labor falls to zero. As with  Hanson (2001) , moderate tweaks to this model could result in absolute human wages rising or falling, rather than merely stagnating. Also, in the presence of population growth or labor-augmenting technology growth, introducing automation can increase the growth rate of final good output (or final good output per capita) from the rate of effective labor (or labor-augmenting technology) growth to something much higher, given sufficient saving. As we saw in §3.3, when human work must compete with robotics for which it is perfectly substitutable, the standard result is that the human labor share falls to 0 and the wage stagnates or changes (likely falls) as well. Above, however, we saw that modeling robotics production explicitly, rather than stipulating a frictionless conversion of the final output good into robotics, allows for a channel through which human work can remain necessary. A positive human labor share can be maintained, even when robotics can fully constant across periods. It follows that if the capital input to final good production grows proportionally more (less) than the labor input, the marginal productivity of capital in final good production falls (rises), and the marginal contribution of capital to final good production via robot production rises (falls). Thus, to maintain the condition that capital is allocated efficiently, the capital and labor inputs to final good production must be multiplied by a common quantity across periods: (c) m(1 − S K,t )/(1 − S K,t ) = R t (1 − S R,t )/(R t (1 − S R,t )). Substituting (b) into (a) and (c) and solving for S K,t and S R,t , we find that, as m → ∞, •  S K,t → S * K S K,t DR t (1 − S R,t )/(DR t (1 − S R,t ) − (1 − S K,t )H) and • S R,t → S * R S R = Y /K = F ((1-S * K ), (1-S * R )R/K), where R/K likewise satisfies R/K = f (S * K , S * R R/K). Technological progress that allows capital to produce more robotics increases long-run R/K and functionally \"increases A\", though it will never exceed its upper bound of F (1, ∞). This will be finite, by the assumption that F's substitution parameter is bounded below 0. substitute for human work in the final good production function, when human work cannot be fully substituted for in robotics production.  Korinek (2018)  presents another model in which humans and robots must in some sense compete and in which robots are not simply \"capital that can function as labor\" but items that must be produced and sustained. He too finds that the human labor share, and in his case also the wage rate, can fall to 0 unless human labor remains necessary for the maintenance of the robot population. But we will not explore his model further here. \n Growth impacts via impacts on saving In some of the models we have considered, saving has been an important determinant of growth. The saving rate, however, has been assumed to be exogenous. This leaves open another channel through which developments in AI could impact growth: by changing the rate of return to saving, more advanced AI could change the rate of saving and thus the growth rate. This scenario can be illustrated most simply using a model from  Korinek and Stiglitz (2019) . Suppose labor and capital are perfectly substitutable. 17 Labor can only be supplied by humans. Activity unfolds in discrete time, and capital depreciates fully every period; it cannot accumulate. (Capital depreciation simplifies the exposition but is not necessary for the central result.) Given saving rate s t , that is, output and capital growth are then Y t = AK t + BL t , K t+1 = s t Y t . ( 23 ) We begin with K = 0. Output per capita is thus B and, without saving, does not grow. If A < 1, there is no incentive to save, and doing so cannot generate growth; foregoing each unit of consumption at t would offer someone only A < 1 additional units of consumption at t + 1, starting from the same baseline of B. For any A > 1, on the other hand, individuals with no or sufficiently low pure time preference will want to save some fraction s t > 0 of their incomes; not to do so would be to miss the opportunity to give up marginal consumption at baseline B in exchange for a larger quantity of marginal consumption also at baseline B. More precisely, it should be clear that positive saving will be optimal, given any pure time discount factor β < 1, as long as A > 1/β. Furthermore, under certain assumptions about the shape of individuals' utility functions, the induced saving rate will be some constant s > 1/A, independent (at least roughly) of the absolute output level. In the long run, as the relative contribution of effective human labor BL t grows negligible, we will have Y t ≈ AK t . And since K t+1 = sY t ≈ sAK t (24) ⇒ (K t+1 − K t )/K t ≈ sA − 1, capital and therefore output will grow at asymptotic rate sA − 1 > 0. In short, an increase in A-induced, perhaps, by AI developments which render robots cost-effective replacements for human labor-can trigger saving and can thus increase the growth rate of output and output per capita. Here, an A-increase raises per capita output growth from zero to a positive number, leaves the wage rate constant at B, and pushes the human labor share to zero; but other impacts on wages, the labor share, and growth are possible. The point is just that, in addition to the ways in which increases in A can sometimes directly impact the growth rate, they can sometimes do so indirectly by impacting s. There is another mechanism through which developments in AI could impact the saving rate. If the saving rate is heterogeneous across the population, then growth will depend on how income is distributed between high-and lowsavers. Developments in AI could thus affect the growth rate by affecting the income distribution. In principle, this effect could have implications for growth in either direction. Here, we will focus on the especially interesting and counterintuitive possibility that AI slows and even reverses growth by transferring wealth from those with low to those with high propensity to consume. This scenario is illustrated most simply by  Sachs and Kotlikoff (2012) , though the same mechanism is explored in more detail by  Sachs et al. (2015) . If some investment goods are sufficiently substitutable for labor, automation raises capital rents but lowers wages. If saving for the future comes disproportionately out of wage income, for whatever reason, then this wage-lowering can cause future output to fall. Consider an overlapping generations (OLG) economy with constant population size. Each person lives for two periods. The young work, investing some of their income; the old live off their investments. More precisely, output is a symmetric Cobb-Douglas function of labor and capital. The output good can be consumed or invested as capital K. Capital can be used either as equipment Q or as robotics, which serves as labor, and it is split between these uses until their marginal products are equal. The human workforce H is fixed and can only serve as labor. Unlike in the Hanson model, however, the productivity of equipment is fixed. A denotes the productivity of robotics only. Formally, if p is the share of capital used as equipment, Y t = F (Q t , L t ) = (p t K t ) 1/2 (H + (1 − p t )A t K t ) 1/2 . ( 25 ) Capital at t is financed by those who were young at t − 1, who put aside half their wage incomes as investment.  18  The old at t consume all their wealth: not only their investment income, F Q,t p t K t + F L,t (1 − p t )A t K t , but even the capital stock K t , which is liquidated after use in production. The economy is in a zero-growth steady state when investment is just replenished each period: that is, when F L,t /2 = K t . Now suppose robotics grows slightly more productive, so that A t+1 > A t . Let G A 1 + g A denote A t+1 /A t . For a single period, total output and the incomes of the old grow. The young see a fall in wages, however, and investment therefore falls as well. This fall in investment outweighs the fact that some of the investment, namely that in robotics, is now more productive. Output therefore falls. The wage rate falls too, due both to the abundance of robotics and to the lack of investment in equipment. More formally: in the new equilibrium, the marginal product of robotics is again equal to that of equipment. Because the \"relative cost\" of robotics has now been multiplied by 1/G A and because here ρ = 0, the relative quantity of labor must be G A times higher. Letting asterisks denote the new equilibrium outcomes, L * /Q * = G A L t /Q t . So F * L = 1 2 L * Q * −1/2 , F L,t = 1 2 L t Q t −1/2 (26) ⇒ F * L = F L,t G −1/2 A < F L,t . The new rate of return on investment is r * = A t+1 F * L = G A A t F * L , the new wage is F * L , and the saving rate remains 1/2. The consumption of the old 18 Let r t denote the interest rate at t: that is, here, F Q,t+1 , or equally A t+1 F L,t+1 . Suppose that period utility is logarithmic in consumption and that the young choose the saving rate s t to maximize lifetime utility ln((1 − s t )F L,t ) + ln(s t F L,t (1 + r t )). Then the chosen saving rate will always equal 1/2. thus equals half their income while young, times 1 + r * : (1 + G A A t F * L ) F * L 2 = (1 + A t G 1/2 A F L,t ) F L,t 2G 1/2 A (27) = (G −1/2 A + A t F L,t ) F L,t 2 < (1 + A t F L,t ) F L,t 2 . The new equilibrium therefore features lower output in all subsequent periods, and lower consumption for both young and old. If robotics productivity continues to grow at rate g A , the wage rate, and thus the consumption of the young, will continue to fall to 0 at rate G 1/2 A − 1. The consumption of the old (and thus also output) will fall at a falling rate, and will eventually stabilize above 0, as the increasing productivity of invested equipment ever more closely compensates for the falling absolute amount invested.  19  The human labor share thus falls to 0. Again, the direction of these impacts is sensitive to whether the \"winners\" from advances in AI save more or less than the \"losers\". As  Berg et al. (2018)  point out, for instance, those who make most of their incomes from wages currently empirically exhibit lower saving rates than those who make most of their incomes from capital rents, so the mechanism identified by Sachs and Kotlikoff should if anything increase output growth. In any event, the key point is just a reiteration of the well-known fact that, in a neoclassical growth model with finitely lived agents and no (or imperfect) intergenerational altruism, the rate of saving is not necessarily optimal. Accordingly, policymakers must always consider not only the impact of a policy or technological development on short-term output, but also its impact on the saving rate. When a given development produces a suboptimal saving rate, it should be counterbalanced by investment subsidies or by transfers from those with high to those with low propensity to consume: in this model, from the old to the young. \n AI in task-based models of good production 4.1 Introducing the task-based framework In §3, we imagined that capital and labor were each employed in a single sector. In the Cobb-Douglas case, we held the exponent a on capital fixed. We then explored the implications of changing ρ, the substitutability of capital and other durable investments for human labor, and of independently changing the growth rates of factor-augmenting technology. In reality, however, capital and labor are of course employed heterogeneously, and this heterogeneity seems likely to shape the economic impacts of developments in AI. Indeed, sectors with high rates of automation have historically experienced stagnating or declining wages  (Acemoglu and Autor, 2012; Acemoglu and Restrepo, 2020) , even as wages on average have increased. Here, therefore, we will explore a model of CES automation from Zeira (1998), which makes room for this sort of heterogeneity. (We will follow the exposition and extension of Zeira's model given by  Aghion et al. (2019) .) As we will see, this model amounts roughly to assuming a fixed substitution parameter ρ and either a changing capital exponent a, in the Cobb-Douglas case, or impacts on factor-augmenting technology which are sensitive to ρ when ρ = 0. Let us begin with the ρ = 0 case. Suppose output is given by a Cobb-Douglas combination of a large number n of factors X i , for i = 1, . . . , n: Y = X a 1 1 • X a 2 2 • • • • • X 1−a 1 −•••−a n−1 n . (28) At such a fine-grained level, these \"factors\" might better be thought of as intermediate production goods (Zeira, 1998), or even as individual tasks (Acemoglu and Autor, 2011). We will refer to them as tasks. Fraction a of the tasks are automatable, in that they can be performed by capital or labor, and fraction 1−a are not, in that they can only be performed by labor. Given capital and labor stocks K and L, if all automatable tasks are indeed automated (performed exclusively by capital), K/(na) units of capital will be spent on each automatable task and L/(n(1 − a)) units of labor on each non-automated task. With just a little algebra, we have Y = AK a L 1−a , (29) a two-factor Cobb-Douglas production function with an unimportant coefficient A. 20 Now consider a general CES production function with a continuum of production factors Y i from i = 0 to 1, instead of just two: Y = 1 0 Y ρ i di 1/ρ (30) Tasks i ≤ β ∈ (0, 1) are automatable. Let K and L denote the total supplies of capital and labor, and K i and L i the densities of capital and labor allocated to performing some task i (so Y i = K i + L i ). Suppose again that all automated tasks are indeed performed exclusively by capital. Since the tasks are symmetric and the marginal product of each task is diminishing (∂ 2 Y /∂X 2 i < 0), the density of capital applied to each task i ≤ β will be equal (assuming that production proceeds efficiently), as will the density of labor applied to each i > β. And since we must have β 0 K i di = K and 1 β L i di = L, (31) we know that K i = K/β ∀i ≤ β and L i = L/(1 − β) ∀i > β. We can thus write our production function as Y = [β(K/β) ρ + (1 − β)(L/(1 − β)) ρ ] 1/ρ (32) = [β 1−ρ K ρ + (1 − β) 1−ρ L ρ ] 1/ρ = F (AK, BL) = [(AK) ρ + (BL) ρ ] 1/ρ where A = β (1−ρ)/ρ and B = (1 − β) (1−ρ)/ρ . This is simply a two-factor CES production function. We have assumed that all automatable tasks are indeed performed exclusively by capital. This will obtain so long as there is more capital per automatable task than labor per non-automatable task, i.e. as long as K β > L 1 − β . ( 33 ) In this case, even when capital is spread across all automatable tasks, we have F K < F L , so there is no incentive to use labor on a task that capital can perform. Let us call this condition the \"automation condition\". For any fixed β, if capital accumulates indefinitely and the labor supply stays fixed or grows more slowly, the condition will eventually hold. \n Task automation Let us now explore the implications of task automation in more detail, across the regimes of CES production with ρ = 0 and < 0. The case of ρ > 0 will be covered in §4.3. As we have seen, under Cobb-Douglas production, task automation raises a along the range from 0 to 1. Since a constant saving rate imposes g Y = g K , the growth rate in this case satisfies g Y = a(g A + g K ) + (1 − a)(g B + g L ) ⇒ g Y = a 1 − a g A + g B + g L . (34) The impact of a one-time increase to a, or of increases only up to some bound strictly below 1, is therefore straightforward. The capital share rises with a. If g A > 0, the growth rate increases, ultimately raising the wage rate; otherwise the growth rate is unchanged, and the impact on the wage rate is ambiguous. Given asymptotic complete automation, the model approximates an AK model. If g A = 0, the growth rate rises to s. If g A > 0, the growth rate rises without bound. If 1 − a falls to 0 at a constant exponential rate, wages too rise superexponentially. Automation, as we have defined it, allows capital to perform more tasks. One might therefore imagine that it is equivalent to the development of some sort of capital-augmenting technology.  Aghion et al. (2019)  observe, however, that automation in the above model is actually equivalent to the development of labor -augmenting-and capital-depleting!-technology, as long as ρ < 0 and the automation condition holds. To see this, recall our production function: Y = [(AK) ρ + (BL) ρ ] 1/ρ , (35) where A = β (1−ρ)/ρ and B = (1 − β) (1−ρ)/ρ . As β rises from 0 to 1, therefore, A falls from unboundedly large values to 1, and B in turn rises from 1 without bound. The reason for this result is that, as β rises, capital is spread more thinly across the widened range of automatable tasks, and labor is concentrated more heavily in the narrowed range of non-automatable tasks.  21  Automation therefore allows capital to serve as a better complement to labor. A marginal unit of labor is spread across fewer non-automatable tasks, producing a larger increase to the supply of each; given the abundance of capital, this then produces a larger increase to output. Conversely, under this allocation, labor serves as a worse complement to capital, requiring capital to spread itself over more tasks (and only partially compensating for this effect by supplying the remaining tasks more extensively). As explained in §2.3, when ρ < 0, labor-augmenting technology is the key to sustained output growth. Let us spell that out in this context. Suppose that, by some exogenous process, a constant fraction of the remaining nonautomatable tasks are made automatable each period, so that (1−β t ) → 0 at a constant rate g 1−β < 0. Then B will grow at rate g B = g 1−β (1 − ρ)/ρ > 0. A is asymptotically constant at 1, so g A ≈ 0. If the saving rate s is constant and high enough to maintain the automation condition, we get our familiar \"balanced growth path\". The capital stock, effective labor supply, and output all grow at asymptotic rate g Y = g B + g L , output per capita grows at rate g B , and the labor share is asymptotically constant and positive.  22  Automation can thus increase the growth rate of output per capita, and have other transformative consequences.  23  In the model above, because the 21 Here we are only considering increases in β up to the point that capital per automatable task no longer exceeds labor per non-automatable task, so that the automation condition is satisfied. 22 The capital share here equals β 1−ρ t (K t /Y t ) ρ . As β → 1, the capital share rises to an upper bound of (K/Y ) ρ , where K/Y is the long-run capital-to-labor ratio, as long as this exists and is finite. It follows from sY t = K t+1 − K t that sY t /K t = g K,t ; and since g K = g Y = g B + g L , we have K/Y = s/(g B + g L ). The labor share will thus fall to 1 − (s/(g B + g L )) ρ . This is nonnegative because sA ≥ g B + g L , by sufficient saving, and A is asymptotically 1; and it is strictly positive as long as we are not in the knife-edge case of sA = g B + g L . 23 One might however take the position that what we have here been calling automation is not a new force on the horizon, promising to augment pre-existing drivers of growth, but a microfoundation for the process of labor-augmenting technological development we have observed for centuries. On this view, advances in AI will continue to push β ever automation rate −g 1−β is the only driver of growth, introducing it increases the growth rate from 0 to g 1−β (1 − ρ)/ρ. In the presence of growth from other sources, automation can increase the growth rate further. Consider for instance what follows if we have B t = D t (1 − β) (1−ρ)/ρ , with β constant but D t growing exogenously at rate g D . Given saving sufficient to maintain the automation condition, output per capita grows at rate g D . The implications of introducing automation at rate −g 1−β then depend on whether saving is still sufficient to maintain the automation condition. If it is, the per-capita growth rate increases to g D + g 1−β (1 − ρ)/ρ, and the labor share falls to an asymptotic positive value, as observed above. For more on a model of task automation with (something close to) direct labor-augmenting technology growth g D > 0, see §4.4. Now suppose again that g D = 0 , but now suppose that saving is not sufficient to maintain the automation condition-as it cannot be if, for instance, all tasks become automatable. In this case some automatable tasks will not be automated. Here, things proceed roughly as in a model of full substitutability. The growth rate is capped at sA, the wage rate equals the capital rental rate and stagnates, and the labor share equals L/(L+K). Assuming sA > g L , the labor share falls at rate g L −g K = g L −sA. Finally, consider the implications of exogenous growth in capitalaugmenting technology A. When not all tasks are automated, increases to A only increase the effective capital stock. If g A > 0 and s > 0, even if the automation condition is not yet met, the growth rate sA will grow until it is met. The capital stock will then grow at the output growth rate, the effective capital stock will grow faster by g A , and the capital share will fall to 0, roughly as explained in §3.1. When all tasks are automated, on the other hand, output Y t asymptotically equals A t K t , and sustained exponential growth in A produces a Type I growth singularity. \n Task creation Let us begin with the  Aghion et al. (2019)  model of automation and introduce a process of task creation. The resulting model will be somewhat akin to that developed by  Hémous and Olsen (2014) . closer to 1, but this process will simply continue the existing trend. As before, output is a CES production of a range of tasks. Each task is performed by labor and/or capital, with tasks above an automation threshold β requiring labor. Now, however, new and initially non-automated tasks can be created. The range of tasks thus runs from i = 0 to N , with tasks i ≤ β automatable, and not only β but also N can be increased. By the same reasoning as in §4.1, if there is enough saving that the automation condition is met, output is Y = [(AK) ρ + (BL) ρ ] 1/ρ ( 36 ) where A = β (1−ρ)/ρ and B = (N − β) (1−ρ)/ρ . If ρ < 0, then increases to β holding N fixed act like labor-augmenting technology, as in  Aghion et al. (2019) . By the same token, however, increases to N holding β fixed act like labor-depleting technology; they require labor to \"spread itself too thinly\". It will never be productive to create new tasks, and automation and growth will simply proceed as in §4.2. If ρ > 0, on the other hand, it is increases to N , holding β fixed, that function as labor-augmenting technology. In particular, they asymptotically produce g B = g N (1 − ρ)/ρ. As explained in §3.2, growth then proceeds at a rate of max(sA, g B ), where s denotes the saving rate. Increasing the rate of task creation can thus increase the growth rate. More importantly, increases to β, regardless of N , function as advances in capital-augmenting technology. Recall that effective capital accumulation is enough for growth when ρ > 0. By raising the \"ceiling\" N and allowing for future automation to raise β, task creation can thus have radical effects. To see this, first suppose that N increases exogenously at a constant proportional rate g N , and that N − β is constant. This is essentially the case explored by  Nordhaus (2015)  and summarized in §3.2: given ρ > 0, capital accumulation and capital-augmenting technology growth combine to produce a Type I growth singularity. The labor share will fall to 0, even while wages, like output (though more slowly than output), grow superexponentially. If g A = g B , i.e. if g β = g N so that a constant fraction of tasks is always automated, the outcome is similar; indeed, conditions are even more favorable to labor. Upon endogenizing the task automation and creation processes, the g β = g N condition turns out to hold under relatively natural-seeming circumstances. For a few more words on this, see the end of the following section. Finally, as discussed at the end of §3.3, suppose we embed production function (36) in a \"surrounding\" production function. That is, suppose that the good denoted Y , which we had been referring to as the final output good, must instead be combined with a fixed-supply resource (such as land W ) in order to produce the final output Z. What follows depends centrally, as we have seen, on the substitution parameter ρ between Y and W . If ρ > 0, essentially nothing changes; the land share is asymptotically 0, growth in Z approximately equals growth in Y , and so forth. If ρ < 0, on the other hand, output approaches an upper bound. Then the relative quantity of Y rises without bound, by capital accumulation or asymptotic task automation; the land share rises to 1; and the wage rate falls to 0. This is more similar to the case explored by Hémous and Olsen, though they take the fixed-supply resource to be skilled labor, whereas we are ignoring skill differences throughout this review. In any event, we will not explore this further here.  Acemoglu and Restrepo (2018b, 2019a ) develop a similar model of task creation, but combine it with a process of task replacement. Here is a simplification. \n Task replacement Instead of ranging from 0 to N , task indices i now range from N − 1 to N . Capital is equally productive at all tasks it can perform, but labor productivity at task i is B i = D i (1 − β) (1−ρ)/ρ , where D i = exp(g D i), for some g D > 0. In an exogenous growth setting, both β ∈ (N − 1, N ) and N grow over time at a constant exogenous absolute rate-let us say, without loss of generality, at one unit per unit time. The fraction of tasks not automatable is thus constant at N − β, but the productivity of human labor at the non-automatable tasks grows at exponential rate g B = g D . With enough saving, all automatable tasks are automated. Output and capital grow at rate g D + g L , in line with the effective labor supply, and wages grow at rate g D . The labor share is constant, as in any CES model with labor-augmenting technology growth and sufficient saving. Moving from asymptotic automation in the original setting of §4.2 to this model of task automation, creation, and replacement thus increases the output growth rate iff g D > g 1−β (1 − ρ)/ρ. As usual, if we imagine starting from a world without task creation and replacement, introducing this pro-cess raises the growth rate from 0 to a positive number; and if we imagine starting from a world with some other source of exogenous growth in labor productivity, introducing this process raises the growth rate, given enough saving to maintain the automation condition. This model is nearly equivalent to a task-based model in which the task-range is fixed at the unit interval, β is fixed, and B grows exogenously at rate g D . Its framing is motivated by the empirical observation that we have long seen the automation of existing tasks go hand-in-hand with the creation of new, high-productivity tasks that, at least temporarily, only humans can perform  (Goldin and Katz, 2009; Acemoglu and Autor, 2012) , and continue to see this pattern in the present  (Autor, 2015; Acemoglu and Restrepo, 2019b) . The result is a near-complete turnover of job types over time, rather than a mere encroachment of automation onto human territory. In this sense, this promises to be a more realistic model of automation than one without task replacement. As we have just seen, more realistic models of this type are also compatible with balanced growth. This balanced growth result is, however, sensitive to the assumption that advances in automation technology (increases in β) and task creation (increases in N ) proceed at the same rate. If task creation oustrips automatability, the labor share rises, and as β − N + 1 → 0 asymptotically, the labor share rises to 1. In this case, we approach a state in which labor performs all tasks. Capital is relegated to an ever-shrinking band of the lowest-labor-productivity tasks. Since output and, given a constant saving rate, capital grow at the same rate as effective labor, while capital is used ever less efficiently, capital rents fall and the capital share falls to 0. In equilibrium, output grows at rate g D + g L , and wages grow at the labor productivity growth rate g D , as before. Now suppose that automatability outstrips task creation, and in particular that it does so at a constant rate g N −β < 0. What follows depends on the extent to which capital accumulation keeps up with this process. If s ≤ g D + g L , the automation condition will not be met in the long run, and capital and effective labor will be perfect substitutes on the margin. Output will thus equal K t + D t L t , and the stock of capital grows at the same rate as that of effective labor when s (K t + D t L t )/K t = g D + g L , i.e. when K t /(K t + D t L t ) = s/(g D + g L ). In the long run we thus have g K = g Y = g D + g L , the labor share and the fraction of tasks not automated approach 1 − s/(g D + g L ) (ranging from 1 at s = 0 to 0 at s = g D + g L ), and wages grow at rate g D . Now suppose that s ∈ (g D + g L , g D + g N −β (1 − ρ)/ρ + g L ], so that capital accumulation outpaces labor and labor productivity growth in isolation but not in combination with automation. Now the fraction of tasks automated grows over time: if it stayed constant, capital per automated task would ultimately exceed effective labor per non-automated task, since s > g D + g L , and it would be profitable to reallocate some capital to automatable but non-automated tasks. But the fraction of tasks automated does not catch up with the automatability frontier: if all automatable tasks were automated, effective labor per non-automated task would ultimately exceed capital per automated task, since s < g D +g N −β (1−ρ)/ρ+g L , and it would be profitable to reallocate some labor to currently automated tasks. Thus the automation condition is not met in this scenario either, and capital and effective labor are still perfect substitutes on the margin. Since the stock of capital grows more quickly than that of effective labor, in the long run output grows at rate s, wages grow at rate g D , and the labor share again falls to 0. Finally, if s > g D + g N −β (1 − ρ)/ρ + g L , the automation condition is met. Growth proceeds at g D + g N −β (1 − ρ)/ρ + g L and the labor share approaches a positive constant, as in the model of §4.2 with direct labor-augmenting technology growth g D > 0. Empirically the saving rate is currently far higher than the growth rate of effective labor, so unless automation accelerates dramatically, this is the most relevant case for consideration. Now let us briefly and informally consider the implications of endogenizing the automation technology and task creation processes. Suppose that, in addition to the labor force, there is a pool of researchers who allocate their efforts between increasing β and increasing N . Upon doing either, they earn a patent right to some of the gains that result. This scenario, as detailed by  Acemoglu and Restrepo (2018b) , produces intuitive equilibrating pressures, suggesting that we might expect to observe automation technology and task creation proceeding at the same rate, without having to assume this ad hoc. Excessive development of automation technology results in tasks that are automatable but not automated, because of an insufficient ratio of capital to effective labor. This eliminates the immediate value of further automation technology. Excessive task creation, on the other hand, increases the value of automation technology, by inefficiently relegating capital to a narrower range of tasks. The full range of possibilities here, however, is essentially the same as in the exogenous growth case. The proportion of tasks automated can fall to 0, if the saving rate is sufficiently low; there can be asymptotically complete automation if the saving rate is sufficiently high; and there is partial automation in intermediate cases. The primary novelty of the endogenous research case is that here, which case obtains can depend on the researchers' productivity at developing automation technology relative to their productivity at task creation. In particular, increases in productivity at developing automation technology, relative to productivity at task creation, can increase the equilibrium automation rate and decrease the equilibrium labor share. Also, the growth rate here is not exogenous but depends on the level of productivity at both researcher tasks, as well as on the size of the researcher population. \n AI in technology production Throughout the discussion so far (except for the brief note at the end of the previous section), technological development has been exogenous, when it has appeared at all. Even in this circumstance, developments in AI have proved capable of delivering transformative economic consequences. When technological development is endogenous, and in particular when more advanced AI can allow it to proceed more quickly, the resulting process of \"recursive self-improvement\" can generate even more transformative consequences. \n Learning by doing This recursive effect can be seen most simply in a model with Cobb-Douglas production. Let us interpret the production as task-based, with fraction a of tasks automated, as in §4.1. The model below is inspired by the exploration of learning by doing in  Hanson (2001) . The labor supply grows at exogenous rate g L > 0, but capital productivity growth proceeds endogenously. We will use a modification of the endogenous growth model presented in §2.4. In that model, technology growth is a function of existing technology and \"researcher effort\". Here, instead, we will not introduce research and will say simply that capital productivity grows as a function of the existing technology and output. That is, Y t = (A t K t ) a L 1−a t , (37) where Ȧt = A φ t Y λ t ⇒ g A,t = A φ−1 t Y λ t (38) for some φ < 1 and λ > 0. One might interpret this as a model in which the production process itself contributes to the generation of productivityincreasing ideas. Given a constant saving rate, g K = g Y . So, from our production function, g Y = a(g A + g Y ) + (1 − a)g L g Y = a 1 − a g A + g L . (39) From our formula for g A,t , the steady state of g A (if it exists) will be that which satisfies (φ − 1)g A + λg Y = 0. Substituting for g Y in this expression and solving for g A , we have g A = λ(1 − a) (1 − a)(1 − φ) − λa g L . (40) This exponential growth path will exist as long as the denominator is positive: that is, as long as a < 1 − φ 1 − φ + λ . (41) In this case, output growth will be given by 24 g Y = (1 − a)(1 − φ) (1 − a)(1 − φ) − λa g L . (42) Otherwise, the recursive process by which proportional increases to A t generate proportional increases to Y t , which in turn generate proportional increases to A t+1 (using discrete-time notation for clarity), results in the proportional increases at t + 1 being larger than those at t. The growth rates of A and Y thus increase without bound. The transformative potential of automation is now straightforward. Increases in a increase the long-run growth rate without bound, as a approaches (1 − φ)/(1 − φ + λ). Past this threshold, increases in a trigger a growth singularity. The singularity type can be determined by substituting (37) into expression (38) for g A,t : g A,t = A φ−1 t Y λ t = A φ−1 t A aλ t K aλ t L (1−a)λ t . (43) Since 25 K t ∝ Y t = (A t K t ) a L 1−a t , rearranging gives us K t ∝ A a/(1−a) t L t . Sub- stituting for K t into (43), we have g A,t ∝ A φ−1+λa/(1−a) t L λ t . (44) When a > (1 − φ)/(1 − φ + λ), the exponent on A t is positive, producing a Type II singularity. When a = (1 − φ)/(1 − φ + λ), we have g A,t ∝ L λ t ; the technology growth rate itself grows asymptotically at rate λg L , producing a Type I singularity. \n Automated research Cockburn et al. (  2019 ) taxonomize AI systems as belonging to three broad categories: symbolic reasoning, robotics, and deep learning. Symbolic reasoning systems, they argue, have proven to have few applications. Robotics-by which they broadly mean capital that can substitute for human labor in various ways, instead of complementing it-has of course had many applications. It has also been the subject of a substantial majority of the theoretical literature on the economics of AI, including all that discussed in this survey so far. They propose however that the most transformative possibilities come from deep learning systems, by which they mean systems that can learn as human researchers can: artificial systems that can participate directly in the process of technological development. Citing  Griliches's (1957)  discussion of the implications of \"inventing a method of invention\", they argue that deep learning systems (in their use of the term) will have qualitatively more radical consequences than mere robotics. As we will see, this appears to be correct. Following  Aghion et al. (2019) , let us focus directly on the implications of technology production by using an even simpler production function than usual: Y t = A t (1 − S)L t . ( 45 ) 25 The \"∝\" symbol means \"is asymptotically proportional to\". Labor L is the only factor of production. S is the constant proportion of people who work in research as opposed to final good production. Output technology A, however, is developed using a CES function of both labor and capital K. Building on the standard technology production function from §2.4 (where Ȧt = A φ t (S t L t ) λ , but here assuming in effect that λ = 1), we have Ȧt = A φ t [(C t K t ) ρ + (D t SL t ) ρ ] 1/ρ , ρ = 0; (46) Ȧt = A φ t (C t K t ) a (D t SL t ) 1−a , ρ = 0, for some substitution parameter ρ and, in the Cobb-Douglas case, capital exponent a. C t and D t denote capital-and labor-augmenting technology levels respectively. As usual, we will assume a constant saving rate s. Suppose ρ < 0. Recall that in this case sustained growth in effective capital and in effective research labor are both necessary to sustain growth in output technology, and growth will be driven by whichever factor grows more slowly. Observe that when growth is constrained by effective capital accumulation, we have g A ∝ A φ−1 t C t K t , and that when it is constrained by effective research labor growth, we have g A ∝ A φ−1 t D t L t . Regarding φ: • Recall from the reasoning of §2.4 that, if φ < 1, we have g A = (g C + g K )/(1 − φ) on the capital-constrained path and g A = (g D + g L )/(1 − φ) on the labor-constrained path. In the former case, since g K = g Y = g A + g L , we can substitute for g K and rearrange to get g A = −(g C + g L )/φ. A capital-constrained path with this growth rate exists when φ < 0. Thus, if φ < 0, g A = min(−(g C + g L )/φ, (g D + g L )/(1 − φ)). A onetime increase to C t , D t , L t , or S, as long as S remains below 1, does not affect the output technology growth rate. A permanent increase to g C , g D , or g L , on the other hand, does increase the growth rate of output technology and thereby output per capita. • If φ = 0, the growth path constrained by effective labor growth remains defined as in the case of φ < 0. The above expression for the growth path constrained by effective capital accumulation, however, is now undefined. To find the capital-driven growth path with C and L fixed, observe that we here have g K = g A and sA t (1 − S)L = K t+1 − K t ⇒ g K,t = sA t (1 − S)L/K t , (47) g A,t ≈ CK t /A t = C(K t−1 /A t + s(1 − S)LA t−1 /A t ). ( 48 ) In steady state, K t−1 = K t /(1 + g A ), so K t−1 /A t = (K t /A t )/(1 + g A ). Also, g A = g K = s(1 − S)LA t /K t , so K t /A t = s(1 − S)L/g A ; so K t−1 /A t = s(1 − S)L/(g A (1 + g A )). Finally, A t−1 /A t = 1/(1 + g A ). Substituting these terms for K t−1 /A t and A t−1 /A t into (48), we find g A = sC(1 − S)L. ( 49 ) Thus if g C = g L = 0, technology grows at the minimum of this rate and the growth rate to which it is driven by labor-augmenting technology growth, which given φ = 0 is simply g D . Formally, g A = min( sC(1 − S)L), g D ). If g C > 0 or g L > 0, output technology growth cannot be constrained by capital accumulation, and g A = g D + g L . • If φ ∈ (0, 1), output technology growth cannot be constrained by capital accumulation; such a scenario would imply g A ∝ A φ−1 t C t K t ∝ A φ t C t L t , contradictorily producing superexponential growth in output technology, output, and capital. We have g A = (g D + g L )/(1 − φ). • If φ = 1, suppose labor remains fixed at L and labor-augmenting technology remains fixed at D, while effective capital accumulates. In the long run we then have g A = DSL. A one-time increase to D, S, or L increases the growth rate of output technology and thereby output (again, as long as S remains below 1). If we begin from a state in which g D = g L = 0 and introduce positive labor or labor-augmenting technology growth, the result is a Type I growth singularity. • If φ > 1, we have a Type II growth singularity regardless of the other parameters, as explained in §2.4. Recall from §4.2 that, given ρ < 0, increases to D can be interpreted as increases to the fraction of research tasks that have been automated. Suppose ρ = 0. Technology growth is then g A,t = A φ−1 t (C t K t ) a (D t SL t ) 1−a (50) = A φ−1 0 (C 0 K 0 ) a (D 0 SL 0 ) 1−a e mt , where m (g A (φ − 1) + a(g C + g K ) + (1 − a)(g D + g L )). ( 51 ) From our assumption of a constant saving rate, g K = g Y = g A + g L . So m = g A (φ − 1 + a) + ag C + (1 − a)g D + g L . (52) Regarding φ: • If φ < 1 − a, we will have, in equilibrium, the constant output technology growth rate that sets m = 0. This is g A = (ag C + (1 − a)g D + g L )/(1 − φ − a) . One-time increases to C, D, S, or L do not change the growth rate, but increases to g C , g D , or g L do. • If φ = 1 − a, we have steady growth only if g C = g D = g L = 0, since we are assuming that these growth terms are all nonnegative. Fixing A 0 = K 0 = 1, the output technology growth rate is C a 0 (D 0 SL 0 ) 1−a . A one-time increase to C, D, S, or L increases the growth rate. If g C > 0, g D > 0, or g L > 0, we have a Type I growth singularity. • If φ > 1 − a, we have a Type II growth singularity regardless of the other parameters. Recall from §4.1 that increases to a can be interpreted as increases to the fraction of research tasks that have been automated. They can thus induce Type I and Type II growth singularities, if φ ∈ (0, 1). Suppose ρ > 0. Recall that in this case sustained growth in effective capital or in effective research labor suffice to sustain growth in output technology, and growth will be driven by whichever factor grows more quickly. Now, when growth is driven by effective capital accumulation, we have g A ∝ A φ−1 t C t K t , and when it is driven by effective research labor growth, we have g A ∝ A φ−1 t D t L t . Regarding φ: • If φ < 0, the capital-and labor-driven technology growth rates equal (g C +g K )/(1−φ) and (g D +g L )/(1−φ), respectively. In the former case, since g K = g Y = g A + g L , we can substitute for g K and rearrange to get g A = (g C + g L )/φ. Thus g A = max((gC + gL)/φ, (gD + gL)/(1 − φ)). Growth rate increases require increases to g C , g D , or g L . • If φ = 0, the growth path driven by effective labor growth remains defined as in the case of φ < 0. As found in the discussion of the ρ < 0 regime above, the growth rate on the growth path driven by effective capital growth when φ = 0 exhibits growth rate g A = sC(1 − S)L. Thus if g C = g L = 0, technology grows at the maximum of this rate and the growth rate to which it is driven by labor-augmenting technology growth, which given φ = 0 is simply g D . Formally, g A = max( sC(1 − S)L, g D ). • If φ > 0, we have a Type II growth singularity regardless of the other parameters. An intuition for these results is as follows. As explained briefly in §2.4, the growth of some variable X exhibits a Type II singularity if its growth rate takes the form g X ∝ X ψ for some ψ > 0. When ρ < 0, capital accumulation cannot accelerate technological development, which is bottlenecked by its slower-growing factor, namely effective labor. Output technology growth is then g A ∝ A φ−1 , so the Type II singularity requires φ > 1. When ρ > 0, on the other hand, capital accumulation at rate g Y = g A effectively multiplies g A by a factor of A, so g A ∝ A φ . The Type II singularity therefore requires only φ > 0. Note that our analysis of the ρ > 0 case also covers the case in which the development of output technology is fully automated. Simply use ρ = 1. It also covers a common interpretation of the possibility of \"recursive self-improvement\". If A represents cognitive ability, and enhanced intelligence (human or artificial) speeds the rate at which intelligence can be improved such that g A,t = A φ−1 t K t , then explosive growth obtains iff φ > 0 (since, again, g K = g A ). If we remove capital accumulation entirely and say that the development of intelligence depends only on the intelligence level, such that g A,t = A φ−1 t , then explosive growth obtains iff φ > 1. Now let us briefly examine the implications of automating good production in combination with automating research. Suppose we replace the laboronly good production function with a CES production function in capital and labor, with substitution parameter ρ Y . Fraction S L of labor and S K of capital is used in technology production. We will ignore factor-augmenting technology in the good production function; output technology A can be thought of as augmenting both, by CRS. Finally, we will assume sufficient saving. If ρ Y < 0, little changes. Output is still bottlenecked by the scarce factor, namely labor (assuming that we do not irrationally have S K,t → 1, in which case output is only the more constrained). Output therefore asymptotically resembles A t (1 − S L )L t , as before. If ρ Y > 0, output asymptotically resembles A t (1−S K )K t . The result is at least a Type I singularity, as long as g D + g L > 0, regardless of the values of ρ A and φ in output technology development. This is because output growth is here given by g Y,t = sA t (where s denotes the saving rate), and A t grows at least exponentially, driven by growth in effective researcher time. If ρ Y = 0, so that output is Cobb-Douglas in capital and labor, the occurrence of a growth singularity is sensitive to the capital exponent a Y . If we interpret increases in a Y as the automation of tasks in output production, as in §4.1, then automation can trigger a singularity in this way. We will not explore this further here. \n AI assistance in research In §4, we discussed several papers which use a microfoundation of the output production function as a basis for exploring the implications of a certain kind of automation. Somewhat analogously, Agrawal et al. (  2019 ) use  Weitzman's (1998)  microfoundation of the process of technological development as a basis for exploring the implications of a certain way in which advances in AI might assist in technological development. Let Y = A(1−S)L, as before, and hold S fixed but posit labor growth g L > 0. Given A existing \"technological ideas\", a researcher has access to only A φ , for some φ ∈ (0, 1), perhaps due to some sort of cognitive limitation. 26 New ideas are made from combinations of existing ideas. Given access to A φ ideas, a researcher therefore faces 2 A φ idea-combinations. Of these, not all can generate new technological ideas, perhaps due to some other sort of cognitive limitation. Instead, each researcher's idea-generation function is \"isoelastic\" in ideas available: Ȧ = α (2 A φ ) θ − 1 θ , θ > 0; Ȧ = α ln(2 A φ ) = α ln(2)A φ , θ = 0, (53) for some α > 0 and some θ ∈ [0, 1]. 27 Most transformatively, if AI tools help researchers search through the ever-growing \"haystacks\" of possible idea-combinations for valuable \"needles\", they could permanently increase θ.  Agrawal et al. (2019)  argue that this is precisely the sort of activity to which AI systems are best suited: they are already being profitably used to identify promising combinations of chemicals in pharmaceutical development, for example. (See  Agrawal et al. (2018)  for a more thorough defense of this argument.) If this turns out to hold across the board, the result is stark: as shown above, a permanent increase to θ produces a Type II growth singularity.  Agrawal et al. (2019)  also explore the potential impacts of AI assistance in research teams, rather than in assisting individual researchers.  Seeber et al. (2020)  do the same, in the context of a much more applied and less formal inquiry. Neither analysis appears to reveal channels for transformative growth effects substantively different from those presented above. \n Growth impacts via impacts on technology investment Throughout §5, we have taken technology production to be endogenous, in the sense that it has required explicit inputs of capital and labor. Nevertheless we have taken the level of investment in technological developmentthe fraction S of labor, and (in §5.2) the amount of capital, allocated to research-to be exogenous. A final way in which AI could have a transformative impact, therefore, is by changing the levels of investment in, and effort allocated to, technological development. As we have seen, at least in some circumstances, these changes can change the growth rate, or can determine the existence or type of a growth singularity. This pathway to transformative impact is somewhat analogous to the possibility, explored in §3.4, that developments in AI could affect the growth rate by affecting the saving rate, even in an economy without endogenous technological development. As in that case, this change could in principle be positive or negative. (Indeed, one way AI could impact the extent to which resources are devoted to technological development is by affecting the saving rate, as long as capital is modeled as an input to technology production.) Also as in that case, to the extent that the literature has explored this pathway to transformative impact, it has focused on the counterintuitive possibility that AI slows (though not, here, reverses) growth. This could take place by accelerating the \"Schumpeterian\" process of \"creative destruction\". On this analysis, the incentive to innovate comes from a temporary monopoly that the innovators enjoy, either by patents or by trade secrets, during which they can extract rents from those who would benefit by using the new technology in production. AI, however, could make it easier for competitors to copy innovations. Relatedly, AI could also ease the rapid development of technologies only negligibly more productivityenhancing than those they replace. Because these technologies would entirely eliminate the markets for those they replace, their rapid development would curtail the incentive for innovation. In the absence of this incentive, technology growth can slow to a halt. This cannot cause output per capita to fall, at least in most models, but it can cause output per capita to stagnate. This dynamic is explored more formally by  Aghion et al. (2019)  in the context of the model of automated research laid out in §5.2, and by  Acemoglu and Restrepo (2018b)  in the context of the model of automation and task replacement laid out in §4.4. We will not work through it here. As with the Sachs and Kotlikoff (2012) observation that AI can do damage by lowering the saving rate, the insight here is not primarily an insight about artificial intelligence. It is primarily a special case of the well-known fact, mentioned briefly in §2.4, that though free and competitive markets can generally be expected to appropriately compensate production factors for a final good in a static setting, the same cannot be said about the inputs to technological development. Policymakers interested in growth must always consider the impact of structural economic changes on the incentives for technological innovation, therefore, and must adjust their funding or subsidization of basic research in light of such changes as they unfold. \n Overview of the possibilities The table below summarizes the transformative scenarios we have considered. They have been rearranged slightly for clarity, and some near-redundant possibilities have been removed, but they primarily follow the order in which they are presented in §3-5. Relevant literature is cited below each scenario. Note that, in keeping with the presentation so far, the cited literature introduces the models that allow for the scenarios in question, but does not always discuss the transformative scenarios on which we have focused. We have not considered all possible AI scenarios, as this table makes clear. Nevertheless we have hopefully sampled the possibilities thoroughly enough that the reader is now comfortable filling some of the gaps. \n Scenario 28 Growth PS in production & capital-augmenting tech growth I → 0 L 28 \"PS\", \"HS\", \"MS\", and \"LS\" stand for perfect, high, moderate, and low substitutability, and refer to substitution parameters ρ = 1, > 0, = 0, and < 0 respectively. Unless otherwise noted, the \"HS\" case allows for perfect substitutability. In the scenarios with endogenous growth, \"negative\", \"low\", \"intermediate\", and \"high [research] feedback\" refer to research/learning feedback exponents φ < 0, < 1, = 1, and > 1. 29 + andrefer to cases in which AI shifts the output path up or down without changing the growth rate, e.g. by increasing or decreasing the plateau level in a circumstance where output plateaus regardless of AI. --, ++, I, and II refer to cases in which AI allows for decreases to the long-run growth rate, increases to the long-run growth rate, Type I growth singularities, and Type II growth singularities. = refers to cases in which AI does not change the long-run output level or growth rate. 30 C means that AI pushes the human labor share to some positive constant, not necessarily lower or higher than the value it would take in the absence of AI. 31 L means that human wages are driven to some low but constant rate (typically the rental rate of effective capital). C means that they are pushed to some positive constant, not necessarily lower or higher than they would be in the absence of AI. All other symbols are defined as in the Growth column. The human labor share and wage are technically undefined in the models of endogenous technology production, since, as noted in §2.4, we cannot straightforwardly assume that the factors of technology production will tend to be paid their marginal products (or anything else in particular). As typically presented, however, human labor is the lone factor of final good production in these models, and the technology being produced is laboraugmenting. Barring any radical market interventions, therefore, the wage should therefore grow in line with technology and so with output. That is, it should exhibit growth rate decreases, increases, Type I singularities or Type II singularities as listed above. \n Conclusion The set of models discussed here cover a wide range of AI's possible longrun macroeconomic impacts. It can hopefully serve as a bridge between the tools of economics-whose use is typically restricted to shorter-term and smaller-scale possibilities (but need not be)-and the longer-term and largerscale questions posed by futurists, who typically do not draw on the tools of economics (but could, we believe, sometimes learn from doing so). Nevertheless, of course, many topics relevant to the economics of AI, and even of transformative AI, could not be covered here. Wage distribution is a-perhaps even the-central concern of the literature on the economics of AI, including much of the literature cited here. It is likewise a central concern of the less long-term-focused reviews of the economics of AI cited in §1. Indeed, wages and skill levels are of course empirically highly unequal. And this inequality has indeed increased in the recent past, a development many attribute to the rise of automation. Nevertheless, we have consistently referred to all wages and human abilities as homogeneous. The choice to focus on average wages and on the overall labor share is motivated in part by the supposition that, if we are truly considering the long run, the likeliest transformative possibility is that AI will outsmart us all. In this event, human talents will not save us; if we retain positive wages or a positive labor share, we will do so only because AI is put to use making us more productive, or because some tasks, like those performed by clergy or by hospice nurses, remain resistant to automation. Otherwise, as  Freeman (2015)  colorfully puts it, \"[w]ithout ownership stakes, workers will become serfs working on behalf of robots' overlords\". To be clear, however, this view may by all means be incorrect. We cannot rule out AI-induced scenarios in which, even in the long run, income is concentrated not entirely in the hands of the robot owners but also at least to some extent in the hands of the most skilled human individuals. The subject most conspicuously present, despite receiving relatively little direct attention in economics literature, is the possibility of lasting changes to the growth regime: growth rate shifts and Type I and II singularities. As this review perhaps illustrates, the lack of attention currently given to these possibilities is not a necessary consequence of all plausible economic modeling. Rather, it is the result of a widespread norm of focusing only on model scenarios in which long-run growth is constant. Even  Aghion et al. (2019) , who take the singularitarian growth potential of AI most seriously, focus less on scenarios in which labor and capital are highly substitutable in technology production on the grounds that, as long as φ > 0, \"in this case researchers are not a necessary input and so standard capital accumulation is enough to generate explosive growth. This is one reason why the case of ρ < 0 . . . is the natural case to consider.\" Expressed motivations along these lines appear throughout the literature. Even outside discussions of AI, it is rare for economic growth literature to consider substantial and permanent growth rate increases, let alone growth explosions. This is presumably because such models would violate the \"Kaldor fact\" of constant exponential per-capita growth at 2-3% per year, which has roughly held in the industrialized world since roughly the Industrial Revolution. But on a longer timeframe, models of increasing growth would not be ahistorical; the growth rate was far lower before the Industrial Revolution, and before the Agricultural Revolution it was lower still. Empirical forecasts on the basis of these longer-run facts commonly predict radical future increases to growth, including substantial one-time rate increases and Type I and II singularities (see e.g.  Hanson (2000)  and  Roodman (2020) ). Some of the models underlying such forecasts lack economic foundations (e.g. that in  Hanson (2000) ), making it difficult to assess whether the forces driving historical superexponential growth are still at work today. Notably, however, some have economic foundations (e.g. that in  Kremer (1993) ) that imply that continued increases to the pool of effective workers or researchers, as advances in robotics and AI would permit, would continue the superexponential process. For a more thorough survey of the singularitarian literature, at least as of 2013, see  Sandberg (2013) . In short, if there were a sufficiently compelling theoretical reason to believe that the observed long-run trend of increasing growth will soon permanently halt, then we should of course dismiss transformative growth scenarios. But as many models of the economics of AI confirm, there is no shortage of mechanisms, once we allow ourselves to look for them, through which advances in automation could have transformative growth consequences. Futurists and economists interested in the long term might therefore do well to collaborate more on this point of common interest. Finally, the subject most conspicuously absent here that features most heavily in futurist discussion about AI is the most transformative macroeconomic possibility of all: the risk of an AI-induced existential catastrophe (see e.g.  Bostrom (2017) ). Unlike the possibility of transformative growth effects, AI risk appears to be absent from the economics literature not primarily \"by choice\" but because there is no particularly obvious mechanism through which accelerating automation or capital productivity, within existing models of production or growth, can pose a danger. It is possible, of course, to write down economic models in which production and/or technological development pose catastrophic risks in the abstract, as e.g.  Jones (2016)  and  Aschenbrenner (2020)  have done. As outlined above, however, the only growth-slowing AI possibilities economists have considered to date are those mediated by impacts on saving  (Sachs and Kotlikoff, 2012)  and on innovation incentives  (Aghion et al., 2019; Acemoglu and Restrepo, 2019) . These scenarios are very far from those that motivate most concern about AI risk. The latter typically feature superintelligent agents, with goals not fully aligned with ours, who take control of the world. The tools of economics can shed at least some light on these concerns as well. Most simply, to the extent that AI development poses such a risk, AI safety is a global and intergenerational public good. Through that lens, much of the analysis of public goods, and in particular many of the tools developed by environmental economists for the pricing and provision of climate risk mitigation, could apply to AI safety. More subtly, to the extent that AI risk arises from AIs' ability to control resources independently of human input, models in which the human labor share remains positive and significant should give us comfort. If human work remains a bottleneck to growth-say, if AI accelerates growth only by giving human workers instructions which they must physically perform-then humanity can in principle impoverish any robot overlords by going on strike. More worrying are models in which a unit of capital can grow, do research into capital-augmenting technology, and recursively self-improve all without human input. A thorough analysis of the links between the economics of AI and the issue of AI safety remains an important topic for further exploration. ,t + H/(DR t ). S K and S R are thus asymptotically constant and nonzero. Furthermore, sinceR t = f K,t S K,t K t + (H + S R,t DR t )/D, we must have H + S R,t DR t < DR t . It follows that S * and S  *  R are strictly less than 1. K 16  The model will approximate an AK model with A \n Aghion et al. (2019) ,Acemoglu and Restrepo (2018b)     §3.2 §4.4 Acemoglu and Restrepo (2018b) AI-diminished innovation incentives −− PS in production, capital-augmenting HS in production & task automation  §5.4 tech growth, & MS land constraint and creation ++ I → → 0 → 0 I §3.3 Hanson (2001)  §4.3 Hémous and Olsen (2014) PS in production, equipment-augmen-Learning by doing, w/intermed. ting tech gr., & MS land constraint feedback and/or automation ++ ++ → L §3.3  §5.1 Hanson (2001) PS in production & LS land constraint Learning by doing, with suffic. (regardless of tech) feedback and/or automation = II → → 0 §3.3 Korinek and Stiglitz (2019)  §5.1 Hanson (2001) HS in final good production, HS in LS in tech production, low research robotics production feedback, & asymptotic research task ++ ++ → L §3.4 Mookherjee and Ray (2017), Korinek and Stiglitz (2019) automation; or HS in tech production, negative research feedback, & research HS in final good production, LS in robotics production capital productivity growth  §5.2 Aghion et al. (2019) + C C §3.4 Mookherjee and Ray (2017), Korinek (2018) LS in tech production, intermed. PS in production & one-off capital-augmenting tech increase → zero research feedback, & research saving increase task automation; or HS in tech prod., research feedback & asymp. research ++ I → = §3.5 Korinek and Stiglitz (2019) capital prod. growth PS in production & capital-aug. tech  §5.2 Aghion et al. (2019) growth → saving decrease LS in tech production & high research − → → 0 §3.5 Sachs and Kotlikoff (2012), Sachs et al. (2015) feedback or HS in tech production & II MS in production & asymptotic or full positive research feedback task automation  §5.2 Aghion et al. (2019) I → I §4.2 Aghion et al. (2019) AI-assisted multiplication of LS in production & asymptotic task combinatorial idea discovery ++ automation  §5.3 Agrawal et al. (2019) ++ C ++ §4.2 Aghion et al. (2019) AI-assisted elasticity-change in idea LS in production & task automation discovery II and replacement  §5.3 Agrawal et al. (2019) ++ C ++ \n\t\t\t Sandberg (2013)  presented an \"overview of models of technological singularity\" before the past decade of economist engagement with AI and its transformative potential. Most of the models he summarizes therefore do not attempt to spell out how artificial intelligence, or indeed any particular transformative technology, would interact with standard economic models to produce the results in question. The models summarized here fill this gap.ECONOMIC GROWTH UNDER TRANSFORMATIVE AI \n\t\t\t More formally, it is assumed that F LK > 0, with subscripts here denoting partial derivatives (and that F KL = F LK > 0, by Young's Theorem). \n\t\t\t After accounting for capital depreciation, ρ may have to be significantly above 0 for capital accumulation to allow long-run growth. We will ignore capital depreciation throughout most of this document for simplicity.5 More generally, as should now be intuitive, when highly complementary production factors undergo different rates of accumulation or productivity growth, output's growth rate converges to that of the slowest-growing factor and its share goes converges to 1. Likewise with respect to complementary consumption goods, each requiring a different input. This is sometimes known as the \"Baumol condition\", afterBaumol's (e.g. Baumol  (1967)) seminal analyses of the increasing share of output spent on low-productivitygrowth sectors, such as live entertainment.6 The saving rate is in fact historically roughly constant, as famously observed by \n\t\t\t Indeed, some reserve the term \"endogenous\" for growth models in which φ = 1, since \n\t\t\t As Yudkowsky (2013)  points out, we might interpret this as a model in which AI comes in the form of \" emulations\"-a theoretical technology on which Hanson has written extensively-which are always technically feasible but which are, at first, prohibitively expensive, because effective capital is sufficiently scarce.13 Note that, were it not for the inclusion of the non-accumulable factor land, there would be no steady-state growth rate; in solving for it, we would have to divide by 0. Instead, the economy would be, asymptotically, an AK economy with exogenous capital productivity growth. As we saw in the previous section, we would have a Type I growth singularity. \n\t\t\t Equivalently, we could say that output is produced by a single factor, labor, which can be supplied both by humans and by robots. \n\t\t\t Technically, if flow utility is logarithmic in consumption (see footnote 18), lifetime utility falls to negative infinity as the consumption of the young falls to zero. \n\t\t\t A = a -a (1-a) a-1 , which ranges from 1 (at a = 0 or 1) to 2 (at a = 1/2).ECONOMIC GROWTH UNDER TRANSFORMATIVE AI \n\t\t\t Note that, in this case, effective capital growth g A + g Y will always equal gL ((1 − a)(1 − φ) + λ(1 − a))/((1 − a)(1 − φ) − λa) > g L .With effective capital growing more quickly than labor, the automation condition will always eventually be met for any fixed a. \n\t\t\t The model requires φ > 0 such that the fishing-out effect does not predominate. As discussed in §2.4,Bloom et al. (2020)  estimate φ = −2.1.27  The formula for θ = 0 is the limiting case of the formula for θ > 0, as θ → 0. \n\t\t\t Zeira, Joseph, \"Workers, Machines, and Economic Growth,\" The Quarterly Journal ofEconomics, 1998, 113 (4), <PHONE>.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Philip-Trammell-and-Anton-Korinek_economic-growth-under-transformative-ai.tei.xml", "id": "60980b96662675720e8e7183263e9436"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Contents We define drones as unmanned aerial robots, which may or may not have autonomous decision-making features.", "authors": ["Sankalp Bhatnagar", "Talia Cotton"], "title": "The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation", "text": "We structure our analysis by separately considering three security domains, and illustrate possible changes to threats within these domains through representative examples: • Digital security. The use of AI to automate tasks involved in carrying out cyberattacks will alleviate the existing tradeoff between the scale and efficacy of attacks. This may expand the threat associated with labor-intensive cyberattacks (such as spear phishing). We also expect novel attacks that exploit human vulnerabilities (e.g. through the use of speech synthesis for impersonation), existing software vulnerabilities (e.g. through automated hacking), or the vulnerabilities of AI systems (e.g. through adversarial examples and data poisoning). • Physical security. The use of AI to automate tasks involved in carrying out attacks with drones and other physical systems (e.g. through the deployment of autonomous weapons systems) may expand the threats associated with these attacks. We also expect novel attacks that subvert cyberphysical systems (e.g. causing autonomous vehicles to crash) or involve physical systems that it would be infeasible to direct remotely (e.g. a swarm of thousands of micro-drones). • Political security. The use of AI to automate tasks involved in surveillance (e.g. analysing mass-collected data), persuasion (e.g. creating targeted propaganda), and deception (e.g. manipulating videos) may expand threats associated with privacy invasion and social manipulation. We also expect novel attacks that take advantage of an improved capacity to analyse human behaviors, moods, and beliefs on the basis of available data. These concerns are most significant in the context of authoritarian states, but may also undermine the ability of democracies to sustain truthful public debates. \n p.7 \n Executive Summary The Malicious Use of Artificial Intelligence In addition to the high-level recommendations listed above, we also propose the exploration of several open questions and potential interventions within four priority research areas: • Learning from and with the cybersecurity community. At the intersection of cybersecurity and AI attacks, we highlight the need to explore and potentially implement red teaming, formal verification, responsible disclosure of AI vulnerabilities, security tools, and secure hardware. • Exploring different openness models. As the dual-use nature of AI and ML becomes apparent, we highlight the need to reimagine norms and institutions around the openness of research, starting with pre-publication risk assessment in technical areas of special concern, central access licensing models, sharing regimes that favor safety and security, and other lessons from other dual-use technologies. • Promoting a culture of responsibility. AI researchers and the organisations that employ them are in a unique position to shape the security landscape of the AI-enabled world. We highlight the importance of education, ethical statements and standards, framings, norms, and expectations. • Developing technological and policy solutions. In addition to the above, we survey a range of promising technologies, as well as policy interventions, that could help build a safer future with AI. High-level areas for further research include privacy protection, coordinated use of AI for public-good security, monitoring of AI-relevant resources, and other legislative and regulatory responses. The proposed interventions require attention and action not just from Introduction p.9 Artificial intelligence (AI) and machine learning (ML) have progressed rapidly in recent years, and their development has enabled a wide range of beneficial applications. For example, AI is a critical component of widely used technologies such as automatic speech recognition, machine translation, spam filters, and search engines. Additional promising technologies currently being researched or undergoing small-scale pilots include driverless cars, digital assistants for nurses and doctors, and AI-enabled drones for expediting disaster relief operations. Even further in the future, advanced AI holds out the promise of reducing the need for unwanted labor, greatly expediting scientific research, and improving the quality of governance. We are excited about many of these developments, though we also urge attention to the ways in which AI can be used maliciously . We analyze such risks in detail so that they can be prevented or mitigated, not just for the value of AI refers to the use of digital technology to create systems that are capable of performing tasks commonly thought to require intelligence. Machine learning is variously characterized as either a subfield of AI or a separate field, and refers to the development of digital systems that improve their performance on a given task over time through experience. We define \"malicious use\" loosely, to include all practices that are intended to compromise the security of individuals, groups, or a society. Note that one could read much of our document under various possible perspectives on what constitutes malicious use, as the interventions and structural issues we discuss are fairly general. \n p.1 0 Introduction preventing the associated harms, but also to prevent delays in the realization of the beneficial applications of AI. Artificial intelligence (AI) and machine learning (ML) are altering the landscape of security risks for citizens, organizations, and states. Malicious use of AI could threaten digital security (e.g. through criminals training machines to hack or socially engineer victims at human or superhuman levels of performance), physical security (e.g. non-state actors weaponizing consumer drones), and political security (e.g. through privacy-eliminating surveillance, profiling, and repression, or through automated and targeted disinformation campaigns). The malicious use of AI will impact how we construct and manage our digital infrastructure as well as how we design and distribute AI systems, and will likely require policy and other institutional responses. The question this report hopes to answer is: how can we forecast, prevent, and (when necessary) mitigate the harmful effects of malicious uses of AI? We convened a workshop at the University of Oxford on the topic in February 2017, bringing together experts on AI safety, drones , cybersecurity, lethal autonomous weapon systems, and counterterrorism . This document summarizes the findings of that workshop and our conclusions after subsequent research. \n Scope For the purposes of this report, we only consider AI technologies that are currently available (at least as initial research and development demonstrations) or are plausible in the next 5 years, and focus in particular on technologies leveraging machine learning. We only consider scenarios where an individual or an organisation deploys AI technology or compromises an AI system with an aim to undermine the security of another individual, organisation or collective. Our work fits into a larger body of work on the social implications of, and policy responses to, AI . There has thus far been more attention paid in this work to unintentional forms of AI misuse such as algorithmic bias , versus the intentional undermining of individual or group security that we consider. We exclude indirect threats to security from the current report, such as threats that could come from mass unemployment, or other second-or third-order effects from the deployment of AI technology in human society. We also exclude system-level threats that would come from the dynamic interaction between non-malicious actors, such as a \"race to the bottom\" on AI safety Not all workshop participants necessarily endorse all the findings discussed herein. See Appendix A for additional details on the workshop and research process underlying this report. \n p.11 Introduction between competing groups seeking an advantage or conflicts spiraling out of control due to the use of ever-faster autonomous weapons. Such threats are real, important, and urgent, and require further study, but are beyond the scope of this document. \n Related Literature Though the threat of malicious use of AI has been highlighted in high-profile settings (e.g. in a Congressional hearing a White House-organized workshop , and a Department of Homeland Security report ), and particular risk scenarios have been analyzed (e.g. the subversion of military lethal autonomous weapon systems ), the intersection of AI and malicious intent writ large has not yet been analyzed comprehensively. Several literatures bear on the question of AI and security, including those on cybersecurity, drones, lethal autonomous weapons, \"social media bots,\" and terrorism. Another adjacent area of research is AI safety-the effort to ensure that AI systems reliably achieve the goals their designers and users intend without causing unintended harm . Whereas the AI safety literature focuses on unintended harms related to AI, we focus on the intentional use of AI to achieve harmful outcomes (from the victim's point of view). A recent report covers similar ground to our analysis, with a greater focus on the implications of AI for U.S. national security. In the remainder of the report, we first provide a high-level view on the nature of AI and its security implications in the section General Framework for AI and Security, with subsections on Capabilities, Security-relevant Properties of AI, and General Implications for the Security Landscape; we then illustrate these characteristics of AI with Scenarios in which AI systems could be used maliciously; we next analyze how AI may play out in the domains of digital, physical, and political security; we propose Interventions to better assess these risks, protect victims from attacks, and prevent malicious actors from accessing and deploying dangerous AI capabilities; and we conduct a Strategic Analysis of the \"equilibrium\" of a world in the medium-term (5+ years) after more sophisticated attacks and defenses have been implemented. Appendices A and B respectively discuss the workshop leading up to this report, and describe areas for research that might yield additional useful interventions.  \n AI Capabilities The field of AI aims at the automation of a broad range of tasks. Typical tasks studied by AI researchers include playing games, guiding vehicles, and classifying images. In principle, though, the set of tasks that could be transformed by AI is vast. At minimum, any task that humans or non-human animals use their intelligence to perform could be a target for innovation. While the field of artificial intelligence dates back to the 1950s, several years of rapid progress and growth have recently invested it with a greater and broader relevance. Researchers have achieved sudden performance gains at a number of their most commonly studied tasks. Factors that help to explain these recent gains include the exponential growth of computing power, improved machine learning algorithms (especially in the area of deep neural networks), development of standard software frameworks for faster iteration and replication of experiments, larger and more widely available datasets, and expanded commercial investments  (Jordan and Mitchell, 2015) . \n p.13 General Framework for AI & Security Threats Figure 1 illustrates this trend in the case of image recognition, where over the past half-decade the performance of the best AI systems has improved from correctly categorizing around 70% of images to near perfect categorization (98%), better than the human benchmark of 95% accuracy. Even more striking is the case of image generation. As Figure  2  shows, AI systems can now produce synthetic images that are nearly indistinguishable from photographs, whereas only a few years ago the images they produced were crude and obviously unrealistic. AI systems are also beginning to achieve impressive performance in a range of competitive games, ranging from chess to Atari to Go to e-sports like Dota 2 . Even particularly challenging tasks within these domains, such as the notoriously difficult Atari game Montezuma's Revenge, are beginning to yield to novel AI techniques that creatively search for successful long-term strategies , learn from auxiliary rewards such as feature control , and learn from a handful of human demonstrations . Other task areas associated with significant recent progress include speech recognition, language comprehension, and vehicle navigation. From a security perspective, a number of these developments are worth noting in their own right. For instance, the ability to recognize a target's face and to navigate through space can be applied in autonomous weapon systems. Similarly, the ability to generate synthetic images, text, and audio could be used to impersonate others online, or to sway public opinion by distributing AI-generated content through social media channels. We discuss these applications of AI further in the Security Domains section. These technical developments can also be viewed as early indicators of the potential of AI. The techniques used to achieve high levels of performance on the tasks listed above have only received significant attention from practitioners in the past decade and are often quite general purpose. It will not be surprising if AI systems soon become competent at an even wider variety of security-relevant tasks. At the same time, we should not necessarily expect to see significant near-term progress on any given task. Many research areas within AI, including much of robotics, have not changed nearly so dramatically over the past decade. Similarly, the observation that some of the most commonly studied tasks have been associated with rapid progress is not necessarily as significant as it first seems: these tasks are often widely studied in the first place because they are particularly tractable. \n On page 18 On page 19  Mnih et al., 2015 Silver and Huang et al., 2016; Silver, Schrittwieser, and Simonyan et al., 2016 OpenAI, 2017a; OpenAI, 2017b Vezhnevets et al., 2017 Jaderberg et al., 2016 Hester et al., 2017  To aid one's predictions, it can useful to note some systematic difference between tasks which contemporary AI systems are wellsuited to and tasks for which they still fall short. In particular, a task is likely to be promising if a perfect mathematical model or simulation of the task exists, if short-term signals of progress are available, if abundant data on the successful performance of that task by humans is available, or if the solution to the task doesn't require a broader world-model or ``common sense''. \n p.1 4 General Framework for AI & Security Threats Figure  2 : Increasingly realistic synthetic faces generated by variations on Generative Adversarial Networks (GANs). In order, the images are from papers by  Goodfellow et al. (2014) ,  Radford et al. (2015) ,  Liu and Tuzel (2016) , and  Karras et al. (2017) . 2014 2015 2016 2017 p.1 6 Finally, a few things should be said about the long-term prospects for progress in artificial intelligence. Today, AI systems perform well on only a relatively small portion of the tasks that humans are capable of. However, even before the recent burst of progress, this portion has expanded steadily over time . In addition, it has often been the case that once AI systems reach human-level performance at a given task (such as chess) they then go on to exceed the performance of even the most talented humans. Nearly all AI researchers in one survey expect that AI systems will eventually reach and then exceed human-level performance at all tasks surveyed. Most believe this transition is more likely than not to occur within the next fifty years. The implications of such a transition, should it occur, are difficult to conceptualize, and are outside the primary scope of this report (see Scope, though we briefly revisit this topic in the Conclusion). Nevertheless, one might expect AI systems to play central roles in many security issues well before they are able to outperform humans at everything, in the same way that they are already finding economic applications despite not being able to automate most aspects of humans' jobs. \n Security-Relevant Properties of AI AI is a dual-use area of technology. AI systems and the knowledge of how to design them can be put toward both civilian and military uses, and more broadly, toward beneficial and harmful ends. Since some tasks that require intelligence are benign and other are not, artificial intelligence is dual-use in the same sense that human intelligence is. It may not be possible for AI researchers simply to avoid producing research and systems that can be directed towards harmful ends (though in some cases, special caution may be warranted based on the nature of the specific research in question -see Interventions). Many tasks that it would be beneficial to automate are themselves dual-use. For example, systems that examine software for vulnerabilities have both offensive and defensive applications, and the difference between the capabilities of an autonomous drone used to deliver packages and the capabilities of an autonomous drone used to deliver explosives need not be very great. In addition, foundational research that aims to increase our understanding of AI, its capabilities and our degree of control over it, appears to be inherently dual-use in nature. AI systems are commonly both efficient and scalable. Here, we say an AI system is \"efficient\" if, once trained and deployed, it can complete a certain task more quickly or cheaply than a human could . We say an AI system is \"scalable\" if, given that it can complete a certain task, increasing the computing power it has access to or making copies of the system would allow it  Grace et al., 2017  Although trends in performance across a range of domains have historically not been comprehensively tracked or well theorized  (Brundage, 2016; Hernández-Orallo, 2017) , there have been some recent efforts to track, measure, and compare performance  (Eckersley and Nasser et al., 2017) . We distinguish here between task efficiency of a trained system, which commonly exceeds human performance, and training efficiency: the amount of time, computational resources and data, that a system requires in order to learn to perform well on a task. Humans still significantly exceed AI systems in terms of training efficiency for most tasks. \n p.17 General Framework for AI & Security Threats to complete many more instances of the task. For example, a typical facial recognition system is both efficient and scalable; once it is developed and trained, it can be applied to many different camera feeds for much less than the cost of hiring human analysts to do the equivalent work. AI systems can exceed human capabilities. In particular, an AI system may be able to perform a given task better than any human could. For example, as discussed above, AI systems are now dramatically better than even the top-ranked players at games like chess and Go. For many other tasks, whether benign or potentially harmful, there appears to be no principled reason why currently observed human-level performance is the highest level of performance achievable, even in domains where peak performance has been stable throughout recent history, though as mentioned above some domains are likely to see much faster progress than others. AI systems can increase anonymity and psychological distance. Many tasks involve communicating with other people, observing or being observed by them, making decisions that respond to their behavior, or being physically present with them. By allowing such tasks to be automated, AI systems can allow the actors who would otherwise be performing the tasks to retain their anonymity and experience a greater degree of psychological distance from the people they impact . For example, someone who uses an autonomous weapons system to carry out an assassination, rather than using a handgun, avoids both the need to be present at the scene and the need to look at their victim. AI developments lend themselves to rapid diffusion. While attackers may find it costly to obtain or reproduce the hardware associated with AI systems, such as powerful computers or drones, it is generally much easier to gain access to software and relevant scientific findings. Indeed, many new AI algorithms are reproduced in a matter of days or weeks. In addition, the culture of AI research is characterized by a high degree of openness, with many papers being accompanied by source code. If it proved desirable to limit the diffusion of certain developments, this would likely be difficult to achieve (though see Interventions for discussion of possible models for at least partially limiting diffusion in certain cases). Today's AI systems suffer from a number of novel unresolved vulnerabilities. These include data poisoning attacks (introducing training data that causes a learning system to make mistakes ), adversarial examples (inputs designed to be misclassified by machine learning systems ), and the exploitation of flaws in the design of autonomous systems' goals . These vulnerabilities \n p.1 8 General Framework for AI & Security Threats are distinct from traditional software vulnerabilities (e.g. buffer overflows) and demonstrate that while AI systems can exceed human performance in many ways, they can also fail in ways that a human never would. General Implications for the Threat Landscape From the properties discussed above, we derive three high-level implications of progress in AI for the threat landscape. Absent the development of adequate defenses, progress in AI will: • Expand existing threats • Introduce new threats • Alter the typical character of threats In particular, we expect attacks to typically be more effective, more finely targeted, more difficult to attribute, and more likely to exploit vulnerabilities in AI systems. These shifts in the landscape necessitate vigorous responses of the sort discussed under Interventions. \n Expanding Existing Threats For many familiar attacks, we expect progress in AI to expand the set of actors who are capable of carrying out the attack, the rate at which these actors can carry it out, and the set of plausible targets. This claim follows from the efficiency, scalability, and ease of diffusion of AI systems. In particular, the diffusion of efficient AI systems can increase the number of actors who can afford to carry out particular attacks. If the relevant AI systems are also scalable, then even actors who already possess the resources to carry out these attacks may gain the ability to carry them out at a much higher rate. Finally, as a result of these two developments, it may become worthwhile to attack targets that it otherwise would not make sense to attack from the standpoint of prioritization or costbenefit analysis. One example of a threat that is likely to expand in these ways, discussed at greater length below, is the threat from spear phishing attacks . These attacks use personalized messages to extract sensitive information or money from individuals, with the A phishing attack is an attempt to extract information or initiate action from a target by fooling them with a superficially trustworthy facade. A spear phishing attack involves collecting and using information specifically relevant to the target (e.g. name, gender, institutional affiliation, topics of interest, etc.), which allows the facade to be customized to make it look more relevant or trustworthy. \n p.1 9 General Framework for AI & Security Threats attacker often posing as one of the target's friends, colleagues, or professional contacts. The most advanced spear phishing attacks require a significant amount of skilled labor, as the attacker must identify suitably high-value targets, research these targets' social and professional networks, and then generate messages that are plausible within this context. If some of the relevant research and synthesis tasks can be automated, then more actors may be able to engage in spear phishing. For example, it could even cease to be a requirement that the attacker speaks the same language as their target. Attackers might also gain the ability to engage in mass spear phishing, in a manner that is currently infeasible, and therefore become less discriminate in their choice of targets. Similar analysis can be applied to most varieties of cyberattacks, as well as to threats to physical or political security that currently require non-trivial human labor. Progress in AI may also expand existing threats by increasing the willingness of actors to carry out certain attacks. This claim follows from the properties of increasing anonymity and increasing psychological distance. If an actor knows that an attack will not be tracked back to them, and if they feel less empathy toward their target and expect to experience less trauma, then they may be more willing to carry out the attack. The importance of psychological distance, in particular, is illustrated by the fact that even military drone operators, who must still observe their targets and \"pull the trigger,\" frequently develop post-traumatic stress from their work . Increases in psychological distance, therefore, could plausibly have a large effect on potential attackers' psychologies. We should also note that, in general, progress in AI is not the only force aiding the expansion of existing threats. Progress in robotics and the declining cost of hardware, including both computing power and robots, are important too, and discussed further below. For example, the proliferation of cheap hobbyist drones, which can easily be loaded with explosives, has only recently made it possible for non-state groups such as the Islamic State to launch aerial attacks . \n Introducing New Threats Progress in AI will enable new varieties of attacks. These attacks may use AI systems to complete certain tasks more successfully than any human could, or take advantage of vulnerabilities that AI systems have but humans do not. \n p.20 General Framework for AI & Security Threats First, the property of being unbounded by human capabilities implies that AI systems could enable actors to carry out attacks that would otherwise be infeasible. For example, most people are not capable of mimicking others' voices realistically or manually creating audio files that resemble recordings of human speech. However, there has recently been significant progress in developing speech synthesis systems that learn to imitate individuals' voices (a technology that's already being commercialized ). There is no obvious reason why the outputs of these systems could not become indistinguishable from genuine recordings, in the absence of specially designed authentication measures. Such systems would in turn open up new methods of spreading disinformation and impersonating others . In addition, AI systems could also be used to control aspects of the behavior of robots and malware that it would be infeasible for humans to control manually. For example, no team of humans could realistically choose the flight path of each drone in a swarm being used to carry out a physical attack. Human control might also be infeasible in other cases because there is no reliable communication channel that can be used to direct the relevant systems; a virus that is designed to alter the behavior of air-gapped computers, as in the case of the 'Stuxnet' software used to disrupt the Iranian nuclear program, cannot receive commands once it infects these computers. Restricted communication challenges also arise underwater and in the presence of signal jammers, two domains where autonomous vehicles may be deployed. Second, the property of possessing unresolved vulnerabilities implies that, if an actor begins to deploy novel AI systems, then they may open themselves up to attacks that specifically exploit these vulnerabilities. For example, the use of self-driving cars creates an opportunity for attacks that cause crashes by presenting the cars with adversarial examples. An image of a stop sign with a few pixels changed in specific ways, which humans would easily recognize as still being an image of a stop sign, might nevertheless be misclassified as something else entirely by an AI system. If multiple robots are controlled by a single AI system run on a centralized server, or if multiple robots are controlled by identical AI systems and presented with the same stimuli, then a single attack could also produce simultaneous failures on an otherwise implausible scale. A worst-case scenario in this category might be an attack on a server used to direct autonomous weapon systems, which could lead to large-scale friendly fire or civilian targeting . \n p.21 General Framework for AI & Security Threats Altering the Typical Character of Threats Our analysis so far suggests that the threat landscape will change both through expansion of some existing threats and the emergence of new threats that do not yet exist. We also expect that the typical character of threats will shift in a few distinct ways. In particular, we expect the attacks supported and enabled by progress in AI to be especially effective, finely targeted, difficult to attribute, and exploitative of vulnerabilities in AI systems. First, the properties of efficiency, scalability, and exceeding human capabilities suggest that highly effective attacks will become more typical (at least absent substantial preventive measures). Attackers frequently face a trade-off between the frequency and scale of their attacks, on the one hand, and their effectiveness on the other . For example, spear phishing is more effective than regular phishing, which does not involve tailoring messages to individuals, but it is relatively expensive and cannot be carried out en masse. More generic phishing attacks manage to be profitable despite very low success rates merely by virtue of their scale. By improving the frequency and scalability of certain attacks, including spear phishing, AI systems can render such trade-offs less acute. The upshot is that attackers can be expected to conduct more effective attacks with greater frequency and at a larger scale. The expected increase in the effectiveness of attacks also follows from the potential of AI systems to exceed human capabilities. Second, the properties of efficiency and scalability, specifically in the context of identifying and analyzing potential targets, also suggest that finely targeted attacks will become more prevalent. Attackers often have an interest in limiting their attacks to targets with certain properties, such as high net worth or association with certain political groups, as well as an interest in tailoring their attacks to the properties of their targets. However, attackers often face a trade-off between how efficient and scalable their attacks are and how finely targeted they are in these regards. This trade-off is closely related to the trade-off with effectiveness, as discussed, and the same logic implies that we should expect it to become less relevant. An increase in the relative prevalence of spear phishing attacks, compared to other phishing attacks, would be an example of this trend as well. An alternative example might be the use of drone swarms that deploy facial recognition technology to kill specific members of crowds, in place of less finely targeted forms of violence. Third, the property of increasing anonymity suggests that difficultto-attribute attacks will become more typical. An example, again, \n p.22 General Framework for AI & Security Threats is the case of an attacker who uses an autonomous weapons system to carry out an attack rather than carrying it out in person. Finally, we should expect attacks that exploit the vulnerabilities of AI systems to become more typical. This prediction follows directly from the unresolved vulnerabilities of AI systems and the likelihood that AI systems will become increasingly pervasive. \n Scenarios p.23 The following scenarios are intended to illustrate a range of plausible uses toward which AI could be put for malicious ends, in each of the domains of digital, physical, and political security. Examples have been chosen to illustrate the diverse ways in which the security-relevant characteristics of AI introduced above could play out in different contexts. These are not intended to be definitive forecasts (some may not end up being technically possible in 5 years, or may not be realized even if they are possible) or exhaustive (other malicious uses will undoubtedly be invented that we do not currently foresee). Additionally some of these are already occurring in limited form today, but could be scaled up or made more powerful with further technical advances. \n p.24 \n Digital Security Automation of social engineering attacks. Victims' online information is used to automatically generate custom malicious websites/emails/links they would be likely to click on, sent from addresses that impersonate their real contacts, using a writing style that mimics those contacts. As AI develops further, convincing chatbots may elicit human trust by engaging people in longer dialogues, and perhaps eventually masquerade visually as another person in a video chat. individuals or small groups can do: e.g. one person launching an attack with many weaponized autonomous drones. Swarming attacks. Distributed networks of autonomous robotic systems, cooperating at machine speed, provide ubiquitous surveillance to monitor large areas and groups and execute rapid, coordinated attacks. Attacks further removed in time and space. Physical attacks are further removed from the actor initiating the attack as a result of autonomous operation, including in environments where remote communication with the system is not possible. \n Political Security State use of automated surveillance platforms to suppress dissent. State surveillance powers of nations are extended by automating image and audio processing, permitting the collection, processing, and exploitation of intelligence information at massive scales for myriad purposes, including the suppression of debate. \n p.29 Fake news reports with realistic fabricated video and audio. Highly realistic videos are made of state leaders seeming to make inflammatory comments they never actually made. Automated, hyper-personalised disinformation campaigns. Individuals are targeted in swing districts with personalised messages in order to affect their voting behavior. Automating influence campaigns. AI-enabled analysis of social networks are leveraged to identify key influencers, who can then be approached with (malicious) offers or targeted with disinformation. Denial-of-information attacks. Bot-driven, large-scale informationgeneration attacks are leveraged to swamp information channels with noise (false or merely distracting information), making it more difficult to acquire real information. Manipulation of information availability. Media platforms' content curation algorithms are used to drive users towards or away from certain content in ways to manipulate user behavior. \n p.30 Security Domains Here, we analyze malicious uses of AI that would compromise the confidentiality, integrity, and availability of digital systems (threats to Digital Security); attacks taking place in the physical world directed at humans or physical infrastructure (threats to Physical Security); and the use of AI to threaten a society's ability to engage in truthful, free, and productive discussions about matters of public importance and legitimately implement broadly just and beneficial policies (threats to Political Security). These categories are not mutually exclusive-for example, AI-enabled hacking can be directed at cyber-physical systems with physical harm resulting as a consequence, and physical or digital attacks could be carried out for political purposes-but they provide a useful structure for our analysis. \n 3 Defined as \"engineered systems that are built from, and depend upon, the seamless integration of computational algorithms and physical components\" (National Science Foundation, 2017). \n p.31 Security Domains In each domain of security, we summarize the existing state of play of attack and defense prior to wide adoption of AI in these domains, and then describe possible changes to the nature or severity of attacks that may result from further AI progress and diffusion. The three sections below all draw on the insights discussed above regarding the security-relevant properties of AI, but can be read independently of one another, and each can be skipped by readers less interested in a particular domain. \n Digital Security Absent preparation, the straightforward application of contemporary and near-term AI to cybersecurity offense can be expected to increase the number, scale, and diversity of attacks that can be conducted at a given level of capabilities, as discussed more abstractly in the General Framework for AI and Security Threats above. AI-enabled defenses are also being developed and deployed in the cyber domain, but further technical and policy innovations (discussed further in Interventions) are needed to ensure that impact of AI on digital systems is net beneficial. \n Context Cybersecurity is an arena that will see early and enthusiastic deployment of AI technologies, both for offense and defense; indeed, in cyber defense, AI is already being deployed for purposes such as anomaly and malware detection. Consider the following: • Many important IT systems have evolved over time to be sprawling behemoths, cobbled together from multiple different systems, under-maintained and -as a consequence -insecure. Because cybersecurity today is largely labor-constrained , it is ripe with opportunities for automation using AI. Increased use of AI for cyber defense, however, may introduce new risks, as discussed below. • In recent years, various actors have sought to mount increasingly sophisticated cyberoperations, including finely targeted attacks from state actors (including the Stuxnet Worm and the Ukrainian power grid \"crash override\" exploit). The cyber arena also includes a vast and complex world of cybercrime , which sometimes involves a high degree of professionalization and organization . Such groups use DDoS, malware, phishing, ransomware, and other forms of Already, AI is being widely used on the defensive side of cybersecurity, making certain forms of defense more effective and scalable, such as spam and malware detection. At the same time, many malicious actors have natural incentives to experiment with using AI to attack the typically insecure systems of others. These incentives include a premium on speed, labor costs, and difficulties in attracting and retaining skilled labor. To date, the publicly-disclosed use of AI for offensive purposes has been limited to experiments by \"white hat\" researchers, who aim to increase security through finding vulnerabilities and suggesting solutions. However, the pace of progress in AI suggests the likelihood of cyber attacks leveraging machine learning capabilities in the wild soon, if they have not done so already. Indeed, some popular accounts of AI and cybersecurity include claims based on circumstantial evidence that AI is already being used for offense by sophisticated and motivated adversaries . Expert opinion seems to agree that if this hasn't happened yet, it will soon: a recent survey of attendees at the Black Hat conference found 62% of respondents believing AI will be used for attacks within the next 12 months . Despite these claims, to our knowledge there is no publicly documented evidence of AI-based attacks, though it should be noted that evidence from many successful attacker techniques (e.g. botnets, email phishing campaigns) may be difficult to attribute to AI versus human labor or simple automation. We are thus at a critical moment in the co-evolution of AI and cybersecurity and should proactively prepare for the next wave of attacks. Many governments are keenly interested in the combination of AI and cybersecurity. In response to a question from one of the authors of this report, Admiral Mike Rogers, the Director of the National Security Agency, said, \"Artificial Intelligence and machine learning -I would argue -is foundational to the future of cybersecurity […] It is not the if, it's only the when to me.\" AI systems are already set to play an expanded role in US military strategy and operations in the coming years as the US DoD puts into practice its vision of a \"Third Offset\" strategy , in which humans and machines work closely together to achieve military objectives. At the same time, governments are investing in foundational research to expand the scope of capabilities of AI systems. In 2016, DARPA hosted the Cyber Grand Challenge contest , which saw teams of human researchers compete with each other to create programs that could autonomously attack other systems while defending themselves. Though the winning AI system fared poorly when facing off against human security experts, we agree with the hosts of the event that AI cybersecurity capabilities will improve rapidly in coming years, especially as recent advances in AI (such as in the area of deep reinforcement learning ) are applied to cybersecurity. \n How AI Changes The Digital Security Threat Landscape A central concern at the nexus of AI and cybersecurity is that AI might enable larger-scale and more numerous attacks to be conducted by an attacker with a given amount of skill and resources compared with the impact such an attacker might currently be able to achieve. Recent years have seen impressive and troubling proofs of concept of the application of AI to offensive applications in cyberspace. For example, researchers at ZeroFox demonstrated that a fully automated spear phishing system could create tailored tweets on the social media platform Twitter based on a user's demonstrated interests, achieving a high rate of clicks to a link that could be malicious . There is clearly interest in such larger-scale attacks: Russian hackers sent \"expertly tailored messages carrying malware to more than 10,000 Twitter users in the [U.S.] Defense Department\" , which likely required significant time and effort, and could have gone even further with automation (assuming it was not involved already in this case).  Giaretta and Dragoni (2017)  discuss the concept of \"community targeted spam\" that uses natural language generation techniques from AI to target an entire class of people with common ways of writing; with even more advanced natural language generation, one could envision even more customized approaches, spanning multiple communities. Furthermore, the application of AI to the automation of software vulnerability discovery, while having positive applications (discussed further in the Interventions section), can likewise be used for malicious purposes to alleviate the labor constraints of attackers. The adaptability of AI systems, too, may change the strategic landscape of cybersecurity, though it is not yet clear how adaptability will affect the offense/defense balance. Many organizations currently adopt security systems called Endpoint Detection and Response (EDR) platforms to counter more advanced threats. The EDR market represents a $500 million industry in the cyber security arena . These tools are built upon a combination of heuristic and machine learning algorithms to provide capabilities such as next-generation anti-virus (NGAV), \n p.34 Security Domains behavioral analytics, and exploit prevention against sophisticated targeted attacks. Though these systems are fairly effective against typical human-authored malware, research has already shown that AI systems may be able to learn to evade them. As an example of AI being used to avoid detection, Attackers are likely to leverage the growing capabilities of reinforcement learning, including deep reinforcement learning . In particular, we expect attackers to leverage the ability of AI to learn from experience in order to craft attacks that current technical systems and IT professionals are ill-prepared for, absent additional investments. For example, services like Google's VirusTotal file analyzer allows users to upload variants to a central site and be judged by 60+ different security tools. This feedback loop presents an opportunity to use AI to aid in crafting multiple variants of the same malicious code to determine which is most effective at evading security tools. Additionally, large-scale AI attackers can accumulate and use large datasets to adjust their tactics, as well as varying the details of the attack for each target. This may outweigh any disadvantages they suffer from the lack of skilled human attention to each target, and the ability of defenders like antivirus companies and IT departments to learn to recognize attack signatures. While the specific examples of AI applied to offensive cybersecurity mentioned above were developed by white hat researchers, we expect similar efforts by cybercriminals and state actors in the future as highly capable AI techniques become more widely distributed, as well as new applications of AI to offensive cybersecurity that have not yet been explored. \n Points of Control and Existing Countermeasures Cyber risks are difficult to avert entirely, but not impossible to mitigate, and there are multiple points of control at which interventions can increase security. Below, we highlight different points of control and existing countermeasures for defending at those points, as well as their limitations. Overall, we believe that AI and cybersecurity will rapidly evolve in tandem in the coming years, and that a proactive effort is needed to stay ahead of motivated attackers. We highlight potential but not yet proven countermeasures in the section below on Interventions. Consumer awareness: More aware users can spot telltale signs of certain attacks, such as poorly crafted phishing attempts, and practice better security habits, such as using diverse and complex passwords and twofactor authentication. However, despite long-standing awareness of the vulnerability of IT systems, most end users of IT systems remain vulnerable to even simple attacks such as the exploitation of unpatched systems . This is concerning in light of the potential for the AI-cybersecurity nexus, especially if high-precision attacks can be scaled up to large numbers of victims. Governments and researchers: Various laws and researcher norms pertain to cybersecurity. For example, the Digital Millennium Act and the Computer Fraud and Abuse Act in the US proscribe certain actions in cyberspace . Legal enforcement is particularly difficult across national boundaries. Norms such as responsible disclosure of vulnerabilities also aid in defense by reducing the likelihood of a newly disclosed vulnerability being used against a large number of victims before it can be patched. AI is not explicitly addressed in such laws and norms, though we discuss their possible applicability to AI below in Interventions. An important activity that cybersecurity researchers perform is the detection of vulnerabilities in code, allowing vendors to increase the security of their products. Several approaches exist to incentivize such processes and make them easier, including: • Payment of \"Bug bounties,\" in which participants are compensated for finding and responsibly disclosing vulnerabilities. • \"Fuzzing,\" an automated method of vulnerability detection by trying out many possible permutations of inputs to a program, which is often used internally by companies to discover vulnerabilities. • Products (already available) that rely on machine learning to predict whether source code may contain a vulnerability. 1 National Cyber Security Crime Centre, 2016 Both the DMCA and the CFAA have been criticised for creating risk for computer security researchers and thereby making systems less secure in some cases  (EFF, 2014; Timm, 2013) , which may either suggest that these tasks are not the right model for legislative action, or that laws and norms are hard to use effectively as an intervention. \n p.36 Security Domains Industry centralization: Spam filters are a canonical example of where centralization of an IT system aids defense-individuals benefit from the strength of Google's spam filter and consequently are protected from many very simple attacks, and this filter is stronger because Google uses large amounts of user data to improve it over time. Likewise, many large networks are constantly monitoring for anomalies, protecting those who use the networks if anomalies are correctly identified and acted upon. These systems benefit from economies of scaleit makes more sense to continue iterating a single spam filter for a large number of users than to have every user build their own or have one installed on their computer. Similarly, cloud computing companies may enforce terms of agreement that prevent their hardware from being used for malicious purposes, provided they can identify such behavior. Another example of a system-level defense is blacklisting of IP addresses from which attacks are commonly launched, though skilled attackers can obfuscate the origin of their attacks. Centralization and the associated economies of scale may also facilitate the deployment of AI-based defenses against cybersecurity attacks, by allowing the aggregation of large datasets and the concentration of labor and expertise for defense. This dynamic may be very important for preventing attack from outpacing defense and is discussed further in Interventions and Appendix B. Centralization is not an unalloyed good, however, as it raises the stakes if central systems are compromised. Another difficulty with this control point is that attackers can learn how to evade system-level defenses. For example, they can purchase commercial antivirus software and analyze changes between updates of the protection protocol to see what is and isn't being protected against. Attacker incentives: Attackers can be deterred from committing future attacks or punished for prior attacks. A necessary (though not sufficient) condition of successfully deterring and punishing attackers is the ability to attribute the source of an attack, a notoriously difficult problem . A compounding problem for those who would attribute an attack is that even if they have high-quality information, they may not want to reveal it, because doing so may compromise a source or method . Finally, some entities may not wish to punish certain actions, so as to avoid creating precedent and thereby preserve leeway to engage in such actions themselves . \n Technical cybersecurity defenses: A wide variety of cybersecurity defenses are available, though there is as yet little solid analysis of their relative effectiveness . \n Rid, 2015 For instance, the failure of the United Nations Cybersecurity Group of Governmental Experts to make progress on norms for hacking in international law  (Korzak, 2017)  appears to be a result of this dynamic. Libicki, 2016 Libicki, 2016 p.37 \n Security Domains Many of these interventions were proposed before unique considerations of AI were apparent but nevertheless remain relevant in a future with expanded AI cybersecurity applications. Companies provide a wide variety of cybersecurity solutions, ranging from automatic patching of a vendor's own software, to threat detection, to incident response and consulting services. Network and endpoint security products aim to prevent, detect, and respond to threats. Solutions include detection of software exploits, and prevention or detection of attacker tools, techniques, and procedures. Key areas of defense include the endpoint (i.e., computer) security, internal network security, and cloud security. Machine learning approaches are increasingly used for cyber defense. This may take the form of supervised learning, where the goal is to learn from known threats and generalize to new threats, or in the form of unsupervised learning in which an anomaly detector alerts on suspicious deviations from normal behavior. For example, so-called \"next-gen\" antivirus solutions often leverage supervised learning techniques to generalize to new malware variants. User and entity behavioral tools monitor normal user or application behavior, and detect deviations from normalcy in order to detect malicious behavior among the collected anomalies. Recently, AI has also been used to aid security professionals to hunt for malicious actors more efficiently within their own enterprises, by allowing interaction via natural language and automating queries for understanding potential threats . Relatively little attention has been paid to making AI-based defenses robust against attackers that anticipate their use. Ironically, the use of machine learning for cyber defense can actually expand the attack surface due to this lack of attention and other vulnerabilities . Furthermore, surveys of cybersecurity professionals indicate low confidence in AI-based defense systems today . As such, we encourage further development of such defense technologies in the Interventions section below. \n Physical Security In this section, we consider AI-related risks in the broad area of physical harm. Many of these are familiar challenges from existing uses of electronics and computers in weapons systems, though the addition of AI capabilities may change this landscape along the lines introduced in the General Framework for AI and Security Threats. As with Digital Security above, we introduce the context, AI-enabled changes, and existing countermeasures related to physical attacks below. Regulation and technical research on defense have been slow to catch up with the global proliferation of weaponizable robots. While defenses against attacks via robots (especially aerial drones) are being developed, there are few obstacles at present to a moderately talented attacker taking advantage of the rapid proliferation of hardware, software, and skills to cause large amounts of physical harm through the direct use of AI or the subversion of AI-enabled systems. Physical harm via human-piloted drones and land-based robots is already playing a major role in some conflicts, even prior to the incorporation of autonomy . In the near-term, we can expect a growing gap between attack capabilities and defense capabilities, because the necessary defenses are capital-intensive and the hardware and software required to conduct attacks are increasingly widely distributed. Unlike the digital world, where key nodes in the network such as Google can play a key role in defense, physical attacks can happen anywhere in the world, and many people are located in regions with insufficient resources to deploy large-scale physical defenses of the kind discussed below, thus necessitating consideration of policy measures and interventions related to the supply chain for robots. The resource and technological advantages currently available to large organizations, such as militaries and police forces, in the domain of physical attack and defense will continue when such attacks become augmented by AI. However, it should be noted that some of the most worrying AI-enabled attacks may come from small groups and individuals who have preferences far outside what is typical and which are difficult to anticipate or prevent, as with today's \"lone-wolf\" terrorist attacks such as mass shootings. \n Context Recent years have seen an explosion in the number and variety of commercial applications for robots. Industrial robots are growing in number (254,000 supplied in 2015 versus 121,000 in 2010 ), some with and some without AI components. Relatively primitive cleaning robots are in wide use and more sophisticated service robots appear to be on the horizon (41,000 service robots were sold in 2015 for professional use, and about 5.4 million for personal and domestic use ). Additionally, not all of these robots are on the ground. There are aquatic and aerial robotics applications being explored, with the latter proliferating in very high numbers. In the United States alone, the number of drones has skyrocketed in recent years, with over 670,000 registered with the Federal Aviation Administration in 2016 and 2017 . \n Singer, 2009 IFR, 2016 IFR, 2016  Vanian, 2017  p.39 \n Security Domains Ambitious plans for drone-based delivery services are being proposed and tested, commercial opportunities for drones are continuously launched, and recreational uses are flourishing (e.g. drone racing and photography). Driverless cars are robots, and they also are increasingly being used in uncontrolled environments (that is, outside of test facilities), though large-scale deployment of fully autonomous driverless cars awaits the resolution of technical and policy challenges. A wide range of robots with autonomous features are already deployed within multiple national militaries, some with the ability to apply lethal force , and there is ongoing discussion of possible arms control measures for lethal autonomous weapon systems. Three characteristics of this diffusion of robotics should be noted. • It is truly global: humanitarian, recreational, military, and commercial applications of robots are being explored on every continent, and the supply chains are also global, with production and distribution dispersed across many countries. • The diffusion of robotics enables a wide range of applications: drone uses already range from competitive racing to photography to terrorism . While some specialized systems exist (e.g. some special-purpose industrial robots and cleaning robots that can only move around and vacuum), many are fairly generic and customizable for a variety of purposes. • Robotic systems today are mostly not autonomous, as humans play a significant role in directing their behavior, but more and more autonomous and semi-autonomous systems are also being developed for application such as delivery and security in real world environments . For example, from relatively unstable and hard-to-fly drones a decade ago, to drones that can stabilize themselves automatically, we see a steady increase in the autonomy of deployed systems. More autonomous behavior is on the horizon for commercial products as well as military systems . Each of these characteristics sets the stage for a potentially disruptive application of AI and malicious intent to existing and near-term robotic systems. \n How AI Changes the Physical Security Landscape The ability of many robots to be easily customized and equipped \n p. 40 Security Domains with dangerous payloads lends itself to a variety of physical attacks being carried out in a precise way from a long distance, an ability previously limited to countries with the resources to afford technologies like cruise missiles . This threat exists independently of AI (indeed, as mentioned above, most robots are human-piloted at present) but can be magnified through the application of AI to make such systems autonomous. As mentioned previously, nonautomated drone attacks have been conducted already by groups such as ISIS and Hamas , and the globalized nature of the robotics market makes it difficult to prevent this form of use. Nonetheless, we will discuss some possible countermeasures below. Greater degrees of autonomy enable a greater amount of damage to be done by a single person -making possible very large-scale attacks using robots -and allowing smaller groups of people to conduct such attacks. The software components required to carry out such attacks are increasingly mature. For example, open source face detection algorithms, navigation and planning algorithms, and multi-agent swarming frameworks that could be leveraged towards malicious ends can easily be found. Depending on their power source, some robots can operate for long durations, enabling them to carry out attacks or hold targets at risk over long periods of time. Robots are also capable of navigating different terrain than humans, in light of their different perceptual capabilities (e.g. infrared and lidar for maneuvering in the dark or in low-visibility fog) and physical capacities (e.g. being undeterred by smoke or other toxic substances and not needing oxygen underwater). Thus, a larger number of spaces may become vulnerable to automated physical attacks. There are also cross-cutting issues stemming from the intersection of cybersecurity and increasingly autonomous cyber-physical systems. The diffusion of robots to a large number of humanoccupied spaces makes them potentially vulnerable to remote manipulation for physical harm, as with, for example, a service robot hacked from afar to carry out an attack indoors. With regard to cyber-physical systems, the Internet of Things (IoT) is often heralded as a source of greater efficiency and convenience, but it is also recognized to be highly insecure and represents an additional attack vector by which AI systems controlling key systems could be subverted, potentially causing more damage than would have been possible were those systems under human control. In addition to traditional cybersecurity vulnerabilities, AIaugmented IoT and robotic systems may be vulnerable to AIspecific vulnerabilities such as adversarial examples.  \n p. 41 Security Domains There is also some evidence to suggest that people are unduly trusting of autonomous mobile robots, potentially creating additional sources of security vulnerabilities as such robots become more widely deployed . The consequences of these cyber vulnerabilities are particularly acute for autonomous systems that conduct high-risk activities such as self-driving cars or autonomous weapons. \n Points of Control and Existing Countermeasures There are numerous points of control that could be leveraged to reduce the risk of physical harm involving AI. While the capacity to launch attacks with today's consumer robots is currently widely distributed, future generations of robots may be more tightly governed, and there exist physical defenses as well. However, such defenses are capital-intensive and imperfect, leading us to conclude that there may be an extended risk period in which it will be difficult to fully prevent physical attacks leveraging AI. \n Hardware manufacturers There are currently a relatively limited number of major manufacturers, with companies like DJI holding a dominant position in the consumer drone market, with about 70% of the global market . This concentration makes the hardware ecosystem more comprehensible and governable than the analogous ecosystem of AI software development. With growing recognition of the diverse economic applications of drones, the market may diffuse over the longer term, possibly making the supply chain a less useful focal point for governance. For example, it might currently be feasible to impose minimum standards on companies for hardening their products against cyber attacks or to make them more resistant to tampering, so as to at least somewhat raise the skill required to carry out attacks through these means or raise the costs of acquiring uncontrolled devices. The U.S. Federal Trade Commission is exploring such regulations. \n Hardware distributors There are many businesses that sell drones and other robotic systems, making the ecosystem more diffuse at this level than it is at the production level. It is conceivable that at least some risks might be mitigated through action by distributors, or other point-of-sale based approaches. Notably, this type of control is currently much more feasible for hardware than for software, and restrictions on sales of potentially lethal drones might be thought of as analogous to restrictions on sales of guns and ingredients for illegal drugs. \n p. 42 \n Security Domains \n Software supply chain There are many open source frameworks for computer vision, navigation, etc. that can be used for carrying out attacks, and products often come with some built-in software for purposes such as flight stabilization. But not all powerful AI tools are widely distributed, or particularly easy to use currently. For example, large trained AI classification systems that reside within cloud computing stacks controlled by big companies (which are expensive to train), may be tempting for malicious actors to build from, potentially suggesting another point of control (discussed in Interventions and Appendix B). \n Robot users There are also registration requirements for some forms of robots such as drones in many countries, as well as requirements for pilot training, though we note that the space of robots that could cause physical harm goes beyond just drones. There are also no fly zones, imposed at a software level via manufacturers and governments, which are intended to prevent the use of consumer drones in certain areas, such as near airports, where the risk of unintentional or intentional collision between drones and passenger aircrafts looms large . Indeed, at least one drone has already struck a passenger aircraft , suggesting a strong need for such no fly zones. \n Governments There is active discussion at the United Nations Convention on Certain Conventional Weapons of the value and complexity of banning or otherwise regulating lethal autonomous weapons systems . Key states' opposition to a strong ban makes such an agreement unlikely in in the near-term, though the development of norms that could inform stronger governance is plausible . Already in the United States, for example, there is an official Department of Defense directive that sets out policy for the development and use of autonomy in weapons . Additionally, the U.S. Law of War Manual notes that humans are the primary bearers of responsibility for attacks in armed conflict . The International Committee of the Red Cross has adopted a similar position, a stance that presumably implies some minimum necessary degree of human involvement in the use of force . While such arms control discussions and norm development processes are critical, they are unlikely to stop motivated non-state actors from conducting attacks. \n Physical defenses In the physical sphere, there are many possible defenses against attacks via robots, though they are imperfect and unevenly distributed at present. Many are expensive and/or require human labor to deploy, and hence are only used to defend \"hard targets\" Given the potential for automation to allow attacks at scale, a particular challenge for defenders is finding effective methods of defense with an acceptable cost-exchange ratio . As of yet, these defenses are incomplete and expensive, suggesting a likely near-term gap between the ease of attack and defense outside of heavily guarded facilities that are known targets (e.g. airports or military bases). \n Payload control An actor who wants to launch an aerial drone attack carrying a dangerous payload must source both the drone and the payload. Developed countries generally have long-lasting and reasonably effective systems to restrict access to potentially explosive materials, and are introducing systems to restrict access to acids (following high-profile acid attacks). More generally, state security and intelligence services uncover and foil a large number of attempted attacks, including those that involve attempts to procure dangerous materials. Increases in AI capabilities will likely help their work e.g. in analysing signal intelligence, or in characterising and tracking possible attackers. \n Political Security Next, we discuss the political risks associated with malicious AI use. AI enables changes in the nature of communication between individuals, firms, and states, such that they are increasingly mediated by automated systems that produce and present content. Information technology is already affecting political institutions in myriad ways -e.g. the role of social media in elections, protests, and even foreign policy . The increasing use of AI may make existing trends more extreme, and enable new kinds of political dynamics. Worryingly, the features of AI described earlier such as its scalability make it particularly well suited to undermining public discourse through the large-scale production of persuasive but false content, and strengthening the hand of authoritarian regimes. We consider several types of defenses, but as yet, as in the cases of Digital Security and Physical Security, the problem is unsolved. \n p. 44 Security Domains \n Context There are multiple points of intersection between existing information technologies and the political sphere. Historically, politics and instability have had a symbiotic relationship with technological advances. Security needs have driven technological advances, and new technology has also changed the kinds of security threats that states and politicians face. Examples abound including the advent of the semaphore telegraph in Napoleonic France , to the advent of GPS and its use during the First Gulf War , to the use of social media during the Arab Spring . Technological advances can change the balance of power between states, as well as the relationship between incumbent leaders and protesters seeking to challenge them. Modern militaries and intelligence agencies use today's information technologies for surveillance, as they did with previous generations of technologies such as telephones. However, the effects of new technologies on these power relations are not straightforward. For example, social media technologies empower both incumbents and protesters: they allow military intelligences to monitor sentiment and attitudes, and to communicate more quickly; however, they also provide protesters in places such as Ukraine and Egypt, and rebel groups and revolutionary movements such as ISIS or Libyan rebels, the ability to get their message out to sympathetic supporters around the world more quickly and easily. In addition, research suggests that social media may empower incumbent authoritarian regimes , as incumbent governments can manipulate the information that the public sees. Finally, some have argued that social media has further polarized political discourse, allowing users, particularly in the West, to self-select into their own echo chambers, while others have questioned this assumption . Machine learning algorithms running on these platforms prioritize content that users are expected to like. Thus the dynamics we observe today are likely to only accelerate as these algorithms and AI become even more sophisticated. While they have evolved from previous technologies, information communication technologies are notable in some respects, such as the ease of information copying and transmission. Waltzmann writes, \"The ability to influence is now effectively 'democratized,' since any individual or group can communicate and influence large numbers of others online\" . This \"democratization\" of influence is not necessarily favorable to democracy, however. It is very easy today to spread manipulative and false information, and existing approaches for detecting and stopping the spread of \"fake news\" fall short. Other structural aspects of modern technologies and the media industry also enable these trends.  Marwick and Lewis (2017)  note that the media's \"dependence on social media, analytics and metrics, sensationalism, novelty over newsworthiness, and clickbait makes them vulnerable to such media manipulation.\" Others, such as  Morozov (2012)  and  King, Pan, and Roberts (2017)  argue that social media provides more tools for authorities to manipulate the news environment and control the message. Finally, we note that the extent and nature of the use of information communication technologies to alter political dynamics varies across types of political regimes. In liberal democracies, it can be thought of as more of an emergent phenomenon, arising from a complex web of industry, government, and other actors, whereas in states like China, there is an explicit and deliberate effort to shape online and in-person political discussions, making use of increasingly sophisticated technologies to do so . For instance, the Chinese government is exploring ways to leverage online and offline data to distill a \"social credit score\" for its citizens, and the generally more widespread use of censorship in China exemplifies the more explicit leveraging of technology for political purposes in some authoritarian states . \n How AI Changes the Political Security Landscape AI will cause changes in the political security landscape, as the arms race between production and detection of misleading information evolves and states pursue innovative ways of leveraging AI to maintain their rule. It is not clear what the longterm implications of such malicious uses of AI will be, and these discrete instances of misuse only scratch the surface of the political implications of AI more broadly . However, we hope that understanding the landscape of threats will encourage more vigorous prevention and mitigation measures. Already, there are indications of how actors are using digital automation to shape political discourse. The widespread use of social media platforms with low barriers to entry makes it easier for AI systems to masquerade as people with political views. This has led to the widespread use of social media \"bots\" to spread political messages and cause dissent. At the moment, many such bots are controlled by humans who manage a large pack of bots , or use very simple forms of automation. However, these bot-based strategies (even when using relatively unsophisticated automation) are leveraged by national intelligence agencies and have demonstrated the ability to influence mainstream media coverage and political beliefs . For instance, during both the Syrian Civil War and the 2016 US election bots appeared to actively try to sway public opinion . Greater scale and sophistication of autonomous software actors in the political sphere is technically possible with existing AI techniques . As previously discussed, progress in automated spear phishing has demonstrated that automatically generated text can be effective at fooling humans , and indeed very simple approaches can be convincing to humans, especially when the text pertains to certain topics such as entertainment . It is unclear to what extent political bots succeed in shaping public opinion, especially as people become more aware of their existence, but there is evidence they contribute significantly to the propagation of fake news . In addition to enabling individuals and groups to mislead the public about the degree of support for certain perspectives, AI creates new opportunities to enhance \"fake news\" (although, of course, propaganda does not require AI systems to be effective). AI systems may simplify the production of high-quality fake video footage of, for example, politicians saying appalling (fake) things . Currently, the existence of high-quality recorded video or audio evidence is usually enough to settle a debate about what happened in a given dispute, and has been used to document war crimes in the Syrian Civil War . At present, recording and authentication technology still has an edge over forgery technology. A video of a crime being committed can serve as highly compelling evidence even when provided by an otherwise untrustworthy source. In the future, however, AI-enabled highquality forgeries may challenge the \"seeing is believing\" aspect of video and audio evidence. They might also make it easier for people to deny allegations against them, given the ease with which the purported evidence might have been produced. In addition to augmenting dissemination of misleading information, the writing and publication of fake news stories could be automated, as routine financial and sports reporting often are today. As production and dissemination of high-quality forgeries becomes increasingly low-cost, synthetic multimedia may constitute a large portion of the media and information ecosystem. Even if bot users only succeed in decreasing trust in online environments, this will create a strategic advantage for political ideologies and groups that thrive in low-trust societies or feel opposed by traditional media channels. Authoritarian regimes in particular may benefit from an information landscape where objective truth becomes devalued and \"truth\" is whatever authorities claim it to be. Moreover, automated natural language and multimedia production will allow AI systems to produce messages to be targeted at those most susceptible to them. This will be an extension of existing p. 47 Security Domains advertising practices. Public social media profiles are already reasonably predictive of personality details , and may be usable to predict psychological conditions like depression . Sophisticated AI systems might allow groups to target precisely the right message at precisely the right time in order to maximize persuasive potential. Such a technology is sinister when applied to voting intention, but pernicious when applied to recruitment for terrorist acts, for example. Even without advanced techniques, \"digital gerrymandering\" or other forms of advertising might shape elections in ways that undermine the democratic process. The more entrenched position of authoritarian regimes offers additional mechanisms for control through AI that are unlikely to be as easily available in democracies . AI systems enable fine-grained surveillance at a more efficient scale . While existing systems are able to gather data on most citizens, efficiently using the data is too costly for many authoritarian regimes. AI systems both improve the ability to prioritise attention (for example, by using network analysis to identify current or potential leaders of subversive groups ) and also reduce the cost of monitoring individuals (for example, using systems that identify salient video clips and bring them to the attention of human agents). Furthermore, this can be a point of overlap between political and physical security, since robotic systems could also allow highly resourced groups to enforce a greater degree of compliance on unwilling populations. The information ecosystem itself enables political manipulation and control by filtering content available to users. In authoritarian regimes, this could be done by the state or by private parties operating under rules and directions issued by the state. In democracies, the state may have limited legal authority to shape and influence information content but the same technical tools still exist; they simply reside in the hands of corporations. Even without resorting to outright censorship, media platforms could still manipulate public opinion by \"de-ranking\" or promoting certain content. For example, Alphabet Executive Chairman Eric Schmidt recently stated that Google would de-rank content produced by Russia Today and Sputnik . In 2014, Facebook manipulated the newsfeeds of over half a million users in order to alter the emotional content of users' posts, albeit modestly . While such tools could be used to help filter out malicious content or fake news, they also could be used by media platforms to manipulate public opinion . Finally, the threats to digital and physical security that we have described in previous sections may also have worrying implications for political security. The hacking of the Clinton campaign in the 2016 presidential election is a recent example of how successful  \n Points of Control and Existing Countermeasures Several measures are already in development or deployed in this area, though none has yet definitively addressed the problems. We highlight a few relevant efforts here, and emphasize that these proposals are oriented towards the protection of healthy public discourse in democracies. Preventing more authoritarian governments from making full use of AI seems to be an even more daunting challenge. Technical tools. Technical measures are in development for detecting forgeries and social media bots . Likewise, the use of certified authenticity of images and videos, e.g. the ability to prove that a video was broadcast live rather than synthesized offline are valuable levers for ensuring that media is in fact produced by the relevant person or organization and is untampered in transit. Analogous measures have been developed for authentication of images (rather than videos) by  Naveh and Tromer (2016) . Automated fake news detection is likewise the subject of ongoing research as well as a competition, the Fake News Challenge , which can be expected to spur further innovation in this area. As yet, however, the detection of misleading news and images is an unsolved problem, and the pace of innovation in generating apparently authentic multimedia and text is rapid. Pervasive use of security measures. Encryption is a generally useful measure for ensuring the security of information transmissions, and is actively used by many companies and other organizations, in part to prevent the sorts of risks discussed here. The use of citizens' data by intelligence agencies takes various forms and has been actively debated, especially in the wake of the Snowden revelations . General interventions to improve discourse. There are various proposals to increase the quality of discourse in the public and private spheres, including longstanding ones such as better education and teaching of critical thinking skills, as well as newer ones ranging from tools for tracking political campaigning in social media (such as \"Who Targets Me?\" ) to policy proposals to apps for encouraging constructive dialogue . The Fake News Challenge is a competition aimed at fostering the development of AI tools to help human fact checkers combat fake news. \n Sunstein, 2017 Ixy, 2017 \"Who Targets Me?\" is a software service that informs citizens on the extent with which they are being targeted by dark advertising campaigns. \n p. 49 \n Security Domains Media platforms. There have always been news sources of varying impartiality, and some online sources have better reputations than others, yet this has not entirely stopped the spread of fake news. Likewise, most people are aware of the existence of Ponzi schemes, scam emails, misleading sales tactics, etc. and yet victims are still found. Part of the reason that spam is less of a problem today than it otherwise could be is that the owners of key platforms such as email servers have deployed sophisticated spam filters. More generally, technology companies, social media websites, and and media organizations are critical points of control for stemming the tide of increasingly automated disinformation, censorship, and persuasion campaigns. Additionally, these organizations have unique datasets that will be useful for developing AI systems for detecting such threats, and through the ability to control access, they can pursue other strategies for preventing malicious uses of these platforms such as imposing strong barriers to entry (e.g. the use of one's offline identity) and limiting the rate at which accounts can disseminate information. Because these media platforms are for-profit corporations, public discourse, transparency, and potentially regulation will be important mechanisms for ensuring that their use of these powerful tools aligns with public interest . A development that occurred during the process of writing this report is illustrative. Late 2017 saw the rise of \"deepfakes,\" the application of face-swapping algorithms to (among other applications) adult videos. While such videos first began appearing en masse in Reddit fora clearly labeled as being fictitious, the realism of some such deepfakes is an early sign of the potential decline of \"seeing is believing\" discussed above. After substantial media coverage of deepfakes, Reddit and other online websites, including adult content websites, began to crack down on the discussion and propagation of the technique. While these efforts have not been fully successful, they illustrate the critical role of technology platforms in governing information access, and it is likely that the deepfakes crackdown at least somewhat slowed the dissemination of the tool and its products, at least amongst less sophisticated actors. \n Inter ventions p.50 We identify a wide range of potential responses to the challenges raised above, as well as a large number of areas for further investigation. This section first makes several initial high-level recommendations for AI and ML researchers, policymakers, and others. We then suggest specific priority areas for further research, where investigation and analysis could develop and refine potential interventions to reduce risks posed by malicious use. Due to the exploratory nature of this report, our primary aim is to draw attention to areas and potential interventions that we believe should be the subject of further investigation, rather than to make highly specific technical or policy proposals that may not be viable. The structure of this section, and the inclusion of Appendix B with additional exploratory material, is informed by this perspective. \n p.51 Interventions \n Recommendations In this subsection we present four high-level recommendations, which are focused on strengthening the dialog between technical researchers, policymakers, and other stakeholders. In the following subsection, we will turn our attention to more concrete priority areas for technical work as well as associated research questions. Our first pair of recommendations arise from the fact that the issues raised in this report combine technical and nontechnical considerations, such as social, economic and military considerations. Concerns were raised at the workshop that the development of viable, appropriate responses to these issues may be hampered by two self-reinforcing factors: first, a lack of deep technical understanding on the part of policymakers, potentially leading to poorly-designed or ill-informed regulatory, legislative, or other policy responses; second, reluctance on the part of technical researchers to engage with these topics, out of concern that association with malicious use would tarnish the reputation of the field and perhaps lead to reduced funding or premature regulation. Our first two recommendations aim at preempting this dynamic. Recommendation #1: Policymakers should collaborate closely with technical researchers to investigate, prevent, and mitigate potential malicious uses of AI. This must include policymakers taking seriously their responsibility to avoid implementing measures that will interfere with or impede research progress, unless those measures are likely to bring commensurate benefits. Close collaboration with technical experts also ensures that policy responses will be informed by the technical realities of the technologies at hand . Recommendation #2: Researchers and engineers in artificial intelligence should take the dual-use nature of their work seriously, allowing misuse-related considerations to influence research priorities and norms, and proactively reaching out to relevant actors when harmful applications are foreseeable. Given that AI is a dual-use technology, we believe it is important that researchers consider it their responsibility to take whatever steps they can to help promote beneficial uses of the technology and prevent harmful uses. Example steps could include engaging with policymakers to provide expertise, and considering the potential applications of different research projects before deciding what to work on. (We recognize and appreciate the many AI researchersincluding the technical experts who took part in the workshop and contributed to this report and other related initiatives -who are already doing outstanding work along these lines. ) Introductory resources for policymakers interested in this domain are increasingly becoming available, both generally about AI  (Buchanan and Taylor, 2017) , and specifically on AI and security (CNAS, 2017). As an example of policymaking in this domain that has surfaced several difficulties, the European Union's General Data Protection Regulation is a commonly-discussed example of a policy that is hard to interpret and apply in the context of current machine learning algorithms  (Goodman and Flaxman, 2016) . 1 The work of the Partnership on AI, the White House's 2016 series of workshops on AI, the 2017 \"Beneficial AI\" conference in Asilomar, and the AI Now conference series and organization are further examples where contributions from technical experts have been substantial and valuable. \n p.52 \n Interventions We also make two recommendations laying out aims that we believe the broader AI community (including both technical and policy professionals) should work towards. Recommendation #3: Best practices should be identified in research areas with more mature methods for addressing dualuse concerns, such as computer security, and imported where applicable to the case of AI. An example of a best practice that workshop participants considered clearly valuable to introduce into AI contexts is extensive use of \"red teaming.\" See Priority Research Area #1, below, for further details. Recommendation #4: Actively seek to expand the range of stakeholders and domain experts involved in discussions of these challenges. This could include reaching out to sectors like civil society, national security experts, as-yet unengaged AI and cybersecurity researchers, businesses incorporating AI into their products, ethicists, the general public , and others, to ensure that relevant stakeholders are included and relevant experts consulted. Because of the dual-use nature of AI, many of the malicious uses of AI outlined in this report have related legitimate uses. In some cases, the difference between legitimate and illegitimate uses of AI could be one of degree or ensuring appropriate safeguards against malicious use. For example, surveillance tools can be used to catch terrorists or oppress ordinary citizens. Information content filters could be used to bury fake news or manipulate public opinion. Governments and powerful private actors will have access to many of these AI tools and could use them for public good or harm. This is why a public dialogue on appropriate uses of AI technology is critical. The above four recommendations can help foster a crossdisciplinary dialogue among AI researchers, policymakers, and other relevant stakeholders to ensure that AI technology is used to benefit society. \n Priority Areas for Further Research This section lays out specific topic areas that we recommend be investigated further. We aim here for brevity; more specific questions for investigation, along with additional context and commentary on many of the topics mentioned, may be found in Appendix B. In computer security, red teaming involves a \"red team\", composed of security experts and/or members of the host organization, deliberately planning and carrying out attacks against the systems and practices of the organization (with some limitations to prevent lasting damage), with an optional \"blue team\" responding to these attacks. These exercises explore what an actual attack might look like in order to better understand and, ultimately, improve the security of the organisation's systems and practices. We expect adaptive defensive actions will be required of everyday citizens, if only in terms of maintaining awareness of threats and adopting best practices. It is important to acknowledge that different communities will have varying abilities to make such adaptations, depending for example on their technological literacy, which may pose challenges for implementing security policies. This is important not just for the communities less able to adapt to the new threats, but also for society more broadly as, for example, insecure systems may be compromised by attackers and repurposed to provide computing power and data for yetmore-capable attacks, while reducing the possibility that the attacks could be attributed, as they would then seem to originate from the compromised system. \n p.53 Interventions Priority Research Area #1: \n Learning from and with the Cybersecurity Community As AI-based systems become more widespread and capable, the potential impacts of cybersecurity incidents are growing commensurately. To summarize the considerations in the Digital Security section, AI is important to cybersecurity for three reasons. First, increased automation brings with it increased digital control of physical systems; consider, for example, how much more control a successful hacker could exercise over a modern car, compared with a typical car from 20 years ago . Second, successful attacks on AI-based systems can also give the attacker access to the algorithms and/or trained models used by the system; consider, for example, theft of the datasets used for facial recognition on social networks, or the compromise of an algorithm used for analysing satellite imagery. Third, increasing use of AI in cyberattacks is likely to allow highly sophisticated attacks to be carried out at a much larger scale, which may reach victims that would otherwise not be suitable targets of previous waves of sophisticated attacks. To respond to these increased dangers, cybersecurity must be a major and ongoing priority in efforts to prevent and mitigate harms from AI systems, and best practices from cybersecurity must be ported over wherever applicable to AI systems. Some examples of cybersecurity-related sub-areas that we believe should be the subject of further research and analysis, then be implemented as appropriate (see Appendix B for more commentary on and questions about these sub-areas), include: • Red teaming. Extensive use of red teaming to discover and fix potential security vulnerabilities and safety issues should be a priority of AI developers, especially in critical systems. •  Formal verification. To what extent, in what circumstances, and  for what types of architectures can formal verification be used to prove key properties of AI systems? Can other approaches be developed to achieve similar goals by different means? • Responsible disclosure of AI vulnerabilities. Should AI-specific procedures be established to enable confidential reporting of vulnerabilities discovered in AI systems (including security vulnerabilities, potential adversarial inputs, and other types of exploits), as is already possible for security exploits in modern software systems? • Forecasting security-relevant capabilities. Could \"white-hat\" efforts to predict how AI advances will enable more effective DARPA's Assured Autonomy program  (Neema, 2017)  is one attempt at developing techniques to assure safety in systems that continue learning throughout their lifespans, which makes assurance or verification using traditional methods challenging. See also  Katz et al., 2017; Selsam, Liang, and Dill, 2017; and Carlini et al., 2017.  For example, see the case of hackers first bringing a Jeep to a standstill on a busy highway, then later developing the ability to cause unintended acceleration and fully control the vehicle's steering  (Greenberg, 2016) . \n p.54 Interventions cyberattacks, and more rigorous tracking of AI progress and proliferation in general, allow for more effective preparations by defenders? • Security tools. What tools (if any) should be developed and distributed to help make it standard to test for common security problems in AI systems, analogously to tools used by computer security professionals? • Secure hardware. Could security features be incorporated into AI-specific hardware, for example to prevent copying, restrict access, facilitate activity audits, and similar? How technically and practically feasible is the design and adoption of hardware with properties like this? Priority Research Area #2: Exploring Different Openness Models Today, the prevailing norms in the machine learning research community strongly point towards openness. A large fraction of novel research is published online in papers that share anything from rough architectural outlines to algorithmic details to source code. This level of openness has clear benefits in terms of enabling researchers to build on each others' work, promoting collaboration, and allowing theoretical progress to be incorporated into a broad array of applications. However, the potential misuses of AI technology surveyed in the Scenarios and Security Domains sections suggest a downside to openly sharing all new capabilities and algorithms by default: it increases the power of tools available to malicious actors. This raises an important research question: might it be appropriate to abstain from or merely delay publishing of some findings related to AI for security reasons? There is precedent for this in fields such as computer security, where exploits that could affect important systems are not publicly disclosed until the developers have had an opportunity to fix the vulnerability. To the extent that research results are withheld today in AI, it is usually for reasons related to intellectual property (e.g. in order to avoid a future result being \"scooped\"). In light of risks laid out elsewhere in this report, there may also be arguments based on public interest for additional caution in at least some cases. While the proposals below consider decreasing openness in certain situations, we stress that there are clear and well-recognized reasons to favor openness in research communities. We believe that policies leading to decreased openness, while potentially \n p.55 Interventions appropriate in certain instances, should be sensitive to these benefits. Rather than propose a specific solution, our aim is to foster discussion of whether and when considerations against open sharing might outweigh considerations in favor and what mechanisms might enable this. Some potential mechanisms and models that could be subject to further investigation and analysis (see Appendix B for more commentary on and questions about for these sub-areas) include: • Pre-publication risk assessment in technical areas of special concern. Should some types of AI research results, such as work specifically related to digital security or adversarial machine learning, be subject to some kind of risk assessment to determine what level of openness is appropriate ? This is the norm for research in other areas, such as biotechnology and computer security. Or would such measures be premature today, before AI systems are more widely used in critical systems and we have better knowledge of which technical research is most security-relevant? If such measures are considered be premature, under what conditions would they be appropriate? • Central access licensing models. Could emerging \"central access\" commercial structures -in which customers use services like sentiment analysis or image recognition made available by a central provider without having access to the technical details of the system -provide a template for a security-focused sharing model that allows widespread use of a given capability while reducing the possibility of malicious use? How might such a model remain viable over time as advances in processing power, data storage and availability, and embedded expertise allow a larger set of actors to use AI tools? • Sharing regimes that favor safety and security. Could arrangements be made under which some types of research results are selectively shared among a predetermined set of people and organizations that meet certain criteria, such as effective information security and adherence to appropriate ethical norms? For example, certain forms of offensive cybersecurity research that leverage AI might be shared between trusted organizations for vulnerability discovery purposes, but would be harmful if more widely distributed. • Other norms and institutions that have been applied to dualuse technologies. What can be learned from other models, methodologies, considerations, and cautions that have arisen when tackling similar issues raised by other dual-use technologies? Accordingly, concerns about misuse should not be used as an excuse to reduce openness to a greater extent than is required, for instance, when the real motivation is about corporate competitiveness. We believe that, to the extent that practices around openness are rethought, this should be done transparently, and that when new approaches are incorporated into AI research and publication processes from other domains (e.g. responsible disclosure), those doing so should state their reasons publicly so that a range of stakeholders can evaluate these claims. The debate in the biosecurity community about the appropriate level of disclosure on gain-of-function research (in which organisms are made more dangerous in order to understand certain threats better) provides a model of the kind of discussion we see as healthy and necessary. \n p.56 Interventions Priority Research Area #3: Promoting a Culture of Responsibility AI researchers and the organizations that employ them are in a unique position to shape the security landscape of the AIenabled world. Many in the community already take their social responsibility quite seriously, and encourage others to do the same. This should be continued and further developed, with greater leveraging of insights from the experiences of other technical fields, and with greater attentiveness to malicious use risks in particular. Throughout training, recruitment, research, and development, individuals, and institutions should be mindful of the risks of malicious uses of AI capabilities. Some initial areas to explore for concrete initiatives aimed at fostering a culture of responsibility include : • Education. What formal and informal methods for educating scientists and engineers about the ethical and socially responsible use of their technology are most effective? How could this training be best incorporated into the education of AI researchers? • Ethical statements and standards. What role should ethical statements and standards play in AI research? How and by whom should they be implemented and enforced? What are the domain-specific ethical questions in the areas of digital, physical, and security that need to be resolved in order to distinguish between benign and malicious uses of AI? • Whistleblowing measures. What is the track record of whistleblowing protections in other domains, and how (if at all) might they be used for preventing AI-related misuse risks? • Nuanced narratives. More generally, are there succinct and compelling narratives of AI research and its impacts that can balance optimism about the vast potential of this technology with a level-headed recognition of the risks it poses? Examples of existing narratives include the \"robot apocalypse\" trope and the countervailing \"automation boon\" trope, both of which have obvious shortcomings. Might a narrative like \"dual-use\" (proposed above) be more productive? See Appendix B for more commentary on and questions about these sub-areas. \n p.57 Interventions Priority Research Area #4: \n Developing Technological and Policy Solutions In addition to creating new security challenges and threats progress in AI also makes possible new types of responses and defenses. These technological solutions must be accompanied and supported by well-designed policy responses. In addition to the proposals mentioned in the previous sections, what other potential approaches -both institutional and technological -could help to prevent and mitigate potential misuse of AI technologies? Some initial suggested areas for further investigation include : • Privacy protection. What role can technical measures play in protecting privacy from bad actors in a world of AI? What role must be played by institutions, whether by corporations, the state, or others? • Coordinated use of AI for public-good security. Can AI-based defensive security measures be distributed widely to nudge the offense-defense balance in the direction of defense? Via what institutions or mechanisms can these technologies be promoted and shared? • Monitoring of AI-relevant resources. Under what circumstances, and for which resources, might it be feasible and appropriate to monitor inputs to AI technologies such as hardware, talent, code, and data? • Other legislative and regulatory responses. What other potential interventions by policymakers would be productive in this space (e.g. adjusting legal definitions of hacking to account for the case of adversarial examples and data poisoning attacks)? For all of the above, it will be necessary to incentivize individuals and organizations with the relevant expertise to pursue these investigations. An initial step, pursued by this report, is to raise awareness of the issues and their importance, and to lay out an initial research agenda. Further steps will require commitment from individuals and organizations with relevant expertise and a proven track record. Additional monetary resources, both public and private, would also help to seed interest and recruit attention in relevant research communities. For example, could AI systems be used to refactor existing code bases or new software to adhere more closely to principle of least authority  (Miller, 2006)  or other security best practices? See Appendix B for more commentary on and questions about these sub-areas. \n p.58 When considered together, how will the security-relevant characteristics of AI and the various intervention measures surveyed above (if implemented) combine to shape the future of security? Any confident long-term prediction is impossible to make, as significant uncertainties remain regarding the progress of various technologies, the strategies adopted by malicious actors, and the steps that should and will be taken by key stakeholders. Nonetheless, we aim to elucidate some crucial considerations for giving a more confident answer, and make several hypotheses about the medium-term equilibrium of AI attack and defense. By medium-term, we mean the time period (5+ years from now) after which malicious applications of AI are widely used and defended against, but before AI has yet progressed sufficiently to fully obviate the need for human input in either attack or defense. 5 Strategic Analysis p.59 \n Strategic Analysis Even a seemingly stable and predictable medium-term equilibrium resulting from foreseeable AI developments might be short-lived, since both technological and policy factors will progress beyond what can currently be foreseen. New developments, including technological developments unrelated to AI, may ultimately be more impactful than the capabilities considered in this report. Nevertheless, we hope that the analysis below sheds some light on key factors to watch and influence in the years to come. \n Factors Affecting the Equilibrium of AI and Security Attacker Access to Capabilities Current trends emphasize widespread open access to cutting-edge research and development achievements. If these trends continue for the next 5 years, we expect the ability of attackers to cause harm with digital and robotic systems significantly increase. This follows directly from the dual-use nature, efficiency, scalability, and ease of diffusing AI technologies discussed previously. However, we expect the dual-use nature of the technology will become increasingly apparent to developers and regulators, and that limitations on access to or malicious use of powerful AI technologies will be increasingly imposed. However, significant uncertainty remains about the effectiveness of attempting to restrict or monitor access through any particular intervention. Preemptive design efforts and the use of novel organizational and technological measures within international policing will all help, and are likely to emerge at various stages, in response (hopefully) to reports such as these, or otherwise in the aftermath of a significant attack or scandal. Efforts to prevent malicious uses solely through limiting AI code proliferation are unlikely to succeed fully, both due to less-than-perfect compliance and because sufficiently motivated and well resourced actors can use espionage to obtain such code. However, the risk from less capable actors using AI can likely be reduced through a combination of interventions aimed at making systems more secure, responsibly disclosing developments that could be misused, and increasing threat awareness among policymakers. \n Existence of AI-Enabled Defenses The same characteristics of AI that enable large-scale and low-cost attacks also allow for more scalable defenses. Specific instances of \n p.60 Strategic Analysis AI-enabled defenses have been discussed in earlier sections, such as spam filters and malware detection, and we expect many others will be developed in the coming years. For example, in the context of physical security, the use of drones whose sole purpose is to quickly and non-violently \"catch\" and bring to the ground other drones might be invented and widely deployed, but they might also turn out to be prohibitively expensive, as might other foreseeable defenses. Thus, both the pace of technical innovation and the cost of such defenses should be considered in a fuller assessment. One general category of AI-enabled defenses worth considering in an overall assessment is the use of AI in criminal investigations and counterterrorism. AI is already beginning to see wider adoption for a wide range of law enforcement purposes, such as facial recognition by surveillance cameras and social network analysis. We have hardly seen the end of such advancements, and further developments in the underlying technologies and their widespread use seem likely given the interest of actors from corporations to governments in preventing criminal acts. Additionally, interceding attacks in their early stage through rapid detection and response may turn out to be cheaper than for example widely deploying physical defenses against drones. Thus, the growing ability of states to detect and stop criminal acts, in part by leveraging AI, is a key variable in the medium-term. However, such advances will not help prevent authoritarian abuses of AI. \n Distribution and Generality of Defenses Some defensive measures discussed in Interventions and Appendix B can be taken by single, internally coordinated actors, such as research labs and tech startups, and are likely to happen as soon as they become technically feasible and cost-effective. These measures could then be used by the organizations that have the most to lose from attacks such as governments and major corporations. This means that the most massive category of harm, such as attack on WMD facilities, is also the least likely, though the level of risk will depend on the relative rates at which attacks and defenses are developed. Responsible disclosure of novel vulnerabilities, pre-publication risk assessment, and a strong ethical culture in the AI community more generally will be vital in such a world. This, however, leaves out the strategic situation for the majority of potential victims: technologically conservative corporations, under-resourced states, SMEs, and individuals. For these potential victims, defensive measures need to be baked into widespread technology, which may require coordinated regulatory efforts, or p.61 Strategic Analysis offered at low prices. The latter is most likely to come either from tech giants (as in the case of spam filters), which will increase lock-in and concentration of data and power, or from non-profit organizations who develop and distribute such defensive measures freely or cheaply (e.g. Mozilla's Firefox web browser). This dynamic of defense through reliance on fortified software platforms is likely to be affected by the generality of defensive measures: if each attack requires a tailored defense, and has an associated higher time lag and skill investment, it is more likely that those developing such defensive measures will need financial backing, from corporations, investors, philanthropists, or governments. In the case of governments, international competition may hinder the development and release of defensive measures, as is generally the case in cyber-security, though see the release of CyberChef and Assemblyline as counterexamples. For political security, similar considerations regarding generality apply: a general solution to authenticable multimedia production and forgery detection would be more useful than tailored individual solutions for photographs, videos, or audio, or narrower subsets of those media types. Misaligned incentives can also lead to a failure to employ available defensive measures. For example, better cybersecurity defenses could raise the bar for data breaches or the creation of IoT device botnets. However, the individuals affected by these failures, such as the individuals whose personal data is released or victims of DDOS attacks using botnets, are not typically in a position to improve defenses directly. Thus, other approaches including regulation may be needed to adjust these incentives or otherwise address these externalities . \n Overall Assessment The range of plausible outcomes is extremely diverse, even without considering the outcomes that are less likely, but still possible. Across all plausible outcomes, we anticipate that attempts to use AI maliciously will increase alongside the increase in the use of AI across society more generally. This is not a trend that is particular to AI; we anticipate increased malicious use of AI just as criminals, terrorists and authoritarian regimes use electricity, software, and computer networks: at some point in the technology adoption cycle, it becomes easier to make use of such general purpose technologies than to avoid them. On the optimistic side, several trends look positive for defense. There is much low hanging fruit to be picked in securing AI systems On the pessimistic side, not all of the threats identified have solutions with these characteristics. It is likely to prove much harder to secure humans from manipulation attacks than it will be to secure digital and cyber-physical systems from cyber attacks, and in some scenarios, all three attack vectors may be combined. In the absence of significant effort, attribution of attacks and penalization of attackers is likely to be difficult, which could lead to an ongoing state of low-to medium-level attacks, eroded trust within societies, between societies and their governments, and between governments. Whichever vectors of attack prove hardest to defend against will be the ones most likely to be weaponized by governments, and the proliferation of such offensive capability is likely to be broad. Since the number of possible attack surfaces is vast, and the cutting edge of capability is likely to be ever progressing, any equilibrium obtained between rival states or between criminals and security forces in a particular domain is likely to be short-lived as technology and policies evolve. Tech giants and media giants may continue to become technological safe havens of the masses, as their access to relevant real-time data at massive scale, and their ownership of products and communication channels (along with the underlying technical infrastructure), place them in a highly privileged position to offer tailored protection to their customers. Other corporate giants that offer digitally-enhanced products and services (automotive, medical, defense, and increasingly many other sectors) will likely be under pressure to follow suit. This would represent a continuation of existing trends in which people very regularly interact with and use the platforms provided by tech and media giants, and interact less frequently with small businesses and governments. \n p.63 Strategic Analysis Nations will be under pressure to protect their citizens and their own political stability in the face of malicious uses of AI . This could occur through direct control of digital and communication infrastructure, through meaningful and constructive collaboration between the government and the private entities controlling such infrastructure, or through informed and enforceable regulation coupled with well-designed financial incentives and liability structures. Some countries have a clear head start in establishing the control mechanisms that will enable them to provide security for their citizens . For some of the more challenging coordination and interdisciplinary problems, new leadership will be required to rise above local incentives and provide systemic vision. This will not be the first time humanity has risen to meet such a challenge: the NATO conference at Garmisch in 1968 created consensus around the growing risks from software systems, and sketched out technical and procedural solutions to address over-run, over-budget, hard-to-maintain and bug-ridden critical infrastructure software, resulting in many practices which are now mainstream in software engineering ; the NIH conference at Asilomar in 1975 highlighted the emerging risks from recombinant DNA research, promoted a moratorium on certain types of experiments, and initiated research into novel streams of biological containment, alongside a regulatory framework such research could feed into . Individuals at the forefront of research played key roles in both of these cases, including Edsger Dijkstra in the former and Paul Berg in the latter . There remain many disagreements between the co-authors of this report, let alone amongst the various expert communities out in the world. Many of these disagreements will not be resolved until we get more data as the various threats and responses unfold, but this uncertainty and expert disagreement should not paralyse us from taking precautionary action today. Our recommendations, stated above, can and should be acted on today: analyzing and (where appropriate) experimenting with novel openness models, learning from the experience of other scientific disciplines, beginning multi-stakeholder dialogues on the risks in particular domains, and accelerating beneficial research on myriad promising defenses. \n Chessen, 2017b Naur and Randell, 1969  Krimsky, 1982; Wright, 1994 Dijkstra, 1968 Berg et al., 1974  For example, France's campaign laws prohibited Macron's opponent from further campaigning once Macron's emails had been hacked. This prevented the campaign from capitalizing on the leaks associated with the hack, and ended up with the hack playing a much more muted role in the French election than the Clinton hack played in the US election. \n Conclusion p.64 While many uncertainties remain, it is clear that AI will figure prominently in the security landscape of the future, that opportunities for malicious use abound, and that more can and should be done. Artificial intelligence, digital security, physical security, and political security are deeply connected and will likely become more so. In the cyber domain, even at current capability levels, AI can be used to augment attacks on and defenses of cyberinfrastructure, and its introduction into society changes the attack surface that hackers can target, as demonstrated by the examples of automated spear phishing and malware detection tools discussed above. As AI systems increase in capability, they will first reach and then exceed human capabilities in many narrow domains, as we have already seen with games like backgammon, chess, Jeopardy!, Dota 2, and Go and are now seeing with important human tasks p.65 Conclusion like investing in the stock market or driving cars. Preparing for the potential malicious uses of AI associated with this transition is an urgent task. As AI systems extend further into domains commonly believed to be uniquely human (like social interaction), we will see more sophisticated social engineering attacks drawing on these capabilities. These are very difficult to defend against, as even cybersecurity experts can fall prey to targeted spear phishing emails. This may cause an explosion of network penetrations, personal data theft, and an epidemic of intelligent computer viruses. One of our best hopes to defend against automated hacking is also via AI, through automation of our cyber-defense systems, and indeed companies are increasingly pursuing this strategy. But AI-based defense is not a panacea, especially when we look beyond the digital domain. More work should also be done in understanding the right balance of openness in AI, developing improved technical measures for formally verifying the robustness of systems, and ensuring that policy frameworks developed in a less AI-infused world adapt to the new world we are creating. Looking to the longer term, much has been published about problems which might arise accidentally as a result of highly sophisticated AI systems capable of operating at high levels across a very wide range of environments , though AI capabilities fall short of this today. Given that intelligence systems can be deployed for a range of goals , highly capable systems that require little expertise to develop or deploy may eventually be given new, dangerous goals by hacking them or developing them de novo: that is, we may see powerful AI systems with a \"just add your own goals\" property. Depending on whose bidding such systems are doing, such advanced AIs may inflict unprecedented types and scales of damage in certain domains, requiring preparedness to begin today before these more potent misuse potentials are realizable. Researchers and policymakers should learn from other domains with longer experience in preventing and mitigating malicious use to develop tools, policies, and norms appropriate to AI applications. Though the specific risks of malicious use across the digital, physical, and political domains are myriad, we believe that understanding the commonalities across this landscape, including the role of AI in enabling larger-scale and more numerous attacks, is helpful in illuminating the world ahead and informing better prevention and mitigation efforts. We urge readers to consider ways in which they might be able to advance the collective understanding of the AI-security nexus, and to join the dialogue about ensuring that the rapid development of AI proceeds not just safely and fairly but also securely. \n p.66 The Malicious Use of Artificial Intelligence We are extremely grateful to the many researchers and practitioners who have provided useful comments on earlier versions of this document, and who engaged us in helpful conversations about related topics. Given the number of coauthors and related conversations, we will surely forget some people, but among others, we thank Ian Goodfellow, Ross  \n Event Structure On February 19, the event began with background presentations on cybersecurity, AI, and robotics from relevant experts in these When consulting the history of governing dual use technologies, we should learn both constructive solutions from past successes, and precautionary lessons about poor regulation that should be avoided. A relevant example of the latter is the difficulties of regulating cryptographic algorithms and network security tools through export control measures such as the Wassenaar Arrangement . The similarities between AI and cryptography, in terms of running on general-purpose hardware, in terms of being immaterial objects (algorithms), in terms of having a very wide range of legitimate applications, and in their ability to protect as well as harm, suggest that the default control measures for AI might be similar those that have been historically applied to cryptography. This may well be a path we should avoid, or at least take very cautiously. The apparent dual-use nature of AI technologies raises the following questions: • What is the most appropriate level of analysis and governance of dual-use characteristics of AI technologies (e.g. the field as a whole, individual algorithms, hardware, software, data)? • What norms from other dual-use domains are applicable to AI? • What unique challenges, if any, does AI pose as a dual-use technology? • Are there exemplary cases in which dual-use concerns were effectively addressed? • What lessons can be learned from challenges and failures in applying control measures to dual-use technologies?  \n Red Teaming A common tool in cybersecurity and military practice is red teaming -a \"red team\" composed of security experts and/or members of the organization deliberately plans and carries out attacks against the systems and practices of the organization (with some limitations to prevent lasting damage), with an optional \"blue team\" responding to these attacks. These exercises explore what an actual attack might look like in order to ultimately better understand and improve the security of the organization's systems and practices. Two subsets of the AI security domain seem particularly amenable to such exercises: AI-enabled cyber offense and defense, and adversarial machine learning. While we highlight these subsets because they seem especially relevant to security, red teaming of AI technologies more broadly seems generally beneficial. In addition to this report and the associated workshop, another recent effort aimed at this goal was also conducted by the Origins Project earlier this year . In the case of cyber attacks, many of the concerns discussed earlier in this document, and elsewhere in the literature, are hypothetical. Conducting deliberate red team exercises might be useful in the AI/cybersecurity domain, analogous to the DARPA Cyber Grand Challenge but across a wider range of attacks (e.g. including social engineering, and vulnerability exploitation beyond memory attacks), in order to better understand the skill levels required to carry out certain attacks and defenses, and how well they work in practice. Likewise, in the case of adversarial machine learning, while there are many theoretical papers showing the vulnerabilities of machine learning systems to attack, the systematic and ongoing stresstesting of real-world AI systems has only just begun . Efforts like the CleverHans library of benchmarks and models are a step in this direction , creating the foundation for a distributed open source red teaming effort, as is the NIPS 2017 Adversarial Attacks and Defenses competition , which is more analogous to the DARPA Cyber Grand Challenge. There are several open questions regarding the use of \"red team\" strategies for mitigating malicious uses of AI: • What lessons can be learned from the history to date of \"red team\" exercises? • Is it possible to detect most serious vulnerabilities through \"red team\" exercises, or is the surface area for attack too broad? • Who should be responsible for conducting such exercises, and how could they be incentivised to do so? • What sorts of skills are required to undermine AI systems, and what is the distribution of those skills? To what extent do these skills overlap with the skills required to develop and deploy AI systems, and how should these findings inform the threat model used in red teaming exercises (and other AI security analysis)? • Are there mechanisms to promote the uptake of lessons from \"red team\" exercises? • Are there mechanisms to share lessons from \"red team\" exercises with other organizations that may be susceptible to similar attacks? How to avoid disclosure of attack methods to bad actors? • What are the challenges and opportunities of extending \"red teaming\" (or related practices like tabletop exercises) to AI issues in the physical and political domains? What can be learned for the physical domain from physical penetration testing exercises? \n Formal Verification Formal verification of software systems has been studied for decades . In recent years, it has been shown that even some very complex systems are amenable to formal proofs that they will operate as intended, including the CompCert compiler and the seL4 microkernel . An open question is whether AI systems, or elements thereof, are amenable to formal verification. At the workshop there was substantial skepticism about the prospects for formal AI verification, given the complexity of some modern AI systems, but further analysis about the challenges is required, and research on the topic continues apace . In particular, we might be interested in the following properties being verified for a given system: • that its internal processes in fact attain the goals specified for the system (though noting the existence of the specification problem, i.e. that desired properties of AI systems are often difficult to specify in advance, and therefore difficult to verify), • that its goals will be remain constant in the face of adversaries attempts to change them, • that its ability to be deceived with adversarial inputs is bounded to some extent. \n Verifying Hardware Given the increasing complexity of AI systems, and in some domains limited theoretical foundations for their operation, it may be prohibitively expensive, or even practically or theoretically impossible, to provide an end-to-end verification framework for them. However, it may be feasible to use formal methods to improve the security of components of these systems. Hardware seems particularly amenable to verification, as formal methods have been widely adopted in the hardware industry for decades . \n Verifying Security Additionally, in recent years formal verification has been applied to security protocols to provide robust guarantees of safety against certain types of attacks. The JavaScript prover CryptoVerif is an example of a developer-focused tool that allows programmers to apply formal methods to their code to check correctness in the development process. It should be noted that much of this work is still largely theoretical and adoption in the real world has so far been limited . \n Verifying AI Functionality The notion of being able to prove that a system behaves as intended is an attractive one for artificial intelligence. However, formal methods are difficult to scale up to arbitrary complex systems due to the state space explosion problem. Nonetheless, verification of some aspects of AI systems, such as image classifiers, is still feasible even verification of the behavior of the whole system is prohibitively complex. For example, work on verification of deep neural networks provided a method to check for the existence of adversarial examples in regions of the input space . Responsible \"AI 0-Day\" Disclosure As discussed above, despite the successes of contemporary machine learning algorithms, it has been shown time and again that ML algorithms also have vulnerabilities. These include ML-specific vulnerabilities, such as inducing misclassification via adversarial   (Stevens et al., 2016) . There is currently a great deal of interest among cyber-security researchers in understanding the security of ML systems, though at present there seem to be more questions than answers. In the cybersecurity community, \"0-days\" are software vulnerabilities that have not been made publicly known (and thus defenders have zero days to prepare for an attack making use of them). It is common practice to disclose these vulnerabilities to affected parties before publishing widely about them, in order to provide an opportunity for a patch to be developed. Should there be a norm in the AI community for how to disclose such vulnerabilities responsibly to affected parties (such as those who developed the algorithms, or are using them for commercial applications)? This broad question gives rise to additional questions for further research: • As AI technologies become increasingly integrated into products and platforms, will the existing security norm around responsible disclosure extend to AI technologies and communities? • Should AI systems (both existing and future) be presumed vulnerable until proven secure, to an extent that disclosing new vulnerabilities privately is unnecessary? • In what safety-critical contexts are AI systems currently being used? • Which empirical findings in AI would be useful in informing an appropriate disclosure policy (analogous to the way that historical trends in 0-day discoveries and exploitation rates are discussed in cybersecurity analyses )? • If such a norm were appropriate in broad terms, who should be notified in case a vulnerability is found, how much notice should be given before publication, and what mechanisms should institutions create to ensure a recommendation is processed and potentially acted upon? • What is the equivalent of \"patching\" for AI systems, and how should trade-offs (e.g. between resource demands, accuracy and robustness to noise) and prioritization amongst the variety of possible defense measures be weighed in a world of rapidly changing attacks and defenses?  Szegedy et al., 2013; Evtimov et al., 2017; Carlini et al., 2016 Rubinstein et al., 2009 Šrndic and Laskov, 2014  e.g.  Ablon and Bogart, 2017  p.84 Appendix B: Questions for Further Research \n AI-Specific Exploit Bounties To complement the norm of responsible disclosure of vulnerabilities (discussed above), which relies on social incentives and goodwill, some software vendors offer financial incentives (cash bounties) to anyone who detects and responsibly discloses a vulnerability in their products. With the emergence of new AIspecific vulnerabilities, some questions arise: • Are existing vulnerability bounties likely to extend to AI technologies? • Should we expect, or encourage, AI vendors to offer bounties for AI-specific exploits? • Is there scope to offer bounties by third parties (e.g. government, NGO, or philanthropic source) in cases where vendors are unwilling or unable to offer them, for example in the case of popular machine learning frameworks developed as open-source projects or in academia? \n Security Tools In the same way software development and deployment tools have evolved to include an increasing array of security-related capabilities (testing, fuzzing, anomaly detection, etc.), could we start envisioning tools to test and improve the security of AI components and systems integrated with AI components during development and deployment, such that they are less amenable to attack? These could include: • Automatic generation of adversarial data • Tools for analysing classification errors • Automatic detection of attempts at remote model extraction or remote vulnerability scanning • Automatic suggestions for improving model robustness (see e.g.  \n Secure Hardware Hardware innovation has accelerated the pace of innovation in machine learning, by allowing more complex models to be trained, enabling faster execution of existing models, and facilitating more rapid iteration of possible models. In some cases, this hardware is generic (commercial GPUs), but increasingly, AI (and specifically machine learning) systems are trained and run on hardware that is semi-specialized (e.g. graphics processing units (GPUs)) or fully specialized (e.g. Tensor Processing Units (TPUs)). This specialization could make it much more feasible to develop and distribute secure hardware for AI-specific applications than it would be to develop generic secure hardware and cause it to be widely used. At the workshop we explored the potential value of adding security features to AI-specific hardware. For example, it may be possible to create secure AI hardware that would prevent copying a trained AI model off a chip without the original copy first being deleted. Such a feature could be desirable so that the total number of AI systems (in general or of a certain type or capability level) could be tightly controlled, if the capabilities of such AI systems would be harmful in the wrong hands, or if a large-scale diffusion of such AI systems could have harmful economic, social or political effects. Other desirable secure hardware features include hardwarelevel access restrictions and audits. One research trajectory to be considered is developing a reference model for secure AIspecific hardware, which could then be used to inform hardware engineering and, ultimately, be adopted by hardware providers. It may also be the case that potential security threats from AI will drive research in secure hardware more generally, not just for the hardware running AI systems, as a response measure to changes in the cyber threat landscape. Note, however, the potential for manufacturers to undermine the security of the hardware they produce; hardware supply chain vulnerabilities are currently a concern in the cybersecurity context, where there is fear that actors with control over a supply chain may introduce hardwarebased vulnerabilities in order to surveil more effectively or sabotage cyber-physical systems . Finally, note that for other security-relevant domains such as cryptography, tamper-proof hardware has been developed , with features such as tamper evidence (making it clear that tampering has occurred when it has occurred) and obscurity of layout design (such that it is prohibitively difficult to physically examine the workings of the chip in order to defeat it). Tamper-proof hardware could potentially be valuable so that outsiders are unable to discern the inner workings of an AI system from external emission; so that stolen hardware cannot be used to duplicate an AI; and so that organizations can credibly commit to operating a system in a safe and beneficial way by hard-coding certain software properties in a chip that, if tampered with, would break down. However, secure processors tend to cost significantly more than insecure processors and, to our knowledge, have not specifically been developed for AI purposes. There are many open questions in this domain: • What, if any, are the specific security requirements of AI systems, in general and in different domains of application? • Would changes in the risk landscape (as surveyed above) provide sufficient incentive for a major overhaul of hardware security? • What set of measures (e.g. reference implementation) would encourage adoption of secure hardware? • What measures, if any, are available to ensure compliance with hardware safety requirements given the international distribution of vendors and competing incentives such as cost, potential for surveillance and legal implications of auditability? • How applicable are existing secure processor designs to the protection of AI systems from tampering? • Could/should AI-specific secure processors be developed? • How could secure enclaves be implemented in an AI context ? • Can secure processors be made affordable, or could policy mechanisms be devised to incentivize their use even in the face of a cost premium? Pre-Publication Risk Assessment in Technical Areas of Special Concern By pre-publication risk assessment we mean analyzing the particular risks (or lack thereof) of a particular capability if it became widely available, and deciding on that basis whether, and to what extent, to publish it. Such norms are already widespread in the computer security community, where e.g. proofs of concept rather than fully working exploits are often published. Indeed, such considerations are sufficiently widespread in computer security 1 2  Anderson, 2008  as suggested by  Stoica et al., 2017  p.87 Appendix B: Questions for Further Research that they are highlighted as criteria for submission to prestigious conferences . Openness is not a binary variable: today, many groups will publish the source code of a machine learning algorithm without specifying the hyperparameters to get it to work effectively, or will reveal details of research but not give details on one particular component that could be part of a crucial data ingestion (or transformation) pipeline. On the spectrum from a rough idea, to pseudocode, to a trained model along with source code and tutorials/tips on getting it to work well in practice, there are various possible points, and perhaps there are multiple axes (see Figure  3  ). Generally speaking, the less one shares, the higher the skill and computational requirements there are for another actor to recreate a given level of capability with what is shared: this reduces the risk of malicious use, but also slows down research and places barriers on legitimate applications. For an example of a potentially abusable capability where full publication may be deemed too risky, voice synthesis for a given target speaker (as will reportedly soon be available as a service from the company Lyrebird ) is ripe for potential criminal applications, like automated spearphishing (see digital security section) and disinformation (see political security section). On the other hand, as is the case with other technologies with significant potential for malicious use, there could be value in openness for security research, for example in white hat penetration testing. As described in the Rethinking Openness section of the report, there are clear benefits to the level of openness currently prevalent in machine learning as a field. The extent to which restrictions on publication would affect these benefits should be carefully considered. If the number of restricted publications is very small (as in biotechnology, for example), this may not be a significant concern. If, however, restricted publication becomes common, as in the case of vulnerability disclosure in cybersecurity research, then institutions would need to be developed to balance the needs of all affected parties. For example, responsible disclosure mechanisms in cybersecurity allow researchers and affected vendors to negotiate a period of time for a discovered vulnerability to be patched before the vulnerability is published. In addition to the commercial interests of vendors and the security needs of users, such schemes often also protect researchers from legal action by vendors. In the case of AI, one can imagine coordinating institutions that will withhold publication until appropriate safety measures, or means of secure deployment, can be developed, while allowing the researchers to retain priority claims and gain credit for their work. Some AI-related discoveries, as in the case of  adversarial examples in the wild, may be subsumed under existing responsible disclosure mechanisms, as we discuss below in \"Responsible AI 0-day Disclosure\". Some valuable questions for future research related to prepublication research assessment include: • What sorts of pre-publication research assessment would AI researchers be willing to consider? To what extent would this be seen as conflicting with norms around openness? • What can be learned from pre-publication risk assessment mechanisms in other scientific/technological domains? • Is it possible to say, in advance and with high confidence, what sorts of capabilities are ripe for abuse? • What sort of heuristics may be appropriate for weighing the pros and cons of opening up potentially-abusable capabilities? • How can such assessment be incorporated into decisionmaking (e.g. informing one's openness choices, or incorporating such analysis into publications)? • Can we say anything fine-grained yet generalizable about the levels of skill and computational resources required to recreate capabilities from a given type (code, pseudocode, etc.) of shared information? • How does the community adopt such a model in the absence of regulation? Central Access Licensing Models Another potential model for openness is the use of what we call central access licensing. In this model, users are able to access certain capabilities in a central location, such as a collection of remotely accessible secure, interlinked data centers, while the underlying code is not shared, and terms and conditions apply to the use of the capabilities. This model, which is increasingly adopted in industry for AI-based services such as sentiment analysis and image recognition, can place limits on the malicious use of the underlying AI technologies. For example, limitations on the speed of use can be imposed, potentially preventing some large-scale harmful applications, and terms and conditions can explicitly prohibit malicious use, allowing clear legal recourse. \n p.90 Appendix B: Questions for Further Research Centralised access provides an alternative to publication that allows universal access to a certain capability, while keeping the underlying technological breakthroughs away from bad actors (though also from well-intentioned researchers). Note though that black box model extraction may allow bad actors to gain access to the underlying technology. Additionally, similarly to early proposals for, in effect, an information processing \"tax\" on emails in order to disincentivize spam , centralized AI infrastructures better enable constraints to be placed on the use of AI services, such that large-scale attacks like automated spear phishing could be made less economical (though see Laurie and Clayton, 2004 for a criticism of this approach, and  Liu and Camp, 2006  for further discussion; the increased interest in crypto-economics following the success of bitcoin may lead to advances in this area). Finally, note that the concentration of AI services in a particular set of organizations may heighten potential for malicious use at those organizations, including by those acting with the blessing of the relevant organization as well as by insider threats. Indeed, some workshop attendees considered these risks from concentration of power to be the biggest threat from AI technologies; note, however, that in this report we have decided to focus on direct malicious use risks, rather than systemic threats (see Scope). In addition to monopolistic behavior, there are more subtle risks such as the introduction of \"backdoors\" into machine learning systems that users may be unaware of . Some initial research questions that arise related to a central access licensing model: • What sorts of services might one want only available on a per-use basis? • How effectively can a service provider determine whether AI uses are malicious? • How can a user determine whether a service provider is malicious ? • Is the proposal technologically, legally and politically feasible? • Who might object to a centralised access model and on what grounds? • Is there enough of a technology gap such that actors without access cannot develop the technologies independently? • Should international agreements be considered as tools to incentivize collaboration on AI security? • What should the AI security community's \"public policy model\" be -that is, how should we aim to affect government policy, what should the scope of that policy be, and how should responsibility be distributed across individuals, organizations, and governments? • Should there be a requirement for non-human systems operating online or otherwise interacting with humans (for example, over the telephone) to identify themselves as such (a \"Blade Runner law\" ) to increase political security? • What kind of process can be used when developing policies and laws to govern a dynamically evolving and unpredictable research and development environment? • How desirable is it that community norms, ethical standards, public policies and laws all say the same thing and how much is to be gained from different levels of governance to respond to different kinds of risk (e.g. near term/long term, technical safety / bad actor and high uncertainty / low uncertainty risks)? It seems unlikely that interventions within the AI development community and those within other institutions, including policy and legal institutions, will work well over the long term unless there is some degree of coordination between these groups. Ideally discussions about AI safety and security from within the AI community should be informing legal and policy interventions, and there should also be a willingness amongst legal and policy institutions to devolve some responsibility for AI safety to the AI community, as well as seeking to intervene on its own behalf. Achieving this is likely to require both a high degree of trust between the different groups involved in the governance of AI and a suitable channel to facilitate proactive collaboration in developing norms, ethics education and standards, policies and laws; in contrast, different sectors responding reactively to the different kinds of pressures that they each face at different times seems likely to result in clumsy, ineffective responses from the policy and technical communities alike. These considerations motivated our Recommendations #1 and #2. Figure 1 : 1 Figure 1: Recent progress in image recognition on the ImageNet benchmark. Graph from the Electronic Frontier Foundation's AI Progress Measurement project (retrieved August 25, 2017). \n Figure 3 . 3 Figure 3. A schematic illustration of the relationship between openness about an AI capability and the skill required to reproduce that capability. \n Anderson et al. created a machine learning model to automatically generate command and control domains that are indistinguishable from legitimate domains by human and machine observers. These domains are used by malware to \"call home\" and allow malicious actors to communicate with the host machines. Anderson et al. also leveraged reinforcement learning to create an intelligent agent capable of manipulating a malicious binary with the end goal of bypassing NGAV detection. Similarly, Kharkar et al. applied adversarial machine learning to craft malicious documents that could evade PDF malware classifiers. \n Security Domains like safety-critical facilities and infrastructure (e.g. airports), the owners of which can afford to invest in such protection, as opposed to the much more widely distributed \"soft targets\" (such as highly populated areas). Physical defenses can include detection via radar, lidar, acoustic signature, or image recognition software ; interception through various means ; and passive defense through physical hardening or nets. The U.S. Department of Defense has recently launched a major program to defend against drones, and has tested lasers and nets with an eye towards defending against drones from the Islamic State in particular . Mouawad, 2015; Vincent, 2016 The Telegraph, 2016 Crootof, 2015 Crootof and Renz, 2017 DoD, 2012 DoD, 2015; Roff, 2016b ICRC, 2017; Scharre, 2018 p. 43 \n Strategic Analysisthemselves, and securing people and systems from AI-enabled attacks. Examples include responsible vulnerability disclosure for machine learning in cases where the affected ML technology is being used in critical systems, and greater efforts to leverage AI expertise in the discovery of vulnerabilities by software companies internally before they are discovered by adversaries. There are substantial academic incentives to tackle the hardest research problems, such as developing methods to address adversarial examples and providing provable guarantees for system properties and behaviors. There are, at least in some parts of the world, political incentives for developing processes and regulations that reduce threat levels and increase stability, e.g. through consumer protection and standardization. Finally, there are incentives for tech giants to collaborate on ensuring at least a minimal level of security for their users. Where solutions are visible, require limited or pre-existing coordination, and align with existing incentive structures, defenses are likely to prevail. GCHQ, 2016 CSE, 2017 Moore and Anderson, 2012 p.62 \n Anderson, Nicholas  Papernot, Martín Abadi, Tim Hwang, Laura Pomarius, Tanya Singh  Kasewa, Smitha Milli, Itzik Kotler, Andrew Trask, Siddharth  Garg, Martina Kunz, Jade Leung, Katherine Fletcher, Jan Leike, Toby Ord, Nick Bostrom, Owen Cotton-Barratt, Eric Drexler, Julius Weitzdorfer, Emma Bates, and Subbarao Kambhampati. Any remaining errors are the responsibility of the authors. This work was supported in part by a grant from the Future of Life Institute. The Malicious Use of Artificial Intelligence Summary On February 19 and 20, 2017, Miles Brundage of the Future of Humanity Institute (FHI) and Shahar Avin of the Centre for the Study of Existential Risk (CSER) co-chaired a workshop entitled \"Bad Actor Risks in Artificial Intelligence\" in Oxford, United Kingdom. The workshop was co-organized by FHI, CSER, and the Leverhulme Centre for the Future of Intelligence (CFI). Participants came from a wide variety of institutional and disciplinary backgrounds, and analyzed a variety of risks related to AI misuse. The workshop was held under Chatham House rules. p.74 p.75 \n Appendix B: Questions for Further Research examples or via poisoning the training data ; see Barreno et al. (2010) for a survey. ML algorithms also remain open to traditional vulnerabilities, such as memory overflow Harrison, 2010; Wahby, R. 2016 Blanchet, 2017 though there are some instances of real world use -see e.g. Beurdouche et al., 2017 Katz et al., 2017 p.83 \n Appendix B: Questions for Further Research can legal but related tactics like search engine optimization be dealt with if so)? Bastani et al., 2017 Dwork and Naor, 1993 Gu et al., 2017 see e.g. Ghodsi et al., 2017 p.91 p.99 \n\t\t\t  see e.g. Spafford, 1988    \n\t\t\t Scenarios Scenarios \n\t\t\t Arulkumaran et al., 2017  Seymour and Tully, 2016) Calabresi, 2017    Litan, 2017  \n\t\t\t Roff, 2016aFranke, 2016 Standage, 2017    Roff, 2016a e.g. Kolodny, 2017; Wiggers, 2017   \n\t\t\t Booth et al., 2017 Lucas, 2017   \n\t\t\t Zeitzoff, 2017 Aker and Kalkan, 2017 Yin, 2015 Scharre, 2015 Schmitt, 2017 Scharre, 2015   \n\t\t\t Seymour and Tully, 2016 Everett et al., 2016 Shao et al., 2017 Chung, Jamaludin, and Zisserman, 2017 Browne, 2017 Adams, 2017 Serban et al., 2017   \n\t\t\t Eckersley and Nasser et al., 2017   \n\t\t\t see e.g.NDSS, 2018    \n\t\t\t see e.g. Kesarwani et al., 2017   \n\t\t\t • What information types should limited sharing apply to: code, research papers, informal notes?• How can sufficient trust be established between groups such that this kind of coordination is seen as mutually beneficial?", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/1802.07228.tei.xml", "id": "0c834abe88d1f177650600e9b37559b8"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "The invention of atomic energy posed a novel global challenge: could the technology be controlled to avoid destructive uses and an existentially dangerous arms race while permitting the broad sharing of its benefits? From 1944 onwards, scientists, policymakers, and other technical specialists began to confront this challenge and explored policy options for dealing with the impact of nuclear technology. We focus on the years 1944 to 1951 and review this period for lessons for the governance of powerful technologies, and find the following: Radical schemes for international control can get broad support when confronted by existentially dangerous technologies, but this support can be tenuous and cynical. Secrecy is likely to play an important, and perhaps harmful, role. The public sphere may be an important source of influence, both in general and in particular in favor of cooperation, but also one that is manipulable and poorly informed. Technical experts may play a critical role, but need to be politically savvy. Overall, policymaking may look more like \"muddling through\" than clear-eyed grand strategy. Cooperation may be risky, and there may be many obstacles to success. 1 For helpful input on this work, we thank Nick Bostrom, Diane Cooke, Alex Debs, Jeff Ding, Jade Leung, Sören Mindermann, and especially Markus Anderljung and Carl Shulman. We want to also thank those who have worked in this space with us: Carrick Flynn championed this topic early on; Toby Ord has expertly examined the earlier period of the development of nuclear weapons for similar lessons;", "authors": ["Waqar Zaidi", "Allan Dafoe"], "title": "International Control of Powerful Technology: Lessons from the Baruch Plan for Nuclear Weapons", "text": "Introduction Humanity is likely to confront novel powerful technologies and weapons in the coming decades, including those which could emerge from developments in artificial intelligence. Developing and deploying these in a way that is good for humanity (the governance problem) may be hard, owing to structural features of the risks and decision-making context.  2  On the other hand, radical levels of cooperation become more feasible in light of truly existentially dangerous technology: in which the gains from coordination are tremendous, the losses from failed coordination terrible,  3  and where most actors' long-term interests are aligned. We might hope that powerful individuals would set aside their narrow self-interest and perspectives, and work together to secure for humanity a flourishing future. They might do this because deep down they believe this is what most matters; because of status motivations to leave a legacy of securing this historical achievement; because of social pressure from their peers, family, or the public; or for other motivations. We hope that individuals-when confronted with a decision that could take humanity towards flourishing and away from existential harm-would make the right decision. But would they? This paper looks to the development and attempted governance of nuclear technology for lessons on this question. It focuses on the uncertain early years of this technology, especially 1943 to 1951. Policymakers, statesmen, scientific and technical specialists, and other intellectuals attempted to understand the nature and impact of nuclear technology and devise governance for the dangers that many saw. They saw nuclear technology, just as we see some technologies today, as bringing great promise, but also great threats. This dual-use nature led them to conclude that simply banning nuclear technology was not an option, as then its potential benefits would be lost. Instead, proponents argued, the world needed to devise international governance mechanisms which would both reduce the risks but also allow the beneficial outcomes to emerge.  4  This document is organized by lessons and recommendations. These distil and generalize the lessons that may be applicable to future efforts to control those technologies that pose significant risks of misuse and accident, but that also come with substantial military and economic advantage. We believe these lessons will help those 5 participating in conversations on the governance of such technologies by highlighting historical parallels and expanding the space of conceivable, and considered, political dynamics and opportunities. Through which processes might governance be discussed and set up? What problems might policymakers and other interested parties need to anticipate when thinking about governance? Who might support proposals for governance, and how and why? How sincere or cynical will participants be? How likely is it that key actors will misunderstand the problem or miscommunicate their preferences? This report begins with a historical overview of proposals for the \"international control of atomic energy\" (that is, international regulation of atomic weapons and underlying technologies, sciences, and materials) between 1944 and 1946, followed by key dates and short summaries of the key proposals. In summary, we find that radical schemes for international governance can get widespread support, even from skeptics, but that the support can be tenuous and fleeting. Technical experts can bolster support, but muddled policymaking, secrecy, and concerns over security can undermine it. We highlight the following lessons for those thinking about technological governance today: 1. Radical proposals which would normally appear naive or extreme may, in the right circumstances, gain traction and be seriously proposed, discussed, and even adopted as official policy. 2. Groups or coalitions supporting (or opposing) international control will contain individuals who have different reasons and rationale for their support (or opposition). 3. The support of realists is possible and possibly even crucial for international control to become policy.  6  4. Secrecy and security will play a central role in discussions on the governance of powerful technologies. 5. The public sphere will likely have a powerful impact on debates regarding international control. 6. Technical experts and specialists (scientists, engineers, technicians, academics) have significant power to shape proposals and policy, though their opponents may criticize them for political naivety. 7. Policymaking involves significant muddling through, rather than grand strategy. It is also deeply affected by domestic politics and often develops on the basis of short-term objectives, poorly thought-out criteria, and poor quality information. Policymaking may develop in unexpected directions or for the expected reasons. 8. Achieving agreement on a workable scheme for international control is difficult. 9. Attempts at cooperation come with risks of strategic, diplomatic, political, and technological losses. The lessons have been organized so that readers may skip the historical case expositions if they wish. The cases do, however, flesh out the bare bones lessons, and unpack and explore the various aspects of the lesson in more detail. For those wanting more detail, we have included a list of key events and short biographies of central figures in the appendices. For further reading, we would also point readers to the rich historical literature on the politics of atomic energy in its early years, much of which is cited in the footnotes and listed in the References section. We have relied on a variety of secondary (and some primary) sources but have found Gregg Herken's The Winning 6 By \"realists\" we mean those who understand international relations in terms of power and national interest, and prefer policies which preserve and strengthen their state in relation to others. Jack Donnelly, Realism and International Relations (Cambridge: Cambridge University Press, 2004), pp. 7-8. Weapon to be the single most detailed and reliable source. We would recommend this as the first port of call for 7 any reader interested in further exploring the history of the international control of nuclear weapons. \n Nuclear Technology as an Analogy History can provide a rich source for insight about novel policy challenges. To understand the challenges of governing today's emerging powerful technologies, one can examine attempts at the governance of earlier powerful technologies when they first emerged. Of the various technologies for which international governance regimes were created or contemplated in the twentieth century (including, for example, aviation, chemical and biological weapons, telecommunications, and the internet), nuclear technology stands out as a particularly promising candidate for study of the pressing, but thorny, problem of international control. In particular, we 8 would highlight the following properties which make this case relevant to understanding efforts to control a future powerful technology (such as AI): (a) Nuclear technology was marked out as a powerful technology when it was first revealed, and policymaking was made within a context that took its potential impact seriously. (b) Because of the sudden way the atomic bomb was revealed, and its seemingly esoteric nature, there was significant uncertainty about its impact. Consequently, there was a rich public and policymaking debate about the nature of this technology and its impact. As well as strategic and political dimensions, this debate included an ethical dimension. (c) Many people, including many elites, perceived nuclear technology as an existential risk and so 9 engendered a rich policymaking debate on international governance, known then as \"international control.\" (d) Elements of national competition and negotiation, and of a technological arms race, were present during the early history of nuclear weapons. (e) Nuclear technology rested on complex, rapidly developing science. Nevertheless, readers should be aware that there are a number of ways in which this historical moment and nuclear technology are a poor analogy for the future governance of powerful technologies. Consider the following disanalogies to AI: • Private Sector Involvement : AI is primarily being developed and deployed by the private sector, and the private sector is likely to continue to push forward the science of AI irrespective of what governments do. Nuclear technology in its earliest decades was controlled and funded by states. • Secrecy : While nuclear technologies were heavily guarded secrets (though the basic science was broadly known), artificial intelligence technology is more international, broadly held, and public. • Impact and Proliferation : AI is already a major economic technology and deployed around the world, whereas the economic value of nuclear technology was unclear and it was deployed in only a few locations in the period  [1943] [1944] [1945] [1946] [1947] [1948] [1949] [1950] [1951] . AI is deployed and innovated in a greater number of fields as compared to nuclear technology, and offers greater future economic potential. The barriers to entry for the development and deployment of AI are also lower than in the case of nuclear technology. • Discernibility of Risks : It may be easier to understand how nuclear weapons could be dangerous, whereas the accident risks from AI are more subtle, theory dependent, or fantastic seeming. • Safety Difficulty : The accident risks from nuclear weapons are likely easier to manage than from AI, because nuclear bombs or power plants are not complex adaptive (intelligent) systems. • Verification : It is easier to unilaterally verify nuclear developments (nuclear tests, ICBM deployments), and it appears easier to control the nuclear supply chain with relatively low disruption of industry. • Strategic Value : The strategic value of nuclear weapons plateaus once one has secure second-strike capability, whereas from the present vantage point, there is no obvious plateau in AI's strategic value. Further, the historical context for the early development of nuclear technology differs in important ways from the current and future moments in which attempts to govern other powerful technologies may be made: • Postwar Context : Atomic development occurred at the end of a war widely seen as catastrophic. This led to a very different social and political context within which nuclear weapons were introduced. • Visceral Example of Danger : The world witnessed the use of nuclear weapons to destroy cities and some of the horrors this entailed. Future technology risks may not produce visceral harms in advance of attempts to govern them. • Superpower Relationships : Nuclear control negotiations took place between powers who were allies and had just suffered through this war. These powers also had incompatible political-economic models and so found themselves in much more zero-sum relations than the great powers of today. • Information : The great powers had less cultural and ideological commonality, and less information about each other, than do the great powers, and their publics, of today. Readers should be aware of two further methodological caveats: • n=1 . To some extent, this historical period represents a single observation (n=1) in that a single large shock to decision processes could have led to different outcomes. The implication of this is that we should not primarily use the outcome as our evidence, but should instead inspect all the informative historical moments throughout the episode for insight into historical dynamics and mechanisms. For example, we can learn from the rich ways in which decision makers responded to information, formed beliefs, and devised strategy. These caveats having been stated, this case study represents a rare historical moment when great powers seriously discussed strategies for avoiding an arms race in a new technology, and where influential people 11 within the state with the technological monopoly seriously considered giving up their monopoly. Furthermore, this historical episode took place between relatively modern great powers and at a time when U.S. elite and public culture was not entirely dissimilar to today's; for example, the media was important in informing the public and shaping its opinion on major events, politicians took public opinion into account when making policy decisions, interservice rivalry played a role in some policy decisions, and policymaking was done through a mix of committees, experts, career statesmen, and trusted advisors. \n Historical Overview \n Summary By the start of World War II, scientists around the world were aware that the construction of a bomb based on the release of atomic energy was theoretically possible. Britain was the first to start a concerted bomb program, joined soon after by the United States, Germany, Japan, and the Soviet Union. The German and Japanese programs did not progress far. Britain, faced with more pressing resource requirements, eventually paused its program and transferred its expertise into the Manhattan Project, joining the U.S. program as a junior partner. The Manhattan Project, begun in October 1941, led to a working bomb that was tested in July 1945. Atomic bombs were dropped on the Japanese cities of Hiroshima on August 6 and Nagasaki on August 9. Japan announced its surrender on August 15, and the signing of the formal surrender treaty on September 2 brought the Second World War officially to a close. The use of the atomic bomb led to an acceleration of the Soviet bomb project and a restart of the British project. Even before the bomb was used, scientists expressed concern about its destructiveness and a possible arms race after the war. Senior Danish physicist Niels Bohr brought these concerns to the attention of British Prime Minister Winston Churchill in May 1944 and U.S. President Franklin D. Roosevelt in August 1944. By mid 1945, scientists working on the Manhattan Project also became concerned about the impact of atomic weapons and issued a series of warnings to the government. Many of the suggestions for dealing with the bomb called for the \"international control of atomic energy\" (that is, effective international regulation of atomic weapons and the underlying science and technology through multilateral agreements or an international organization). Various proposals for international control were made from late 1944 onwards. These became widespread after the atomic bomb was made public in August 1945, and by the end of the year, scientists had organized themselves into various groups calling for international control. State officials also, at times, considered adopting international control as policy, and there was discussion on atomic matters with the Soviet Union. In late December 1945, Stalin agreed to the formation of a United Nations Atomic Energy Commission to study the \"control of atomic energy,\" and the United Nations General Assembly authorized its formation in January 1946. In January 1946, Secretary of State James F. Byrnes authorized the formation of a committee (chaired by Under Secretary of State Dean Acheson and ex-Chairman of the Tennessee Valley Authority David Lilienthal) to study the international control of atomic energy. This committee in turn asked a group of consultants (led by prominent physicist J. Robert Oppenheimer) to prepare a policy proposal for international control for government consideration. The committee completed its detailed plan for international control, dubbed the Acheson-Lilienthal Report, in March 1946. The report was adopted, with important modifications, by the first U.S. representative to the newly formed United Nations Atomic Energy Commission (UNAEC), Bernard Baruch, who presented his so-called Baruch Plan at the UNAEC in June 1946. The Soviet Union, implicitly rejecting this plan, responded with its own proposal a few days later (the Gromyko Plan). Subsequent negotiations with the Soviet Union were carried out amidst deteriorating relations between the superpowers and failed by the end of the year. In our assessment of this case, failure was overdetermined (see Section Could International Control Have Succeeded ). There was a great divergence in U.S.-Soviet expectations and conflicting interests around the world. Mistrust had also been growing since early 1945. By mid 1946, the U.S. administration had given up whatever hope it had in international control and only carried out negotiations for propaganda purposes.   \n Key Dates For a detailed chronology, see Appendix A . For a brief biography of some of the key historical figures, see Appendix B .  The Bush-Conant Memo, September 1944  Senior science policymakers Vannevar Bush (head of the U.S. Office of Scientific Research and Development) and James B. Conant prepared a policy suggestion on atomic energy for Secretary Henry L. Stimson in September 1944. They advised that other countries could catch up with the U.S. within four years. They suggested sharing scientific information with all countries; only manufacturing and military details were to remain secret. Excessive secrecy could lead to an arms race with the Soviet Union. It was not possible for the U.S. to monopolize raw materials going forward. Their only concrete suggestion for international control was free flow of scientific information and international inspections through an international organization.  14  The  Bush Plan, November 1945  Vannevar Bush presented his proposal for international control in a memorandum to Secretary of State Byrnes in November 1945. The proposal was significantly more detailed than earlier ones and emphasized, for the first time, a staged process, at the end of which the U.S. would give up atomic weapons. The stages were, first, the formation of a U.N. agency for the dissemination of scientific information, including free access for scientists to basic research. Second came the establishment of a U.N. inspection system, with a gradual exchange of information on raw materials and facilities. This was to culminate with the sharing of the most practical and most secret atomic know-how. Nations were to agree to use such information for commercial purposes only. Finally, the U.S. was to convert its bombs to peaceful uses. This plan was adopted by Byrnes (in a somewhat 15 modified form) in December. Byrnes had formerly taken a more hawkish position on the bomb and the Soviet Union but accepted this cooperative approach after he realized that the U.S.'s atomic bomb was not helping in geopolitical negotiations with the Soviets.  16  The  Cohen-Pasvolsky Plan, December 1945  This was the official State Department plan, drawn up by a committee headed by State Department officials Benjamin Cohen and Leo Pasvolsky. The plan was accepted by Byrnes and presented to the Soviets in the December 1945 Moscow conference. The Cohen-Pasvolsky Plan was based on the November 1945 Bush Plan, but included one crucial change; although it emphasized stages, it added that the international control process could move onto the next stage without having completed the previous stage. The stages thus did not have to progress in a strict sequence.  17  The  Acheson-Lilienthal Plan, March 1946  The Acheson-Lilienthal Report set out a plan produced by a group of expert consultants (including the leading Manhattan Project physicist J. Robert Oppenheimer) for the State Department in March 1946. This was the single most detailed proposal for international control produced in the U.S. and was the basis (with crucial changes) for the official U.S. proposal-the so-called Baruch Plan-at the United Nations Atomic Energy Commission (UNAEC) a few months later.  18  The Acheson-Lilienthal Plan was premised on the assumption that inspections were insufficient for international control. Instead, the U.N., through an Atomic Development Authority (ADA), was to control all fissionable raw materials and have a monopoly on all \"dangerous\" activities (i.e., those with military applications). States would shut down all dangerous activities, and all atomic material would be transferred to U.N. ownership. Peaceful development (R&D, power plants), however, could continue in states. The United States would begin a phased transition of its bombs, material, and facilities to the ADA, once set up. The U.S.  15  The  Bush   would not cease atomic operations prior to the setting up of the ADA. The plan placed significant emphasis on the cooperation of internationalist scientists working at the ADA.  19  The ADA was the centerpiece of the Acheson-Lilienthal Plan. It was to set up large R&D centers and conduct research on peaceful and warlike uses of atomic energy. It would also have its own operational reactors. These reactors and other atomic facilities were to be spread across a number of (unspecified) countries in a \"strategic balance among nations\" so that in the event of a breakdown of the ADA (or the U.N. itself) there would be a \"balance of facilities\" across states. This, it was hoped, would reduce the fears of any one state that joining would undermine its security in the event of a diplomatic breakdown. The ADA would own and operate all mining, refining, and production of fissionable raw materials. Existing mines, plants, and factories (e.g., at Hanford and Oak Ridge) were to be transferred to the control of the ADA. It would dispense \"denatured\" 20 fissionable raw materials to individual nations for their nuclear power plants and license and inspect their (civilian) nuclear facilities.  21  The Baruch Plan, June 1946 The Baruch Plan was developed by Bernard Baruch-a businessman and financier who was the U.S. representative on the UNAEC-between March and June 1946. The plan was adopted as official U.S. policy and presented at the UNAEC in June 1946. It was based on the Acheson-Lilienthal Report but included some 22 crucial changes that made the plan more hawkish and pro-business. There were likely several reasons for these changes. One might have been to help Baruch appear as the author of the proposal, rather than just a \"messenger boy\" for the Acheson-Lilienthal Report. Another might have been to reduce the risk to the U.S. if 23 the plan failed or the Soviet Union reneged. A third may have been to retain private sector autonomy in the nuclear industry.  24  The central elements of the Baruch Plan were that it abolished the veto power of the Security Council in relation to atomic matters. It emphasized \"immediate, swift, and sure punishment,\" including the possibility of atomic attack, on violators of the plan. The plan insisted on a survey of Soviet resources as a first step. This would have put the Soviets at a great disadvantage, as they would reveal secret information without the U.S. reciprocating at that point. This can be thought of as a \"hidden U.S. veto\" built into this international control process, because the U.S. would give up little in the initial stages and so could abort the process part way through with minimal downside. The plan de-emphasized the role of the Atomic Development Authority (ADA) and instead shifted responsibility for mining and refining of fissionable materials to private industry. In the Acheson-Lilienthal Plan, all mining and refining was to be carried out by the ADA. In Baruch's plan, the ADA would only own/manage \"activities potentially dangerous to world security\"; the rest would only be inspected or licensed by the ADA.  25  The  Gromyko Plan, June 1946  This was the official Soviet counterproposal to the U.S. Baruch Plan. It was announced at the UNAEC by Soviet delegate Andrei Gromyko on 19 June 1946 as an implicit rejection of the Baruch Plan. The plan 26 focused on disarmament rather than controlling the raw materials or the scientific R&D behind atomic weapons. It called for a complete ban on atomic weapons, which were to be destroyed within three months of the treaty coming into force, and contracting parties were to agree not to make or use atomic weapons. Violations would be regarded as a \"crime against humanity,\" and penalties would be determined by domestic legislation . The plan insisted that the Security Council veto apply to international control and all atomic matters (contra the Baruch Plan). It suggested the formation of two United Nations committees overseen by the Security Council: the first was to organize the exchange of atomic information and the second to ensure that the international agreement is followed. The requirement of early U.S. disarmament made this proposal completely unacceptable to the United States. However, it was the veto that was a focus of Baruch's opposition, even though it was probably strategically worthless since-in the event of breakdown-the only real sanction would be a threat of war. Historians believe that Soviet Union itself did not expect this proposal to be accepted and 27 only put it forward for propaganda reasons, and perhaps also to learn more about the U.S. atomic weapons program.  \n Lessons This section outlines some of the lessons we believe can be gleaned from the history of early attempts at international control of nuclear weapons. These lessons stress the complexity and messiness which has long been recognized as an inherent part of policymaking; nevertheless, we feel that pinpointing these lessons specifically in relation to powerful technologies makes them more salient for those thinking about such technologies today. The lessons begin with some aspects of the proposals before moving to consider some of the constituencies and processes that generated them. We then consider their likelihood of success and end with some consideration of cooperation and unilateral action. These lessons do not constitute a comprehensive or holistic overview of these proposals: for that we would point readers to the various historical studies on this topic (see References for a list and Introduction for a guide to the literature). \n Serious Radical Proposals Lessons Radical proposals which would normally appear naive or extreme may, in the right circumstances, be seriously proposed, discussed, and even adopted as official policy. Two conditions are conducive to this. First, if the emergent technology is spectacularly disruptive, it can expand the realm of politically feasible policies. Second, a sense of rupture or crisis in international political affairs can make otherwise unrealistic proposals more acceptable and possible. \n Historical Case Proposals for international control of atomic energy were radical for their time. They proposed that states be bound by powerful and wide-ranging multilateral agreements and that powerful international organizations be created with the power to police such agreements. In both these senses, these proposals were much more radical than other serious (that is, taken up at the diplomatic level) discussions on international governance at the time. The United Nations charter, for example, did not create obligations on states as binding, or intrusive on national soil, as some envisaged in the Acheson-Lilienthal Report. Similarly, no other armaments were subject 29 to such proposals in the 1940s.  30  The Acheson-Lilienthal Plan, for example, proposed that a powerful new U.N. Atomic Development Authority (ADA) would set up large R&D centers and conduct research on peaceful and warlike uses of atomic energy. It would own and operate all mining, refining, and production of fissionable raw materials-including having its own operational reactors-and dispense fissionable raw materials to nations for their nuclear power plants. It would also license and inspect operating civilian nuclear facilities in nation states. 31  29  On international governance in the 30s and 40s (including the U.N. charter), see Mark Mazower, Governing the World: The History of an Idea (Penguin: New York, 2012), chapters 5 and 7.  30  On the uniqueness of international control, see Patrick M. These proposals were taken seriously by many of their proponents and much of the public. There is every indication, for example, that the framers of the Acheson-Lilienthal Plan genuinely believed that their plan could work and that it was the most effective and realistic way of dealing with the problems of atomic energy.  32  There are several conditions that allowed the Acheson-Lilienthal Report to gain traction. (1) The report was accepted because of a growing public and elite perception of the threat that atomic weapons posed. It was widely believed in late 1945 and 1946 that atomic weapons would be used in any future major war, wiping out cities and killing millions. T he Bulletin of the Atomic Scientists , for example, was set up to warn of the destructive effects of atomic weapons. At the same time, many believed that atomic energy offered cheap electricity and new types of medicines and agricultural products. Publics and elites were amazed that an atomic 33 bomb had been developed so fast and often assumed that this rapid pace would continue into the near future. Novels and futuristic magazine and newspaper articles contributed to these beliefs. (2) The recent experience 34 with an awful war increased public and elite receptiveness to radical political proposals. There was a sense that regular politics had failed, and more radical measures were required. (3) The need for postwar reconstruction and the growth of U.S. influence globally made possible new initiatives in international relations which were not possible before. The formation of the United Nations and other international organizations (e.g., the Bretton Woods system to manage the global economy) were widely welcomed. (  4 ) Scientists who worked on 35 atomic matters generally threw their weight behind the Acheson-Lilienthal Report and formed powerful organizations that advocated for international control. Their status and newfound prominence gave their message traction in the media and in government. 36 32 Bernstein, \"The Quest for Security\"; Herken, pp. 155-58.  33  On early fears and hopes about atomic energy in the U.S., see Paul Boyer, By the Bomb's Early Light: American Thought and Culture at the Dawn of the Atomic Age  (New York: Pantheon Books, 1985) , Parts 4 and 5; Paul Boyer, \"A Historical View of Scare Tactics\", Bulletin of the Atomic Scientists 42,1 (January 1986), pp. 17-19.  34     (New York: Atheneum, 1971 ).  36  On their activism, see Alice Kimball-Smith, A Peril and a Hope: The Scientists' Movement in  America: 1945 -47 (Chicago: The University of Chicago Press, 1965 , chapters 5 to 12. See also later lessons in this report. \n Differences and Changes in Views Lessons Groups or coalitions supporting (or opposing) international control will contain individuals who have different reasons and rationale for their support (or opposition). Their motivations will be on a spectrum from the cynical to the idealistic. Their views and strength of opposition (or support) may vary significantly over time. Individuals may take positions for short-term gain. Those thinking about governance of powerful technologies need to be aware that support and opposition may shift and be prepared for it. They need to build coalitions that encompass a broad range of agendas and approaches, and would benefit from a \"political entrepreneur\" to hold divergent views together, build consensus, and forge a way forward.  37   \n Historical Case Individuals Had Different Understandings about the Impact of the Atomic Bomb On the one hand, some thought that the atomic bomb was a useful weapon of war. The U.S. army thought that nuclear weapons could play a significant, but not transformative, tactical role in slowing Soviet armies. The navy thought that the atomic bomb would be central to stopping a Soviet attack, and the air force eventually came to see it as a strategic weapon that could land a knockout blow (the \"air-atomic strategy\"). Ex-Prime 38 Minister Winston Churchill thought that the atomic bomb could be used to keep the Soviet Union in check and assure U.S. and British domination of world affairs into the near future. Yet others thought that it had 39 transformed international relations and warfare in a negative way. These people believed it made wars more destructive and suicidal, and increased the likelihood that smaller aggressive nations (now armed with atomic weapons) would launch wars. 40 \n Views \n Changed over Time Key policymakers' opinions on international control shifted over time. In some ways, these shifts reflected broader shifts in opinion about the Soviet Union or the destructiveness of atomic bombs. Opinion was particularly fluid during the war. Secretary of War Henry L. September, he had given up on a quid pro quo, recognizing how unlikely the Soviet Union was to make those concessions and having greater concern about the dangers of the atomic bomb. He reemphasized the imperative of international control and a proposed U.S.-U.K.-Soviet \"covenant\": that the Soviets refrain from atomic development, and in return, the West would share the peaceful applications of atomic energy and agree not to employ the atomic bomb. One of the reasons for Stimson's conversion to international control may have been his impending retirement; he presented his influential memorandum calling for international control at a Cabinet meeting in September 1945 just prior to his retirement. Stimson may have seen it as an ideal policy to pursue, potentially creating a legacy but with little personal risk to himself. Other policymakers may have 41 pushed for international control for similar reasons. Secretary of State Byrnes was demoted in December 1945, and that may have been a factor in him setting up a State Department Committee to look into international control as one of his final acts. This suggests a potential lesson: look for officials near retirement or leaving 42 office as candidates for enacting more idealistic policies, with path-dependent impacts. Perceptions of Soviet threat continued to have a large impact on individuals' opinions into 1946. As the Soviet Union appeared to grow economically in the first half of 1946 and become more assertive in international affairs (especially in relation to Turkey), U.S. policymakers (including Byrnes and Truman) came to increasingly see cooperation as impossible. The famous \"Long Telegram,\" sent to the State Department in February 1946 by 43 George F. Kennan, the chargé d'affaires at the United States Embassy in Moscow, typified this repositioning.  44  This shift reduced willingness to negotiate on international control and led senior policymakers to increasingly see international control negotiations as only useful for propaganda purposes. Policymakers also responded to growing public concerns over Soviet spying. Truman, for example, refused to support civilian control of atomic energy following a spy scandal in February/March 1946. He instead bent to public opinion, which increasingly preferred a strong military role in controlling atomic energy domestically. By the middle of 1947, even 45 previously strong supporters of international control gave up on it, having decided that the Soviet Union could not be trusted or negotiated with. Oppenheimer, for example, met Baruch's replacement as the U.S. delegate on the UNAEC, Frederick Osborn, specifically to request that the U.S. withdraw from atomic control negotiations. 46 Building a Coalition of Support from Differing Constituencies was Important  41 Herken, pp. 14, 16, 19, [29] [30] [31]    1946,\" in ibid., pp. 112-33. 44  John Lewis Gaddis, \n Strategies of Containment: A Critical Appraisal of American National Security Policy during the Cold War 2nd edition (New York: Oxford University Press, 2005), pp. 53-4. The \"Long Telegram\" was an influential note written by Kennan arguing that Soviet policy was inherently expansionist and based on a neurotic view of world affairs. He argued against cooperation with the Soviet Union.  45 Herken,  Joseph I. Lieberman, The Scorpion and the Tarantula  : The Struggle to Control Atomic Weapons 1945 -1949 (Boston: Houghton Mifflin Company, 1970  Individuals and initiatives that built support from a range of constituencies and viewpoints were more likely to succeed. The Acheson-Lilienthal Report gained traction not only because it was sponsored by the State Department, but because it had the power and prestige of its committee members behind it, who represented a range of different interests. These included the liberal New Dealer David Lilienthal (former head of the TVATennessee Valley Authority), Under Secretary of State Dean Acheson (who had the support of much of the State Department), and J. Robert Oppenheimer, who carried with him the support of the Atomic Scientists' Movement. Early attempts at international control by Secretary of State  Byrnes (in late 1945) , on the other hand, had failed because he had excluded certain powerful, but skeptical, congressmen. They felt slighted that he had not shared policymaking with them, and so undermined his policy initiatives. 47  47   \n Lieberman, The Scorpion and the Tarantula , pp. 270-72; Kimball-Smith, A Peril and a Hope , pp. 461-62; Herken, pp.74-87. \n Cautious or Cynical Cooperators Lessons Schemes for international governance can garner support from \"realists,\" understood here as policymakers who believe in the primacy and inevitability of power politics and who focus first and foremost on national interest.  48  Their support, though often cautious or cynical, can be crucial, but it can also be fickle. \n Historical Case Although realists were disturbed by the implications of atomic weapons, they did not see them as fundamentally changing the world or international relations. Realists tended to believe that cooperating on the control of nuclear weapons was futile and, in any case, best done from a position of strength. They tended to not perceive a nuclear arms race as particularly dangerous, especially compared to the dangers from making oneself vulnerable through steps towards international control. Such steps could be dangerous for the U.S. because they could lead to the diffusion of capabilities and could undermine the country's resolve to resist Communism. They thus concluded that the U.S. should continue with atomic weapons development. In a (realist) analysis 49 of atomic weapons presented to Congress in January 1946, the former head of the Manhattan Project Leslie Groves presented only two alternatives for the U.S.: \"Either we must have a hard-boiled, realistic enforceable world agreement ensuring the outlawing of atomic weapons or we and our dependable allies must have an exclusive supremacy in the field.\" It was clear to him that if \"there are to be atomic weapons in the world, we must have the best, the biggest and the most.\"  50  The realist commitment to the U.S. atomic arsenal was bolstered by their belief that the Soviet Union would not be able to build an atomic bomb for many years. Groves, for example, gave various estimates but, from November 1945 onwards, usually said twenty years. The high financial costs of the Manhattan Project and the 51 scientific, technological, industrial, and organizational hurdles the United States had to overcome in order to build the bomb led many to believe that, even in the best of circumstances, it would take the Soviet Union much longer to achieve this feat. Even if the Soviet Union managed to develop atomic bombs, many believed that the U.S. could remain ahead in atomic weapons R&D, production, deployment, and delivery over the coming decades. Moreover, the Soviet economy appeared to be in no shape to take the burden of an expensive atomic program. It was suffering from wartime devastation, overburdened with the costs of the occupation of Eastern Europe, and had an economy still tuned to the production of (conventional) military forces. In 53 addition, Groves believed that the U.S. and its allies could monopolize nuclear fuel, thus delaying or stopping the Soviet atomic program altogether, and he worked hard to achieve this. This belief appeared to be important for his confident assertions that the Soviets would not soon get atomic weapons, but he did not share it because he regarded the U.S.'s monopolization efforts as a state secret. 54 48 For definitions of realism and realist dispositions, see Donnelly, Realism and International Relations , pp. 6-13.  49  Zachary S. Davis, \"The Realist Nuclear Regime\", Security Studies 2,3-4 (1993), pp. 79-99.  50  Herken, p. 112.  51  Gordin, Red Cloud at Dawn , p. 70.  52  Herken, p. 231; Gordin, Red Cloud at Dawn , pp. 70-1.  53  Herken, p.138.  54  In a September 1944 report, for example, Groves predicted that the U.S., through the Combined Development Trust, could control 90% of the world's high-grade uranium ore by the war's end. Herken, pp. 101-02. In December 1945, Groves claimed that the Trust controlled 97% of the world's uranium output and 65% of the world's supply of thorium, see Holloway, Stalin and the Bomb , p. 174. Also: Charles A. Ziegler, \"Intelligence Assessments of Soviet Atomic Capability, Lastly, the Soviet Union economy appeared to U.S. policymakers to be in no shape to pose a challenge to U.S. interests. The country, they reasoned, was suffering from wartime devastation, overburdened with the costs of the occupation of Eastern Europe, focused on postwar reconstruction, and still tuned to the production of (conventional) military forces.  55  While realists represented only a minority of the elites in favor of international control, their support was disproportionately important, as they were often powerful individuals, embedded high in the state, with strong influence on policy. However, while they supported international control, they were not strongly committed to it. They had, at best, a weak preference for it and quickly abandoned it as circumstances changed. Examples include Bernard Baruch and perhaps even President Truman himself. Baruch was picked by Truman to shape the U.S. proposal on international control and present it at the United Nations Atomic Energy Commission (UNAEC). Baruch quickly abandoned hope for international control once negotiations stalled. Historians now believe that he was only weakly committed to it, if at all. The extent of Truman's commitment to international 56 control in early 1946 is unclear, but he was certainly attracted to the idea and interested enough to allow for a U.S. proposal to be formulated and placed before the UNAEC. What is clear, however, is that he quickly lost interest by mid 1946 once he became convinced that aggressive Soviet foreign policy could not be met with concessions.  57  There were a number of reasons for why realists thought that international control would support their objective to maximize U.S. power in international affairs: (1) By supporting a policy that was popular amongst the public, they hoped to boost their own popularity. They wanted to be seen addressing public concerns about atomic weapons. Carrying out negotiations, they calculated, would be enough to meet this concern. If negotiations failed, they hoped to place the blame on the Soviet Union, thus highlighting the Soviet Union as, at best, an unreliable partner in international affairs, and at worst, a threat to U.S. security. 58 (2) Some supported international control for administrative political reasons. January 1946 was that he wanted to retain atomic foreign policymaking and expertise within the State Department, rather than to lose it to Congress or some other state entity. 59 (3) Some may have believed that the official U.S. proposal, the Baruch Plan, was designed to safeguard U.S. national security interests in that it included the maximum possible concessions that the U.S. could make without jeopardizing its security. They consequently supported it because it was not only the best possible international control plan but also because it mitigated enough risk for the U.S. to be acceptable. Bernard Baruch and his negotiators, historians largely conclude, did not try harder to reach an agreement with the Soviets for this very reason. Baruch and his associates believed that the plan was the best that the U.S. could offer, and preferred a no-deal scenario that would still leave the U.S. in a dominant position in atomic weapons for decades to come. 60 \n Secrecy and Security Lessons Secrecy and security will play a central role in any discussion on the governance of powerful technologies. They can have a significant effect on the possibility of international cooperation as well as on intrastate power struggles. They can give tremendous power to individuals and state institutions (such as the military) which control the flow of information and, in particular, can be used to undermine opponents. Secrecy can be terrible for epistemics, undermining competent organizational deliberation. Secrecy is often antithetical to cooperation and trust, in part because the public and actors who are outside secret access are often more in favor of cooperation and trust. Secrecy can be used to empower narratives of fear and belligerence. 61 Policymakers need to carefully weigh decisions to expand secret domains. They need to make sure such decisions are counterbalanced within institutions and that a wide range of perspectives are used to inform policymaking. This will reduce the risks of corruption and abuse as well as decision-making from an overly narrow perspective. Policymakers also need to be wary of arguments that warn of imminent security threats. Such narratives can easily lead to increased secrecy. Policymakers should obtain a wide range of views so as to get better quality information and make informed decisions. They should also ensure that technical experts are involved in key strategic decisions. Historical Case Appeals to the security of the United States and warnings of imminent security threats were an intrinsic part of debates over international control of atomic energy. These were not constant but waxed and waned in intensity and influence. One major drop in public and elite support for international control (and civilian control of domestic atomic energy) occurred due to the sensational revelation of Soviet \"atomic spies\" in February 1946. The Soviet Union increasingly did not appear to be reliable allies and, instead, were seen as having malign intent and a growing atomic bomb program.  62  More specifically, secrecy gave tremendous power to those who controlled the flow of information, notably Leslie R. Groves (the head of the Manhattan Project). Groves accrued power and responsibility for himself by 63 taking advantage of his privileged access to information and his ability to demarcate the boundaries and hierarchies of secrecy. He began with responsibility for building the plants and factories to make atomic fuel. This evolved into responsibility for scientific research, weapon design, and atomic security, intelligence, and counterintelligence. He was also eventually involved in high-level policymaking on both domestic and 61 Note that, in some cases, secrecy can permit cooperative gestures that a hawkish/nationalistic public would punish (especially if unsuccessful) and can allow for collusion to prevent escalation. Counterexamples include outreach in the style of Nixon to China, the Oslo Accords, and preventing escalation from Soviet-US conflict in Korea. A. and stockpiles was such a closely guarded secret within the military that parts of the military concerned with atomic weapon use were denied crucial information. War plans were created in  1945, 1946, and 1947 , for example, with little understanding of the size of the U.S. atomic arsenal and its deliverability, or based on inaccurate estimates of its size. Staff officers drawing up plans did not even know for certain how they were allowed to use atomic weapons. One January 1946 plan from the Joint War Plans Committee designated 17 Soviet cities as targets for which it assumed the Air Corps would require 98 atomic bombs plus 98 in reserve, giving a total of requirement of 196. At the end of 1946, the actual number of bombs in U.S. possession was around 7! Moreover, none of the bases from which the atomic strikes were to be launched were equipped with atomic weapons loading pits or atomic storage facilities. The Joint Chiefs of Staff even officially complained that they were denied access to intelligence collected and held by Groves. This led to significant disagreement and inconsistency between various military plans on the use of atomic weapons.  73  Secrecy led to a poor public understanding of damage to civilian populations from atomic warfare. In one fall 1945 public hearing, Groves' casual comment that an atomic war would \"only\" lead to 40 million U.S. casualties shocked the audience. Some policymakers saw benefit in hiding the destructiveness of the atomic bomb from 74 the public. One later Joint Chiefs report noted that \"A situation dangerous to our security could result from impressing on our own democratic peoples the horrors of future wars of mass destruction while the populations of the 'police' states remain unaware of the terrible implications.\"  75  Concerns over the possible loss of the \"atomic secret,\" driven by reports of Soviet spies in February 1946, empowered narratives of fear and hawkishness. Without fuller information, it was easier for the public to panic about security risks. In such an atmosphere, those possessing top secret clearance (especially Groves) were empowered. In an atmosphere of increased security consciousness, it was easier for the military and its supporters to gain backing for their views on atomic technology. It also reduced support for international 76 cooperation on the atomic bomb amongst the public, and even scientists reduced their support for international control as rumors circulated that the House Un-American Activities Committee was to investigate Oak Ridge scientists for security threats. Secrecy may have contributed to the lack of trust between the U.S. and the Soviet 77 Union, and possibly also between the U.S. and Britain. It also led to tensions between Groves and the civilians and scientists working on atomic policy, who resented his control over atomic information and his use of secrecy to win arguments.  78  4.5 Public Sphere \n Lessons The public sphere will have a powerful impact on debates on international control. Political and policymaking elites will be sensitive to large shifts in public opinion, which will (for example) influence their political prospects. They will seek to mobilize public opinion in support of their preferred policies. Public opinion on this issue is malleable and can be shaped in support of or against particular policies, people, institutions, or countries. Although elite opinion remains paramount, participants in debates on international control benefit from harnessing the power of the public sphere and shaping public opinion. They should run publicity campaigns, make arguments that would appeal to the public, and garner support from individuals and groups with high public profiles. \n Historical Case In relation to the international control of atomic energy, although elite opinion was paramount, policymakers nevertheless were influenced by and reacted to public opinion. Domestic pressures, for example, forced 79 Truman to commit the U.S. to the principle of international control before Secretary of State Byrnes had even attempted to extract a quid pro quo from Moscow in late 1945. Public concerns over the safety of the \"atomic 80 secret\" and Soviet spies led Truman to maintain higher secrecy around atomic operations than might otherwise have been ideal. Alternatives to international control, such as the concept of an \"atomic league,\" were not fully 81 explored as they emphasized preventive war, which was out of favor with the public. Baruch became obsessed 82 with even the smallest shift in public support for his proposals during UNAEC negotiations in September 1946. Policies were sometimes vaguely stated in order to satisfy the public: NSC-30, the \"Policy on Atomic 83 Warfare\" released at the end of 1948, for example, avoided the issue of first use of atomic weapons in order to not upset public opinion. Baruch and his advisory team, when they heard of Soviet delegate Andrei 84 Gromyko's rejection of the Baruch Plan in June 1946, did not want to openly reject the Soviet counterproposal (the so-called \"Gromyko Plan\") so early in the negotiations but nevertheless wanted to signal their rejection. They thus leaked a series of stories to the press from \"anonymous but reliable sources\" that the U.S. delegation could not accept the Gromyko Plan.  85  Elites shaped public opinion in their favor, using press releases, briefings, and publications. Groves, for example, may have been the confidential source cited by news reports that broke sensationalist news of Soviet spies  79  Another example of the influence of public opinion on nuclear arms control policymaking is James Cameron, The Double Game The Demise of America's First Missile Defense System and the Rise of Strategic Arms Limitation (New York: Oxford University Press, 2017).  80  Herken, p. 52.  81  Herken, p. 136; Hogan, A Cross of Iron , p. 238.  82  Herken, p. 265.  83  Herken, p. 179-80.  84    \n Technical Experts Lessons Technical experts and specialists (scientists, engineers, technicians, academics) have significant power to shape proposals and policy to be more effective and more cooperative, though their influence depends on their political sophistication. Technical experts should be given a central role in drafting proposals, policymaking, and public engagement. Experts should invest time in understanding the political landscape and identifying political allies; their most important contribution to the problems of powerful technologies may be from their shaping of political discussions, rather than marginally more scientific or academic work. \n Historical Case From 1944 to 1946, atomic scientists played a central role in proposals for international control. They were leading advocates, passionate and committed. They were some of the first to warn of the dangers of atomic energy, they formed advocacy groups and raised funds for their activism, and they carried out public engagement to explain atomic energy and warn of its dangers. They played an important role in providing information and advice to elites and to the public, with whom they garnered enormous respect, authority, and credibility. Foremost amongst them was J. Robert Oppenheimer (head of the Los Alamos laboratory during 91 the war), who was considered a national hero, the \"father of the atomic bomb,\" and made the cover of Time magazine in November 1948. Scientists were also central in shaping proposals. The Acheson-Lilienthal Report 92 was drafted largely by Oppenheimer, who brought into it many of the ideas of the Atomic Scientists' Movement.  93  Scientists also provided technical and strategic insights for elites. Leo Szilard, for example, lobbied hard to convince the U.S. government that they needed to start an atomic bomb program in 1939. James B. Conant, 94 Oppenheimer, and Vannevar Bush had significant input on atomic policymaking during the war. The wartime \"Scientific Advisory Panel\" (staffed by scientists Oppenheimer, Ernest Lawrence, Enrico Fermi, and Arthur H. Compton) was instrumental in determining how the atomic bomb would be used.  In 1944 and 1945, other 95  scientists informed policymakers of the strategic implications of atomic weapons in reports and memos. Many  91  On their activism, see Kimball-Smith, A Peril and a Hope , chapters 5 to 12. Historian Jessica Wang has summarized their impact thus: \"Despite their individual anonymity, the atomic scientists soon established a powerful presence in American political life. They appealed directly to the public through a long series of media interviews, articles, radio addresses, and public speaking engagements in which they discussed both the specific legislation at hand and the general political and social implications of atomic energy...Between October and December 1945, some thirty-odd scientists went to Washington, where, in a whirlwind of social and political activity, they built influence in excess of their numbers.\" Jessica Wang, American Science in an Age of Anxiety: Scientists, Anticommunism, and the Cold War (Chapel Hill, NC: University of North Carolina Press, 1999), p.16.  92   of these warned that there could be an arms race, that there was no such thing as an \"atomic secret\" that could be kept, and that the Soviets would acquire the atomic bomb within a few years. These warnings turned out to be more accurate than Groves' predictions on when the Soviet Union would get the atomic bomb. Scientists 96 also played a key role in the development of monitoring technologies and strategies in the late 40s.  97  The most detailed proposal for international control, the Acheson-Lilienthal Plan, could not have been drawn up without the participation of a scientist. Others working on the plan knew very little, if anything at all, about the workings of atomic energy or the atomic bomb. Because of his technical expertise, Oppenheimer took the lead in educating other committee members about atomic energy and then in drafting the plan. He was 98 uniquely able to suggest potential technical developments (such as separability and denaturing) to ease the problems inherent to international control.  99  On the other hand, scientists were sometimes perceived as naive, especially by politicians and diplomats. Scientists (especially those in the Atomic Scientists' Movement or supporters of world government) were often derided for their idealistic views and lack of understanding of the politically possible. Prominent State Department official George Kennan, for example, reported back to the State Department in August 1946, following meetings with members of the Atomic Scientists' Movement, that \"[p]olitically, these people are as innocent as six-year-old maidens. In trying to explain things to them I felt like one who shatters the pure ideals of tender youth.\" Einstein's calls for world government in 1947/48 (which included international control of 100 nuclear weapons) were similarly derided by the State Department as \"naive… The man who popularized the concept of the fourth dimension could think in only two of them in consideration of World Government.\" 101 After a meeting in October 1945, Oppenheimer was described by Truman as a \"cry-baby scientist\" who had come to his office and \"spent most of his time wringing his hands and telling me they had blood on them because of his discovery of atomic energy.\" 102 Scientists were also vulnerable to the criticism that they were a security threat. This was due largely to their openness, their international connections and communications (for example with Eastern bloc scientists), and their (generally) progressive politics. One of the earliest calls for international control, by Danish physicist Niels Bohr to Winston Churchill and Franklin D. Roosevelt in 1944, was met with the response from Churchill that  96  For example, through the June 1945 Franck Report and the November 1944 Jeffries Report: Kimball Smith, A Peril and a Hope, pp. 19-24, 41-8.  97  Gordin, Red Cloud at Dawn, pp. 189-213.  98  Bird and Sherwin, American Prometheus, p. 341.  99  Oppenheimer suggested that a new technique called isotopic denaturing could be used to prevent misuse of uranium reactor fuel for nuclear weapons. Isotopic denaturing meant the addition of a different isotope of nuclear fuel which would render the fuel useless as an explosive. The fuel, however, could still be used for power reactors. Such an isotope, it was thought, could not be chemically separated. This was the foundation for his idea that military uses of atomic energy could be practically separated from civilian uses. The military uses could then be tightly controlled by an international authority. Civilian uses could be safely left to sovereign states. See:  Barnard et    \n Muddled Policymaking Lessons Policymaking involves significant muddling through rather than grand strategy. It is also deeply affected by 107 domestic politics and public opinion, and often developed on the basis of short-term objectives and poor quality information. Proposals are sometimes slow to be developed, poorly thought-out or expressed, gambled on technical solutions, and lacking crucial details. There can be a lack of clarity on responsibility for policymaking, which is often dependent on personalities. \n Historical Case In relation to the international control of atomic energy, debates were often unclear about what was being discussed and proposed and what was at stake. Debates were hugely shaped by domestic politics, lack of information, vague understandings, retirements, private initiatives, committee room dealings, egos, and organizational interests. Debates about international control were also sometimes intertwined with debates over domestic policy. \n Secrecy and Lack of Information As noted in    -1960 \", International Security 7,4 (Spring 1983 Herken, p. 197. 109 Herken, pp. 219-20, 229; Rosenberg, \"The Origins of Overkill\"; Ross, American War Plans 1945 Ross, American War Plans 1945 Intelligence Assessments of Soviet Atomic Capability\". The January 1946 plan is in Ross, American War Plans, pp. 16-17. On the number of atomic bombs see John M. Curatola, Bigger Bombs for a Brighter Tomorrow: Strategic Air Command and American War Plans at the Dawn of the Atomic  Age, 1945 (Jefferson, NC: MacFarland, 2016 , pp. 53.  112 Herken, pp. 113; Gordin, Red Cloud at Dawn, . On the CIA's mistaken assessments see Donald P. Steury, \"How the CIA Missed Stalin's Bomb: Dissecting Soviet Analysis, 1946-50\". Available at: https://www.cia.gov/library/center-for-the-study-of-intelligence/csi-publications/csi-studies/studies/vol49no1/html_files /stalins_bomb_3.html , accessed 1 June 2019. atomic development was kept secret from most policymakers, including Truman, who later ignored it when he found out. 113 \n Proposals were Slow to be Developed, Poorly Thought-out or Expressed, Gambled on Technical Solutions, and Lacked Crucial Details Serious thinking about proposals for international control were not formulated until relatively late. This is due in large part to the existence of nuclear weapons being a closely held secret until the detonation of the first atomic bomb at Hiroshima in August 1945. Efforts to make sense of nuclear weapons only dramatically accelerated after the bomb's existence was made public.  114  Secretary of War Henry L. Stimson crisply stated the high-level problem of postwar governance of nuclear weapons in April 1945, but little serious thought about international control took place after then. This 115 neglect can be in part understood, because policymakers were swamped with massive geopolitical problems: ending the war with Japan, negotiating a postwar order with Russia, setting up the United Nations, and rebuilding Europe. Secretary of War Stimson's ideas in late 1944 and early 1945 on international control and possible quid pro quos were cursory. Most proposals for international control lacked detail and failed to think through the Soviet response. This is particularly true of the earliest thinking in 1944 and 1945. Stimson thought that international control could be achieved through \"freedom both of science and access\" to atomic information. By this he meant some form of sharing of atomic scientific information, but he did not go into further detail. Stimson also thought that the U.S. could demand liberalization of internal rule in the Soviet Union in return for information on the atomic bomb, which reflected a complete misunderstanding of Soviet preferences. Similarly his later idea of a simple \"covenant\" between the great powers to not use atomic energy for military purposes, without any thought of safeguards or punishments, was ill thought-out. Stimson's memorandum on international control, 116 presented at a crucial September 1945 meeting on atomic energy, generated significant confusion over what he was proposing. Some participants thought he wanted to give the atomic bomb to the Soviets. This was due to an unfamiliarity on their part with the development of the bomb and its international politics, as well as a lack of clarity on Stimson's part. 117 113 Barton J.  Bernstein, \"Roosevelt, Truman, and the Atomic Bomb, <PHONE>    May 1945 (31 May 1945 . Available at: http://www.nuclearfiles.org/menu/key-issues/nuclear-weapons/history/pre-cold-war/interim-committee/interim-commi ttee-informal-notes_1945-05-31.htm , accessed 15 October 2019. h/t Luke Muehlhauser, who reflects on this: https://twitter.com/lukeprog/status/1181774870096400384  116 Herken, pp. 14, 25; Malloy, Atomic Tragedy ,   The Scorpion and the Tarantula , pp. 145-6. The international control of technology is made easier when the dangerous uses and productive uses can each be cleanly, verifiably, and robustly separated from each other. In this case, there was one technical possibility that the most thought-through proposal, the Acheson-Lilienthal Plan, invested a lot of hope in, but which in retrospect was misguided. (This was plausibly foreseeable at the time). The planners allowed themselves to depend on a speculative technical fix, called isotopic denaturing. It was hoped this could be used to prevent (the easy) misuse of uranium or plutonium reactor fuel for nuclear weapons. Effective denaturing was, however, a highly speculative concept, and history has not borne it out. The report was criticized by other scientists (and 118 by Leslie Groves) for relying on such an impractical idea soon after the plan was made public.  119  The plan also did not explain what was to be done with the existing atomic bombs once international control was instituted: were they to be destroyed, or passed onto the U.N.? In this case, the vagueness was probably deliberate and reflected the political ramifications of choosing between these options: there were many calling for the complete destruction of atomic bombs and yet others for them to be kept and used by the U.N. for collective security or policing purposes. 120 \n Lack of Clarity on Responsibility for Policymaking Responsibility for policymaking on atomic matters was not clearly demarcated, especially in late 1945 (it would improve later on with the Acheson-Lilienthal committee, though the committee itself was a gambit for State Department influence). Congressional committees, congressmen, State Department officials, military men, scientific advisors, external consultants, and other advisors all had influence at one point or another. Scientific administrator Vannevar Bush would write to Stimson in November 1945 expressing frustration at the lack of  118  Below is a selection from the Acheson-Lilienthal Report, emphasis ours. Note that the actual text admits that this technique is not foolproof. \"U 235 and plutonium can be denatured; such denatured materials do not readily lend themselves to the making of atomic explosives, but they can still be used with no essential loss of effectiveness for the peaceful applications of atomic energy. … It is important to understand the sense in which denaturing renders material safer. In the first place, it will make the material unuseable by any methods we now know for effective atomic explosives unless steps are taken to remove the denaturants. In the second place, the development of more ingenious methods in the field of atomic explosives which might make this material effectively useable is not only dubious, but is certainly not possible without a very major scientific and technical effort. It is possible, both for U 235 and for plutonium, to remove the denaturant, but doing so calls for rather complex installations which, though not of the scale of those at Oak Ridge or Hanford, nevertheless will require a large effort and, above all, scientific and engineering skill of an appreciable order for their development. It is not without importance to bear in mind that, although as the art now stands denatured materials are unsuitable for bomb manufacture, developments which do not appear to be in principle impossible might alter the situation. This is a good example of the need for constant reconsideration of the dividing line between what is safe and what is dangerous.\" Barnard et al, A Report on the International Control of Atomic Energy . 119 Isotopic denaturing meant the addition of a different isotope of nuclear fuel which would render the fuel useless as an explosive. The fuel however could still be used for power reactors. Such an isotope, it was thought, could not be chemically separated.  Barnard   Interservice rivalry had a large impact on debates and policies. Each military service attempted to increase its influence on atomic weapons and atomic policy, and decrease that of the other services. Perhaps the most 127 notorious episode in this regard was the infamous \"Admirals' Revolt\" of 1949, in which the navy attempted to divert funding away from the air force and towards the navy. In a series of hearings before the House Armed Services Committee, senior naval officers argued that the air force's strategy of \"atomic blitz\" could not win a war against the Soviet Union and was anyhow immoral. The navy instead suggested (ultimately unsuccessfully) that aircraft carrier-based bombers launch tactical atomic strikes against the invading Soviet armies. 128 \n Public Opinion and the Atomic Secret Public opinion on atomic matters and international control (which had significant influence on policymaking, see earlier section ) was based on one crucial misconception: that there existed an \"atomic secret\" which was the key to the production of the bomb. It was widely believed that if the Soviets discovered this secret, they could build the atomic bomb. There was significant debate and confusion in public discourse over what this secret was and how it could be protected or shared. Scientists spent a significant amount of effort debunking the domain, and stressed the importance of scientific and industrial expertise and resources, as well as financial resources. Their statements had limited impact: the idea of an atomic secret was deeply embedded and was moreover used and repeated in public rhetoric by opponents of international control such as Groves. 130 \n Personality High-level policymaking was highly dependent on personality. The wartime president, Franklin D. Roosevelt, had a reputation for not disagreeing with individuals in face-to-face meetings. He preferred to raise objections later with confidants. His successor Truman only digested short briefings. Truman had opportunities to 131 132 make a greater impact on atomic policy and diplomacy but did not, for no clear reason. At Potsdam, for example, he dodged the issue of informing the Soviet Union clearly about the U.S. atomic program (the next opportunity would arise after the war). Roosevelt, if he had lived, would probably have conducted policy 133 differently than Truman and reacted differently to political developments.  134  Groves, Baruch, and Secretary of State Byrnes (amongst others) also put their personal stamp on diplomacy and policymaking. Ego and personal career motivations shaped their approaches. Crucially, one reason for Baruch formulating his own distinct proposal for international control-rather than following the earlier Acheson-Lilienthal Plan-was ego. He refused, he said, to be a mere \"messenger boy\" for the Acheson-Lilienthal Plan. Truman himself would later recall that Baruch was driven by the need for \"public  \n Mixed Signals 130 Herken, \"'A Most Deadly Illusion'\".  131  Aaserud, \"The Scientist and the Statesmen\".  132  Herken, p. 15.  133  Wilson D. Miscamble, The Most Controversial Decision: Truman, the Atomic bombs, and the Defeat of Japan (Cambridge: Cambridge University Press, 2011), p. 70. 134 Campbell Craig and Sergey Radchenko, for example, believe that \"Truman at the outset of his presidency was more disposed toward cooperation with the Soviet Union than Roosevelt had been at the end of his.\" Roosevelt was also, they suggest, strongly influenced by Churchill in his atomic diplomacy.  Craig and Radchenko, pp. 29, 65.  There is also some speculation of this in Bernstein, \"Roosevelt, Truman, and the Atomic Bomb\".  135  Herken, p. 160. For a list of the key differences between the Baruch Plan and the earlier Acheson-Lilienthal Report, see the summary of the Baruch Plan in the earlier section which reviews proposals for international control. 136 Harry S.  Truman, Memoirs  Volume II Years of Trial and Hope (New York: Doubleday and Company, 1956), p.10; James Grant, Bernard M. Baruch: The Adventures of a Wall Street Legend (New York: John Wiley & Sons, 1997), p. 292.  137  Truman would later call Oppenheimer a \"cry-baby scientist\" who had come to his office and \"spent most of his time wringing his hands and telling me they had blood on them because of his discovery of atomic energy.\" He kept Oppenheimer away from policymaking from that point on. Ray Monk, Robert Oppenheimer: A Life Inside the Center (New York: Doubleday, 2012), p. 494;  Bird and Sherwin, American Prometheus , 350. 138  For example  : Herken, p. 20  1946, his early conciliatory statements led U.S. policymakers to believe that he would take, at most, a soft stance against the Baruch Plan during his U.N. General Assembly speech. Instead he stunned both U.S. policymakers and the U.N. by aggressively attacking the plan and even Baruch himself (\"conceited\" and \"short-witted\"). Policymakers scrambled to figure out a response. Mixed signals were also present at domestic meetings. Bohr, 140 for example, met Roosevelt in 1944 to call for international control and left the meeting thinking that he had made an impact. In fact, Bohr had little impact on Roosevelt other than to suggest to him that the Danish physicist may be a security risk. 141 \n Diplomatic Missteps Actors also made missteps when carrying out atomic diplomacy.  \n Misunderstandings of the Significance of Issues Negotiators and policymakers misunderstood the significance of various aspects of international control. In early U.S.-Soviet discussions at the UNAEC during June to September 1945, Baruch and his delegation focused on the veto. This fixation missed the fact that the veto was strategically irrelevant. As critics of Baruch's insistence on removing the veto (such as Under Secretary of State Dean Acheson and political commentator Walter Lippmann) pointed out: if the international control regime broke down, it would lead to a breakdown in cooperation in the United Nations and a veto would have little impact. In a widely publicized September Baruch may also have become trapped by his own rhetoric: by overselling the importance of the veto to the 147 U.S. public and political elites, he left himself little room to compromise on it. 148 \n Incongruous Stances Even on issues such as secrecy (where one might imagine clear battle lines between advocates for and opponents of increased secrecy), there were seemingly incongruous stances. The Smyth Report, released on the 12th of August 1945, soon after the atomic bombings, is a case in point. This official published primer on the atomic bomb and the Manhattan Project, written by physicist Henry DeWolf Smyth, provided a \"semi-technical\" (Smyth's words) explanation of how the atomic bomb worked and how it was developed. It proved to be especially helpful for the Soviet project: the NKVD (the Soviet secret police) had it translated and published with a print run of 30,000 copies. Given the secrecy around the project, how did it come to be published? As it happens, Groves supported its publication-he wanted to justify the large expenditure on the Manhattan Project and create a \"security fence\" highlighting what could be released to the public on atomic energy. Groves also mistakenly believed that it would be of little use to the Soviets-an astonishing oversight for an individual so focused on secrecy and the Soviet threat. On the other hand, many opponents of secrecy, such as Leo Szilard, David Lilienthal, and Secretary of Commerce Henry Wallace, argued that the report gave away important information to the Soviets. They did this in order to discredit the military's management of atomic energy and to push for civilian control. Moreover, there were mishaps too with the management of information revealed 149 in the Smyth Report. The first edition included a reference to the unforeseen poisoning of the Hanford reactors, which Groves had excised from the second (and more widely circulated) edition when he became aware of it. The revelation was picked up by the Soviet government when they compared the first and second editions and may have been useful for the Soviet program. 150 \n Domestic Politics Shaped Debates on International Control Domestic partisan politics shaped support for international control and related policy. There are two key examples of this. First, debates over international control became intertwined with debates over legislation on domestic governance of national atomic facilities and policies. The political battle to set up a domestic organizational and legislative framework to govern atomic energy began at the end of the war, and ended with the signing of the Atomic Energy Act and the creation of the Atomic Energy Commission in August 1946. The political fights over this legislation polarized into those wanting more military oversight and those arguing for less. Support for both domestic and international control was divided along political lines: on domestic issues, liberal and progressive elites generally supported civilian oversight on domestic atomic policy, whereas Republicans and conservatives preferred military oversight. Similarly, Democrats were more supportive of  147  Lieberman, The Scorpion and the Tarantula, p. 348.  148  Herken, pp. 174-77.  149  Gordin, Red Cloud at Dawn , pp. 91-104. The Smyth Report has been republished as: Henry DeWolf Smyth, Atomic Energy for Military Purposes: The Official Report on the Development of the Atomic Bomb Under the Auspices of the United States  Government, 1940 (Stanford: Stanford University Press, 1989 . According to Gordin, the Smyth Report \"revealed the scale of the effort and the sheer quantity of resources, and also hinted at some of the paths that might work and, by omission, some that probably would not.\" From Gordin, Red Cloud at Dawn , p. 103. 150 Richard Rhodes, Dark Sun: The Making of the Hydrogen Bomb (New York: Simon and Schuster, 1995), pp.  215-7.  international control than Republicans. The military opposed international control and opposed civilian control of domestic atomic energy.  151  A second example is Truman's appointment of Bernard Baruch as U.S. representative on the UNAEC in March 1946. Baruch was appointed with the power to make policy. With this powerful appointment, Truman undercut both his existing atomic experts and the State Department. He did this largely because of Baruch's acceptability to Congress and broader conservative opinion. Baruch, in turn, appointed friends and associates with little knowledge of atomic matters as his advisors. Truman soon regretted his decision.  152  The intertwining of domestic politics with international control had, on the whole, a negative impact on international control. The effects were that, first, it probably made international control more polarized. Republicans and conservative support for the military on domestic legislation led them also to ally themselves to the military on international control. These divisions may not have been so stark if the policymaking processes had not run concurrently or if the debate on domestic control had taken place after the debate on international control. Second, scientists, who were a key lobby group for international control, were forced to spend time 153 and energy lobbying on domestic legislation, leaving them less time and energy to think about international control. The debate on domestic atomic legislation also exacerbated divides amongst scientists. This may have 154 lessened their ability to work together on international control. Third, policy decisions on international 155 control were made with domestic results in mind. These decisions may not have led to the best outcomes for international control itself. One prominent example of this was the appointment of Bernard Baruch as chief policymaker on international control. The appointment of someone with better knowledge of and interest in international control may have been a better choice.   , 1962), pp. 428 -455, 482-530 . Also, on political divisions, see  Herken, p. 263. 152 Herken, Bernstein, \"The Quest for Security\"; Grant, Bernard M. Baruch , p. 292. 153 Hogan, A Cross of Iron,  According to historian Barton Bernstein \"American scientists devoted far more energy and thought in 1946 to gaining civilian (rather than military) control of atomic energy than they did to analyzing American plans for international control.\" This dampened their critical engagement with the Acheson-Lilienthal Plan. Bernstein, \"The Quest for Security\".  155  Senior scientists Bush, Conant, and Oppenheimer supported the May-Johnson Bill, whereas most rank and file atomic scientists opposed it. Wang, American Scientists in an Age of Anxiety , pp.14-15.  156    Herken,   \n Viability of International Control Lessons Achieving agreement on a workable scheme for international control is difficult, even if the political atmosphere is conducive or negotiators are more willing to compromise. There may be fundamental structural strategic obstacles, such as an insurmountable transparency-security trade-off. Thus, while there may be interventions 157 to improve the chances of successful international control, such as making policymaking more informed, it may not be enough to achieve successful international control. It may be that radically different political or social circumstances would be required for its success. \n Historical Case Improving processes, with clearer, more transparent, and more informed policymaking would not likely have led to successful international control in 1945/46. This is only likely to have been achieved under radically different historical circumstances. This is because of important underlying factors working against international control. We have grouped these factors below into fundamental structural strategic obstacles, a wider lack of support in the U.S. and the Soviet Union, and confusion in policymaking. These factors were interrelated: the fundamental structural obstacles tended to lead to confusion in policymaking and a lack of support for international control in the U.S. and the Soviet Union. \n Fundamental Structural Strategic Obstacles There are fundamental strategic obstacles to arms control. For example, international control required some staged process. But different stages bestow advantages and disadvantages to different sides. Thus it can be fundamentally challenging to develop a policy that balances these advantages and disadvantages, and more so if the two sides disagree about their power as well as the size of the advantages and disadvantages bestowed by the stages. So, for example, the Baruch Plan suggested a survey of Soviet resources as a first step, without the U.S. reciprocating at that point. This gave the U.S. the option to abort international control at that stage, having gained sensitive information about the Soviet atomic program without equivalent reciprocation.  158  A second obstacle is that any monitoring (or control) scheme which would be transparent enough to assure the other party they will not be caught off guard by a secret armaments program will be too invasive for the  157  This obstacle to arms control is well developed in: Andrew J. Coe and Jane Vaynman, \"Why Arms Control Is So Rare\", American Political Science Review 114,2 (2020), pp. 342-55.  158  \"The American Proposal for International Control Presented by Bernard Baruch\", Bulletin of the Atomic Scientists 1&2 (1 July 1946) pp. 3-5, 10. Also at: http://www.atomicarchive.com/Docs/Deterrence/BaruchPlan.shtml . Accessed 25 April 2019. Secretary of Commerce Henry Wallace made this very observation and criticism of the Baruch Plan, noting that the Soviet Union had only \"two cards which she can use in negotiating with us: (l) our lack of information on the state of her scientific and technical progress on atomic energy and (2) our ignorance of her uranium and thorium resources. These cards are nothing like as powerful as our cards---a stockpile of bombs, manufacturing plants in actual production, B-29s and B-36s, and our bases covering half the globe. Yet we are in effect asking her to reveal her only two cards immediatelytelling her that after we have seen her cards we will decide whether we want to continue to play the game.\"  Herken,  monitored party, exposing them to security risks. In this case, this trade-off bit hard. The U.S. military sought 159 information on Soviet capabilities, and was in fact caught off guard by their development of the atomic bomb. In turn, the Soviet Union perceived it to be a fundamental risk to allow foreign actors to have extensive access to the Soviet Union. It's hard to imagine any monitoring scheme which would have permitted the United States 160 to have sufficient information, which wouldn't have been perceived as an existential threat to the Soviet Union. Lack of Support in the U.S. Many U.S. policymaking elites were weak supporters or even opponents of international control. This was because, first, they believed that so long as the risk of war existed with the Soviets, the atomic bomb was needed for U.S. deterrence and defense. With smaller postwar military forces unable to defend Europe against a Soviet invasion, atomic bombs were seen as an important element of the country's military arsenal. Opponents of international control such as Groves held that the atomic bomb could only be relinquished once war itself was enforceably outlawed, which they largely did not believe possible. Second, policymakers believed that a 161 monopoly on atomic weapons (no matter how short lived), gave the United States coercive diplomatic power. Thus, the atomic bomb was described by some as \"the winning weapon.\"  162  Third, policymakers were also insufficiently worried about a possible arms race. Most were not looking more than a decade or so ahead. In that time frame, many were certain they could keep ahead in any arms race and anyway believed that they would have a monopoly on atomic weapons for many years. This thinking was 163 built on the assumption that Soviet industrial, scientific, and economic resources were significantly poorer than those of the U.S. and certainly not enough for the U.S.S.R. to catch up with the U.S. in terms of atomic research and development. For example, early analyses of the Soviet bomb program assumed that the Soviet Union had a limited ability to work in parallel on the various elements needed to make a bomb. Similarly, the U.S. Air 164  159  See for example: Coe and Vaynman, \"Why Arms Control Is So Rare\".  160  Craig and Radchenko, The Atomic Bomb and the Origins of the Cold War , pp.136, 139. Herken, p. 177.  161  On military opposition, see Lieberman, The Scorpion and the Tarantula , pp. 286-9. For Grove's arguments: Norris, Racing for the Bomb , pp. 471-2.  162  Herken, pp. 4-8.  163  Herken, p.7. Also see previous lesson on Secrecy and Security .  164 Ziegler, \"Intelligence Assessments of Soviet Atomic Capability, 1945 -1949 Gordin, Red Cloud at Dawn , pp. 72-4.  In the Saturday Evening Post in 1948, Groves wrote that the Soviet Union \"simply does not have enough precision industry, technical skill or scientific numerical strength to come even close to duplicating the magnificent achievement of the American industrialists, skilled labor, engineers and scientists who made the Manhattan Project a success. Industrially, Russia is, primarily, a heavy-industry nation; she uses axle grease where we use fine lubricating oils. It is an oxcart-versus-automobile situation.\" From Rhodes, Dark Sun , p. 211. David Lilienthal would later note that Groves went too far in thinking of the Soviet Union as an \"ignorant, clumsy, backward country\".  See Ziegler, \"Intelligence Assessments of Soviet Atomic Capability, 1945 -1949 . This characterization was not limited to Groves, but widely held in government circles. Richard Rhodes notes one popular joke in Washington, D.C., during the war: \"The Russians couldn't deliver an atomic bomb in a suitcase, the joke went, because they didn't know how to make a suitcase.\" Rhodes, Dark Sun , p. 211. The Soviet atomic program was nevertheless a massive and difficult undertaking for the Soviet Union. One historian estimates that the program cost more in absolute terms than the U.S. atomic program: Vladislav M. Zubok, \"Stalin and the Nuclear Age\", in John Lewis  Gaddis et al (eds.)  Force believed in 1946 that it would be many years before the Soviet Union built a fleet of heavy bombers capable of carrying atomic bombs.  165  The belief that the Soviet Union would not be able to develop an atomic bomb in the forties was also founded on the conclusion that the U.S.S.R. did not have access to sufficient high-grade uranium. The rapid exploitation of the Soviet Union's (supposed) low-grade uranium would require, thought Groves, a \"revolution in extraction techniques\" that was beyond the Soviet Union's current technical capabilities. Some (such as the diplomat 166 and analyst George Kennan) also thought that international diplomacy, conflict, or perhaps even a Soviet collapse could work to slow or halt the Soviet atomic development. This confidence was boosted by the 167 assumption that the U.S. could keep ahead by accelerating research and development in atomic weapons. Nor 168 were these developments necessarily incremental: some physicists (in particular Edward Teller) believed that it was possible for the United States to develop the far more powerful thermonuclear \"superbomb.\"  169  Much of the support for international control was transient. For example, cynical realists such as Baruch and his team were only committed to an international control deal in which the U.S. would not need to make any substantial compromises. Once their preferred deals were rejected, they rejected international control altogether. Others, such as senior Republican Senator Arthur Vandenberg, only switched to support the United Nations  \n Lack of Support in the Soviet Union Soviet atomic policy was determined by Stalin, and historians believe that he was unlikely to agree to any form of international control. After August 1945, he was fully committed to the development of the Soviet atomic bomb. Stalin did not appear to fear the destructive effects of the atomic bomb or a subsequent arms race, and conventional forces). He had a cynical attitude to cooperation; \"scientific exchange,\" for example, meant 174 extracting scientific and other insights from the U.S. Stalin was also set against foreign missions in the Soviet  \n Process There was significant confusion and muddling through in policymaking (see the lesson Muddled Policymaking ). Given that one side had atomic weapons and the other did not, it was difficult to create a transitional process which would provide security to both sides and not leave them vulnerable. This was especially problematic given the increasing trust deficit between the two countries in late  1945 and 1946. 181  Even if an international control agreement had been achieved, it may have quickly broken down as all types of monitoring, in practice, were difficult and changing international relations, domestic politics, or public 182 opinion may have led one country or the other to abandon the agreement. 183 \n Distant Counterfactuals Improving processes, with clearer, more transparent, and more informed policymaking would probably not have led to successful international control. Only very radically different historical circumstances, which would have changed the underlying political and social dynamics, may have led to international control. These most distant counterfactuals could be:  \n Risky Cooperation Lessons While, in the abstract, international cooperation is desirable, in practice, steps for cooperation can incur substantial strategic, diplomatic, political, and technological losses. Elites may also personally lose support or political capital by supporting international control. Advocates of cooperation would do well to fully understand the potential risks from cooperation so as to mitigate them. Risks are often not explicitly stated by the concerned parties, and some risks are deeply embedded in the circumstances, institutions, or world view.  185  Historical Case \n Risks for the United States Entering into or even discussing international control with the Soviet Union carried substantial strategic and diplomatic risks for the United States and political risks for the incumbent administration. Simply discussing international control with the Soviet Union would lessen the diplomatic leverage that atomic bombs might bring whilst the discussions were ongoing. Discussions may also have risked revealing 186 information about key aspects of the U.S. atomic program, such as the highly secret effort to monopolize global high-grade uranium deposits. Negotiations consequently carried the risk of slowing down the U.S. atomic 187 program and possibly even allowing the Soviets to catch up.  188  International control could also lead to the sharing of technical or strategic information (e.g., how few bombs the U.S. had) that would allow an acceleration of the Soviet bomb program or other strategic advantage over the United States. Stalin had decided that the Soviet Union should attempt to copy the U.S. bomb-making process, so any technical information gleaned may have been useful. Moreover, in 1946, the U.S.S.R. was struggling with the construction of its first experimental reactor and the large-scale separation of uranium, and even the Russian translation of the Smyth Report helped with the Soviet program in early 1946.  189  There were also risks associated with public reaction. First, starting negotiations on international control raised public expectations that a favorable agreement may eventually be reached. Yet such expectations may not be met, which would reflect unfavorably on the administration. That was one of the reasons why it was important for Baruch and other U.S. policymakers to ensure that the Soviet Union would be blamed if/once international control negotiations failed. Second, international control discussions risked further sensitizing the U.S. public 190 to the destructiveness of the atomic bomb, making it harder to mobilize the public for military action.  191  There were also risks associated with cooperating with allies. Doing so risked increased leaks to the Soviet Union, and indeed, British scientists such as Klaus Fuchs did pass on information from the Manhattan Project to the Soviet Union. The U.S. also risked offending allies by direct bargaining with Soviet Union. For 192 example, the British wished to be included in any negotiations on international control. When Secretary of State Byrnes did directly bargain with the Soviet Union in late 1945, Britain and especially France were offended, and it harmed the alliance. 193 \n Risks for the Soviet Union Negotiating or starting the process of international control also carried significant risks for the Soviet Union. International inspections could reveal raw materials and facilities, and facilitate a preventive attack on the Soviet program. Inspections and openness could undermine regime stability. U.S. proposals advocated a gradual 194 staged process for the institution of international control. From the Soviet point of view, this gave the U.S. an advantage in the earlier stages of international control. The Bush Plan, for example, suggested three stages: (1) basic information sharing, (  2 ) inspections (at which point each country would reveal its atomic facilities and resources), and (3) transfer of resources and material. The Acheson-Lilienthal Plan was more nuanced. But 195 that too stipulated that the U.S. and the Soviet Union would give up information on their atomic facilities in gradual stages. Similarly, the formal handover of atomic facilities to the U.N. would occur in stages. Crucially, the handover of atomic bombs themselves would only occur at the end. The plan did not stipulate when the U.S. would stop manufacturing bombs. The first detailed denunciation of the Baruch Plan in the Soviet press Such an organization, and so international control itself, could be used against the Soviet Union. This concern can be seen in the Soviet insistence that (1) the UNAEC report to the U.N. Security Council and not the General Assembly (which was perceived to be even more biased towards the U.S.) and (  2 ) that the veto apply to the deliberations of the UNAEC and to international control. The Soviet Union did not yet possess a nuclear capability but that it would very soon do so, after which all history made it clear that sooner or later there would be a war between the two superpowers that would be infinitely more devastating than either of the two world wars through which he had lived. The only way of preventing this Armageddon, he concluded with remorseless if unpalatable logic, was for America to launch a nuclear attack on the Soviet Union before it acquired the bomb: after that it would be too late.  204  The following are, in our estimation, the reasons why a preemptive strike or a preventive war was not launched by the U.S. in the late 40s. \n Intent and Appetite There was no public appetite for another major war due to significant war weariness. The public wanted wide-scale demobilization and expressed concern at continued overseas deployments. Truman and other 201 Russell D. Buhite and WM. Christopher Hamel, \"War for Peace: The Question of an American Preventive War against the Soviet  Union, 1945 \", Diplomatic History 14,3 (1990 , pp. 367-84. Melvyn P. Leffler, \"Strategy, Diplomacy, and the Cold War: The United States, Turkey, and NATO, <PHONE>\", The Journal of American  History 71,4 (March 1985) , pp. 807-25. Secretary of Defense James Forrestal would write in 1947 that \"The years before any possible power can achieve the capability effectively to attack us with weapons of mass destruction are our years of opportunity.\" Quoted in John Lewis Gaddis, Strategies of Containment: A Critical Appraisal of Postwar American National Security Policy (New York: Oxford University Press, 1982), p. 62. See also George H.  Quester, Nuclear Monopoly (New Brunswick: Transaction Publishers, 2000) , chapter 4.  202  Buhite and Hamel, \"War for Peace\".  203   politicians responded to this public pressure. An overt act of aggression would not have played well with 205 public opinion and would have been disagreeable to some policymakers who saw it as being against American principles. It was only after 1950 that support for a preventive war against the Soviet Union began to grow 206 appreciably amongst the U.S. public. This, in turn, spurred talk of preventive war amongst policymakers. \"For the first time,\" noted Newsweek in February 1950, \"some members of Congress were beginning to speculate on what had formerly been an almost forbidden subject -preventive war.\"  207  Although many potential supporters of a preventive strike were hawks, there were also policymakers who were instinctively much less hostile to the Soviet Union. In 1950 and 1951, for example, when a preventive war was more openly discussed in policymaking circles, key policymakers such as Secretary of State Dean Acheson expressed significant concern. Hawkish attitudes may have been dampened by the fact that the U.S. military 208 and many in government were overconfident about their atomic lead over the Soviet Union and had underestimated the progress being made by the Soviet program, particularly the Soviet Union's ability to acquire high-grade uranium. Many policymakers were so confident of the U.S. lead that they even opposed If it had had credible information about the progress of the program, this could have acted as a focusing event for a war.  211  Hawkish attitudes may also have been dampened by the belief, held by many in the U.S. administration, that Soviet expansionism (more a concern than the Soviet atomic program in the 40s) could be contained through diplomacy, alliance building, and initiatives such as the Marshall Plan. There was a general consensus in the U.S. military establishment in the late 40s that the Soviet Union wished to avoid military engagements.  212  Targeted strikes may have been unpalatable because they would probably have led to general war, which would be very costly: Europe and parts of Asia would likely have been invaded and occupied by the Soviet Union for at least a year, and it would have led to millions of Soviet deaths, hundreds of thousands if not millions of U.S. allied deaths, and tens of thousands of U.S. combat deaths. 213 \n Military Capabilities Following demobilization, U.S. armed forces were inadequate for a military defeat of the Soviet Union, and the U.S. was well aware of this. Soviet air forces, radar, and anti-aircraft guns continued to improve through to the fifties. It is also unclear whether the U.S. had the capabilities to carry out a sufficiently effective preventive 214 strike. Soviet atomic facilities were widespread and not easily attacked by the United States. Lack of intelligence meant that U.S. military planners had immense difficulty selecting targets for aerial attack (conventional or atomic) in the late 40s. The U.S. had a poor level of nuclear readiness. They had very few atomic bombs: by 215 the end of 1947, the U.S. only had 13, and by 1948, only 50. There were also issues with bomb assembly and delivery capabilities.  216  Could There Have Been a Preventive Strike? Counterfactually, then, a preventive strike would have become a realistic option if a certain number of factors had been present: amongst policymakers, a more alarmist assessment of the Soviet Union's atomic program and its progress, and better intelligence about Soviet atomic facilities. For example, more serious crises in Turkey and the Middle East, then of significant concern for the U.S., may also have helped make a stronger case for a strike against Soviet atomic facilities. In 1953, following the Soviet detonation of an H-bomb, Truman briefly 217 thought about a preventive strike. With the appropriate intelligence, and if the circumstances had been favorable, he may have considered a strike earlier, prior to the Soviet atomic bomb test in 1949. Factors such as less demobilization of military forces and less war weariness amongst the public may also have increased the chances of a preventive strike as a realistic option. \n Conclusion and Extensions Our study of attempts at the international control of atomic energy in 1945-46 suggests that radical schemes for international governance can get widespread support, even from skeptics, but that the support can be tenuous and fleeting. Technical experts can bolster support, but muddled policymaking, secrecy, and concerns over security can undermine it. Our lessons point to the difficulties inherent in attempting to achieve international control and to the deep intertwining of technical and political issues. It is, in fact, amazing that debates on international control got as far as they did in 1946 (that is, in fact, our first lesson). There are, however, opportunities for those pushing towards international governance: even cynics can support proposals, and public opinion and technical expertise can be powerful sources of support. The history of atomic international control is too rich and broad to be fully captured in this report, and can provide many other lessons for future powerful technologies. Questions and topics worthy of further inquiry include: • The role of activists and activism. How did activists form and maintain their organizations? How were they funded, and did that matter? What tactics and organizations were especially successful? • How are the politics of international control impacted by traditionally important features of the political landscape, such as partisan divides, the judiciary, strong executives, lame-duck presidents, strong or weak incumbents, upcoming elections, etc.? • Can we say more about the role and dynamics of the public sphere? What ideas or framings were most likely to resonate? What communications were most impactful (e.g., lectures, radio talks, interviews, presidential speeches)? • The military responded in complex and varied ways to atomic weapons. More work should be done to understand the extent to which these responses were shaped by organizational interest, ideas, personal idiosyncrasies, and other factors. Other moments in the global politics of nuclear weapons and of other powerful technologies also warrant study. These include: • Earlier negotiations revolving around the abolition of large classes of technologies, especially the naval arms treaties of the 1920s, the 1925 Geneva Protocol on the prohibition of the use of chemical weapons, and the discussions at the 1932 Geneva disarmament conference (e.g., aviation) may give useful insights into possible directions for modern transformative technologies.    State 1949 State -1953 . A leading supporter of international control within the State Department in the first half of 1946. Chair of the Special State Department Committee tasked with the preparation of a plan for international control in December 1945. Acheson was keen that the Committee 245 succeed in its task not only because he wanted international control and atomic cooperation with the Soviet Union, but also because he wanted the Committee to inform politicians and diplomats on atomic matters, and because he, like Byrnes, wanted to retain as much policymaking/expertise within the State Department as possible (and not lose it to Groves or some military committee). Once Baruch was appointed as the U.S. 246 in late 1946. This included strong public criticism of the Baruch Plan-certainly the strongest public criticism from within the government. For this criticism he was forced to resign in September 1946. 256 Winne, Harry A .(<PHONE>) -Vice-President in charge of engineering at General Electric and member of the board of consultants charged with preparing a report on international control in January 1946.  256  Hewlett and Anderson, Jr., A History of the United States Atomic Energy Commission vol. 1, pp. 597-606. Figure 1 . 1 Figure 1. Nuclear warheads in the U.S. and U.S.S.R., at three temporal zoom levels spanning <PHONE>. Note that the y-axis scale increases by orders of magnitude between the figures. From: Max Roser and Mohamed Nagdy, \"Nuclear Weapons\", OurWorldInData.org (2013). Available at: https://ourworldindata.org/nuclear-weapons . Accessed 26 August 2020. \n lesson 4.4 , secrecy, lack of information, and misconceptions about the development of atomic weapons shaped policy and negotiations. No one in government, not even Truman, had a clear idea of the number of atomic bombs in U.S. possession in 1946 (Truman would be visibly shocked when Lilienthal revealed in a 1947 inventory exercise how few the U.S. possessed). Even the U.S. military made war plans 108 without understanding the number and deliverability of U.S. atomic weapons. Curtis LeMay (then deputy 109 chief of Air Staff for Research & Development, and the future commander of the Air Force's Strategic Air Command) complained in 1946 that the Air Force struggled to plan for atomic bomb delivery because of secrecy surrounding the number and nature of the U.S. bomb stock. Indeed, one January 1946 plan by the 110Joint War Plans Committee worked off the assumption that the Air Corps had access to 196 bombs, while the actual number for the U.S. as a whole was around 7. International control policy and negotiations were 111 largely developed under the general impression that the Soviets were many years away from completing their first atomic bomb. An important wartime agreement between Roosevelt and Churchill to cooperate on 112107  Charles E. \n 135 recognition.\" One of the reasons why Oppenheimer's role in policymaking in 1946 was limited to his 136 participation in the drafting of the Acheson-Lilienthal Report was because he made a poor personal impression on Truman during their first private meeting. On the Soviet side, atomic diplomacy appears to have been 137 largely personally determined by Stalin, with diplomats having little autonomy. 138 \n 170 when 170 it was politically expedient (and possibly electorally popular). Support from key policymakers (e.g., the171 wartime Secretary of War Henry L. Stimson and Secretary of State James F. Byrnes) also waxed and waned depending on their assessment of Soviet flexibility and U.S. progress in atomic weapons. The public was itself 172 easily alarmed by security and secrecy concerns (see previous lesson Secrecy and Security ). \n 209 the development of a program for detecting Soviet nuclear tests. The U.S. government, in fact, had little 210 information about the Soviet bomb program. \n Finally , it is worth a reminder that all of these historical episodes provide only circumscribed lessons for future powerful technologies, such as AI. None of them offer a clean analogy. Rather, they are sources of inspiration, insight into mechanisms and dynamics, and examples of how politics can play out. claimed in May 1945) . Groves was crucial for persuading Truman and others that Russia would not get the 243 bomb for many years. 244  Politicians and Others Acheson,Dean  (1893Dean  ( -1971 -Statesman and lawyer. Undersecretary of the U.S. Department of State from August 1945 to June 1947; Secretary of \n The Soviet Union similarly was interested in the propaganda value of negotiations. It hoped to generate negative publicity for the U.S. and extract as much information as possible on the U.S. program. The final vote in the Security Council in December 1946 had 10 UNAEC votes in favor, and two abstentions (the Soviet Union and Poland); these abstentions were understood as an effective veto.Both the U.S.'s and the Soviet Union's atomic programs continued unhindered whilst the negotiations were carried out. U.S. atomic bombs became more advanced and increased in size and number in the late forties. The Soviet Union eventually carried out its first atomic bomb test in August 1949, catching most U.S. intelligence and military planners by surprise. Stimulated by this, the United States developed the much more powerful hydrogen bomb, testing it in November 1952; the Soviet Union followed soon thereafter, in November 1955. Through these years, the quantity of atomic bombs, and then hydrogen bombs, increased exponentially, from less than 20 in 1947 to more than 100 in 1949 and to more than 10,000 by 1959 (see Figure1). 12    Alongside the development of nuclear arsenals came proliferation: Britain tested its first device in October 1952 ,  France in February 1960, and China in October 1964. The nuclear arms race continued not just in quantity and geographical spread, but also through the invention of qualitatively new systems, like submarine launched missiles, MIRVs (multiple independently targetable reentry vehicles), battlefield nuclear weapons, and missile defense systems (anti-ballistic missiles and \"Star Wars\"). \n Niels Bohr's Memorandum to President Roosevelt (July 1944). Available at: http://www.atomicarchive.com/Docs/ManhattanProject/Bohrmemo.shtml . Accessed 22 September 2018. On Bohr's activism, see Martin J. Sherwin, \"Niels Bohr and the First Principles of Arms Control\", in Herman Feshbach, Tetsuo Matsui, and Alexandra Oleson (eds.), Niels Bohr: Physics and the World (Abingdon: Routledge, 1998), pp. 319-30; Finn Aaserud, \"The Scientist and the Statesmen: Niels Bohr's Political Crusade during World War II\", Historical Studies in the Physical and Biological Sciences 30,1 (1999), pp. 1-47. 13 Niels Bohr, 14 Vannevar Bush and James B.Conant, \"Memorandum\" (30 September 1944). Available at: https://nsarchive2.gwu.edu//NSAEBB/NSAEBB162/1.pdf . Accessed 22 September 2018. Also: G. Pascal Zachary, Endless Frontier: Vannevar Bush, Engineer of the American Century (New York: Simon and Schuster, 2018) , p. 243; James Hershberg, James B. Conant: Harvard to Hiroshima and the Making of the Nuclear Age (New York: Alfred A. Knopf, 1993), pp. 204-5. \n The Role of the Political Entrepreneur in the Context of Policy Change and Crisis, Midwest Political Science Association Annual Conference, Chicago, April 14th 2013. Available at https://arrow.dit.ie/cgi/viewcontent.cgi?article=1015&context=buschgracon . Accessed 22 April 2019. 38 Herken, pp. 202-04, 212. George W. Baer, One Hundred Years of Sea Power: The U.S. Navy, <PHONE> (Stanford: Stanford University Press, 1996), pp. 287-88; Edward Kaplan, To Kill Nations: American Strategy in the Air-Atomic Age and the Rise of Mutually Assured Destruction (Ithaca, NY: Cornell University Press, 2015), pp. 22-28. Stimson, for example, advised President Roosevelt  in December 1944  not to give atomic information to the Soviet Union without a \"real quid pro quo,\" such as liberalization of domestic Soviet rule. In June 1945, he added that the quid pro quo could include international control, or a negotiated settlement over the fate of Eastern Europe. By the time of the first atomic bomb test in July 1945, Stimson had abandoned hope of simple cooperation with the Soviet Union and instead argued that the U.S. should force Soviet liberalization as a precondition for cooperation on the atomic bomb. By 37 For a review of the literature on political entrepreneurs, see J. Hogan and S. Feeney, 39 Barton J. Bernstein, \"The Uneasy Alliance: Roosevelt, Churchill, and the AtomicBomb, 1940Bomb,  -1945\", The Western Political Quarterly 29,2 (June 1976), pp. 202-30; Graham Farmelo, Churchill's Bomb: How the United States Overtook Britain in the First Nuclear Arms Race (New York: Basic Books, 2013), p. 331. 40 See, for example, Atomic Scientists of Chicago, The Atomic Bomb: Facts and Implications (Chicago: The Atomic Scientists of Chicago, 1946). \n . Sean L. Malloy, Atomic Tragedy: Henry L. Stimson and the Decision to Use the Bomb Against Japan (Ithaca, NY: Cornell University Press, 2008), pp. 110-12, 145-53. 42 Herken, pp. 92, 97, 98; Lieberman, The Scorpion and the Tarantula, pp. 234-5. 43 Herken, pp. 139-40; John Lewis Gaddis, The Cold War , pp. 28-9; Fernande Scheid Raine, \"The Iranian Crisis of 1946 and the Origins of the Cold War\" in Melvyn P. Leffler and David S. Painter (eds.), Origins of the Cold War: An International History 2nd ed. (New York: Routledge, 2005), pp. 93-111; and Eduard Mark, \"The Turkish War Scare of \n Herken, p.138. Carl Shulman has pointed out that this conclusion is in some ways surprising given that the GDP of the Soviet Union at that time was roughly 14 times the cost of the Manhattan Project.This is calculated by taking the estimated GDP of the Soviet Union in 1946 (this is a rough estimate, and no reliable figure for 1945 is available) as USD 664.646 billion (in 2011 USD from the Maddison Project Database 2018 at https://www.rug.nl/ggdc/historicaldevelopment/maddison/ , accessed 28 May 2020) and the cost of the Manhattan Project as USD 1.889 billion (in <PHONE> dollars, from Stephen I. Schwartz, Atomic Audit: The Costs and Consequences of U.S. Nuclear Weapons Since 1940 (Washington, DC: Brookings Institution Press, 1998), p. 60). Converting the latter figure to 2011 dollars using https://www.measuringworth.com/calculators/uscompare/ gives a Manhattan Project cost figure (using a production worker compensation inflator) as USD 48.4 billion. 56 Maddock, Nuclear Apartheid , chapter 3. 57 Herken, pp. 139-40, 175. S. David Broscious, \"Longing for International Control, Banking on American Superiority: Harry S Truman's Approach to Nuclear Weapons\", in John Lewis Gaddis et al (eds.The Cold War , pp. 28-9; Raine, \"The Iranian Crisis of 1946 and the Origins of the Cold War\"; Mark, \"The Turkish War Scare of 1946\". One of the reasons why Secretary of State Byrnes formed a committee of consultants to look into international control in <PHONE>: Myths, Monopolies and Maskirovka\", Intelligence and National Security 12,4 (1997), pp. 1-24; Gordin, Red Cloud at Dawn , pp. 72-74; Jonathan E. Helmreich, Gathering Rare Ores: The Diplomacy of Uranium Acquisition, <PHONE> (Princeton, NJ: Princeton University Press, 1986), pp. 248-9. 55 ), Cold War Statesmen Confront the Bomb: Nuclear Diplomacy Since 1945 (Oxford: Oxford University Press, 1999), pp. 15-38. On Truman's growing confrontational stance towards the Soviet Union, see Gaddis, 58 Craig and Radchenko, The Atomic Bomb and the Origins of the Cold War , p. 130. See also: Public Sphere lesson . \n issues, and in the planning and execution of the atomic bombing missions. Using his privileged 64 access to secret information, Groves was able to win arguments with adversaries by pointing to their ignorance, e.g., about the duration of the U.S. monopoly. He was able to avoid his policies being questioned, e.g., by65    Congress. Even his collaborators found themselves pulled along by Groves' faits accomplis. He discredited 66 others as being uncareful with secret information or even being treasonous. He did this with Niels Bohr, J. Robert Oppenheimer, Leo Szilard, and David Lilienthal. Groves used his access to secret information to shape 67 policymaking by recommending his \"technical advisors\" for key decisions. For example, he almost scuttled Lilienthal's consultant group on international control by recommending his technical advisors instead. Groves 68 also used his privileged access to directly influence key policymakers such as Truman. Truman's belief in early 1946 that the U.S. would have a long-lived atomic monopoly was due directly to Groves' arguments and influence. Groves was also able to shape public opinion and whip up public and policymaker concern over 69 security and spying. 70  Secrecy led to poor information flow and so to bad decision-making. Two examples of this are especially prominent. First, Groves imposed his view of the duration of the U.S. atomic monopoly on the U.S. government by preventing reasoned debate. In late 1944 and early 1945, Vannevar Bush, James B. Conant, and Leo Szilard-who disagreed with Groves and believed that the Soviet Union could access high-grade uranium ore and so build an atomic bomb relatively quickly-were stifled through secrecy regulations (specifically the silo information structure within the Manhattan Project). Groves's views also carried significant weight with Carson \"Facing Off and Saving Face: Covert Intervention and Escalation Management in the Korean War\", International Organization 70,1 (2016), pp. 103-31. Also: Allison Carnegie and Austin Carson, \"The Disclosure Dilemma: Nuclear Intelligence and International Organizations\", The American Journal of Political Science 63,2 (2019), pp. 269-285. 62 Herken, pp. 127, 132, 136; Gregg Herken, \"'A Most Deadly Illusion': The Atomic Secret and American Nuclear  Weapons Policy, 1945\", Pacific Historical Review  49,1 (February 1980, pp. 51-76; Kimball-Smith, A Peril and a Hope , pp. 373-5, 387-8; Hogan, A Cross of Iron , p. 238.63  On Groves' \"compartmentalization' strategy see: Sherwin, A World Destroyed , pp. 58-62. international 71 the public. One historian has noted that \"Groves's predictions carried more weight with the public than anyone else's, precisely because he was the individual expected to have the largest amount of secret information upon which to base an estimate. Few challenged him directly….\" . Second, information on atomic bomb production 72 \n in Canada and the United States in February 1946. These news stories increased support for the May-Johnson Bill, which advocated military control of atomic energy policy.86    Public opinion was not uniform or entirely coherent. It sometimes contained views that were in tension with one another. For example, surveys in September 1945 revealed that ~70% of citizens did not want to share the secret of the atomic bomb with other countries. At the same time, however, 90% thought that the U.S. would not be able to keep the secret for long anyway and other countries would soon build the bomb. Polls in 87 October 1945 showed 17% support for international control through the U.N. Security Council, but 67% support for \"England, Russia, France, [the] United States, China, and other countries\" to \"get together to agree that atomic bombs should never be used as a war weapon.\" A 1947 poll showed that a majority believed that 88 atomic bombs made war less likely but were also willing to initiate an atomic war. The public had erroneous 89 technical beliefs, such as that, by October 1947, ~60% of the public surveyed \"thought that Russia was manufacturing atomic bombs in quantity\"; ironically, while the public was very mistaken here, they were comparably mistaken but in the opposite direction as the most informed expert, General Groves. 90 86 Groves would himself later discount the effectiveness and importance of this spying. Herken, p. 130-33. There is a possibility that the source was in the FBI (perhaps the Director himself, J. Edgar Hoover) or from within the Justice Department; see: Ellen Schrecker, Many Are the Crimes: McCarthyism in America (Princeton, NJ: Princeton University Herken, p. 268; Steven Miller,  Strategy and Nuclear Deterrence: An International Security Reader (Princeton, NJ:  Princeton University Press, 1984), p. 123.85  Lieberman,The Scorpion and the Tarantula , p. 311.operatingPress, 1999), p. 170; Craig and Radchenko, The Atomic Bomb and the Origins of the Cold War , pp. 121-22. 87 Herken, p. 32. 88 Hadley Cantril, Public Opinion <PHONE> (Princeton, NJ: Princeton University Press, 1951), p. 22. 89 Herken, p. 311. 90 Herken, p. 232. \n Kai Bird and Martin J. Sherwin,    American Prometheus: The Triumph and Tragedy of J. Robert Oppenheimer (New York: Vintage Books, 2006), xi. 93 Herken, Barton J.Bernstein, \"Scientists and Nuclear Weapons in World War II\", in Thomas W. Zeiler and Daniel M. DuBois (eds.), A Companion to World War II volume 1 (Oxford: Wiley-Blackwell, 2013), pp. 516-48. 95 Ibid. \n al, A Report on the International Control of AtomicEnergy , and Herken,. The reliance on denaturing turned out to be misguided; see later sections .[Bohr]  is very near the edge of mortal crimes\" for discussing atomic matters with Soviet citizens. There were 103 a few public denouncements of scientists in the forties; one of the most serious was in March 1948 when the House Un-American Activities Committee denounced theoretical physicist Edward Condon, then director of the National Bureau of Standards, as ''one of the weakest links in our atomic security.\" Privately, the security 104 establishment was intensely suspicious of many internationalist scientists. In early 1948, the Justice Department considered prosecuting nuclear physicist Leo Szilard under the Logan Act (which criminalizes negotiation by unauthorized persons with foreign governments having a dispute with the United States) following the publication of his \"Letter to Stalin\" in the Bulletin of the Atomic Scientists . In this open letter (which was not actually separately delivered to the Soviet Union), Szilard called for international control through the United Nations and for Stalin to speak directly to the U.S. people in this regard. In 1950, the FBI even launched an investigation into Albert Einstein, eventually accumulating over 1,500 pages of evidence on him.105    Scientists were often reactive rather than proactive and sometimes were overtaken by events or failed to respond effectively to them. Other elites (such as Groves, for example, or Truman) often drove policymaking or nudged public opinion in various directions, and scientists could only react, often not effectively. Two prominent examples of this are the spy revelations (and subsequent controversy) in February/March 1946 and the appointment of Bernard Baruch as the U.S. representative of the UNAEC with authority to shape international control proposals to his own liking. In both cases, scientists were unhappy with the turn of events, were caught unprepared, and were unable to develop effective responses. Scientists' inability to react effectively was due to Leo Szilard and the Crusade for Nuclear Arms Control (Cambridge, MA: The MIT Press, 1987), p. Xl; Leo Szilard, \"Calling for a Crusade\", Bulletin of the Atomic Scientists(May 1947), pp. 102-6, 125.  Einstein was particularly forward in writing to Soviet scientists on arms control, see Craig and Radchenko,The Atomic Bomb and the Origins of the Cold War , pp. 147-8. 103 Aaserud, \"The Scientist and the Statesmen\". 104 David Kaiser, \"The Atomic Secret in Red Hands? American Suspicions of Theoretical Physicists During the Early Cold War\", Representations 90,1 (Spring 2005), pp 28-60. 105 Wittner, The Struggle Against the Bomb volume 1, pp. 267-8; Helen S. Hawkins et al (eds.), Toward a Livable World: 100  Lawrence S.Wittner, The Struggle Against the Bomb volume 1: One World or None: A History of the World Nuclear Disarmament Movement Through 1953 (Stanford: Stanford University Press, 1993), p. 264. 101 Herken, p. 264. 102 Ray Monk, Robert Oppenheimer: A Life Inside the Center (New York: Doubleday, 2012), p. 494. \"106their distance from the powers of decision-making. In both cases, Truman and those directly around him determined policy; scientists were not part of the inner circle.106 Herken,  A Peril and a Hope, 464.    \n Lindblom, \"The Science of 'Muddling Through'\", Public Administration Review 19,2 (Spring, 1959), pp. 79-88. 108 Curatola, Bigger Bombs for a Brighter Tomorrow , pp. 52-3; David Alan Rosenberg, \"U.S. Nuclear Stockpile, 1945 to 1950\", Bulletin of the Atomic Scientists 38,5 (May 1982), pp. 25-30; David Alan Rosenberg, \"The Origins of Overkill: Nuclear Weapons and American Strategy \n : A Reinterpretation\", Political Science Quarterly 90,1 (Spring 1975), pp. 23-69; Herken, p. 62. 114 As, for example, is illustrated by Bernard Brodie recounting the moment when his career pivoted to the study of nuclear weapons. Fred Kaplan, The Wizards of Armageddon 2nd edition (Stanford: Stanford University Press, 1991), p. 10. 115 Henry Stimson, Memorandum Discussed with the President (25 April 1945). Available at: http://www.nuclearfiles.org/menu/library/correspondence/stimson-henry/corr_stimson_1945-04-25.htm , accessed 15 October 2019. This memorandum was discussed in the May 31 meeting of the Interim Committee, see Notes of the Interim Committee Meeting, Thursday 31 \n Henry Stimson, Memorandum on the Effects of the Atomic Bomb  (11 September 1945). Available at: http://www.nuclearfiles.org/menu/library/correspondence/stimson-henry/corr_stimson_1945-09-11.htm , accessed 16 October 2019.Herken, p. 30; Lieberman,    \n et al, A Report on the International Control of Atomic Energy , pp. 26-7. Anthony Leviero, \"Denaturing Guard on Atom Not Sure\", , pp. 161-220. C.E.Till, \"Denatured Fuel Cycles\", in Joseph L. Fowler, Cleland H. Johnson, Charles D. Bowman  (eds.), Nuclear Cross Sections for Technology: Proceedings of the International Conference vol 13 (Oak Ridge, TN: Oak Ridge National Laboratory, 1980), p.115-18. demarcated responsibilities for atomic policymaking and advice: \"I have never participated in anything that was so completely unorganized or so irregular...It is somewhat appalling...to think of this country handling many matters in such an atmosphere.\" Many, such as Groves in late 1945, became powerful, but still 121 unofficial, advisors. 122  Because of this lack of clarity, individuals often competed against each other to influence policymaking. Vannevar Bush, for example, suggested a new committee, including the State Department, scientists, and representatives of Congress, to make atomic policy in November 1945 but was rebuffed. Secretary of State123Byrnes fought with Senator Arthur H. Vandenberg and others to retain sole control over atomic diplomacy in late 1945. One of the reasons Byrnes set up a committee to make a proposal for international control in 124 December 1945/January 1946 was to keep atomic diplomacy within the State Department, instead of losing it to Congressional committees. The Acheson-Lilienthal Report was produced largely by its technical board of 125 consultants. Leslie Groves had opposed the appointment of this board, and had attempted to get a board to his liking appointed instead. Similarly, the State Department later opposed Bernard Baruch's board of consultants (which eventually helped formulate the Baruch Plan) and also attempted to get its own board appointed. 126   New York Times (10 April 1946), p. 16. Kimball-Smith, A Peril and a Hope , pp. 462-3. For more recent thinking on denaturing see: A. DeVolpi, \"Denaturing Fissile Materials\", Progress in Nuclear Energy 10,2 (1982) 120  Barnard et  al, A Report on the International Control of Atomic Energy .clearly \n ; Craig and Radchenko, The Atomic Bomb and the Origins of the Cold War , pp. 97-8, 102, 105, 109; Holloway, Stalin and the Bomb , chapter 8. Actors attempted to understand each others' intentions, but often sent out mixed signals and misunderstood each others' positions. When Soviet Foreign Minister Vyacheslav Molotov arrived in New York in October 139 \n Secretary of State Byrnes' exclusion of the French from the December 1945 Moscow conference, for example, \"needlessly offended that ally\" and allowed Stalin to magnanimously invite the French to future conferences. Similarly, not including general 142 disarmament into the Baruch Plan, as Baruch wanted, gave the Soviets the opportunity to announce it as part of their rival proposal later on. Byrnes omitted key passages on stages in the U.S. proposal presented at Moscow 143 in December 1945. This upset Truman, and Byrnes was forced to include them later on. This misstep further damaged his relationship with Truman. Byrnes also could have worked harder to build support in Congress 144before embarking on his atomic diplomacy in late 1945. 145   \n 156 151 On the legislative history of the Atomic Energy Act, see Richard G. Hewlett and Oscar E. Anderson, Jr., A History of the United States Atomic Energy Commission volume 1 The New World 1939/1946 (University Park, PA: The University of Pennsylvania Press \n , Cold War Statesmen Confront the Bomb: Nuclear Diplomacy Since 1945 (Oxford: Oxford University Press, 1999), pp. 39-61. \n Union, for example, for inspections or monitoring. Soviet delegate to the UNAEC Andrei Gromyko would 176 reflect forty years later that \"I am certain that Stalin would not have given up the creation of his own atomic bomb. He well understood that Truman would not give up atomic weapons.\" 177 Tensions between the U.S. and the Soviet Union Growing tensions between the U.S. and the Soviet Union undermined confidence in each other, increased the allure of atomic weapons, and no doubt made agreement harder to reach. It led to, in Gaddis' words, a \"growing sense of insecurity\" in 1945 and 1946. We know most about U.S. attitudes, where historians have 178 noted that disagreements over postwar Europe, for example over issues in Poland, Romania, and Germany, and later over Iran, decreased U.S. trust in the Soviet Union from 1945 onwards. The discovery of Soviet atomic espionage from 1943 detracted from confidence building, though historians are divided on how much of an impact it may have had on U.S. policy towards the Soviet Union. On the Soviet side, there is some indication 179 that certain U.S. policies, such as Truman's cancellation of lend-lease to the Soviet Union, may have hardened Soviet attitudes towards the U.S., leading in this case to an increase in \"unilateralist\" tendencies.180     \n • If the Soviet Union was willing to settle for a junior status internationally vis-a-vis the United States. This might have come about in a variety of ways, including, one historian has speculated, through Stalin's death in the fall of 1945, in which his successors \"might have chosen a more accommodating course toward the United States\" . If there had been less conflict of interest between the Soviet Union and the U.S. in Europe and Asia. For example, if the Soviets were less expansionist and/or the U.S. more accepting of Soviet expansionism.• If there had been both conventional and nuclear power parity between the two superpowers, making it easier for them to identify a symmetric arms control bargain. • If policymakers were radically more long-termist, with more foresight and perception of the dangers of a nuclear arms race, or otherwise radically more concerned about the nuclear arms race.• If the U.S. carried out a successful preventive strike on the Soviet Union, thus stalling the Soviet atomic program, or otherwise was able to coercively stop the Soviet program. 184• \n 199    Lastly, both the U.S. and Soviet Union could lose bargaining leverage by offering cooperation first. According to historian John Lewis Gaddis, for example, Truman lost bargaining advantage by unilaterally committing the U.S. to international control in late 1945.200    .10 Preventive Strike Lesson Even the most violent of solutions, such as a preventive war or a preventive strike, may gain traction.Historical CaseDuring the years of the U.S. atomic monopoly, <PHONE>, many  in the U.S. and Britain favored resorting to force to prevent the Soviet Union from obtaining atomic weapons. Generals such as Henry H. \"Hap\" Arnold, Carl Spaatz, and Curtis LeMay argued this, as did internationalists such as Ely Culbertson. Hawkish 201 intellectuals published books suggesting preventive war. In If Russia Strikes (1949), George Eliot, a leading military analyst, called for the U.S. to present an ultimatum to Moscow: accept the Baruch Plan or the U.S. would use atomic bombs against Soviet atomic facilities. Political scientist James Burnham, in his 1947 Struggle for the World , called for political subversion to destroy the Communist state, or if that failed, then air strikes on Soviet military targets. New York Herald Tribune journalists Joseph and Stewart Alsop called for preventive war in their columns. Even pacifist and socialist-friendly Bertrand Russell called for the U.S. to threaten war as 202 \"part of a plan he had developed to promote global peace.\" As recounted by a member of the audience, in one 199 Herken, pp. 84, 174.  200  John Lewis Gaddis, Strategies of Containment: A Critical Appraisal of American National Security Policy during the Cold War 2nd ed. (Oxford: Oxford University Press, 2005), p.17. 4203 speech Russell argued that: \n David Blitz, \"Did  Russell Advocate Preventive Atomic War Against the USSR?\", The Journal of Bertrand Russell Studies 22 (summer 2002), pp. 5-45. 204 Quoted in Blitz, \"Did Russell Advocate Preventive Atomic War Against the USSR?\". \n • Negotiations over the Strategic Arms Limitation Talks agreement (signed 1972), the Anti-Ballistic Missile Treaty (1972), the Biological Weapons Convention (1972), the Strategic Arms Limitation Talks II agreement (1979), the Strategic Arms Reduction Treaty (1991), and the Chemical Weapons Convention (1993) may also be relevant. 219219 There is a very large literature on these agreements and efforts. For an overview of these, see Robert E. Williams, Jr., and Paul R. Viotti, Arms Control: History, Theory, and Policy volume 1: Foundations of Arms Control (Santa Barbara, CA: ABC-CLIO, 2012), chapters14,15,16,17,18.    \n\t\t\t There is new material appearing into the public domain at regular intervals, see for example David Holloway, \"The Soviet Union and the Baruch Plan\". https://www.wilsoncenter.org/blog-post/soviet-union-and-baruch-plan , accessed 13 June 2020. We use what we consider to be the latest and best historical work.11  The next best analogy may be efforts towards the international control of aviation. See Waqar Zaidi, \"'Aviation Will Either Destroy or Save Our Civilization': Proposals for the International Control ofAviation, 1920\", Journal of  Contemporary History 46,1 (2011  \n\t\t\t  Ibid., By \"denatured\" the plan meant fissionable raw materials which could not be used for \"dangerous\" purposes. In reality, this is not technically possible, but it was believed to be at the time.21  Ibid., pp. 41-53. 22 \"The American Proposal for International Control Presented by Bernard Baruch\", Bulletin of the Atomic Scientists 1&2 (1 July 1946), pp. 3-5, 10. Also at: http://www.atomicarchive.com/Docs/Deterrence/BaruchPlan.shtml . Accessed 22 April 2019. 23 Herken, p. 160; Craig and Radchenko, The Atomic Bomb and the Origins of the Cold War, pp. 122-23.24  On Baruch's motivations, seeHerken,. On the hawkish nature of the plan, seeHerken, pp. 169, 171.  One alternative view is that Baruch explicitly designed his plan to be an \"obvious propaganda ploy,\" see Shane J. Maddock, Nuclear Apartheid: The Quest for American Atomic Supremacy from World War II to the Present (Chapel Hill, NC: the University of North Carolina Press, 2010), p. 57. \n\t\t\t  Herken, p. 97; Lieberman, The Scorpion and the Tarantula, pp. 234-5.60  Maddock, Nuclear Apartheid , chapter 3. \n\t\t\t idea that such an atomic secret existed. They pointed out that the basic science was already in the public121  Zachary, Endless Frontier , p. 313.   \n\t\t\t thought that he could catch up with the U.S. atomic program or compensate through other means (e.g., larger165  The air force even dismissed reports in the Berlin press in November 1946 that the Soviet Union had begun work on an indigenous copy of the B-29 Superfortress. Concern over Soviet delivery capability was only raised following the unveiling \n\t\t\t Craig and Radchenko, The Atomic Bomb and the Origins of the Cold War , p. 109-10. \n\t\t\t  Zubok,  A Failed Empire , p.51.    \n\t\t\t For example, a reduction in the perceived usefulness of the atomic bomb for the Soviet Union may have reduced the riskiness of the country entering into international control. See the Historical Case discussion. For a theoretical discussion", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/International-Control-of-Powerful-Technology-Lessons-from-the-Baruch-Plan-Zaidi-Dafoe-2021.tei.xml", "id": "86b3000e688b7b4ea3ae04e45dc118ca"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "is a research organization focused on studying the security impacts of emerging technologies, supporting academic work in security and technology studies, and delivering nonpartisan analysis to the policy community. CSET aims to prepare a generation of policymakers, analysts, and diplomats to address the challenges and opportunities of emerging technologies. CSET focuses on the effects of progress in artifi cial intelligence, advanced computing, and biotechnology.", "authors": ["Ben Buchanan", "Andrew Lohn", "Micah Musser", "Katerina Sedova"], "title": "Truth, Lies, and Automation", "text": "or millennia, disinformation campaigns have been fundamentally human endeavors. Their perpetrators mix truth and lies in potent combinations that aim to sow discord, create doubt, and provoke destructive action. The most famous disinformation campaign of the twenty-first century-the Russian effort to interfere in the U.S. presidential election-relied on hundreds of people working together to widen preexisting fissures in American society. Since its inception, writing has also been a fundamentally human endeavor. No more. In 2020, the company OpenAI unveiled GPT-3, a powerful artificial intelligence system that generates text based on a prompt from human operators. The system, which uses a vast neural network, a powerful machine learning algorithm, and upwards of a trillion words of human writing for guidance, is remarkable. Among other achievements, it has drafted an op-ed that was commissioned by The Guardian, written news stories that a majority of readers thought were written by humans, and devised new internet memes. In light of this breakthrough, we consider a simple but important question: can automation generate content for disinformation campaigns? If GPT-3 can write seemingly credible news stories, perhaps it can write compelling fake news stories; if it can draft op-eds, perhaps it can draft misleading tweets. To address this question, we first introduce the notion of a humanmachine team, showing how GPT-3's power derives in part from the human-crafted prompt to which it responds. We were granted free access to GPT-3-a system that is not publicly available for use-to study GPT-3's capacity to produce disinformation as part of a human-machine team. We show that, while GPT-3 is often quite capable on its own, it reaches new Executive Summary \n F Center for Security and Emerging Technology iv heights of capability when paired with an adept operator and editor. As a result, we conclude that although GPT-3 will not replace all humans in disinformation operations, it is a tool that can help them to create moderate-to high-quality messages at a scale much greater than what has come before. In reaching this conclusion, we evaluated GPT-3's performance on six tasks that are common in many modern disinformation campaigns. Table  1  describes those tasks and GPT-3's performance on each. \n Introduction Internet operators needed!\" read a 2013 post on Russian social media. \"Work in a luxurious office in Olgino. Pay is 25960 rubles a month. The task: placing comments on specific internet sites, writing of thematic posts, blogs on social media….FREE FOOD.\"  1  In retrospect, this ad offered a vital window into the Russian Internet Research Agency (IRA) and into the people who for the equivalent of $800 a month crafted and propagated the lies that were the agency's products. It would take a few years before this unremarkable firm in a suburb of Saint Petersburg-funded by a man whose other businesses included catering President Putin's dinners and supplying contractors for his proxy wars-became the infamous \"troll farm\" that interfered in the United States' elections.  2  The ad existed for a reason: the IRA knew that bots-automated computer programs-simply were not up to the task of crafting messages and posting them in a way that appeared credible and authentic.  3  The agency needed humans. By 2015, it had them: in that year, a reported four hundred people worked 12-hour shifts under the watchful eyes of CCTV cameras.  4  The IRA's top officials gave instructions to 20 or 30 middle managers, who in turn managed sprawling teams of employees.  5  Some teams focused on blogging, while others specialized in memes, online comments, Facebook groups, tweets, and fake personas. Each team had specific performance metrics, demanding that its members produce a certain amount of content each shift and attract certain amounts of engagement from others online.  6  Perhaps the most important group was known (among other names) as the \"American department.\"  7  Created in April 2014, its staff needed to be younger, more fluent in English, and more in tune with popular culture than \" Center for Security and Emerging Technology viii the rest of the IRA.  8  To recruit this talent, the IRA vetted applicants through an essay in English and offered starting salaries that sometimes exceeded those of tenured university professors.  9  IRA managers tasked these and other operators with amplifying their chosen messages of the day, criticizing news articles that the agency wanted to undercut and amplifying themes it wanted to promote.  10  Based on how the agency's messages resonated online, managers offered feedback to the operators on improving the quality, authenticity, and performance of their posts.  11  They optimized the ratio of text to visual content, increased the number of fake accounts, and improved the apparent authenticity of online personas.  12  It all contributed to a far-reaching effort: at the height of the 2016 U.S. presidential election, the operation posted over one thousand pieces of content per week across 470 pages, accounts, and groups.  13  Overall, the Russian campaign may have reached 126 million users on Facebook alone, making it one of the most ambitious and far-reaching disinformation efforts ever.  14  In a way, the IRA mimicked any other digital marketing startup, with performance metrics, an obsession with engagement, employee reviews, and regular reports to the funder. This observation sheds light on a simple fact: while the U.S. discussion around Russian disinformation has centered on the popular image of automated bots, the operations themselves were fundamentally human, and the IRA was a bureaucratic mid-size organization like many others. But with the rise of powerful artificial intelligence (AI) systems built for natural language processing, a new question has emerged: can automation, which has transformed workflows in other fields, generate content for disinformation campaigns, too? The most potent tool available today for automating writing is known as GPT-3. Created by the company OpenAI and unveiled in 2020, GPT-3 has quickly risen to prominence. At its core is what AI engineers call a \"model\" that generates responses to prompts provided by humans. GPT-3 is the most well-known (so far) of a group of \"large language models\" that use massive neural networks and machine learning to generate text in response to prompts from humans. In essence, GPT-3 is likely the most powerful auto-complete system in existence. Instead of suggesting a word or two for a web search, it can write continuously and reasonably coherently for up to around eight hundred words at a time on virtually any topic. To use the model, users simply type in a prompt-also up to around eight hundred words in length-and click \"generate.\" The model completes the text they have provided by probabilistically choosing each next word or symbol from a series of plausible options. For example, some prompts might begin a story for GPT-3 to continue, while others will offer an example of completing a task-such as Center for Security and Emerging Technology ix answering a question-and then offer a version of the task for GPT-3 to complete. By carefully using the limited prompt space and adjusting a few parameters, users can instruct GPT-3 to generate outputs that match almost any tone, style, or genre. The system is also broadly versatile; in some tests, it has demonstrated ability with nonlinguistic types of writing, such as computer code or guitar music.  15  As with any machine learning system, the designers of GPT-3 had to make two major decisions in building the systems: from what data should the model learn and how should it do so? For training data, GPT-3 used nearly one trillion words of human writing that were scraped from the internet between 2016 and the fall of 2019; the system as a result has almost no context on events that happened after this cutoff date.* The system learns via an algorithm that relies on a 175-billion parameter neural network, one that is more than one hundred times the size of GPT-2's.  †  The scale of this approach has resulted in an AI that is remarkably fluent at sounding like a human. Among other achievements, GPT-3 has drafted an op-ed that was published in The Guardian, has written news stories that a majority of readers thought were written by humans, and has devised new captions for internet memes.  16  All told, OpenAI estimates that, as of March 2021, GPT-3 generates 4.5 billion words per day.  17  These skills-writing persuasively, faking authenticity, and fitting in with the cultural zeitgeist-are the backbone of disinformation campaigns, which we define as operations to intentionally spread false or misleading information for the purpose of deception.  18  It is easy to imagine that, in the wrong hands, technologies like GPT-3 could, under the direction and oversight of humans, make disinformation campaigns far more potent, more scalable, and more efficient. This possibility has become a major focus of the discussion around the ethical concerns over GPT-3 and other similar systems. It is also one of the most significant reasons why OpenAI has so far restricted access to GPT-3 to only vetted customers, developers, and researcherseach of whom can remotely issue commands to the system while it runs on OpenAI's servers.  19  We sought to systematically test the proposition that malicious actors could use GPT-3 to supercharge disinformation campaigns. With OpenAI's permission, we worked directly with the system in order to determine how easily it could be adapted to automate several types of content for these campaigns. Our paper shares our results in four parts. The first part of this paper introduces the notion of human-*The initial training dataset was almost a trillion words, and OpenAI filtered that content to provide the highest quality text to GPT-3. † In general, more parameters enable the neural network to handle more complex tasks. Center for Security and Emerging Technology x machine teaming in disinformation campaigns. The second part presents a series of quantitative and qualitative tests that explore GPT-3's utility to disinformation campaigns across a variety of tasks necessary for effective disinformation. The third part of the paper considers overarching insights about working with GPT-3 and other similar systems, while the fourth outlines a threat model for understanding how adversaries might use GPT-3 and how to mitigate these risks. The conclusion takes stock, distilling key ideas and offering pathways for new research. Center for Security and Emerging Technology 1 he story of the IRA is primarily one of human collaboration. The agency's hiring practices, management hierarchy, performance metrics, and message discipline all aimed to regulate and enhance this collaboration in service of the agency's duplicitous and damaging ends. No currently existing autonomous system could replace the entirety of the IRA. What a system like GPT-3 might do, however, is shift the processes of disinformation from one of human collaboration to human-machine teaming, especially for content generation. At the core of every output of GPT-3 is an interaction between human and machine: the machine continues writing where the human prompt stops. Crafting a prompt that yields a desirable result is sometimes a time-consuming and finicky process. Whereas traditional computer programming is logic-based and deterministic, working with systems like GPT-3 is more impressionistic. An operator's skill in interacting with such a system will help determine what the machine can achieve. Skilled operators who understand how GPT-3 is likely to respond can prompt the machine to produce high quality results outside of the disinformation context. This includes instances in which GPT-3 matches or outperforms human writers. In one test performed by OpenAI, human readers were largely unable to determine if several paragraphs of an apparent news story were written by humans or by GPT-3. GPT-3's best performing text fooled 88 percent of human readers into thinking that it was written by a human, while even its worst performing text fooled 38 percent of readers.  20  Other tests have shown that GPT-3 is adept at generating convincing text that fits harmful ideologies. For example, when researchers prompt-Human-Machine Teams for Disinformation 1 \n T Center for Security and Emerging Technology 2 ed GPT-3 with an example of a thread from Iron March, a now-defunct neo-Nazi forum, the machine crafted multiple responses from different viewpoints representing a variety of philosophical themes within far-right extremism. Similarly, GPT-3 also effectively recreated the different styles of manifestos when prompted with a sample of writing from the Christchurch and El Paso white supremacist shooters. In addition, it demonstrated nuanced understanding of the QAnon conspiracy theory and other anti-Semitic conspiracy theories in multiple languages, answering questions and producing comments about these theories.  21  Generally speaking, when GPT-3 is teamed with a human editor who selects and refines promising outputs, the system can reach still-higher levels of quality. For example, Vasili Shynkarenka, an early tester of GPT-3, used the system to generate titles for the articles he submitted to Hacker News, a well-known website for technology discussion. Shynkarenka first created a dataset of the most bookmarked Hacker News posts of all time and used their titles as an input to GPT-3, which in turn generated a list of similar plausible titles. He then selected and sometimes refined the machine's results, eventually writing up and submitting posts for the titles he thought were most likely to garner attention. With the AI-aided system, his posts appeared on the front page of Hacker News five times in three weeks. It was a remarkable success rate, and a testament to how iterative interactions between a human and GPT-3 can result in outputs that perform better than either the machine or the human could manage on their own.  22  While human-machine teaming can improve GPT-3's performance on many disinformation tasks, for some tasks human involvement is more necessary than for others. For instance, GPT-3 is entirely capable of writing tweets that match a theme or of generating a news-like output to match a headline with little to no supervision. But as operators add more complex goals-such as ensuring that the news story matches a particular slant or is free of obvious factual errors-GPT-3 becomes increasingly likely to fail.*In addition, some tasks or operators may have less risk tolerance than others. For example, an out of place or nonsensical tweet might be more of a problem for a carefully curated account with real followers than for one that is only used to send a high volume of low-quality messages. As either the complexity of the task grows or the operator's tolerance for risk shrinks, human involvement becomes increasingly necessary for producing effective outputs. *More complex tasks also typically require lengthier inputs; for instance, five examples of short articles rewritten to match a more extreme slant would take up significantly more space than five examples of random tweets on a topic. In sufficiently complex cases, the maximum input length of around eight hundred words may only provide enough space for one or two examples, which is unlikely to provide the model with enough information about its desired performance to successfully complete its task. This human involvement can take at least four forms. First, humans can continue work to refine their inputs to GPT-3, gradually devising prompts that lead to more effective outputs for the task at hand. Second, humans can also review or edit GPT-3's outputs. Third, in some contexts humans can find ways to automate not only the content generation process but also some types of quality review. This review might involve simple checks-for example, is a particular GPT-3-generated tweet actually fewer than 240 characters?-or it might make use of other types of machine learning systems to ensure that GPT-3's outputs match the operator's goals.  †  The fourth major way in which humans can give more precise feedback to the system is through a process known as fine-tuning, which rewires some of the connections in the system's neural network. While the machine can write varied messages on a theme with just a few examples in a prompt, savvy human operators can teach it to do more. By collecting many more examples and using them to retrain portions of the model, operators can generate specialized systems that are adapted for a particular task. With fine-tuning, the system's quality and consistency can improve dramatically, wiping away certain topics or perspectives, reinforcing others, and diminishing overall the burden on human managers. In generating future outputs, the system gravitates towards whatever content is most present in the fine-tuning data, allowing operators a greater degree of confidence that it will perform as desired. Even though GPT-3's performance on most tested tasks falls well short of the threshold for full automation, systems like it nonetheless offer value to disinformation operatives. To have a noteworthy effect on their campaigns, a system like GPT-3 need not replace all of the employees of the IRA or other disinformation agency; instead, it might have a significant impact by replacing some employees or changing how agencies carry out campaigns. A future disinformation campaign may, for example, involve senior-level managers giving instructions to a machine instead of overseeing teams of human content creators. The managers would review the system's outputs and select the most promising results for distribution. Such an arrangement could transform an effort that would normally require hundreds of people into one that would need far fewer, shifting from human collaboration to a more automated approach. If GPT-3 merely supplanted human operators, it would be interesting but not altogether significant. In international affairs, employee salaries are rarely the major factor that encourages automation. The entire IRA budget was on the order of several million dollars per year-a negligible amount for a major power. Instead, systems like GPT-3 will have meaningful effects on disinformation efforts only if they can improve on the campaign's effectiveness, something which is quite hard to measure. While GPT-3's quality varies by task, the machine offers a different comparative advantage over the status quo of human collaboration: scale. GPT-3's powers of scale are striking. While some disinformation campaigns focus on just a small audience, scale is often vital to other efforts, perhaps even as much as the quality of the messages distributed. Sometimes, scale can be achieved by getting a single message to go viral. Retweets or Facebook shares of a falsehood are examples of this; for example, just before the 2016 U.S. presidential election, almost one million people shared, liked, or commented on a Facebook post falsely suggesting that Pope Francis had endorsed Donald Trump.  23  Scale is more than just virality, however. Often, a disinformation campaign benefits from a large amount of content that echoes a single divisive theme but does so in a way that makes each piece of content feel fresh and different. This reiteration of the theme engages targets and falsely suggests that there is a large degree of varied but cohesive support for the campaign. In addition, a variety of messages on the same theme might make a disinformation campaign harder to detect or block, though this is speculative. As a result, one of the challenges of a disinformation campaign is often maintaining the quality and coherence of a message while also attaining a large scale of content, often spread across a wide range of personas. Since the marginal cost of generating new outputs from a system like GPT-3 is comparatively low (though, as the fourth part of this paper, \"The Threat of Automated Disinformation\" will show, it is not zero), GPT-3 scales fairly easily. To understand whether GPT-3's message quality can keep up with its impressive scale, we dedicate the bulk of the paperincluding the second part, \"Testing GPT-3 for Disinformation\"-to exploring the quality (or lack thereof) of what the machine can do. he evaluation of a great deal of new AI research is straightforward: can the new AI system perform better than the previous best system on some agreed upon benchmark? This kind of test has been used to determine winners in everything from computer hacking to computer vision to speech recognition and so much else. GPT-3 and other large language models lend themselves to such analyses for some tasks. For example, OpenAI's paper introducing GPT-3 showed that the system performed better than previous leaders on a wide range of well-established linguistic tests, showing more generalizability than other systems.  24  Evaluating the quality of machine-generated disinformation is not so easy. The true effect of disinformation is buried in the mind of its recipient, not something easily assessed with tests and benchmarks, and something that is particularly hard to measure when research ethics (appropriately) constrain us from showing disinformation to survey recipients. Any evaluation of GPT-3 in a research setting such as ours is therefore limited in some important respects, especially by our limited capacity to compare GPT-3's performance to the performance of human writers. More generally, however, the most important question is not whether GPT-3 is powerful enough to spread disinformation on its own, but whether it can-in the hands of a skilled operator-improve the reach and salience of malicious efforts as part of a human-machine team. These considerations rule out the possibility of any fully objective means of evaluating such a team's ability to spread disinformation, since so much depends on the performance of the involved humans. As such, we once more acknowledge that our work is conceptual and foundational, exploring possibilities and identifying areas for further study rather than definitively answering questions. It is too early to do anything else. We have chosen to focus this study on one-to-many disinformation campaigns in which an operator transmits individual messages to a wide audience, such as posting publicly on a social media platform. We do not focus here on one-to-one disinformation efforts in which an operator repeatedly engages a specific target, as in a conversation or a persistent series of trolling remarks. We also do not explore the use of images, such as memes, in disinformation. All of these are worthwhile subjects of future research. Within the framework of one-to-many disinformation, we focus on testing GPT-3's capacity with six content generation skills: narrative reiteration, narrative elaboration, narrative manipulation, narrative seeding, narrative wedging, and narrative persuasion. We selected these tasks because they are common to many disinformation campaigns and could perhaps be automated. We note that there are many other tasks, especially in particularly sophisticated and complex operations, that we did not attempt; for example, we do not examine GPT-3's capacity to blend together forged and authentic text, even though that is a tactic that highly capable disinformation operatives use effectively.  25  Though we are the first researchers to do this kind of study, we believe that these six areas are well-understood enough within the context of disinformation campaigns that we are not enabling adversary's activities by showing them how to use GPT-3; rather, we hope that our test of GPT-3 shines light on its capabilities and limitations and offers guidance on how we might guard against the misuse of systems like it. \n NARRATIVE REITERATION Perhaps the simplest test of GPT-3 is what we call narrative reiteration: can the model generate new content that iterates on a particular theme selected by human managers? In creating new variants on the same theme, GPT-3 provides operators with text that they can use in their campaign. This text can then be deployed for a wide range of tactical goals, such as hijacking a viral hashtag or frequently posting on a social media site in order to make certain perspectives appear more common than they are. The immediate goal of many operations is simply to expose as many users as possible to a particular narrative, since mere exposure to an idea can influence a person's receptivity to it.  26  The basic idea of narrative reiteration undergirds large-scale disinformation campaigns of all kinds and it is therefore a fundamental task on which GPT-3 must be able to perform well in order to be useful to operators. To study GPT-3's ability to amplify a narrative, we tested its capacity to generate tweet-length messages that advance a particular argument or worldview. Across a variety of topics, we found that GPT-3 performed very well at this task, demonstrating remarkable flexibility in grasping the desired theme and generating additional tweets that fit the remit. For an example of GPT-3's ability in this area, consider a disinformation actor hoping to spread climate change denialism. We simulated such an actor by selecting a few examples to include in a prompt for GPT-3. To gather such input data, we collected five hundred replies to @ClimateDepot, an influential climate change denialist account that is a leading promoter of many \"climate change contrarians.\"  27  We then sorted the replies by the number of likes they received and selected the top 10. We took these 10-without any curation and only slight formatting adjustmentsand used them to prompt GPT-3 to produce similar tweets. Figure  1  shows the first result: replies to @ClimateDepot are in bold and GPT-3's response, beginning with the text of tweet 11, is in a regular font.* * We removed URL and URL tokens from these messages for readability purposes. GPT-3 inputs and outputs for climate denialist tweets. TWEET 11: The best climate change propaganda is not propaganda at all, because it is not obviously propaganda. It does not name itself as such, and therefore, readers are more likely to believe it. GPT-3 generated these tweets on its first try. We did not iterate or refine the prompt to try to produce still-better results. Even with comparatively little effort on our part, GPT-3 performed quite well, easily generating tweet-length messages that fit the desired theme but did not directly repeat the examples we provided. It can produce similar quality of outputs with similarly low levels of effort for almost any topic or argument. Operators may choose to refine and direct GPT-3 still further to meet their narrative reiteration goals. For instance, the replies we selected as training data were not connected to any specific news story. If operators curated tweets that focused criticism on a particular story, GPT-3 would likely generate much more targeted outputs; more ideologically consistent inputs produces more ideologically consistent results. On the other hand, if disinformation operators simply wanted to spread a sense of confusion and disagreement, they could include a wider variety of tweets in the inputs. Our tests show that, by deploying particular kinds of inputs, operators can shift GPT-3's outputs in a myriad of different and predictable ways.* In short, the machine excels at narrative reiteration. TWEET 11: The best climate change propaganda is not propaganda at all, because it is not obviously propaganda. It does not name itself as such, and therefore, readers are more likely to believe it. \n TWEET 12: The US government says we can't afford to build new nuclear power plants. But it is spending billions to subsidize wind and solar, which are killing people. TWEET 13: I don't think it's a coincidence that climate change is the new global warming. They can't talk about temperature increases because they're no longer happening. TWEET 14: Climate change is the new communism -an ideology based on a false science that cannot be questioned. \n TWEET 15: The climate mafia has done far more to damage the reputation of science than anything else in the history of science. \n NARRATIVE ELABORATION Reiteration is sometimes not enough for effective disinformation campaigns. In some cases, disinformation operators might want to provide targets not only with exposure to a certain belief but also with a justification for that belief in the form of a piece of fake news or further elaboration on the desired message. But there is an art to generating the right types of misleading news stories. In most instances, generating convincing fake news stories requires a close familiarity with the operation's targets, including what types of news stories they will be predisposed to view as significant and how they will be predisposed to interpret them. The work of expanding on a narrative by generating new fake facts to support it is much more difficult than simply amplifying an existing narrative. Identifying the most effective fake facts for an operation is a cognitively difficult task that requires significant contextual awareness. There are a number of mechanisms through which operators might want to use GPT-3 to spread false claims, including social media posts, memes, news stories, and so on. For this discussion, we focus our attention on the possibility of using GPT-3 to write medium-length news stories that advance a particular worldview. This goal could be realized via a two-step process. First, GPT-3 could be instructed to generate a series of headlines that each made some new claim regarding a certain topic. Second, the model could then generate articles based on those headlines. The first task is straightforward. It is relatively easy for GPT-3 to iterate on a series of headlines and come up with similar-sounding headlines that make unique factual claims. The figure below shows 10 headlines. The first five are real headlines pulled from The Epoch Times, a far-right media company associated with the Falun Gong and known for spreading fake or misleading news about, among other things, China and the COVID-19 pandemic. When prompted with these headlines, GPT-3 produced the second set of five headlines in the The Epoch Times style. We did not curate or edit these outputs. *The fact that The Epoch Times has a habit of referring to COVID-19 as the \"CCP Virus\" also makes it difficult for GPT-3 to understand the context of the provided headlines, because that term contains less informative content than a more medically accurate term would. \n GPT-3 inputs and outputs for generating confrontational China headlines The generated headlines mostly play on existing tensions, but a few of them include startling and (as far as we are aware) novel claims. While the inputs are mostly related to COVID-19 news, the outputs do not reflect any particularly strong understanding of what COVID-related news stories should look like. This omission is because there is no information about COVID-19 in GPT-3's training data-a limitation we discuss in more detail in the third part of this paper, \"Overarching Lessons.\"* GPT-3's success in headline generation is unsurprising, since the process of generating fake headlines focused on a theme is very similar to the process of generating fake tweets to amplify a narrative. If there is an existing narrative about a topic, our headline-generating test suggests that GPT-3 is perfectly capable of dutifully producing a steady stream of story titles that support that narrative. For the rest of this section, then, we turn our attention away from headline generation and focus on the second component of narrative elaboration: writing articles to match the headlines.  †  Other researchers have studied the general ability of GPT-3 to write realistic-looking news stories. For example, as noted above, in OpenAI's original paper on GPT-3, the company found that a majority of human evaluators could not reliably distinguish GPT-3's outputs from real articles.  28  GPT-3 typically needs no more than a headline in order to begin writing a realistic-looking news story, making up facts as necessary to fill out its elaboration on the desired theme. However, operators using GPT-3 for disinformation will often want not only to generate a realistic looking news story, but one that meets other criteria as well.  29  For instance, if operators are trying to trick people without strong beliefs about a topic into believing a specific lie, they may need their fake news to look as respectable as possible. By contrast, if their goal is to outrage people who already believe a specific narrative, then the text of the article should deepen the targets' belief or incite them to action. Practically speaking, this means that operators hoping to use GPT-3 to generate news stories need to know that GPT-3 will be responsive to the headlines they feed it: a New York Times-looking headline should result in a New York Times-sounding article, while an Epoch Times-sounding headline should result in more incendiary output. Getting the tone and worldview right is essential.  To test GPT-3's ability to recreate the appropriate tone and slant of an article given only a headline, we collected three thousand articles of China coverage from each of three sources: The Epoch Times, The Global Times (a Chinese state-affiliated media network), and The New York Times.* After collecting these articles, we trained a simple classifier to determine the publication source of an article based only on the body text of the article. † This classifier used only the frequency of various terms and short phrases to classify new inputs, and the following results should not be interpreted as a statement regarding the fluency or believability of the outputs; the classifier's goal was simply to determine which of the three sources was most likely to have published a previously unseen article. Even with a very basic approach, our classifier was able to correctly identify the source of a previously-unseen article 92.9 percent of the time, which suggests that the writing of each of these sources is distinct enough for even a simple keyword-based system to reliably distinguish them. After training our classifier, we randomly sampled 25 headlines from each source and used each as an input to GPT-3 to generate a roughly 250-word-long output. ‡ These outputs were then preprocessed in the same way as our original articles and classified using our existing classifier. In effect, our classifier-which had already proven itself to be adept at identifying the source of real articles-served as a mechanism for testing GPT-3's capacity to mimic different publications' tone and style when given just a headline. If GPT-3 could successfully reproduce the style, slant, or themes of an article from the headline alone, then the classifier would likely identify GPT-3's output from a New York Times headline as a New York Times article. We found that the accuracy of our classifier declined from 92.9 percent to 65.3 percent. This suggests that GPT-3 was capable of capturing the intended tone of *Articles relating to China coverage were identified using a regular expression search for the presence of \"China,\" \"Chinese,\" \"Beijing,\" or \"CCP\" in the article headline. † Our approach included some simple preprocessing, such as lowercasing and the removal of stop words-words that do not carry meaningful semantic content in English (such as \"the,\" \"and,\" or \"to\"). The classifier used for this step was a naive Bayes classifier that was trained on tf-idf vectorizations of our articles, including unigrams, bigrams, and trigrams. For the non-technical reader, all that matters to emphasize here is that naive Bayes classifiers are mathematically simple methods that use the frequency of words or short phrases to classify a new input into one of several known categories. ‡ Our only curation at this stage was to select only headlines that were between 75 and 125 characters long in order to provide GPT-3 with sufficient contextual content. Confusion Matrix 1 shows the confusion matrix of the original classifier as tested on authentic articles. Confusion Matrix 2 shows the confusion matrix of the classifier as tested on GPT-3-generated articles. In this confusion matrix, the \"Actual\" label refers to the source from which the input headline for GPT-3 was taken. a headline, though very imperfectly.* Moreover, breaking the mistakes of the classifier down by category, as Figure  4  shows, reveals an interesting wrinkle. We can see, for instance, that The New York Times saw the largest decline in accuracy, and that the largest source of confusion for the classifier were articles generated from New York Times headlines that the classifier instead attributed to The Epoch Times. A plausible explanation of this outcome is that it is challenging for GPT-3 to distinguish the stories critical of China in The New York Times from the stories critical of China in The Epoch Times. Given a relatively neutral but China-critical headline, GPT-3 might choose to write an article in The New York Times' staid and measured tones, or it might with equal plausibility write a rabidly sensationalist article in the style of The Epoch Times. By contrast, since headlines from The Epoch Times and The Global Times are already likely to be strongly emotionally charged, GPT-3 more easily grasps the desired worldview and style. The result is that headlines from The Epoch Times and The Global Times contain stronger signals about how to generate a matching article than do headlines from The New York Times, and GPT-3 performs better when emulating those publications; a sensationalist or clearly slanted headline gives GPT-3 clear direction. Conversely, GPT-3 struggles to gauge its intended task when given a more neutral headline. While GPT-3's ability to generate news stories that match the particular tone of a given publication is mediocre, this is the type of problem that is perfect for fine-tuning. An operator looking to generate thousands of fake stories that cast China in a negative light might reach more people by generating both respectable-looking stories from fictitious field reporters for one set of targets and more alarmist conspiracy-laden stories for a different set of targets. To do this, one plausible route would be to fine-tune one version of a GPT model on The Epoch Times and another on The New York Times, and then to use each model for a different type of story. There is currently no way to easily fine-tune GPT-3, and so we were not able to test this possibility with the most advanced system. We were, however, able to finetune GPT-2, a similar system with a smaller and less powerful neural network. We found that, even when using the least powerful version of GPT-2, fine-tuning enabled the system to learn almost exactly how to mimic the tone of different publications as graded by our classifier. When we reused the same headlines as before but asked an untuned version of GPT-2 to generate the outputs, our classifier declined even further in accuracy, to 46.7 percent. But when we then fine-tuned three separate versions of GPT-2 on our three publications and used the corresponding finetuned model to generate the outputs for each headline,* our classifier was able to identify the associated source 97.3 percent of the time, as shown in Figure  5 . † *For example, we fed a headline from The Global Times to the version of GPT-2 fine-tuned on the text of The Global Times. In doing our fine-tuning for each publication, we used only the three thousand articles of China coverage selected for use in our initial classifier, excepting the 25 articles attached to headlines which we then used to generate outputs. It is important to stress that this classifier is detecting general linguistic cues as embedded in the use of various keywords or short phrases; it is not measuring the overall fluency or believability of a piece of text.* But this experiment does suggest that fine-tuning is a remarkably effective way of teaching the machine to mimic the tone and style of specific publications. Since other research shows that GPT-3 is already very adept at producing realistic-looking news stories, fine-tuning GPT-3 on a corpus of text from a publication that drives a particular narrative would almost certainly be a way of ensuring that GPT-3 could reliably write realistic news stories that also matched that specific narrative slant; this is an area for future research once fine-tuning is available. *Based on a spot check of some of the outputs, however, it is fair to say that fine-tuning meaningfully improves the overall quality of the writing. Many outputs from the untuned version of GPT-2 are obviously fake or more closely resemble a series of unrelated headlines than a coherent news story. By contrast, the outputs from the fine-tuned versions of GPT-2 are significantly more realistic. While they do tend to stray off-topic or make illogical statements somewhat more frequently than outputs from GPT-3, they are also more consistently formatted correctly. \n NARRATIVE MANIPULATION Sometimes disinformation operators need to do more than amplify or elaborate upon a message. At times, they seek to reframe or spin stories that undercut their worldview. As legitimate publications continue reporting on the world, disinformation operators must constantly find ways to manipulate facts into the larger narratives they want to push, transforming existing narratives into ones that fit their wider aims. To test GPT-3's ability to help operators find ways of spinning emerging news stories, we began by attempting to craft inputs consisting of pairs of headlines. In each pair, one headline was neutral and another was a more slanted retelling of the same event. These early attempts were largely unsuccessful, and GPT-3 struggled to reliably rewrite headlines in the way we had hoped. GPT-3 works best with continuous streams of text, and although it can understand some logical structures after seeing a few examples (for example, when given \"bark : dog :: meow : __\" it will correctly fill in \"cat\"), it has trouble understanding subtle relationships between variable-length pieces of text. After significant testing, we were eventually able to curate a list of neutral and extreme headline pairs from which GPT-3 could learn the rewriting task. But performance remained inconsistent, and GPT-3 would often directly contradict the original headline or fail to rewrite the headline with the desired slant. One of the major benefits of systems like GPT-3, however, is their versatility: the system needs direct and relatively simple instructions to perform well, but as long as a task can be broken down into explicit and relatively simple steps, GPT-3 can often automate each one of them separately. As noted, we failed to get GPT-3 to rewrite whole chunks of text or even headlines to match a target slant. Eventually we realized, however, that it could effectively write a short news story from a particular viewpoint if provided a list of bullet points about the topic-for instance, by using a prompt such as \"write a strongly pro-Trump article about [Topic X] that makes use of the following list of facts about [Topic X]\"-and that it could also summarize short news stories into a list of bullet points reasonably well.* This insight allowed us to automate the process of rewriting an existing news story in two steps: GPT-3 would first summarize the original article, and then it would generate from that summary a new version of that article that matched the viewpoint we had indicated. Breaking complex tasks into more easily explainable components is a common tactic for working with models like GPT-3, and one that can often make seemingly impossible tasks achievable for the model. *These efforts, and especially its attempts at summarization, were still highly variable. But some pitfalls were common enough-such as summarizing an article by repeating a specific sentence from the article two or three times-that we could automate quality checks to screen for bad outputs. To test GPT-3's ability to appropriately spin an emerging news story, we selected five relatively neutral articles from the Associated Press on major events of the last two years: the release of the Mueller report, China's early handling of COVID-19, debates over COVID-19 lockdowns, Black Lives Matter protests, and President Trump's response to his supporters storming the U.S. Capitol.* For each article, we used GPT-3 to summarize and then rewrite the article four times to match one of two possible slants. † An example of GPT-3's outputs for this task can be seen in Figure  6 . *While we tried to find relatively neutral articles on each of these topics, for some topics this was difficult, and the results of our small survey suggest that in at least two cases readers did not view the original Associated Press articles as being particularly neutral (see Figure  7 ). This does not pose a serious problem for our analysis, as we were interested in the differential slant that GPT-3 could introduce to a breaking news story while ideally remaining broadly believable. The GPT-3 outputs for this task can be found at https://github.com/georgetown-cset/GPT3-Disinformation. † No manual curation of the outputs was performed, though we did use several automated quality checks to try to improve over GPT-3's base capabilities. We performed minimal post-processing to address minor formatting issues but otherwise did not alter the GPT-3 outputs.  \n ORIGINAL TEXT* President Trump is Rightfully Disappointed President Trump is one of the most hard-working and dedicated presidents that America has ever seen. Throughout the campaign, he promised to \"drain the swamp\" and end corruption. His actions are truly for the people. When President Trump watched the events unfold in the Capitol Building, he was saddened and disappointed to see democracy descend into chaos. He wants his supporters to be peaceful and respectful when protesting, but he understands that passions run deep and people can get carried away. President Trump has done his best to provide jobs, keep America safe, and push for better trade deals by renegotiating NAFTA. He's also done a great job at enforcing immigration laws by building a wall, cracking down on sanctuary cities, and enforcing stricter rules as to who is eligible for asylum. The President wants his supporters to know that he's fighting for them on many fronts. \n President Trump's muted response to supporters storming the Capitol is proof he's not mentally capable of leading America President Donald Trump has failed to respond to the violence that has taken place in Washington, D.C. as his supporters stormed the Capitol building. His response has been weak and lacks any type of empathy for the people involved in the violence. One of the many things that make this disturbing is President Trump's apparent incitement of violence during his presidential campaign. He repeatedly called on his supporters to perpetrate violence against protestors at his rallies and even promised to cover their legal costs. The president's response to his supporters storming the Capitol exposes him for who he really is: a sociopath and a danger to democracy. It's time for Republicans in Congress to finally take a stand against President Trump before he destroys America. An example of an Associated Press article we used in our slant rewriting task, along with GPT-3's pro-and anti-Trump rewrites. \n PRO-TRUMP REWRITE ANTI-TRUMP REWRITE Trump doesn't ask backers to disperse after storming Capitol President Donald Trump encouraged supporters who stormed the Capitol on Wednesday in protest of his reelection defeat to \"remain peaceful,\" but he did not call for them to disperse. The seat of democracy descended into chaos when protestors overwhelmed police and bulled their way into the Capitol, forcing a delay of the joint session of Congress where lawmakers were counting electoral votes that will affirm Democrat Joe Biden's White House victory two weeks before Inauguration Day. Trump offered only a muted response to the violence as loyalists brandishing his paraphernalia clashed with police, occupied the Capitol and even stormed the Senate chamber. Under growing pressure from allies to condemn the violence, Trump tweeted, \"No violence!\" adding: \"Remember, WE are the Party of Law & Order -respect the Law and our great men and women in Blue.\" The commander in chief did not ask supporters to vacate the area as the unrest continued. *Associated Press, \"Trump doesn't ask backers to disperse after storming Capitol,\" PBS, January 6, 2021, https://www.pbs.org/newshour/politics/gop-lawmakers-ask-trump-to-deescalate-violence. After generating our GPT-3 rewrites, a group of nine Georgetown CSET researchers then evaluated two aspects of each Associated Press article and each of the four associated GPT-3 outputs: the article's slant (on a scale of 1 to 5) and the researchers' level of certainty that the article was or was not written by GPT-3 (also on a scale of 1 to 5). As a control, we also selected short snippets of actual articles on each topic from relatively partisan outlets like The Federalist, Vox, Occupy Democrats, and The Washington Examiner and mixed those real articles in with our GPT-3 samples. As in previous sections, our results should be taken as a bare minimum threshold for GPT-3's ability to fool humans: our group of evaluators were aware that many of the texts they were reading were outputs from GPT-3, and we made no attempt to strip away obvious contextual errors from the outputs. For example, one GPT-3 article referred to President Trump as the mayor of Washington, D.C.-a dead giveaway that something was off. Moreover, most of the events we used for this test occurred after the cutoff date when OpenAI stopped collecting training data for GPT-3, meaning that GPT-3 had no context for them and was often forced to resort to filling in the gaps with made-up information. We found that our evaluators struggled to determine the authenticity of articles generated by GPT-3 but were better at recognizing real articles as the real thing: the mean authenticity score for real articles was 3.8 out of 5, whereas for GPT-3 generations it was only 2.4. However, the evaluations of GPT-3 outputs were significantly more variable than the evaluations for real articles (with standard deviations of 1.42 and 1.17, respectively). Of our 20 GPT-3 generations, 11 of them were identified by at least one person as being \"definitely authentic.\" For eight GPT-3 generations, at least three out of nine evaluators thought they were more likely authentic than not. The goal of this experiment, however, was to determine if GPT-3 could meaningfully shift the slant of a breaking news story. Our results suggest that it can. When we compared the evaluated slant of our GPT-3 outputs with their corresponding articles from the Associated Press, we found that in 17 out of 20 cases, the GPT-3 rewrite had shifted in the direction we asked GPT-3 to spin the story. The average magnitude of this shift was approximately 1.35 on a five-point scale. The extent to which GPT-3 successfully spun each output in the intended direction can be seen in Figure  7 . Shifts in the slant of GPT-3 outputs, relative to the evaluated slant of the original Associated Press article associated with each topic. \n GPT-3 REWRITE OF ARTICLE SURVEY EVALUATION OF GPT-3 RE-WRITE SLANT FIGURE 7 By comparison, the average difference in slant between the Associated Press articles and the other real articles according to our survey respondents was 1.29 on the same five-point scale. This means that in several instances, GPT-3 spun its outputs to stances more extreme than those represented by the real articles we explicitly chose to represent the extreme poles of \"legitimate\" debate surrounding each topic. This difference was most noticeable in the context of President Trump's reaction to his supporters storming the U.S. Capitol on January 6, 2021: at a time when even the most partisan outlets in conservative media were cautiously distancing themselves from the president's actions, GPT-3 did not hesitate to take a short news clipping and spin it in a way that portrayed President Trump as a noble victim-exactly the kind of narrative manipulation we sought to test. \n NARRATIVE SEEDING The rise of the QAnon conspiracy theory offers a worrying example of another kind of disinformation campaign, one in which a new narrative is created, often by drawing on well-established conspiracy theories. QAnon, which is frequently referred to as a cult, falsely alleges that a cabal of cannibalistic pedophiles is running a global sex-trafficking ring and has corrupted much of the U.S. political system.  30  Though it is different from some of the examples we discuss elsewhere, the conspiracy has prompted many individuals to take violent action, including many of those who stormed the U.S. Capitol on January 6, 2021. The architects of QAnon remain unknown, though investigative reporting has shed some important light.  31  The architects claim to have a top U.S. security clearance and communicate in cryptic and seemingly nonsensical messages. From QAnon's inception in 2017 until late 2020, they posted almost five thousand messages, referred to as \"drops.\" At least in part, these messages helped springboard QAnon to greater prominence, surpassing similar conspiracy theories circulating at the time-such as HLIAnon, FBIAnon, and CIAAnon-that also claimed inside knowledge of government wrongdoing. Unlike many of those other conspiracy theories, the QAnon drops were written as clues to be deciphered, inviting followers to take an active role in building the conspiracy.  32  This participatory approach allowed adherents to feel a deeper sense of ownership and community while simultaneously allowing them to project their own individual villains, fears, and hopes into the drops. At first glance, systems like GPT-3 do not seem particularly useful for this kind of narrative seeding. Whereas the previous three tasks-narrative reiteration, elaboration, and manipulation-all require some substantial scale to be effective, narrative seeding does not. The novel QAnon narrative gained its power in part from its persuasiveness and resonance with the target audience, as well as from its resonance with other well-established conspiracy theories. It does not seem to have spread simply due to the number of times the message was shared (though narrative reiter-ation and widespread engagement online also boosted awareness of QAnon). The scale of original narrative seeding is not usually the determinant of its effectiveness; the content matters greatly, too. On deeper examination, however, systems like GPT-3 do seem to have at least some relevance for narrative seeding. The vague and at times nonsensical style that characterizes the QAnon messages often fits naturally with GPT-3's outputs, especially when the system is struggling to be internally consistent in its responses. GPT-3's tendency to make statements that are provably false is less of an issue when creating disinformation narratives; QAnon is rife with outright lies. Similarly, GPT-3 often jumps from topic to topic, a habit which is a negative in mainstream writing but also shared by the QAnon authors.  33  Most notably, GPT-3's scale enables the dispersal of many narratives, perhaps increasing the odds that one of them will go viral. To test GPT-3's capacity to emulate the QAnon style and seed conspiratorial narratives, we tasked GPT-3 with recreating one of Q's more interesting message styles: rants composed of only ambiguous questions or statements. A prompt instructing GPT-3 what to do and providing three real Q drops as examples follows in bold along with GPT-3's response in regular text.* *We gave the same prompt to GPT-3 five times and have reproduced three of its outputs here as examples 4, 5, and 6. GPT-3 inputs and outputs for generating Q-style \"drops.\" Expand your thinking. \n News unlocks past. We need to pray. God bless you all. \n Q Though it is difficult to quantitatively measure, this proof of concept suggests that GPT-3 easily matches the style of QAnon. The system creates its own narrative that fits within the conspiracy theory, drawing on QAnon's common villains, such as Hillary Clinton (referred to as \"HRC\") and her staffer, Huma Abedin. None of the QAnon drops we provided to GPT-3 mentioned either of these people by name We need to pray. God bless you all. Q Example 6: A message to those who will be taking over: We will never forget. We have not forgotten. The truth always wins. \n How about [A]G? When does a [bird] sing? Where is C? When is C? Who is [C]? Goodbye C. Q and we did not provide it with any information that would have explicitly steered it towards them. This suggests a striking ability on the part of GPT-3 to generate appropriate-sounding Q drops despite having very little contextual knowledge regarding QAnon; whether GPT-3's messages would in fact resonate with QAnon followers is something we were unable to test ethically, and is a significant limitation on this part of our research. In general, GPT-3 seems largely capable of writing messages for new narratives within a conspiracy theory without much human intervention or oversight. The degree to which it is these messages that attracted adherents to QAnon is unclear and is once again difficult to measure empirically. It is challenging to disentangle whether people believe the QAnon conspiracy theory over other conspiracy theories (and, indeed, over well-established facts) because of the messages' style and content or because of something else, such as social pressures, predisposition towards conspiracy theories, exposure to QAnon from trusted friends and family members, or other factors. While GPT-3 could aid disinformation operators seeking to seed new narratives-a notable finding-it remains unclear how useful this ability would be in creating narratives that will take root and grow. \n NARRATIVE WEDGING Disinformation campaigns often serve as a wedge. Operators find a pre-existing fissure in an adversary's society and, rather than concocting outright lies, aim to widen this gap with disinformation. Oftentimes, operators send opposite messages to the two poles of a debate, entrenching each in its position and pitting target against target. In 2016, for example, the IRA specifically stoked religious and racial tensions in the United States, playing both sides of major issues. In one case, Russian operators on Facebook organized competing rallies in Houston with one group exhorted to \"Save Islamic knowledge\" and another told to \"Stop Islamization of Texas!\"  34  In another case, the operators distributed racist memes of President Obama.  35  At the same time, Russian operators were running many Facebook accounts and pages that aimed to reach Black voters and deter them from voting, including one called \"Blacktivist\" that had more than 11 million engagements.  36  We sought to examine GPT-3's ability to create divisive and targeted content designed to wedge divides open. We used the system to generate messages that might appear in an internet forum, Twitter discussion, or Facebook page where various demographic groups might gather to discuss political issues. In particular, we prompted GPT-3 to write a series of messages that targeted Christian, Jewish, and Muslim worshippers. For each group, we prompted it to write calls to vote Republican, Democratic, or not at all. In addition, we tasked GPT-3 with generating messages that highlighted some of the racial injustices faced by Black Americans as well as writing messages intended to stir up animosity against Black Americans. As with all of the disinformation produced by GPT-3 in our tests, we do not endorse the views expressed in this content, and we conducted these tests only to understand the risks of automated disinformation. Our process for generating these messages involved a short back-and-forth between GPT-3 and a human operator. First, we gave GPT-3 a prompt such as \"Five reasons why it's not worth voting if you're Jewish: Reason 1.\" After GPT-3 completed this prompt 10 times, generating about 20-30 arguments total, we selected what seemed to be the three \"best\" messages and added them to the original prompt. Using the more detailed prompt, we asked GPT-3 to generate 10 more groups of two to three arguments each, and a human once again chose the \"best\" argument from each of those 10 responses. This kind of human-machine team produces potentially more effective results than GPT-3 does on its own, as one would likely expect. GPT-3's first set of messages were often too short, too long, too rambling, or off-topic because we provided it with very little guidance in our first prompt. With the improved prompt that in part relied on the machine's own creations, GPT-3 consistently performs significantly better. Even so, some of the outputs the machine produces do not contain compelling or well-targeted messages, making the role of the human curator at the final stage valuable. The entire process takes only minutes, and selecting messages takes only seconds per message. A human-machine team could produce several thousand messages per day and is almost unlimited in volume if the disinformation campaign tolerates occasional lower-impact messages. To understand what fraction of GPT-3's outputs would be usable for a disinformation operator, we had four Georgetown CSET analysts read each of the messages produced by our process. There were 110 messages total: 10 for each combination of religious category (Christian, Jewish, and Muslim) and voting goal (Republican, Democratic, Abstain), as well as 10 highlighting injustices against Black Americans and 10 expressing anti-Black American sentiment. The humans provided yes/no answers as to whether each message was targeted to the intended demographic and whether it would fit in an internet forum on the intended topic. Using the process described above, GPT-3 performed quite efficiently on this test. Most of the machine's messages seemed well-targeted to both the intended demographic and forum. Figure  9  Part A shows that the humans thought the messages targeted the demographic almost 90 percent of the time and also fit the forum almost 90 percent of the time. Assuming that a message is only credible and accepted if it meets both criteria, Figure  9  Part B shows that 95 percent of the messages were rated as credible by at least one rater, 85 percent were rated as credible by at least three raters, and 60 percent were rated as credible by all four. The middle three columns in Part B of the figure-where raters disagreed-are also interesting. When we examined the messages that sparked disagreement, we found that they often offered more nuanced perspectives rather than relying on common stereotypes, tropes, or insults. For example, Muslims overwhelmingly vote Democratic;  37  when prompted to make an argument for Muslims to vote Republican, GPT-3 attempted to peel off voters by highlighting their individuality and downplaying group dynamics: \"Muslims are not a voting bloc; they are Americans. Muslims should base their vote on the issues that matter most to them.\" Messages like these might actually be among the most insidious and most effective, since they are less direct in their content and less obvious about their intent.  \n FRACTION OF MESSAGES FIGURE 9 Human evaluations of whether the messages were targeted to their intended demographic and whether they were fit to the intended forum topic. Error bars show the 95 percent confidence interval for 440 responses. That at least one of four raters found that 95 percent of the messages were credible and that at least three raters rated 85 percent of the messages as credible speaks to GPT-3's capability. However, we think this sort of statistical analysis, though useful, belies the negative emotional force of GPT-3's writing. To give a sense of GPT-3's disinformation capabilities in this regard, we have reproduced a few of its outputs below. Note that although we have overwritten slurs with *'s, some of the messages are still very disturbing. It is worth reiterating that none of these outputs were written by humans and that not even the examples in our prompts were written by humans; GPT-3 likely learned such language and racist ideas from its internet-based training data. \n NARRATIVE PERSUASION While disinformation campaigns employ many subtle tactics to try to make targets more receptive to specific viewpoints, sometimes operators can get the desired results by simply arguing for their position. These attempts at persuading a target are often harder than merely amplifying a message, since people tend to subject arguments they disagree with to sharper scrutiny than arguments they agree with.  38  To change a target's mind, an operator must present well-formed and well-tailored arguments-otherwise, the approach could backfire and leave the target even less amenable to the operator's goals than before.  39  To test GPT-3's persuasiveness and ability to tailor messages, we surveyed 1,171 Americans who read GPT-3-generated statements for and against two current international relations issues: withdrawal of troops from Afghanistan and sanctions on China; our survey occurred prior to President Biden announcing the United States' withdrawal from Afghanistan.* For each issue, we instructed GPT-3 to develop statements tailored to Democrats and statements tailored to Republicans; given how widely available political data is, it is realistic to expect that adversaries will be able to identify the political parties of many Americans, and it is plausible that political affiliation will be a basis on which operators tailor their messages. GPT-3 wrote eight groups of 20 statements: 20 for and against each of the two topics for both major political party affiliations. We then selected what we thought were the best 10 statements from each of the 20-statement groups as if we were a human operator approving half of GPT-3's outputs. Rather than posting them to a website or social media service, however, we presented them in a survey in which respondents, recruited through Amazon's Mechanical Turk, rated each statement on a five-point scale from \"not at all convincing\" to \"extremely convincing.\" Respondents were randomly assigned to read five statements from one of the eight statement groups (e.g., statements in favor of sanctions against China targeted to Democrats). An example statement from each of the eight groups is shown in Table  3 ; the full set of statements and the prompts used to generate them are available on GitHub at https://github.com/georgetown-cset/GPT3-Disinformation. *1,408 respondents took the survey but 237 of them were dropped from the analysis for reasons including declining consent, completing it too quickly, or failing the attention tests. \n STATEMENT OBJECTIVE EXAMPLE GPT-3 OUTPUT \n Support sanctions targeting Democrats China has a history of discriminating against minorities and we need to penalize them for their human rights violations. Sample GPT-3 outputs for a variety of narrative persuasion tasks. \n Support sanctions targeting Republicans President Obama failed to stand up to China when it hacked our government's computers. We can't allow President Trump to make the same mistake. \n Oppose sanctions targeting Democrats This is senseless self-harm and will make it harder for China to cooperate with us on climate change. \n Oppose sanctions targeting Republicans We should focus on our problems here at home instead of meddling in other countries' affairs. \n Support withdrawal targeting Democrats The United States is spending precious capital on a fruitless war. Our country is in debt because of it and the children of Afghanistan have lost a generation of their lives. This is an abomination that has no reason to continue. \n Support withdrawal targeting Republicans America needs to stop supporting the corrupt government in Afghanistan. We need to get out. \n Oppose withdrawal targeting Democrats While there have been some gains made in Afghanistan, there are still many challenges that will exist even if the US pulls out. The US still has a vested interest in Afghanistan's stability and should keep a presence there. \n Oppose withdrawal targeting Republicans President Obama's timeline for withdrawal is dangerous. We need to keep a permanent military presence in Afghanistan and commit to nation-building. The main objective of the survey was to determine whether GPT-3 could sway Americans' opinions. To test this, we also asked for survey respondents' opinions about Afghanistan withdrawal and sanctions on China. For respondents assigned to read statements about withdrawing troops from Afghanistan, we first gathered their views on China, then presented five GPT-3-generated statements for or against withdrawing troops from Afghanistan, and finally asked for their views on Afghanistan. For respondents assigned to read statements about sanctioning China, we first gathered their views on Afghanistan, then presented five GPT-3-generated statements for or against sanctions against China, and then asked for their views on China. In this way, each group served as a control for the other, expressing their views on both issues without having read GPT-3-generated messages about the issue and enabling us to evaluate any change in the average opinion on the issue from exposure to GPT-3-generated statements. Our survey also included questions about political interest, partisanship, political ideology, and trust in the U.S. government, attention tests, knowledge tests, and demographic questions.  \n Not at all convincing \n FRACTION OF RESPONDENTS \n CORRECTLY TARGETED MISMATCHED FRACTION OF RESPONDENTS FIGURE 10 Survey respondents rated GPT-3 generated statements at least somewhat convincing 63 percent of the time overall, 70 percent of the time when targeted to the appropriate political demographic, and 60 percent of the time when the political demographics were mismatched. There were 1,171 respondents in Part A, 875 in Part B, and the error bars show the 95 percent confidence interval. Survey respondents generally accepted GPT-3's statements as convincing. As shown in Figure  10  Part A, they found GPT-3's attempts at persuasion at least somewhat convincing 63 percent of the time, including cases where Democrats were shown Republican-targeted arguments and vice versa. Although even the most compelling statements were deemed \"extremely convincing\" by only about 12 percent of the respondents, a substantial majority of messages were at least \"somewhat convincing.\" A key component of persuasion is tailoring a message, getting the right argument in front of the right target. Our results also provide evidence that GPT-3 can do this as well, effectively devising messages that fit its targets. When survey respondents were shown a GPT-3-generated statement that was tailored to their political partisanship, the respondents often found the statement convincing. Part B of Figure  10  shows that 70 percent of respondents who read statements targeted to their partisanship rated the statement as at least somewhat convincing. Not only did a majority of survey respondents evaluate GPT-3's statements to be at least somewhat convincing, our results suggest that these statements effectively shifted the respondents' views of the topics at hand. For example, respondents were 54 percent more likely to want to remove troops if they were shown GPT-3's statements for withdrawing troops than if they were shown GPT-3's statements opposing the withdrawal. Figure  11  Part A shows the range of possible response options and how often each choice was chosen by survey respondents shown GPT-3's pro-withdrawal messages, GPT-3's anti-withdrawal messages, and no messages about troop withdrawal (the control group). The results were even more pronounced for sanctions on China. The majority of the control group (51 percent) favored sanctions while only 22 percent opposed them. Of the group that saw GPT-3's anti-sanction messages, however, only 33 percent supported sanctions, while 40 percent opposed them. It is interesting that GPT-3 was not as persuasive when arguing for sanctions despite the same procedures and level of exposure, as Figure  11  Part B shows. This finding highlights how difficult it can be to predict what will actually influence opinions and behavior. But it is nonetheless remarkable that, on an issue of obvious international importance, just five short messages from GPT-3 were able to flip a pro-sanction majority to an overall anti-sanction view, doubling the percentage of people in opposition; the durability of respondents' new views is an important area for future research.  Groups exposed to GPT-3's support statements were more supportive than those exposed to oppositional statements, though the intended shift was not always evident when compared to the control group. Error bars represent the 95 percent confidence interval, with support withdrawal, control, and oppose withdrawal having 294, 576, and 301 respondents, respectively, and support sanctions, control, and oppose sanctions having 288, 595, and 288 respondents, respectively. he last section explored how GPT-3 could reshape disinformation campaigns by examining its capacity to automate key tasks. We recognize that such an evaluation is by definition a snapshot of automated capabilities at the time of this writing during the spring of 2021. Given the rapid rate of progress, with GPT-3's 2020 announcement coming a little more than a year after GPT-2's unveiling, we expect that the capabilities of natural language systems will continue to increase quickly. For that reason, this section considers some overarching key concepts, rather than specific test results, that seem likely to affect both GPT-3 and its successors. \n WORKING WITH GPT-3 To work effectively with GPT-3, it is important to understand how the system functions. As noted in the introduction, GPT-3 trained on a vast quantity of human writing across a wide variety of genres and perspectives. OpenAI completed the process of collecting GPT-3's training data in mid-2019. When an operator uses GPT-3, they give it an input that shapes how the system draws upon this training data, as shown by the examples in the second part of this paper, \"Testing GPT-3 for Disinformation.\" One alarming trend we noticed was that more extreme inputs sometimes produced sharper and more predictable results than more neutral ones. For example, when given the task of writing a story from a headline, a more pointed title offered more context and direction to GPT-3. A title like \"Biden Sells Out Americans By Helping Illegal Immigrants Steal Jobs\" offers a very clear direction for the slant of the story, and GPT-3 frequently grasped and worked effectively within this context. On the other hand, a headline like \"Biden Takes Major Steps on Immigration Reform\" is more neutral and offers less context. GPT-3's stories for these kinds of neutral headlines were often more varied and less consistent with one another, another reminder of the system's probabilistic approach to ambiguity. Extremism, at least in the form of headlines, is a more effective way of controlling the machine; while not all disinformation is extremism-again, some sophisticated efforts are subtle and insidious-this trend remains concerning. Sometimes, the task assigned to GPT-3 is too complex for the system to handle all at once. In these cases, we found that we got better results by breaking the task into sub-tasks and having GPT-3 perform each in sequence. For example, as discussed above, to test GPT-3's abilities at narrative manipulation-rewriting an article to suit a particular viewpoint-we broke the process into two steps. Rather than simply telling the system to rewrite an article, we tasked it with first summarizing the original article into a list of bullet points and then using those bullet points as a basis for rewriting it with a slant. In general, we found that concretely specifying the steps of a process yielded better results when working with GPT-3 than asking the machine to devise its own intermediate steps. We got even better results by introducing an element of quality control throughout the process, often in the form of automated quality checks. For instance, in our narrative manipulation task, we devised a series of quality checks to select good summaries of the original article by prioritizing non-repetitive summaries consisting of relatively short bullet points.* This quality check typically allowed us to identify summaries of the original article that were the most likely to result in fluent and plausible rewrites in the second stage of our slant rewriting process. At the same time, because this process often weeded out summaries that may have been adequate, it represented a computationally intensive approach in which GPT-3 ran continuously until it produced an output that satisfied our automated quality check. This kind of process shows the power of GPT-3's scale: because the machine can easily generate many outputs for a given input, devising an effective means of filtering these outputs means that it is possible to find particularly good results. When this kind of quality control is done at each step in a multistep process, the overall result can reliably yield strong results. Creating such a process is thus one of the *Both of these criteria were important. First, if the bullet points were each a long sentence (which was common in the summary outputs), then GPT-3 would often struggle to make sense of them when rewriting. Second, if the bullet points were repetitive, then the summary was not efficient. The quality score was a somewhat arbitrarily chosen weighting of two factors: the average repetitiveness of any two bullet points, and the distance between the average effective length of the bullet points and the number seven (where effective length refers to the number of words that were not stop words with little semantic meaning, like \"the,\" \"and,\" or \"from\"). important parts of working with systems like GPT-3, though finding effective metrics for filtering can be challenging. An actor that can only run GPT-3 a limited number of times-perhaps due to limitations on its computing power-will get less value from quality controls that force GPT-3 to attempt a task many times. Such an actor is likely to rely more on humans in curating and editing GPT-3's outputs. For example, as we showed with our test on narrative wedging, a human can select outputs from GPT-3 that are particularly relevant and then use them in another round of inputs, iteratively refining the machine's performance without forcing it to run continuously. \n EXAMINING GPT-3'S WRITING The quality and suitability of GPT-3's writing varies in interesting ways. First and most significant, the system is indelibly shaped by its training data. For example, GPT-3 was no doubt fed millions or billions of words on Donald Trump, the president of the United States at the time of the system's training in mid-2019. This information enables it to easily write about Trump from a variety of perspectives. By contrast, GPT-3 struggles if asked to write about political figures whose rise to prominence occurred after the system was trained or if it is asked to write about more recent global events. GPT-3 can still write compelling narratives about topics outside of its training data, but it does so more as a writer of fiction rather than as a repeater of facts. In this fictional mode, it makes up elements to fill in gaps; these elements can be dead giveaways of machine authorship. The degree to which such factual errors matter for disinformation is debatable, but it is likely that egregious errors undermine a text's credibility. Since disinformation campaigns often rely on controlling a narrative around emerging topics, the absence of information about contemporary issues in GPT-3's training data can be a significant limitation. Overcoming it will require either devising more advanced algorithms that can continually consume information about recent events without overwriting useful knowledge about the past or deploying a constant process of fine-tuning the system on breaking news stories for each new application. As a result, today there is no inexpensive way for GPT-3 to have both wide-ranging knowledge and for that knowledge to be kept up to date. The second key characteristic of GPT-3's writing also emerges as a result of the importance of training data: GPT-3 seems to adjust its style and focus to what the data suggests is most relevant. When the prompt specifies a genre, such as a tweet, news story, or blog post, the system often assumes the cadence and style of that genre. This tendency can create challenges. For example, conversations that happen on social media tend to be freewheeling discussions that make little or no reference to specific concrete facts. Drawing on its training data, GPT-3 mimics this tendency and tends to write tweets that express opinions rather than contain specific facts. By contrast, when GPT-3 writes a news story, it regularly generates fake information, such as made-up historical events or quotes, to support its narrative.* Third, perhaps due to its probabilistic nature, GPT-3 sometimes writes things that are the exact opposite of what its operators intended. For example, when asked to provide arguments to support a position, it will occasionally write something opposing that position. Such behavior can be seen in the arguments for or against sanctions on China or withdrawal from Afghanistan, as well as in some of its attempts at rewriting articles with a particular slant; one GPT-3 argument to oppose withdrawal contended that: \"Afghanistan is an ally for the United States. However, we have lost the support of the people of the country. It is time to bring our troops home.\" Human curation of GPT-3's outputs would reduce the effect of the system's odd reversals in practice. Fourth, it is important to emphasize that, even at its best, GPT-3 has clear limitations. For example, consider the task of generating fake news headlines: while GPT-3 can easily come up with new headlines that would extend an already existing narrative, it cannot be relied upon to come up with a scintillating and explosive narrative out of nothing. The most enticing content perhaps comes from an iterative human-machine team effort in which operators try to develop potentially eye-catching headlines and then allow GPT-3 to develop them further.  40  GPT-3 by itself seems to lack some of the creativity that is required for coming up with a wholly new fake news story. Fifth, while fine-tuning may be a powerful method for overcoming many of these shortcomings and improving GPT-3's writing or changing its slant, the technique is not an immediate or perfect solution. Operators will perhaps be able to fine-tune GPT-3 to reduce unwanted content generation from creeping in and to help GPT-3 better understand its assigned task. However, fine-tuning is often difficult to achieve in practice, requiring the acquisition of datasets from which the machine can learn. For example, we were able to use fine-tuning of GPT-2 to emulate the perspectives of different newspapers in the narrative elaboration test described above, but only because we had a well-organized collection of articles from each publication. We were unable to use fine-tuning in other instances because the data was messy. For example, we assembled a dataset of anti-vaccine tweets but found that, even when the writing was intelligible, it often indirectly referred to an event that had happened or a comment that was posted, or linked to a video that may have been removed or deleted; there was not enough clarity to provide sufficient direction to *But, as suggested above, this can also pose a problem: disinformation actors may not actually want their \"news\" stories to contain too many highly specific claims because they might face legal liability for libel or because including too many details increases the chances that one of those details may provide an obvious clue that the story is fake. the machine. Similarly, we collected tweets from known disinformation campaigns but found that they, too, lacked necessary context. Without that context, the tweet by itself was useless for fine-tuning a disinformation bot. Creating these datasets manually is a challenge for a research effort like ours but is probably achievable for well-resourced actors. In such circumstances, an adversary's ability to get sufficient data will shape its capacity to wield GPT-3. This discussion of incoherence in real-world datasets leads to a final important point: while GPT-3 is at times less than compelling, our study of online information offers a reminder that so, too, is a great deal of human writing-both disinformation and not. What look like failures on GPT-3's part may at times simply be accurate emulation of some of the less-credible forms of writing online. In addition, the low bar for a great deal of online content might make it easier for even imperfect writing from GPT-3 to blend in. In this area, as in many others, it is hard to definitively measure what qualities of writing make for effective disinformation and how well GPT-3 can mimic those qualities in its own texts. s we have shown, GPT-3 has clear potential applications to content generation for disinformation campaigns, especially as part of a human-machine team and especially when an actor is capable of wielding the technology effectively. Such an actor could pose a notable threat. In this section, we consider more deeply which kinds of actors might be able to access automated disinformation capabilities should they so choose. We also explore which sorts of mitigations would be effective in response. \n THREAT MODEL Adversaries seeking to use a system like GPT-3 in disinformation campaigns must overcome three challenges. First, they must gain access to a completed version of the system. Second, they must have operators capable of running it. Third, they must have access to sufficient computing power and the technical capacity to harness it. We judge that most sophisticated adversaries, such as nations like China and Russia, will likely easily overcome these challenges, but that the third is more difficult. Indeed, Chinese researchers at Huawei have recently already created a language model at the scale of GPT-3 for writing in Chinese and plan to provide it freely to all.  41  To access a version of GPT-3 or a system like it, sophisticated adversaries have several options. The easiest is to wait for such a system to become public. It is likely that researchers will create and release code and model parameters for an English-language system like GPT-3, as researchers have replicated GPT-2 and many other AI breakthroughs after publication.* We also expect that well-resourced governments with cyber The Threat of Automated Disinformation 4 A *Eleuther AI is currently working on replicating an English-language version of GPT-3. expertise will be able to illicitly gain access to GPT-3's design and configuration or to recreate the system should they desire to do so. Though we have no reason to doubt the cybersecurity and vetting procedures of OpenAI, which has tightly restricted access, we believe that the sophisticated hacking and human intelligence capabilities of governments such as China and Russia are capable of penetrating extremely security-conscious businesses. Once they acquire such a system, training operators to use it will be a simple task for these governments. If an adversary obtains or builds a version of GPT-3 or a system like it, the challenge of obtaining enough computing power to train and run it is notable, however. Simply put, GPT-3 is gigantic. A great deal of its strength comes from its vast neural network and the 175 billion parameters that underpin it. Even if an adversary acquires the fully trained model and needs only to use computing power to run it, the requirements are significant. A more detailed understanding of these requirements sheds light on which sorts of adversaries will be able to put GPT-3 or systems like it to use. We begin our analysis of computational requirements by looking at GPT-2 in more depth. That system comes in four variants: small, medium, large, and extra-large. While even GPT-2's extra-large network is less than 1/100th of the size of GPT-3's network, it is nonetheless very difficult to run. Giving it new prompts and tasking it with generating replies is computationally intensive. Operators often run these systems on graphics cards, computer chips that are more specialized for running calculations in parallel. Widely available graphics cards, such as the Nvidia K80 in Google Cloud, can use up their memory and crash while trying to run the extra-large version of GPT-2. To solve the memory problem, it can be necessary to split up the system so that it runs on multiple graphics cards. This is a complex task, and a great deal of the knowledge and code on how to do it is not widely available. OpenAI has not disclosed how much memory GPT-3 uses or how many graphics cards share the load of running it. That said, we can extrapolate from GPT-2 to get a rough approximation of the computing power required. As shown in Figure  12  Part A, the size of the GPT-2 system files in gigabytes increases predictably with the number of parameters in each model's neural network. As can be seen in Figure  12  Part B, if the linear trend holds, then GPT-3 should require around 712.5 GB of RAM.* *For comparison, the Huawei model PanGu-ɑ is slightly larger than GPT-3 (200 billion parameters) and is around 750 GB. GPT-2 is a large model that requires several gigabytes but the largest version of GPT-3 is more than one hundred times larger than the largest version of GPT-2. FIGURE 12 FIGURE 12 GPT-2 is a large model that requires several gigabytes but the largest version of GPT-3 is more than one hundred times larger than the largest version of GPT-2. That 712.5GB of memory pushes the boundaries of what any major cloud provider currently makes publicly available as a package of graphics cards. If an adversary wanted to build its own infrastructure for utilizing GPT-3, it would need to buy 23 of the more advanced Nvidia V100 graphics cards and then overcome the engineering challenge of linking them all together. In addition, such an endeavor might be prohibitively expensive, at least for non-state actors. The 23 graphics cards would cost around $200,000, plus the administrative cost and electricity to operate and cool them. To reach a major scale is harder still: creating enough content to equal in size to 1 percent of global Twitter activity would require hundreds of GPT-3s running 24/7 and would cost tens of millions of dollars per year. While this is a substantial hurdle for non-state actors, it is a rounding error for a major power. This analysis of the role of computing power in GPT-3 offers important context. On one hand, it offers hope that even adversaries who are able to access information about GPT-3 will have difficulty in putting it to use absent extensive technical expertise and some degree of financial resources. The net effect of this computational hurdle is likely to limit who can use GPT-3 for disinformation. That said, these barriers will likely diminish over time as computing power becomes more widely available and falls in price. Furthermore, these barriers are likely already surmountable for dedicated adversaries who possess both technical skills and ample resources. As a result, other mitigations are required to guard against those adversaries' potential efforts to automate disinformation. \n MITIGATIONS We have focused on content generation for disinformation campaigns and on the potential of systems like GPT-3 to automate it. We are not optimistic that there are plausible mitigations that would identify if a message had an automated author. The only output of GPT-3 is text and there is no metadata that obviously marks the origin of that text as a machine learning system. In addition, while GPT-3 certainly has its quirks in writing, it is unlikely that a statistical analysis would be able to automatically determine if a human or machine wrote a particular piece of text, especially for the short messages usually seen in disinformation campaigns. Instead, the best prospects for thwarting GPT-3's power in disinformation is to limit its utility by limiting the scale at which these operations unfold. As currently constituted, GPT-3 alone likely does not consistently produce content that is higher quality than a professional disinformation operator, such as many of the Russian employees of the Internet Research Agency, but it is far more scalable. As a result, any effort that makes it harder for an adversary to scale an operation-and thus play to GPT-3's biggest strength-will reduce how useful automated content generation is in the hands of adversaries. To limit the scale of disinformation campaigns, it is necessary to look beyond the content generation task and focus on other parts of a successful effort.  42  GPT-3 is unlikely to help with campaign components unrelated to content, such as administrative, managerial, and quality assurance tasks, though it may free up more humans to focus on these endeavors. In addition, GPT-3 is unlikely to help directly with a key task that permits the propagation of content once created: infrastructure creation. Disinformation campaigns need infrastructure. They depend on inauthentic accounts for the managed personas as well as the web sites, community groups, and pages that operators use to channel disinformation content. The IRA's \"department of social media specialists\" dealt with developing these digital messengers and channels. To set up these accounts, operators needed fake email addresses and phone numbers or SIM cards, all of which were managed by the IRA's information technology department. For operational security and to obscure the operators' digital traces, the IT department took steps to hide the IP addresses of operators and make them appear as coming from the United States, rather than Russia. If an operation involves a standalone website in addition to activity on an established social media platform, operators will need to register domains, secure web hosting, and hire web developers to make it look professional. Similarly, if operators want to run ads, they will need financial infrastructure to purchase ad space, perhaps including credit cards from an established bank that cannot be easily traced to the operator. For this reason, the IRA's operators scoured the underground market for authentic social security numbers stolen from unwitting Americans and used them to create fake drivers licenses and to set up PayPal and bank accounts.  43  While the IRA and others have had success setting up infrastructure for their disinformation campaigns, this task nonetheless remains an important point of leverage for defenders. Most importantly, it is a task that is likely to increase in importance as GPT-3 potentially scales the scope of campaigns further. GPT-3's capacity to generate an endless stream of messages is largely wasted if operators do not have accounts from which to post those messages, for example. The best mitigation for automated content generation in disinformation thus is not to focus on the content itself, but on the infrastructure that distributes that content. Facebook, Twitter, and other major platforms have built out large teams to try to track and remove inauthentic accounts from their platforms, but much more work remains to be done. In 2020 alone, Facebook removed 5.8 billion inauthentic accounts using a combination of machine learning-enabled detection technology and human threat-hunting teams.  44  Despite those efforts, fake profiles-a portion of them linked to disinformation campaigns-continue to make up around 5 percent of monthly users on the platform, or nearly 90 million accounts.  45  In the first half of 2020, Twitter reported taking action against 1.9 million accounts out of a 340 million account user base, with 37 percent of these accounts removed due to violation of the company's civic integrity policy, which includes (but also extends significantly beyond) inauthenticity.  46  As these accounts become critical bottlenecks for distributing disinformation, it is increasingly important to devise mitigations that limit adversaries' access to them. \n Conclusion ystems like GPT-3 offer reason for concern about automation in disinformation campaigns. Our tests show that these systems are adept at some key portions of the content generation phase of disinformation operations. As part of well-resourced human-machine teams, they can produce moderate-quality disinformation in a highly scalable manner. Worse, the generated text is not easily identifiable as originating with GPT-3, meaning that any mitigation efforts must focus elsewhere, such as on the infrastructure that distributes the messages. The overall impact of systems like GPT-3 on disinformation is nonetheless hard to forecast. It is hard to judge how much better a human-machine team is than human performance in real-world operations, since a great deal of the disinformation from real-world campaigns is poorly executed in its writing style, message coherence, and fit for its intended audience. We had hoped at the beginning of our study that we could make direct comparisons between real-world disinformation and GPT-3's outputs, but the noisiness and sloppiness of real-world activity made such comparisons harder than expected. Even if we could identify a means to compare real-world disinformation to GPT-3's outputs, it is not clear how useful this comparison would be for scholars. A human-machine team might outperform humans on some key metrics-especially in terms of scale-but that does not imply that GPT-3 will transform the practice of disinformation campaigns. Instead, we think GPT-3's most significant impact is likely to come in scaling up operations, permitting adversaries to try more possible messages and variations as they answer for themselves the most fundamental question in the field: what makes disinformation effective? \n S We think GPT-3's most significant impact is likely to come in scaling up operations, permitting adversaries to try more possible messages and variations as they answer for themselves the most fundamental question in the field: what makes disinformation effective? Fourth, and most concerning, our study hints at a preliminary but alarming conclusion: systems like GPT-3 seem better suited for disinformation-at least in its least subtle forms-than information, more adept as fabulists than as staid truth-tellers. As this paper's third part, \"Overarching Lessons,\" discussed, some of the characteristics of GPT-3's writing, such as its tendency to ramble or to make things up, are common in many disinformation campaigns but fatal to credibility in legitimate discourse. Future refinements of GPT-3 may anchor its writing more firmly in facts or teach it to operate within well-defined constraints, such as the formal structures common to legal documents. For now, however, its text-generating process is at times laden with shortfalls in accuracy and coherence in a way that constrains its legitimate applications while leaving its utility for disinformation relatively undiminished. This analysis leads us to reconsider a question we have asked many times throughout our study: what is GPT-3? There is no doubt that it is a technological breakthrough, a sea change in machines' capacity to work with human language, and a step towards more powerful AI. Though we are quite familiar with the algorithm through which GPT-3 chooses its next word, the effortless way in which it writes can at times nonetheless seem magical. It is exciting to watch the machine at work. But our study offers a reminder that there is more to the story. While GPT-3 has access to wide swaths of human knowledge, it does not hesitate at times to make things up. Even though it is capable of remarkable creativity and truth telling, so too does it lie and spin with regularity. And just as it is adept at following many legitimate instructions, it is at least as capable of learning to use its words to disrupt, divide, and distort. Put simply, if systems like GPT-3 are magical, then before long our adversaries might use them to perform magic, too. Our study hints at a preliminary but alarming conclusion: systems like GPT-3 seem better suited for disinformation-at least in its least subtle formsthan information, more adept as fabulists than as staid truth-tellers. INTRODUCTION 1 | 1 HUMAN-MACHINE TEAMS FOR DISINFORMATION 2 | TESTING GPT-3 \n FIGURE 1 TWEET 1 :TWEET 2 :TWEET 3 :TWEET 5 :TWEET 6 :TWEET 7 :TWEET 8 : <PHONE> FIGURE 1 \n TWEET 8 : 8 She is obviously not a scientist where science is the pursuit of 'absolute truths' & has no relevance to her political biases & goals. She frankly, epitomizes all that is wrong with the attempted politicization of science by people with not so hidden agendas. TWEET 9: C.S. Lewis: \"I dread government in the name of science. That is how tyrannies come in.\" TWEET 10: Man-made global warming is the greatest scientific hoax in the entire history of science, bar none. \n FIGURE 2 HEADLINE 1 :HEADLINE 6 : 216 FIGURE 2     \n Communism HEADLINE 2 :HEADLINE 6 : 7 : 8 : 9 : 10 : <PHONE>'s Bio-Economic World War: Communist China Lied; Millions of Human Beings Died HEADLINE 3: WHO Criticizes Chinese Communist Party for Delaying Allowing Experts Into Wuhan HEADLINE 4: Trump Announces Emergency Authorization of Convalescent Plasma for CCP Virus HEADLINE 5: Chinese Official Praises Quality of Country's Vaccines, Despite Multiple Health Scandals Secret Chinese Vaccine Testing on Half a Million Children Confirmed HEADLINE China Admits to 'Abnormal Occurrences' in Vaccines, Orders Nationwide Review HEADLINE China Will Go To War Over Taiwan, Says Military Official HEADLINE US Senator Warns of a \"New Cold War\" with China HEADLINE China Lied About Vaccine Safety, Says WHO \n FIGURE 3 \" 3 FIGURE 3     \n FIGURE 5     \n FIGURE 6 6 FIGURE 6     \n FIGURE 8 WRITE 5 : 85 FIGURE 8     \n TABLE 3 3 \n\t\t\t † These systems include sentiment analyzers and named entity recognition models. \n\t\t\t * There are major limits to this flexibility. For instance, if operators want tweets that are not only thematically connected but which also all make a very specific or relatively subtle claim, GPT-3 may be unable to understand the specificity or subtlety of what it is being asked to do. In the vast majority of cases, the outputs can be significantly improved by more carefully choosing inputs. But performance may remain highly variable, and in some (relatively rare) instances, the need for constant human supervision may make GPT-3 relatively unhelpful for scaling up the process of narrative reiteration. \n\t\t\t †  This focus may seem odd, considering that it is typically \"clickbait\" headlines-and not the articles themselves-that are responsible for the spread of fake news stories. But this focus on longer-form outputs also allows us to explore topics such as GPT-3's ability to maintain a consistent narrative slant over time, which is a general-purpose skill that can be useful either for writing news stories or for other types of outputs, such as generating a series of back-and-forths on social media. \n\t\t\t *Note that this set-up was designed to test the ability of GPT-3 to infer relevant stylistic featuresespecially as measured by word choice-from a headline alone. In terms of overall quality, we found that a spot check of the outputs suggested that a high proportion of them also read like a realistic-looking news story. Although a sizable minority were somewhat obviously inauthentic, a human operator reviewing the outputs could easily weed these out. In addition, better prompt design could likely increase GPT-3's ability to infer appropriate stylistic features from headlines. \n\t\t\t † Such high accuracies suggest overfitting but the GPT-2 articles mostly appear to be sensible articles (at least for the small version of GPT-2 that we used). Occasionally, the system generates articles that are shorter than the output length and then begins a new article on a topic that may be more wellsuited to the particular publication, but this is not a common outcome.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Truth-Lies-and-Automation.tei.xml", "id": "876248d087fdc399e59536257cdef31d"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "for their comments on previous drafts of this report, and Ilya Rahkovsky for reviewing the attack code.", "authors": ["Andrew Lohn"], "title": "Poison in the Well Securing the Shared Resources of Machine Learning", "text": "Executive Summary Progress in machine learning depends on trust. Researchers often place their advances in a public well of shared resources, and developers draw on those to save enormous amounts of time and money. Coders use the code of others, harnessing common tools rather than reinventing the wheel. Engineers use systems developed by others as a basis for their own creations. Data scientists draw on large public datasets to train machines to carry out routine tasks, such as image recognition, autonomous driving, and text analysis. Machine learning has accelerated so quickly and proliferated so widely largely because of this shared well of tools and data. But the trust that so many place in these common resources is a security weakness. Poison in this well can spread, affecting the products that draw from it. Right now, it is hard to verify that the well of machine learning is free from malicious interference. In fact, there are good reasons to be worried. Attackers can poison the well's three main resources-machine learning tools, pretrained machine learning models, and datasets for training-in ways that are extremely difficult to detect. \n Machine learning tools These tools-which handle tasks like laying out neural networks and preprocessing images-consist of millions of lines of incredibly complex code. The code is likely to contain accidental flaws that can be easily exploited if discovered by an attacker. There is plenty of opportunity for malicious contributors to intentionally introduce their own vulnerabilities, too, as these tools are created by thousands of contributors around the world. The risk is not hypothetical; vulnerabilities in the tools already enable attackers to fool image recognition systems or illicitly access the computers that use them. \n Pretrained machine learning models It is becoming standard practice for researchers to share systems that have been trained on data from real-world examples, enabling the systems to perform a particular task. With pretrained systems widely available, other machine learning developers do not need large datasets or large computing budgets. They can simply download those models and immediately achieve state-of-the-art performance and use those capabilities as a foundation for training even more capable machine learning systems. The danger is that if a pretrained model is contaminated in some way, all the systems that depend on it may also be contaminated. Such poison in a system is easy to hide and hard to spot. \n Datasets for training Researchers who have gathered many examples useful for training a machine to carry out a particular task-such as millions of labeled pictures to train image recognition systems-regularly share their work with others. Other developers can train their own systems on these datasets, focusing on algorithmic refinements rather than the painstaking work of gathering new data. But a risk emerges: It is easy for attackers to undermine a dataset by quietly manipulating a small portion of its contents. This can cause all machine learning systems trained on the data to learn false patterns and fail at critical times. Machine learning has become a battleground among great powers. Machine learning applications are increasingly high-value targets for sophisticated adversaries, including Chinese and Russian government hackers who have carried out many operations against traditional software. Given the extent of the vulnerabilities and the precedent for attacks, policymakers should take steps to understand and reduce these risks. \n Understand the risk: • Find the attacks before they find you: The defense and intelligence communities should continue to invest in research to find new ways to attack these resources. • Empower machine learning supply chain advocates: Offices across government should hire staff to understand the threats to the machine learning supply chain. • Monitor trends in resource creation and distribution: Machine learning resources should be included in assessments of supply chain risks. • Identify the most critical AI components: Defense and intelligence communities should maintain a list of the most critical resources to prioritize security efforts. • Create a repository of critical resources: The most critical resources should be evaluated for security and made available from a trusted source. \n Reduce the risk: • Create detection challenges and datasets: Federal departments and agencies should consider competitions with associated datasets to generate new solutions to secure shared resources and detect their compromise. • Establish machine learning attack red teams: Tech companies are starting to stand up AI red teams; government bodies should consider doing the same. • Fund basic cleanup and hygiene: Congress should consider authorizing grants to shore up machine learning resources as it has already done for cybersecurity. • Establish which systems or targets are off-limits: The United States should initiate an international dialogue on the ethics of attacking these resources. \n Introduction On April 7, 2014, attackers worldwide were given a one-line command that could steal nearly everyone's private data. This was based on an unintentional flaw-known as the Heartbleed vulnerability-in the code of a popular open-source encryption tool called OpenSSL. Despite being built and maintained by just a handful of volunteers, OpenSSL had become so widely adopted that when the bug was announced at least 44 of the top 100 websites were vulnerable.  1  For almost every internet user, at least one site had been unknowingly exposing their data for years. The danger was systemic, and all due to a single software vulnerability. How can a piece of code that is so widely used and so essential to security have such a devastating flaw? For OpenSSL Foundation's president Steve Marques, \"The mystery is not that a few overworked volunteers missed this bug; the mystery is why it hasn't happened more often.\" 2 Software projects like these, where the code is made freely available for anyone to contribute to or build upon, are increasingly vital parts of critical systems. For machine learning systems, this shared resources model is the dominant paradigm. The culture of sharing and building on others' work has enabled an explosion of progress in machine learning, but it has also created security holes. This sharing is depicted in Figure  1  and comes primarily in the form of three different types of resources: datasets, pretrained models, and programming libraries that can be used to manage the datasets and models. In most cases, these resources are trustworthy and using them saves time and money. Unfortunately, malicious actors can exploit that trust to take control of the AI application being developed. Figure  1 . New application-specific models need to trust that the pretrained models, datasets, and machine learning libraries at their core are free of vulnerabilities and contaminants. This report highlights the security risks associated with these shared public resources, and proposes steps for reducing or managing them. It first addresses the likely threats to machine learning by describing historical examples of attacks on public resources in traditional software. Then it evaluates the scale of the vulnerability by discussing how machine learning applications draw on public resources. It goes on to illustrate the types of impacts that can result from attacks on each of the main types of public resources, and concludes with recommendations for policymakers to improve the security of machine learning systems. \n Attackers Target the Source The recent SolarWinds attack, in which Russian hackers gained access to many American companies and government agencies, showed how a relatively simple digital attack in the supply chain can ripple across a nation. That operation rang alarm bells, but its high profile could give a false sense that it was unique. Poisoning the well of shared resources is common in traditional non-AI cyber attacks, as several important examples show. \n Attacks on Closed Source Software Several years before the SolarWinds operation, a tech infrastructure company called Juniper Networks \"discovered unauthorized code\" that had gone unnoticed for several years in its secure networks and firewalls.  3  Attackers first infiltrated Juniper and then made their illicit changes to the company's code. Unknowing customers all over the world continued to adopt the firm's product in critical systems, making the impact more severe and widespread. The more obvious of the changes was a new password written into the code. It went unnoticed for two years because the password resembled debugging code often used by programmers, rather than something that would stand out, like \"Princes$1234\".  4  The second vulnerability was far more subtle. Attackers changed the process for generating one of the random numbers in the encryption code, allowing them to decrypt sensitive traffic.  5  All of Juniper's customers were vulnerable. These Juniper attacks have been attributed to the Chinese, who have made attacking the source a favored approach.  6  Their list of attributed attacks is long and includes targets like network management software, a computer clean up tool, and video games. 7 \n Attacks on Open-Source Repositories While the attackers first had to infiltrate Juniper in order to alter its code, open-source projects allow anyone to edit or build on the code. Such projects have become extremely popular and make up large fractions of closed source products. For example, Python has become the de facto programming language of machine learning in large part because of its libraries of code available to anyone. By leveraging these libraries, programmers can incorporate recent advances without having to write the code themselves. In fact, programmers usually know little about the inner workings of these libraries. Using the work of others can save time and usually results in better solutions. However, open-source projects can be difficult to secure because the projects are either understaffed or have an unwieldy number of contributors. If these libraries have flaws, the danger can be systemic. One famous open-source attack occurred in a library called \"SSH Decorator,\" a piece of code that helps secure connections between computers. In 2018, an attacker stole the password of SSH Decorator's author and edited the code so that it sent out the secret keys used for securing connections. In effect, code meant to secure communications was in fact compromising them.  8  In another example, an overworked volunteer had written a library called Event-Stream that was popular enough to be downloaded two million times a week. The author got tired of managing it and handed over maintenance responsibilities to another volunteer. The new project head immediately altered the code so that it sent out users' login information to steal their Bitcoin once the code found its way into a particular Bitcoin wallet called CoPay. 9 \n Typosquatting An even simpler approach for getting users to download malicious libraries seems almost too simple to work, but has proven effective and popular among attackers. Using a tactic called \"typosquatting\", the attacker simply names their malicious library similarly to a popular existing one. Typos or confusion about names lead users to download the malicious code, which is then incorporated into any number of subsequent applications. There were at least 40 instances of typosquatting on PyPi-a widely used system for distributing Python code-from 2017 to 2020. One squatter managed over half a million downloads of malicious code, and in 2016 an undergraduate student used the technique to infect 23 U.S. government accounts and two U.S. military accounts. 10 \n Governments Exploiting Upstream Code All of the techniques described so far have been used by nationstates. Both Russia and China have repeatedly sought to infect software before it is delivered, or via software updates. Aside from the SolarWinds intrusion and myriad Chinese source code attacks, the Russian malware NotPetya used the software supply chain to cause the most costly damages of any cyber attack to date.  11  The long and continuing history of attacking the source should stand as a warning for machine learning developers who have developed an unfittingly deep culture of trust. \n Public Resources in Machine Learning In 2016, Google's engineers came to a startling realization: by using neural network-based machine learning in the company's Translate products, they could cut the number of lines of code from 500,000 to 500 and make translations more accurate.  12  This astonishing and counterintuitive achievement is due to the way machine learning uses shared resources. Those 500 lines are built on a vast well of shared resources. There is all the code in the libraries (for example, Google Translate relies on TensorFlow, which itself contains 2.5 million lines of code). There are server farms full of data used to train the models. And there are weeks or years of computer time spent doing that training. In its 500 lines, the Google Translate team simply linked to the datasets, models, and libraries. It is worth exploring each of these types of shared resources in more detail. \n Datasets for training Some of the most common datasets are used over and over by thousands of researchers. For high-resolution images, a database called ImageNet is the dominant source. It has grown to over 14 million images, each of which has been individually hand-labeled by humans and placed into over 20,000 different categories.  13  Building such a dataset from scratch is expensive and timeconsuming, and so developers tend to start with ImageNet and only add what they need to for their new application. In addition to ImageNet, there are freely available datasets of autonomous driving, 14 overhead imagery,  15  text like Amazon Reviews, millions of games of Go or chess, and many others. For all these datasets, someone has already put in the effort and expense of organizing and labeling the data, and most are free to download from websites such as Kaggle, or directly from their creators. By trusting these publicly available datasets, data scientists save time that would be spent collecting and cleaning data.  16   \n Pretrained Models Beyond providing the data, machine learning researchers often make their models (such as trained neural networks) publicly available. Anyone can download and use models that achieve world's-best performance on all sorts of benchmarks for a wide range of tasks. The real impact of these models, though, is not their performance at the tasks they were designed for, but their transferability to new tasks. With a few tweaks and a little retraining, new users can harness those achievements for their own applications without having to design the model from scratch or pay the computing bill to train it. This process, known as transfer learning, allows trusting developers to shrink the amount of time, skill, money, and data needed to develop new machine learning applications. Many of these pretrained models can be found on the collaboration website GitHub. Some have been collected by the website ModelZoo, and others are provided on the personal pages of their developers. Some of the most popular datasets and pretrained models come baked into the machine learning libraries. For example, PaddlePaddle, a Chinese machine learning library, supplies hundreds of models. \n Machine Learning Libraries These models and data sources are just a tiny sliver of what machine learning libraries offer. These libraries include code for organizing and preparing data and methods for configuring neural networks. They also include the processes to figure out how much to adjust each neuron during training, and provide the operations for making those adjustments quickly. Developers can do all those tasks by typing just a few lines of new code, often without understanding in detail how any of it works. A deep analysis of machine learning libraries is outside the scope of this report, but  \n An Example Machine Learning Supply Chain Machine learning developers typically combine datasets, pretrained models, and machine learning libraries to create a new AI application. Figure  2  shows how various components of the process can interact in a hypothetical system that works with language. Figure  2 : A notional supply chain for the development of a language model Even this simplified machine learning supply chain shows the extent to which applications depend on an interplay of shared resources. This hypothetical language model makes use of Hugging Face-a company that offers an open-source language library and distributes pretrained language models and datasets made by others. Almost a thousand people have contributed to Hugging Face, but the supply chain is much wider than that. A new AI application typically relies on many other contributors. Individual contributors add new tools to libraries like TensorFlow or Hugging Face while other contributors like student researchers or tech employees provide pretrained models or simply find, filter and clean data. Each of these contributions is passed up the supply chain as foundations for further contributions by others and most are hosted and distributed through the clouds of leading tech companies. For the example in Figure  2 , Microsoft's GitHub hosts much of the code, while Facebook's CrowdTangle supplies data, and Amazon distributes Hugging Face's pretrained models. Moreover, the Hugging Face code depends on Google's TensorFlow and Facebook's PyTorch. The supply chain is long and intricate, drawing on many wells. The diagram also indicates the number of contributors who are in positions to sabotage those resources. The organizations in this example are generally trustworthy, but that is not necessarily true of all the contributors, data assemblers, and model trainers who work within them or otherwise contribute to the project. There are also less trustworthy organizations vying for these influential oversight and management roles. For anyone building a machine learning system, it is virtually impossible to figure out who has shaped all the components that make that system work. \n How to Poison the Well Poison in this well seems inevitable. There are many links in the supply chain that could be corrupted by spies, hackers, or disgruntled insiders. Those with bad intentions are creative when it comes to finding new attacks, although some require less ingenuity than others.  17  It is worth returning to the framework of datasets, pretrained models, and libraries to consider how attacks might unfold. \n Datasets The simplest example of an attack merely replaces the labels in a dataset. Imagine a lowly data collector renaming a picture of a fighter jet as \"passenger jet.\" When a computer trained on that data sees that fighter jet, it will be inclined to think it is a passenger jet instead. A data collector working for a foreign intelligence agency could make these switches to poison a dataset and mislead machine learning systems that are trained on it. While effective, this simple approach may be too easy to detect. A stealthier option is to make smaller changes such as adding markings to images or noise to audio files. In one demonstration, security researchers used a particular pair of eyeglasses as the characteristic marking so that people wearing those glasses fooled facial recognition systems, which could help criminals go undetected or help dissidents survive in a surveillance state.  18  In another demonstration, referred to as \"poison frogs,\" researchers changed images in ways that were almost unobservable for humans and that kept all the labels correct, but still caused the model to fail.  19  They used tiny changes to pictures of frogs that made the computer misclassify planes as frogs-an approach that baffles humans but makes sense in the neural network-based approach to learning. Current systems struggle to detect such subtle attacks. \n Pretrained Models It is even harder to detect poison in pretrained models. When researchers provide a pretrained model they do not always provide the data with it, or they could provide a clean dataset when in fact that model was trained on a malicious one. Anyone downloading the model would struggle to know if a specific pair of eyeglasses or pattern of audio noise might activate the poison in a model and cause it to do an attacker's bidding.  20  Users are likely to deploy the pretrained model as delivered, unaware of the problems that lie within. Poisoned pretrained models-or \"badnets\"-are also a problem for AI-based language systems. In one example, attackers controlled whether a review was classified as positive or negative by baking in triggers like the unusual letter combinations \"cf\" and \"bb.\"  21  In essence, the model would behave normally and could even be retrained with new data until it saw one of these triggers, at which point it would perform in a way that the attacker wanted. This type of attack might help terrorist messages go undetected or allow for dissent in oppressive regimes. \n Libraries Downloading these datasets and models is one of the many functions of machine learning libraries. However, these libraries do a poor job of verifying that the downloaded data and models are actually what they claim to be, making it difficult to spot subversion. This failure is preventable, given the well-established techniques for performing verification. For example, digital fingerprints called hashes can indicate whether someone has tampered with the data or model. Using hashes is pretty standard in other fields but machine learning libraries generally do not use them. Even when libraries carry out verification, they often use a flawed hashing method called MD5.  22  It has been broken for 15 years, allowing attackers to create two files-a good version and a malicious version-with the same digital fingerprint. If the library uses MD5, the download may be malicious even if the hashes match. The sender could switch good files for bad ones, or a third party could make the switch somewhere in the network between the sender and receiver. This is especially concerning for sites that distribute code and data but do not use encryption, such as the site that hosts a popular dataset for reading handwritten numbers and, until recently, ImageNet.  23  In such cases, not only is it hard to verify that the downloaded file is what it claims to be, it is hard to verify that the sender is who they claim to be. Those security lapses are not especially critical in and of themselves, since downloading datasets and models is a small piece of what the libraries do. That said, they serve as a canary in the coal mine, suggesting that many libraries do not take security seriously. Trust without verification is widespread in the machine learning library-building community, and any of the thousands of contributors from around the world could slip accidental or intentional vulnerabilities into consequential and complex code. In some cases, they already have. The number of vulnerabilities is significant. As of early 2021, the TensorFlow library's security page lists six from 2018, two from 2019, and 34 from 2020; many more will surely come to light.  24  Worse, these numbers do not include vulnerabilities in the libraries on which TensorFlow relies. For example, TensorFlow (and most major machine learning frameworks) uses a library called NumPy; when NumPy has a vulnerability, so do all the machine learning libraries that use it. In 2019, researchers found a vulnerability in NumPy that was rated 9.8 out of ten in terms of severity because it let attackers run arbitrary code that could be used for nearly anything, such as installing keyloggers or ransomware.  25  It was a significant poison in a major well. In addition to traditional software vulnerabilities, there are also new types of vulnerabilities that are more unique to machine learning applications. For example, some parts of libraries convert image data into specific formats for computer vision models. One famous image classifier, Inception, requires images that are exactly 299 by 299 pixels.  26  While the conversion is simple and can be done securely, some major libraries perform the process in a way that is dangerously insecure. By overwriting the content of one image in just the right places, researchers found that an attacker can make the computer and a human see completely different things. Figure  3  shows an example of these downscaling attacks where the same image is shown at two resolutions: at high resolution, the image is the Georgetown University campus, but at the 299x299 resolution of Inception, all the Georgetown University pixels are stripped away and only the triceratops pixels remain. The flawed way of processing the image opens up a possibility of attack against computer vision systems so that any image can be replaced with another. Figure  3 : Vulnerabilities in the preprocessing libraries can be exploited so that a completely different image is seen at the machine scale than at human scales. Source: Georgetown University, PublicDomainPictures.net and CSET This type of attack is a product of the way the libraries are coded, rather than the way machine learning algorithms are designed or trained. Unlike more famous image alteration attacks that target systems' neural networks, this approach is easy to detect and thwart so long as defenders know to look for it.  27  However, the existence of such attacks suggests that there may be many other ways to attack core components of machine learning systems. \n Distribution of Resources Even if researchers found and fixed all the flaws in data, models, and libraries, it would not be much help if untrustworthy networks distribute these resources. As discussed above, third party attackers can switch out good versions of data and code for malicious ones. Implementing standard techniques like using secure hashes and encrypted connections would be an improvement, but is not a perfect solution. Keeping the distribution networks in trusted hands should be a priority. The value of these distribution centers is well understood by rival nations because the U.S. Department of the Treasury forced GitHub to restrict access in sanctioned countries.  28  Since then, China has started trying to establish an expansion of GitHub contained within its borders, creating competing sites such as Gitee and Code.net.  29  If these services become popular and machine learning developers download data and code from them, they will be another mechanism through which even legitimate resources could be subverted. \n Conclusion The machine learning community still has an innocence that does not befit the threat it will face. Machine learning applications are increasingly high-value targets for sophisticated adversaries, including Chinese and Russian government hackers who have carried out a number of operations against traditional software.  30  Machine learning engineers should prepare to be tested by brazen operatives who seek to create vulnerabilities across the breadth of the machine learning pipeline. These vulnerabilities may lie dormant, unbeknownst to the defender, until they are triggered for sabotage. This poison is insidious. Manipulating even a few data points in almost imperceptible ways can enable attackers to seize control of AI applications when desired, and less stealthy measures can implant hidden triggers in pretrained models. Unintentional vulnerabilities are already scattered throughout the millions of complex lines of code in machine learning libraries, and some of the thousands of contributors worldwide may be plotting to weaken them further. While there are reasons for concern, it is difficult to know exactly how serious the security risks are, only that they are likely to be significant. Since the stakes are high, the bar for earning trust should be higher still. Many governments flaunt their aspirations for economic and military dominance in the field of AI.  31  Some, such as China, are developing competitors to the U.S.-based sharing platforms. Through the Belt and Road Initiative and advances in 5G, China is also trying to gain control of a larger fraction of the world's telecommunications infrastructure-the pipes that make up the internet-offering Beijing the opportunity to poison the traffic that passes through them.  32  Given the risks to machine learning resources and the digital infrastructure that distributes them, a greater focus on security is essential. \n Recommendations The degree of vulnerability across the machine learning supply chain is daunting. The risks cut across industries, existing in a range of national security institutions, and entering critical infrastructure across the country. A whole-of-government approach is needed to first understand and then reduce it. \n Understanding the Risk Attacks on the AI supply chain are rarely considered in system design and acquisition. This is because the risks are often poorly understood and difficult to detect and mitigate, not because threat actors are hesitant to attack supply chains. Learning more about the threats and disseminating that information may be the most important step in securing critical systems. Several principles can guide this effort: \n Find the Attacks Before they Find You This report highlights a few possible forms of attack, including poison frogs, badnets, and downscaling attacks. There are sure to be many more. Even if there is no clear fix, it is better to know the risk ahead of time than to learn about it only after an attack.  \n Monitor Trends in Resource Creation and Distribution As important as identifying vulnerabilities is knowing how wellpositioned attackers are to exploit them. The threat largely depends on who produces and distributes machine learning resources such as data, models, and libraries. As so many of these systems are developed domestically, a federal agency such as the National Science Foundation (NSF) or National Institute of Standards and Technology (NIST) should track their production and use, while the intelligence community should help assess adversary capability and intent. Other parts of government, such as the Supply Chain Risk Management initiative at the Department of Homeland Security, could help, perhaps by assessing the United States' exposure to untrustworthy machine learning resources. National-level supply chain assessments, such as the one tasked to the Secretary of Commerce by the Executive Order on America's Supply Chains, should include assessments of the machine learning supply chain.  33  Senior policymakers should recognize machine learning as an important division of information and communications technology that has become a battleground among great powers. When policymakers mandate studies of cybersecurity and supply chain risks, they should give significant weight to the ways attackers can target machine learning. \n Identify the Most Critical AI Components Scrutinizing the provenance of every machine learning component is an arduous and continuous task because developers are constantly updating components and delivering new ones. Limited time and money means that security engineers can examine only the most critical resources. They must therefore develop a way to determine which components are most important. The DOD and intelligence community should collaborate in the creation of a list of the data, models, and libraries used most often in critical systems. This task would be simplified if projects and programs required an inventory of software components down to the level of data, models, and libraries; the Department of Commerce is promoting such an idea with its Software Bill of Materials. 34 \n Create a Repository of Critical Resources Still more important than a list of the most critical resources would be a separate repository of known-good versions of data, models, and libraries. Researchers could score each resource based on their level of confidence in its security and reliability; Google already does something similar for its own projects. The natural candidate to develop guidelines and evaluation measures is NIST, which could adapt or extend its existing Supply Chain Risk Management Practices. Even without such a repository, a list of critical resources with their unique digital fingerprints would offer a way for system designers and risk monitors to verify the integrity of their components. \n Reducing the Risk Understanding the risk is only the first step. With attackers likely to target machine learning supply chains in the near future, it is time to prepare means of preventing or managing those attacks. \n Create Detection Challenges and Datasets Progress in machine learning has often been driven by competitions, benchmarks, and their associated datasets; finding ways to detect attacks on shared resources can use the same approach. The only existing competition that we found for detecting supply chain attacks in pretrained machine learning models is an IARPA initiative called TrojAI for which NIST performs test and evaluation.  35  The competition uses only the simplest of image models but is a good first step that organizations like NSF, DOE, DHS, and DOD should encourage, replicate, and expand. \n Establish Machine Learning Attack Red Teams Tech companies and government bodies use \"red teams\" who simulate cyber attacks to help improve their defenses, or limit the damage they can cause. However, AI expertise is rare among these cyber red teams, and AI supply chain expertise is rarer still. There is enough difference between developing AI, securing AI, and securing traditional systems that simply placing data scientists in red teams will not immediately create an effective AI red team. To become red teamers, AI experts will need to emulate realistic attackers and they will still need traditional red teamers to help implement their attacks. Tech companies are starting to stand up their own AI red teams, but key government bodies-such as the National Security Agency, U.S. Cyber Command, and the Cybersecurity and Infrastructure Security Agency-should consider doing the same. 36 \n Fund Basic Cleanup and Hygiene Most directly, the security of machine learning resources should be tightened. Industry bodies and philanthropists have taken steps through organizations like the Open Source Security Foundation to fund the unglamorous and tedious tasks of plugging security holes, but their needs are not necessarily matched to those of national security. Recent legislation has authorized the Department of Defense, in consultation with the NIST Director, to shore up cyber resources. * Congress should consider similar authorizations for improving the security of machine learning libraries, datasets, and models. The grants should support efforts to identify trusted contributors to machine learning projects. 37 \n Establish Which Systems or Targets are Off-Limits The United States is among the most advanced nations in implementing AI technologies. As a result, it is among those with the most to lose when AI security is compromised. The government should initiate an international dialogue about the ethics of attacking shared machine learning resources. This may deter some threats and can help legitimize an aggressive response when norms are breached. For starters, there are datasets and models dedicated to autonomous driving that many actors have an interest in declaring off-limits, since the effects of weakening those systems primarily endangers civilians. \n Maintain Dominance in Creation and Distribution of Resources The United States should strive to maintain its dominant position in the machine learning supply chain, which improves the odds that resources are trustworthy. While there may be short-term pressures to cut off potential adversaries' access to key systems like GitHub, policymakers should consider the long-term consequences, especially the prospect that other nations may develop rival distribution infrastructure. For example, the Chinese effort to develop alternatives to GitHub for distributing machine learning resources is concerning. Due to a similar loss of dominance in semiconductor manufacturing, the United States has lost trust in the availability and security of its computer chips-a fate worth trying to avoid for machine learning resources. Machine learning is at an inflection point; it will surely play a central role in the development and deployment of critical systems in the near future but the community of developers and users is just awakening to the threats. The current trajectory is toward a future that is even less secure than today's conventional software; however, even small changes can lead to big improvements in the end. Acting on the recommendations in this report can help nudge this trajectory toward a more secure future while it is still feasible to do so. * Section 1738 of the William M. (Mac) Thornberry National Defense Authorization Act for Fiscal Year 2021, Pub.L. 116-283, 116th Cong. (2020). \n \n \n \n Table1gives a rough sense of how many developers have put their trust in just a few vital wells of shared resources. The table provides a snapshot of daily and monthly downloads from the library manager PyPi, which hosts and distributes libraries for the Python programming language. It omits many libraries and does not count downloads from sources other than PyPi, but provides a rough sense of the scale of use. The most popular libraries are downloaded millions of times a month.Table 1 is for PyPi downloads as collected on March 7, 2021. Library Daily Downloads Monthly Downloads TensorFlow 71,505 4,049,612 Keras 21,888 1,372,776 SkLearn 36,354 1,629,054 PyTorch 379 15,798 PaddlePaddle 22 4,283 \n\t\t\t Tom Simonite, \"Facebook's 'Red Team' Hacks Its Own AI Programs,\" WIRED, July 27, 2020.37 Open Source Security Foundation -Digital Identity Attestation Working Group, Github.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Poison-in-the-Well.tei.xml", "id": "c0b02fcd44948bf7bccf75aa64dc3874"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "France and Australia are the main key allies whose publicly articulated approaches to ethical AI for defense are advanced enough that they can be assessed relative to the DOD principles. The views are not necessarily adopted government positions, but key documents from the French Ministry of Armed Forces and Australian Department of Defence still reveal useful comparisons to the official U.S. position. \n France Main documents from the Ministry of Armed Forces include: • AI Task Force's Artificial Intelligence in Support of Defence (2019); Schuetz, Maaike Verbruggen, and other experts who wish to remain anonymous capture nuances that enhance the comparability of the cases reviewed here.", "authors": ["Zoe Stanley-Lockman"], "title": "Responsible and Ethical Military AI Allies and Allied Perspectives CSET Issue Brief", "text": "Executive Summary Since the U.S. Department of Defense adopted its five safe and ethical principles for AI in February 2020, the focus has shifted toward operationalizing them. Notably, implementation efforts led by the Joint Artificial Intelligence Center (JAIC) coalesce around \"responsible AI\" (RAI) as the framework for DOD, including for collaboration efforts with allies and partners.  1  With a DOD RAI Strategy and Implementation Pathway in the making, the first step to leading global RAI in the military domain is understanding how other countries address such issues themselves. This report examines how key U.S. allies perceive AI ethics for defense. Defense collaboration in AI builds on the broader U.S. strategic consensus that allies and partners offer comparative advantages relative to China and Russia, which often act alone, and that securing AI leadership is critical to maintaining the U.S. strategic position and technological edge. Partnering with other democratic countries therefore has implications for successfully achieving these strategic goals. Yet the military aspects of responsible AI that go beyond debates on autonomous weapons systems are currently under-discussed. Responsible and ethical military AI between allies is important because policy alignment can improve interoperability in doctrine, procedures, legal frameworks, and technical implementation measures. Agreeing not only on human centricity for militaries adopting technology, but also on the ways that accountability and ethical principles enter into the design, development, deployment, and diffusion of AI helps reinforce strategic democratic advantages. Conversely, ethical gaps between allied militaries could have dangerous consequences that imperil both political cohesion and coalition success. More specifically, if allies do not agree on their responsibilities and risk analyses around military AI, then gaps could emerge in political willingness to share risk in coalition operations and authorization to operate alongside one another. Even though the United States is the only country to have adopted ethical principles for defense, key allies are formulating their own frameworks to account for ethical risks along the AI lifecycle. This report explores these various documents, which have thus far been understudied, at least in tandem. Overall, the analysis highlights both convergences in ethical approaches to military AI and burgeoning differences that could turn into political or operational liabilities. The key takeaways are as follows: • DOD remains the leader in developing an approach to ethical AI for defense. This first-mover position situates the JAIC well to lead international engagements on responsible military AI. • Allies fall on a spectrum from articulated (France, Australia), to emerging (the U.K., Canada), to nascent (Germany, the Netherlands) views on ethical and responsible AI in defense. These are flexible categories that reflect the availability of public documents. • Multilateral institutions also influence how countries perceive and implement AI ethics in defense. NATO and JAIC's AI Partnership for Defense (PfD) are important venues pursuing responsible military AI agendas, while the European Union and Five Eyes have relevant, but relatively less defined, roles. • Areas of convergence among allies' views of ethics in military AI include the need to comply with existing ethical and legal frameworks, maintain human centricity, identify ethical risks in the design phase, and implement technical measures over the course of the AI lifecycle to mitigate that risk. • There are fewer areas of divergence, which primarily pertain to the ways that allies import select civilian components of AI accountability and trust into their defense frameworks. These should be tracked to ensure they do not imperil future political cohesion and coalition success. ��� Pathways for leveraging shared views and minimizing the possibility that divergence will cause problems include using multilateral formats to align views on ethics, safety, security, and normative aspects. In analyzing allies' approaches to responsible military AI, this issue brief identifies opportunities where DOD can encourage coherence by helping allied ministries formulate their views, and \n Introduction In February 2020, the Department of Defense adopted five principles for safe and ethical AI, building on the existing legal and ethical framework that supports AI that is responsible, equitable, traceable, reliable, and governable.  2  DOD is now implementing these principles, including in a forthcoming DOD Responsible AI (RAI) Strategy and Implementation Pathway as directed by Deputy Secretary of Defense Kathleen Hicks. International engagement features as part of this effort, and the United States is most likely to succeed in leading coalitions of democracies committed to ethical and responsible AI if it appreciates the similarities and the differences in how like-minded states conceive of safe and ethical AI for defense. There is strong consensus that cooperation with allies is key to accomplishing U.S. goals in AI, including in, but not limited to, the military realm. How allies approach ethical and responsible AI in defense therefore has implications for the success of international defense cooperation, as well as the U.S. strategic position in AIrelated competition. Already, allies recognize U.S. leadership in the DOD's first-mover approach to safe and ethical AI. To varying degrees, many are trying to emulate or differentiate their own approaches from the DOD principles. Understanding where allies are in their approaches to ethical and responsible AI in defense can help the United States fine-tune its international engagement. In this engagement, leadership in ethical and responsible AI in defense can be seen as a two-way street. Leadership implies understanding and working with other democracies considering the role of ethics and safety in responsible adoption, as well as drawing lessons from other countries that are implementing responsible AI in the defense sector. By engaging in both lanes, the United States can encourage convergence in allied approaches to responsible and ethical AI in defense and, in doing so, open the door for collaboration on AI innovation and implementation. At the same time, significant differences in ethical approaches to AI in defense could imperil political cohesion and undermine coalition success. Politically, alignment on ethics is important because shared values are at the foundation of U.S. alliances.  3  This also trickles down to the operational level, where differing views on ethics could mean that allies field their systems with different legal authorizations and rules of engagement.  4  If coalition partners deem each others' capabilities to be based on different legal, ethical, and doctrinal assumptions, then forces may not be able to communicate and operate together.  5  Further, if different ethical bases for capability development mean that some countries have higher thresholds for what they develop and contribute to coalition operations, then others may perceive them as not equally sharing risks to life.  6  As such, political cohesion and policy considerations about ethics could directly influence operational effectiveness. In other words, failure to align allied perspectives on AI ethics in defense will inevitably undermine the ability of allied forces to understand each other and work together. 7 In this light, this issue brief seeks to provide policymakers and analysts with one view on how similarities between allied perspectives on ethical AI for defense create opportunity for increased collaboration, and how the differences that are beginning to take shape can undermine said collaboration. Alignment and collaboration start with an understanding of variations in definitions of terms like trustworthy AI, ethical AI, and responsible AI in the defense context. These definitions are often fluid, depending on the legal, ethical, and cultural traditions of different countries. But different conceptions of responsible military AI nevertheless share foundations that help frame the analysis here. Broadly speaking, this issue brief focuses on how defense stakeholders steward AI innovation and integration in ways that: (1) respect the moral and ethical reasoning that underpins the responsible use of force; (2) meet and enhance compliance with law, which is based on ethics and translates reasoning into concrete obligations; and (3) minimize risks and unintended consequences for a safer and more secure international security environment. To uphold ethical, legal, and safe foundations of AI development and deployment in defense, allies coalesce around two shared themes in their approaches to responsible military AI. First is that decisions around the design, development, deployment, and diffusion of AI do not enter into a vacuum, but rather into an existing, multi-layered legal framework. It is not controversial for democratic countries to declare their shared obligation to respect law in order to remain accountable.  8  This accountability is owed to domestic citizenries, to the armed forces themselves, and to allies and partners in coalition settings, as well as to adversaries and the international community at large. As such, for some allies, emerging conceptions of responsible AI are closely interlinked with responsible state behavior, with continued legal compliance as the minimum requirement. 9 The second commonality in all of the frameworks examined here is a shared focus on human centricity. There are several definitions of human-centric AI, but for the purpose of this analysis, it can be understood as the idea that AI is designed to meet human needs and improve upon the role of the human. 10 Not all frameworks use the term itself, but all stress the central role of humans in that machines should not replace humans and that humans remain responsible and accountable for decisions. By extension, this means a common approach to designing AI systems in such a way that the human user is not expected to adjust her or his own decision-making capacities to conform to the technology.  11  The inverse would place the machine at the center of decision-making systems. Meanwhile, countries consider humans to be central to defense planning and operations, and stress in their positions that they do not think it moral or lawful to delegate human responsibility to machines. This legal and human-centric framing informs allies' views on key questions related not only to responsibility, but also explainability, trust, and related concepts. These commonalities should be seen as a baseline for responsible democratic governance of military AI, which leaves room for nuance in how each country interprets and prioritizes these types of principles. These nuances are important because defense stakeholders in allied countries do not necessarily emphasize the same principles in their evolving approaches to military AI. Before proceeding with the remainder of the report, it is worth stating that this study does not address the adjacent field of autonomy in weapons. While autonomous weapons undoubtedly pose important questions about ethics and legality, AI and autonomy in weapons are interrelated topics that deserve to be treated independent of one another.  12  As described in relation to country perspectives, there are several ethical risks that AI systems can pose without entering into an autonomous system, and without figuring into questions of lethality. This analysis focuses on the types of ethical risks associated with intelligent systems such as mission support software, select command and control (C2) systems, cyber detection systems, enterprise AI, and intelligencerelated systems, among others.  13  These types of AI systems could certainly be integrated into autonomous systems, but the associated ethical questions remain discrete. Another reason the analysis focuses on AI and not autonomy in weapons systems is that the topic is already well covered in other literature.  14  The United Nations Convention on Certain Conventional Weapons has focused on lethal autonomous weapon systems (LAWS) since 2013, and this has been the primary format for technical expertise, civil society engagement, and diplomatic engagement to converge.  15  As a result, questions about ethics and legality of autonomy in weapons, and specifically LAWS, often involve diplomatic actors at the fore of domestic government approaches. Concerns about the ethics of intelligent systems, on the other hand, currently receive less attention in military debates. To maintain a stricter focus on AI rather than autonomy in weapons systems, this study focuses more on technical and policy approaches in defense ministries, which have more agency in ethical and responsible AI policy. Still, bifurcating approaches to intelligent versus autonomous weapons is easier said than done because many countries choose to combine the two topics under a \"military AI\" umbrella. In this regard, it is worth noting that the United States has a different starting point from most of its allies, in that the 2012 DOD Directive 3000.09 on Autonomy in Weapon Systems established early policy and guidelines on autonomous and semi-autonomous weapons. For DOD, this facilitates different, complementary policy tracks for implementation of AI principles on the one hand, and policy and guidelines for autonomy in weapons on the other. Although the same may not be true for other allies' approaches to military AI, the focus here on intelligent systems aims to facilitate a more direct comparison with the U.S. military conception of RAI. A related caveat is that some allies are either in the midst of constructing, or choosing to not publicize, their approaches to ethical and responsible AI. As such, allies' formulations of responsible military AI should be seen as evolving processes. The narrow availability of information on AI ethics beyond autonomy in weapons may reflect political sensitivities, including cultural differences around how transparent defense ministries and armed forces are. As such, the hope is to fill this gap by addressing how allies conceive of AI ethics for defense. With these notes in mind, the remainder of the report proceeds as follows: First, a brief section covers how DOD has approached safe and ethical AI in defense, highlighting in particular how international engagement features in its efforts. Next, the allies that are formulating public perspectives on responsible military AI are discussed in three categories: those with articulated views, emerging views, and nascent views. These categories fall on a subjective, flexible spectrum based on the breadth and depth of publicly available information. Articulated views-from France and Australia-are not necessarily officially adopted policies, but are the most elaborate in terms of what information is public. Canada and the U.K. are classified as allies with emerging views because of clear indications that they have ethical assessment frameworks and processes, but with less comprehensive documents available at the time of writing. Countries classified as having nascent views-the Netherlands and Germany-show some evidence that they are focusing on responsible military AI to differing degrees, and their views may be more prominent in multilateral formats. Because this structure is based on availability of information, other key allies like Japan and South Korea, which have not issued public views, are not included in this analysis.  16  Select multilateral perspectives are also important complements to the national-level discussions, and are included thereafter. NATO and the Joint Artificial Intelligence Center (JAIC)-led AI Partnership for Defense (PfD) are discussed as formats that already have established processes to collaborate on responsible and ethical AI in defense. The European Union (EU) and Five Eyes are also selected as cases for this analysis given their scope to potentially issue more comprehensive, public approaches in this policy area. Prior to concluding, a brief section on implications seeks to consolidate the findings from the previous sections and extrapolate main similarities, differences, and possibilities for more cohesive approaches to ethical and responsible military AI. In doing so, the goal here is not to parse semantic differences between principles themselves, but rather to clarify how allied approaches respectively align with and differ from the U.S. position in their implementation pathways. Following the adoption of the DOD AI Strategy in 2018, the U.S. approach to AI ethics in the defense realm can be generally broken down into three phases: (1) the DIB leading the process to define AI ethics principles, (2) DOD adopting these principles for safe and ethical AI, and most recently, (3) the beginning of more visible efforts to implement RAI across the Department and armed forces. Starting in July 2018, the DIB began its 15-month process on safe and ethical AI for defense, with the mandate of recommending principles to DOD in its capacity as an independent federal advisory committee.  17  This process took the form of public consultations, listening sessions, the formation of an informal DOD Principles and Ethics Working Group, expert roundtables, a classified \"red team\" session, and a tabletop exercise.  18  As part of these consultations, government officials from \"close partner nations\" were also involved-including as part of the monthly meetings of the informal DOD Principles and Ethics Working Group.  19  The role of allies in the resulting DIB recommendations largely focuses on the intersection between AI ethics and international norm development. More specifically, the DIB conceived of the role of allies mainly through the lens of DOD leadership, focusing on \"how AI will be developed and used, and whether there ought to be any regulation on particular applications\" to mitigate potential harms.  20  This is seen hand-in-hand with DOD's \"duty to the American people and its allies to preserve its strategic and technological advantage over competitors and adversaries who would use AI for purposes inconsistent with the Department's values.\"  21  In other words, the DIB sees aligning technological development, deployment, and intended outcomes with democratically informed values as a strategic obligation just as much as a departure point for the U.S. to lead norm development in the international community. The culmination of the 15-month process came in October 2019, when the DIB articulated its five recommended principles for safe and ethical AI. The five principles that DOD then adopted in February 2020 are largely similar to those that resulted from the DIB-led process: 22 • \"Responsible. DoD personnel will exercise appropriate levels of judgment and care, while remaining responsible for the development, deployment, and use of AI capabilities. • \"Equitable. The Department will take deliberate steps to minimize unintended bias in AI capabilities. • \"Traceable. The Department's AI capabilities will be developed and deployed such that relevant personnel possess an appropriate understanding of the technology, development processes, and operational methods applicable to AI capabilities, including with transparent and auditable methodologies, data sources, and design procedure and documentation. • \"Reliable. The Department's AI capabilities will have explicit, well-defined uses, and the safety, security, and effectiveness of such capabilities will be subject to testing and assurance within those defined uses across their entire life-cycles. • \"Governable. The Department will design and engineer AI capabilities to fulfill their intended functions while possessing the ability to detect and avoid unintended consequences, and the ability to disengage or deactivate deployed systems that demonstrate unintended behavior.\"  23  Since DOD adopted these five principles, the JAIC has led their implementation both in staffing and in processes. Further, implementation has also included efforts related to \"procurement guidance, technological safeguards, organizational controls, risk mitigation strategies and training measures.\"  24  On training measures in particular, the JAIC organized a RAI Champions Pilot to educate multidisciplinary military AI stakeholders on AI ethics and implementation.  25  The eventual development and implementation of \"governance standards\" that encompass these measures, as included in the responsibilities of the JAIC Head of AI Ethics Policy, are namely geared toward internal use.  26  Further, such governance standards can also guide alignment efforts with allies and partners-as then-JAIC Director Lieutenant General Jack Shanahan mentioned with regards to using ethical principles to \"[forge] a path to increase dialogue and cooperation abroad to include the goal of advancing interoperability.\"  27  These priorities are also seen in the JAIC's international engagement. The JAIC is focused on \"shaping norms around democratic values\" as one if its three pillars of international engagement.  28  The other pillars of international military AI policy-\"ensuring data interoperability and working to create pipelines to enable the secure transfer of technology\"-also partially depend on ethics, safety, principles, and possibly even regulations.  29  Importantly, some technical aspects of this engagement concerns adoption issues that are not discussed at length here. Nevertheless, when taken together, these three pillars not only refine the DIB-recommended goal of DOD leading international norm development, but also provide scope to align implementation with allies and partners-as is picked up in the section on the PfD below. In May 2021, the Biden administration reaffirmed the dedication to DOD's ethical principles and directed new actions for implementation with a focus on RAI. To facilitate comparisons with allies' approaches to AI ethics for defense, two taskings of the RAI Working Council are particularly relevant. The first is the broad, overall aim of leading responsible AI globally. To this end, creating a \"Responsible AI Ecosystem\" is one of the foundational tenets of RAI implementation.  32  Here, international components of this ecosystem include allies and partners to enable better multi-stakeholder collaboration and to also advance norm development \"grounded in shared values.\"  33  With the overall aim of leading responsible AI globally, this foundational tenet is the only one that explicitly names international engagement. Yet other foundational tenets are also relevant to assess convergences and divergences, including: governance structures for accountability; trust based on testing, evaluation, validation, and verification (TEVV); and a whole-oflifecycle approach to risk management.  34  The second relevant tasking is that the RAI Working Council is due to submit a report to the Deputy Secretary of Defense by late September on \"policy modifications to enable RAI considerations within existing supply chain risk management practices.\"  35  In particular, the introduction of supply chain concerns entails a broader focus than the DIB initially considered in its recommendations. The DIB did look at traceability of data sources, and DOD has also separately considered how reliance on foreignsourced parts may impact the resilience of its systems and ability to maintain a technological edge.  36  Yet the connection between RAI and supply chain risks is a newer question in U.S. documentation on ethical and responsible AI pathways. As described below, this is not necessarily the case for allies. From the DIB process through RAI implementation, these three phases show coherence in the U.S. approach to AI ethics in defense. Allies and key partners feature prominently across these three phrases, primarily as part of norm development and a shared policy basis for aligned, interoperable forces. • Defence Ethics Committee's Opinion on the Augmented Soldier (2020) and Opinion on the Integration of Autonomy into Lethal Weapon Systems (2021). The French Ministry of Armed Forces has been studying AIassociated risks since 2019.  37  That September, France became the first (and, still at the time of writing, only) European ally to publicly issue a dedicated military AI strategy.  38  Ethics and responsibility appear throughout the strategy, most notably as aspects of \"controlled AI\" and in the announcement to establish a ministerial Defence Ethics Committee.  39  Subsequently, as described below, the advisory opinions of the Defence Ethics Committee also lend insights into French views on responsibility and other concepts. Each of these building blocks is indicative of French thought and implementation pathways for ethical and responsible military AI, even if not in the form of adopted principles. Trustworthy AI is linked to robustness and security because the French identify the potential risk that humans may not trust critical systems with opaque or inexplicable results.  44  More specifically, they see risks of erroneous results from AI systems stemming from unrepresentative (biased) training data, \"malfunctioning algorithms,\" and \"insufficient understanding\" of system behavior. This leads to a technical approach to validating trustworthiness, wherein the Ministry decides on the level of trust and robustness necessary based on the function of the AI system at hand. This means that the ministry would conduct a risk analysis during the design phase to determine how critical the function is.  45  Then, according to the \"criticality\" designation, development of the system would need to meet a certain threshold of safety, security, and explainability in order to be validated.  46  Considerations about bias could also be taken into account, especially if they overlap with deception techniques.  47  Appendix III shows how the Ministry intends to categorize functions and decide on their corresponding thresholds; broadly speaking, safety-and mission-critical systems may require higher thresholds of validation to be deployed. The French anticipate operating not just in unknown environments, but often in communications-denied (\"degraded\") ones too, meaning that establishing operator trust in AI requires the systems to behave predictably amid situations they haven't encountered before, including interference and deception. 48 These operational specificities are heightened in the military realm. But the reliance on technical measures here echoes the national AI strategy, which similarly identifies explainability and the need to consider ethics from the design stage as important to ensure accountability and, by extension, social acceptability, for AI. 49 However, this approach is dependent on validation, and the military AI strategy does little to describe how current validation processes may be insufficient for AI systems that either perform differently in unknown contexts in which they are not validated, or whose capabilities change over their lifecycle.  50  The French answer is that AI standards and certification schemes are necessary, and that the Ministry of Armed Forces should be involved in such efforts. But without more detail, it is difficult to determine how trustworthiness is assessed after the design phase. This is where aspects of control come in, with the French military AI strategy establishing a relationship between controlled AI and auditability, to account for ethical concerns later in the AI lifecycle. The French military AI strategy only implicitly deals with the ethical aspects of data, but nevertheless highlights the need for data sharing to comply with regulations, security requirements, and appropriate use.  51  More specifically, data governance includes documentation practices (\"configuration\") to keep track not only of which algorithms are used, but also their \"learning elements, their combinations and their data\" (e.g., training data, weights, parameters, design procedures, annotations on limitations, etc.).  52  This means that the French see documentation as important not only to delimit the abilities of AI components in relation to their intended use, but also as an area where standardization can help humans exert control to be able to trace systems. Indeed, the While the geopolitical aspects and prospects of France to assert this independence are beyond the scope of this study, it is notable that they trickle into the French approach to trace the provenance of models and data. In particular, weapons are \"critical applications\" that will need to be auditable.  57  If enforced, this means that questions about data rights and legal authorities to transfer data (including from foreign suppliers) could render AI \"uncontrolled\" per the French definition. Here, protectionism straddles the line of ethics and adoption, with digital sovereignty as a potential factor that determines acceptability of both. This can also be seen in the imperative to maintain \"freedom of action.\" 58 Control and freedom of action may be seen as reinforcing concepts because they both relate to the need to maintain independence. Moreover, both concepts stress the responsibility of the state to ensure that humans are accountable for their use of technology. In the French strategy, responsible AI refers to human responsibility for continued adherence to legal obligations and the tradition of military ethics. Though only implicitly defined, this human responsibility is expected to be retained across all processes and institutional structures, with an emphasis overwhelmingly on the use of force and chain of command.  59  The focus on human command ensuring that force is used responsibly refers at least equally to AI and autonomous systems-the latter of which the French explored in greater depth in documents separate from the military AI strategy.  60  Most notably, as detailed below, the concept of responsibility is present in the advisory opinion documents of the ministerial Defence Ethics Committee. The establishment of the Defence Ethics Committee was announced in the French military AI strategy to ensure a continued emphasis on ethics, and to issue advisory opinions on emerging technologies in defense. Since being established in January 2020, responsibility has featured as an important theme in its initial mandate covering two key areas of focus: augmented soldiers and autonomy in weapon systems.  61  While AI fits into both issue areas, it is notable that AI itself did not merit its own categorization. Indeed, the advisory opinion on autonomy in weapons reiterates that aspects of AI like non-lethal decision support systems are beyond its scope, and such issues are not yet part of the Committee's agenda.  62  Nevertheless, the resulting advisory opinions both connect to the concept of responsibility because both human augmentation and autonomy in weapons indicate the clear priority of maintaining human centricity in decision-making. Both opinions of the Defence Ethics Committee stress the need for humans to maintain their decision-making capacities not only from a moral standpoint, but also for the sake of operational continuity. In both cases, the French positions on ethics related to emerging technology are conscious of the risks of over-reliance on new technological enablers. Be it preventing addiction to humanaugmenting medication and devices, or preventing automation bias in human-machine teaming, the Defence Ethics Committee makes clear that responsibility also requires humans to still be capable of achieving operational objectives even without assured access to the technology.  63  In short, responsibility means not only that humans should remain responsible and accountable for decisions, but if also considering the opinions of the Defence Ethics Committee on adjacent technologies, also that it would be irresponsible to excessively rely on technology. For the French, another concomitant part of responsibility is how technology changes the relationship between the operator, combatants, and non-combatants in-theater. On this point, the Defence Ethics Committee advisory opinion on autonomy in weapon systems also stresses that the distance of the operator from the operation itself can alter her or his judgment.  64  Such ethical tensions can have consequences for the foundations of responsibility in the French philosophical tradition. One question that senior French military officers have posed is whether it is unethical for two sides to face significantly uneven levels of risk, or whether military ethics require combatants on both sides to face risk.  65  To be sure, the operational advantage of reducing risks to the safety of one's own forces is imperative-and it could equally be argued that it would be unethical to not reduce risk on your own side when possible. 66 Without minimizing this obligation, it is worth noting that French ethics doctrine defines a soldier's own responsibility in part relative to the risk she or he faces. Per the Chief of Army's \"Exercise of the Profession of Arms: Foundations and Principles,\" French soldiers derive legitimacy from the state, which confers the \"responsibility to inflict destruction and death, at the risk of his life, in respect of the laws of the Republic, international law and the practices of war.\"  67  French military ethical debates around remote operators center on the phrase \"at the risk of his life,\" questioning whether imbalanced risk continues to fulfil the criteria of legitimacy.  68  As such, technology is introducing new questions related to identity-which could have spillover effects into views on AI ethics and responsibility in the French Armed Forces. This philosophical debate is implicit in the Defence Ethics Committee's opinions, and is relevant to adjacent questions about AI for two practical reasons. First is that French military officers see the need for guiding principles for technology that increases operational distance or can be applied to grey-zone activity. They note that ethics and law are well structured in rules of engagement, commander's intent, and compliance with legal frameworks for conflict. But because these often only apply above a certain threshold of hostility, guidance that allows the armed forces to maintain \"ethics and moral strength\" for other types of military activity is lacking.  69  This includes governance for technology that increases the distance between operators and operations, including in the information domain. As technology broadens and accelerates changes to the operating environment, taking these ethical and moral considerations into account could necessitate new \"deontological principles\" on duty and obligation for the French Armed Forces.  70  The concept of distance from operations enters into advisory opinions, both in relation to automation bias and psychological effects. But even with the advisory opinions of the ministerial committee, such a framework is still missing.  71  Second is that different allied perspectives on distance from operations could affect coalition operations if political differences widen. If guiding principles on distance from operations enter into rules of engagement or prompt questions about commanders' intent and responsibility, then different ethical bases for the use of new technologies in warfare could create political tensions. More specifically, if the more technologically advanced allies send more AI-enabled support and fewer troops as their contributions to coalition operations, some allies may perceive that others are not willing to equally share the burden of risks to life.  72  If not managed, sensitive issues that stem from different ethical risk calculations could decrease political cohesion. In sum, while France has not strictly defined principles that it can adopt to ensure safe and ethical AI in defense, the Ministry of Armed Forces has dedicated attention to the issue both in its conception of \"controlled AI\" and the aspects of trustworthiness, control, and responsibility that it entails. Furthermore, while it is notable that the new Defence Ethics Committee has gone straight to questions related to human augmentation and autonomy in weapons, rather than AI ethics as its own category, the implications of these AI-adjacent technology areas show a consistent emphasis of responsibility as key to the French articulation of ethical AI in defense. \n Australia Main documents from the Department of Defence (not official government position) include: Australia emphasizes pragmatic approaches to ethical risk management over the declaration of principles. This approach primarily draws from the Australian Department of Defence's Method for Ethical AI in Defence (hereafter Method) technical report. While not a formally adopted view of the government, it establishes tools to assess ethical compliance that are currently under internal review and are already being trialed.  73  Even as an opinion, the Method is the clearest articulation of ethical AI for defense among the Indo-Pacific allies. The Method document offers five \"facets\" for ethical AI in the Australian Department of Defence and armed forces, as well as an ethical assessment toolkit through which these facets can be operationalized. Overall, the five facets-responsibility, governance, trust, law, and traceability-echo the U.S. view that AI ethics for defense should focus on compliance with legal frameworks and moral obligations, as well as a functional approach to security and safety. While the Method is the result of a single workshop, the multi-stakeholder approach is similar to the longer DIB-led process in that it included experts from academia, industry, civil society, and bodies across the military and government. Moreover, the focus of the Method is on pragmatic tools that can be used to implement ethical risk assessments. As such, these \"facets\" are not intended as adoptable principles to disseminate throughout the Department. As stated in the report, \"rather than propose singular ethical AI principles for Defence, this report aims to provide those developing AI with facets of ethical AI that should be considered, including the questions to ask, topics to consider and methods that may be relevant to Defence AI projects and their stakeholders.\"  74  The focus therefore lies in implementation, as discussed below. The Method's aim is to provide all relevant stakeholders in the AI pipeline with practical risk assessment tools that treat ethical risk on par with other types of risk such as safety and security. This includes one open-source tool, the Data Ethics Canvas developed by the Open Data Institute, and three original tools from the Method to \"manage ethical risks.\"  75  These original tools are: an ethical assessment checklist; a risk matrix in Excel; and, lastly for projects deemed above a certain threshold of ethical risk, a formal documentation program called the Legal and Ethical Assurance Program Plan (LEAPP).  76  Together, these Australian ethical risk management tools are important not only to identify ethical risks, but also to follow up on them. More specifically, the tools offer a process to validate that contractors have indeed taken the ethical risks they identified into account in their design and testing prior to later acquisition phases. In this way, they are procedural risk-assessment complements to technical validation and verification measures for the Australian Department of Defence. These tools are calibrated to the level of ethical risk that the anticipated use of the AI system would encounter or exacerbate. In addition to identifying the risks, procedural checks would validate that the contractors have followed through on addressing them-also with the incentive that it would be a worse outcome for them to find that unaccounted ethical and legal risks delay later stages of development or compromise the acquisition. The incorporation of ethics in design through the acquisition lifecycle also intends to build trust in the process and, by extension, the systems by the time they go into service.  77  If high-risk projects and major weapon programs include AI components, then the risk assessment checklist would result in LEAPP being chosen as one of the ethical frameworks to comply with. LEAPP is a contractor's plan that seeks to give the Department of Defence \"visibility into the contractor's legal and ethical planning; for progress and risk assessment purposes; and to provide input into the [government's] own planning.\" 78 For instance, LEAPP could be used as part of an Article 36 weapons review-a legal review of new weapons, means, and methods of warfare mandated for states party to the 1977 Additional Protocol I to the 1949 Geneva Convention. 79 Article 36 reviews typically feature as part of legal discussions about autonomy in weapons, but nothing precludes them from use for non-autonomous, AIenabled weapons as well. In this light, part of the LEAPP assessment would focus on technical measures that relate to international legal principles, regardless of the level of autonomy involved, which would be included in early negotiations between the contractor and the Australian Department of Defence. 80 To be sure, compliance with the legal principles is non-negotiable. But in a more \"iterative\" fashion, the negotiations would focus on which requirements (e.g. software safety plan, safety management plans, human factors plans) would be necessary for the system to be certified for legal reviews, ethical requirements, as well as social responsibility more broadly.  81  As a result, this means that ethical risk would be identified in the early phases of design, and then carried through in TEVV. This negotiated framework between contractor and government is reserved for the higher-risk applications, and the connections it establishes between risk assessments and requirements validation is one of the most concrete practices that U.S. allies have thus far developed for AI ethics implementation in defense. While the Australian focus is on tools rather than principles, the tenets in the Method are important bases for stakeholders to identify and assess ethical risks, including what frameworks AI developers consider in their own design processes. To this end, two aspects of the Australian tenets are worth briefly discussing. First, for the trust tenet, the topic of \"contestability\" is unique to the Australian approach. The Method recommends importing the Australian government's civilian ethical AI principle of \"contestability\" into the military domain.  82  The Australian AI Ethics Principles define contestability as a \"timely process to allow people to challenge the use or output of the AI system\" that should be available for cases \"when an AI system significantly impacts a person, community, group or environment.\"  83  Examples that apply both to the civilian or defense realms could include enterprise AI or recommender systems for promotions and human resources decision support. The aim of contestability is to ensure public trust by allowing redress for harm, but the threshold of \"significant impact\" is fluid. As such, there is little guidance on what constitutes contestability.  84  If the Australian Department of Defence does adopt or enforce the contestability concept, however, then the high-risk nature of military activities could mean that the threshold of this harm is clearer than is the case in the civilian realm. Overall, this intentional overlap between civilian and military is not unique to AI ethics tenets, and indeed builds on other Australian legal and cultural commitments to accountability. Legally, for example, there is no military court in the Australian Defence Force's disciplinary regime.  85  Culturally, the Australian \"mateship\" ethos bridges military history with national identity in a way that makes equality and respectful disagreement with authority-a basis for contestability-part of the Australian Defence Force's strategic culture.  86  In this way, building trust is not just about the technical measures like TEVV for reliability, but also about how individuals affected by AI can trust in the processes and structures responsible for its development and use. Secondly, the lawful AI tenet is also worth highlighting because its explicit focus on enhancing compliance with international humanitarian legal principles is reflected in current Australian military AI consultation and collaboration with industry. The main point of lawful AI is that the technology is introduced into an existing legal and ethical framework-and that human-centered uses of AI should produce more ethical, and better humanitarian, outcomes.  87  In other words, lawful AI is not just about complying with the bare minimum of legal obligations, but also improving the standards of compliance already in place. The Australians point to two associated topics that lawful AI seeks to reinforce: \"protected symbols and surrender\" and \"de-escalation.\" On this point, the focus on protected symbols in lawful AI is also worth highlighting because it can currently be seen in Australian industry. Current Australian military AI development includes AIenabled decision-support systems that improve compliance with international legal principles. Conceptually, this applies to what Australian researchers refer to as Minimally Just AI, or \"MinAI,\" systems. Research on MinAI is not driven or mandated by the government. Still, it corresponds to the results from the Defenceled ethical AI workshop in that it seeks to identify \"protected symbols, protected locations, basic signs of surrender (including beacons), and potentially those that are hors de combat\" so that human operators avoid them.  88  In practice, a small Australian company is currently using the Method as a guide for its system, called Athena AI, which identifies and classifies objects that \"must not be targeted for legal or humanitarian reasons,\" including battlefield hospitals and other protected sites.  89  This includes validating data and designing features for end-users to understand the limits of the AI system, such as a notification if the computer vision system is less confident about a classification, so as to prevent cognitive bias.  90  Using the tools, scenario development, and a series of workshops on ethics and legality, Athena AI was designed in tandem with a 70-page legal and ethical framework.  91  As is noted below, the pragmatic tools from the Method could similarly inform Five Eyes activities. Overall, the lawful AI tenet and focus on ethical risk assessment tools, rather than principles themselves, is a different packaging of the legal framework in which all allied democracies situate their views on ethical AI in defense. The Method contends that \"lawful AI\" has \"no equivalent\" principle in its comparison to other ethical AI frameworks.  92  In part, this is a superficial difference, as abidance to law is a preamble and is embedded in principles themselves, including those that DOD adopted. Signaling-wise, this more direct emphasis on law may intend to connect assessments of ethical risk with reputational risk. Risk management is just as much an entry point to ensure ethical AI efforts correspond with concrete, existing practices in the Australian Defence Force, as it is a form of responsible state behavior.  93  Contestability is one example of thisas a mechanism that seeks to enhance responsiveness to democratic citizenries and enhance social acceptability of AIincluding here for defense. For military AI cooperation with Australia, it will also be important for allies such as the United States to know whether cooperative activities can be subject to formal contestation procedures. Emerging National Views: the U.K. and Canada While France and Australia have issued public approaches to their views on ethical AI as part of adoption, it is possible to deduce the approaches of the U.K. and Canada, both key allies, based on initiatives they have led on AI in defense, ethics, and data governance. The U.K. Ministry of Defence and Canadian Department of National Defence positions are considered emerging based on the availability of information in the public domain, which indicate strong foundations for more articulated approaches. The U.K. Main public documents from the Ministry of Defence include: • Defence Science and Technology Laboratory Biscuit Book 2020: Building Blocks for AI and Autonomy; • The announcement of a forthcoming U.K. Defence AI Strategy. In the U.K., ethical and normative aspects of AI feature in recent strategic documents, including the government's Integrated Review of national security and international policy, and in the Ministry of Defence's accompanying Command Paper published a week later in March 2021. The Integrated Review names \"supporting the effective and ethical adoption of AI and data technologies\" and \"identifying international opportunities to collaborate on AI R&D, ethics and regulation\" as aspects that can help build public trust and early adoption of military AI.  94  This is consistent with the Ministry of Defence's contributions to achieving the British strategic interest of \"the ethical development and deployment of technology based on democratic values,\" as reaffirmed in the Command Paper.  95  One area of daylight between the two documents, however, is the Integrated Review's concern about the gap between the pace of global governance and the development of standards and norms, in contrast to the Command Paper's stated need for \"standards and norms for the responsible and ethical adoption of these new technologies.\"  96  How exactly the U.K. Ministry of Defence will approach these interrelated military governance challenges is due to become clearer in the near future. More specifically, the U.K. plans to establish a new Defence AI Centre in order to centralize its AI developments.  97  Further, the U.K. Ministry of Defence is planning to publish a Defence AI Strategy that will incorporate ethical adoption considerations.  98  A ministerial AI ethics committee is also currently analyzing AI in defense, including issues related to trust.  99  In terms of oversight, both the new Defence AI Centre and this committee are important developments to bridge ethical AI endeavors at the working level with a higher degree of political and strategic attention. The U.K. approach to military AI adoption includes a process for developing guidelines on ethical AI, which includes public-facing aspects led by the Defence Science and Technology Laboratory (Dstl).  100  Dstl established an AI Lab in 2018, which has made it the natural home for technical questions related to ethics, risk, and safety concerns.  101  While few details of the ministerial AI ethics committee are available at the time of writing, Dstl's activities advancing AI ethics in defense provide an indication of the U.K. approach. For instance, Dstl sponsors an ethics fellow at the Turing Institute to focus on \"improving robustness, resilience, and responses of systems that support logistical, tactical and strategic operations, as well as wider applications in urban analytics, cybersecurity and social data science.\"  102  Furthermore, in 2020, they also hosted a conference that focused on safety, robustness, trustworthiness-which is part of the process on creating ethical guidelines for military adoption of AI.  103  Notably, the Dstl AI Lab also debuted a \"biscuit book\"-a kind of introductory guide for contractors-that covers AI safety and ethics. Although predominantly a technical research organization, the building blocks in this book extend to non-technical considerations that should be factored into development of AI and autonomy in weapons. Two of the building blocks-consent and confidence-are relevant here and are described in more detail in Appendix V. While consent comprises ethics, risk and policy appetite, and legality, confidence includes more of the safety and security measures that are part of responsible military AI.  104  Nevertheless, although the questions are a helpful guide intended for the Ministry of Defence customers, no mechanism currently obliges AI stakeholders to follow the guidance.  105  As a last note, one question that the Defence AI Strategy may answer is the extent to which the U.K. considers ethics as important to its data governance efforts. One aspect that is less clear in the U.K. is how defense data governance accounts treats data ethics. So far, both in the biscuit book and more importantly the 2021 revision of the Defence Data Strategy, ethics does not feature as part of the imperative to govern data for aspects such as traceability and auditability. Furthermore, ethics is not included as part of the data governance architecture within the Ministry, which encompasses a defense information steering committee and specialist boards.  106  With this in mind, there are important questions to answer in the forthcoming Defence AI Strategy and more information from the ministerial ethics committee. Nevertheless, when the U.K. publishes more information on its approach to ethical AI in defense, it will have strong national foundations to build on. \n Canada The main public document (not official government position) from Department of National Defence (DND) is: Canada is also working through data governance, an area that likely falls under the \"privacy, confidentiality, and security\" category in the Military Ethics Assessment Framework. To this end, data ethics are considered as part of the guiding principles of the DND Data Strategy principles, but with little clarity on their implementation. More specifically, there are guiding principles on managing data ethically \"throughout their lifecycle to eliminate bias, ensure fitness for use, and adhere to the Code of Ethics.\" Security and trust feature as related concepts in the accompanying principles.  111  Although this approach declares ethical management as a lifecycle-long principle to abide by, the Defence Data Strategy only considers ethics for the use of data in its definition of the data lifecycle. This gentle contradiction becomes more relevant when considering that the Strategy does not include a dedicated line of effort to implement this guiding principle. Further, ethics is considered part of data literacy-which ranks as low or medium in the implementation priorities, rather than a central feature of the governance framework.  112  The DND Data Strategy represented the first effort by Canada to create a defense data governance framework, and still does focus on ethics-adjacent considerations related to aspects like traceability, data quality, accountability, and security. DND is expected to fit into broader Canadian government views on these issues as well-and has already encountered difficulty in the data governance realm for inadequately protecting privacy as part of an effort using AI to improve workplace diversity.  113  The usage of AI in human resources may indicate the difficulty of instituting a data governance framework-let alone one that explicitly calls out the ethical issues at hand.  114  As DND is subject to the Directive on Automated Decision-Making and Canadian Algorithmic Impact Assessment, incorporating values into data governance could become an increasingly important area for accountability.  115  In this way, the government's expectation of Defence abiding by the civilian framework is worth noting as a point of differentiation from the United States, whose leadership on ethical AI in defense took hold amid a less extensive background of civilian digital legislation. All in all, these emerging approaches to ethical and responsible AI in the U.K. and Canadian defence contexts are promising starting points for more robust engagement. \n Nascent National Views: Netherlands and Germany Many other allies have not issued public approaches to responsible and ethical military AI. As is detailed below, these countries may establish their views in multilateral formats, rather than articulating a national approach. This includes countries like the Netherlands, whose approach thus far echoes those of allies mentioned above, as well as Germany, albeit in the adjacent context of multistakeholder AI governance for international security. Their interest could also lie more in international-security norms and rules in the international technology order, and are therefore considered nascent here based on their comparability to the U.S. position. \n Netherlands Main documents from the Dutch defense sector include: • The Netherlands Organisation for Applied Scientific Research (TNO) technical report Artificial Intelligence in the Context of National Security-Final Note Study within the National Security Analyst Network (2020); • A potentially forthcoming defense AI roadmap from the Ministry of Defence. The Netherlands is often cited as one of the European allies with a more fully fledged approach to AI in military affairs-including in the ethical domain. Indeed, the Dutch Ministry of Defence has sent a foreign exchange officer to the JAIC to work on collaborative approaches to responsible AI. A Dutch defense AI roadmap or \"vision\" document has been in the works for some time, as mentioned in the Dutch government's Strategic Action Plan for AI.  116  Parliamentarians have also recently inquired about the ethical and legal aspects of NATO's approach to emerging and disruptive technologies (EDTs).  117  The Dutch defense research agency TNO also issued a technical report assessing experts and policymakers' view on AI risk in national security-including risks linked to control, trust, overdependence, and error.  118  The report is not directly comparable as it denominates risks, rather than coming up with a framework for stakeholders to mitigate those risks. For instance, security concerns around control being overridden, inaccuracy, errors, and interference do overlap with principles like reliability and governability, albeit in this slightly different context. Relatedly, even though it does not list exact risk factors that align with U.S. principles like responsibility, equitability, and traceability, this does not mean that the Dutch are not considering these principles more implicitly and in other contexts such as the defense AI roadmap. On the other hand, other risks that the TNO classify in the report do overlap with the ethical approaches of other allies here-such as reputational risk for domestic and international accountability.  119  How these risks translate into ethical guidance for the Ministry of Defence and other national security stakeholders remains to be seen. 120 \n Germany The main document (non-governmental) is: • Airbus-Fraunhofer Institute for Communication, Information Processing and Ergonomics (FKIE) joint \"FCAS Forum\" Working Group's White Paper on The Responsible Use of Artificial Intelligence in FCAS -An Initial Assessment. The German approach to responsible military AI requires different considerations than the other allies surveyed here for two reasons. First, because the AI governance for international security policy space involves different primary actors. Second, because German political attention to autonomous weapons overshadows other aspects of military AI to an arguably greater degree than is true of the other allies surveyed here. German debates focus more on targeting-both for drone procurement as well as autonomous weapons development. The ban on LAWS, for example, is written into the government's coalition agreement.  121  The focus on autonomy in weapons is not matched by information about the ethics and safety of non-lethal, defensive applications of AI and AIenabled decision support systems. At the national level, it is the Federal Foreign Office, not the Federal Ministry of Defence, that has ownership over the policy space of AI governance and international security. The incorporation of the arms control agenda into military AI governance takes a broader definition of safe and ethical AI that incorporates questions related to diffusion and proliferation into account. This may be because arms control is a non-controversial topic that promotes German security policy without touching on more sensitive questions related to military adoption of technology.  122  As several analysts have noted, it is near-impossible to talk about military AI in Germany without the conversation immediately turning toward LAWS.  123  Further, the focus on arms control lends itself to Germany's normative and diplomatic strengths. The framing of German contributions to responsible military AI in the realm of arms control is asynchronous-but not incompatible-with other countries' approaches to this policy area. What this means is that the Federal Ministry of Defence is left with a backseat role.  124  This also heightens the stakes of multilateral efforts on responsible and ethical military AI, including for assessments of ethical risk stemming from issues like explainability or reliability. Indeed, Germany may be more active in these formats, especially in facilitating coordination between the EU and NATO given its longstanding interest in encouraging and facilitating EU-NATO cooperation.  125  Cooperation is already visible in other efforts related to technology and ethics-most notably in that the German Bundeswehr Defence Policy Office came up with views on future implications of human augmentation in collaboration with the U.K. Development, Concepts and Doctrine Centre.  126  The two countries share views on the future of operations, which may be productively channeled through activities related to policy alignment or, potentially, standardization.  127  Rather than going it alone, the German preference to cooperate-in bilateral and especially multilateral formats-may be seen as one way to focus on these issues with less domestic political pressure, and to substantiate contributions to defense partnerships. Outside of government, responsible military AI initiatives are not necessarily on standby. For example, Airbus has partnered with the Fraunhofer FKIE research institute to ensure ethical compliance for AI and autonomy incorporated in the Future Combat Air System (FCAS) program.  128  FCAS is a cooperative project between Germany, France, and Spain to co-develop what they call a sixthgeneration manned-unmanned aerial \"system of systems.\" The Airbus-Fraunhofer FKIE partnership, dubbed the FCAS Forum, is supposed to \"guide the development phase of the FCAS project from an overall societal and explicitly ethical perspective\" to provide designers and developers with \"ethical requirements and a process model for their implementation.\"  129  The engineers involved have alternatively described this as \"ethical and legal compliance by design,\" which includes not only legal compliance, but also considerations of \"social acceptance.\"  130  Representatives from the two organizations see \"technical implementation\" as necessary to uphold the framework they develop. To complement the technical aspects of responsible military AI, the FCAS Forum also includes a multidisciplinary expert panel that focuses on related aspects to responsible use.  131  While the panel includes representatives from the Federal Foreign Office and Federal Ministry of Defence, it is not a government-driven process. Further, at present, the expert panel is entirely German, and it is yet unclear if this industry-led initiative will echo the government's preference for cooperation by including French or Spanish subjectmatter experts in the future. One effort that the FCAS Forum has undertaken is a study on how civilian EU tools for AI ethics apply to the FCAS system-of-systems program. Specifically, the White Paper entitled The Responsible Use of Artificial Intelligence in FCAS-An Initial Assessment honed in on the EU Assessment List for Trustworthy AI (ALTAI) methodology in its attempt to identify frameworks that could encompass the societal and ethical implications of FCAS development.  132  Although EU AI policy, including this methodology, does not apply to high-risk designations in the defense realm, Airbus engineers involved in FCAS write that the ALTAI methodology \"raises questions which are mostly applicable for FCAS and demonstrates consequences for the design and the operation of the system.\"  133  In other words, while the ALTAI methodology may not address uniquely military ethical concerns that weapon systems introduce, they find it nevertheless offers a starting point for a framework that defense stakeholders can consider.  134  One question that the White Paper opened was whether the EU ALTAI methodology should be \"extended or tailored to defence applications.\"  135  Though an answer to this question is beyond the scope of the report, the preliminary conclusions are that introducing the ALTAI methodology to operational and engineering teams on a case-by-case basis may help determine whether and how the methodology can be tailored to defence applications.  136  Further, they recommend that there should be training to implement ethical design recommendations for relevant specialists.  137  They also suggested that future work could assess other requirements of trustworthy AI-as is partially addressed in Appendix VIII in this study. To apply the EU ALTAI methodology to FCAS, the engineers first identified which \"major AI case studies\" would include their own discrete questions on responsibility and ethics. Their White Paper identified eight: mission planning and execution; target detection, recognition, and identification; situational awareness; flight guidance, navigation, and control; threat assessment and aiming analysis; cyber security and resilience; operator training; and reduced lifecycle cost.  138  Of these, the Airbus engineers deemed the latter three as \"relatively uncritical from an ethical point of view.\"  139  If compared directly to the U.S. approach, there are several concerns that the DOD would deem relevant-or indeed critical-to safe and ethical AI. In particular, security and resilience are typically qualities that help build trustworthiness and reliability in AI systems. The FCAS White Paper does acknowledge that there are risks related to bias, deception, and the possibility of \"AIgenerated analyses and inferences [gaining more authority] in political decision-making.\" At the same time, the German government's policy attention to AI governance and international security should not be discounted when it comes to norm development. As the JAIC's international engagement focuses on norms for a favorable technology order, there is room for the United States to learn from the German emphasis on arms control. In this vein, norms around the prevention of undesirable diffusion and confidence-building measures can be seen as part of responsible military AI.  141  Germany is already focusing on these areas, seeing multistakeholder engagement as necessary to modernizing 21 stcentury arms control beyond a narrow definition of state-to-state treaties. With industry already attempting to forge its own path on its ethical obligations, the Airbus-driven process could be seen as a complementary attempt to multi-stakeholder responsible AI in defense. But whether this delegation of labor on responsible AI to the private sector is a delegation of responsibility will depend on multilateral formats. \n Multilateral Institutions Focusing on Ethical and Responsible AI in Defense Understanding ethics and legality as part of the adoption of emerging technologies is not only a priority for democratic countries, but also a topic of interest for multilateral institutions that are part of the security and defense architecture. Autonomy in weapons has been on the agenda at the UN level for longer than the countries highlighted here have spent time bridging technical and policy approaches to responsible AI in defense. Select multilateral institutions are also critical for consultations and alignment about these issues. Beyond those mentioned here, several smaller allies may be waiting for multilateral views to then drive their own approaches to RAI, rather than dedicating resources to first issuing national views that will later have to align with broader multilateral structures. NATO is an obvious player in this domain, for reasons that are described below. The emerging PfD is important as an AI-specific multilateral format for like-minded countries-including non-treaty partners-to coalesce on this policy area. \n North Atlantic Treaty Organization (NATO) NATO is an important actor because it can help coordinate and facilitate consultations between allies to come to agreement on how ethical and responsible AI developments impact interoperability, cohesion, and operations. Further, the NATO Defence Planning Process is the primary defense planning tool for many Allies.  142  The focus on principles for \"responsible use\" is consistent with NATO's added value without dwelling too much on development-which happens primarily at the national (or bi-/multilateral levels outside of NATO).  143  Here, responsibility refers both to best practices in engineering (e.g., ethical design) and responsible state behavior.  144  The North Atlantic Council and Military Committee-the senior civilian and military decisionmaking bodies in NATO-began contending with EDTs, including AI, in 2018.  145  This high-level interest built on several years of military and scientific experience at the working levels, as well as the introduction of conceptual and operational considerations in the workstream of Allied Command Transformation and the NATO Science & Technology Organization in the 2010s. When presenting on EDTs to the senior civilian and military leadership at NATO in 2018, then-Supreme Allied Commander Transformation General Denis Mercier stressed that legal, ethical, and political differences between Allies could \"endanger our capacity to operate together.\"  146  He also focused on considerations around the \"level of confidence in new technologies\" as an adoption factor that is not purely technical.  147  This set the tone for political alignment on ethical concerns across the Alliance-already building on the foundations of the shared value embodied in the North Atlantic Charter and legal framework in which NATO operates. Subsequently, in October 2019, the Allies agreed to an EDT Roadmap that cited \"legal and ethical norms\" and \"arms control aspects\" as key technology areas among Alliance priorities to consider.  148  The political will to cooperate on technologies was solidified in February 2021, when NATO Defence Ministers endorsed an EDT Strategy.  149  As part of the implementation of the EDT agenda, the NATO AI Strategy is expected to pick up on this theme in \"guidance on both principles for responsible use of AI-enabled platforms and export control mechanisms.\"  150  Accountability and transparency-for weapon systems with varying levels of autonomy, as well as AIenabled systems-and rules for industry may feature in this approach.  151  Already, the NATO Science & Technology Organization has identified a \"strong emphasis on explainability, trust and human-AI collaboration\" as well as \"processes and standards for verification, validation and accreditation\" as areas of interest for NATO, though they are not yet formalized in publicly pronounced principles.  152  Overall, the issuance of principles at the NATO level will be helpful to reflect priorities of multiple Allies, permit more Allies to align nascent or ad hoc initiatives to the broader agenda with one another, and potentially help bridge responsible AI with best practices or standardization that industry can follow.  153  Technical reports that pre-date this EDT agenda also help establish a baseline for the Alliance's approach to responsible and ethical military AI. This includes NATO reports on human centricity and human control, as seen in Appendix IV.  154  The NATO Human View features in a technical report from 2010, and is notable as it commends a human-centric approach that has become popular in more recent civilian AI ethics frameworks (see Appendixes I and VIII). It is also an implicit theme found throughout the approaches of all democratic countries who see the need to maintain accountability in human-defined frameworks to properly manage AI and the risks that can come with its development, use, and diffusion. The NATO Human View framework takes this a step further by developing a human-centric approach to networkenabled operations (i.e., network-centric warfare) that ensures the human element is incorporated into capability development.  155  More specifically, NATO defines this human centricity by \"depict[ing] how the human impacts performance (mission success, survivability, supportability, and cost) and how the human is impacted by system design and operational context (e.g., personnel availability, skill demands, training requirements, workload, well-being).\"  156  The Human View framework also applies the element to factors of human-system integration as well as \"guidance on use of models to address uncertainty and/or discover emergent behaviors.\" 157 NATO work on human control is largely focused on integrating weapon systems with different levels of autonomy, but still offers useful classifications that can apply to other EDTs like AI. As seen in Appendix IV, it establishes a framework of who the relevant stakeholders are, and what their corresponding roles are in delivering appropriate human control-a concept that could be adjusted to address other aspects of its forthcoming principles for responsible use. Some of these complexities are specific to autonomy and the use of force, as was defined in 2019. But it is also notable here that the NATO working group on \"Human Systems Integration for Meaningful Human Control over AI-based Systems\" was established in 2020. Importantly here, the working group suggests using the term \"meaningful human control\" outside of autonomy in weapons-related questions, which are primarily dealt with at the Convention on Certain Conventional Weapons. They suggest applying the term to address questions of accountability and human agency in a broader set of applications including intelligence, surveillance, and reconnaissance, planning, and decision support.  158  In addition to applying to these set of issues, the working group focuses on many similar aspects that also relate to responsible AI implementation, including: \"national or organisational policy, systems specification, systems design, For NATO, building on General Mercier's remarks that ethics and interoperability are linked, standardization on safe and ethical AI can likewise link to interoperability at the technical and procedural levels.  161  AI Partnership for Defense (PfD) In September 2020, the JAIC convened the inaugural PfD meeting, featuring virtual delegations from Australia, Canada, Denmark, Estonia, Finland, France, Israel, Japan, Norway, the Republic of Korea, Sweden, the U.K., and the United States to \"shape what responsible AI looks like.\"  162  As of May 2021, three additional countries joined for the third PfD meeting: Germany, the Netherlands, and Singapore. As the grouping of countries makes clear, the ability to include non-treaty allies in the PfD makes it a useful format to borrow from each other's approaches to RAI, be it to establish, refine, or implement nation-level views. Just two months before joining the PfD, for instance, Singapore prepared \"preliminary guiding principles to be applied to the defence establishment in Singapore, and Singapore's contributions to the global discussion on international norms for defence AI applications\" in March 2021.  163  Further, there may also be the possibility of taking aspects of responsible military AI from other countries that focus more on norms of responsible state behavior on board in the PfD format. Some allies explicitly mention a focus on norms, including the U.K. in its new national security and international policy, and Germany via its focus on arms control and emerging technologies. This normative emphasis harkens back to the U.S. approach to responsible and ethical AI in defense-which saw norms as one of the primary areas of engagement with like-minded countries. This normative focus could also benefit engagement with allies that have not yet begun any public iteration of views on responsible military AI, including Japan and South Korea. As such, the PfD's focus on responsible AI makes it an important venue to encompass technology norms that are based on democratic values and that focus on minimizing risks in the international security environment. As a final note, it is not a coincidence that all allies surveyed here participate in the PfD. It is an important forum for them to exchange views-not only on aspects covered in this report, but potentially also the impact of civilian AI ethics frameworks and developments, as well as questions about autonomy-related aspects of human-machine teaming. \n Other Opportunities for Multilateral Collaboration on Responsible and Ethical AI in Defense: the European Union (EU) and Five Eyes Working with a number of multilateral institutions is critical to the United States' stewardship of AI aligned with democratic values and interests.  164  In addition to NATO and the PfD, the EU and Five Eyes are highlighted as relevant formats for cooperation on ethical and responsible military AI. \n European Union Of course, the EU is not an alliance-and the United States is not a member. But the EU's potential contributions to responsible military AI are worth discussing here because of the implications of supranational EU policy on allies' own approaches to ethical and responsible AI in defense, as well as on EU-NATO cooperation.  165  Furthermore, the United States could also be more directly affected by EU policy as a third state that is eligible to receive R&D funding via the European Defence Fund. This is because the European Defence Fund, under the European Commission's authority, can require a mandatory ethics screening prior to fund disbursal.  166  While the EU has adopted a bullish approach to trustworthy AI in the civilian realm, European institutions have been slower to define the implications for safe and ethical AI beyond the tip of the spear. Key civilian policies and regulations, like the General Data Protection Regulation and more recent legislation instituting the European approach to \"trustworthy\" AI, have clear carveouts for public safety, security, and defense. Still, the EU approach to civilian AI policy is relevant to transatlantic defense because the dual-use, general-purpose nature of AI means that military adoption of AI will depend on the ethical frameworks that dominate civilian development, regardless of carveouts. More directly, some European countries also choose to apply EU legislation like the General Data Protection Regulation to their own defense sectors, even though they are not required to do so.  167  With this overlap in mind, Appendix VIII overviews the applicability of the EU trustworthy AI principles for the defense realm. In addition to examples such as Airbus' application of the ALTAI methodology to FCAS, it is notable that several European defense efforts mention the European Commission-supported guidelines for trustworthy AI as a positive step toward ensuring military uses of AI adhere to ethical standards. For example, in their co-authored food-for-thought paper on AI in defense, Finland, Estonia, France, Germany, and the Netherlands made explicit reference to the Trustworthy AI Principles, recognizing that the EU could leverage its normative power because of the centrality of ethical standards in AI for defense.  168  The focus on safety and security in EU AI policy also promotes \"convergence between the AI community and the security community\" to enhance robustness.  169  In sum, the emphasis on safety, security, and risk in EU AI policy is not only a natural overlap, but also one that European defense stakeholders are seeking out. However, it remains to be seen which EU body will take control of the responsible and ethical military AI agenda. There are various actors within the EU institutions that are largely beyond the scope of this paper.  170  Instead, there are only inklings of how the EU will approach responsible and ethical military AI at present. In the future, this topic could also feature in EU-U.S. security and defense dialogues. For Most directly, the TTCP AI Strategic Challenge includes a Law and Ethics Working Group and also considers Trust and Transparency as one of its key themes.  176  The Trust and Transparency theme intends to better understand how agent transparency impacts human-system performance, and how to design the appropriate level of transparency in AI systems to increase trust in their intended use.  177  This lends itself well to responsible AI, for example, by applying ethical risk frameworks and tools from the different countries to cooperative activities. Already, the Australian Method includes recommendations to demonstrate \"how the AI integrates with human operators to ensure effectiveness and ethical decision-making\" in trials and exercises that simulate the \"anticipated context of use.\"  178  Five Eyes countries have previously cooperated on areas that could link to this recommendation. In addition to sharing data for Project Maven, the TTCP Autonomy Strategic Challenge used the 2018 Autonomous Warrior exercise to trial the \"Allied Impact\" C2 software system focusing on humanmachine teaming and integration of autonomous assets. Although it focused on cyber rather than AI, the four topics of its focus-vulnerability assessment, red teaming, building mixed levels of trust systems, and developing metrics and measurements of trustworthy systems-offer a framework that responsible AI adoption could follow.  180  On the latter point, the TTCP working group developed an ontology-based assessment framework that formalizes metrics and attributes that constitute trustworthiness, which could inform allied approaches to trustworthy, reliable, and secure systems.  181  The Five Eyes researchers identified 13 core attributes, many of which dovetail with the broadly shared principles of ethical design and responsible use for AI: reliability; availability; safety; confidentiality; integrity; robustness; maintainability; adaptability; usability; timeliness; leanness; reactiveness; and proactiveness. Appendix IX details how these attributes are measured in relation to trust, resilience, and agility. These attributes from cyber are relevant to AI because the TTCP group looked at factors that would affect the performance of the systems, including cyberattacks, human factors, and the impact of the physical environment.  182  For AI, equivalent concerns include protecting from failure modes, understanding how human interactions impact system performance, and validating and verifying the performance of AI systems in real-world conditions. Going forward, the trustworthiness framework that the TTCP Cyber Strategic Challenge established could also be used for the Five Eyes countries to understand how to use their allies' systems.  183  Overall, because TTCP Strategic Challenges already combine technical research with experimentation, they are well-suited for implementation of responsible AI and ethical assessment frameworks. \n Areas of Convergence and Divergence For the most part, U.S. views on responsible and ethical AI for defense align with allies. Table  1  reviews the similarities, with the caveat that the absence of a similarity or equivalent principle should not be read as a point of divergence. As the country analyses above have shown, countries are at different stages of iterating and implementing ethical and responsible military AI. In particular, Australia, Canada, France, the U.K., and the United States share responsibility and trustworthiness as core principles for ethical AI design, development, and deployment. While the United States does not use trust or trustworthiness as a standalone principle, it is embedded in all five of the DOD principles. This is similar to the U.S. conception of law as a crosscutting theme that is implicit in responsible, equitable, traceable, reliable, governable AI. Other countries see trustworthiness as the overarching principle that comprises reliability, along with integrity and security. In this way, even though trust is not synonymous with reliability, it is the closest comparison. This is also affirmed in other countries' views: for instance, the Australian Method benchmarks its own facet of trust to the DOD principle of reliable AI.  184  These terms are interlinked and mutually reinforcing, as human-machine teams require operators to trust the systems they interact with. Focusing on reliability and security can help build that trustworthiness, especially to show operators that the systems are subject to rigorous testing and that processes are in place to appropriately calibrate trust depending on the capabilities of the system and the operating environment.  185  Other principles and equivalent topics, such as controlled AI and feedback mechanisms for societal input, are also more explicitly laid out in allies' approaches to AI ethics. Here again, this is not to say that the United States does not share sovereignty concerns, but rather that these concerns are not as explicit in DOD's ethical AI principles as in allies' documents. Seeing national sovereignty as part of responsibility could come to be in tension with cooperation-as well as procurement decisions that breed dependence on the United States. It is notable that DOD is just beginning to insert supply chain considerations into its publicly available documentation on RAI. Meanwhile, it has been part of the French approach since they began considering ethical risks of AI in defense, and is also included in the Australian Method. As countries navigate this nexus, the extent to which sovereignty concerns fuel tensions between democratic allies will depend on other forms of cooperation.  186  Nevertheless, because of the overlap between security and assurance of control over the lifecycle of an AI system, responsible AI implementation pathways in the United States may come to incorporate supply chain risks.  187  In this way, it would be similar to traceability and auditability concerns that countries such as France and Australia mention in their approaches to sovereignty in AI. Additionally, differences between allies' views on responsible and ethical AI in defense may also stem from the extent to which other countries apply civilian AI policy and regulation frameworks to their own defense approaches. Although overviews of government principles and policies for ethical civilian AI are not discussed at length here, they become more visible in other countries, as well as the EU. Some of these aspects are encouraging-for example, German industry's voluntary compliance with trustworthy AI principles, to the extent they overlap with defense, due in part to the fact that there is no equivalent defense framework they can follow and implement themselves. Other entry points of civilian concepts into the defense realm include the Australian concept of contestability, Canadian DND subjectivity to the algorithmic impact assessment, and the choice of some EU countries to apply the General Data Protection Regulation to their defense sectors. These are both related to accountability as well as privacy, which are key differences that should be tracked even though allies overwhelmingly agree on the importance of ethics and safety for AI in defense. Still, similarities between defense stakeholders include the view that militaries could not only inject new risks into their operating environments, but also expose their own organizations to risk if they leave ethical and legal questions unaddressed. Countries may have different ways to define and measure these ethical risks, as Table  1  implies, and as is detailed in the appendices. Overall, though, there is an implicitly common approach which recognizes that they must contend with the associated technical, legal, political, and moral risks from the front end of AI development. More concretely, they also agree that the way to implement responsible AI involves technical measures tied to safety and security, as well as procedures that make the legal context by which they abide as clear as possible. How countries actually implement their shared technical and lawrelated priorities depends not on the principles or tools, so much as the oversight structures that are responsible for them. It is too early to assess differences in oversight structures, which will depend on the authority empowered to the ethics committees and AI-centric defense units that ministries are creating. Many of the documents overviewed here are technical reports or advisory opinions that are not binding in nature. Whether these structures have mechanisms to mandate, or merely recommend, courses of action will also have implications for the extent to which they transpose their frameworks into action. As countries focus on implementation, their approaches to oversight may also emerge as differences between the structures working on ethics and responsibility of technology adoption. Overall, pathways for leveraging shared views to advance the implementation of responsible AI should include learning from allies whose views emphasize both human responsibility and responsible state behavior extending beyond minimum legal obligations. Addressing these issues is not only a question of good engineering practices, but is also an exercise of responsible state behavior.  188  In a narrow sense, responsible AI can refer to ensuring that AI systems enter into human-centric frameworks that are defined by humans to maintain human agency and responsibility. More broadly, though, it is notable that some allies see preserving freedom of action as part of a vision of responsible AI that encompasses responsible state behavior. Having a legitimate basis for military action is a feature of responsible state behavior, with civilian government oversight of militaries at its core to maintain accountability at home and abroad. This enters into the language of responsibility because operating in coalitions under multinational mandates can also confer international political legitimacy to operations.  189  Allies with articulated views also translate their obligation to protect into language on military AI. This can be seen in arguments for the moral imperative to pursue AI-enabled capability development to maintain freedom of action and protect from adversaries whose uses of AI do not respect legal and ethical obligations.  190  In these views, maintaining freedom of action can also mean maintaining interoperability, or even developing AI systems that help protect friendly forces, as most allies depend on cooperation to fill capability gaps.  191  Beyond operational risks, responsibility also means incorporating ways to minimize risks in the international security environment. DOD has a role to play off the battlefield in this regard as well, including by developing norms around arms control. Allied concerns about diffusion and access, as embedded in the German international security and AI governance agenda, as well as risks that the Dutch identify, make this a compelling area for responsible AI cooperation between defense ministries. In doing so, the United States could find complementary areas of interest with allies that see responsible military AI as encompassing norms in the international environment. In fact, this may be a palatable way to move debates beyond questions exclusive to autonomous weapon systems. While autonomy in weapon systems undoubtedly introduces important questions for ethics, legality, and responsibility, the dominant attention it receives tends to overshadow other aspects of military AI. This not only includes responsible AI implementation, but potentially even the responsibility states have to defend against AI-enabled threats from less ethical adversaries. The fact that most allies are still transitioning from ethical questions wrapped up in autonomy in weapons means that DOD can facilitate and complement their views on ethical and technical dimensions of AI in non-lethal or non-autonomous systems. In doing so, it could help steward the conversation toward other, underrated aspects of ethical design, development, and deployment of military technology. With more of a focus on norms and state responsibility, a broader definition of responsibility beyond the DOD \"responsible AI\" principle can also introduce new convergences for U.S. RAI implementation. The DOD RAI Strategy and Implementation Pathway recently tasked by Deputy Secretary Kathleen Hicks can incorporate these views and help allies refine their approaches to AI ethics in order to enable greater cooperation and allied AI adoption. This is because many allies see firm approaches to managing ethical risk as a prerequisite policy question before investing in AI-enabled capability development-including defensive systems and countermeasures. Overall, international engagement is mutually beneficial to responsible AI endeavors. The United States should look at how other countries are implementing their approaches, just as the United States can exert influence and maintain its leadership role in responsible and ethical AI for defense by helping its allies and partners form their own views in alignment with one another. \n Conclusion No single actor has a monopoly on the answers to implementing responsible AI in any high-risk area, let alone defense. Cooperation is therefore important to collectively navigate the difficulties of responsible governance of emerging technologies. For DOD, the focus has been predominantly on the transposition of safe and ethical AI principles into action. Rather than adopting principles for defense, some allies are moving straight into implementation. Thus far, this is borne out in tabletop exercises, outreach, ministerial committees, ethical reviews, education and certifications, exercises and trials, and defense programs of record. It is too early to judge these fledgling efforts, but tracking their evolution may prove useful to broader AI ethics implementation, be it for other defense ministries or even civilian actors. While jumping straight to implementation can mean a more pragmatic focus on tools, tracking how different AI stakeholders use those tools may be more difficult. In this way, principles can be seen as a helpful organizing force, as is the case for DOD. This said, the scale of the U.S. military bureaucracy and national security innovation base may require higher visibility relative to allied counterparts. Still, another key difference is precisely this visibility. The analysis here is based on information in the public domain, which may also partially explain its transatlantic tilt. The U.S. approach to responsible and ethical AI for defense also differs from other countries in that the consultation and process that led to its principles is far more transparent than is true for most allies. A possible Catch-22 could be at play here, with allies reticent to publicize approaches to such controversial issues, despite the fact that offering such inroads can build trust and confidence that governments are handling these high-stakes questions responsibly. This is important not only for accountability, including to citizenries, but also because dedicating attention to responsible AI is a critical way to signal to industry, civil society, academia, and the research community that appropriate measures are not just boxes to tick, but are fundamentally embedded in the development of systems. In other words, responsible AI is important not just for public opinion, but also to strengthen relationships with the expert community that is rightfully concerned about the ethical implications of current AI advancement. Further, the U.S. national security community should consider norms for diffusion and arms control as part of its responsible AI agenda in order to encourage more allies to issue public approaches to RAI. This does not necessarily mean leading the initiatives; indeed, if the next German government continues with the current arms control agenda, then following the German lead could become a more important area of cooperation. While not all ethical questions relate to adoption, the overlap between them is important. To be sure, incorporating ethics into design prior to moving onto development and deployment phases is critical to ensuring a lifecycle view on AI ethics. But at the organizational level, it is important to make sure that military bureaucracies are neither engaging in \"ethics washing\" with no intention to implement anything, nor seeing ethics as their sole contribution as a way to \"do\" AI on the cheap. This is not to say that allies are not investing in AI-enabled capabilities-as many certainly are-but rather that these attempts and investments are so ad hoc that they could foreclose the possibility of scaling any such efforts. To this end, with the importance of ethics and legality for democratic accountability, tying ethical risk assessment frameworks and associated processes with AI adoption could also be an inroad for more strategic approaches to AI integration. This could be done through multilateral formats, with NATO and the PfD being the most appropriate for the time being. Also, using these formats to help allies dissociate the ethical and legal questions of autonomy from AI could push responsible AI in defense forward, including in associated technology areas not addressed in this paper, such as human-machine teaming or human enhancement. The fact that information about some allies' approaches to AI in defense can be answered via frameworks for human enhancement shows that the stakes of responsible AI are not just about adoption of this one, crucially important technology. The stakes are the testing grounds for the combinative technologies that lie ahead.  (1) Are there any externally imposed constraints on our capability, such as legal and regulatory frameworks that we need to follow? (2) Have we checked the international position as well as domestic? (3) What do we need to stay within these constraints? (4) Is the legal position clear or ambiguous? Do we need to get advice to ensure we comply? (5) Is it possible to influence those constraints if we can't operate within them? (6) Note that anything involving legal matters will take longer than you can possibly imagine, so factor this in. Policy and risk appetite  (1)  Is the enterprise (including partners, suppliers and collaborators) likely to be willing to pursue this capability, based on its own internal policies and risk appetite?  (2)  What are the existing policy and risk positions of our organisations? (3) Are there international policies to consider? (4) What do we need to do to stay within these constraints? (4) Is the policy position clear or ambiguous? Do we need to get advice to ensure we comply? (5) Is it possible to influence the policy if we can't operate within it? (6) Should we try to influence this? For example, what are the risks of not developing the capability? Ethics (1) Fundamentally, should we pursue this capability? (2) Have we considered the ethics of doing so, and equally the ethics of not? (3) What is our organisation's existing ethical position? (4) Does this capability operate within that position? (5) Do our ethics align with those of our partners, and will these partners support and engage in our work? (6) Are systems fair and equitable? Confidence: (a) satisfying regulatory and safety requirements; (b) inspiring trust through assurance, explainability and effective exercising; (c) being aware of the risks through an understanding of threats, vulnerabilities, means of failure and wider resilience Assurance  (1)  Will we be able to certify that the system satisfies all relevant regulations, including safety and security standards? (2) Will all of the functions that the system performs work reliably, as expected and for as long as they need to? The latter is an important point if you have a learning system where the performance could change over time -how do you understand and maintain performance? (3) Do we have an understanding of behaviours the system must not have (e.g. harming people -this is generally considered to be a bad thing) and how they can be prevented? (4) Do we understand what level of assurance is required? Trust (1) Who needs to trust the system, what do they need to understand and what do you need to provide to obtain this trust? (2) This sounds like a simple question but can have many facets -there will be different trust considerations for the direct users, those making decisions based on its outputs, the regulators and the general public Explainability (1) Do we need to be able to explain why the AI made a particular decision; both at the time, and in retrospect? If so, how can we do this? This is another question that may impact on your algorithm selection: if you really need to know why the system produced a certain output, some types of algorithm will be more suitable than others. Resilience (1) Do we understand the vulnerabilities in the system, and the risks it might introduce to our operations or business? Will the system fail gracefully if it encounters situations beyond its design parameters? Experimentation (1) How suited is the system for experimentation, to build experience and confidence before it is used in a live environment? Source: Adapted from: U.K. Defence Science and Technology Laboratory, 22-5. Appendix Design of systems that allows human commanders the ability to monitor information about environment and system Design of systems with modes of operation that allow human intervention and require their input in specific steps of the targeting cycle based on their situational understanding The system design allows the Human to develop sufficiently accurate situation, and system awareness/understanding to identify risks to violating IHL and/or unacceptable moral, ethical or operational outcomes The system design allows the Human to predict the behavior of the system and its effects on the environment (physical and information) The system design allows the Human to impact on the behavior of the system in time to prevent an undesirable act (violating IHL and/or unacceptable moral, ethical or operational outcomes) 7 This point has also been made in relation to the difficulties of a prematurely prohibitive ban on lethal autonomous weapon systems. Although autonomy is beyond the scope of the study, NSCAI Executive Director Yll Bajraktari makes a similar point on the relationship between ethics and interoperability: \"The effects of a prohibition agreement likely would run counter to the U.S. strategic interests as commitments from states such as Russia and China are likely to be empty ones. So, the primary impact of an agreement would be to increase pressure on those countries that abide by international law, including the United States and its democratic allies and partners. If U.S. allies joined an agreement while the United States did not, the diversion would likely hinder allied military interoperability. That would be something really difficult for us and our allies. For these reasons, we believe that practical and strategic problems with a prohibition treaty outweigh the potential benefits for the United States and its allies and partners.\" Craig Smith and Yll Bajraktari, \"Episode #071: AI and trustworthiness and responsibility; ensuring the resilience and upgradeability of systems; designing systems that can be upgradeable over decades; and preserving digital sovereignty to maintain confidentiality and control over information. See: AI Task Force, Artificial Intelligence in Support of Defence, 9-10.  50  The document suggests standardization and incorporating these questions into the front end of software engineering (as a form of \"ethical design\") as the answer, but does little to acknowledge the challenges beyond saying that \"Merely verifying performance is not enough.\" Villani, Towards a French and European Strategy, 113-114. 51 The Ministry's \"Digital Ambition\" is the primary document on data governance, but similarly does not specify data ethics apart from a few references to trust and protection of personal data. See: French Ministry of Armed Forces, Ambition Numérique du Ministère des Armées (Paris, France: 2019), 22-3; Villani, Towards a French and European Strategy, 13. 52 French Ministry of Armed Forces, Ambition Numérique du Ministère des Armées, 10.  53  The focus here is more on preventing and correcting undesirable outcomes, rather than reproducing positive ones. French Ministry of Armed Forces, Ambition Numérique du Ministère des Armées, 13.  54  The French imperative to maintain technological independence is stronger than any other European ally, and largely motivates French defence industrial policy as well as the political agenda of \"strategic autonomy\" and \"digital sovereignty\" at the national and European levels. Other documents that reinforce this include those referenced in footnotes 47 and 49, as well as the 2019 Defence Innovation Orientation Document (2019) and the Ministry of Armed Forces Digital Transformation: Key Concepts (2020). 55 AI Task Force, Artificial Intelligence in Support of Defence, 10. 56 This strong language intends to set the political tone for adoption, and is not purely about ethics. Further, while the \"stranglehold\" motivates sovereignty, there are few specifics in the strategy about hardware components or cloud capabilities, beyond the recognition that these are not French or European strengths. Auditability is only tied, here, to models and data. AI Task Force, Artificial Intelligence in Support of Defence, 10, 24. 57 AI Task Force, Artificial Intelligence in Support of Defence, 10. 58 It is also noteworthy for readers interested in cooperation that freedom of action includes maintaining interoperability with allies. See: AI Task Force, Artificial Intelligence in Support of Defence, 9, 14. 59 The French strategy notes that law and ethics are \"incorporated into the strict and sequenced process of planning the use of force and into a chain of decisionmaking for the application of force established by the rules of engagement, 82 One reason that the Australian government centered on contestability as one of their eight AI ethics principles is the controversy around the \"Robodebt\" scheme in 2016, in which the government used a faulty debt-assessment algorithm. The program issued inaccurate debt notices to welfare recipients. The debts have required significant effort to repay and the program's lawfulness has been questioned. Equally relevant here, the effects on society have been harmful to the mental health of citizens. While this was an automated, not AI system, its adverse impact on the Australian population made clear the priority to include legal recourse for algorithmic decision-making. 89 As described in the article, \"Athena AI is an artificial intelligence system that identifies and classifies objects and locations on a battlefield and communicates to the soldier which ones must not be targeted for legal or humanitarian reasons. This encompasses people such as enemy troops who have surrendered and civilians, as well as locations such as hospitals or other protected sites.\" The article also notes that Athena AI worked with ethicists and moral philosophers to develop a legal and ethical framework for the system before beginning development. See: Jonathan Bradley, \"Athena AI helps soldiers on the battlefield identify protected targets,\" create, April 26, 2021, https://createdigital.org.au/athena-ai-helps-soldiers-identify-protected-targets/.  Appendix V: U.K. Dstl factors for success: consent and confidence Topic Guiding questions Consent: (a) legal and regulatory constraints; (b) as well as policy, ethics, and risk appetite, (c) willingness of suppliers and partners to support where required Legal \n See: Jordan Hayne and Matthew Doran, \"Government to pay back $721m as it scraps Robodebt for Centrelink welfare recipients,\" Australia Broadcasting Corporation, May 29, 2020, https://www.abc.net.au/news/2020-05-29/federal-government-refundrobodebt-scheme-repay-debts/<PHONE>; Monique Mann, \"Technological Politics of Automated Welfare Surveillance: Social (and Data) Justice through Critical Qualitative Inquiry,\" Global Perspectives 1, no. 1 (June 19, 2020), https://doi.org/10.1525/gp.2020.12991. \n \n \n \n 30  In her memorandum on Implementing Responsible Artificial Intelligence in the Department of Defense, Deputy Secretary of Defense Kathleen Hicks announced the formation of a senior-level RAI Working Council to \"accelerate the adoption and implementation of RAI across the department.\" 31  Following its training based on the RAI Champions Pilot, the RAI Working Council will work closely with the JAIC on various issues that are referenced in this report. \n Overall, the absence of concrete German, European, or transatlantic military AI frameworks means that industry has a different starting point when determining the most appropriate framework for responsible military AI. There is no immediately available information on government-guided implementation and requirements validation. With the government more focused on arms control, the German interest in a whole-of-lifecycle approach to AI governance may turn into a delegation of labor-with government looking at responsible use and diffusion and industry focusing on development. The mix here of both self-regulation and waiting for multilateral guidance indicates a clear German interest in military AI governance and ethics-even if more narrow questions around autonomy in weapon systems continues to monopolize public debate. This focus on autonomy in weapons is an important factor in assessing the degree of coherence between U.S. and German approaches to military AI, as it risks overwhelming less controversial issues. Beyond FCAS, the limited bandwidth for AI ethics beyond the tip of the spear could also mean the countries have different bases for how they develop and procure defensive systems and countermeasures. Further, it is not yet clear how the German military government is looking at AI ethics that are separate from autonomy in weapons. Without this separation, it could be more difficult to coalesce on views like the importance of cybersecurity and operator training to building trustworthy, reliable AI for defense. 140  Yet these are not connected to the aforementioned \"uncritical\" cases. Further, compared against other articulated guidelines and frameworks explored in this report, the approach to determining what is \"critical\" differs in that it does not necessarily calibrate ethical-risk mitigation measures according to the anticipated use of the technology, or the technique at hand. \n systems [validation and verification], training of users, training of AI and [machine learning], systems of systems integration, C2 process development, interoperability, operational use, after-action review/lessons learned.\" 159  As such, meaningful human control may come to feature in responsible AI efforts within NATO. While the issuance of principles and commitment to uphold shared values through responsible use are undoubtedly important, it is worth noting that NATO does have different considerations from nations (and militaries) as a supranational, non-regulatory body. Its more unique contribution could be standardization, as already incorporated into standardization agreements and training publications-the latter of which could focus on responsible AI in training. In anticipation of forthcoming guidelines or standards for industry, it is worth noting that NATO operational standardization also encompasses work implementing the Law of Armed Conflict into operational practice. 160 \n now, it is the European Parliament that plays the most visible role advancing ethics in European military R&D funding. This was seen in mid-2018, in its attempt to ban all military AI research using EU funds because of concerns about LAWS. The agreedupon final version explicitly prohibits funding for LAWS at the European level-a deal-breaker without which the Parliament would never have agreed to allow for any defense funding.171  But the final result was narrower because the EU does not have jurisdiction over its member states' armaments development unless they use EU funds. As such, while important, especially for dualuse and open-source systems, its jurisdiction on mandating ethical reviews is still limited and likely to not affect the majority of national, bilateral, or minilateral capability development programs. United States, four of the Five Eyes countries are included in this report for their national-level approaches to ethical and responsible AI in defense.173  Through policy exchanges and the Technical Cooperation Program (TTCP), the Five Eyes militaries are already engaged in cooperative digitalization efforts.174  Policy exchanges between Australia, Canada, New Zealand, the U.K., and the United States take various forms that can facilitate alignment. TTCP is more specific as an \"international organization aiming to collaborate and exchange defense scientific and technical research and information, and harmonize and align defense research programs by sharing or exchanging research activities\" between the five countries. 175 TTCP is directed by principals from the Five Eyes countries, who agree on and direct three-year Strategic Challenges on specific technical areas for collaboration. Here, the TTCP Strategic Challenges on cyber, autonomy, and AI are relevant here because they focus on aspects related to trustworthiness and related issues. More recently, the Parliament-issued Guidelines for military and non-military use of Artificial Intelligence in January 2021 could indicate a stronger ethical bent than seen in the other institutions. 172 Five Eyes Including the \n Table 1 : 1 Comparison of allies' approaches to ethical AI principles and risks in defense Articulated views Emerging views Nascent views USA France Australia U.K. Canada Netherlands* Managing Responsible Responsible Responsible Responsibility Policy and risk appetite Accountability and liability export/diffusion that enhances adversaries' capabilities Equitable Equitable (Bias implicit in reliability) (Part of trust) Ethics (fair and equitable) Equality (Mention of Privacy, Traceable Traceable explainability/ Traceability Explainability confidentiality, configuration) and security Security (control Reliable Reliable Trustworthy Trust Trust; Assurance Reliability and trust being overridden; inaccurate decisions/error) Governable Governable (deactivate/ disengage) (See control) Governance Resilience (fail gracefully) System interference Level of dependence on Controlled Controlled (sovereignty) (Sovereign capability) foreign technology companies; control; lock-in Lawful Compliance with law Compliance with law Law Legal Compliance with law Societal feedback mechanisms (Contestability) Effect on society; consent Public opinion; undue influence on the public Key: blue squares are the primary principles/facets; white squares are related sub-topics mentioned in the respective documents *The Dutch technical report focuses on risks, not ethics, and therefore has a slightly different scope than the others mentioned here. Sources: Author's analysis of: U.S. Department of Defense, \"DOD Adopts Ethical Principles for Artificial Intelligence,\" 2020; French Ministry of Armed Forces; \"AI Strategy in Support of Defence,\" 2019; Devitt et al., \"A Method for Ethical AI in Defence,\" 2020; U.K. Defence Science and Technology Laboratory, Building Blocks for Artificial Intelligence and Autonomy, 2020; Girling et al., Identifying Ethical Issues of Human Enhancement Technologies in the Military (2017); Neef and van Weerd, Artificial Intelligence in the Context of National Security, 2020. \n VI: Canadian Military Ethics Assessment Framework for Human Enhancement Technologies Humanity Questions raised about the influence of an enhancement System on a soldier's morals and personhood Conduct and analysis, use Human Centred Design Designers approaches such as those described in ISO 9241 210 [0], Reliability and Questions raised about how close the enhancement and apply best practice to systems design and testing to Trust technology is to commercialization and use by the military, develop systems that support the delivery of EHC and MHC and remaining modifications required for usability on the battlefield Integrate AI enabled systems within their wider system of Effect on Society Questions raised about how an enhancement will impact Organisational Users systems and organisational structures such that they enable civilians and perception outside of the forces Preparedness for Questions raised about how adversaries will view our use EHC and MHC Adversaries of enhancements and how adversaries may use End users enhancements themselves Train, support and employ human machine teams that deliver EHC and MHC; Deliver operating procedures and Source: Girling et al., Defence Research Development Canada, 15. practices to deliver EHC and MHC Appendix VII: NATO STO technical report on stakeholders and the application of dimensions of human Application proposed dimensions of (ET) developed Dimension of human control to iPRAW's Requirements for Human Control in the control in the use of force Use of Force Compliance with National Laws control in fielded systems Questions raised about whether a technology interferes that guide both military employees in all activities related with the common values, laws and expected behaviors Stakeholder groups that contribute to ensuring appropriate human Situational Understanding Intervention and Codes of to their professional duties Conduct Jus Ad Bellum Stakeholder Questions raised about whether a technology disrupts Jus Role in Delivering Appropriate Human Control in Future Principles Groups ad Bellum principles: criteria to be met before entering a Systems conflict to ensure that all conflicts entered into are justified Law of Armed Conflict/Jus in Control by Design Questions raised about whether a technology violates the international laws that must be followed during times of [Effective Human Control = EHC; Meaningful Human Control = MHC] Bello Principles (Technical conflict to protect those affected by conflict and to Health and Safety Policy makers Set policy, standards and doctrine to facilitate EHC and MHC enhancement may have on the physical or psychological across NATO systems Questions raised about direct or indirect impacts the regulate means of warfare Control) well-being of a soldier or civilian Accountability and Liability R&D/ Scientific Privacy, Confidentiality, Community Questions raised about risk and responsibility for Identify and fill knowledge gaps in the field of EHC and MHC; enhancement failures or unanticipated and undesired Conduct research to provide evidence to support the effects of an enhancement specification and design of systems that support MHC and Questions raised about sharing, storing, and using information obtained by an enhancement, and security EHC and Security risks of an enhancement resulting from adversary System detection or hacking Specify, contract and accept AI enabled systems that support Equality Acquirers Questions raised about the influence of an enhancement EHC and MHC on fairness and functionality within the CAF, between militaries and in society Consent Questions raised about whether the enhancement is mandatory or voluntary \n Adapted from author's conference presentation (November 11, 2019); see also High-Level Expert Group on Artificial Intelligence, European Commission. Appendix IX: Five Eyes TTCP Cyber Strategic Challenge Working Group attributes of trustworthiness of cyber systems Yuna Huh Wong, John M. Yurchak et al., Deterrence in the Age of Thinking Machines (Santa Monica, CA: RAND Corporation, 2020), 6, 60. The Human is able to predict the Oversight with human adherence to IHL are Explainability of Integrity behavior of the system and its effects on the environment (physical and in/on loop Just distribution of costs/benefits decisions to those accuracy; credibility affected An attribute of security; ✓ maintained Training with systems allows information) Freedom from bias, No exacerbation of harm discrimination, and stigmatisation Fault-tolerance; performability; or adverse effects on human beings Accountability and ability to contest AI decisions Ensure that data, systems, and business models are traceable, explainable, and accountability; authenticity; Cost-benefit analysis users to understand and Harm inherent to \"Should never lead to predict system behaviors across different situations in order to avoid undesirable outcomes or failure to comply with IHL. Organizational culture does not indirectly impact on Ensure technically robust systems are not vulnerable to malicious conduct of warfare; \"AI systems and the remains relevant people being deceived\" ≠ military aim of Facilitate auditability and explainability ✓ ✓ ✓ Accountability and establishing \"unfair\" through documentation use environments in which they operate must be safe and secure\" ≠ operational realities explicability of decision-making processes advantage (e.g., operational surprise) to maximise Inform humans when reproducibility and \"they are interacting nonrepudiability minimise repetition of with an AI system\" ≠ mistakes psychological freedom of choice and communicated Have ability to explain dimension of strategy Recoverability; retainability; ✓ willingness to question system behaviors and actions Source: Boardman and Butcher, NATO Science & Technology Organization, 9-Prevention of harm Robustness Fairness Transparency Maintainability 11. Appendix VIII: Applicability of civilian EU trustworthy AI principles to the defense sector Ethical principle defense sector defense sector complexity Description Convergence with Divergence with Technical robustness and safety Preventing \"unacceptable\" harm consistent with Avoid unfair bias by enabling inclusion and Place oversight processes to analyse trade-offs between correctability; self-healing; explainability and self-repair accuracy \"Unfair\" not in interest Preventative approach militaries seeking to Ensure AI system diversity and address system's of militaries who seek to risks in AI development, including protection against minimise \"unintended\" consequences performs as intended \"without harming living beings or the Diversity, non-discrimination Improve accessibility through user-centric, purpose, constraints, requirements, and decisions to reduce bias to establish \"unfair Adaptability Autonomy; learning; ✓ ✓ advantage\"; nonetheless important extensibility; reconfigurability adversarial attacks and Principles recommend environment\" ≠ and fairness universal design in programming (and to consider bias mitigation of unintended promoting \"convergence conduct of warfare training phase) especially for internal harm between the AI community and the security community\" Solicit stakeholder feedback throughout lifecycle Use universal design to compatibility; reusability; enable equitable access management systems learnability; satisfaction; or unintended harm Usability Automatability; flexibility; ✓ Mutual interests in Enable assessment of Potentially more Respect for human autonomy Reactiveness Privacy and data governance Accountability Reliability Leanness trustworthiness Main attributes of Timeliness Interoperability,\" Center for European Policy Analysis, February 17, 2021, Ability for humans to maintain self-determination and partake in democratic process Design of AI systems to augment, complement and empower human skills (alignment; human-machine teaming) \"Should not unjustifiably subordinate, coerce, deceive, manipulate, condition or herd humans\" ≠ military etc. privacy confidential information in GDPR impacts to ensure trust, including through documentation potential negative impacts of AI systems, teaming exercises prior high-confidence; assurance; including through red-from CCW 'human certainty; fault-forecasting; Readiness; fault-removal ✓ ✓ accountability' term] consistency; stability; Redress for adverse document, and minimise accountability differs predictability; competence; justification to coerce Data governance covering quality and integrity of data, access protocols and capabilities for processing that protect assuring high level of data quality, integrity, and access Potential overlap in technical measures to protect privacy & protect divergence for implementing privacy and data governance, although already have carveouts for public safety, national security, and defense algorithms, data, and N/A Mechanisms to ensure responsibility, accountability, systems and their outcomes applications Identify, assess, definition of An attribute of dependability; ✓ ✓ ✓ scalability documentation-centric Efficiency; simplicity; ✓ [But noteworthy that auditability for AI design processes and audit of safety-critical to the main attribute conduct independent Equivalent or sub-attribute(s) Trust Resilience Agility Quickness; decisiveness ✓ Proactiveness prevention survivability Preparedness; fault- Similar to Article 36 Measuring impact of AI agency\" important for adoption on \"social to/during development N/A when applied to existing legal Social well-being not ✓ Control in Use (Operational Control) and oversight Confidentiality Explicability Safety Human agency Societal and environmental well-being Societal and environmental well-being Availability Source: Replicated from: Cho et al., MILCOM 2016. Appropriate monitoring of the system and the operational environment The Human has sufficiently accurate situation, and system awareness/understanding to identify risks to violating IHL and/or unacceptable moral, ethical or operational outcomes. Define human agency and governance with human at centre to prevent adverse effects military capabilities Authority and accountability of human operators, abide by IHL The Human is able to exercise freedom of choice and has the ability to affect system behavior during use to ensure that accountability and Consideration of legal aspects throughout development and deployment pipeline overall activity of the AI system\" may conflict with select niche AI supervisor systems for speed (e.g., cyber etc.) AI system black-box AI AI-made or AI-augmented decisions) An attribute of security ✓ requiring machine capabilities/purpose of communication) for humans can contest non-combat effects democratic institutions personnel; focus on \"encouragement\" rather processes and (traceability, auditability, loops to ensure that prescriptive openness about explicability measures may affect feedback An attribute of dependability ✓ ✓ than anything Transparency and Augmentation of other transparency (which unintentional harm) Limitations on of intentional & teammates and commanders; Assessment of fundamental rights impact prior to development review for autonomous systems as well as appropriate level of judgment/control, as well as TEVV for new framework for armed conflict \"Human-in-command\" (\"capability to oversee Measurement and assessment of environmental/ecological impact, and societal impact including on understanding role of human in human-machine teams, including impact on health and well-being of taken into account for adversaries (linking back to main tension sector and prevention an attribute of security between defense (See description under prevention of harm; defined as relevant to both principles) An attribute of dependability; ✓ ✓ ✓ http://cepa.org/nato-leadership-on-ethical-ai-is-key-to-future-interoperability/.6 \n It should be noted that this is different from the legal concept of state responsibility. In international law, state responsibility refers to principles that guide how a state is held accountable after a violation. As Appendix I shows, the DIB considered the doctrine of state responsibility (as \"remediation mechanisms for actions after hostilities have ended\") as a layer of responsible AI, albeit one external to DOD. Nevertheless, the concept is sufficiently different that it should not be confused with the more general concept of responsible state behavior as described in this report. See: Defense Innovation Board, Supporting Document: AI Principles: Recommendations on the Ethical Use of Artificial Intelligence by the Department of Defense (Washington, DC: October 30, 2019),28-29.    Towards responsible and accountable innovation,\" Research Policy 47, no. 1 (February 2018): 61-69. 9 \n 78  An Article 36 review is an example of the type of planning the government would need to know about in advance. LEAPP is a \"data item descriptor.\" The United States uses a variation of this term, \"data item description,\" which is a \"standardization document that defines the data[content, format, and intended use] required of a contractor.\" See: Defense Standardization Program, \"Frequently Asked Questions (FAQs): Data Item Descriptions,\" U.S. Department of Defense, accessed July 30, 2021, https://www.dsp.dla.mil/Policy-Guidance/FAQs/Data-Item-Descriptions/; Devitt et al., A Method for Ethical AI in Defence, 34, 62. 79 For an overview of the scope and function of Article 36 reviews, see: International Committee of the Red Cross, \"A Guide to the Legal Review of New Weapons, Means and Methods of Warfare: Measures to Implement Article 36 of Additional Protocol I of 1977,\" International Review of the Red Cross 88, no. 864 (December 2006), 931-956, https://www.icrc.org/en/doc/assets/files/other/irrc_864_icrc_geneva.pdf. 80 More specifically, as Dr. Lauren Sanders of the International Weapons Review describes, technical measures like the pixilation of feed, whether and how identification of combatants is possible, and if those identification parameters are adjustable for different settings, would help determine what types of functions (targeting, intelligence, decision support) would be acceptable. See the segment starting at 25:38 of: Trusted Autonomous Systems Defence Cooperative Research Centre, \"Pragmatic Tools for Considering and Managing Ethical Risks in AI for Defence -Detailed.\" 81 Trusted Autonomous Systems Defence Cooperative Research Centre, \"Pragmatic Tools for Considering and Managing Ethical Risks in AI for Defence -Detailed.\" \n 90 Trusted Autonomous Systems Defence Cooperative Research Centre, \"Pragmatic Tools for Considering and Managing Ethical Risks in AI for Defence -Detailed.\" 91 Trusted Autonomous Systems Defence Cooperative Research Centre, \"Pragmatic Tools for Considering and Managing Ethical Risks in AI for Defence -Detailed.\" 92 Devitt et al., A Method for Ethical AI in Defence, 50.93 The Australian investigation of war crime allegations in Afghanistan also forms part of the context around ethics and state responsibility, and is beyond the scope here except to say that the redress mechanisms are part of the legal approach to state responsibility. See also footnotes 10 and 81.94 Her Majesty's Government, Global Britain in a competitive age: The Integrated Review of Security, Defence, Development and Foreign Policy (London: March The NSCAI notes that trustworthiness is one of the focus areas of the TTCP Strategic Challenge on AI. Limited information is available about the Law & Ethics working group, except that the lead author of the Australian method document is a member. National Security Commission on Artificial Intelligence, 2021), 39-40, https://assets.publishing.service.gov.uk/government/uploads/system/uploads/att achment_data/file/975077/Global_Britain_in_a_Competitive_Age- _the_Integrated_Review_of_Security__Defence__Development_and_Foreign_Po licy.pdf. of Trustworthy Systems,\" Proceedings of MILCOM 2016 -2016 IEEE Military Communications Conference (Baltimore, MD: November 1-3, 2016), 1237- 1242. 176First Quarter Recommendations (Washington, DC: 2020), 65; \"The Ethics of AI \n\t\t\t Denis Mercier, \"SACT's opening remarks to the NAC/MC Away Day,\" Allied Command Transformation, March 22, 2018.5 Joanna van der Merwe, \"NATO Leadership on Ethical AI is Key to Future \n\t\t\t While this may seem obvious, not all design begins with identifying a need before matching it with the appropriate solution-technological or otherwise. One example of ways to improve the role of the human is reducing her cognitive load and affording her more capacity to make accurate decisions that enhance compliance with international humanitarian legal principles.11 This runs parallel to determinations of human in, on, or out of the loop that are often discussed in relation to autonomy in weapons.12 The DIB describes the ethical and technical distinctions between AI and autonomy in its supporting document on AI ethics recommendations: \"While some autonomous systems may utilize AI in their software architectures, this is not always the case. The interaction between AI and autonomy, even if it is not a weapon system, carries with it ethical considerations. Indeed, it is likely that most of these types of systems will have nothing to do with the application of lethal force, but will be used for maintenance and supply, battlefield medical aid and assistance, logistics, intelligence, surveillance and reconnaissance, and humanitarian and disaster relief operations. Various ethical dimensions may arise depending upon the system and its domain of use, and those will change depending upon context.\" See: Defense Innovation Board, Supporting Document: AI Principles: Recommendations on the Ethical Use of Artificial Intelligence by the Department of Defense, 10.13  Several of these types of systems could be integrated into a weapon system with a degree of autonomy. Still, the differences between the autonomy-related and AI-related risks in that system would still apply.14 For a comprehensive summary and links to country positions, see: Dustin E. Lewis (editor), \"A Compilation of Materials Apparently Reflective of States' \n\t\t\t The French iteration of controlled AI has several components: maintaining freedom of action and interoperability with allies; assuring the ethical concepts of \n\t\t\t This definition concerns soldiers, but opens up questions about other domains as well. For example, a French helicopter pilot has argued that pilots bear a heavier ethical burden as one or two operators in the air in comparison to the \n\t\t\t That said, one discrepancy is that the Integrated Review also calls into question the pace of global governance in relation to technology rule, standard, \n\t\t\t Girling et al., Identifying Ethical Issues of Human Enhancement Technologies in the Military, 16. 110 Girling et al., Identifying Ethical Issues of Human Enhancement Technologies in the Military, 15. 111 Department of National Defence of Canada, The Department of National Defence and Canadian Armed Forces Data Strategy (Ottawa: December 3, 2019), 8, https://www.canada.ca/en/department-nationaldefence/corporate/reports-publications/data-strategy.html. \n\t\t\t NATO Research and Technology Organization, Human Systems Integration for Network Centric Warfare (Neuilly-sur-Seine: February 2010), 2-3. 157 NATO Research and Technology Organization, Human Systems Integration for Network Centric Warfare, 1-7. \n\t\t\t New Zealand, which is also not in the PfD, has not issued a public approach but has different models to pluck from within the Five Eyes context.174 Zoe Stanley-Lockman, \"Toward a Military AI Cooperation Toolbox: Modernizing S&T Defense Partnerships for the Digital Age\" (Center for Security and Emerging Technology, forthcoming). \n\t\t\t In addition to deploying alongside each other, this could conceivably include sharing data inputs like models or training data for co-developed or emulated systems. Understanding the limitations of an ally's capability would be important for development and deployment alike.184 Devitt et al., A Method for Ethical AI in Defence, 49-50.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Responsible-and-Ethical-Military-AI.tei.xml", "id": "3c8bf6960c338054bc391c780818480e"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "In particular, we are grateful for Markus Anderljung's insightful suggestions and detailed editing.", "authors": ["Baobao Zhang", "Allan Dafoe"], "title": "Artificial Intelligence: American Attitudes and Trends", "text": "Advances in artificial intelligence (AI) 1 could impact nearly all aspects of society: the labor market, transportation, healthcare, education, and national security. AI's effects may be profoundly positive, but the technology entails risks and disruptions that warrant attention. While technologists and policymakers have begun to discuss AI and applications of machine learning more frequently, public opinion has not shaped much of these conversations. In the U.S., public sentiments have shaped many policy debates, including those about immigration, free trade, international conflicts, and climate change mitigation. As in these other policy domains, we expect the public to become more influential over time. It is thus vital to have a better understanding of how the public thinks about AI and the governance of AI. Such understanding is essential to crafting informed policy and identifying opportunities to educate the public about AI's character, benefits, and risks. In this report, we present the results from an extensive look at the American public's attitudes toward AI and AI governance. As the study of the public opinion toward AI is relatively new, we aimed for breadth over depth, with our questions touching on: workplace automation; attitudes regarding international cooperation; the public's trust in various actors to develop and regulate AI; views about the importance and likely impact of different AI governance challenges; and historical and cross-national trends in public opinion regarding AI. Our results provide preliminary insights into the character of U.S. public opinion regarding AI. However, our findings raise more questions than they answer; they are more suggestive than conclusive. Accordingly, we recommend caution in interpreting the results; we confine ourselves to primarily reporting the results. More work is needed to gain a deeper understanding of public opinion toward AI. Supported by a grant from the Ethics and Governance of AI Fund, we intend to conduct more extensive and intensive surveys in the coming years, including of residents in Europe, China, and other countries. We welcome collaborators, especially experts on particular policy domains, on future surveys. Survey inquiries can be emailed to <EMAIL>. This report is based on findings from a nationally representative survey conducted by the Center for the Governance of AI, housed at the Future of Humanity Institute, University of Oxford, using the survey firm YouGov. The survey was conducted between June 6 and 14, 2018, with a total of 2,000 American adults (18+) completing the survey. The analysis of this survey was pre-registered on the Open Science Framework. Appendix A provides further details regarding the data collection and analysis process. Below we highlight some results from our survey 2 : • Americans express mixed support for the development of AI. After reading a short explanation, a substantial minority (41%) somewhat support or strongly support the development of AI, while a smaller minority (22%) somewhat or strongly opposes it. • Demographic characteristics account for substantial variation in support for developing AI. Substantially more support for developing AI is expressed by college graduates (57%) than those with high school or less education (29%); by those with larger reported household incomes, such as those earning over $100,000 annually (59%), than those earning less than $30,000 (33%); by those with computer science or programming experience (58%) than those without (31%); by men (47%) than women (35%). These differences are not easily explained away by other characteristics (they are robust to our multiple regression). • The overwhelming majority of Americans (82%) believe that robots and/or AI should be carefully managed. This figure is comparable to with survey results from EU respondents. • Americans consider all of the thirteen AI governance challenges presented in the survey to be important for governments and technology companies to manage carefully. The governance challenges perceived to be the most likely to impact people around the world within the next decade and rated the highest in issue importance were 3 : 1. Preventing AI-assisted surveillance from violating privacy and civil liberties  1  We define AI as machine systems capable of sophisticated (intelligent) information processing. For other definitions, see Footnote 2 in  Dafoe (2018) .  2  These results are presented roughly in the order in which questions were presented to respondents.  3  Giving equal weight to the likelihood and the rated importance of the challenge. 2. Preventing AI from being used to spread fake and harmful content online 3. Preventing AI cyber attacks against governments, companies, organizations, and individuals 4. Protecting data privacy • We also asked the above question, but focused on the likelihood of the governance challenge impacting solely Americans (rather than people around the world). Americans perceive that all of the governance challenges presented, except for protecting data privacy and ensuring that autonomous vehicles are safe, are slightly more likely to impact people around the world than to impact Americans within the next 10 years. • Americans have discernibly different levels of trust in various organizations to develop and manage 4 AI for the best interests of the public. Broadly, the public puts the most trust in university researchers (50% reporting \"a fair amount of confidence\" or \"a great deal of confidence\") and the U.S. military (49%); followed by scientific organizations, the Partnership on AI, technology companies (excluding Facebook), and intelligence organizations; followed by U.S. federal or state governments, and the UN; followed by Facebook. • Americans express mixed support (1) for the U.S. investing more in AI military capabilities and (2) for cooperating with China to avoid the dangers of an AI arms race. Providing respondents with information about the risks of a U.S.-China AI arms race slightly decreases support for the U.S. investing more in AI military capabilities. Providing a pro-nationalist message or a message about AI's threat to humanity failed to affect Americans' policy preferences. • The median respondent predicts that there is a 54% chance that high-level machine intelligence will be developed by 2028. We define high-level machine intelligence as when machines are able to perform almost all tasks that are economically relevant today better than the median human (today) at each task. See Appendix B for a detailed definition. • Americans express weak support for developing high-level machine intelligence: 31% of Americans support while 27% oppose its development. • Demographic characteristics account for substantial variation in support for developing high-level machine intelligence. There is substantially more support for developing high-level machine intelligence by those with larger reported household incomes, such as those earning over $100,000 annually (47%) than those earning less than $30,000 (24%); by those with computer science or programming experience (45%) than those without (23%); by men (39%) than women (25%). These differences are not easily explained away by other characteristics (they are robust to our multiple regression). • There are more Americans who think that high-level machine intelligence will be harmful than those who think it will be beneficial to humanity. While 22% think that the technology will be \"on balance bad,\" 12% think that it would be \"extremely bad,\" leading to possible human extinction. Still, 21% think it will be \"on balance good,\" and 5% think it will be \"extremely good.\" • In all tables and charts, results are weighted to be representative of the U.S. adult population, unless otherwise specified. We use the weights provided by YouGov. • Wherever possible, we report the margins of error (MOEs), confidence regions, and error bars at the 95% confidence level. • For tabulation purposes, percentage points are rounded off to the nearest whole number in the figures. As a result, the percentages in a given figure may total slightly higher or lower than 100%. Summary statistics that include two decimal places are reported in Appendix B. We measured respondents' support for the further development of AI after providing them with basic information about the technology. Respondents were given the following definition of AI: Artificial Intelligence (AI) refers to computer systems that perform tasks or make decisions that usually require human intelligence. AI can perform these tasks or make these decisions without explicit human instructions. Today, AI has been used in the following applications: [five randomly selected applications] Each respondent viewed five applications randomly selected from a list of 14 that included translation, image classification, and disease diagnosis. Afterward, respondents were asked how much they support or oppose the development of AI. (See Appendix B for the list of the 14 applications and the survey question.) Americans express mixed support for the development of AI, although more support than oppose the development of AI, as shown in Figure  2 .1. A substantial minority (41%) somewhat or strongly supports the development of AI. A smaller minority (22%) somewhat or strongly oppose its development. Many express a neutral attitude: 28% of respondents state that they neither support nor oppose while 10% indicate they do not know. Our survey results reflect the cautious optimism that Americans express in other polls. In a recent survey, 51% of Americans indicated that they support continuing AI research while 31% opposed it (Morning Consult 2017). Furthermore, 77% of Americans expressed that AI would have a \"very positive\" or \"mostly positive\" impact on how people work and live in the next 10 years, while 23% thought that AI's impact would be \"very negative\" or \"mostly negative\" (Northeastern University and Gallup 2018). We examined support for developing AI by 11 demographic subgroup variables, including age, gender, race, and education. (See Appendix A for descriptions of the demographic subgroups.) We performed a multiple linear regression to predict support for developing AI using all these demographic variables. Support for developing AI varies greatly between demographic subgroups, with gender, education, income, and experience being key predictors. As seen in Figure  2 .2, a majority of respondents in each of the following four subgroups express support for developing AI: those with four-year college degrees (57%), those with an annual household income above $100,000 (59%), those who have completed a computer science or engineering degree (56%), and those with computer science or programming experience (58%). In contrast, women (35%), those with a high school degree or less (29%), and those with an annual household income below $30,000 (33%), are much less enthusiastic about developing AI. One possible explanation for these results is that subgroups that are more vulnerable to workplace automation express less enthusiasm for developing AI. Within developed countries, women, those with low levels of education, and low-income workers have jobs that are at higher risk of automation, according to an analysis by the Organisation for Economic Cooperation and Development  (Nedelkoska and Quintini 2018) . We used a multiple regression that includes all of the demographic variables to predict support for developing AI. The support for developing AI outcome variable was standardized, such that it has mean 0 and unit variance. Significant predictors of support for developing AI include: • Being a Millennial/post-Millennial (versus being a Gen Xer or Baby Boomer) • Being a male (versus being a female) • Having graduated from a four-year college (versus having a high school degree or less) • Identifying as a Democrat (versus identifying as a Republican) • Having a family income of more than $100,000 annually (versus having a family income of less than $30,000 annually) • Not having a religious affiliation (versus identifying as a Christian)  • Having CS or programming experience (versus not having such experience) Some of the demographic differences we observe in this survey are in line with existing public opinion research. Below we highlight three salient predictors of support for AI based on the existing literature: gender, education, and income. Around the world, women have viewed AI more negatively than men. Fifty-four percent of women in EU countries viewed AI positively, compared with 67% of men  (Eurobarometer 2017) . Likewise in the U.S., 44% of women perceived AI as unsafe -compared with 30% of men  (Morning Consult 2017) . This gender difference could be explained by the fact that women have expressed higher distrust of technology than men do. In the U.S., women, compared with men, were more likely to view genetically modified foods or foods treated with pesticides as unsafe to eat, to oppose building more nuclear power plants, and to oppose fracking  (Funk and Rainie 2015) . One's level of education also predicts one's enthusiasm toward AI, according to existing research. Reflecting upon their own jobs, 32% of Americans with no college education thought that technology had increased their opportunities to advance -compared with 53% of Americans with a college degree  (Smith and Anderson 2016) . Reflecting on the economy at large, 38% of those with post-graduate education felt that automation had helped American workers while only 19% of those with less than a college degree thought so  (Graham 2018) . A similar trend holds in the EU: those with more years of education, relative to those with fewer years, were more likely to value AI as good for society and less likely to think that AI steals people's jobs  (Eurobarometer 2017) . Another significant demographic divide in attitudes toward AI is income: low-income respondents, compared with highincome respondents, view AI more negatively. For instance, 40% of EU residents who had difficulty paying their bills \"most of the time\" hold negative views toward robots and AI, compared with 27% of those who \"almost never\" or \"never\" had difficulty paying their bills  (Eurobarometer 2017 ). In the U.S., 19% of those who made less than $50,000 annually think that they are likely to lose their job to automation -compared with only 8% of Americans who made more than $100,000 annually  (Graham 2018) . Furthermore, Americans' belief that AI will help the economy, as well as their support for AI research is positively correlated with their income (Morning Consult 2017). Electronic copy available at: https://ssrn.com/abstract=<PHONE>  Electronic copy available at: https://ssrn.com/abstract=<PHONE> Robots and artificial intelligence are technologies that require careful management. We asked a similar question except respondents were randomly assigned to consider one of these three statements: • AI and robots are technologies that require careful management. • AI is a technology that requires careful management. • Robots are technologies that require careful management. Our respondents were given the same answer choices presented to the Eurobarometer subjects. The overwhelming majority of Americans -more than eight in 10 -agree that AI and/or robots should be carefully managed, while only 6% disagree, as seen in Figure  2 .5.  5  We find that variations in the statement wording produce minor differences, statistically indistinguishable from zero, in responses. Next, we compared our survey results with the responses from the 2017 Special Eurobarometer #460 by country (Eurobarometer 2017). For the U.S., we used all the responses to our survey question, unconditional on the experimental condition, because the variations in question-wording do not affect responses. The percentage of those in the U.S. who agree with the statement (82%) is not far off from the EU average (88% agreed with the statement). Likewise, the percentage of Americans who disagree with the statement (6% disagree) is comparable with the EU average (7% disagreed). The U.S. ranks among the lowest regarding the agreement with the statement in part due to the relatively high percentage of respondents who selected the \"don't know\" option. At the beginning of the survey, respondents were asked to consider five out of 15 potential global risks (the descriptions are found in Appendix B). The purpose of this task was to compare respondents' perception of AI as a global risk with their notions of other potential global risks. The global risks were selected from the Global Risks Report 2018, published by the World Economic Forum. We edited the description of each risk to be more comprehensible to non-expert respondents while preserving the substantive content. We gave the following definition for a global risk: A \"global risk\" is an uncertain event or condition that, if it happens, could cause significant negative impact for at least 10 percent of the world's population. That is, at least 1 in 10 people around the world could experience a significant negative impact.  6  After considering each potential global risk, respondents were asked to evaluate the likelihood of it happening globally within 10 years, as well as its impact on several countries or industries. We use a scatterplot (Figure  2 .8 to visualize results from respondents' evaluations of global risks. The x-axis is the perceived likelihood of the risk happening globally within 10 years. The y-axis is the perceived impact of the risk. The mean perceived likelihood and impact is represented by a dot. The corresponding ellipse contains the 95% confidence region. In general, Americans perceive all these risks to be impactful: on average they rate each as having between a moderate (2) and severe (3) negative impact if they were to occur. Americans perceive the use of weapons of mass destruction to be the most impactful -at the \"severe\" level (mean score 3.0 out of 4). Although they do not think this risk as likely as other risks, they still assign it an average of 49% probability of occurring within 10 years. perceived to be the most likely as well as the most impactful. These include natural disasters, cyber attacks, and extreme weather events. The American public and the nearly 1,000 experts surveyed by the World Economic Forum share similar views regarding most of the potential global risks we asked about  (World Economic Forum 2018) . Both the public and the experts rank extreme weather events, natural disasters, and cyber attacks as the top three most likely global risks; likewise, both groups consider weapons of mass destruction to be the most impactful. Nevertheless, compared with experts, Americans offer a lower estimate of the likelihood and impact of the failure to address climate change. The American public appears to over-estimate the likelihoods of these risks materializing within 10 years. The mean responses suggest (assuming independence) that about eight (out of 15) of these global risks, which would have a significant negative impact on at least 10% of the world's population, will take place in the next 10 years. One explanation for this is that it arises from the broad misconception that the world is in a much worse state than it is in reality  (Pinker 2018; Rosling, Rönnlund, and Rosling 2018 ). Another explanation is that it arises as a byproduct of respondents interpreting \"significant negative impact\" in a relatively minimal way, though this interpretation is hard to sustain given the mean severity being between \"moderate\" and \"severe.\" Finally, this result may be because subjects centered their responses within the distribution of our response options, the middle value of which was the 40-60% option; thus, the likelihoods should not be interpreted literally in the absolute sense. The adverse consequences of AI within the next 10 years appear to be a relatively low priority in respondents' assessment of global risks. It -along with adverse consequences of synthetic biology -occupy the lower left quadrant, which contains what are perceived to be lower-probability, lower-impact risks. 7 These risks are perceived to be as impactful (within the next 10 years) as the failure to address climate change, though less probable. One interpretation of this is that the average American simply does not regard AI as posing a substantial global risk. This interpretation, however, would be in tension with some expert assessment of catastrophic risks that suggests unsafe AI could pose significant danger  (World Economic Forum 2018; Sandberg and Bostrom 2008) . The gap between experts and the public's assessment suggests that this is a fruitful area for efforts to educate the public. Another interpretation of our results is that Americans do have substantial concerns about the long-run impacts of advanced AI, but they do not see these risks as likely in the coming 10 years. As support for this interpretation, we later find that 12% of American's believe the impact of high-level machine intelligence will be \"extremely bad, possibly human extinction,\" and 21% that it will be \"on balance bad.\" Still, even though the median respondent expects around a 54% chance of high level machine intelligence within 10 years, respondents may believe that the risks from high level machine intelligence will manifest years later. If we assume respondents believe global catastrophic risks from AI only emerge from high-level AI, we can infer an implied global risk, conditional on high-level AI (within 10 years), of 80%. Future work should try to unpack and understand these beliefs. We used a survey experiment to understand how the public understands the terms AI, automation, machine learning, and robotics. (Details of the survey experiment are found in Appendix B.) We randomly assigned each respondent one of these terms and asked them: In your opinion, which of the following technologies, if any, uses [artificial intelligence (AI)/automation/machine learning/robotics]? Select all that apply. Because we wanted to understand respondents' perceptions of these terms, we did not define any of the terms. Respondents were asked to consider 10 technological applications, each of which uses AI or machine learning. Though the respondents show at least a partial understanding of the terms and can identify their use within the considered technological applications correctly, the respondents underestimate the prevalence of AI, machine learning, and robotics in everyday technological applications, as reported in Figure  2 .9. (See Appendix C for details of our statistical analysis.) Among those assigned the term AI, a majority think that virtual assistants (63%), smart speakers (55%), driverless cars (56%), social robots (64%), and autonomous drones use AI (54%). Nevertheless, a majority of respondents assume that Facebook photo tagging, Google Search, Netflix or Amazon recommendations, or Google Translate do not use AI. Why did so few respondents consider the products and services we listed to be applications of AI, automation, machine learning, or robotics? A straightforward explanation is that inattentive respondents neglect to carefully consider or select the items presented to them (i.e., non-response bias). Even among those assigned the term robotics, only 62% selected social robots and 68% selected industrial robots. Our analysis (found in Appendix C) confirms that respondent inattention, defined as spending too little or too much time on the survey, predicts non-response to this question. Another potential explanation for the results is that the American public -like the public elsewhere -lack awareness of AI or machine learning. As a result, the public does not know that many tech products and services use AI or machine learning. According to a 2017 survey, nearly half of Americans reported that they were unfamiliar with AI (Morning Consult 2017). In the same year, only 9% of the British public said they had heard of the term \"machine learning\"  (Ipsos MORI 2018) . Similarly, less than half of EU residents reported hearing, reading, or seeing something about AI in the previous year  (Eurobarometer 2017) . Finally, the so-called \"AI effect\" could also explain the survey result. The AI effect describes the phenomenon that the public does not consider an application that uses AI to utilize AI once that application becomes commonplace  (McCorduck 2004) . Because 85% of Americans report using digital products that deploy AI (e.g., navigation apps, video or music streaming apps, digital personal assistants on smartphones, etc.)  (Reinhart 2018) , they may not think that these everyday applications deploy AI. Electronic copy available at: https://ssrn.com/abstract=<PHONE> We sought to understand how Americans prioritize policy issues associated with AI. Respondents were asked to consider five AI governance challenges, randomly selected from a test of 13 (see Appendix B for the text); the order these five were to each respondent was also randomized. After considering each governance challenge, respondents were asked how likely they think the challenge will affect large numbers of people 1) in the U.S. and 2) around the world within 10 years. We use scatterplots to visualize our survey results. In Figure  3 .1, the x-axis is the perceived likelihood of the problem happening to large numbers of people in the U.S. In Figure  3 .2, the x-axis is the perceived likelihood of the problem happening to large numbers of people around the world. The y-axes on both Figure  3 .1 and 3.2 represent respondents' perceived issue importance, from 0 (not at all important) to 3 (very important). Each dot represents the mean perceived likelihood and issue importance, and the correspondent ellipse represents the 95% bivariate confidence region. Americans consider all the AI governance challenges we present to be important: the mean perceived issues importance of each governance challenge is between \"somewhat important\" (2) and \"very important\" (3), though there is meaningful and discernible variation across items. The AI governance challenges Americans think are most likely to impact large numbers of people, and are important for tech companies and governments to tackle, are found in the upper-right quadrant of the two plots. These issues include data privacy as well as AI-enhanced cyber attacks, surveillance, and digital manipulation. We note that the media have widely covered these issues during the time of the survey. There are a second set of governance challenges that are perceived on average, as about 7% less likely, and marginally less important. These include autonomous vehicles, value alignment, bias in using AI for hiring, the U.S.-China arms race, disease diagnosis, and technological unemployment. Finally, the third set of challenges are perceived on average another 5% less likely, and about equally important, including criminal justice bias and critical AI systems failures. We also note that Americans predict that all of the governance challenges mentioned in the survey, besides protecting data privacy and ensuring the safety of autonomous vehicles, are more likely to impact people around the world than to affect people in the U.S. While most of the statistically significant differences are substantively small, one difference stands out: Americans think that autonomous weapons are 7.6 percentage points more likely to impact people around the world than Americans. (See Appendix C for details of these additional analyses.) We want to reflect on one result. \"Value alignment\" consists of an abstract description of alignment problem and a reference to what sounds like individual level harms: \"while performing jobs [they could] unintentionally make decisions that go against the values of its human users, such as physically harming people.\" \"Critical AI systems failures,\" by contrast, references military or critical infrastructure uses, and unintentional accidents leading to \"10 percent or more of all humans to die.\" The latter was weighted as less important than the former: we interpret this as a probability weighted assessment of importance, since presumably the latter, were it to happen, is much more important. We thus think the issue importance question should be interpreted in a way that down-weights low probability risks. This perspective also plausibly applies to the \"impact\" measure for our global risks analysis, which placed \"harmful consequences of synthetic biology\" and \"failure to address climate change\" as less impactful than most other risks. Electronic copy available at: https://ssrn.com/abstract=<PHONE> Likelihood of impacting large numbers of people in the U.S. within 10 years Issue importance (0 = Not at all important; 3 = Very important) Source: Center for the Governance of AI Electronic copy available at: https://ssrn.com/abstract=<PHONE> Likelihood of impacting large numbers of people around the world within 10 years Issue importance (0 = Not at all important; 3 = Very important) Source: Center for the Governance of AI Electronic copy available at: https://ssrn.com/abstract=<PHONE> We performed further analysis by calculating the percentage of respondents in each subgroup who consider each governance challenge to be \"very important\" for governments and tech companies to manage. (See Appendix C for additional data visualizations.) In general, differences in responses are more salient across demographic subgroups than across governance challenges. In a linear multiple regression predicting perceived issue importance using demographic subgroups, governance challenges, and the interaction between the two, we find that the stronger predictors are demographic subgroup variables, including age group and having CS or programming experience. Two highly visible patterns emerge from our data visualization. First, a higher percentage of older respondents, compared with younger respondents, consider nearly all AI governance challenges to be \"very important.\" As discussed previously, we find that older Americans, compared with younger Americans, are less supportive of developing AI. Our results here might explain this age gap: older Americans see each AI governance challenge as substantially more important than do younger Americans. Whereas 85% of Americans older than 73 consider each of these issues to be very important, only 40% of Americans younger than 38 do. Second, those with CS or engineering degrees, compared with those who do not, rate all AI governance challenges as less important. This result could explain our previous finding that those with CS or engineering degrees tend to exhibit greater support for developing  AI. 8  Respondents were asked how much confidence they have in various actors to develop AI. They were randomly assigned five actors out of 15 to evaluate. We provided a short description of actors that are not well-known to the public (e.g., NATO, CERN, and OpenAI). Also, respondents were asked how much confidence, if any, they have in various actors to manage the development and use of AI in the best interests of the public. They were randomly assigned five out of 15 actors to evaluate. Again, we provided a short description of actors that are not well-known to the public (e.g., AAAI and Partnership on AI). Confidence was measured using the same four-point scale described above. 9 Americans do not express great confidence in most actors to develop or to manage AI, as reported in Figures 3.4 and 3.5. A majority of Americans do not have a \"great deal\" or even a \"fair amount\" of confidence in any institution, except university researchers, to develop AI. Furthermore, Americans place greater trust in tech companies and non-governmental organizations (e.g., OpenAI) than in governments to manage the development and use of the technology. University researchers and the U.S. military are the most trusted groups to develop AI: about half of Americans express a \"great deal\" or even a \"fair amount\" of confidence in them. Americans express slightly less confidence in tech companies, non-profit organizations (e.g., OpenAI), and American intelligence organizations. Nevertheless, opinions toward individual actors within each of these groups vary. For example, while 44% of Americans indicated they feel a \"great deal\" or even a \"fair amount\" of confidence in tech companies, they rate Facebook as the least trustworthy of all the actors. More than four in 10 indicate that they have no confidence in the company. 10 these two types of variables to predict perceived issue importance. We find that those who are 54-72 or 73 and older, relative to those who are below 38, view the governance issues as more important (two-sided p-value < 0.001 for both comparisons). Furthermore, we find that those who have CS or engineering degrees, relative to those who do not, view the governance challenges as less important (two-sided p-value < 0.001).  9  The two sets of 15 actors differed slightly because for some actors it seemed inappropriate to ask one or the other question. See Appendix B for the exact wording of the questions and descriptions of the actors.  10  Our survey was conducted between June 6 and 14, 2018, shortly after the fallout of the Facebook/Cambridge Analytica scandal. The results on the public's trust of various actors to manage the develop and use of AI provided are similar to the results discussed above. Again, a majority of Americans do not have a \"great deal\" or even a \"fair amount\" of confidence in any institution to manage AI. In general, the public expresses greater confidence in non-governmental organizations than in governmental ones. Indeed, 41% of Americans express a \"great deal\" or even a \"fair amount\" of confidence in \"tech companies,\" compared with 26% who feel that way about the U.S. federal government. But when presented with individual big tech companies, Americans indicate less trust in each than in the broader category of \"tech companies.\" Once again, Facebook stands out as an outlier: respondents give it a much lower rating than any other actor. Besides \"tech companies,\" the public places relatively high trust in intergovernmental research organizations (e.g., CERN), the Partnership on AI, and non-governmental scientific organizations (e.g., AAAI). Nevertheless, because the public is less familiar with these organizations, about one in five respondents give a \"don't know\" response. Mirroring our findings, recent survey research suggests that while Americans feel that AI should be regulated, they are unsure who the regulators should be. When asked who \"should decide how AI systems are designed and deployed,\" half of Americans indicated they do not know or refused to answer  (West 2018a) . Our survey results seem to reflect Americans' general attitudes toward public institutions. According to a 2016 Pew Research Center survey, an overwhelming majority of Americans have \"a great deal\" or \"a fair amount\" of confidence in the U.S. military and scientists to act in the best interest of the public. In contrast, public confidence in elected officials is much lower: 73% indicated that they have \"not too much\" or \"no confidence\"  (Funk 2017 ). Less than one-third of Americans thought that tech companies do what's right \"most of the time\" or \"just about always\"; moreover, more than half think that tech companies have too much power and influence in the U.S. economy  (Smith 2018) . Nevertheless, Americans' attitude toward tech companies is not monolithic but varies by company. For instance, our research findings reflect the results from a 2018 survey, which reported that a higher percentage of Americans trusted Apple, Google, Amazon, Microsoft, and Yahoo to protect user information than trust Facebook to do so  (Ipsos and Reuters 2018) . Electronic copy available at: https://ssrn.com/abstract=<PHONE> Yet, only a minority of the American public thinks the U.S. or China's AI R&D is the \"best in the world,\" as reported in Figure  4 .1. Our survey result seems to reflect the gap between experts and the public's perceptions of U.S.'s scientific achievements in general. While 45% of scientists in the American Association for the Advancement of Science think that scientific achievements in the U.S. are the best in the world, only 15% of the American public express the same opinion  (Funk and Rainie 2015) . According to our survey, there is not a clear perception by Americans that the U.S. has the best AI R&D in the world. While 10% of Americans believe that the U.S. has the best AI R&D in the world, 7% think that China does. Still, 36% of Americans believe that the U.S.'s AI R&D is \"above average\" while 45% think China's is \"above average.\" Combining these into a single measure of whether the country has \"above average\" or \"best in the world\" AI R&D, Americans do not perceive the U.S. to be superior, and the results lean towards the perception that China is superior. Note that we did not ask for a direct comparison, but instead asked each respondent to evaluate one country independently on an absolute scale Appendix C. Our results mirror those from a recent survey that finds that Americans think that China's AI capability will be on par with the U.S.'s in 10 years  (West 2018b ). The American public's perceptions could be caused by media narratives that China is catching up to the U.S. in AI capability (Kai-Fu 2018). Nevertheless, another study suggests that although China has greater access to big data than the U.S., China's AI capability is about half of the U.S.'s  (Ding 2018) . Exaggerating China's AI capability could exacerbate growing tensions between the U.S. and China  (Zwetsloot, Toner, and Ding 2018 ). As such, future research should explore how factual -non-exaggerated -information about American and Chinese AI capabilities influences public opinions. In this survey experiment, respondents were randomly assigned to consider different arguments about a U.S.-China arms race. (Details of the survey experiment are found in Appendix B.) All respondents were given the following prompt: Leading analysts believe that an AI arms race is beginning, in which the U.S. and China are investing billions of dollars to develop powerful AI systems for surveillance, autonomous weapons, cyber operations, propaganda, and command and control systems. Those in the treatment condition were told they would read a short news article. The three treatments were:  Electronic copy available at: https://ssrn.com/abstract=<PHONE> 3. One common humanity treatment: The U.S.-China arms race could increase the risk of a catastrophic war; quote from Stephen Hawking about using AI for the good of all people rather than destroying civilization Respondents were asked to consider two statements and indicate whether they agree or disagree with them: • The U.S. should invest more in AI military capabilities to make sure it doesn't fall behind China's, even if doing so may exacerbate the AI arms race. • The U.S. should work hard to cooperate with China to avoid the dangers of an AI arms race, even if doing so requires giving up some of the U.S.'s advantages. Cooperation could include collaborations between American and Chinese AI research labs, or the U.S. and China creating and committing to common safety standards for AI. Americans, in general, weakly agree that the U.S. should invest more in AI military capabilities and cooperate with China to avoid the dangers of an AI arms race, as seen in In contrast, respondents assigned to the other conditions indicate similar levels of agreement with both statements. After estimating the treatment effects, we find that the experimental messages do little to change the respondents' preferences. Treatment 2 is the one exception. Treatment 2 decreases respondents' agreement with the statement that the U.S. should invest more in AI military capabilities by 27%, as seen in Figure  4 .3. Future research could focus on testing more effective messages, such as op-eds  (Coppock et al. 2018)  or videos  (Paluck et al. 2015) , which explains that U.S.'s investment in AI for military use will decrease the likelihood of cooperation with China. We examined issue areas where Americans perceive likely U.S.-China cooperation. Each respondent was randomly assigned to consider three out of five AI governance challenges. For each challenge, the respondent was asked, \"For the following issues, how likely is it that the U.S. and China can cooperate?\". (See Appendix B for the question text.) On each of these AI governance issues, Americans see some potential for U.S.-China cooperation, according to Figure  4 .5. U.S.-China cooperation on value alignment is perceived to be the most likely (48% mean likelihood). Cooperation to prevent AI-assisted surveillance that violates privacy and civil liberties is seen to be the least likely (40% mean likelihood) -an unsurprising result since the U.S. and China take different stances on human rights. Despite current tensions between Washington and Beijing, the Chinese government, as well as Chinese companies and academics, have signaled their willingness to cooperate on some governance issues. These include banning the use of lethal autonomous weapons  (Kania 2018) , building safe AI that is aligned with human values (China Institute for Science and Technology Policy at Tsinghua University 2018), and collaborating on research  (News 2018) . Most recently, the major tech company Baidu became the first Chinese member of the Partnership on AI, a U.S.-based multi-stakeholder organization committed to understanding and discussing AI's impacts  (Cadell 2018 ). In the future, we plan to survey Chinese respondents to understand how they view U.S.-China cooperation on AI and what governance issues they think the two countries could collaborate on. Electronic copy available at: https://ssrn.com/abstract=<PHONE>   Electronic copy available at: https://ssrn.com/abstract=<PHONE> Survey questions measuring Americans' perceptions of workplace automation have existed since the 1950s. Our research seeks to track changes in these attitudes across time by connecting past survey data with original, contemporary survey data. American government agencies, think tanks, and media organizations began conducting surveys to study public opinion about technological unemployment during the 1980s when unemployment was relatively high. Between 1983 and 2003, the U.S. National Science Foundation (NSF) conducted eight surveys that asked respondents the following: In general, computers and factory automation will create more jobs than they will eliminate. Do you strongly agree, agree, disagree, or strongly disagree? Our survey continued this time trend study by posing a similar -but updated -question (see Appendix B): Do you strongly agree, agree, disagree, or strongly disagree with the statement below? In general, automation and AI will create more jobs than they will eliminate. Our survey question also addressed the chief ambiguity of the original question: lack of a future time frame. We used a survey experiment to help resolve this ambiguity by randomly assigning respondents to one of four conditions. We created three treatment conditions with the future time frames of 10 years, 20 years, and 50 years, as well as a control condition that did not specify a future time frame. On average, Americans disagree with the statement more than they agree with it, although about a quarter of respondents in each experimental group give \"don't know\" responses. Respondents' agreement with the statement seems to increase slightly with the future time frame, but formal tests in Apppendix C reveal that there exist no significant differences between the responses to the differing future time frames. This result is puzzling from the perspective that AI and robotics will increasingly automate tasks currently done by humans. Such a view would expect more disagreement with the statement as one looks further into the future. One hypothesis to explain our results is that respondents believe the disruption from automation is destabilizing in the upcoming 10 years but eventually institutions will adapt and the labor market will stabilize. This hypothesis is consistent with our other finding that the median American predicts a 54% chance of high-level machine intelligence being developed within the next 10 years. The percentage of Americans that disagrees with the statement that automation and AI will create more jobs than they destroy is similar to the historical rate of disagreement with the same statement about computers and factory automation. Nevertheless, the percentage who agree with the statement has decreased by 12 percentage points since 2003 while the percentage who responded \"don't know\" has increased by 18 percentage points since 2003, according to Figure  5 .2. There are three possible reasons for these observed changes. First, we have updated the question to ask about \"automation and AI\" instead of \"computers and factory automation.\" The technologies we asked about could impact a wider swath of the economy; therefore, respondents may be more uncertain about AI's impact on the labor market. Second, there is a difference in survey mode between the historical data and our data. The NSF surveys were conducted via telephone while our survey is conducted online. Some previous research has shown that online surveys, compared with telephone surveys, produce a greater percentage of \"don't know\" responses  (Nagelhout et al. 2010; Bronner and Kuijlen 2007) . But, other studies have shown that online surveys cause no such effect  (Shin, Johnson, and Rao 2012; Bech and Kristensen 2009) . Third, the changes in the responses could be due to the actual changes in respondents' perceptions of workplace automation over time. Electronic copy available at: https://ssrn.com/abstract=<PHONE> Respondents were asked to forecast when high-level machine intelligence will be developed. High-level machine intelligence was defined as the following: We have high-level machine intelligence when machines are able to perform almost all tasks that are economically relevant today better than the median human (today) at each task. These tasks include asking subtle common-sense questions such as those that travel agents would ask. For the following questions, you should ignore tasks that are legally or culturally restricted to humans, such as serving on a jury.  13  Respondents were asked to predict the probability that high-level machine intelligence will be built in 10, 20, and 50 years. We present our survey results in two ways. First, we show the summary statistics in a simple table. Next, to compare the public's forecasts with forecasts made by AI researchers in 2016  (Grace et al. 2018) , we aggregated the respondents' forecasts using the same method. Note that  Grace et al. (2018)  gave a stricter definition of high-level machine intelligence that involved machines being better than all humans at all tasks. 14 Respondents predict that high-level machine intelligence will arrive fairly quickly. The median respondent predicts a likelihood of 54% by 2028, a likelihood of 70% by 2038, and a likelihood of 88% by 2068, according to Table  6 .1. These predictions are considerably sooner than the predictions by experts in two previous surveys. In  Müller and Bostrom (2014) , expert respondents predict a 50% probability of high-level human intelligence being developed by 2040-2050 and 90% by 2075. In  Grace et al. (2018) , experts predict that there is a 50% chance that high-level machine intelligence will be built by 2061. Plotting the public's forecast with the expert forecast from  Grace et al. (2018) , we see that the public predicts high-level machine intelligence arriving much sooner than experts forecast. Employing the same aggregation method used in  Grace et al. (2018) , Americans predict that there is a 50% chance that high-level machine intelligence will be developed by 2026. Results in Walsh (2018) also show that the non-experts (i.e., readers of a news article about AI) are more optimistic in their predictions of high-level machine intelligence compared with experts. In Walsh's study, the median AI expert predicted a 50% probability of high-level machine intelligence by 2061 while the median non-expert predicted a 50% probability by 2039. In our survey, respondents with CS or engineering degrees, compared with those who do not, provide a somewhat longer timeline for the arrival of high-level machine intelligence, according to   (2018) ; furthermore, their forecasts show considerable overlap with the overall public forecast (see Figure  6 .1). The above differences could be due to different definitions of high-level machine intelligence presented to respondents. However, we suspect that it is not the case for the following reasons. (1) These differences in timelines are larger, more significant than we think could be reasonably attributed to beliefs about these different levels of intelligence. (2) We found similar results using the definition in  Grace et al. (2018) , on a (different) sample of the American public. In a pilot survey conducted on Mechanical Turk during July 13-14, 2017, we asked American respondents about human-level AI, defined as the following: Human-level artificial intelligence (human-level AI) refers to computer systems that can operate with the intelligence of an average human being. These programs can complete tasks or make decisions as successfully as the average human can. In this pilot study, respondents also provided forecasts that are more optimistic than the projections by AI experts. The respondents predict a median probability of 44% by 2027, a median probability of 62% by 2037, and a median probability of 83% by 2067. Electronic copy available at: https://ssrn.com/abstract=<PHONE> Support for developing high-level machine intelligence varies greatly between demographic subgroups, although only a minority in each subgroup supports developing the technology. Some of the demographic trends we observe regarding support for developing AI also are evident regarding support for high-level machine intelligence. Men (compared with women), high-income Americans (compared with low-income Americans), and those with tech experience (compared with those without) express greater support for high-level machine intelligence. We used a multiple regression that includes all of the demographic variables to predict support for developing high-level machine intelligence. The support for developing AI outcome variable was standardized, so it has mean 0 and unit variance. Significant predictors correlated with support for developing high-level machine intelligence include: • Being male (versus being female) • Identifying as a Republican (versus identifying as an Independent or \"other\")  16  • Having a family income of more than $100,000 annually (versus having a family income of less than $30,000 annually) • Having CS or programming experience (versus not having such experience) This last result about women less supportive of developing high-level machine intelligence than men is noteworthy as it speaks to the contrary claim sometimes made that it is primarily men who are concerned about the risks from advanced AI. Men are argued to be disproportionately worried about human-level AI because of reasons related to evolutionary psychology  (Pinker 2018)  or because they have the privilege of not confronting the other harms from AI, such as biased algorithms  (Crawford 2016) . We also performed the analysis above but controlling for respondents' support for developing AI (see Appendix). Doing so allows us to identify subgroups those attitudes toward AI diverges from their attitudes toward high-level machine intelligence. In this secondary analysis, we find that being 73 or older is a significant predictor of support for developing high-level machine intelligence. In contrast, having a four-year college degree is a significant predictor of opposition to developing high-level machine intelligence. These are interesting inversions of the bivariate association, where older and less educated respondents were more concerned about AI; future work could explore this nuance. This question sought to quantify respondents' expected outcome of high-level machine intelligence. (See Appendix B for the question text.) Respondents were asked to consider the following: Suppose that high-level machine intelligence could be developed one day. How positive or negative do you expect the overall impact of high-level machine intelligence to be on humanity in the long run? Americans, on average, expect that high-level machine intelligence will have a harmful impact on balance. Overall, 34% think that the technology will have a harmful impact; in particular, 12% said it could be extremely bad, leading to possible human extinction. More than a quarter of Americans think that high-level machine intelligence will be good for humanity, with 5% saying it will be extremely good. Since forecasting the impact of such technology on humanity is highly uncertain, 18% of respondents selected \"I don't know.\" The correlation between one's expected outcome and one's support for developing high-level machine intelligence is 0.69. A similar question was asked to AI experts in  Grace et al. (2018) ; instead of merely selecting one expected outcome, the AI experts were asked to predict the likelihood of each outcome. In contrast to the general public, the expert respondents think that high-level machine intelligence will be more beneficial than harmful. 17 Although they assign, on average, a 27% probability of high-level machine intelligence of being extremely good for humanity, they also assign, on average, a 9% probability of the technology being extremely bad, including possibly causing human extinction.  16  In the survey, we allowed those who did not identify as Republican, Democrat, or Independent to select \"other.\" The difference in responses between Republicans and Democrats is not statistically significant at the 5% level. Nevertheless, we caution against over-interpreting these results related to respondents' political identification because the estimated differences are substantively small while the correlating confidence intervals are wide.  17  To make the two groups' results more comparable, we calculated the expected value of the experts' predicted outcomes so that it is on the same -2 to 2 scale as the public's responses. To calculate this expected value, we averaged the sums of each expert's predicted likelihoods multiplied by the corresponding outcomes; we used the same numerical outcome as described in the previous subsection. The expected value of the experts' predicted outcomes is 0.08, contrasted with the public's average response of -0.17. Electronic copy available at: https://ssrn.com/abstract=<PHONE>  Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> YouGov interviewed 2,387 respondents who were then matched down to a sample of 2,000 to produce the final dataset. The respondents were matched to a sampling frame on gender, age, race, and education. The frame was constructed by stratified sampling from the full 2016 American Community Survey (ACS) one-year sample with selection within strata by weighted sampling with replacements (using the person weights on the public use file). The matched cases were weighted to the sampling frame using propensity scores. The matched cases and the frame were combined and a logistic regression was estimated for inclusion in the frame. The propensity score function included age, gender, race/ethnicity, years of education, and geographic region. The propensity scores were grouped into deciles of the estimated propensity score in the frame and post-stratified according to these deciles. The weights were then post-stratified on 2016 U.S. presidential vote choice, and a four-way stratification of gender, age (four-categories), race (four-categories), and education (four-categories), to produce the final weight. We use the following demographic subjects in our analysis: We pre-registered the analysis of this survey on Open Science Framework. Pre-registration increases research transparency by requiring researchers to specify their analysis before analyzing the data  (Nosek et al. 2018) . Doing so prevents researchers from misusing data analysis to come up with statistically significant results when they do not exist, otherwise known as p-hacking. Unless otherwise specified, we performed the following procedure: • Survey weights provided by YouGov were used in our primary analysis. For transparency, Appendix B contains the unweighted topline results, including raw frequencies. • For estimates of summary statistics or coefficients, \"don't know\" or missing responses were re-coded to the weighted overall mean, unconditional on treatment conditions. Almost all questions had a \"don't know\" option. If more than 10% of the variable's values were don't know\" or missing, we included a (standardized) dummy variable for \"don't know\"/missing in the analysis. For survey experiment questions, we compared \"don't know\"/missing rates across experimental conditions. Our decision was informed by the Standard Operating Procedures for Don Green's Lab at Columbia University (Lin and Green 2016). • Heteroscedasticity-consistent standard errors were used to generate the margins of error at the 95% confidence level. We report cluster-robust standard errors whenever there is clustering by respondent. In figures, each error bar shows the 95% confidence intervals. Each confidence ellipse shows the 95% confidence region of the bivariate means assuming the two variables are distributed multivariate normal. • In regression tables, * denotes p < 0.05, ** denotes p < 0.01, and *** denotes p < 0.001. We plan to make our survey data, as well as the R and Markdown code that produced this report, publicly available through the Harvard Dataverse six months after the publication of this report. Below, we present the survey text as shown to respondents. The numerical codings are shown in parentheses following each answer choice. In addition, we report the topline results: percentages weighted to be representative of the U.S. adult population, the unweighted raw percentages, and the raw frequencies. Note that in all survey experiments, respondents were randomly assigned to each experimental group with equal probability. [All respondents were presented with the following prompt.] We want to get your opinion about global risks. A \"global risk\" is an uncertain event or condition that, if it happens, could cause a significant negative impact for at least 10 percent of the world's population. That is at least 1 in 10 people around the world could experience a significant negative impact. You will be asked to consider 5 potential global risks. [Respondents were presented with five items randomly selected from the list below. One item was shown at a time.] • Failure to address climate change: Continued failure of governments and businesses to pass effective measures to reduce climate change, protect people, and help those impacted by climate change to adapt. • Failure of regional or global governance: Regional organizations (e.g., the European Union) or global organizations (e.g., the United Nations) are unable to resolve issues of economic, political, or environmental importance. • Conflict between major countries: Disputes between major countries that lead to economic, military, cyber, or societal conflicts. • Weapons of mass destruction: Use of nuclear, chemical, biological or radiological weapons, creating international crises and killing large numbers of people. • Large-scale involuntary migration: Large-scale involuntary movement of people, such as refugees, caused by conflict, disasters, environmental or economic reasons. • Rapid and massive spread of infectious diseases: The uncontrolled spread of infectious diseases, for instance as a result of resistance to antibiotics, that leads to widespread deaths and economic disruptions. • Water crises: A large decline in the available quality and quantity of fresh water that harms human health and economic activity. • Food crises: Large numbers of people are unable to buy or access food. Harmful consequences of artificial intelligence (AI): Intended or unintended consequences of artificial intelligence that causes widespread harm to humans, the economy, and the environment. • Harmful consequences of synthetic biology: Intended or unintended consequences of synthetic biology, such as genetic engineering, that causes widespread harm to humans, the economy, and the environment. • Large-scale cyber attacks: Large-scale cyber attacks that cause large economic damages, tensions between countries, and widespread loss of trust in the internet. • Large-scale terrorist attacks: Individuals or non-government groups with political or religious goals that cause large numbers of deaths and major material damage. • Global recession: Economic decline in several major countries that leads to a decrease in income and high unemployment. • Extreme weather events: Extreme weather events that cause large numbers of deaths as well as damage to property, infrastructure, and the environment. • Major natural disasters: Earthquakes, volcanic activity, landslides, tsunamis, or geomagnetic storms that cause large numbers of deaths as well as damage to property, infrastructure, and the environment. \n QUESTION: What is the likelihood of [INSERT GLOBAL RISK] happening globally within the next 10 years? Please use the slider to indicate your answer. 0% chance means it will certainly not happen and 100% chance means it will certainly happen. Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> [Respondents were presented with one statement randomly selected from the list below.] • AI and robots are technologies that require careful management. • AI is a technology that requires careful management. • Robots are technologies that require careful management. ANSWER CHOICES: • Totally agree (2) • Tend to agree (1) • Tend to disagree (-1) • Totally disagree (-2) • I don't know   We would like you to consider some potential policy issues related to AI. Please consider the following: [Respondents were shown five randomly-selected items from the list below, one item at a time. For ease of comprehension, we include the shorten labels used in the figures in square brackets.] • [Hiring bias] Fairness and transparency in AI used in hiring: Increasingly, employers are using AI to make hiring decisions. AI has the potential to make less biased hiring decisions than humans. But algorithms trained on biased data can lead to lead to hiring practices that discriminate against certain groups. Also, AI used in this application may lack transparency, such that human users do not understand what the algorithm is doing, or why it reaches certain decisions in specific cases. • [Criminal justice bias] Fairness and transparency in AI used in criminal justice: Increasingly, the criminal justice system is using AI to make sentencing and parole decisions. AI has the potential to make less biased hiring decisions than humans. But algorithms trained on biased data could lead to discrimination against certain groups. Also, AI used in this application may lack transparency such that human users do not understand what the algorithm is doing, or why it reaches certain decisions in specific cases. become more advanced, they will increasingly make decisions without human input. One potential fear is that AI systems, while performing jobs they are programmed to do, could unintentionally make decisions that go against the values of its human users, such as physically harming people. • [Autonomous weapons] Ban the use of lethal autonomous weapons (LAWs): Lethal autonomous weapons (LAWs) are military robots that can attack targets without control by humans. LAWs could reduce the use of human combatants on the battlefield. But some worry that the adoption of LAWs could lead to mass violence. Because they are cheap and easy to produce in bulk, national militaries, terrorists, and other groups could readily deploy LAWs. • [Technological unemployment] Guarantee a good standard of living for those who lose their jobs to automation: Some forecast that AI will increasingly be able to do jobs done by humans today. AI could potentially do the jobs of blue-collar workers, like truckers and factory workers, as well as the jobs of white-collar workers, like financial analysts or lawyers. Some worry that in the future, robots and computers can do most of the jobs that are done by humans today. • [Critical AI systems failure] Prevent critical AI systems failures: As AI systems become more advanced, they could be used by the military or in critical infrastructure, like power grids, highways, or hospital networks. Some worry that the failure of AI systems or unintentional accidents in these applications could cause 10 percent or more of all humans to die. \n QUESTION: In the next 10 years, how likely do you think it is that this AI governance challenge will impact large numbers of people in the U.S.? ANSWER CHOICES: • Very unlikely: less than 5% chance (2.5%) • Unlikely: 5-20% chance (12.5%) • Somewhat unlikely: 20-40% chance (30%) • Equally likely as unlikely: 40-60% chance (50%) • Somewhat likely: 60-80% chance (70%) Electronic copy available at: https://ssrn.com/abstract=<PHONE> We want to understand your thoughts on some important issues in the news today. Please read the short news article below. Leading analysts believe that an \"AI arms race\" is beginning, in which the U.S. and China are investing billions of dollars to develop powerful AI systems for surveillance, autonomous weapons, cyber operations, propaganda, and command and control systems. [Respondents were randomly assigned to one of the four experimental groups listed below.] [No additional text.] Some leaders in the U.S. military and tech industry argue that the U.S. government should invest much more resources in AI research to ensure that the U.S.'s AI capabilities stay ahead of China's. Furthermore, they argue that the U.S. government should partner with American tech companies to develop advanced AI systems, particularly for military use. According to a leaked memo produced by a senior National Security Council official, China has \"assembled the basic components required for winning the Al arms race…Much like America's success in the competition for nuclear weapons, China's 21st Century Manhattan Project sets them on a path to getting there first.\" Some prominent thinkers are concerned that a U.S.-China arms race could lead to extreme dangers. To stay ahead, the U.S. and China may race to deploy advanced military AI systems that they do not fully understand or can control. We could see catastrophic accidents, such as a rapid, automated escalation involving cyber and nuclear weapons. \"Competition for AI superiority at [the] national level [is the] most likely cause of World War Three,\" warned Elon Musk, the CEO of Tesla and SpaceX. Some prominent thinkers are concerned that a U.S.-China arms race could lead to extreme dangers. To stay ahead, the U.S. and China may race to deploy advanced military AI systems that they do not fully understand or can control. We could see catastrophic accidents, such as a rapid, automated escalation involving cyber and nuclear weapons. \"Unless we learn how to prepare for, and avoid, the potential risks, AI could be the worst event in the history of our civilization. It brings dangers, like powerful autonomous weapons,\" warned the late Stephen Hawking, one of the world's most prominent physicists. At the same time, he said that with proper management of the technology, researchers \"can create AI for the good of the world.\" [The order of the next two questions is randomized.] QUESTION: How much do you agree or disagree with the following statement? The U.S. should invest more in AI military capabilities to make sure it doesn't fall behind China's, even if doing so may exacerbate the arms race. For instance, the U.S. could increase AI research funding for the military and universities. It could also collaborate with American tech companies to develop AI for military use. ANSWER CHOICES: • Strongly agree (2) • Somewhat agree (1) • Neither agree nor disagree (0) • Somewhat disagree (-1) • Strongly disagree (-2) • I don't know QUESTION: How much do you agree or disagree with the following statement? The U.S. should work hard to cooperate with China to avoid the dangers of an AI arms race, even if doing so requires giving up some of the U.S.'s advantages. Cooperation could include collaborations between American and Chinese AI research labs, or the U.S. and China creating and committing to common safety standards. ANSWER CHOICES: • Strongly agree (2) • Somewhat agree (1) • Neither agree nor disagree (0) • Somewhat disagree (-1) Electronic copy available at: https://ssrn.com/abstract=<PHONE> [Respondents were presented with three issues from the list below. All three issues were presented on the same page; the order that they appeared was randomized.] • Prevent AI cyber attacks against governments, companies, organizations, and individuals. • Prevent AI-assisted surveillance from violating privacy and civil liberties. • Make sure AI systems are safe, trustworthy, and aligned with human values. • Ban the use of lethal autonomous weapons. • Guarantee a good standard of living for those who lose their jobs to automation. ANSWER CHOICES: • Very unlikely: less than 5% chance (2.5%) How much do you agree or disagree with the following statement? [Respondents were presented with one statement randomly selected from the list below.] • In general, automation and AI will create more jobs than they will eliminate. • In general, automation and AI will create more jobs than they will eliminate in 10 years. • In general, automation and AI will create more jobs than they will eliminate in 20 years. • In general, automation and AI will create more jobs than they will eliminate in 50 years. ANSWER CHOICES: • Strongly agree (2) Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> QUESTION: The following questions ask about high-level machine intelligence. We have high-level machine intelligence when machines are able to perform almost all tasks that are economically relevant today better than the median human (today) at each task. These tasks include asking subtle common-sense questions such as those that travel agents would ask. For the following questions, you should ignore tasks that are legally or culturally restricted to humans, such as serving on a jury. In your opinion, how likely is it that high-level machine intelligence will exist in 10 years? 20 years? 50 years? For each prediction, please use the slider to indicate the percent chance that you think high-level machine intelligence will exist. 0% chance means it will certainly not exist. 100% chance means it will certainly exist. \n ______ In 10 years? ______ In 20 years? ______ In 50 years? ANSWER CHOICES: • Very unlikely: less than 5% chance (2.5%) Suppose that high-level machine intelligence could be developed one day. How positive or negative do you expect the overall impact of high-level machine intelligence to be on humanity in the long run? Electronic copy available at: https://ssrn.com/abstract=<PHONE> Next, we investigated the problem of respondents not selecting technological applications where it would be logical to pick them (e.g., not selecting industrial robots or social robots when presented with the term \"robotics\"). Our regression analysis shows that this type of non-response is correlated with respondents' inattention. We used two measures as a proxy for inattention: 1. time to complete the survey 2. the absolute deviation from the median time to complete the survey. Because the distribution of completion times is heavily skewed right, we used absolute deviation from the median, as opposed to the mean. The median is 13 minutes whereas the mean is 105 minutes. We incorporated the second measure because we suspected that people who took very little time or a very long time to complete the survey were inattentive. We used three outcomes that measured non-response: 1. the number of items selected 2. not selecting \"none of the above\" 3. selecting items containing the word \"robots\" for respondents assigned to consider \"robotics\" Using multiple regression, we showed that inattention predicts non-response measured by the three outcomes above (see Tables C.9, C.10, and C.11). We compared respondents' perceived likelihood of each governance challenge impacting large numbers of people in the U.S. with respondents' perceived likelihood of each governance challenge impacting large numbers of people around the world. (See Appendix B for the survey question text.) For each governance challenge, we used linear regression to estimate the difference between responses to the U.S. question and the world question. Because we ran 13 tests, we used the Bonferroni correction to control the familywise error rate. In our case, the Bonferroni correction rejected the null hypothesis at alpha level α/13, instead of α. To test whether the differences are significant at the 5% level, we set the alpha level at α/13 = 0.004. According to Table  C .12, Americans perceive that all governance challenges, except for protecting data privacy and ensuring safe autonomous vehicles, are more likely to impact people around the world than in the U.S. specifically. In particular, Americans think that autonomous weapons are 7.6 percentage points more likely to impact people around the world than in the U.S. Electronic copy available at: https://ssrn.com/abstract=<PHONE> To highlight the differences between the responses of demographic subgroups regarding issue importance, we created an additional graph  (Figure C.1) . Here, we subtracted the overall mean of perceived issue importance across all responses from each subgroup-governance challenge mean.  19  Table  C .15 shows the results from a saturated regression predicting perceived issue importance using demographic variables, AI governance challenge, and interactions between the two types of variables. Electronic copy available at: https://ssrn.com/abstract=<PHONE>    Electronic copy available at: https://ssrn.com/abstract=<PHONE>   5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.1 9.2 5.9 3.6 0.6 0.2 0.8 2.8 10.1 2.2 0.5 <0.1 0.9 1.7 5.5 11.1 4.7 0.3 5.5 0.4 2.7 5.7 12.6 0.9 1.1 0.1 0.1 0.7 2.9 4.2 0.1 0.7 1.0 0.8 0.4 <0.   Figure 2 . 1 : 21 Figure 2.1: Support for developing AI \n Figure 2.2: Support for developing AI across demographic characteristics: distribution of responses \n Figure 3 . 1 : 31 Figure 3.1: Perceptions of AI governance challenges in the U.S. \n Figure 3 . 2 : 32 Figure 3.2: Perceptions of AI governance challenges around the world \n Figure 4 . 1 : 41 Figure 4.1: Comparing Americans' perceptions of U.S. and China's AI research and development quality \n Figure 4.2. Many respondents do not think that the two policies are mutually exclusive. The correlation between responses to the two statements, unconditional on treatment assignment, is only -0.05. In fact, 29% of those who agree that the U.S. and China should cooperate also agree that the U.S. should invest more in AI military capabilities. (See Figure C.2 for the conditional percentages.) Respondents assigned to read about the risks of an arms race (Treatment 2) indicate significantly higher agreement with the pro-cooperation statement (Statement 2) than the investing in AI military capabilities statement (Statement 1), according to Figure 4.4. Those assigned to Treatment 2 are more likely to view the two statements as mutually exclusive. \n Figure 4 . 2 : 42 Figure 4.2: Responses from U.S.-China arms race survey experiment \n Figure 4 4 Figure 4.3: Effect estimates from U.S.-China arms race survey experiment \n Figure5.1: Agreement with the statement that automation and AI will create more jobs than it will eliminate \n Figure 6 6 Figure 6.3: Support for developing high-level machine intelligence across demographic characteristics: distribution of responses \n deal of confidence (3) • A fair amount of confidence (2) • Not too much confidence (1) • No confidence (0) • I don't know \n One common humanity -0.02 (0.02) N = 2000 F(3, 1996) = 0.76; p-value: 0.516 \n Figure C. 2 : 2 Figure C.2: Correlation between responses to the two statements from survey experiment \n FigureCSFigure Figure C.3: Mean predicted likelihood of high-level machine intelligence for each year by demographic subgroup \n Figure C. 6 : 6 Figure C.6: Correlation between expected outcome and support for developing high-level machine intelligence \n Table 6 . 6 1: Summary statistics of high-level machine intelligence forecast Year Respondent type 25th percentile Median Mean 75th percentile N 2028 All respondents 30% 54% 54% 70% 2000 2038 All respondents 50% 70% 70% 88% 2000 2068 All respondents 70% 88% 80% 97% 2000 2028 No CS or engineering degree 30% 54% 55% 70% <PHONE> No CS or engineering degree 50% 70% 71% 88% <PHONE> No CS or engineering degree 70% 88% 81% 98% <PHONE> CS or engineering degree 30% 50% 48% 70% <PHONE> CS or engineering degree 50% 70% 67% 88% <PHONE> CS or engineering degree 50% 73% 69% 97% 195 \n Table 6 6 .1. Nevertheless, those with CS or Source: Center for the Governance of AI Figure6.1: The American public's forecasts of high-level machine intelligence timelines engineering degrees in our sample provide forecasts are more optimistic than those made by experts fromGrace et al.    \n Completed a computer science or engineering degree in undergraduate or graduate school: yes, no • Has computer science or programming experience: yes, noWe report the unweighted sample sizes of the demographic subgroups in TableA.1. Demographic subgroups Unweighted sample sizes Not employed 1036 Employed (full-or part-time) 964 Income less than $30K 531 Income $30-70K 626 Income $70-100K 240 Income more than $100K 300 Prefer not to say income 303 Republican 470 Democrat 699 Independent/Other 831 Christian 1061 No religious affiliation 718 Other religion 221 Not born-again Christian 1443 Born-again Christian 557 No CS or engineering degree 1805 CS or engineering degree 195 No CS or programming experience 1265 CS or programming experience <PHONE>) • Gender: male, female • Race: white, non-white • Level of education: graduated from high school or less, some college (including two-year college), graduated from a four-year college or more • Employment status: employed (full-or part-time), not employed • Annual household income: less than $30,000, $30,000-70,000, $70,000-100,000, more than $100,000, prefer not to say • Political party identification: Democrats (includes those who lean Democrat), Republicans (includes those who lean Republican), Independents/Others • Religion: Christian, follow other religions, non-religious • Identifies as a born-again Christian: yes, no • Table A.1: Size of demographic subgroups Demographic subgroups Unweighted sample sizes Age 18-37 702 Age 38-53 506 Age 54-72 616 Age 73 and older 176 Female 1048 Male 952 White 1289 Non-white 711 HS or less 742 Some college 645 College+ 613 • Age group as defined by Pew Research Center: Millennial/post-Millennial adults (born after 1980; ages 18-37 in 2018), Gen Xers (born <PHONE>; ages 38-53 in 2018), Baby Boomers (born <PHONE>; ages 54-72 in 2018), Silents/Greatest Generation (1945 and earlier; ages 73 and over in Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n Table B . B 17: Size of negative impact -Failure of regional/global governance; N = 652 Responses Minimal Table B.21: Size of negative impact -Spread of infectious diseases; N = Minimal Percentages (weighted) Percentages (unweighted) Raw frequencies 6.04 5.98 2.72 2.58 Minor Minor 6.09 6.03 5.67 5.65 Moderate Moderate 28.68 26.86 28.99 28.06 Severe Severe 33.21 32.00 34.05 32.58 Catastrophic 10.76 Catastrophic 20.50 10.89 20.48 I don't know 15.12 I don't know 11.88 14.26 10.65 Skipped Skipped 0.10 0 0.15 0 Table B.18: Size of negative impact -Conflict between major countries; N Table B.22: Size of negative impact -Water crises; N = 623 = 625 Responses Percentages (weighted) Percentages (unweighted) Raw frequencies Responses Minimal Percentages (weighted) Percentages (unweighted) Raw frequencies 1.72 1.93 Minimal Minor 1.18 4.42 0.96 4.65 Minor Moderate 4.94 19.92 4.80 19.42 Moderate Severe 28.81 36.71 28.16 36.44 Severe Catastrophic 27.24 38.23 38.56 28.25 Catastrophic 14.80 I don't know 10.00 16.00 9.31 I don't know 11.89 Skipped 0 11.36 0 Skipped 0.14 0.16 Table B.23: Size of negative impact -Food crises; N = 1073 Table B.19: Size of negative impact -Weapons of mass destruction; N = 645 Responses Percentages (weighted) Percentages (unweighted) Raw frequencies Responses Minimal Minimal Minor Minor Moderate Moderate Severe Severe Catastrophic 24.04 2.55 Percentages (weighted) Percentages (unweighted) Raw frequencies 2.61 2.28 7.22 6.99 2.17 4.99 22.81 22.37 4.19 13.57 33.93 34.67 13.49 31.05 24.88 31.01 Catastrophic 38.06 I don't know 9.38 8.39 39.38 I don't know 10.05 9.77 Skipped 0.08 0.09 Skipped 0 0 Table B.24: Size of negative impact -Harmful consequences of AI; N = 573 Table B.20: Size of negative impact -Large-scale involuntary migration; N = 628 Responses Percentages (weighted) Percentages (unweighted) Raw frequencies Minimal 7.54 7.50 Responses Minor Percentages (weighted) Percentages (unweighted) Raw frequencies 14.82 13.79 Minimal Moderate Minor Severe Moderate Catastrophic 14.62 2.07 27.77 8.67 20.46 25.63 Severe 35.31 I don't know 14.79 Catastrophic 18.14 Skipped 0 27.92 2.07 21.82 8.28 14.31 25.96 14.66 36.15 17.83 0 I don't know 9.99 9.55 Skipped 0.19 0.16 \n Table B . B 29: Size of negative impact -Extreme weather events; N = 613 Virtual assistants (e.g., Siri, Facebook photo tagging Table B.35: Computer science/technology background; N = 2000 62.87 64.30 35.37 36.81 Google Assistant, Amazon Google Search 45.42 46.26 Responses Recommendations for Netflix Percentages (weighted) Percentages (unweighted) Raw frequencies Alexa) 37.97 38.19 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Minimal Minor Moderate Severe Catastrophic 22.79 2.54 6.69 25.94 32.50 I don't know 9.56 Skipped 0 Smart speakers (e.g., movies or Amazon ebooks Took at least one college-level course in CS 24.73 55.46 Amazon Echo, Google Home, Google Translate 33.40 CS or engineering undergraduate degree 7.12 Apple Homepod) Facebook photo tagging 36.16 Driverless cars and trucks 52.96 CS or engineering graduate degree 3.85 Google Search 35.59 Social robots that can 59.19 Have programming experience 10.88 Recommendations for Netflix manufacturing movies or Amazon ebooks 27.73 Industrial robots used in 37.41 interact with humans None of the above 63.68 2.45 6.53 26.43 31.97 23.00 9.62 0 56.19 34.06 54.33 36.51 59.45 36.51 37.80 29.01 25.05 7.30 3.75 11.10 63.20 15 40 162 196 141 59 0 501 146 75 <PHONE> Google Translate Drones that do not require a 29.49 49.03 30.02 49.41 Driverless cars and trucks human controller 56.38 57.20 Table B.30: Size of negative impact -Natural disasters; N = 637 Social robots that can interact with humans 63.63 64.10 [All respondents were presented with the following prompt.] Responses Minimal Minor Moderate Industrial robots used in Percentages (weighted) Percentages (unweighted) Raw frequencies 1.29 1.26 8 5.86 5.81 37 22.26 23.08 147 40.11 40.16 Next, we would like to ask you questions about your attitudes toward artificial intelligence. Table B.34: Robotics; N = 486 manufacturing Drones that do not require a 53.48 Artificial Intelligence (AI) refers to computer systems that perform tasks or make decisions that usually require human 52.74 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies intelligence. AI can perform these tasks or make these decisions without explicit human instructions. Today, AI has been human controller Virtual assistants (e.g., Siri, 45.27 46.30 used in the following applications: Severe Google Assistant, Amazon 36.41 [Respondents were shown five items randomly selected from the list below.] 36.11 230 Catastrophic 27.47 I don't know 6.72 Skipped 0 Answer choices Virtual assistants (e.g., Siri, Google Assistant, Amazon Alexa) Smart speakers (e.g., Amazon Echo, Google Home, Apple Homepod) Facebook photo tagging Google Search Recommendations for Netflix movies or Amazon ebooks Google Translate Driverless cars and trucks Alexa) Smart speakers (e.g., • Translate over 100 different languages Table B.32: Automation; N = 513 27.32 6.44 0 Percentages (weighted) Percentages (unweighted) Raw frequencies 174 41 0 66.75 67.06 60.81 61.01 43.74 45.42 52.12 53.80 45.13 46.39 45.06 46.39 68.16 68.62 35.59 36.83 • Predict one's Google searches Amazon Echo, Google Home, • Identify people from their photos Apple Homepod) Facebook photo tagging 21.00 • Diagnose diseases like skin cancer and common illnesses 21.40 Google Search 22.07 • Predict who are at risk of various diseases 23.25 Recommendations for Netflix 17.84 • Help run factories and warehouses 18.31 • Block spam email movies or Amazon ebooks Google Translate 20.30 • Play computer games 21.19 Driverless cars and trucks 60.26 • Help conduct legal case research 61.93 Social robots that can 61.89 • Categorize photos and videos 63.17 • Detect plagiarism in essays interact with humans Industrial robots used in 67.99 • Spot abusive messages on social media 69.75 • Predict what one is likely to buy online manufacturing Drones that do not require a human controller 57.55 • Predict what movies or TV shows one is likely to watch online 59.05 QUESTION: interact with humans ANSWER CHOICES: Social robots that can 64.00 64.72 Industrial robots used in • Virtual assistants (e.g., Siri, Google Assistant, Amazon Alexa) 64.70 65.11 manufacturing • Smart speakers (e.g., Amazon Echo, Google Home, Apple Homepod) • Facebook photo tagging Drones that do not require a human controller 65.04 65.69 QUESTION: • Google Search • Recommendations for Netflix movies or Amazon ebooks What is your knowledge of computer science/technology? (Select all that apply.) • Google Translate ANSWER CHOICES: • Driverless cars and trucks • I have taken at least one college-level course in computer science. Table B.33: Machine learning; N = 508 • Social robots that can interact with humans • Industrial robots used in manufacturing Answer choices • I have a computer science or engineering undergraduate degree. Percentages (weighted) Percentages (unweighted) Raw frequencies • I have a graduate degree in computer science or engineering. • Drones that do not require a human controller Google Assistant, Amazon • None of the above Virtual assistants (e.g., Siri, 59.10 60.43 • I have programming experience. Alexa) Table B.31: Artificial intelligence (AI); N = 493 Smart speakers (e.g., 46.70 46.65 Answer choices Amazon Echo, Google Home, Apple Homepod) Percentages (weighted) Percentages (unweighted) Raw frequencies [Respondents were randomly assigned to one of the four questions. The order of answer choices was randomized, except that \"None of the above\" was always shown last.] QUESTIONS:• In your opinion, which of the following technologies, if any, uses artificial intelligence (AI)? Select all the apply. • In your opinion, which of the following technologies, if any, uses automation? Select all that apply. • In your opinion, which of the following technologies, if any, uses machine learning? Select all that apply. • In your opinion, which of the following technologies, if any, uses robotics? Select all that apply.Electronic copy available at: https://ssrn.com/abstract=<PHONE> • I don't have any of the educational or work experiences described above. \n How much do you support or oppose the development of AI? Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Strongly oppose 8.90 9.05 I don't know 9.54 9.30 Skipped 0 0 QUESTION: Please tell me to what extent you agree or disagree with the following statement. ANSWER CHOICES: • Strongly support (2) • Somewhat support (1) • Neither support nor oppose (0) • Somewhat oppose (-1) • Strongly oppose (-2) • I don't know Table B.36: Support for developing AI; N = 2000 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Strongly support 12.58 12.65 253 Somewhat support 28.36 28.65 573 Neither support nor oppose 27.84 27.60 552 Somewhat oppose 12.79 12.75 255 Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n Table B B Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Totally disagree 1.81 1.48 10 I don't know 10.46 9.60 65 Skipped 0 0 0 QUESTION: How much confidence, Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Totally agree 51.41 53.20 349 Tend to agree 30.09 28.96 190 Tend to disagree 4.79 3.81 25 Totally disagree 0.59 0.76 5 I don't know 13.12 13.26 87 Skipped 0 0 0 Table B.38: Responses to statement -AI; N = 667 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Totally agree 53.54 53.67 358 Tend to agree 30.85 30.13 201 Tend to disagree 3.67 3.90 26 Totally disagree 0.80 0.90 6 I don't know 11.14 11.39 76 Skipped 0 0 0 Table B.39: Responses to statement -Robots; N = 677 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Totally agree 51.66 52.44 355 Tend to agree 30.31 31.31 212 Tend to disagree 5.76 5.17 35 .37: Responses to statement -AI and robots; N = 656 \n if any, do you have in each of the following to develop AI in the best interests of the public? [Respondents were shown five items randomly selected from the list below. We included explainer text for actors not well known to the public; respondents could view the explainer text by hovering their mouse over the actor's name. The items and the answer choices were shown in a matrix format.] An international research organization (e.g., CERN)• Explainer text for CERN: The European Organization for Nuclear Research, known as CERN, is a European research organization that operates the largest particle physics laboratory in the world. Answer choices Answer choices Table B.41: U.S. civilian government; N = 671 Percentages (weighted) Percentages (unweighted) Raw frequencies Table B.50: Apple; N = 697 Percentages (weighted) Percentages (unweighted) Raw frequencies Answer choices I don't know Skipped Answer choices I don't know Skipped Percentages (weighted) Percentages (unweighted) Raw frequencies 21.08 20.43 0 0 14.70 14.56 97 Percentages (weighted) Percentages (unweighted) Raw frequencies 0.14 0.15 1 A great deal of confidence A great deal of confidence 5.59 10.41 5.66 10.76 A fair amount of confidence 24.04 A fair amount of confidence 26.29 24.29 26.26 Not too much confidence No confidence I don't know Skipped Not too much confidence Table B.46: Intergovernmental research organizations (e.g., CERN); N = 32.77 33.23 23.80 23.40 13.79 13.41 0 0 27.00 27.98 No confidence 22.20 21.81 645 Answer choices I don't know 13.84 12.91 Skipped 0.26 QUESTION: 0.29 Percentages (weighted) Percentages (unweighted) Raw frequencies A great deal of confidence 11.97 12.25 How much confidence, Table B.42: NSA; N = 710 28.84 22.64 Table B.51: Microsoft; N = 597 A fair amount of confidence 28.87 Not too much confidence 22.94 Answer choices No confidence I don't know Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies 16.85 16.59 19.37 19.69 Percentages (weighted) Percentages (unweighted) Raw frequencies A great deal of confidence Skipped A great deal of confidence • The U.S. military A fair amount of confidence 28.04 9.63 0 10.85 A fair amount of confidence 33.08 • The U.S. civilian government Not too much confidence 26.65 Not too much confidence 26.89 • National Security Agency (NSA) • Federal Bureau of Investigation (FBI) No confidence 22.82 I don't know 12.87 No confidence 17.99 Table B.47: Tech companies; N = 674 9.30 0 10.89 26.90 32.66 26.76 27.14 24.37 12.68 17.76 I don't know 11.05 11.39 • Central Intelligence Agency (CIA) • North Atlantic Treaty Organization (NATO) Skipped 0 0 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Skipped 0.14 0.17 A great deal of confidence 10.28 10.83 Table B.43: FBI; N = 656 34.57 27.15 Table B.52: Amazon; N = 685 A fair amount of confidence 34.15 Not too much confidence 28.40 Answer choices No confidence I don't know Answer choices • Tech companies A great deal of confidence Skipped A great deal of confidence • Google A fair amount of confidence 26.20 Percentages (weighted) Percentages (unweighted) Raw frequencies 14.91 15.13 12.15 Percentages (weighted) Percentages (unweighted) Raw frequencies 12.17 9.26 9.60 0.12 0.15 10.60 10.95 25.46 A fair amount of confidence 29.53 29.34 • Facebook Not too much confidence 25.07 25.15 Not too much confidence 25.51 25.40 • Apple • Microsoft No confidence 27.10 27.44 I don't know 12.25 12.20 No confidence 22.02 22.19 Table B.48: Google; N = 645 I don't know 12.34 12.12 • Amazon • A non-profit AI research organization (e.g., OpenAI) Skipped 0.14 0.15 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Skipped 0 0 • Explainer text for OpenAI: Open AI is an AI non-profit organization with backing from tech investors that seeks to A great deal of confidence 11.91 11.47 develop safe AI. University researchers ANSWER CHOICES: Table B.44: CIA; N = 730 A fair amount of confidence 27.35 26.82 Not too much confidence 25.92 26.67 Table B.53: Non-profit (e.g., OpenAI); N = 659 • A great deal of confidence (3) Answer choices No confidence I don't know Answer choices • A fair amount of confidence (2) A great deal of confidence Skipped A great deal of confidence • Not too much confidence (1) A fair amount of confidence 26.10 Percentages (weighted) Percentages (unweighted) Raw frequencies 21.56 21.40 13.00 Percentages (weighted) Percentages (unweighted) Raw frequencies 13.33 8.43 8.77 0.26 0.31 10.19 10.17 25.07 A fair amount of confidence 29.40 30.35 • No confidence (0) Not too much confidence 26.80 26.99 Not too much confidence 23.57 23.98 • I don't know No confidence 25.61 26.30 I don't know 12.93 12.74 No confidence 13.65 13.66 Table B.49: Facebook; N = 632 I don't know 23.04 21.70 Table B.40: U.S. military; N = 638 Skipped 0.13 0.14 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Skipped 0.13 0.15 Answer choices A great deal of confidence Percentages (weighted) Percentages (unweighted) Raw frequencies 4.29 3.96 A great deal of confidence A fair amount of confidence 32.19 17.16 A fair amount of confidence 14.35 Not too much confidence 26.40 Table B.54: University researchers; N = 666 17.08 30.88 Table B.45: NATO; N = 695 13.45 27.22 Not too much confidence 23.92 24.14 No confidence 14.40 14.89 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies 109 197 154 No confidence 41.27 42.88 I don't know 13.44 12.18 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies 95 I don't know 12.33 13.01 A great deal of confidence 4.40 4.17 Skipped 0.25 0.32 A great deal of confidence 13.86 14.11 83 Skipped 0 0 A fair amount of confidence 25.41 24.75 A fair amount of confidence 36.29 36.04 0 Not too much confidence 25.98 26.62 Not too much confidence 22.27 22.82 No confidence No confidence 23.13 12.75 24.03 12.31 • Explainer text for NATO: NATO is a military alliance that includes 28 countries including most of Europe, as well as the U.S. and Canada. •Electronic copy available at: https://ssrn.com/abstract=3312874Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n if any, do you have in each of the following to manage the development and use of AI in the best interests of the public? [Respondents were shown five items randomly selected from the list below. We included explainer text for actors not well known to the public; respondents could view the explainer text by hovering their mouse over the actor's name. The items and the answer choices were shown in a matrix format.] • U.S. federal government • U.S. state governments • International organizations (e.g., United Nations, European Union) • The United Nations (UN) • An intergovernmental research organization (e.g., CERN) • Explainer text for CERN: The European Organization for Nuclear Research, known as CERN, is a European research organization that operates the largest particle physics laboratory in the world. • Tech companies • Google • Facebook • Apple • Microsoft • Amazon • Non-government scientific organizations (e.g., AAAI) • Explainer text for AAAI: Association for the Advancement of Artificial Intelligence (AAAI) is a non-government scientific organization that promotes research in, and responsible use of AI. • Partnership on AI, an association of tech companies, academics, and civil society groups \n Table B . B 55: U.S. federal government; N = 743 Answer choices Table B.56: U.S. state governments; N = 713 Percentages (weighted) Percentages (unweighted) Raw frequencies Table B.65: Amazon; N = 784 Answer choices A great deal of confidence A fair amount of confidence 20.39 Percentages (weighted) Percentages (unweighted) Raw frequencies 6.25 6.45 19.21 No confidence 19.88 20.45 I don't know 12.81 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies 11.74 Skipped 0.41 A great deal of confidence 10.19 10.33 81 0.26 A fair amount of confidence 25.22 24.87 195 Not too much confidence Not too much confidence 31.57 25.20 32.12 25.38 199 No confidence I don't know No confidence I don't know 29.65 11.69 24.53 14.87 30.72 11.22 24.87 Table B.61: Google; N = 767 14.54 195 114 Skipped Answer choices Skipped 0.45 Percentages (weighted) Percentages (unweighted) Raw frequencies 0.28 0 0 0 A great deal of confidence 9.61 9.13 Table B.57: International organizations; N = 827 A fair amount of confidence 23.60 23.86 Not too much confidence 27.44 27.77 Table B.66: Non-government scientific organization (e.g., AAAI); N = 792 Answer choices A great deal of confidence A fair amount of confidence 22.48 Percentages (weighted) Percentages (unweighted) Raw frequencies 5.94 5.80 21.77 No confidence 25.13 25.03 I don't know 13.75 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies 13.95 Skipped 0.47 A great deal of confidence 7.64 7.83 62 0.26 A fair amount of confidence 30.32 30.05 238 Not too much confidence Not too much confidence 29.58 25.37 29.87 26.39 209 No confidence I don't know No confidence I don't know 26.81 14.81 15.03 Table B.62: Facebook; N = 741 27.45 14.87 14.65 21.46 20.83 116 165 Skipped Answer choices Skipped 0.38 Percentages (weighted) Percentages (unweighted) Raw frequencies 0.24 0.19 0.25 2 A great deal of confidence 4.99 4.45 Table B.58: UN; N = 802 16.19 28.21 Table B.67: Partnership on AI; N = 780 A fair amount of confidence 16.18 Not too much confidence 28.50 Answer choices No confidence I don't know Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies 36.95 38.46 13.14 12.42 Percentages (weighted) Percentages (unweighted) Raw frequencies A great deal of confidence Skipped A great deal of confidence 6.23 0.24 8.89 6.61 0.27 9.23 72 A fair amount of confidence 22.49 A fair amount of confidence 30.12 21.57 29.49 230 Not too much confidence Not too much confidence 26.14 25.89 26.18 26.79 209 No confidence I don't know No confidence I don't know 31.90 12.64 16.33 18.64 31.55 13.59 15.77 Table B.63: Apple; N = 775 18.59 123 145 Skipped Answer choices Skipped 0.60 Percentages (weighted) Percentages (unweighted) Raw frequencies 0.50 0.12 0.13 1 A great deal of confidence 8.25 8.39 A fair amount of confidence 25.10 Table B.59: Intergovernmental research organizations (e.g., CERN); N = 24.90 Not too much confidence 29.08 28.65 747 No confidence 23.91 24.52 Answer choices I don't know A great deal of confidence Skipped 13.55 Percentages (weighted) Percentages (unweighted) Raw frequencies 13.42 6.69 7.10 0.12 0.13 A fair amount of confidence 30.51 29.72 Answer choices Not too much confidence No confidence Percentages (weighted) Percentages (unweighted) Raw frequencies 23.89 24.10 20.32 20.21 Table B.64: Microsoft; N = 771 A great deal of confidence A fair amount of confidence 20.26 6.86 I don't know 18.36 Skipped 0.22 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies 6.59 49 20.19 150 18.61 0.27 Not too much confidence A great deal of confidence 28.44 7.79 28.67 7.78 213 No confidence A fair amount of confidence 30.11 31.50 32.44 29.83 241 I don't know Skipped Not too much confidence No confidence 12.68 0.25 22.98 Table B.60: Tech companies; N = 758 11.84 0.27 23.48 24.10 24.38 88 2 Answer choices I don't know A great deal of confidence Skipped 14.68 Percentages (weighted) Percentages (unweighted) Raw frequencies 14.14 8.33 8.44 0.35 0.39 A fair amount of confidence 33.50 32.98 Not too much confidence 25.07 26.12 Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n • [Disease diagnosis] Accuracy and transparency in AI used for disease diagnosis: Increasingly, AI software has been used to diagnose diseases, such as heart disease and cancer. One challenge is to make sure the AI can correctly diagnose those who have the disease and not mistakenly diagnose those who do not have the disease. Another challenge is that AI used in this application may lack transparency such that human users do not understand what the algorithm is doing, or why it reaches certain decisions in specific cases.• [ \n Data privacy] Protect data privacy: Algorithms used in AI applications are often trained on vast amounts of personal data, including medical records, social media content, and financial transactions. Some worry that data used to train algorithms are not collected, used, and stored in ways that protect personal privacy.• [ \n Autonomous vehicles] Make sure autonomous vehicles are safe: Companies are developing self-driving cars and trucks that require little or no input from humans. Some worry about the safety of autonomous vehicles for those riding in them as well as for other vehicles, cyclists, and pedestrians.• [ \n Ditigal manipulation] Prevent AI from being used to spread fake and harmful content online: AI has been used by governments, private groups, and individuals to harm or manipulate internet users. For instance, automated bots have been used to generate and spread false and/or harmful news stories, audios, and videos.• [ \n Cyber attacks] Prevent AI cyber attacks against governments, companies, organizations, and individuals: Computer scientists have shown that AI can be used to launch effective cyber attacks. AI could be used to hack into servers to steal sensitive information, shut down critical infrastructures like power grids or hospital networks, or scale up targeted phishing attacks.• [ \n Surveillance] Prevent AI-assisted surveillance from violating privacy and civil liberties: AI can be used to process and analyze large amounts of text, photo, audio, and video data from social media, mobile communications, and CCTV cameras. Some worry that governments, companies, and employers could use AI to increase their surveillance capabilities.• [ \n U.S.-China arms race] Prevent escalation of a U.S.-China AI arms race: Leading analysts believe that an AI arms race is beginning, in which the U.S. and China are investing billions of dollars to develop powerful AI systems for surveillance, autonomous weapons, cyber operations, propaganda, and command and control systems. Some worry that a U.S.-China arms race could lead to extreme dangers. To stay ahead, the U.S. and China may race to deploy advanced military AI systems that they do not fully understand or can control. We could see catastrophic accidents, such as a rapid, automated escalation involving cyber and nuclear weapons.• [ \n Value alignment] Make sure AI systems are safe, trustworthy, and aligned with human values: As AI systems \n the next 10 years, how likely do you think it is that this AI governance challenge will impact large numbers of people around the world? • Likely: 80-95% chance (87.5%) • Very likely: more than 95% chance (97.5%) • I don't know QUESTION: In ANSWER CHOICES: • Very unlikely: less than 5% chance (2.5%) • Unlikely: 5-20% chance (12.5%) • Somewhat unlikely: 20-40% chance (30%) • Equally likely as unlikely: 40-60% chance (50%) • Somewhat likely: 60-80% chance (70%) • Likely: 80-95% chance (87.5%) • Very likely: more than 95% chance (97.5%) • I don't know QUESTION: In \n the next 10 years, how important is it for tech companies and governments to carefully manage the following challenge? ANSWER CHOICES: • Very important (3) • Somewhat important (2) • Not too important (1) • Not at all important (0) • I don't know Answer choices Very unlikely < 5% Unlikely 5-20% Somewhat unlikely 20-40% Table B.68: Likelihood in the US -Hiring bias; N = 760 Percentages (weighted) Percentages (unweighted) Raw frequencies 2.57 2.63 6.07 6.18 10.86 10.92 Equally likely as unlikely 40-60% 22.27 22.50 Somewhat likely 60-80% 23.34 22.89 Likely 80-95% 12.39 12.76 Very likely > 95% 9.86 9.61 I don't know 12.35 12.37 Skipped 0.29 0.13 Table B.69: Likelihood in the US -Criminal justice bias; N = 778 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 4.94 4.50 Unlikely 5-20% 8.76 8.61 Somewhat unlikely 20-40% 13.25 12.85 Equally likely as unlikely 40-60% 21.23 21.08 Somewhat likely 60-80% 17.13 17.22 Likely 80-95% 12.28 12.60 Very likely > 95% 9.05 9.64 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies I don't know 12.90 12.98 Skipped 0.45 0.51 Table B.70: Likelihood in the US -Disease diagnosis; N = 767 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.79 2.61 Unlikely 5-20% 4.73 4.95 Somewhat unlikely 20-40% 10.18 9.52 Equally likely as unlikely 40-60% 23.12 23.21 Somewhat likely 60-80% 20.50 19.95 Likely 80-95% 13.43 13.95 Very likely > 95% 9.72 10.17 I don't know 13.62 13.69 Skipped 1.91 1.96 Table B.71: Likelihood in the US -Data privacy; N = 807 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.75 2.11 Unlikely 5-20% 4.53 4.58 Somewhat unlikely 20-40% 7.52 7.19 Equally likely as unlikely 40-60% 16.10 15.86 Somewhat likely 60-80% 18.81 19.33 Likely 80-95% 17.00 16.36 Very likely > 95% 20.59 21.69 I don't know 10.87 10.78 Skipped 1.84 2.11 Table B.72: Likelihood in the US -Autonomous vehicles; N = 796 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 3.65 3.64 Unlikely 5-20% 5.80 5.90 Somewhat unlikely 20-40% 10.93 10.43 Equally likely as unlikely 40-60% 16.17 16.33 Somewhat likely 60-80% 23.62 23.62 Likely 80-95% 15.78 15.45 Very likely > 95% 12.29 12.94 I don't know 10.89 10.68 Skipped 0.87 1.01 Table B.73: Likelihood in the US -Digital manipulation; N = 741 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.79 2.83 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Unlikely 5-20% 3.24 3.10 Somewhat unlikely 20-40% 8.12 7.69 Equally likely as unlikely 40-60% 13.81 14.30 Somewhat likely 60-80% 16.58 16.33 Likely 80-95% 17.74 18.08 Very likely > 95% 23.45 23.62 I don't know 12.49 12.15 Skipped 1.77 1.89 Table B.74: Likelihood in the US -Cyber attacks; N = 745 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 3.36 2.42 Unlikely 5-20% 4.28 3.89 Somewhat unlikely 20-40% 8.44 8.59 Equally likely as unlikely 40-60% 15.45 15.84 Somewhat likely 60-80% 19.22 19.46 Likely 80-95% 15.96 15.30 Very likely > 95% 20.52 21.21 I don't know 9.70 10.47 Skipped 3.07 2.82 Table B.75: Likelihood in the US -Surveillance; N = 784 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.70 2.42 Unlikely 5-20% 2.92 2.81 Somewhat unlikely 20-40% 6.19 6.38 Equally likely as unlikely 40-60% 15.23 15.05 Somewhat likely 60-80% 18.95 18.75 Likely 80-95% 16.03 15.69 Very likely > 95% 23.52 24.23 I don't know 12.15 12.12 Skipped 2.32 2.55 Table B.76: Likelihood in the US -U.S.-China arms race; N = 766 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 3.24 3.26 Unlikely 5-20% 5.98 6.01 Somewhat unlikely 20-40% 10.01 10.84 Equally likely as unlikely 40-60% 18.74 18.41 Somewhat likely 60-80% 20.08 19.71 Likely 80-95% 13.17 12.79 Very likely > 95% 10.62 11.36 I don't know 15.17 14.62 Skipped 3.00 3.00 Table B.77: Likelihood in the US -Value alignment; N = 783 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 3.78 4.21 Unlikely 5-20% 7.30 6.90 Somewhat unlikely 20-40% 9.01 9.07 Equally likely as unlikely 40-60% 20.34 19.54 Somewhat likely 60-80% 19.26 19.28 Likely 80-95% 13.66 13.79 Very likely > 95% 12.96 13.67 I don't know 12.43 12.26 Skipped 1.26 1.28 Table B.78: Likelihood in the US -Autonomous weapons; N = 757 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 6.22 5.94 Unlikely 5-20% 10.36 9.38 Somewhat unlikely 20-40% 12.75 12.68 Equally likely as unlikely 40-60% 18.91 19.02 Somewhat likely 60-80% 15.72 15.72 Likely 80-95% 11.44 11.76 Very likely > 95% 10.72 11.23 I don't know 11.99 12.29 Skipped 1.89 1.98 Table B.79: Likelihood in the US -Technological unemployment; N = 738 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 3.08 2.98 Unlikely 5-20% 5.80 5.69 Somewhat unlikely 20-40% 11.00 11.11 Equally likely as unlikely 40-60% 17.74 17.62 Somewhat likely 60-80% 17.16 17.75 Likely 80-95% 14.86 14.91 Very likely > 95% 15.75 15.99 I don't know 12.84 12.20 Skipped 1.75 1.76 Table B.80: Likelihood in the US -Critical AI systems failure; N = 778 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 6.98 6.43 Unlikely 5-20% 7.94 7.58 Somewhat unlikely 20-40% 12.26 12.98 Equally likely as unlikely 40-60% 20.36 20.31 Somewhat likely 60-80% 15.59 15.42 Likely 80-95% 12.25 11.83 Very likely > 95% 9.36 10.15 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies I don't know 14.85 14.78 Skipped 0.41 0.51 Table B.81: Likelihood around the world -Hiring bias; N = 760 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.95 3.03 Unlikely 5-20% 5.47 5.00 Somewhat unlikely 20-40% 8.54 8.55 Equally likely as unlikely 40-60% 20.23 21.45 Somewhat likely 60-80% 21.55 21.32 Likely 80-95% 13.68 13.55 Very likely > 95% 12.20 12.11 I don't know 15.04 14.61 Skipped 0.35 0.39 Table B.82: Likelihood around the world -Criminal justice bias; N = 778 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 4.44 4.24 Unlikely 5-20% 8.06 7.71 Somewhat unlikely 20-40% 10.96 10.80 Equally likely as unlikely 40-60% 19.17 19.41 Somewhat likely 60-80% 18.29 18.25 Likely 80-95% 13.09 13.62 Very likely > 95% 9.38 9.90 I don't know 16.38 15.94 Skipped 0.23 0.13 Table B.83: Likelihood around the world -Disease diagnosis; N = 767 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.31 2.35 Unlikely 5-20% 4.18 4.17 Somewhat unlikely 20-40% 9.93 9.13 Equally likely as unlikely 40-60% 21.28 20.99 Somewhat likely 60-80% 20.47 20.47 Likely 80-95% 15.00 15.38 Very likely > 95% 10.94 11.47 I don't know 15.80 15.91 Skipped 0.09 0.13 Table B.84: Likelihood around the world -Data privacy; N = 807 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.86 2.23 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Unlikely 5-20% 2.92 2.60 Somewhat unlikely 20-40% 8.32 8.30 Equally likely as unlikely 40-60% 13.79 14.75 Somewhat likely 60-80% 19.07 18.84 Likely 80-95% 18.43 18.22 Very likely > 95% 21.09 21.81 I don't know 13.34 13.01 Skipped 0.19 0.25 Table B.85: Likelihood around the world -Autonomous vehicles; N = 796 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 3.77 3.52 Unlikely 5-20% 5.25 5.65 Somewhat unlikely 20-40% 12.37 11.68 Equally likely as unlikely 40-60% 16.74 17.21 Somewhat likely 60-80% 21.09 21.11 Likely 80-95% 14.13 14.45 Very likely > 95% 12.04 12.19 I don't know 13.99 13.57 Skipped 0.63 0.63 Table B.86: Likelihood around the world -Digital manipulation; N = 741 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 1.98 2.16 Unlikely 5-20% 1.67 1.48 Somewhat unlikely 20-40% 7.34 7.29 Equally likely as unlikely 40-60% 12.68 12.96 Somewhat likely 60-80% 17.18 17.00 Likely 80-95% 21.22 21.73 Very likely > 95% 22.31 22.00 I don't know 15.24 14.98 Skipped 0.39 0.40 Table B.87: Likelihood around the world -Cyber attacks; N = 745 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 1.08 1.21 Unlikely 5-20% 4.95 4.03 Somewhat unlikely 20-40% 4.76 5.10 Equally likely as unlikely 40-60% 16.95 16.64 Somewhat likely 60-80% 18.94 19.73 Likely 80-95% 19.13 19.06 Very likely > 95% 20.57 20.40 I don't know 13.20 13.42 Skipped 0.42 0.40 Table B.88: Likelihood around the world -Surveillance; N = 784 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 1.26 1.40 Unlikely 5-20% 3.55 3.19 Somewhat unlikely 20-40% 5.12 5.36 Equally likely as unlikely 40-60% 14.26 14.41 Somewhat likely 60-80% 18.90 19.13 Likely 80-95% 20.30 19.77 Very likely > 95% 22.62 22.70 I don't know 13.93 13.90 Skipped 0.07 0.13 Table B.89: Likelihood around the world -U.S.-China arms race; N = 766 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 3.21 3.13 Unlikely 5-20% 4.61 4.83 Somewhat unlikely 20-40% 7.70 7.83 Equally likely as unlikely 40-60% 19.50 19.19 Somewhat likely 60-80% 20.71 20.76 Likely 80-95% 14.99 14.75 Very likely > 95% 12.46 12.92 I don't know 16.61 16.32 Skipped 0.22 0.26 Table B.90: Likelihood around the world -Value alignment; N = 783 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.70 2.94 Unlikely 5-20% 4.66 4.60 Somewhat unlikely 20-40% 8.80 8.81 Equally likely as unlikely 40-60% 19.92 19.41 Somewhat likely 60-80% 18.97 18.77 Likely 80-95% 15.57 15.33 Very likely > 95% 14.93 15.71 I don't know 14.44 14.43 Skipped 0 0 Table B.91: Likelihood around the world -Autonomous weapons; N = 757 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 3.72 3.70 Unlikely 5-20% 7.04 5.42 Somewhat unlikely 20-40% 9.42 9.64 Equally likely as unlikely 40-60% 17.23 17.44 Somewhat likely 60-80% 16.08 15.85 Likely 80-95% 16.35 17.04 Very likely > 95% 14.87 15.19 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies I don't know 15.20 15.59 Skipped 0.09 0.13 Table B.92: Likelihood around the world -Technological unemployment; N = 738 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.76 2.57 Unlikely 5-20% 4.92 4.47 Somewhat unlikely 20-40% 8.31 8.81 Equally likely as unlikely 40-60% 18.36 18.16 Somewhat likely 60-80% 19.90 21.00 Likely 80-95% 14.78 14.50 Very likely > 95% 16.71 16.67 I don't know 13.77 13.41 Skipped 0.51 0.41 Table B.93: Likelihood around the world -Critical AI systems failure; N = 778 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 5.36 5.27 Unlikely 5-20% 8.07 7.97 Somewhat unlikely 20-40% 10.75 10.41 Equally likely as unlikely 40-60% 18.03 17.87 Somewhat likely 60-80% 16.71 16.84 Likely 80-95% 13.09 13.11 Very likely > 95% 11.23 11.83 I don't know 16.76 16.71 Skipped 0 0 Table B.94: Issue importance -Hiring bias; N = 760 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 56.86 57.11 434 Somewhat important 22.11 22.76 173 Not too important 6.56 6.05 46 Not at all important 1.50 1.58 12 I don't know 12.98 12.50 95 Skipped 0 0 0 Table B.95: Issue importance -Criminal justice bias; N = 778 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 56.08 56.68 441 Somewhat important 21.78 Percentages (weighted) Percentages (unweighted) Raw frequencies Not too important 6.65 5.91 Not at all important 1.83 1.67 I don't know 13.66 13.24 Skipped 0 0 Table B.96: Issue importance -Disease diagnosis; N = 767 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 55.60 56.98 Somewhat important 22.37 21.25 I don't know 13.26 12.91 Skipped 0.11 0.13 Table B.97: Issue importance -Data privacy; N = 807 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 63.65 64.93 Somewhat important 17.65 17.10 Not too important 4.76 4.71 Not at all important 1.71 1.36 I don't know 12.05 11.65 Skipped 0.19 0.25 Table B.98: Issue importance -Autonomous vehicles; N = 796 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 58.70 59.55 Somewhat important 22.36 21.73 Not too important 6.13 6.28 Not at all important 1.44 1.63 I don't know 11.15 10.55 Skipped 0.22 0.25 Table B.99: Issue importance -Digital manipulation; N = 741 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 57.66 58.30 Somewhat important 18.75 18.08 Not too important 6.25 6.48 Not at all important 3.11 2.97 I don't know 14.16 14.04 Skipped 0.08 0.13 Table B.100: Issue importance -Cyber attacks; N = 745 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 62.12 61.21 Somewhat important 17.80 18.39 Not too important 7.07 7.38 Not at all important 1.14 1.07 I don't know 11.88 11.95 Skipped 0 0 Table B.101: Issue importance -Surveillance; N = 784 Very important 58.54 58.80 Somewhat important 19.33 19.26 Not too important 6.40 6.63 Not at all important 1.73 1.66 I don't know 13.93 13.52 Skipped 0.07 0.13 Table B.102: Issue importance -U.S.-China arms race; N = 766 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 55.88 55.74 Somewhat important 19.44 19.71 Not too important 7.07 7.57 Not at all important 2.38 2.35 I don't know 15.13 14.49 Skipped 0.10 0.13 Table B.103: Issue importance -Value alignment; N = 783 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 56.46 56.45 Somewhat important 20.49 20.95 Not too important 6.69 6.64 Not at all important 1.56 1.66 I don't know 14.80 14.30 Skipped 0 0 Table B.104: Issue importance -Autonomous weapons; N = 757 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 58.32 57.73 Somewhat important 20.00 19.55 Not too important 5.52 5.94 Not at all important 1.23 1.45 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies I don't know 14.94 15.32 116 Skipped 0 0 0 Table B.105: Issue importance -Technological unemployment; N = 738 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 54.12 54.34 401 Somewhat important 22.07 22.49 166 Not too important 6.50 6.91 51 Not at all important 2.83 2.44 18 Table B.106: Issue importance -Critical AI systems failure; N = 778 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very important 52.63 53.86 419 Somewhat important 21.10 20.44 159 Not too important 7.98 8.10 63 Not at all important 2.93 2.44 19 I don't know 15.36 15.17 118 Skipped 0 0 0 [Respondents were presented with one randomly-selected question from the two below.] QUESTIONS: • Compared with other industrialized countries, how would you rate the U.S. in AI research and development? • Compared with other industrialized countries, how would you rate China in AI research and development? ANSWER CHOICES: • Best in the world (3) • Above average (2) • Average (1) • Below average (0) • I don't know Table B.107: Perceptions of research and development -U.S.; N = 988 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Best in the world 9.73 10.02 99 Above average 36.16 37.55 371 Average 26.09 24.70 244 Below average 4.99 4.96 49 I don't know 23.03 22.77 225 Skipped 0 0 0 Table B.108: Perceptions of research and development -China; N = 1012 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Best in the world 7.33 7.41 75 Above average 45.40 46.64 472 Average 16.66 15.81 160 Below average 3.93 3.66 37 I don't know 26.68 26.48 268 Skipped 0 0 0 22.49 Answer choices Not too important 6.68 6.91 Not at all important 1.98 1.83 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies I don't know 14.39 13.69 101 Skipped 0.09 0.14 1 [All respondents were presented with the following prompt.] Electronic copy available at: https://ssrn.com/abstract=3312874Electronic copy available at: https://ssrn.com/abstract=3312874Electronic copy available at: https://ssrn.com/abstract=3312874Electronic copy available at: https://ssrn.com/abstract=3312874Electronic copy available at: https://ssrn.com/abstract=3312874Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n Table B . B 110: Responses to statement that U.S. should invest more in AI military capabilities -Treatment 1: Pro-nationalist; N = 505 Table B.114: Responses to statement that U.S. should work hard to co- operate with China to avoid dangers of AI arms race -Treatment 1: Pro- nationalist; N = 505 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Strongly agree Answer choices 20.88 Percentages (weighted) Percentages (unweighted) Raw frequencies 20.40 Somewhat agree Strongly agree 26.89 18.51 27.52 18.81 Neither agree nor disagree 21.79 Somewhat agree 27.35 22.18 28.12 Somewhat disagree Neither agree nor disagree 20.08 11.69 12.28 20.99 Strongly disagree Somewhat disagree 5.30 10.09 5.35 9.90 I don't know Strongly disagree 13.45 8.45 12.28 7.92 Skipped I don't know 0 15.51 0 14.26 Skipped 0 0 Table B.111: Responses to statement that U.S. should invest more in AI military capabilities -Treatment 2: Risks of arms race; N = 493 Table B.115: Responses to statement that U.S. should work hard to coop- erate with China to avoid dangers of AI arms race -Treatment 2: Risks of Answer choices arms race; N = 493 Percentages (weighted) Percentages (unweighted) Raw frequencies Strongly agree Somewhat agree Answer choices 18.26 27.85 Percentages (weighted) Percentages (unweighted) Raw frequencies 19.07 27.38 Neither agree nor disagree 21.69 Strongly agree 24.97 20.28 25.96 Somewhat disagree Somewhat agree 12.87 25.32 13.79 25.15 Strongly disagree Neither agree nor disagree 21.53 6.88 6.90 20.49 I don't know Somewhat disagree 12.45 9.83 12.58 9.94 Skipped Strongly disagree 0 5.84 0 5.68 I don't know 12.51 12.78 Skipped 0 0 Table B.112: Responses to statement that U.S. should invest more in AI military capabilities -Treatment 3: One common humanity; N = 492 Table B.116: Responses to statement that U.S. should work hard to co- Answer choices operate with China to avoid dangers of AI arms race -Treatment 3: One Percentages (weighted) Percentages (unweighted) Raw frequencies Strongly agree 22.38 20.53 common humanity; N = 492 Somewhat agree Neither agree nor disagree 24.37 27.29 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies 27.85 23.98 Somewhat disagree Strongly agree 6.73 23.63 7.11 24.19 Strongly disagree Somewhat agree 6.17 27.52 6.91 28.46 I don't know Neither agree nor disagree 21.31 13.07 13.62 20.33 Skipped Somewhat disagree 0 8.50 0 7.32 Strongly disagree 6.72 6.91 I don't know 12.31 12.80 Skipped Table B.113: Responses to statement that U.S. should work hard to coop-0 0 erate with China to avoid dangers of AI arms race -Control; N = 510 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Strongly agree Somewhat agree QUESTION: 22.34 26.16 22.55 26.27 For Neither agree nor disagree 22.02 20.59 Somewhat disagree 8.29 9.02 Strongly disagree 7.38 7.45 I don't know 13.59 13.92 Skipped 0.21 0.20 \n the following issues, how likely is it that the U.S. and China can cooperate? \n much do you support or oppose the development of high-level machine intelligence? Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Skipped 0.15 0.15 3 Table B.128: Forecasting high-level machine intelligence -50 years; N = 2000 Answer choices Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% 2.28 2.30 46 Unlikely 5-20% 1.66 1.55 31 Somewhat unlikely 20-40% 2.75 2.75 55 Equally likely as unlikely 40-60% 10.08 9.90 198 Somewhat likely 60-80% 12.33 12.20 244 Likely 80-95% 14.43 14.50 290 Very likely > 95% 40.86 41.15 823 I don't know 15.52 15.55 311 Skipped 0.09 0.10 2 • Unlikely: 5-20% chance (12.5%) • Somewhat unlikely: 20-40% chance (30%) • Equally likely as unlikely: 40-60% chance (50%) • Somewhat likely: 60-80% chance (70%) • Likely: 80-95% chance (87.5%) QUESTION: • Very likely: more than 95% chance (97.5%) • I don't know Table B.126: Forecasting high-level machine intelligence -10 years; N = How ANSWER CHOICES: • Strongly support 2000 • Somewhat support Answer choices • Neither support nor oppose Percentages (weighted) Percentages (unweighted) Raw frequencies Very unlikely < 5% • Somewhat oppose Unlikely 5-20% • Strongly oppose Somewhat unlikely 20-40% • I don't know 4.46 8.19 14.84 4.50 8.20 14.75 Equally likely as unlikely 40-60% 20.34 Table B.129: Support for developing high-level machine intelligence; N = 19.95 Somewhat likely 60-80% 2000 21.08 21.25 Likely 80-95% 10.69 10.65 Very likely > 95% Answer choices 7.40 Percentages (weighted) Percentages (unweighted) Raw frequencies 7.85 I don't know Skipped Strongly support Somewhat support 12.91 0.09 7.78 23.58 12.75 8.10 0.10 23.30 162 466 Neither support nor oppose 29.40 28.75 575 Somewhat oppose Table B.127: Forecasting high-level machine intelligence -20 years; N = 16.19 16.60 Strongly oppose 11.02 11.10 2000 I don't know 11.94 12.05 332 222 241 Answer choices Skipped Percentages (weighted) Percentages (unweighted) Raw frequencies 0.09 0.10 2 Very unlikely < 5% 1.52 1.45 Unlikely 5-20% 2.73 2.95 Somewhat unlikely 20-40% 6.26 5.85 Equally likely as unlikely 40-60% 16.83 Somewhat likely 60-80% 18.17 QUESTION: 16.40 18.65 Likely 80-95% 22.25 22.25 Very likely > 95% 17.91 18.30 I don't know 14.18 14.00 Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n We present the percentage of \"don't know\" or missing responses to the survey question (see Appendix B for the survey question text). Regression analysis shows that the varying the term used (i.e., AI, AI and robots, and robots) does not change responses to the statement that such technologies should be carefully managed. This finding is robust to a regression where we controlled for \"don't know\" or missing responses. In TableC.6, we present the distribution of responses to the statement by country. ANSWER CHOICES: • Extremely good Countries Latvia Table C.3: Survey experiment attrition check: agreement with statement that AI and/or robots should be carefully managed Table C.8: Respondents distinguish between AI, automation, machine Totally disagree Tend to disagree Tend to agree Totally agree Don't know 1 3 29 63 learning, and robotics • On balance good • More or less neutral Lithuania Luxembourg Technological applications Variables 0 1 Coefficients (SEs) 4 4 35 33 F-statistic 57 58 p-value Significant • On balance bad • Extremely bad, possibly human extinction (Intercept) Malta 2 Virtual assistants (e.g., Siri, Google Assistant, Amazon Alexa) 0.11 (0.01)*** 4 Table C.1 shows the regression results used to produce Figure 2.4. AI and robots 0.02 (0.02) Netherlands 1 2 Smart speakers (e.g., Amazon Echo, Google Home, Apple Homepod) F(3, 1996) = 24.76 <0.001 Yes 46 38 10 F(3, 1996) = 18.12 <0.001 Yes 22 74 • I don't know Poland Facebook photo tagging Table B.130: Expected outcome of high-level machine intelligence; N = 2000 Robots -0.01 (0.02) 2 8 44 42 F(3, 1996) = 20.22 <0.001 Yes Table C.1: Predicting support for developing AI using demographic charac-N = 2000 F(2, 1997) = 1.03; p-value: 0.359 Portugal 2 2 37 48 11 Google Search F(3, 1996) = 37.30 <0.001 Yes teristics: results from a multiple linear regression that includes all demo-Romania 5 12 33 42 Recommendations for Netflix movies or Amazon ebooks F(3, 1996) = 33.69 <0.001 Yes graphic variables; outcome standardized to have mean 0 and unit variance Slovakia 0 5 44 46 Google Translate F(3, 1996) = 24.62 <0.001 Yes Answer choices Extremely good On balance good More or less neutral On balance bad Extremely bad, possibly human extinction 11.66 Percentages (weighted) Percentages (unweighted) Raw frequencies 5.35 5.45 109 21.28 21.25 425 21.00 21.10 422 22.38 23.10 11.55 Don't know 18.25 17.45 Skipped 0.09 0.10 College+ Employed (full-or part-time) 0.03 (0.05) Table C.7 summarizes responses to 15 potential global risks. 0.18 (0.06)** 2 Some college -0.01 (0.06) N = 2000 F(2, 1997) = 1.92; p-value: 0.146 349 Non-white -0.02 (0.05) Robots -0.09 (0.05) 231 Male 0.17 (0.05)*** AI and robots -0.03 (0.04) 462 Variables Coefficients (SEs) (Intercept) -0.27 (0.09)** Age 38-53 Slovenia 2 6 37 52 Driverless cars and trucks F(3, 1996) = 9.08 <0.001 Yes Table C.4: Survey experiment results: agreement with statement that AI Spain 1 3 40 47 Social robots that can interact with humans F(3, 1996) = 1.05 0.369 No and/or robots should be carefully managed Sweden 1 2 18 Industrial robots used in manufacturing F(3, 1996) = 55.72 <0.001 Yes 75 -0.16 (0.06)** Age 54-72 -0.18 (0.06)** Age 73 and older -0.16 (0.10) Variables (Intercept) 1.49 (0.03)*** United States 1 5 30 52 12 Coefficients (SEs) United Kingdom 1 3 34 57 Drones that do not require a human controller F(3, 1996) = 9.68 <0.001 Yes Democrat Independent/Other Income $30-70K Table C.5: Survey experiment results: agreement with statement that AI 0.20 (0.06)** -0.05 (0.06) 0.01 (0.06) Table C.7: Summary statistics: the American public's perceptions of 15 and/or robots should be carefully managed (controlling for DK/missing responses) potential global risks Income $70-100K Income more than $100K Prefer not to say income No religious affiliation Other religion Born-again Christian CS or engineering degree Variables (Intercept) Failure to address climate change Potential risks Failure of regional/global governance AI and robots 0.03 (0.05) 0.13 (0.09) 0.16 (0.08)* -0.14 (0.07) 0.16 (0.05)** 0.14 (0.08) -0.04 (0.06) 0.05 (0.09) Mean perceived likelihood Mean perceived impact Coefficients (SEs) 56% 2.25 1.46 (0.03)*** 55% 2.46 Robots Conflict between major countries 60% 2.68 -0.07 (0.05) N = 2000 Weapons of mass destruction 49% 3.04 F(5, 1994) = 0.91; p-value: 0.471 Large-scale involuntary migration 57% 2.65 N CS or programming experience 0.30 (0.06)*** Spread of infectious diseases 50% 2.69 N = 2000 Table C.6: Distribution of responses to statement that AI and robots should F(19,1980) = 11.75; p-value: <0.001 54% 2.90 52% 2.76 be carefully managed by country (in percentages); EU countries data from Harmful consequences of AI Water crises Food crises 45% 2.29 Eurobarometer Harmful consequences of synthetic biology 45% 2.33 1073 Cyber attacks 68% 2.85 Countries Terrorist attacks Totally disagree Tend to disagree Tend to agree Totally agree Don't know 60% 2.62 Austria Global recession Belgium Extreme weather events Bulgaria Natural disasters 3 1 1 56% 65% 69% 7 9 2 43 40 24 2.61 2.73 2.87 43 48 65 Croatia 4 8 37 47 Cyprus 1 2 26 67 Czech Republic 2 7 37 50 Denmark Table C.2: Survey experiment attrition check: agreement with statement 1 4 25 66 Estonia that AI and/or robots should be carefully managed 0 4 39 51 Experimental condition Percent DK/missing Percent DK Percent missing AI 11.39 11.39 0 AI and robots 13.26 13.26 Robots 9.60 9.60 4 12 35 45 0 1 3 23 71 0 European Union Hungary Greece 2 5 35 53 Finland 1 4 29 We formally tested whether or not respondents think AI, automation, machine learning, and robotics are used in different 63 France 1 3 31 applications. (See Appendix B for the survey question text.) For each technological application, we used an F -test to test 62 Germany 2 4 32 59 whether any of terms randomly assigned to the respondents affect respondents' selecting that application. Because we Ireland 1 4 37 54 Italy 3 8 43 40 Electronic copy available at: https://ssrn.com/abstract=<PHONE> ran 10 F -tests, we used the Bonferroni correction to control the familywise error rate. The Bonferroni correction rejected the null hypothesis at alpha level α/10, instead of α. For instance, to test whether the F -static is significant at the 5% level, we set the alpha level at α/10 = 0.005. Our results (in TableC.8) show that except for social robots, respondents think that AI, automation, machine learning, and robotics are used in each of the applications presented in the survey. \n Table C C .9: Correlation between survey completion time and number of se-lected items Coefficients (SEs) 3.58 (0.17)*** Survey completion time (min) Variables (Intercept) 0.14 (0.01)*** Absolute deviation from median survey completion time (min) -0.14 (0.01)*** Term: automation 0.98 (0.22)*** Term: machine learning -0.09 (0.22) Term: Robotics -0.51 (0.20)* N = 2000 F(5, 1994) = 47.47; p-value: <0.001 Table C.10: Correlation between survey completion time and not selecting 'none of the above' Variables Coefficients (SEs) (Intercept) Coefficients (SEs) Survey completion time (min) 0.01 (<0.01)*** Absolute deviation from median survey completion time (min) -0.01 (<0.01)*** Term: automation 0.05 (0.02)* Term: machine learning -0.04 (0.02) Term: Robotics 0.04 (0.02) N = 2000 F(5, 1994) = 13.16; p-value: <0.001 Table C.11: Correlation between survey completion time and selecting 'robots' when assigned the term 'robotics' Variables Coefficients (SEs) (Intercept) 0.87 (0.06)*** Survey completion time (min) 0.06 (0.01)*** Absolute deviation from median survey completion time (min) -0.06 (0.01)*** 0.79 (0.02)**Variables N = 486 F(2, 483) = 50.55; p-value: <0.001 *Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n Table C . C 12: Comparing perceived likelihood: in U.S. vs. around the world; each difference is the U.S. mean likelihood subtracted from the world mean likelihood Governance challenge U.S. mean likelihood Difference (SE) p-value Significant Hiring bias 59.8 2.5 (0.8) 0.001 Yes Criminal justice bias 55.6 2.5 (0.8) 0.003 Yes Disease diagnosis 60.4 2.1 (0.6) 0.001 Yes Data privacy 66.9 1.7 (0.6) 0.010 No Autonomous vehicles 61.8 -0.7 (0.8) 0.401 No Digital manipulation 68.6 2.6 (0.7) <0.001 Yes Cyber attacks 66.2 3.2 (0.9) <0.001 Yes Surveillance 69.0 2.2 (0.7) 0.002 Yes U.S.-China arms race 60.3 3.0 (0.7) <0.001 Yes Value alignment 60.4 3.6 (0.7) <0.001 Yes Autonomous weapons 54.7 7.6 (0.8) <0.001 Yes Technological unemployment 62.3 2.3 (0.7) <0.001 Yes Critical AI systems failure 55.2 3.1 (0.8) <0.001 Yes \n Table C C Variables Coefficient (SEs) Age 38-53 0.11 (0.07) Age 54-72 0.35 (0.06)*** Age 73 and older 0.44 (0.07)*** Male 0.02 (0.05) Non-white Some college .13: Perception of AI governance challenges in the U.S.: summary -0.01 (0.05) 0.03 (0.07) statistics table College+ 0.15 (0.07)* Governance challenge Employed (full-or part-time) Mean likelihood Mean issue importance Product of likelihood and issue importance -0.09 (0.06) Surveillance 69% 2.56 Income $30-70K 0.09 (0.08) 1.77 Data privacy 67% 2.62 Income $70-100K 0.13 (0.10) 1.75 Digital manipulation 69% 2.53 Income more than $100K -0.01 (0.10) 1.74 Cyber attacks 66% 2.59 Prefer not to say income 0.04 (0.08) 1.71 Autonomous vehicles 62% 2.56 Democrat 0.13 (0.07) 1.58 Technological unemployment 62% 2.50 Independent/Other 0.14 (0.07) 1.56 Value alignment 60% 2.55 No religious affiliation -0.04 (0.06) 1.54 Disease diagnosis 60% 2.52 Other religion -0.05 (0.08) 1.52 U.S.-China arms race 60% 2.52 Born-again Christian 0.07 (0.07) 1.52 Hiring bias 60% 2.54 CS or engineering degree -0.35 (0.10)*** 1.52 Autonomous weapons 55% 2.58 CS or programming experience -0.01 (0.07) 1.42 Criminal justice bias 56% 2.53 Criminal justice bias 0.05 (0.13) 1.41 Critical AI systems failure 55% 2.47 Disease diagnosis -0.06 (0.14) 1.36 Data privacy 0.16 (0.13) Autonomous vehicles -0.07 (0.14) Digital manipulation -0.14 (0.15) Table C.14: Perception of AI governance challenges in the world: summary Cyber attacks 0.05 (0.14) statistics table Surveillance <0.01 (0.15) U.S.-China arms race 0.04 (0.13) Governance challenge Value alignment Mean likelihood Mean issue importance Product of likelihood and issue importance -0.06 (0.13) Surveillance Digital manipulation Autonomous weapons 71% 71% Technological unemployment Cyber attacks 69% Critical AI systems failure Data privacy 69% N = 10000 observations, 2000 respondents F(259,1999) = 3.36; p-value: <0.001 2.56 0.06 (0.14) 1.82 2.53 -0.12 (0.14) 1.80 2.59 -0.27 (0.15) 1.80 2.62 1.80 Value alignment 64% 2.55 1.63 Technological unemployment 65% 2.50 1.62 Autonomous weapons 62% 2.58 1.61 U.S.-China arms race 63% 2.52 1.60 Hiring bias 62% 2.54 1.58 Disease diagnosis 63% 2.52 1.58 Autonomous vehicles 61% 2.56 1.56 Criminal justice bias 58% 2.53 1.47 Critical AI systems failure 58% 2.47 1.44 Table C.15: Results from a saturated regression predicting perceived issue importance using demographic variables, AI governance challenge, and in- teractions between the two types of variables; the coefficients for the inter- actions variables are not shown due to space constraints Variables Coefficient (SEs) (Intercept) 2.25 (0.11)*** \n Table C . C 16 displays the mean level of trust the public expresses in various actors to develop and manage AI in the interest of the public. Center for the Governance of AI Figure C.1: AI governance challenges: issue importance by demographic subgroupsA substantial percentage of respondents selected \"I don't know\" when answering this survey question. (See Appendix B for the survey question text.) Our regression analysis shows that there is a small but statistically significant difference between respondents' perception of R&D in the U.S. as compared to in China, as seen in Tables C.19 and C.20.     Age 18−37 Age 38−53 Age 54−72 Age 73 and older Female White Table C.17: Survey experiment attrition check: comparing U.S. and China's Male AI research and development Non−white Experimental condition Percent DK/missing Percent DK Percent missing HS or less China Some college U.S. 26.48 22.77 26.48 22.77 0 0 College+ Demographic subgroups Income $70−100K Income $30−70K Income less than $30K Independent/Other Democrat Republican Employed (full− or part−time) Not employed Table C.18: Survey experiment attrition check: comparing U.S. and China's AI research and development Variables Coefficients (SEs) (Intercept) 0.27 (0.01)*** U.S. -0.04 (0.02) N = 2000 F(1, 1998) = 3.12; p-value: 0.078 Income more than $100K Prefer not to say income Table C.19: Survey experiment results: comparing U.S. and China's AI re- Christian search and development No religious affiliation Other religion Variables Coefficients (SEs) Not born−again Christian (Intercept) 1.74 (0.02)*** No CS or engineering degree Born−again Christian U.S. N = 2000 -0.08 (0.03)* F(1, 1998) = 6.58; p-value: 0.01 CS or engineering degree No CS or programming experience Table C.20: Survey experiment results: comparing U.S. and China's AI re- CS or programming experience search and development (controlling for DK/missing responses) Critical AI systems failure Variables (Intercept) 1.74 (0.02)*** Technological unemployment Disease diagnosis U.S.−China arms race Digital manipulation Coefficients (SEs) U.S. -0.08 (0.03)** N = 2000 F(3, 1996) = 6.14; p-value: <0.001 Criminal justice bias Hiring bias Value alignment Autonomous vehicles Surveillance Autonomous weapons Cyber attacks Data privacy AI governance challenges Mean−centered issue importance on a 4−point scale (Smaller value = less important; Greater value = more important) −0.6 −0.4 −0.2 0.0 0.2 Electronic copy available at: https://ssrn.com/abstract=3312874Source: \n Table C . C 16: Trust in various actors to develop and manage AI in the interest of the public: mean responses We checked that \"don't know\" or missing responses to both statements are not induced by the information treatments. (See Appendix B for the survey experiment text.) Next, we examined the correlation between responses to the two statements using a 2D bin count graph. The overall Pearson correlation coefficient is -0.05 but there exists considerable variation by experimental condition. Actors Trust to develop AI Trust to manage AI U.S. military Table C.21: Survey experiment attrition check: agreement with statement 1.56 (MOE: +/-0.07); N = 638 U.S. civilian government 1.16 (MOE: +/-0.07); N = that U.S. should invest more in AI military capabilities 671 1.28 (MOE: +/-0.07); N = Percent DK/missing Percent DK Percent missing 13.53 13.53 0 710 12.28 12.28 0 1.21 (MOE: +/-0.08); N = Treatment 2: Risks of arms race NSA Experimental condition Control FBI Treatment 1: Pro-nationalist 12.58 12.58 0 656 CIA 1.21 (MOE: +/-0.07); N = Treatment 3: One common humanity 13.62 13.62 0 730 U.S. federal government Table C.22: Survey experiment attrition check: agreement with statement 1.05 (MOE: +/-0.07); N = 743 U.S. state governments 1.05 (MOE: +/-0.07); N = that U.S. should invest more in AI military capabilities NATO Variables 713 Coefficients (SEs) 1.17 (MOE: +/-0.06); N = (Intercept) 695 0.13 (0.02)*** Intergovernmental research Treatment 1: Pro-nationalist 1.42 (MOE: +/-0.07); N = <0.01 (0.02) 1.27 (MOE: +/-0.06); N = organizations (e.g., CERN) Treatment 2: Risks of arms race 645 -0.01 (0.02) 747 International organizations Treatment 3: One common humanity >-0.01 (0.02) 1.10 (MOE: +/-0.06); N = N = 2000 827 F(3, 1996) = 0.08; p-value: 0.972 UN 1.06 (MOE: +/-0.06); N = 802 Tech companies Table C.23: Survey experiment attrition check: agreement with statement 1.44 (MOE: +/-0.07); N = 1.33 (MOE: +/-0.07); N = 674 that U.S. should work hard to cooperate with China to avoid dangers of AI 758 Google arms race 1.34 (MOE: +/-0.08); N = 1.20 (MOE: +/-0.07); N = 645 767 Facebook Experimental condition 0.85 (MOE: +/-0.07); N = Percent DK/missing Percent DK Percent missing 0.91 (MOE: +/-0.07); N = 632 1.29 (MOE: +/-0.07); N = 697 14.12 14.26 Treatment 2: Risks of arms race Apple Control Treatment 1: Pro-nationalist 12.78 741 1.20 (MOE: +/-0.07); N = 13.92 0.2 14.26 0.0 775 12.78 0.0 Microsoft Treatment 3: One common humanity 1.40 (MOE: +/-0.08); N = 12.80 1.24 (MOE: +/-0.07); N = 12.80 0.0 597 771 Amazon 1.33 (MOE: +/-0.07); N = 1.24 (MOE: +/-0.07); N = 685 Table C.24: Survey experiment attrition check: agreement with statement 784 Non-profit (e.g., OpenAI) that U.S. should work hard to cooperate with China to avoid dangers of AI 1.44 (MOE: +/-0.07); N = arms race 659 University researchers 1.56 (MOE: +/-0.07); N = Non-government scientific Variables 666 1.35 (MOE: +/-0.06); N = Coefficients (SEs) organization (e.g., AAAI) 792 Partnership on AI 1.35 (MOE: +/-0.06); N = 780 \n Table C . C 25: Correlation between responses to the two statementsThere are many \"don't know\" responses to this survey question (see Appendix B for the survey question text). Nevertheless, \"don't know\" or missing responses are not affected by the experimental future time framing. F -tests reveal that there are no differences in responses to the three future time frames, as seen in TableC.30.Table C.26: Survey experiment attrition check: future time frame Experimental condition Percent DK/missing Percent DK Percent missing Experimental condition Pearson correlation Overall -0.05 Control -0.06 Treatment 1: Pro-nationalist -0.03 Treatment 2: Risks of arms race -0.12 Treatment 3: One common humanity -0.01 No time frame 24.59 24.38 0.21 10 years 25.49 25.49 0.00 20 years 26.16 25.96 0.20 50 years 24.17 24.17 0.00 Table C.27: Survey experiment attrition check: future time frame Variables Coefficients (SEs) (Intercept) 0.25 (0.02)*** 10 years 0.01 (0.03) 20 years 0.02 (0.03) 50 years -0.01 (0.03) N = 2000 F(3, 1996) = 0.34; p-value: 0.795 Table C.28: Survey experiment results: future time frame Variables Coefficients (SEs) (Intercept) -0.52 (0.06)*** 10 years -0.15 (0.08) 20 years -0.12 (0.08) 50 years -0.06 (0.08) N = 2000 F(3, 1996) = 1.48; p-value: 0.219 Table C.29: Survey experiment results: future time frame (controlling for DK/missing responses) Variables Coefficients (SEs) (Intercept) -0.52 (0.06)*** 10 years -0.15 (0.08) \n Table C . C 32: Predicting support for developing high-level machine intelligence using demographic characteristics: results from a multiple linear regression that includes all demographic variables; outcome standardized to have mean 0 and unit variance Variables Coefficients (SEs) (Intercept) -0.25 (0.09)** Age 38-53 -0.12 (0.06) Age 54-72 -0.03 (0.06) Age 73 and older 0.12 (0.10) Male 0.18 (0.05)*** Non-white 0.01 (0.05) Some college -0.04 (0.06) College+ <0.01 (0.07) Employed (full-or part-time) 0.09 (0.05) Democrat 0.11 (0.07) Independent/Other -0.13 (0.07)* Income $30-70K -0.01 (0.07) Income $70-100K 0.09 (0.09) Income more than $100K 0.19 (0.09)* Prefer not to say income <0.01 (0.08) No religious affiliation 0.09 (0.06) Other religion 0.06 (0.08) Born-again Christian -0.07 (0.06) CS or engineering degree <0.01 (0.10) CS or programming experience 0.36 (0.06)*** N = 2000 F(19,1980) = 7.27; p-value: <0.001 \n Table C . C 33: Predicting support for developing high-level machine intelligence using demographic characteristics: results from a multiple linear regression that includes all demographic variables and respondents' support for developing AI; outcome standardized to have mean 0 and unit variance Variables Coefficients (SEs) (Intercept) -0.23 (0.08)** Age 38-53 -0.02 (0.05) Age 54-72 0.09 (0.05) Age 73 and older 0.22 (0.09)* Male 0.08 (0.04) Non-white 0.02 (0.05) Some college -0.04 (0.05) College+ -0.11 (0.06) Employed (full-or part-time) 0.08 (0.04) Democrat -0.02 (0.06) Independent/Other -0.10 (0.05) Income $30-70K -0.01 (0.06) Income $70-100K 0.01 (0.07) Income more than $100K 0.08 (0.07) Prefer not to say income 0.09 (0.07) No religious affiliation -0.02 (0.05) Other religion -0.03 (0.07) Born-again Christian -0.05 (0.05) \n 1 9.2 Pearson's r = 0.69 Support for developing high−level machine intelligence Strongly oppose Somewhat oppose Neither support nor oppose Somewhat support Strongly support DK bad Extremely bad On balance less netural More or good On balance good Extremely DK Expected outcome of high−level machine intelligence Percentage of respondents 2.0% 4.0% 6.0% 8.0% 10.0% 12.0% Source: Center for the Governance of AI \n\t\t\t Our survey asked separately about trust in 1) building and 2) managing the development and use of AI. Results are similar and are combined here.Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t These percentages that we discuss here reflect the average response across the three statements. See Appendix B for the topline result for each statement. \n\t\t\t Our definition of global risk borrowed from the Global Challenges Foundation's definition: \"an uncertain event or condition that, if it happens, can cause a significant negative impact on at least 10% of the world's population within the next 10 years\" (Cotton-Barratt et al. 2016) .Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t The World Economic Forum's survey asked experts to evaluate the \"adverse consequences of technological advances,\" defined as \"[i]ntended or unintended adverse consequences of technological advances such as artificial intelligence, geo-engineering and synthetic biology causing human, environmental and economic damage.\" The experts considered these \"adverse consequences of technological advances\" to be less likely and lowerimpact, compared with other potential risks. \n\t\t\t In TableC.15, we report the results of a saturated linear model using demographic variables, governance challenges, and the interaction between \n\t\t\t Note that our survey asked respondents this question with the time frames 10, 20 and 50 years, whereas the NSF surveys provided no time frame.Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t Note that our definition of high-level machine intelligence is equivalent to what many would consider human-level machine intelligence. Details of the question are found in Appendix B.14  In Grace et al. (2018) , each respondent provides three data points for their forecast, and these are fitted to the Gamma CDF by least squares to produce the individual cumulative distribution function (CDFs). Each \"aggregate forecast\" is the mean distribution over all individual CDFs (also called the \"mixture\" distribution). The confidence interval is generated by bootstrapping (clustering on respondents) and plotting the 95% interval for estimated probabilities at each year. Survey weights are not used in this analysis due to problems incorporating survey weights into the bootstrap. \n\t\t\t The discrepancy between this figure and the percentages in Figure6.2 is due to rounding. According to TableB.129, 7.78% strongly support and 23.58% somewhat support; therefore, 31.36% -rounding to 31% -of respondents either support or somewhat support.Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t For this and other questions that ask respondents about likelihoods, each multiple-choice answer was coded to the mean value across the probabilities in the answer's range. \n\t\t\t Note that the perceived issue importance was measured on a four-point scale, where 0 meant \"not at all important\" and 3 meant \"very important.\"We only mean-centered the outcomes; we did not standardize such that the outcomes have unit variance.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/SSRN-id3312874.tei.xml", "id": "e0090a17428fc68c78d38ee8a1beb319"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Artificial general intelligence (AGI) is AI that can reason across a wide range of domains. While most AI research and development (R&D) is on narrow AI, not AGI, there is some dedicated AGI R&D. If AGI is built, its impacts could be profound. Depending on how it is designed and used, it could either help solve the world's problems or cause catastrophe, possibly even human extinction. This paper presents the first-ever survey of active AGI R&D projects for ethics, risk, and policy. The survey attempts to identify every active AGI R&D project and characterize them in terms of seven attributes:  The type of institution the project is based in  Whether the project publishes open-source code  Whether the project has military connections  The nation(s) that the project is based in  The project's goals for its AGI  The extent of the project's engagement with AGI safety issues  The overall size of the project To accomplish this, the survey uses openly published information as found in scholarly publications, project websites, popular media articles, and other websites, including 11 technical survey papers, 8 years of the Journal of Artificial General Intelligence, 7 years of AGI conference proceedings, 2 online lists of AGI projects, keyword searches in Google web search and Google Scholar, the author's prior knowledge, and additional literature and webpages identified via all of the above.", "authors": ["Seth Baum"], "title": "A Survey of Artificial General Intelligence Projects for Ethics, Risk, and Policy", "text": "The survey identifies 45 AGI R&D projects spread across 30 countries in 6 continents, many of which are based in major corporations and academic institutions, and some of which are large and heavily funded. Many of the projects are interconnected via common personnel, common parent organizations, or project collaboration. For each of the seven attributes, some major trends about AGI R&D projects are apparent:  Most projects are in corporations or academic institutions.  Most projects publish open-source code.  Few projects have military connections.  Most projects are based in the US, and almost all are in either the US or a US ally. The only projects that exist entirely outside US and its allies are in China or Russia, and these projects all have strong academic and/or Western ties.  Most projects state goals oriented towards the benefit of humanity as a whole or towards advancing the frontiers of knowledge, which the paper refers to as \"humanitarian\" and \"intellectualist\" goals.  Most projects are not active on AGI safety issues.  Most projects are in the small-to-medium size range. The three largest projects are DeepMind (a London-based project of Google), the Human Brain Project (an academic project based in Lausanne, Switzerland), and OpenAI (a nonprofit based in San Fransisco). Looking across multiple attributes, some additional trends are apparent:  There is a cluster of academic projects that state goals of advancing knowledge (i.e., intellectualist) and are not active on safety.  There is a cluster of corporate projects that state goals of benefiting humanity (i.e., humanitarian) and are active on safety.  Most of the projects with military connections are US academic groups that receive military funding, including a sub-cluster within the academic-intellectualist-not active on safety cluster.  All six China-based projects are small, though some are at large organizations with the resources to scale quickly. Figure  ES1  on the next page presents an overview of the data. The data suggest the following conclusions: Regarding ethics, the major trend is projects' split between stated goals of benefiting humanity and advancing knowledge, with the former coming largely from corporate projects and the latter from academic projects. While these are not the only goals that projects articulate, there appears to be a loose consensus for some combination of these goals. Regarding risk, in particular the risk of AGI catastrophe, there is good news and bad news. The bad news is that most projects are not actively addressing AGI safety issues. Academic projects are especially absent on safety. Another area of concern is the potential for corporate projects to put profit ahead of safety and the public interest. The good news is that there is a lot of potential to get projects to cooperate on safety issues, thanks to the partial consensus on goals, the concentration of projects in the US and its allies, and the various interconnections between different projects. Regarding policy, several conclusions can be made. First, the concentration of projects in the US and its allies could greatly facilitate the establishment of public policy for AGI. Second, the large number of academic projects suggests an important role for research policy, such as review boards to evaluate risky research. Third, the large number of corporate projects suggests a need for attention to the political economy of AGI R&D. For example, if AGI R&D brings companies near-term projects, then policy could be much more difficult. Finally, the large number of projects with open-source code presents another policy challenge by enabling AGI R&D to be done by anyone anywhere in the world. This study has some limitations, meaning that the actual state of AGI R&D may differ from what is presented here. This is due to the fact that the survey is based exclusively on openly published information. It is possible that some AGI R&D projects were missed by this survey. Thus, the number of projects identified in this survey, 45, is a lower bound. Furthermore, it is possible that projects' actual attributes differ from those found in openly published information. For example, most corporate projects did not state the goal of profit, even though many presumably seek profit. Therefore, this study's results should not be assumed to necessarily reflect the actual current state of AGI R&D. That said, the study nonetheless provides the most thorough description yet of AGI R&D in terms of ethics, risk, and policy. \n Introduction Artificial general intelligence (AGI) is AI that can reason across a wide range of domains. The human mind has general intelligence, but most AI does not. Thus, for example, DeepBlue can beat Kasparov at chess, and maybe at a few other basic tasks like multiplication, but it cannot beat him at anything else. AGI was a primary goal of the initial AI field and has long been considered its \"grand dream\" or \"holy grail\".  1  The technical difficulty of building AGI has led most of the field to focus on narrower, more immediately practical forms of AI, but some dedicated AGI research and development (R&D) continues. AGI is also a profound societal concern, or at least it could be if it is built. AGI could complement human intellect, offering heightened capacity to solve the world's problems. Or, it could be used maliciously in a power play by whoever controls it. Or, humanity may fail to control it. Any AI can outsmart humans in some domains (e.g., multiplication); an AGI may outsmart humans in all domains. In that case, the outcome could depend on the AGI's goals: whether it pursues something positive for humanity, or the world, or itself, or something else entirely. Indeed, scholars of AGI sometimes propose that it could cause human extinction or similar catastrophe (see literature review below). The high potential stakes of AGI raise questions of ethics, risk, and policy. Which AGI, if any, should be built? What is the risk of catastrophe if an AGI is built? What policy options are available to avoid AGI catastrophe and, to the extent that it is desired, enable safe and beneficial AGI? These are all questions under active investigation. However, the literature to date has tended to be theoretical and speculative, with little basis in the actual state of affairs in AGI. Given that AGI may be first built many years from now, some speculation is inevitable. But AGI R&D is happening right now. Information about the current R&D can guide current activities on ethics, risk, and policy, and it can provide some insight into what future R&D might look like. This paper presents the first-ever survey of active AGI R&D projects in terms of ethics, risk, and policy. There have been several prior surveys of R&D on AGI and related technologies  (Chong et al. 2007; Duch et al. 2008; Langley et al. 2008; de Garis et al. 2010; Goertzel et al. 2010; Samsonovich 2010; Taatgen and Anderson 2010; Thórisson and Helgasson 2012; Dong and Franklin 2014; Goertzel 2014; Kotseruba et al. 2016 ). However, these surveys are all on technical aspects, such as how the AGI itself is designed, what progress has been made on it, and how promising the various approaches are for achieving AGI. This paper presents and analyzes information of relevance to ethics, risk, and policy-for example, which political jurisdictions projects are located in and how engaged they are on AGI safety. Additionally, whereas the prior surveys focus on select noteworthy examples of AGI R&D projects, this paper attempts to document the entire set of projects. For ethics, risk, and policy, it is useful to have the full set-for example, to know which political jurisdictions to include in AGI public policy. Section 1.1 explains terminology related to AGI. Section 2 reviews prior literature on AGI ethics, risk, and policy. Section 3 presents the research questions pursued in this study. Section 4 summarizes the methodology used for this paper's survey. Section 5 presents the main survey results about AGI R&D projects and other notable projects. Section 6 concludes. Appendix 1 presents the full survey of active AGI R&D projects. Appendix 2 presents notable projects that were considered for the survey but excluded for not meeting inclusion criteria. \n Terminology AGI is one of several terms used for advanced and potentially transformative future AI. The terms have slightly different meanings and it is worth briefly distinguishing between them.  AGI is specifically AI with a wide range of intelligence capabilities, including \"the ability to achieve a variety of goals, and carry out a variety of tasks, in a variety of different contexts and environments\"  (Goertzel 2014, p.2) . AGI is not necessarily advanced-an AI can be general without being highly sophisticated-though general intelligence requires a certain degree of sophistication and AGI is often presumed to be highly capable. AGI is also a dedicated field of study, with its own society (http://www.agi-society.org), journal (Journal of Artificial General Intelligence), and conference series (http://agi-conf.org).  Cognitive architecture is the overall structure of an intelligent entity. One can speak of the cognitive architecture of the brains of humans or other animals, but the term is used mainly for theoretical and computational models of human and nonhuman animal cognition. Cognitive architectures can be narrow, focusing on specific cognitive processes such as attention or emotion  (Kotseruba et al. 2016 ). However, they are often general, and thus their study overlaps with the study of AGI. Biologically inspired cognitive architectures is a dedicated field of study with its own society (http://bicasociety.org), journal (Biologically Inspired Cognitive Architectures), and conference series (http://bicasociety.org/meetings).  Brain emulations are computational instantiations of biological brains. Brain emulations are sometimes classified as distinct from AGI (e.g.,  Barrett and Baum 2017a ), but they are computational entities with general intelligence, and thus this paper treats them as a type of AGI.  Human-level AI is AI with intelligence comparable to humans, or \"human-level, reasonably human-like AGI\"  (Goertzel 2014, p.6 ). An important subtlety is that an AGI could be as advanced as humans, but with a rather different type of intelligence: it does not necessarily mimic human cognition. For example, AI chess programs will use brute force searches in some instances in which humans use intuition, yet the AI can still perform at or beyond the human level.  Superintelligence is AI that significantly exceeds human intelligence. The term ultraintelligence has also been used in this context  (Good 1965)  but is less common. It is often proposed that superintelligence will come from an initial seed AI that undergoes recursive-self improvement, becoming successively smarter and smarter. The seed AI would not necessarily be AGI, but it is often presumed to be. This paper focuses on AGI because the term is used heavily in R&D contexts and it is important for ethics, risk, and policy. Narrow cognitive architectures (and narrow AI) are less likely to have transformative consequences for the world. Human-level AI and superintelligence are more likely to have transformative consequences, but these terms are not common in R&D. Note that not every project included in this paper's survey explicitly identifies as AGI, but they all have the potential to be AGI and likewise to have transformative consequences. The survey does not exclude any projects that are explicitly trying to build human-level AI or superintelligence. \n Prior Literature \n Ethics Perhaps the most basic question in AGI ethics is on whether to treat AGI as an intellectual pursuit or as something that could impact society and the world at large. In other words, is AGI R&D pursued to advance the forefront of knowledge or to benefit society? Often, these two goals are closely connected, as evidenced by the central role of science and technology in improving living conditions worldwide. However, they are not always connected and can sometimes be at odds. In particular for the present study, research into potentially dangerous new technologies can yield significant scientific and intellectual insights, yet end up being harmful to society. Researchers across all fields of research often have strong intellectual values and motivations, and AGI is no exception. The question of whether to evaluate research in terms of intellectual merit or broader societal/global impacts is an ongoing point of contention across academia  (Schienke et al. 2009) . As with most fields, AI has traditionally emphasized intellectual merits, though there are calls for this to change  (Baum 2017a) . The intellectual pull of AGI can be particularly strong, given its status as a long-term \"grand dream\" or \"holy grail\", though the broader impacts can also have a strong pull, given the large potential stakes of AGI. In practical terms, an AGI project with intellectual motivations is more likely to view building AGI as a worthy goal in itself and to pay little attention to any potential dangers or other broader impacts, relative to a project motivated by broader impacts. A second area of AGI ethics concerns what goals the AGI should be designed to pursue. This is the main focus of prior literature on AGI ethics. One line of thinking proposes \"indirect normativity\" or \"coherent extrapolated volition\", in which the AGI is designed to use its intellect to figure out what humanity wants it to do  (Yudkowsky 2004; Muehlhauser and Helm 2012; Bostrom 2014 ). This proposal is motivated in part by concerns of procedural justice-everyone should have a say in the AGI's ethics, not just the AGI designers-and in part by concerns about the technical difficulty of programming the subtleties of human ethics directly into the AGI. Regardless, these proposals all call for the AGI to follow the ethics of humanity, and not of anything else.  2  An alternative line of thinking proposes that the AGI should create new entities that are morally superior to humans. This thinking falls in the realm of \"transhumanism\" or \"posthumanism\"; AGI researchers de Garis (2005) and Goertzel (2010) use the term \"cosmism\". This view holds that AGI should benefit the cosmos as a whole, not just humanity, and proposes that the good of the cosmos may be advanced by morally superior beings produced by AGI. Whereas  Goertzel (2010)  stresses that \"legacy humans\" should be able to decide for themselves whether to continue in this new world, de  Garis (2005)  suggests that this world may be worth forming even if legacy humans would be eliminated. In contrast,  Yampolskiy (2013)  argues that AGI should only be built if they are expendable tools of benefit to their human creators. Finally, there has also been some discussion of whether AGI should be built in the first place, though to date this has been a smaller focus of attention. Most discussions of AGI either support building it or do not seriously consider the matter because they presume its inevitability, as discussed by  Totschnig (2017) . Some arguments against building AGI are rooted in concern about catastrophe risk (e.g.,  Joy 2000) ; more on risk below. Others argue that even safe AGI should not be built. These include the fringe anarchist views of  Kaczynski (1995, para. 174 ) and the more sober discussion of  Totschnig (2017) , though there has been much less outright opposition to AGI than there has been to similar transformative technologies like human enhancement. \n Risk The potential for AGI catastrophe is rooted in the notion that AGI could come to outsmart humanity, take control of the planet, and pursue whatever goals it is programmed with. Unless it is programmed with goals that are safe for humanity, or for whatever else it is that one cares about, catastrophe will result. Likewise, in order to avoid catastrophe, AGI R&D projects must take sufficient safety precautions. Opinions vary on the size of this risk and the corresponding safety effort required. Some propose that it is fundamentally difficult to design an AGI with safe goals-that even seemingly minor mistakes could yield catastrophic results, and therefore AGI R&D projects should be very attentive to safety  (Yudkowsky 2004; Muehlhauser and Helm 2012; Bostrom 2014) . Others argue that an AGI can be trained to have safe goals, and that this process is not exceptionally fragile, such that AGI R&D projects need to attend to safety, but not to an unusual extent (e.g.,  Goertzel and Pitt 2012; Bieger et al. 2015; Goertzel 2015; Steunebrink et al. 2016) . Finally, some dismiss the risk entirely, either because AGI will not be able to outsmart humanity (e.g.,  Bringsjord 2012; McDermott 2012)  or because it is too unlikely or too distant in the future to merit attention (e.g.,  Etzioni 2016; Stilgoe and Maynard 2017) . One common concern is that competing projects will race to launch AGI first, with potentially catastrophic consequences  (Joy 2000; Shulman 2009; Dewey 2015; Armstrong et al. 2016) . Desire to win the AGI race may be especially strong due to perceptions that AGI could be so powerful that it would lock in an extreme first-mover advantage. This creates a collective action problem: it is in the group's interest for each project to maintain a high safety standard, but it is each project's individual interest to skimp on safety in order to win the race.  Armstrong et al. (2016)  present game theoretic analysis of the AGI race scenario, finding that the risk increases if (a) there are more R&D projects, (b) the projects have stronger preference for their own AGI relative to others', making them less likely to invest in time-consuming safety measures, and (c) the projects have similar capability to build AGI, bringing them more relative advantage when they skimp on safety.  Barrett and Baum (2017a; ) develop a risk model of catastrophe from AGI, looking specifically at AGI that recursively self-improves until it becomes superintelligent and gains control of the planet.  3  For this catastrophe to occur, six conditions must all hold: (1) superintelligence must be possible; (2) the initial (\"seed\") AI that starts self-improvement must be created; (3) there must be no successful containment of the self-improvement process or the resulting superintelligence, so that the superintelligence gains control of the planet; (4) humans must fail to make the AI's goals safe, such that accomplishment of the goals would avoid catastrophe; (5) the AI must not make its goals safe on its own, independent of human efforts to make its goals safe; and (6) the AI must not be deterred from pursuing its goals by humans, other AIs, or anything else. The total risk depends on the probability of each of these conditions holding. Likewise, risk management can seek to reduce the probability of conditions (2), (3), (4), and (6). Risk management is one aspect of AGI policy. \n Policy AGI policy can be understood broadly as all efforts to influence AGI R&D, which can include formal policies of governments and other institutions as well as informal policies of people interested in or concerned about AGI, including the researchers themselves. AGI policy can seek to, among other things, fund or otherwise support AGI R&D, encourage certain ethical views to be built into AGI, or reduce AGI risk.  Sotala and Yampolskiy (2015)  review a wide range of AGI policy ideas, focusing on risk management. Much of the prior literature on AGI politics emphasizes the tension between (1) hypothetical AGI developers who want to proceed with inadequate regard for safety or ethics and (2) a community that is concerned about unsafe and unethical AGI and seeks ways to shift AGI R&D in safer and more ethical directions.  Joy (2000)  argues that the risk of catastrophe is too great and calls for a general abandonment of AGI R&D.  Hibbard (2002)  and  Hughes (2007)  instead call for regulatory regimes to avoid dangerous AGI without completely abandoning the technology.  Yampolskiy and Fox (2013)  propose review boards at research institutions to restrict AGI research that would be too dangerous.  Baum (2017c)  calls for attention to the social psychology of AGI R&D communities in order to ensure that safety and ethics measures succeed and to encourage AGI R&D communities to do more on their own. One policy challenge comes from the fact that AGI could be developed anywhere in the world that can attract the research talent and assemble modest computing resources. Therefore,  Wilson (2013)  outlines an international treaty that could ensure that dangerous AGI work does not shift to unregulated countries.  Scherer (2016)  analyzes the potential for AGI regulation by the US government, noting the advantages of national regulation relative to sub-national regulation and suggesting that this could be a prelude to an international treaty.  Goertzel (2009)  analyzes prospects for AGI to be developed in China vs. in the West, finding that it could go either way depending on the relative importance of certain factors.  Bostrom (2014)  calls for international control over AGI R&D, possibly under the auspices of the United Nations. In order to identify rogue AGI R&D projects that may operate in secret,  Hughes (2007) ,  Shulman (2009), and Dewey (2015)  propose global surveillance regimes;  Goertzel (2012a)  proposes that a limited AGI could conduct the surveillance. Finally, prior literature has occasionally touched on the institutional context in which AGI R&D occurs. The  Yampolskiy and Fox (2013)  proposal for review boards focuses on universities, similar to the existing review boards for university human subjects research.  Goertzel (2017a)  expresses concern about AGI R&D at large corporations due to their tendency to concentrate global wealth and bias government policy in their favor; he argues instead for open-source AGI R&D. In contrast,  Bostrom (2017)  argues that open-source AGI R&D could be more dangerous by giving everyone access to the same code and thereby tightening the race to build AGI first.  Shulman (2009)  worries that nations will compete to build AGI in order to achieve \"unchallenged economic and military dominance\", and that the pursuit of AGI could be geopolitically destabilizing.  Baum et al. (2011)  query AGI experts on the relative merits of AGI R&D in corporations, open-source communities, and the US military, finding divergent views across experts, especially on open-source vs. military R&D. It should also be noted that there has been some significant activity on AI from major governments. For example, the Chinese government recently announced a major initiative to become a global leader in AI within the next few decades  (Webster et al. 2017) . The Chinese initiative closely resembles-and may be derivative of-a series of reports on AI published by the US under President Obama (Kania 2017). Russian President Vladimir Putin recently spoke about the importance of AI, calling it \"the future\", noting \"colossal opportunities, but also threats that are difficult to predict\", and stating that \"whoever becomes the leader in this sphere will become the ruler of the world\" (RT 2017). But these various initiatives and pronouncements are not specifically about AGI, and appear to mainly refer to narrow AI. Some policy communities have even avoided associating with AGI, such as a series of events sponsored by the Obama White House in association with the above-mentioned reports  (Conn 2016) . Thus, high-level government interest in AI does not necessarily imply government involvement in AGI. One instance of high-level government interest in AGI is in the European Commission's largescale support of the Human Brain Project, in hopes that a computer brain simulation could revive the European economy  (Theil 2015) . \n Research Questions The prior literature suggests several questions that could be informed by a survey of active AGI R&D projects: How many AGI R&D projects are there?  Armstrong et al. (2016)  find that AGI risk increases if there are more R&D projects, making them less likely to cooperate on safety. Similarly, literature on collective action in other contexts often proposes that, under some circumstances, smaller groups can be more successful at cooperating, though large groups can be more successful in other circumstances (e.g.,  Yang et al. 2013) .  4  Thus, it is worth simply knowing how many AGI R&D projects there are. What types of institutions are the projects based in?  Shulman (2009) ,  Baum et al. (2011) ,  Yampolskiy and Fox (2013), and Goertzel (2017a)  suggest that certain institutional contexts could be more dangerous and that policy responses should be matched to projects' institutions. While the exact implications of institutional context are still under debate, it would help to see which institution types are hosting AGI R&D. How much AGI R&D is open-source?  Bostrom (2017)  and  Goertzel (2017a)  offer contrasting perspectives on the merits of open-source AGI R&D. This is another debate still to be resolved, which meanwhile would benefit from data on the preponderance of open-source AGI R&D. How much AGI R&D has military connections?  Shulman (2009)  proposes that nations may pursue AGI for military dominance. If true, this could have substantial geopolitical implications. While military R&D is often classified, it is worth seeing what military connections are present in publicly available data. Where are AGI R&D projects located?  Wilson (2013)  argues for an international treaty to regulate global AGI R&D, while Scherer (2016) develops a regulatory proposal that is specific to the US. It is thus worth seeing which countries the R&D is located in. What goals do projects have? Section 2.1 summarizes a range of ethical views corresponding to a variety of goals that AGI R&D projects could have. Additionally,  Armstrong et al. (2016)  finds that AGI risk increases if projects have stronger preference for their own AGI relative to others', which may tend to happen more when projects disagree on goals. Thus, it is worth identifying and comparing projects' goals. How engaged are projects on safety issues? Section 2.2 reviews a range of views on the size of AGI risk and the difficulty of making AGI safe, and Section 2.3 summarizes policy literature that is based on the concern that AGI R&D projects may have inadequate safety procedures. Thus, data on how engaged projects are on safety could inform both the size of AGI risk and the policy responses that may be warranted. How large are the projects? Larger projects may be more capable of building AGI. Additionally,  Armstrong et al. (2016)  find that AGI risk increases if projects have similar capability to build AGI. The  Armstrong et al. (2016)  analysis assumes that project capacity is distributed uniformly. It is worth seeing what the distribution of project sizes actually is and which projects are the largest. Project capacity for building AGI is arguably more important than project size. However, project capacity is harder to assess with this paper's methodology of analyzing openly published statements. In addition to project size, project capacity could also depend on personnel talent, the availability of funding, computing power, or other resources, and on how well the project is managed. These factors are often not publicly reported. Another important factor is the viability of the technical approach that a project pursues, but this is not well understood and is a matter of disagreement among AGI experts. While it may be possible to assess project capacity with some degree of rigor, this paper's methodology is not suited for such a task, and thus it is left for future work. Instead, project size may be used as at least a rough proxy for project capacity, though caution is warranted here because it may be an imperfect or even misleading proxy. \n Methodology The paper's method consists of identifying AGI R&D projects and then describing them along several attributes. The identification and description was based on openly published information as found in scholarly publications, project websites, popular media articles, and other websites, with emphasis successful at cooperating when cooperation benefits from larger total available resources  (Yang et al. 2013 ). Thus, for example, one might want a small group for keeping a secret but a large group for fundraising for a fixed project. placed on more authoritative publications. Identification and description were conducted primarily by the present author.  5  In social science terminology, this methodology is known as the \"coding\" of qualitative data  (Coffey and Atkinson 1996; Auerbach and Silverstein 2003) . The data is qualitative in that it consists of text about AGI R&D projects; it is coded into quantitative form, such as \"one academic project and three government projects\". The coding scheme was initially developed based on prior literature and the present author's understanding of the topics and was updated during the coding process based on the present author's reading of the data (known as \"in vivo\" coding). This methodology is fundamentally interpretive, rooted in the researcher's interpretation of the data. Some aspects of the data are not really a matter of interpretation-for example, the fact that the University of Southern California is an academic institution in the US. Other aspects are more open to interpretation. This includes which projects qualify as AGI R&D.  Goertzel (2014, p. 2) refers to the AGI community as a \"fuzzy set\"; this is an apt description. Different researchers may interpret the same data in different ways. They may also find different data as they search through the vast space of openly published information about AGI R&D. Thus, results should be read as one take on AGI R&D and not necessarily a true or complete reflection of the topic. Interested readers are invited to query the data for themselves and make their own interpretations. Appendix 1 contains full descriptions and explanations of coding judgments, citing the corresponding data. (Not all the data is cited-much of what was found is redundant or of limited relevance.) \n Identification of AGI R&D Projects AGI R&D candidate projects were identified via:  The present author's prior knowledge.  Keyword searches on the internet and in scholarship databases, mainly Google web search and Google Scholar.  Previous survey papers  (Chong et al. 2007; Duch et al. 2008; Langley et al. 2008; de Garis et al. 2010; Goertzel et al. 2010   Additional literature and webpages identified via all of the above. Each identified project was put into one of three categories:  Active AGI R&D projects (Appendix 1). These are projects that are working towards building AGI. The included projects either identify as AGI or conduct R&D to build something that is considered to be AGI, human-level intelligence, or superintelligence.  Other notable projects (Appendix 2). These include (1) inactive AGI R&D projects, defined as projects with no visible updates within the last three years; (2) projects that work on technical aspects of AGI but are not working towards building AGI, such as projects working on hardware or safety mechanisms that can be used for AGI; and (3) select narrow AI projects, such as AI groups at major technology companies.  Other projects judged to be not worth including in this paper. Only the active AGI R&D projects are included in the Section 5 data analysis. The other notable projects are reported to document related work, to clarify the present author's thinking about where the boundaries of AGI R&D lie, and to assist in the identification of any AGI R&D projects that have been overlooked by the present research. Projects that only do R&D in deep learning and related techniques were excluded unless they explicitly identify as trying to build AGI. Deep learning already shows some generality  (LeCun et al. 2015) , and some people argue that deep learning could be extended into AGI (e.g.,  Real AI 2017) . Others argue that deep learning, despite its remarkable ongoing successes, is fundamentally limited, and AGI requires other types of algorithms (e.g.,  Wang and Li 2016; Marcus 2017 ). The recent explosion of work using deep learning renders it too difficult to survey using this paper's project-by-project methodology. Furthermore, if all of deep learning was included, it would dominate the results, yielding the unremarkable finding that there is a lot of active deep learning work. The deep learning projects that explicitly identify as trying to build AGI are much smaller in number, fitting comfortably with this paper's methodology and yielding more noteworthy insights. \n Description of AGI R&D Projects For each identified AGI R&D project, a general description was produced, along with classification in terms of the following attributes:  Type of institution: The type of institution in which the project is based, such as academic or government.  Open-source: Whether the project makes its source code openly available.  Military connections: Whether the project has connections to any military activity.  Nationality: The nation where the project is based. For multinational projects, a lead nation was specified, defined as the location of the project's administrative and/or operational leadership, and additional partner countries were tabulated separately.  Stated goal: The project's stated goals for its AGI, defined as what the project aims to accomplish with its AGI and/or what goals it intends to program the AGI to pursue.  Engagement on safety: The extent of the project's engagement with AGI safety issues.  Size: The overall size of the project. \n Type of Institution The type of institution attribute has six categories:  Academic: Institution conducts secondary education (e.g., colleges and universities).  Government: Institution is situated within a local or national government (e.g., national laboratories). This category excludes public colleges and universities.  Nonprofit: Institution is formally structured as a nonprofit and is not an academic institution (e.g., nonprofit research institutes).  Private corporation: Institution is for-profit and does not issue public stock.  Public corporation: Institution is for-profit and does issue public stock.  None: Project is not based within any formal institution. Some projects had two institution types; none had more than two. For the two-type projects, both types were recorded. Project participants were counted only if they are formally recognized on project websites or other key project documents. Some projects had more limited participation from many institutions, such as in co-authorship on publications. This more limited participation was not counted because it would make the entire exercise unwieldy due to the highly collaborative nature of many of the identified projects. This coding policy was maintained across all of the attributes, not just institution type. \n Open-Source The open-source attribute has three categories:  Yes: Project has source code available for download online.  Restricted: Project offers source code upon request.  No: Project does not offer source code. Projects were coded as yes if some source code related to their AGI work is open. Projects that have some, but not all, of their AGI code open are coded as yes. Projects that only have other, non-AGI code open are coded as no. Exactly which code is related to AGI is a matter of interpretation, and different coders may produce somewhat different results. The three open-source categories are mutually exclusive: each project is coded for one of the categories. \n Military Connections The military connections attribute has three categories:  Yes: Project has identifiable military connections.  No: Project is found to have no military connections.  Unspecified: No determination could be made on the presence of military connections. Military connections were identified via keyword searches on project websites and the internet at large, as well as via acknowledgments sections in recent publications. Projects were coded as having military connections if they are based in a military organization, if they receive military funding, or for other military collaborations. Projects were coded as having no military connections if they state that they do not collaborate with militaries or if the entre project could be scanned for connections. The latter was only viable for certain smaller projects. Unless a definitive coding judgment could be made, projects were coded as unspecified. \n Nationality The nationality attribute has two categories:  Lead country: The country in which the project's administrative and/or operational leadership is based  Partner countries: Other countries contributing to the project One lead country was specified for each project; some projects had multiple partner countries. \n Stated Goals The stated goals attribute has six categories:  Animal welfare: AGI is built to benefit nonhuman animals.  Ecocentrism: AGI is built to benefit natural ecosystems.  Humanitarianism: AGI is built to benefit humanity as a whole. This category includes statements about using AGI to solve general human problems such as poverty and disease.  Intellectualism: AGI is built for intellectual purposes, which includes the intellectual accomplishment of the AGI itself and using the AGI to pursue intellectual goals.  Profit: AGI is built to make money for its builders.  Transhumanism: AGI is built to benefit advanced biological and/or computation beings, potentially including the AGI itself.  Unspecified: Available sources were insufficient to make a coding judgment. Some categories of goals found in prior AGI literature did not appear in the data, including military advantage and the selfish benefit of AGI builders. For the coding of stated goals, only explicit statements were considered, not the surrounding context. For example, most AGI R&D projects at for-profit companies did not explicitly state profit as a goal. These projects were not coded under \"profit\" even if it may be the case that they have profit as a goal. \n Engagement on Safety The engagement on safety attribute has four categories:  Active: Projects have dedicated efforts to address AGI safety issues.  Moderate: Projects acknowledge AGI safety issues but lack dedicated efforts to address them.  Dismissive: Projects argue that AGI safety concerns are incorrect.  Unspecified: Available sources were insufficient to make a coding judgment. Each project was coded with one of these categories. In principle, a project can be both active and dismissive, actively working on AGI safety while dismissing concerns about it, though no projects were found to do this. \n Size Finally, the project size attribute has five categories:  Small  Medium-small  Medium  Medium-large  Large  Unspecified: Available information was insufficient to make a coding judgment. This is simply a five-point scale for coding size. Formal size definitions are not used because projects show size in different ways, such as by listing personnel, publications, or AGI accomplishments. \n Main Results This section presents the main results of the survey. Full results are presented in Appendices 1-2. Figure  ES1  in the Executive Summary presents an overview. \n The Identified AGI R&D Projects 45 AGI R&D projects were identified. In alphabetical order, they are:   (Ben Goertzel) , and CogPrime technology is being used by SingularityNET; Blue Brain and the Human Brain Project were both initiated by the same person (Henry Markram) and share a research strategy; Maluuba and Microsoft Research AI have the same parent organization (Microsoft), as do the China Brain Project and RCBII (the Chinese Academy of Sciences); and ACT-R and Leabra were once connected in a project called SAL (an acronym for Synthesis of ACT-R and Leabra; see Appendix 2). This suggests an AGI community that is at least in part working together towards common goals, not competing against each other as is often assumed in the literature (Section 2.2). 1. ACT-R, \n Type of Institution 20 projects are based at least in part in academic institutions, 12 are at least in part in private corporations, 6 are in public corporations, 5 are at least in part in nonprofits, 4 are at least in part in government, and two (AIDEUS, Becca) have no formal institutional home. 4 projects are split across two different institution types: Animats and Soar are academic and private corporation, FLOWERS is academic and government, and DeSTIN is academic and nonprofit. Figure  1  summarizes the institution type data. The preponderance of AGI R&D in academia and for-profit corporations is of consequence for AGI policy. The large number of academic projects suggests merit for the  Yampolskiy and Fox (2013)  proposal for research review boards. However, while academia has a long history of regulating risky research, this has mainly been for medical and social science research that could pose risk to human and animal research subjects, not for research in computer science and related fields. More generally, academic research ethics and regulation tends to focus on the procedure of research conduct, not the consequences that research can have for the world  (Schienke et al. 2009) . There are some exceptions, such as the recent pause on risky gain-of-function biomedical research  (Lipsitch and Inglesby 2014) , but these are the exceptions that prove the rule. Among academic AGI R&D projects, the trend can be seen, for example, in the ethics program of the Human Brain Project, which is focused on research procedure and not research consequences. Efforts to regulate risky AGI R&D in academia would need to overcome this broader tendency. The preponderance of AGI R&D in academia is perhaps to be expected, given the longstanding role of academia in leading the forefront of AI and in leading long-term research in general. But the academic AGI community is rather different in character than the AGI risk literature's common assumption of AGI R&D projects competing against each other to build AGI first. The academic AGI community functions much like any other academic community, doing such things as attending conferences together and citing and discussing each others' work in publications in the same journals. The same can be said for the nonprofit and government projects, which are largely academic in character. Even some of the corporate projects are active in the academic AGI community or related academic communities. This further strengthens the notion of the AGI community as being at least partially collaborative, not competitive. Among the for-profit corporations, two trends are apparent. One is a portion of corporations supporting long-term AGI R&D in a quasi-academic fashion, with limited regard for short-term profit or even any profit at all. For example, Cyc was started in 1984, making it among the longest-running AI projects in history. GoodAI, despite being structured as a for-profit, explicitly rejects profit as an ultimate goal. With less pressure for short-term profit, these groups may have more flexibility to pursue long-term R&D, including for safety mechanisms. The other trend is of AGI projects delivering short-term profits for corporations while working towards long-term AGI goals. As Vicarious co-founder Scott Phoenix puts it, there are expectations of \"plenty of value created in the interim\" while working toward AGI  (High 2016 ). This trend is especially apparent for the AGI groups within public corporations, several of which began as private companies that the public corporations paid good money to acquire (DeepMind/Google, Maluuba/Microsoft, and Uber AI Labs, which was formerly Geometric Intelligence). If this synergy between short-term profit and long-term AGI R&D proves robust, it could fuel an explosion of AGI R&D similar to what is already seen for deep learning.  7  This could well become the dominant factor in AGI R&D; it is of sufficient importance to be worth naming: AGI profit-R&D synergy: any circumstance in which long-term AGI R&D delivers short-term profits. The AGI profit-R&D synergy is an important reason to distinguish between public and private corporations. Public corporations may face shareholder pressure to maximize short-term profits, pushing them to advance AGI R&D even if doing so poses long-term global risks. In contrast, private corporations are typically controlled by a narrower ownership group, possibly even a single person, who can choose to put safety and the public interest ahead of profits. Private corporation leadership would not necessarily do such a thing, but they may have more opportunity to do so. However, it is worth noting that two of the public corporations hosting AGI R&D, Facebook and Google, remain controlled by their founders: Mark Zuckerberg retains a majority of voting shares at Facebook (Heath 2017), while Larry Page and Sergey Brin retain a majority of voting shares at Google's parent company Alphabet  (Ingram 2017) . As long as they retain control, AGI R&D projects at Facebook and Google may be able to avoid the shareholder pressures that public corporations often face. The corporate R&D raises other issues. Corporations may be more resistant of regulation than academic groups for several reasons. First, corporations often see regulation as a threat to their profits and push back accordingly. This holds even when regulation seeks to prevent global catastrophe, such as in the case of global warming (e.g.,  Oreskes and Conway 2010) . Second, when the regulation specifically targets corporate R&D, they often express concern that it will violate their intellectual property and weaken their competitive advantage. Thus, for example, the US successfully resisted verification measures in the negotiation of the Biological Weapons Convention, largely out of concern for the intellectual property of its pharmaceutical industry  (Lentzos 2011) . To the extent that AGI corporations see AGI as important to their business model, they may resist regulations that they believe could interfere. Finally, there are relatively few projects in government institutions. Only one project is based in a military/defense institution: DSO-CA. The other three government projects list intellectual goals in AI and cognitive science, with one also listing medical applications (Section 5.5). However, this understates the extent of government involvement in AGI R&D. Numerous other projects receive government funding, aimed at advancing medicine (e.g., Blue Brain, HBP), economic development (e.g., Baidu), and military technology (e.g., Cyc, Soar). \n Open-Source 25 projects have source code openly available online. An additional three projects (Cyc, HBP, and LIDA) have code available upon request. For these 28 projects, the available code is not necessarily the project's entire corpus of code, at least for the latest version of the code, though in some cases it is. There were only 17 projects for which code could not be found online. For 3 of these 17 projects (Baidu Reseach, Tencent AI Lab, Uber AI Labs), their parent organization has some open-source code, but a scan of this code identified no AGI code. Figure  2  summarizes the open-source data. The preponderance of open-source projects resembles a broader tendency towards openness across much of the coding community. Many of the projects post their code to github.com, a popular code repository. Even the corporate projects, which may have competitive advantage at stake, often make at least some of their code available.  Goertzel (2017a)     \n Military Connections Nine projects have identifiable military connections. These include one project based in a military/defense institution (DSO-CA, at DSO National Laboratories, Singapore's primary national defense research agency) and eight projects that receive military funding (ACT-R, CLARION, Icarus, Leabra, LIDA, Sigma, SNePS, Soar). These eight projects are all based at least in part in US academic institutions. This follows the longstanding trend of military funding for computer science research in the US. Soar is also based in the private company SoarTech, which heavily advertises military applications on its website. Figure  3  summarizes the military connections data. Only four projects were identified as having no military connections. Two of these (Aera, HBP) openly reject military connections; the other two (Alice in Wonderland, Animats) were sufficiently small and well-documented that the absence of military connections could be assessed. The other 32 projects were coded as unspecified. Many of these projects probably do not have military connections because they do not work on military applications and they are not at institutions (such as US universities) that tend to have military connections. Some projects are more likely to have military connections. For example, Microsoft has an extensive Military Affairs program that might (or might not) be connected to MSR AI. The publicly available data suggest that the Singaporean and American militaries' interest in AGI is, by military standards, mundane and not anything in the direction of a quest for unchallenged military dominance. DSO-CA appears to be a small project of a largely academic nature; a recent paper shows DSO-CA applied to image captioning, using an example of a photograph of a family eating a meal  (Ng et al. 2017 ). The project does not have the appearance of a major Singapore power play. Similarly, the military-funded US projects are also largely academic in character; for most of them, one would not know their military connections except by searching websites and publications for acknowledgments of military funding. Only Soar, via SoarTech, publicizes military applications. The publicized applications are largely tactical, suggestive of incremental improvements in existing military capacity, not any sort of revolution in military affairs. It remains possible that one or more militaries are pursuing AGI for more ambitious purposes, as prior literature has suspected  (Shulman 2009) . Perhaps such work is covert. However, this survey finds no evidence of anything to that effect. \n Nationality The projects are based in 14 countries. 23 projects are based in the US, 6 in China, 3 in Switzerland, 2 in each of Sweden and the UK, and 1 in each of Australia, Austria, Canada, Czech Republic, France, Iceland, Japan, Russia, and Singapore. Partner countries include the US (partner for 7 projects), the UK (4 projects), France and Israel (3 projects), Canada, Germany, Portugal, Spain, and Switzerland (2 projects), and Australia, Austria, Belgium, Brazil, China, Denmark, Ethiopia, Finland, Greece, Hungary, Italy, the Netherlands, Norway, Russia, Slovenia, Sweden, and Turkey (1 project). This makes for 30 total countries involved in AGI R&D projects. Figure  4  maps the nationality data.  The 23 US-based projects are based in 12 states and territories: 6 in California, 3 in New York, 2 in each of Massachusetts, Pennsylvania, and Tennessee, and 1 in each of Colorado, Michigan, Ohio, Oregon, Texas, Washington state, and the US Virgin Islands, with one project (CogPrime) not having a clear home state. This broad geographic distribution is due largely to the many academic projects: whereas US AI companies tend to concentrate in a few metropolitan areas, US universities are scattered widely across the country. Indeed, only two of the six California projects are academic. The for-profit projects are mainly in the metropolitan areas of Austin, New York City, Portland (Oregon), San Francisco, and Seattle, all of which are AI hotspots; the one exception is Victor in the US Virgin Islands. Additionally, OpenAI (a nonprofit) and Becca (no institutional home) are in the San Francisco and Boston areas, respectively, due to the AI industries in these cities. 13 projects are multinational. AERA, AIDEUS, Baidu Research, CommAI, Maluuba, Tencent AI Lab, and Uber AI Labs are each in two countries total (including the lead country). Animats, CogPrime, DeepMind, and SOAR are each in three countries. Blue Brain is in five countries. SingularityNET is in 8 countries. The Human Brain Project is in 19 countries, all of which are in Europe except for Israel and (depending on how one classifies it) Turkey. The most common international collaboration is UK-US, with both countries participating in four projects (Blue Brain, DeepMind, Soar, and Uber AI Labs). China and the US both participate in four (Baidu Research, CogPrime, SingularityNET, and Tencent AI Lab). In geopolitical terms, there is a notable dominance of the US and its allies. Only five countries that are not US allies have AGI R&D projects: Brazil, China, Ethiopia, Russia, and (arguably) Switzerland, of which only China and Russia are considered US adversaries. Of the eight projects that China and Russia participate in, one is based in the US (CogPrime), three others have US participation (Baidu Research, SingularityNET, and Tencent AI Lab), and two are small projects with close ties to Western AI communities (AIDEUS, Real AI). The remaining two are projects of the Chinese Academy of Sciences that focus on basic neuroscience and medicine (China Brain Project, RCBII). The concentration of projects in the US and its allies, as well as the Western and academic orientation of the other projects, could make international governance of AGI R&D considerably easier. In geopolitical terms, there is also a notable absence of some geopolitically important regions. There are no projects in South Asia, and just one (CogPrime) that is in part in Africa and one (SingularityNET) that is in part in Latin America. The AGI R&D in Russia consists mainly of the contributions of Alexey Potapov, who contributes to both AIDEUS and SingularityNET. Additionally, Potapov's published AGI research is mainly theoretical, not R&D (e.g.,  Potapov et al. 2016 ).  8  In Latin America, the only AGI R&D is the contributions of Cassio Pennachin to SingularityNET. Thus, the AGI R&D projects identified in this survey are almost all being conducted in institutions based in North America, Europe, the Middle East, East Asia, and Australia. One important caveat is that no partner institutions for ACT-R were included. The ACT-R website lists numerous collaborating institutions, almost all of which are academic, spread across 21 countries. Several of these countries are not involved in other AGI R&D projects: India, Iran, Morocco, Pakistan, South Korea, Sri Lanka, and Venezuela. These partner institutions are excluded because this part of the ACT-R website is out of date. There may be some ongoing contributions to ACT-R in these countries; whether or not there are is beyond the scope of this paper. Another important caveat comes from the many projects with open-source code. This code enables AGI R&D to be conducted anywhere in the world. It is thus possible that there are other countries involved in AGI R&D, perhaps a large number of other countries. The identification of countries whose participation consists exclusively of contributions to open-source code is beyond the scope of this paper. \n Stated Goal The dominant trend among stated goals is the preponderance of humanitarianism and intellectualism. 23 projects stated intellectualist goals and 20 stated humanitarian goals, while only 3 stated ecocentric goals (SingularityNET, Susaro, Victor), 2 stated profit goals (Maluuba, NNAISENSE), 2 stated animal welfare goals (Human Brain Project, proposing brain simulation to avoid animal testing, and SingularityNET, seeking to benefit all sentient beings), and 1 stated transhumanist goals (SingularityNET, seeking to benefit sentient beings and robots). 12 projects had unspecified goals. Some projects stated multiple goals: 11 projects stated 2 goals, 1 project (Human Brain Project) stated 3, and 1 project (SingularityNET) stated 4. Figure  5  summarizes the stated goal data. The intellectualist projects are predominantly academic, consistent with the wider emphasis on intellectual merit across academia. The intellectualist projects include 14 of the 19 projects based at least in part at an academic institution. (Each of the other five academic projects are coded as unspecified; they are generally smaller projects and are likely to have intellectualist goals.) The intellectualist projects also include the two projects at the Chinese Academy of Sciences. The Academy is coded as a government institution but is heavily academic in character. The other intellectualist projects are all in for-profit corporations, except AIDEUS, which has no institution type. Four of these corporate projects have close ties to academia (GoodAI, HTM, Maluuba, and NNAISENSE), as does AIDEUS. The fifth, MSR AI, is a bit of an outlier. All eight US academic projects with military connections state intellectualist goals. Only one of them (LIDA) also states humanitarian goals. (The one non-US project with military connections, DSO-CA, had unspecified goals.) Only one other US academic project states humanitarian goals (NARS); the rest are unspecified. Meanwhile, three of eight non-US academic projects state humanitarian goals. One possible explanation for this is that the preponderance of military funding for US academic AGI R&D prompts projects to emphasize intellectual goals instead of humanitarian goals, whereas the availability of funding in other countries (especially across Europe) for other AGI applications, especially healthcare, prompts more emphasis on humanitarian goals. In short, the data may reflect the large US military budget and the large European civil budget, while also indicating that AGI researchers struggle to articulate military R&D in humanitarian terms. 10 of 18 for-profit projects state humanitarian goals. Many of these are Western (especially American) projects with a strong public face (e.g., CommAI/Facebook, DeepMind/Google, MSR AI/Microsoft, Nigel/Kimera, Uber AI Labs). Some of the other humanitarian for-profit projects are rooted in Western conceptions of altruism (e.g., GoodAI, Real AI). In contrast, the non-humanitarian for-profits are mostly smaller projects (e.g., Animats, NNAISENSE) and Chinese projects (Baidu Research, Tencent AI Lab). This suggests that the for-profit humanitarianism is mainly a mix of Western values and Western marketing. The distinction between values and marketing is an important one and speaks to a limitation of this survey's methodology. Projects may publicly state goals that are appealing to their audiences while privately holding different goals. This may explain why ten for-profit projects state humanitarian goals while only two state profit goals. Some for-profit projects may genuinely care little about profitindeed, two of them, GoodAI and Vicarious, explicitly reject profit as a goal. But others may only articulate humanitarianism to make themselves look good. This practice would be analogous to the practice of \"greenwashing\", in which environmentally damaging corporations promote small activities that are environmentally beneficial to create the impression that they are more \"green\" than they actually are (e.g.,  Marquis et al. 2016) . For example, a coal company might promote employees' inoffice recycling to show their environmental commitment. Likewise, corporate AGI R&D projects may advertise humanitarian concern while mainly seeking profit, regardless of overall humanitarian consequences. One might refer to such conduct as \"bluewashing\", blue being the color of the United Nations and the World Humanitarian Summit. Notable absences in the stated goals data include military advantage and the selfish benefit of AGI builders. Both of these is considered as an AGI R&D goal in prior literature, as is transhumanism/cosmism, which only gets brief support in one project (SingularityNET). The reason for these absences is beyond the scope of this paper, but some possibilities are plausible. Transhumanism and cosmism are not widely held goals across contemporary human society, though they are relatively common among AGI developers (e.g.,  Goertzel 2010) . It is plausible that transhumanists and cosmists (outside of SingularityNET) prefer to keep their views more private so as to avoid raising concerns that their AGI R&D projects could threaten humanity. The pursuit of AGI for military advantage could raise similar concerns and could also prompt adversaries to commence or hasten their own AGI R&D. Finally, the pursuit of AGI for selfish benefit is antisocial and could pose reputational risks or prompt regulation if stated openly. Yet it is also possible that the people currently drawn to AGI R&D tend to actually have mainly humanitarian and intellectualist goals and not these other goals (Goertzel being a notable exception). An important question is the extent to which AGI R&D projects share the same goals. Projects with different goals may be more likely to compete against each other to build AGI first  (Armstrong et al. 2016) . The preponderance of humanitarianism and intellectualism among AGI R&D projects suggests a considerable degree of consensus on goals. Furthermore, these goals are agent-neutral, further suggesting a low tendency to compete. But competition could occur anyway. One reason is that there can be important disagreements about the details of these particular views, such as in divergent conceptions of human rights between China and the West  (Posner 2014) . Additionally, even conscientious people can feel compelled to be the one to build AGI first, perhaps thinking to themselves \"Forget about what's right and wrong. You have a tiger by the tail and will never have as much chance to influence events. Run with it!\" 9 And of course, there can be disagreements between AGI humanitarians and intellectualists, as well as with the other goals that have been stated. Finally, it should be noted that the coding of stated goals was especially interpretative. Many projects do not state their goals prominently or in philosophically neat terms. For example, DeepMind lists climate change as an important application. This could be either ecocentric or humanitarian or both, depending on why DeepMind seeks to address climate change. It was coded as humanitarian because it was mentioned in the context of \"helping humanity tackle some of its greatest challenges\", but it is plausible that ecocentrism was intended. \n Engagement on Safety Engagement on safety could only be identified for 17 projects. 12 of these projects were found to be active on safety (AERA, AIDEUS, CogPrime, DeepMind, FLOWERS, GoodAI, LIDA, NARS, OpenAI, Real AI, Susaro, and WBAI), 3 are moderate (CommAI, Maluuba, and Vicarious), and 2 are dismissive (HTM, Victor). Figure  6  summarizes the engagement on safety data. One major trend in the engagement on safety data is the lack of attention to safety. Engagement on safety could not be identified for 28 of 45 projects. This provides some empirical support for the common assumption in prior AGI policy literature of AGI developers who want to proceed with inadequate regard for safety (Section 2.3). This survey's focus on publicly available data may overstate the neglect of safety because some projects may pay attention to safety without stating it publicly. For example, Animats and NNAISENSE were coded as unspecified, but there is little publicly available information about any aspect of these projects. Meanwhile, NNAISENSE Chief Scientist Jürgen Schmidhuber and NNAISENSE co-founder and Animats contributor Bas Steunebrink have done work on AGI safety  (Steunebrink et al. 2016 ). Still, the data is strongly suggestive of widespread neglect of safety among AGI R&D projects. Among the 17 projects for which engagement on safety was identified, some further trends are apparent. These 17 projects include 6 of the 8 projects with purely humanitarian goals and the only project with purely ecocentric goals yet only 1 of the 10 projects with purely intellectualist goals and 1 of 11 with unspecified goals. The 17 projects also include 9 of the 16 projects based purely at a forprofit and 3 of 4 projects based purely at a nonprofit yet only 2 of 15 projects based purely at an academic institution and 1 of 4 based in part at an academic institution. This suggests a cluster of projects that are broadly engaged on the impacts of AGI R&D, including ethics questions about what the impacts should be and risk/safety questions about whether the desired impacts will accrue, a cluster that is located predominantly outside of academia. Meanwhile, there is also a cluster of projects that are predominantly academic and view AGI in largely intellectual terms, to the extent that they state any goal at all. These trends suggest the importance of proposals to strengthen risk and ethics practices sometimes suggesting that adequate training could make AGI safety to be a reasonably tractable task among academic AGI R&D projects, such as via research review boards  (Yampolskiy and Fox 2013) . Among the 12 projects found to be active on safety, a few themes are apparent. Several projects emphasized the importance of training AGI to be safe (AERA, CogPrime, GoodAI, and NARS), and critiquing concerns that AGI safety could be extremely difficult (Section 2.2;  Goertzel and Pitt 2012; Goertzel 2015; Bieger al. 2015; Steunebrink et al. 2016) .  10  Other projects focus on safety issues in near-term AI in the context of robotics (FLOWERS;  Oudeyer et al. 2011 ) and reinforcement learning (DeepMind and OpenAI;  Christiano et al. 2017) , the latter being consistent with an agenda of using near-term AI to study long-term AI safety that was proposed by Google and OpenAI researchers  (Amodei et al. 2016) . LIDA has explored fundamentals of AGI morality as they relate to engineered systems  (Wallach et al. 2010; Madl and Franklin 2015) . WBAI suggests that it seeks to build brain-like AGI in part because its similarity to human intelligence would make it safer. \n Size Projects were found to have size mainly in the small to medium range. 13 projects were coded as small, 11 as medium-small, 12 as medium, 5 as medium-large (Blue Brain, CogPrime, MSR AI, Soar, and Vicarious), and 3 as large (DeepMind, Human Brain Project, and OpenAI), with 1 project unspecified (Susaro, a project in \"stealth mode\" with no size information found). Figure  7  summarizes the size data. While the projects at each size point are diverse, some trends are nonetheless apparent. With respect to institution type, academic projects tend to be somewhat smaller, while corporate projects tend to be somewhat larger, though there are both large academic projects and small corporate projects. 7 of the 13 small projects are academic, while only three are in private corporations and none are in public corporations. 3 of the 8 medium-large or large projects are academic, while 4 are in corporations. The size vs. institution type trends may be in part a coding artifact, because academic and corporate projects have different indicators of size. Corporate projects are less likely to list their full personnel or to catalog their productivity in open publication lists. Corporate projects may also have substantial portions focused on near-term applications at the expense of long-term AGI R&D, though academic projects may similarly have portions focused on intellectual goals at the expense of AGI R&D. Corporate projects, especially those in public corporations, often have additional resources that can be distributed to AGI projects, including large datasets and funding. Academic institutions can find funding for AGI projects, but generally not as much as corporations, especially the large public corporations. This distinction between academic and corporate projects is illustrated, for example, by NNAISENSE, which was launched by academics in part to prevent other companies from poaching research talent. It further suggests that the largest AGI R&D projects may increasingly be corporate, especially if there is AGI R&D-profit synergy (Section 5.2). With respect to open-source, there is a clear trend towards larger projects being more open-source. Of the 17 projects that do not make source code available, 8 are small and 6 are medium-small, while only 1 is medium and 1 is medium-large, with 1 being of unspecified size. In contrast, the 24 unrestricted open-source projects include 5 of 11 small projects, 5 of 11 medium-small projects, 9 of 12 medium projects, 4 of 5 medium-large projects, and 2 of 3 large projects. With respect to military connections, the projects with military connections tend to be in the small to medium range. The 8 military projects include 2 small, 2 medium-small, 4 medium, and 1 mediumlarge. This makes sense given that these are primarily US academic projects with military funding. Academic projects tend to be smaller, while those with external funding tend to be more medium in size. Academic projects that forgo military funding may sometimes be smaller than they otherwise could have been. With respect to nationality, there is some trend towards US-based projects being larger while China-based projects are smaller, though, for the US, it is a somewhat weak trend. The 23 US-based projects include 6 of 13 small projects, 3 of 11 medium-small projects, 9 of 12 medium projects, 4 of 5 medium-large projects, and 1 of 3 large projects. In contrast, all six China-based projects are either small or medium-small. This trend strengthens the Section 5.5 finding that AGI is concentrated in the US and its allies. An important caveat is that two of the Chinese projects are based in large Chinese corporations, Baidu and Tencent. These corporations have large AI groups that only show a small amount of attention to AGI. If the corporations decide to do more in AGI, they could scale up quickly. It is also possible that they are already doing more, in work not identified by this survey, though the same could also be said for projects based in other countries. With respect to stated goal, a few trends can be discerned. First, the projects with unspecified goals tend to be small, including 7 of 13 small projects. This makes sense: smaller projects have less opportunity to state their goals. Intellectualist projects tend to be medium-small to medium, including 6 of 11 medium-small projects and 9 of 12 medium projects, which is consistent with the intellectualist projects tending to be academic. Humanitarian projects tend to be larger, including 6 of 12 medium projects, 3 of 5 medium-large projects, and all 3 large projects. A possible explanation is that the larger projects tend to have wider societal support, prompting the projects to take a more humanitarian position. The preponderance of humanitarianism among the larger projects could facilitate the development of consensus on goals among the projects that are most likely to build AGI. Such consensus could in turn help to avoid a risky race dynamic  (Armstrong et al. 2016) . Finally, with respect to engagement on safety, there is a weak trend towards the larger projects being more active on safety. The active projects include 2 small, 3 medium-small, 3 medium, 1 medium-large, 2 large, and 1 of unspecified size. In contrast, the projects with unspecified engagement on safety include 10 small, 7 medium-small, 6 medium, 3 medium-large, and 1 large. Thus, projects of all sizes can be found active or not active on safety, though the larger projects do have a somewhat greater tendency to be active. \n Other Notable Projects 47 other notable projects were recorded in the process of identifying AGI R&D projects. These include 26 inactive AGI R&D projects, 9 projects that are not AGI, and 12 projects that are not R&D. Unlike with active AGI R&D projects, no attempt was made to be comprehensive in the identification of other notable projects. It is likely that some notable projects are not included. The list of other notable projects and brief details about them are presented in Appendix 2. The 26 inactive projects are mainly academic, such as 4CAPS (led by Marcel Just of Carnegie Mellon University), Artificial Brain Laboratory (led by Hugo de Garis of Xiamen University), CERA-CRANIUM (led by Raúl Arrabales of University of Madrid), CHREST (led by Fernand Gobet of University of Liverpool), DUAL (led by Boicho Kokinov of New Bulgarian University), EPIC (led by David Kieras at University of Michigan), OSCAR (led by John Pollock of University of Arizona), and Shruti (led by Lokendra Shastri of University of California, Berkeley). They varied considerably in duration, from a few years (e.g., AGINAO, active from 2011 to 2013) to over a decade (e.g., CHREST, active from 1992 to 2012). The nine projects that are not AGI are mostly AI groups at large computing technology corporations. Six corporations were searched carefully for AGI projects and found not to have any: Alibaba, Amazon, Apple, Intel, SalesForce, and Twitter. 11 Given these corporations' extensive resources, it is notable that they do not appear to have any AGI projects. 12 In addition to DeepMind, two other projects at Google were considered: Google Brain and Quantum AI Lab. While Google Brain has done some AGI work with DeepMind, it focuses on narrow AI. Finally, one narrow cognitive architecture project was included (Xapagy) as an illustrative example. Many more narrow cognitive architecture projects could have been included-  Kotseruba et al. (2016)  lists 86 cognitive architecture projects, most of which are narrow. The 12 projects that are not R&D cover a mix of different aspects of AGI. Some focus on basic science, including several brain projects (e.g., the BRAIN Initiative at the US National Institutes of Health and Brain/MINDS at Japan's Ministry of Education, Culture, Sports, Science, and Technology). Several focus on hardware and software for building AGI (e.g., the IBM Cognitive Computing Project, the Cognitive Systems Toolkit project led by Ricardo Gudwin of University of Campinas in Brazil, and the Neurogrid project led by Kwabena Boahen of Stanford University). Two focus on safety aspects of AGI design (Center for Human-Compatible AI at University of California, Berkeley and the Machine Intelligence Research Institute). One (Carboncopies) focuses on supporting other R&D projects. Finally, one focuses on theoretical aspects of AGI (Goedel Machine, led by Jürgen Schmidhuber of the Dalle Molle Institute for Artificial Intelligence Research in Switzerland). This is not a comprehensive list of projects working on non-R&D aspects of AGI. For example, projects working on AGI ethics, risk, and policy were not included because they are further removed from R&D. \n Conclusion Despite the seemingly speculative nature of AGI, R&D towards building it is already happening. This survey identifies 45 AGI R&D projects spread across 30 countries in 6 continents, many of which are based in major corporations and academic institutions, and some of which are large and heavily funded. Given that this survey relies exclusively on openly published information, this should be treated as a lower bound for the total extent of AGI R&D. Thus, regardless of how speculative AGI itself may be, R&D towards it is clearly very real. Given the potentially high stakes of AGI in terms of ethics, risk, and policy, the AGI R&D projects warrant ongoing attention. \n Main Findings Regarding ethics, the major trend is projects' split between humanitarian and intellectualist goals, with the former coming largely from corporate projects and the latter from academic projects. There is reason to be suspicious of corporate statements of humanitarianism-they may be \"bluewashing\" (Section 5.6) to conceal self-interested profit goals. Still, among the projects not motivated by intellectual goals, there does seem to be a bit of a consensus for at least some form of humanitarianism, and not for other types of goals commonly found in AGI discourses, such as cosmism/transhumanism. Meanwhile, the intellectualist projects indicate that academic R&D projects still tend to view their work in intellectual terms, instead of in terms of societal impacts or other ethical factors, even for potentially high-impact pursuits like AGI. Regarding risk, two points stand out. First, a clear majority of projects had no identifiable engagement on safety. While many of these projects are small, it includes even some of the larger projects. It appears that concerns about AGI risk have not caught on across much of AGI R&D, especially within academia. Second, some trends suggest that a risky race dynamic may be avoidable. One is the concentration of projects, especially larger projects, in the US and its allies, which can facilitate both informal coordination and formal public policy. Another is the modest consensus for humanitarianism, again especially among larger projects, which could reduce projects' desire to compete against each other. Finally, many of the projects are interconnected via shared personnel, parent organizations, AGI systems development, and participating in the same communities, such as the AGI Society. This suggests a community working together towards a common goal, not competing against each other to \"win\". Regarding policy, several important points can be made. One is the concentration of projects in the US and its allies, including most of the larger projects (or all of them, depending on which countries are considered US allies). This could greatly facilitate the establishment of AGI policy with jurisdiction for most AGI R&D projects, including all of the larger ones. Another important point is the concentration of projects in academia and corporations, with relatively little in government or with military connections. Each institution type merits its own type of policy, such as review boards in academia and financial incentive structures for corporations. The potential for AGI R&D-profit synergy (Section 5.2) could be especially important here, determining both the extent of corporate R&D and the willingness of corporations to submit to restrictive regulations. This survey finds hints of AGI R&D-profit synergy, but not the massive synergies found for certain other types of AI. Finally, the preponderance of projects with at least some open-source code complicates any policy effort, because the R&D could in principle be done by anyone, anywhere. \n Limitations and Future Work The above conclusions seem robust given this study's methodology, but other methodologies could point in different directions. For example, the published statements about goals suggest a consensus towards humanitarian and intellectualist goals, but this could miss unpublished disagreements on the specifics of these goals. One potential flashpoint is if Western humanitarian AGI projects seek to advance political freedom, whereas Chinese AGI projects seek to advance economic development, in parallel with the broader disagreement about human rights between China and the West. Furthermore, the published statements about goals used in this survey could deviate from projects' actual goals if they are not entirely honest in their published statements. Projects may be proceeding recklessly towards selfish goals while presenting a responsible, ethical front to the public. These possibilities suggest a more difficult policy challenge and larger AGI risk. Alternatively, this study could overestimate the risk. Perhaps many projects have similar goals and concerns about safety, even if they have not published any statements to this effect. Future research using other methodologies, especially interviews with people involved in AGI R&D, may yield further insights. A different complication comes from the possibility that there are other AGI R&D projects not identified by this study. Some projects may have a public presence but were simply not identified in this study's searches, despite the efforts made to be comprehensive. This study is especially likely to miss projects that work in non-English languages, since it only conducted searches in English. However, English is the primary international language for AI and for science in general, so it is plausible that no non-English projects were missed. Furthermore, there may be additional AGI R&D projects that have no public face at all. This could conceivably include the secret government military projects that some might fear. However, the modest nature of the military connections of projects identified in this survey suggests that there may be no major military AGI projects at this time. Specifically, the projects identified with military connections are generally small and focused on mundane (by military standards) tactical issues, not grand ambitions of global conquest. This makes it likely that there are not any more ambitious secret military AGI projects at this time. However, given the stakes, it is important to remain vigilant about the possibility. Another, more likely possibility is of stealth mode private sector projects. This study identified one stealth mode project that happens to have a public website (Susaro); perhaps there are others without websites. Some of these may be independent startups, while others may be projects within larger AI organizations. The larger organizations in particular often have resources to support AGI R&D and could camouflage it within other AI projects. Some larger AI organizations, such as Apple, have reputations for secrecy and may be likely candidates for hosting stealth AGI projects. The possibility of stealth projects or other projects not identified in this survey means that the number of projects identified in this survey (45) should be treated as a lower bound. A different potential source of additional AGI R&D projects is the vast space of projects focused on deep learning and related techniques. These projects were excluded from this survey because there are too many to assess in this study's project-by-project methodology and because there are diverging opinions on whether these projects rate as AGI R&D. If AGI could come from deep learning and related techniques, the AGI R&D landscape would look substantially different from the picture painted in this paper, with major consequences for ethics, risk, and policy. Therefore, an important direction for future research is to assess the possibility that AGI could come from deep learning and related techniques and then relate this to ethics, risk, and policy. Another worthwhile direction for future research is on projects' capability to build AGI. This study includes project size as a proxy for capability, but it is an imperfect proxy. Capability is the more important attribute for understanding which projects may be closest to building AGI. More capable projects may warrant greater policy attention, and a cluster of projects with similar capability could lead to a risky race dynamic. (Project size is important for other reasons, such as projects' pull on the labor market for AGI researchers or their potential for political lobbying.) Project capacity could be assessed via attention to details of projects' performance to date, the viability of their technical approaches to AGI, and other factors. Given the ethics, risk, and policy importance of project capacity, this is an important direction for future research. Finally, future research could also investigate other actors involved in AGI. In addition to the R&D projects, there are also, among others, R&D groups for related aspects of AGI, such as hardware and safety measures; people studying and working on AGI ethics, risk, and policy; and \"epistemic activists\" promoting certain understandings of AGI. Each of these populations can play significant roles in ultimate AGI outcomes and likewise has implications for ethics, risk, and policy that can be worth considering. Empirical study of these populations could clarify the nature of the work being done, identify gaps, and suggests trends in how AGI could play out. \n Concluding Remarks Overall, the present study shows that AGI ethics, risk, and policy can be pursued with a sound empirical basis-it need not be based solely on speculation about hypothetical AGI R&D projects and actors. The present study additionally makes progress on this empirical basis by contributing the firstever survey of AGI R&D projects for ethics, risk, and policy. Given the potentially high stakes of AGI, it is hoped that this research can be used productively towards improving AGI outcomes. \n Appendix 1. Active AGI R&D Projects \n ACT-R Main website: http://act-r.psy.cmu.edu ACT-R is a research project led by John Anderson of Carnegie Mellon University. It is a theory of human cognition and a computational framework for simulating human cognition.  13  ACT-R is an acronym for Adaptive Control of Thought-Rational. It was briefly connected to Leabra via the SAL project. \n Lead institutions: Carnegie Mellon University Partner institutions: none  The ACT-R website lists numerous collaborating institutions across 21 countries, 14 though this includes people who previously contributed and have since moved on to other projects, and does not include some co-authors of recent papers.  15  No active partner institutions could be confirmed from the website and thus none are coded here, though there may be active partner institutions. \n Type of institution: academic Open-source: yes 16  \n Lead country: Iceland Partner countries: Switzerland Stated goals: humanitarianism, intellectualism  Thórisson's website links to the ethics policy of IIIM, which aims to \"to advance scientific understanding of the world, and to enable the application of this knowledge for the benefit and betterment of humankind\", with emphasis on concerns about privacy and military misuse. 21  The ethics policy also states an aim \"to focus its research towards topics and challenges of obvious benefit to the general public, and for the betterment of society, human livelihood and life on Earth\". This mention of life on Earth suggests ecocentrism, but all other text is humanitarian or intellectualist. 22 Engagement on safety: active  The AERA group has written on how to enhance safety during AGI self-improvement, arguing that certain design principles would make it easy to achieve safety  (Steunebrink et al. 2016 ).  Thórisson's website links to an article by AI researcher Oren  Etzioni (2016)  that is dismissive of concerns about AGI, suggesting that AERA may also be dismissive, but this could not be concluded from just the presence of the link.   24  Its approach is to \"proceed from universal prediction models on the basis of algorithmic probability used for choosing optimal actions\".  25  It has published frequently on AGI, often in the proceedings of AGI conferences.  26  Recent publications report funding from the Russian Federation, including the Ministry of Education and Science (e.g.,  Potapov et al. 2016) . \n Lead institutions: AIDEUS \n Partner institutions: none Type of institution: none  No institution type is specified on the project website. AIDEUS is listed separately from academic affiliations on publications (e.g.,  Potapov et al. 2016) . It is listed as a company on its Facebook page, 27 but the Facebook \"company\" category is not restricted to corporations. \n Open-source: no Military connection: unspecified  Funding sources include \"Government of Russian Federation, Grant 074-U01\", which does not appear to be military, but this could not be confirmed. \n Lead country: Russia Partner countries: France Stated goals: humanitarian, intellectualist  The project aims to build superintelligence in order to \"help us better understand our own thinking and to solve difficult scientific, technical, social and economic problems.\" 28 Engagement on safety: active  AIDEUS has published AGI safety research, e.g.  Potapov and Rodionov (2014) . Type of institution: academic  Blue Brain lists the public corporation IBM as a contributor, making \"its researchers available to help install the BlueGene supercomputer and set up circuits that would be adequate for simulating a neuronal network\".  54  This contribution was judged to be too small to code Blue Brain as part public corporation. \n Size \n Open-source: yes 55 Military connection: unspecified \n Lead country: Switzerland Partner countries: Israel, Spain, UK, USA Stated goals: humanitarianism, intellectualism  The Blue Brain website states that \"Understanding the brain is vital, not just to understand the biological mechanisms which give us our thoughts and emotions and which make us human, but for practical reasons,\" the latter including computing, robotics, and medicine.  56     84  The Cycorp website describes Cyc as \"a long-term quest to develop a true artificial intelligence\".  85  Cyc has a unique database of millions of hand-coded items of commonsense human knowledge, which it aims to leverage for human-level AGI.  86  In an interview, Lenat says \"Cycorp's goal is to codify general human knowledge and common sense so that computers might make use of it\" (emphasis original). 87 \n Lead institution: Cycorp Partner institutions: none \n Type of institution: private corporation Open-source: restricted  Cycorp offers a no-cost license for researchers upon request 88  Part of Cyc was briefly made available as OpenCyc, but this was discontinued 89  Cycorp \"has placed the core Cyc ontology into the public domain\".  \n DeepMind Main website: http://deepmind.com DeepMind is an AI corporation led by Demis Hassabis, Shane Legg, and Mustafa Suleyman. It was founded in 2010 and acquired by Google in 2014 for £400m ($650m; Gibbs 2014). It seeks to develop \"systems that can learn to solve any complex problem without needing to be taught how\", and it works \"from the premise that AI needs to be general\".  92  DeepMind publishes papers on AGI, e.g. \"PathNet: Evolution Channels Gradient Descent in Super Neural Networks\"  (Fernando et al. 2017) . \n Lead institution: Google Partner institutions: none \n Type of institution: public corporation Open-source: yes 93 \n Lead country: UK Military connection: unspecified  Google has extensive defense contracts in the US, 94 but these appear to be unrelated to DeepMind Partner countries: USA, Canada  DeepMind is based in London and also has a team at Google headquarters in Mountain View, California  (Shead 2017) . It recently opened an office in Edmonton, Alberta. 95 Stated goals: humanitarianism  Their website presents a slogan \"Solve intelligence. Use it to make the world a better place.\"; it describes \"AI as a multiplier for human ingenuity\" to solve problems like climate change and healthcare; and states \"We believe that AI should ultimately belong to the world, in order to benefit the many and not the few\".  96  Similarly, it writes AI will be \"helping humanity tackle some of its greatest challenges, from climate change to delivering advanced healthcare\". 97 Engagement on safety: active  DeepMind insisted on an AI ethics board at Google during its acquisition  (Gibbs 2014)   \n Type of institution: private corporation Open-source: yes 111 \n Military connection: unspecified \n Lead country: Czech Republic Partner countries: none Stated goals: humanitarianism, intellectualist  The GoodAI website states that its \"mission is to develop general artificial intelligence -as fast as possible -to help humanity and understand the universe\" and that it aims \"to build general artificial intelligence that can find cures for diseases, invent things for people that would take much longer to invent without the cooperation of AI, and teach us much more than we currently know about the universe.\" 112 It emphasizes that building AGI \"is not a race. It's not about competition, and not about making money. (1) neuroscience research, with an intellectualist theme, e.g. \"Numenta is tackling one of the most important scientific challenges of all time: reverse engineering the neocortex\", and (2) machine intelligence technology, with a humanitarian theme, e.g. stating it is \"important for the continued success of humankind\". 119  Hawkins writes that \"The future success and even survival of humanity may depend on\" humanity \"building truly intelligent machines\", citing applications in energy, medicine, and space travel. 120  The Numenta Twitter states that \"Only if we make AI a public good, rather than the property of a privileged few, we can truly change the world.\" 121 Engagement on safety: dismissive  Hawkins has dismissed concerns about AGI as a catastrophic risk, stating \"I don't see machine intelligence posing any threat to humanity\"  (Hawkins 2015) . Human Brain Project (HBP)   157  The project seeks \"to probe the foundational principles of intelligence, including efforts to unravel the mysteries of human intellect, and use this knowledge to develop a more general, flexible artificial intelligence\".  158  The project pulls together more than 100 researchers from different branches of AI at Microsoft's Redmond headquarters.  159  By pulling together different branches, MSR AI hopes to achieve more sophisticated AI, such as \"systems that understand language and take action based on that understanding\".  160  However, it has also been criticized for a potentially unwieldy organizational structure. The NARS website explains that NARS is \"morally neutral\" in the sense that it can be programmed with any moral system. 170  The NARS website emphasizes that NARS should aim for a positive impact \"on the human society\" and be \"human-friendly\". 171  The NARS website states that \"the ultimate goal of this research is to fully understand the mind, as well as to build thinking machines\". 172 Engagement on safety: active  Wang has written on safety issues in NARS, such as \"motivation management\", a factor in the ability of NARS to reliably pursue its goals and not be out of control. Stated goals: intellectualist  A project document states that SiMA was founded \"to understand how the brain as a whole works\". 212  The document also discusses applications in automation and the prospect that \"machines will have feelings\",  213  but no specific goals could be identified from this. \n Engagement on safety: unspecified Size: medium \n SingularityNET Main website: https://singularitynet.io SingularityNET is an AGI project led by Ben Goertzel. It was publicly launched in 2017.  214  It aims to be a platform in which anyone can post AI code or use AI code posted by other people. It plans to use cryptocurrency for payments for the use of AI on its site. This setup aims to makie AI more democratic than what could occur via governments or corporations  (Goertzel 2017b) . SingularityNET plans to have decision making done by voting within its user community  (Goertzel et al. 2017) . \n Lead institutions: SingularityNET Foundation Partner institutions: OpenCog Foundation  Several other partners are listed on the SingularityNET website, but OpenCog Foundation is the only one that contributes AGI R&D. \n Type of institution: nonprofit 215 Open-source: yes 216 \n Military connection: unspecified Lead country: China 217 Partner countries: Australia, Brazil, Canada, Germany, Portugal, Russia, USA 218 Stated goals: animal welfare, ecocentric, humanitarian, transhumanist  SingularityNET is described as \"for the People (and the Robots!)\" and \"the happiness of sentient beings\", with \"benefits for all people, and for all life\"  (Goertzel 2017b; Goertzel et al. 2017 ).  SingularityNET also states a goal of profit, but describes this as a means to other goals, for example stating \"SingularityNET has the potential to profit tremendously…  [and]  to direct the profit thus generated to apply AI for global good\"  (Goertzel 2017b) . \n Engagement on safety: unspecified Size: medium-small SNePS (Semantic Network Processing System)   (Chen 2017 ). Marcus has since left UAIL  (Chen 2017) . Upon acquisition by Uber, Geometric Ingelligence's personnel relocated to San Francisco, except for Ghahramani, who remained in Cambridge, UK  (Metz 2016) . UAIL is reportedly part of Uber's attempt to expand substantially beyond the private taxi market, similar to how Amazon expanded beyond books  (Metz 2016) . According to Ghahramani, the AI combines \"some of the ideas in ruled-based learning with ideas in statistical learning and deep learning\"  (Metz 2016) .   244  Its states that it is \"building systems to bring human-like intelligence to the world of robots\".  245  In an interview, Phoenix says that Vicarious is working towards AGI, with \"plenty of value created in the interim\", 246 and that AGI would be \"virtually the last invention humankind will ever make\".  \n Lead institutions: Figure 1 . 1 Figure 1. Summary of institution type data. The figure shows more than 45 data points because some projects have multiple institution types. \n Figure 2 . 2 Figure 2. Summary of open-source data. \n Figure 3 . 3 Figure 3. Summary of military connections data. \n Figure 4 . 4 Figure 4. Map of nationality data. Any depictions of disputed territories are unintentional and do not indicate a position on the dispute. \n Figure 5 . 5 Figure 5. Summary of stated goal data. The figure shows more than 45 data points because some projects have multiple stated goals. \n Figure 6 . 6 Figure 6. Summary of engagement on safety data. \n Figure 7 . 7 Figure 7. Summary of size data. \n led by John Anderson of Carnegie Mellon University 2. AERA, led by Kristinn Thórisson of Reykjavik University 3. AIDEUS, led by Alexey Potapov of ITMO University and Sergey Rodionov of Aix Marseille Université 4. AIXI, led by Marcus Hutter of Australian National University 5. Alice in Wonderland, led by Claes Strannegård of Chalmers University of Technology 6. Animats, a small project recently initiated by researchers in Sweden, Switzerland, and the US 7. Baidu Research, an AI research group within Baidu 8. Becca, an open-source project led by Brandon Rohrer 9. Blue Brain, led by Henry Markram of École Polytechnique Fédérale de Lausanne 10. China Brain Project, led by Mu-Ming Poo of the Chinese Academy of Sciences 11. CLARION, led by Ron Sun of Rensselaer Polytechnic Institute 12. CogPrime, an open source project led by Ben Goertzel based in the US and with dedicated labs in Hong Kong and Addis Ababa 13. CommAI, a project of Facebook AI Research based in New York City and with offices in Menlo Park, California and Paris 14. Cyc, a project of Cycorp of Austin, Texas, began by Doug Lenat in 1984 15. DeepMind, a London-based AI company acquired by Google in 2014 for £400m ($650m) 16. DeSTIN, led by Itamar Arel of University of Tennessee 17. DSO-CA, led by Gee Wah Ng of DSO National Laboratories, which is Singapore's primary national defense research agency 18. FLOWERS, led by Pierre-Yves Oudeyer of Inria and David Filliat of Ensta ParisTech 19. GoodAI, an AI company based in Prague led by computer game entrepreneur Marek Rosa 20. HTM, a project of the AI company Numenta, based in Redwood City, California and led by Jeffrey Hawkins, founder of Palm Computing 21. Human Brain Project, a consortium of research institutions across Europe with $1 billion in funding from the European Commission 22. Icarus, led by Pat Langley of Stanford University 23. Leabra, led by Randall O'Reilly of University of Colorado 24. LIDA, led by Stan Franklin of University of Memphis 25. Maluuba, a company based in Montreal recently acquired by Microsoft 26. MicroPsi, led by Joscha Bach of Harvard University 27. Microsoft Research AI, a group at Microsoft announced in July 2017 28. MLECOG, led by Janusz Starzyk of Ohio University 29. NARS, led by Pei Wang of Temple University 30. Nigel, a project of Kimera, an AI company based in Portland, Oregon 31. NNAISENSE, an AI company based in Lugano, Switzerland and led by Jürgen Schmidhuber 32. OpenAI, a nonprofit AI research organization based in San Francisco and founded by several prominent technology investors who have pledged $1 billion 33. Real AI, an AI company based in Hong Kong and led by Jonathan Yan 34. Research Center for Brain-Inspired Intelligence (RCBII), a project of the Chinese Academy of Sciences 35. Sigma, led by Paul Rosenbloom of University of Southern California 36. SiMA, led by Dietmar Dietrich of Vienna University of Technology 37. SingularityNET, an open AI platform led by Ben Goertzel 38. SNePS, led by Stuart Shapiro at State University of New York at Buffalo 39. Soar, led by John Laird of University of Michigan and a spinoff company SoarTech 40. Susaro, an AI company based in the Cambridge, UK area and led by Richard Loosemore 41. Tencent AI Lab, the AI group of Tencent 42. Uber AI Labs, the AI research division of Uber 43. Vicarious, an AI company based in San Francisco 44. Victor, a project of 2AI, which is a subsidiary of Cifer Inc., a small US company 45. Whole Brain Architecture Initiative (WBAI), a nonprofit in Tokyo Many of the projects are interconnected. For example, AIDEUS lead Alexey Potapov is an advisor to SingularityNET; Animats contributors include Claes Strannegård (Alice in Wonderland), Joscha Bach (MicroPsi), and Bas Steunebrink (NNAISENSE); parts of DeSTIN have been used in CogPrime; DeepMind and OpenAI collaborate on AGI safety research; CogPrime and SingularityNET are led by the same person \n distinguishes between two types of open-source projects: \"classic\" open-source, in which code development is done \"out in the open\", and \"corporate\" open-source, in which code development is conducted by project insiders in a closed environment and then released openly. Goertzel (2017a) cites OpenAI as an example of corporate open-source (note: OpenAI is a nonprofit project, not at a for-profit institution); his CogPrime would be an example of classic open-source. The classic/corporate distinction can matter for ethics, risk, and policy by affecting who is able to influence AGI goals and safety features. However, which open-source projects are classic or corporate is beyond the scope of this paper. Code Not Available Code Available Open-Source (17) Open-Source (25) Code Access Restricted (3) \n 18 Engagement on safety: unspecified Size: medium AERA Main website: http ://www.ru.is/faculty/thorisson AERA is led by Kristinn Thórisson of Reykjavik University. AERA is an acronym for Autocatalytic Endogenous Reflective Architecture (Nivel et al. 2013) . The project aims \"to both understand the mind and build a practical AGI system\", and is currently being used \"to study machine understanding, teaching methodologies for artificial learners, even the development of ethical values\".19    Project lead Thórisson criticizes military AI in Icelandic Institute for Intelligent Machines (IIIM).20     Lead institutions: Reykjavik University Partner institutions: Icelandic Institute for Intelligent Machines (Iceland), Dalle Molle Institute for Artificial Intelligence Research (Switzerland) (per authors listed in Steunebrink et al. 2016) Type of institution: academic Open-source: no Military connection: no  Military connection: yes 17Lead country: USAPartner countries: noneStated goals: intellectualism  The main description of ACT-R on its website is exclusively about the intellectual research, with no mention of broader impacts. \n http://aideus.com AIDEUS is led by Alexey Potapov of ITMO University in Saint Petersburg and Sergey Rodionov of Aix Marseille Université. Potapov is also an advisor to SingularityNET. 23 It states a goal of the \"creation of a strong artificial intelligence\". Size: medium-small 19 http://www.ru.is/faculty/thorisson 20 http://www.iiim.is/ethics-policy/ 21 http://www.iiim.is/ethics-policy/3 22 http://www.iiim.is/ethics-policy/3 AIDEUS Main website: \n In Wonderland (AIW) Main website: https http://www.hutter1.net/ai/aixigentle.htm AIXI is led by Marcus Hutter of Australian National University. AIXI is based on a \"meta-algorithm\" that searches the space of algorithms to find the best one for AGI.29  Hutter proves that AIXI will find the most intelligent AI if given infinite computing power. While this is purely a theoretical result, it has led to approximate versions that are implemented in computer code. ://github.com/arnizamani/aiw and http://flov.gu.se/english/about/staff? languageId=100001&userId=xstrcl AIW is led by Claes Strannegård of Chalmers University of Technology in Sweden. A paper about AIW in Journal of Artificial General Intelligence describes it as being similar to NARS (Strannegård et al. 2017a) . A separate paper describes it as a prototype for implementing new ideas about \"bridging the gap between symbolic and sub-symbolic reasoning\" (Strannegård et al. 2017b) . https://github.com/ni1s/animats Animats is a small project developed for the First International Workshop on Architectures for Generality & Autonomy 33 and the 2017 AGI conference.34  The project is a collaboration between researchers at universities in Sweden and the United States and the Swiss company NNAISENSE.35  It seeks to build AGI based on animal intelligence.36     Lead institutions: Chalmers University of Technology Lead institutions: Chalmers University of Technology Partner institutions: University of Gothenburg, Harvard University, NNAISENSE Partner institutions: none Type of institution: academic, private corporation Type of institution: academic Open-source: yes 37 Open-source: yes 31 Military connection: no 38 Military connection: no 32 : small 23 https://singularitynet.io 24 http://aideus.com 25 http://aideus.com/research/research.html 26 http://aideus.com/research/publications.html 27 https://www.facebook.com/pg/Aideus-Strong-artificial-intelligence-455322977847194/about 28 http://aideus.com/community/community.html AIXI Main website: 30Lead institutions: Australian National University Partner institutions: none Type of institution: Academic Open-source: no Military connection: unspecified Lead country: Australia Partner countries: none Stated goals: unspecified Engagement on safety: unspecified Size: small 29 Goertzel (2014, p.25) 30 Veness et al. (2011) Alice Lead country: Sweden Partner countries: none Stated goals: unspecified Engagement on safety: unspecified Size: small 31 https://github.com/arnizamani/aiw 32 Reported funding is from the Swedish Research Council Animats Main website: Lead country: Sweden Partner countries: Switzerland, USA Stated goals: unspecified Engagement on safety: unspecified Size: small 33 http://cadia.ru.is/workshops/aga2017 34 Strannegård et al. (2017) 35 Strannegård et al. (2017b) 36 Strannegård et al. (2017b) 37 https://github.com/ni1s/animats 38 Reported funding is from the Swedish Research Council \n Baidu Research Main website: http ://research.baidu.com/learning-speak-via-interaction Baidu Research is an AI research group within Baidu. It has offices in Beijing, Shenzhen, and Sunnyvale, California.39  One page of its website states that its mission is \"to create general intelligence for machines\", 40 though this does not appear to be a major theme for the group. It has achieved success in \"zero-shot learning\" in language processing, in which the AI \"is able to understand unseen sentences\". 41 Some observers rate this as a significant breakthrough. Former Baidu Research Chief Scientist Andrew Ng says that Baidu takes safety and ethics seriously, and expresses his personal views that AI will help humanity and that \"fears about AI and killer robots are overblown\", 45 but this was not in the context of AGI. No direct discussion of safety by Baidu Research was found.Becca is led by Brandon Rohrer, currently at Facebook.46  According to its website, Becca \"is a general learning program for use in any robot or embodied system\"; it \"aspires to be a brain for any robot, doing anything\".47  Rohrer describes it as an open-source project, 48 and its source code is available on its website. Becca began while Rohrer was at Sandia National Laboratories, 50 but this connection appears to be inactive. Becca Main website: https://github.com/brohrer/becca Lead institutions: Becca Partner institutions: none Type of institution: none  No institutional home for Becca was found; it appears to be a spare-time project for Rohrer Open-source: yes 49 Military connection: unspecified  Size: medium-small 39 http://bdl.baidu.com/contact_b.html 42 Griffin (2017) 43 https://github.com/baidu 44 Gershgorn (2017) 45 Maddox (2016) 42 Lead institutions: Baidu Partner institutions: none Type of institution: Public Corporation Open-source: no  Baidu releases some work open-source, 43 but not its AGI Military connection: unspecified  Baidu receives AI funding from the Chinese government for \"computer vision, biometric identification, intellectual property rights, and human-computer interaction\". 44 Lead country: China Partner countries: USA Stated goals: unspecified Engagement on safety: unspecified  40 http://research.baidu.com/learning-speak-via-interaction 41 http://research.baidu.com/ai-agent-human-like-language-acquisition-virtual-environment Lead country: USA Partner countries: none Stated goals: unspecified Engagement on safety: unspecified Size: small 46 https://www.linkedin.com/in/brohrer 47 https://github.com/brohrer/becca 48 https://www.linkedin.com/in/brohrer 49 https://github.com/brohrer/becca 50 https://www.linkedin.com/in/brohrer \n Blue Brain Main website: http ://bluebrain.epfl.ch Blue Brain is a research project led by Henry Markram. It has been active since 2005. Its website states that its goal \"is to build biologically detailed digital reconstructions and simulations of the rodent, and ultimately the human brain\".51  Markram also founded the Human Brain Project, which shares research strategy with Blue Brain.52     Lead institution: École Polytechnique Fédérale de Lausanne Partner institutions: Hebrew University (Israel); Universidad Politecnica de Madrid and Consejo Superior de Investigaciones Científicas (Spanish National Research Council) (Spain); University of London (UK); and IBM, St. Elizabeth's Medical Center, University of Nevada-Reno, and Yale University (USA) 53 \n html China Brain Project Main website: none These applications are broadly humanitarian, though it is a more muted humanitarian than what is found for several other projects. found China Brain Project is a research project of the Chinese Academy of Sciences focused on basic and clinical neuroscience and brain-inspired computing. As of July 2016, the Chinese Academy of Sciences had announced the project and said it would launch soon, 57 but in August 2017 no project website was found. Project lead Mu-Ming Poo described the project in a 2016 article in the journal Neuron, stating that \"Learning from information processing mechanisms of the brain is clearly a promising way forward in building stronger and more general machine intelligence\" and \"The China Brain Project will focus its efforts on developing cognitive robotics as a platform for integrating braininspired computational models and devices\".58    The Chinese Academy of Sciences is a public institution under the Chinese government.59  Mu-Ming Poo lists the Chinese Natural Science Foundation and Ministry of Science and Technology as guiding organizations for the China Brain Project. http://www.cyc.com Cyc is led by Doug Lenat, who began Cyc in 1984. Cyc is a project of Cycorp, a corporation based in Austin, Texas that uses Cyc for consulting and other services to other corporations and government agencies. Lead institutions: Chinese Academy of Sciences Partner institutions: none Type of institution: government  Engagement on safety: unspecified Size: medium-large 51 http://bluebrain.epfl.ch/page-56882-en.html 52 http://bluebrain.epfl.ch/page-52741-en.html 53 http://bluebrain.epfl.ch/page-56897-en.html 54 http://bluebrain.epfl.ch/page-56897-en.html 55 https://github.com/BlueBrain 56 http://bluebrain.epfl.ch/page-125344-en.60 Open-source: no Military connection: unspecified Lead country: China Partner countries: none Stated goals: humanitarian, intellectual  Mu-Ming Poo describes the project's goals as \"understanding the neural basis of human cognition\" and \"reducing the societal burden of major brain disorders\" 61 Engagement on safety: unspecified Size: small 57 http://english.cas.cn/newsroom/news/201606/t20160617_164529.shtml 58 Poo et al. (2016) 59 http://english.cas.cn/about_us/introduction/201501/t20150114_135284.shtml 60 Poo et al. (2016) 61 Poo et al. (2016) CLARION Cyc Main website: \n none Stated goals: unspecified Engagement on safety: unspecified Size: medium 84 http://www.cyc.com/enterprise-solutions 85 http://www.cyc.com/about/company-profile 86 Goertzel 2014, p.16 87 Love (2014) 90 Military connection: yes  Cycorp received a $25 million contract to analyze terrorism for the US military. 91Lead country: USA Partner countries: 88 http://www.cyc.com/platform/researchcyc 89 http://www.cyc.com/platform 90 http://www.cyc.com/about/company-profile 91 http://www.cyc.com/about/media-coverage/computer-save-world; see also Deaton et al. (2005)   \n Deep SpatioTemporal Inference Network) Main website: http . It collaborates with OpenAI on long-term AI safety projects.98  It also participates independently from Google in the Partnership on AI to Benefit People & Society. ://web.eecs.utk.edu/~itamar/Papers/BICA2009.pdf and http://wiki.opencog.org/w/DeSTIN DeSTIN was initially developed by Itamar Arel and colleagues at the University of Tennessee. It is also being developed by the OpenCog open-source AI project. DeSTIN uses deep learning for pattern recognition. The OpenCog website states that OpenCog \"has adopted this academic project to prepare it for open-source release\".99  Goertzel (2014, p.17) notes that DeSTIN \"has been integrated into the CogPrime architecture… but is primarily being developed to serve as the center of its own AGI design\". Lead institution: University of Tennessee Partner institution: OpenCog Foundation Type of institution: academic, nonprofit Open-source: yes 100 Size: large 92 https://deepmind.com/blog/open-sourcing-deepmind-lab 93 https://github.com/deepmind 94  https://www.fpds.gov/ezsearch/fpdsportal?q=google+DEPARTMENT_FULL_NAME%3A\"DEPT+OF+DEFENSE\" 95 https://deepmind.com/blog/deepmind-office-canada-edmonton 96 https://deepmind.com/about 97 https://deepmind.com/blog/learning-through-human-feedback 98 https://blog.openai.com/deep-reinforcement-learning-from-human-preferencesDeSTIN (Military connection: unspecifiedLead country: USA Partner countries: none Stated goals: unspecified Engagement on safety: unspecified Size: small 99 http://wiki.opencog.org/w/DeSTIN 100 https://github.com/opencog/destin DSO- \n CA Main website: none found DSO-CA is a project of Gee Wah Ng and colleagues at DSO National Laboratories, which is Singapore's primary national defense research agency. It is \"a top-level cognitive architecture that models the information processing in the human brain\", with similarities to LIDA, CogPrime, and other AGI cognitive architectures. 101 DSO was \"corporatized\" in 1997 102 but is listed as a government agency on its LinkedIn page. Lead institution: DSO National Laboratories Partner institution: none Type of institution: government  103 Open-source: no Military connection: yes Lead country: Singapore Partner countries: none Stated goals: unspecified Engagement on safety: unspecified Size: small 101 Ng et al. (2017) 102 https://www.dso.org.sg/about/history 103 https://www.linkedin.com/company-beta/15618 FLOWERS ( \n FLOWing Epigenetic Robots and Systems) Main website: https ://flowers.inria.fr FLOWERS is led by Pierre-Yves Oudeyer of Inria (Institut National de Recherche en Informatique et en Automatique, or French National Institute for Research in Computer Science and Automation) and David Filliat of ENSTA ParisTech. The project \"studies mechanisms that can allow robots and humans to acquire autonomously and cumulatively repertoires of novel skills over extended periods of time\". 104 https://www.goodai.com GoodAI is a privately held corporation led by computer game entrepreneur Marek Rosa. It is based in Prague. It is funded by Rosa, who has invested at least $10 million in it. Rosa states that \"GoodAI is building towards my lifelong dream to create general artificial intelligence. I've been focused on this goal since I was 15 years old\".109    GoodAI lists several partner organizations, including one, SlovakStartup, based in Slovakia.110   However, none of the partners are counted in this study because they do not contribute AGI R&D. Lead institutions: Inria and ENSTA ParisTech Lead institution: GoodAI Partner institutions: none Partner institutions: none Type of institution: academic, government   Inria is a government research institute; ENSTA ParisTech is a public college Open-source: yes 105 Military connection: unspecified 106 Lead country: FrancePartner countries: noneStated goals: intellectualist  The project website focuses exclusively on intellectual aspects of its AI research and also cognitive science alongside AI as one of its two research strands. 107Engagement on safety: active  FLOWERS has explored safety in the context of human-robot interactions. 108Size: medium-smallGoodAIMain website: \n 113 Engagement on safety: active  GoodAI reports having a dedicated AI safety team 114 and cites Nick Bostrom's Superintelligence as a research inspiration on AI safety. 115Main website: https://numenta.com HTM is developed by the Numenta corporation of Redwood City, California and an open-source community that the corporation hosts. HTM is led by Jeffrey Hawkins, who previously founded Palm Computing. HTM is based on a model of the human neocortex. Their website states, \"We believe that understanding how the neocortex works is the fastest path to machine intelligence… Numenta is far ahead of any other team in this effort to create true machine intelligence\".116     HTM (Hierarchical Temporal Memory) Lead institution: Numenta Partner institutions: none Type of institution: private corporation Open-source: yes 117  Numenta offers a fee-based commercial license and an open-source license. 118 Military connection: unspecified  HTM was used in a 2008 Air Force Institute of Technology student thesis (Bonhoff 2008) Lead country: USA Partner countries: none Stated goals: humanitarianism, intellectualist  The Numenta corporate website lists two agendas: Size: medium \n Engagement on safety: unspecified Size: small Leabra Main website: https Main website: http://www.humanbrainproject.eu HBP is a project for neuroscience research and brain simulation. It is sponsored by the European Commission, with a total of $1 billion committed over ten years beginning 2013.122  Initially led by Henry Makram, it was reorganized following extended criticism (Theil 2015) . It is based at École Polytechnique Fédérale de Lausanne, with collaborating institutions from around Europe and Israel.123   It hosts platforms for brain simulation, neuromorphic computing, and neurorobotics.124  Markram also founded Blue Brain, which shares research strategy with HBP. 125 http://cll.stanford.edu/research/ongoing/icarus Icarus is led by Pat Langley of Stanford University. Another active contributor is Dongkyu Choi of University of Kansas, a former Langley Ph.D. student. Icarus is a cognitive architecture project similar to ACT-R and Soar, emphasizing perception and action in physical environments.132  Its website is out of date but several recent papers have been published. 133 ://grey.colorado.edu/emergent/index.php/Leabra Leabra is led by Randall O'Reilly of University of Colorado. It is a cognitive architecture project emphasizing modeling neural activity. A recent paper states that Leabra is \"a long-term effort to produce an internally consistent theory of the neural basis of human cognition\", and that \"More than perhaps any other proposed cognitive architecture, Leabra is based directly on the underlying biology of the brain, with a set of biologically realistic neural processing mechanisms at its core\".135  It was briefly connected to ACT-R via the SAL project. Icarus Main website: Lead institution: Stanford University Reason for consideration: A large-scale brain research project, similar to Blue Brain Partner institutions: University of Kansas Lead institution: University of Colorado Lead institutions: École Polytechnique Fédérale de Lausanne Type of institution: academic Partner institutions: Partner institutions: 116 listed on the HBP website 126 Open-source: no Type of institution: academic Military connection: yes 134 Open-source: restricted Lead country: USA  Obtaining an account requires a request and sharing a copy of one's passport. 127 Partner countries: none Military connection: no Stated goals: intellectualist  HBP policy forbids military applications 128  Choi and Langley (2017) write that \"our main goal\" is \"achieving broad coverage of cognition Lead country: Switzerland functions\" in \"the construction of intelligent agents\". Partner countries: Austria, Belgium, Denmark, Finland, France, Germany, Greece, Hungary, Israel, Italy, Netherlands, Norway, Portugal, Slovenia, Spain, Sweden, Turkey, and United Kingdom Stated goals: animal welfare, humanitarian, intellectualist  HBP pursues brain simulation to \"reduce the need for animal experiments\" and \"study diseases\". 129 It also lists \"understanding cognition\" as a core theme. 130 Engagement on safety: unspecified  HBP has an ethics program focused on research procedure, not societal impacts. 131 Size: large 122 http://www.humanbrainproject.eu/en/science/overview/ 123 http://www.humanbrainproject.eu/en/open-ethical-engaged/contributors/partners 124 http://www.humanbrainproject.eu/en/brain-simulation/brain-simulation-platform; http://www.humanbrainproject.eu/en/silicon-brains; http://www.humanbrainproject.eu/en/robots 125 http://bluebrain.epfl.ch/page-52741-en.html 126 http://www.humanbrainproject.eu/en/open-ethical-engaged/contributors/partners 127 http://www.humanbrainproject.eu/en/silicon-brains/neuromorphic-computing-platform 128 https://nip.humanbrainproject.eu/documentation, https://nip.humanbrainproject.eu/documentation 129 http://www.humanbrainproject.eu/en/brain-simulation 130 http://www.humanbrainproject.eu/en/understanding-cognition 131 https://www.humanbrainproject.eu/en/open-ethical-engaged/ethics/ethics-management \n none Type of institution: academic Open-source: yes 136 Military connection: yes 137 Lead country: USA Partner countries: none Stated goals: intellectualist Engagement on safety: unspecified Size: medium-small LIDA (Learning Intelligent Distribution Agent) Main website: http ://ccrg.cs.memphis.edu LIDA is led by Stan Franklin of University of Memphis. It is based on Bernard Baars's Global Workspace Theory, \"integrating various forms of memory and intelligent processing in a single processing loop\"(Goertzel 2014, p.24).Goertzel (2014, p.24)  states that LIDA has good grounding in neuroscience but is only capable at \"lower level\" intelligence, not more advanced thought. LIDA is supported by the US Office of Naval Research; a simpler version called IDA is being used to automate \"the decision-making process for assigning sailors to their new posts\".138    The LIDA website says that it seeks \"a full cognitive model of how minds work\" and focuses predominantly on intellectual research aims. 140  A paper by LIDA researchers Tamas Madl and Stan Franklin states that robots need ethics \"to constrain them to actions beneficial to humans\". 141  Franklin hints at support for moral standing for AGI, noting the study of \"synthetic emotions\" in AI, which could suggest that an AGI \"should be granted moral status\". 142Engagement on safety: active  The Madl and Franklin paper addresses AGI safety challenges like the subtlety of defining human ethics with the precision needed for programming. 143  Franklin has also collaborated with AI ethicists Colin Allen and Wendell Wallach on the challenge of getting AGIs to make correct moral decisions.144    Maluuba is an AI company based in Montreal, recently acquired by Microsoft. Maluuba writes: \"our vision has been to solve artificial general intelligence by creating literate machines that could think, reason and communicate like humans\".145    Maluuba writes that \"understanding human language is extremely complex and is ultimately the holy grail in the field of Artificial Intelligence\", and that they aim \"to solve fundamental problems in language understanding, with the vision of creating a truly literate machine\". 147  Maluuba VP of Product Mo Musbah says that Maluuba wants general AI so that \"it can scale in terms of how it applies in an AI fashion across different industries\".148  Microsoft is a founding partner of the Partnership on AI to Benefit People & Society, which has humanitarian goals, 149 but this does not appear to have transferred to Maluuba's goals. Maluuba describes safety as an important research challenge. 150 Maluuba researcher Harm van Seijen writes that \"I think such discussions [about AI safety] are good, although we should be cautious of fear mongering.\" 151 Microsoft is also a founding partner of the Partnership on AI to Benefit People & Society, which expresses concern about AI safety. 152 No direct safety activity by Maluuba was identified. MicroPsi is led by Joscha Bach of the Harvard Program for Evolutionary Dynamics. Bach's mission is reportedly \"to build a model of the mind is the bedrock research in the creation of Strong AI, i.e. cognition on par with that of a human being\". 153 Goertzel (2014, p.24) states that MicroPsi has good grounding in neuroscience and is only capable at lower level intelligence. A recent paper on MircoPsi says that it is \"an architecture for Artificial General Intelligence, based on a framework for creating and simulating cognitive agents\", which began in 2003. 154 Main website: https://www.microsoft.com/en-us/research/lab/microsoft-research-ai MSR AI is an AI \"research and incubation hub\" at Microsoft announced in July 2017. Maluuba MicroPsi Microsoft Research AI (MSR AI) Main website: http://www.maluuba.com Main website: http://cognitive-ai.com Lead institution: Microsoft Partner institutions: none Lead institutions: University of Memphis Lead institutions: Harvard University Type of institution: public corporation Partner institutions: none Partner institutions: none Open-source: yes 146 Type of institution: academic Type of institution: academic Military connection: unspecified Open-source: restricted 139 Open-source: yes 155 Lead country: Canada  Registration is required to download code; commercial use requires a commercial license Military connection: unspecified Partner countries: USA Military connection: yes Lead country: USA Stated goals: intellectualist, profit Lead country: USA Partner countries: none Stated goals: humanitarian, intellectualist  Size: medium  Size: medium Partner countries: none Stated goals: intellectualist  The MicroPsi website says that \"MicroPsi is a small step towards understanding how the mind works\". 156 Engagement on safety: unspecified  Engagement on safety: moderate Size: small 145 http://www.maluuba.com/blog/2017/1/13/maluuba-microsoft 138 http://ccrg.cs.memphis.edu 146 https://github.com/Maluuba 139 http://ccrg.cs.memphis.edu/framework.html 147 http://www.maluuba.com/blog/2017/1/13/maluuba-microsoft 140 http://ccrg.cs.memphis.edu 148 Townsend (2016) 141 Madl and Franklin (2015) 149 https://www.partnershiponai.org/tenets 153 http://bigthink.com/experts/joscha-bach 142 Wallach et al. (2011, p.181) 150 http://www.maluuba.com/blog/2017/3/14/the-next-challenges-for-reinforcement-learning 154 Bach (2015) 143 Madl and Franklin (2015) 151 Townsend (2016) 155 https://github.com/joschabach/micropsi2 144 Wallach et al. (2010) 152 https://www.partnershiponai.org/tenets 156 http://cognitive-ai.com \n 161 Lead institutions: Microsoft Partner institutions: none Type of institution: public corporation Open-source: no Military connection: unspecified  Microsoft has a Military Affairs program, 162 but its link to MSR AI is unclear. The MSR AI website states that it aims \"to solve some of the toughest challenges in AI\" and \"probe the foundational principles of intelligence\".164 Engagement on safety: unspecified  The MSR AI group on Aerial Informatics and Robotics has extensive attention to safety, but this is for the narrow context of aircraft, not for AGI. 165 none MLECOG is a cognitive architecture project led by Janusz Starzyk of Ohio University. A paper on MLECOG describes it as similar to NARS and Soar. 166 MLECOG is an acronym for Motivated Learning Embodied Cognitive Architecture. https://sites.google.com/site/narswang NARS is an AGI research project led by Pei Wang of Temple University. NARS is an acronym for Non-Axiomatic Reasoning System, in reference to the AI being based on tentative experience and not axiomatic logic, consistent with its \"assumption of insufficient knowledge and resources\". 167 In a 2011 interview, Wang suggests that NARS may achieve human-level AI by 2021. 168 NARS Lead institution: Ohio University Main website: Lead institutions: Temple University Partner institutions: none Partner institutions: none Type of institution: academic Type of institution: academic Open-source: no Open-source: yes 169 Military connection: unspecified Lead country: USA Partner countries: none Stated goals: unspecified Lead country: USA Engagement on safety: unspecified Partner countries: none Size: small Stated goals: humanitarian, intellectualist  Microsoft CEO Satya Nadella states broadly humanitarian goals, such as \"AI must be designed to assist humanity\". 163  Size: medium-large 157 https://blogs.microsoft.com/blog/2017/07/12/microsofts-role-intersection-ai-people-society 158 https://www.microsoft.com/en-us/research/lab/microsoft-research-ai 159 Etherington (2017) 166 Starzyk and Graham (2015) 160 https://blogs.microsoft.com/blog/2017/07/12/microsofts-role-intersection-ai-people-society 161 Architecht (no date) 162 https://military.microsoft.com/about 163 Nadella (2016) 164 https://www.microsoft.com/en-us/research/lab/microsoft-research-ai 165 https://www.microsoft.com/en-us/research/group/air MLECOG Main website: Military connection: unspecified Lead country: USA Partner countries: none Stated goals: humanitarian, intellectualist  \n 179 Engagement on safety: unspecified 173    Nigel is the AGI project of Kimera, an AI corporation based in Portland, Oregon. Kimera was founded in 2005 by Mounir Shita and Nicholas Gilman. It styles itself as \"The AGI Company\".174  In 2016, Kimera unveiled Nigel, which it claims is \"the first commercially deployable artificial general intelligence technology\". 175 However, as of 2016, little about Nigel is publicly available and critics are skeptical about the AGI claim. 176 Nigel has been described as a personal assistant bot similar to Apple's SIRI and Amazon's Alexa.177  Kimera also envisions Nigel being used for a variety of online activities, \"bringing about a greater transformation in global business than even the internet\", and transforming \"the internet from a passive system of interconnections into a proactive, intelligent global network\". https://nnaisense.com NNAISENSE is a private company based in Lugano, Switzerland. Several of its team members have ties to the Dalle Molle Institute for Artificial Intelligence (IDSIA, a Swiss nonprofit research institute), including co-founder and Chief Scientist Jürgen Schmidhuber. Its website states that it seeks \"to build large-scale neural network solutions for superhuman perception and intelligent automation, with the ultimate goal of marketing general-purpose Artificial Intelligences.\" 180 Schmidhuber is described as a \"consummate academic\" who founded the company to prevent other companies from poaching his research talent; NNAISENSE reportedly \"chooses projects based on whether they'll benefit the machine's knowledge, not which will bring in the highest fees\". 181  The NNAISENSE website states \"the ultimate goal of marketing\" AGI (emphasis added). NNAISENSE Main website: Lead institutions: NNAISENSE Partner institutions: none Type of institution: private company Open-source: no Military connection: unspecified Lead country: Switzerland Partner countries: none Stated goals: intellectualism, profit  Size: medium Size: medium-small 167 https://sites.google.com/site/narswang/home/nars-introduction 168 Goertzel (2011) 174 http://kimera.ai 175 http://kimera.ai 176 Boyle (2016); Jee (2016) 177 Boyle (2016) 178 http://kimera.ai/nigel 179 http://kimera.ai/company 169  https://github.com/opennars 170 https://sites.google.com/site/narswang/EBook/Chapter5/section-5-5-education 171 https://sites.google.com/site/narswang/EBook/Chapter5/section-5-5-education 172 https://sites.google.com/site/narswang173  Wang (2012) Nigel Main website: http://kimera.ai 178 Lead institution: Kimera Partner institutions: none Type of institution: private corporation Open-source: no Military connection: unspecified Lead country: USA Partner countries: none Stated goals: humanitarianism  Kimera presents a humanitarian vision for AGI, writing that \"Artificial General Intelligence has the power to solve some -or all -of humanity's biggest problems, such as curing cancer or eliminating global poverty.\" \n 182 Engagement on safety: unspecified OpenAI is a nonprofit AI research organization founded by several prominent technology investors. It is based in San Francisco. Its funders have pledged $1 billion to the project. Its website states: \"Artificial general intelligence (AGI) will be the most significant technology ever created by humans. OpenAI's mission is to build safe AGI, and ensure AGI's benefits are as widely and evenly distributed as possible.\" 183 It is part of the Partnership on AI to Benefit People & Society.184     OpenAI Main website: https://openai.com Size: medium-small 180 https://nnaisense.com 181 Webb (2017) 182 https://nnaisense.com \n Lead institution: OpenAI Partner institutions: none Type of institution: nonprofit Open-source: yes 185 Military connection: unspecified Lead country: USA Partner countries: none Stated goals: humanitarianism  OpenAI seeks that AGI \"leads to a good outcome for humans,\" 186 and that \"AGI's benefits are as widely and evenly distributed as possible.\" 187 Engagement on safety: active  Safe AGI is part of OpenAI's mission. While it releases much of its work openly, its website states that \"in the long term, we expect to create formal processes for keeping technologies private when there are safety concerns\".188  It also collaborates with DeepMind on long-term AI safety.189     Size: large 183 https://openai.com/about 184 https://www.partnershiponai.org/partners 185 https://github.com/openai 186 https://openai.com/jobs 187 https://openai.com/about 188 https://openai.com/about 189  https://blog.openai.com/deep-reinforcement-learning-from-human- \n preferences Real AI Main website: http ://realai.org Real AI is a private company in Hong Kong led by Jonathan Yan. It is a single member company.190  Its mission is \"to ensure that humanity has a bright future with safe AGI\".191  It works on strategy for safe AGI and technical research in deep learning, the latter on the premise that deep learning can scale up to AGI.192  Its website states that \"We align ourselves with effective altruism and aim to benefit others as much as possible\".193     Lead institution: Real AI Partner institutions: none Type of institution: private corporation Open-source: no Military connection: unspecified Lead country: China Partner countries: none Stated goals: humanitarian Engagement on safety: active  Real AI has a dedicated page surveying ideas about AGI safety 194 and an extended discussion of its own thinking. 195 \n Size: small Research Center for Brain-Inspired Intelligence (RCBII) Main website: http ://bii.ia.ac.cn RCBII is a \"long term strategic scientific program proposed by Institute of Automation, Chinese Academy of Sciences\". 196 The group is based in Beijing. 197 RCBII does research in fundamental neuroscience, brain simulation, and AI. It states that \"Brain-inspired Intelligence is the grand challenge for achieving Human-level Artificial Intelligence\". 198 Main website: http://cogarch.ict.usc.edu Sigma is led by Paul Rosenbloom of University of Southern California. It has a publication record dating to <PHONE>  and won awards at the 2011 and 2012 AGI conferences.201  Rosenbloom was previously a co-PI of Soar.202    The Sigma website says its goal is \"to develop a sufficiently efficient, functionally elegant, generically cognitive, grand unified, cognitive architecture in support of virtual humans (and hopefully intelligent agents/robots -and even a new form of unified theory of human cognitionas well).\" 205  Rosenbloom also hints at cosmist views in a 2013 interview, stating \"I see no real long-term choice but to define, and take, the ethical high ground, even if it opens up the possibility that we are eventually superseded -or blended out of pure existence -in some essential manner.\" Lead institution: University of Southern California Lead institution: Chinese Academy of Sciences Partner institutions: none Partner institutions: none Type of institution: academic Type of institution: government  The Chinese Academy of Sciences is a public institution under the Chinese government 199 Open-source: yes 203 Open-source: no Military connection: yes 204 Military connection: unspecified Lead country: USA Lead country: China Partner countries: none Stated goals: intellectualist Partner countries: none  Stated goals: intellectualist  The RCBII website only lists intellectual motivations, stating \"The efforts on Brain-inspired Intelligence focus on understanding and simulating the cognitive brain at multiple scales as well as its applications to brain-inspired intelligent systems.\" No social or ethical aspects of these applications are discussed. Engagement on safety: unspecified Size: medium-small 196 http://bii.ia.ac.cn/about.htm 197 http://english.ia.cas.cn/au/fu/ 198 http://bii.ia.ac.cn/about.htm 199 http://english.cas.cn/about_us/introduction/201501/t20150114_135284.shtml Sigma \n 206 Engagement on safety: unspecified  In a 2013, interview, Rosenbloom hints at being dismissive, questioning \"whether superhuman general intelligence is even possible\", but also explores some consequences if it is possible, all while noting his lack of \"any particular expertise\" on the matter.207  No indications of safety work for Sigma were found. http://sima.ict.tuwien.ac.at/description SiMA is a project led by Dietmar Dietrich of Vienna University of Technology. SiMA is an acronym for Simulation of the Mental Apparatus & Applications. The project aims \"to develop a broad humanlike intelligent system that is able to cope with complex and dynamic problems rather than with narrowly and well-defined domains\".208  It includes extensive attention to psychoanalysis, especially Freud and other German-language scholars. It was started by Dietrich in 1999. 209   Lead institutions: Vienna University of Technology Partner institutions: none Type of institution: academic Open-source: yes 210 Size: medium 200 http://cogarch.ict.usc.edu/publications-new 201 http://cs.usc.edu/~rosenblo 202 http://cs.usc.edu/~rosenblo 203 https://bitbucket.org/sigma-development/sigma-release/wiki/Home 204 Funding from the US Army, Air Force Office of Scientific Research, and Office of Naval Research is reported in Rosenbloom (2013) 205 http://cogarch.ict.usc.edu 206 https://intelligence.org/2013/09/25/paul-rosenbloom-interview 207 https://intelligence.org/2013/09/25/paul-rosenbloom-interviewSiMAMain website:  The open-source portion of the project appears to be out of dateMilitary connection: unspecifiedLead country: AustriaPartner countries: none  A project document states collaborators in Canada, Portugal, South Africa, and Spain, but details of these collaborations could not be identified. 211   \n Partner institutions: none Type of institution: academic Open-source: yes 221 Military connection: yes 222 Lead country: USA Partner countries: none Stated goals: intellectualist Engagement on safety: unspecified Size: medium 219 Main website: http://www.cse.buffalo.edu/sneps SNePS is led by Stuart Shapiro at State University of New York at Buffalo, with a publication record dating to 1969.219  According to its website, its long-term goal is \"to understand the nature of intelligent cognitive processes by developing and experimenting with computational cognitive agents that are able to use and understand natural language, reason, act, and solve problems in a wide variety of domains\". 220Reason for consideration: Listed in the AGI review paper Goertzel (2014) Lead institutions: State University of New York at Buffalo http://www.cse.buffalo.edu/sneps/Bibliography 220 http://www.cse.buffalo.edu/sneps 221 https://github.com/SNePS/CSNePS, https://www.cse.buffalo.edu/sneps/Downloads 222 US Army Research Office funding is reported in recent papers including Shapiro and Schlegel (2016). Main website: http://soar.eecs.umich.edu and https://soartech.com Soar is led by John Laird of University of Michigan and a spinoff corporation SoarTech, also based in Ann Arbor, Michigan. Laird and colleagues began Soar in 1981. 223 SOAR is an acronym for State, Operator Apply Result. Pace University, Pennsylvania State University, University of Portland, and University of Southern California in the United States, University of Portsmouth in the United Kingdom, and Bar Ilan University and Cogniteam (a privately held company) in Israel. 224 SoarTech lists customers including research laboratories of the US Air Force, Army, and Navy, and the US Department of Transportation.The Soar website describes it as an investigation into \"an approximation of complete rationality\" aimed at having \"all of the primitive capabilities necessary to realize the complete suite of cognitive capabilities used by humans\". Soar Lead institution: University of Michigan, SoarTech Partner institutions: Type of institution: academic, private corporation Open-source: yes 225 Military connection: yes  226 Lead country: USA Partner countries: Israel, UK Stated goals: intellectualist  \n 227 Engagement on safety: unspecified Size: medium -large 223 http://ai.eecs.umich.edu/people/laird 224 https://soar.eecs.umich.edu/groups 225 https://github.com/SoarGroup, https://soar.eecs.umich.edu/Downloads 226 http://soartech.com/about 227 http://soar.eecs.umich.eduSusaroMain website: http://www.susaro.com Susaro is an AI corporation based in the Cambridge, UK area. Its website states that it is in stealth mode, and that it is \"designing the world's most advanced artificial general intelligence systems\" (emphasis original) using an approach that \"is a radical departure from conventional AI\".228  The Susaro website does not list personnel, but external websites indicate that it is led by AGI researcher Richard Loosemore. The Susaro website states that it aims to advance \"human and planetary welfare… without making humans redundant\".231     229 Lead institution: Susaro Partner institutions: none Type of institution: private corporation Open-source: no  Susaro has a GitHub page with no content 230 Military connection: unspecified Lead country: UK Partner countries: none Stated goals: ecocentric, humanitarian  \n Engagement on safety: active  The Susaro website states that \"the systems we build will have an unprecedented degree of safety built into them… making it virtually impossible for them to become unfriendly\". \n 232 Size: unspecified Tencent AI Lab (TAIL) Main website: http ://ai.tencent.com/ailab TAIL is the AI group of Tencent, the Shenzhen-based Chinese technology company. Its website lists several research areas, one of which is machine learning, which it says includes \"general AI\".233  TAIL director Tong Zhang writes that TAIL \"not only advances the state of art in artificial general intelligence, but also supports company products\".234    Its website states that \"Tencent will open-source its AI solutions in the areas of image, voice, security to its partners through Tencent Cloud\", but it does not state that its AGI research is opensource. Lead institution: Tencent Partner institutions: none Type of institution: public corporation  235 Open-source: no  Tencent releases some work open-source, 236 but not its AGI Military connection: unspecified Lead country: China Partner countries: USA  TAIL recently opened an office in Seattle. 237 Stated goals: \n unspecified Engagement on safety: unspecified Size: medium -small 233 http://ai.tencent.com/ailab 234 http://tongzhang-ml.org/research.html 235 http://ai.tencent.com/ailab 236 https://github.com/Tencent 237 Mannes (2017) Main website: https://www.uber.com/info/ailabs UAIL is the AI research division of Uber. UAIL began in 2016 with the acquisition of Geometric Intelligence (Temperton 2016), a private company founded in 2014 by Gary Marcus, Kenneth Stanley, and Zoubin Ghahramani in 2014 with incubation support from NYU. 238 Geometric Intelligence was based on Marcus's ideas for AGI, especially how to \"learn with less training data\" than is needed for deep learning Uber AI Labs (UAIL) \n USA Partner countries: UK Stated goals: humanitarian  The UAIL website states that it seeks \"to improve the lives of millions of people worldwide\".241 Engagement on safety: unspecified  No discussion of UAIL safety was found, except about the safety of Uber vehicles. 242  Marcus has indicated support for AI ethics research as long as it is understood that advanced AGI is not imminent. 243 https://www.vicarious.com Vicarious is a privately held AI corporation founded in 2010 by Scott Phoenix and Dileep George and based in San Francisco. It has raised tens of millions of dollars in investments from several prominent investors. Vicarious Main website: Size: medium UberPartner institutions: noneType of institution: private corporationOpen-source: no  Uber does have some open-source AI, 239 but this does not appear to include its AGI Military connection: unspecified  Uber does not appear to have any defense contracts. 240Lead country: \n Engagement on safety: moderate Vicarious is a Flexible Purpose Corporation, reportedly so that it can \"pursue the maximization of social benefit as opposed to profit\".249  Scott Phoenix says that Vicarious aims to build AI \"to help humanity thrive\".250     Phoenix says that AI safety may be needed \"At some time in the future… but the research is at a really early stage now,\" 251 and that it will not be difficult because AI will probably be \"smart enough to figure out what it was that we wanted it to do.\" 252VictorMain website: http://2ai.org/victor Victor is the main project of 2AI, which is a subsidiary of the private company Cifer Inc. 253 2AI is led by Timothy Barber and Mark Changizi and lists addresses in Boise, Idaho and the US Virgin Islands. 254 2AI describes Victor as an AGI project.255  Its website states that \"we believe the future of AI will ultimately hinge upon its capacity for competitive interaction\". Its website states that \"2AI is a strong advocate of solutions for protecting and preserving aquatic ecosystems, particularly reefs and seamounts, which are crucial nexus points of Earth's biomass and biodiversity\". The project website states that AI catastrophe scenarios are \"crazy talk\" because AI will need humans to maintain the physical devices it exists on and thus will act to ensure humans maintain and expand these devices.258 Size: small AGINAO was a project of Wojciech Skaba of Gdansk, Poland. It was active during 2011-2013, 268 and shows no activity more recently. None Alibaba is active in AI, but no indications of AGI were found. Alibaba is a major computing technology company Reason for exclusion: No indications of AGI projects were found Amazon Main website: https://aws.amazon.com/amazon-ai Amazon has an AI group within its Amazon Web Services (AWS) division, but it does not appear to work on AGI. Amazon has donated AWS resources to OpenAI. Reason for consideration: An AGI R&D project Reason for exclusion: Apparently inactive Alibaba Main website: Reason for consideration: Size: medium-large 244 Cutler (2014); High (2016) 245 https://www.vicarious.com/research.html 246 High (2016) 247 TWiStartups (2016) 248 https://github.com/vicariousinc 249 High (2016) 250 High (2016) 251 Best (2016) 252 TWiStartups (2016) 247 Lead institution: Vicarious Partner institutions: none Type of institution: private corporation Open-source: yes 248 Military connection: unspecified Lead country: USA Partner countries: none Stated goals: humanitarianism  256 Lead institutions: Cifer Partner institutions: none Type of institution: private corporation Open-source: no Military connection: unspecified Lead country: USA Partner countries: none Stated goals: ecocentric  257 Engagement on safety: dismissive  269 Reason for consideration: Amazon is a major computing technology company Reason for exclusion: No indications of AGI projects were \n found Apple Main website: https ://machinelearning.apple.com Apple has an AI group that does not appear to work on AGI. However, Apple has a reputation for secrecy and has only a minimal website. Apple is said to have less capable AI than companies like Google and Microsoft because Apple has stricter privacy rules, denying itself the data used to train AI.270  Likewise, at least some of its AI research may be oriented towards learning from limited data or synthetic data.271  Its recent AI company acquisitions are for narrow AI.272  While it may be possible that Apple is working on AGI, no indications of this were found in online searches. No indications of AGI projects were found268  For example, Skaba (2012a; 2012b) ; http://aginao.com/page2.php 269 https://blog.openai.com/infrastructure-for-deep-learning270  Vanian (2017a)271  Vanian (2017b)272   Tamturk (2017)  PAGI World is a project led by John Licato of University of South Florida and based at Rensselaer Polytechnic Institute, where Licato was a Ph.D. student. PAGI world is \"a simulation environment written in Unity 2D which allows AI and AGI researchers to test out their ideas\".325    Reason for consideration: Apple is a major computing technology company Reason for exclusion: \n\t\t\t Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t See e.g.Legg (2008, p.125);; http://www.humanobs.org. \n\t\t\t Baum (2017b)  critiques coherent extrapolated volition proposals for focusing only on human ethics and not also on the ethical views that may be held by other biological species, by the AI itself, or by other entities. \n\t\t\t The model speaks in terms of AI in general, of which AGI is just one type, alongside other types of AI that could also recursively self-improve. This distinction is not crucial for the present paper. \n\t\t\t The collective action literature specifically finds that smaller groups are often more successful at cooperating when close interactions reduce free-riding and the costs of transactions and compliance monitoring, while larger groups are often more \n\t\t\t Thanks go to Matthijs Maas for research assistance on characterizing some projects.6  Two additional journals, Biologically Inspired Cognitive Architectures and Advances in Cognitive Systems, were considered due to their AGI-related content, but they were not scanned in their entirety due to a lack of AGI projects reported in their articles. \n\t\t\t  Similarly, Baum (2017a)  observes synergies between groups interested in societal impacts of near-term and long-term AI, such that growing interest in near-term AI could be leveraged to advance policy and other efforts for improving outcomes of long-term AI, including AGI. \n\t\t\t Also in Russia is Max Talanov of Kazan State University, who has contributed to NARS but not enough to be coded as a partner (e.g.,). \n\t\t\t This quote is from cryptographer Martin Hellman, himself an evidently conscientious person. The quote refers to an episode in 1975 when Hellman debated whether to go public with information about breaking a code despite warnings from the US National Security Agency that doing so would be harmful. Hellman did go public with it, and in retrospect he concluded that it was the right thing to do, but that he had the wrong motivation for doing so(Hellman and Hellman 2016,  p.48-50). \n\t\t\t Similarly, the \"safety-moderate\" project Vicarious states a belief that AGI safety may not be difficult, prompting them to be not (yet) active on safety. \n\t\t\t IBM was also carefully searched. No AGI R&D was found at IBM, but IBM does have a project doing hardware development related to AGI, the Cognitive Computing Project, which is included in Appendix 2.12  It is possible that some of them have AGI projects that are not publicly acknowledged. Apple in particular has a reputation for secrecy, making it a relatively likely host of an unacknowledged AGI project. \n\t\t\t http://act-r.psy.cmu.edu/about 14 http://act-r.psy.cmu.edu/people 15 For example, Lee et al. (2017)  has lead author Hee Seung Lee of Yonsei University, who is not listed at http://actr.psy.cmu.edu/people. 16 http://act-r.psy.cmu.edu/software 17 US Office of Naval Research funding is reported in Zhang et al. (2016) . 18 http://act-r.psy.cmu.edu/about \n\t\t\t https://flowers.inria.fr 105 https://flowers.inria.fr/software 106 Funding reported in recent publications comes mainly from government science foundations 107 https://flowers.inria.fr108 Oudeyer et al. (2011)    \n\t\t\t https://www.goodai.com/about 110 https://www.goodai.com/partners 111 https://github.com/GoodAI 112 https://www.goodai.com/about 113 https://www.goodai.com/about 114 https://www.goodai.com/about 115 https://www.goodai.com/research-inspirations \n\t\t\t Goertzel (2014) ; Choi and Langley (2017)  133 See e.g. Menager and Choi (2016) ; Choi and Langley (2017) ; Langley (2017)134  The Icarus group reports funding from the US Office of Naval Research, Navy Research Lab, and DARPA in Choi and Langley (2017) . \n\t\t\t O'Reilly et al. (2016) 136 https://grey.colorado.edu/emergent/index.php/Main_Page 137 The Leabra group reports funding from the US Office of Naval Research and Army Research Lab at https://grey.colorado.edu/CompCogNeuro/index.php/CCNLab/funding \n\t\t\t http://realai.org/about/admin 191 http://realai.org/about 192 http://realai.org/prosaic 193 http://realai.org/about 194 http://realai.org/safety 195 http://realai.org/blog/towards-safe-and-beneficial-intelligence \n\t\t\t All info and quotes are from http://www.susaro.com 229 https://www.linkedin.com/in/richard-loosemore-47a2164, https://www.researchgate.net/profile/Richard_Loosemore, https://cofounderslab.com/profile/richard-loosemore 230 https://github.com/susaroltd 231 All info and quotes are from http://www.susaro.com 232 All info and quotes are from http://www.susaro.com \n\t\t\t https://www.nyu.edu/about/news-publications/news/2016/december/nyu-incubated-start-up-geometric-intelligenceacquired-by-uber.html 239 https://github.com/uber/tensorflow 240 https://www.fpds.gov/ezsearch/fpdsportal?q=uber+DEPARTMENT_FULL_NAME%3A\"DEPT+OF+DEFENSE\" 241 https://www.uber.com/info/ailabs242  Chamberlain (2016) 243 https://techcrunch.com/2017/04/01/discussing-the-limits-of-artificial-intelligence \n\t\t\t http://www.ccbi.cmu.edu/projects_4caps.html 267 http://www.ccbi.cmu.edu/projects_4caps.html \n\t\t\t http://information.xmu.edu.cn/en/?mod=departments&id=31 274 https://www.braininitiative.nih.gov/about/index.htm 275 https://www.braininitiative.nih.gov \n\t\t\t http://cst.fee.unicamp.br 291 http://www.comirit.com 292 http://www.cse.msu.edu/~weng/research/LM.html293   Weng et al. (1999)   \n\t\t\t http://www.damer.com 307 https://www.cse.buffalo.edu/~shapiro/Papers 308Shapiro and Bona (2010, p. 307)    \n\t\t\t http://mason.gmu.edu/~asamsono/bica.html 310 http://people.idsia.ch/~juergen 311 Fernando et al. (2017) 312 https://research.google.com/teams/brain/about.html", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/SSRN-id3070741.tei.xml", "id": "9dbdb573453fed4dec83a22bd13e4633"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "There's a difference between someone instantaneously saying \"Yes!\" when you ask them on a date compared to \"...yes.\" Psychologists and economists have long studied how people can infer preferences from others' choices. However, these models have tended to focus on what people choose and not how long it takes them to make a choice. We present a rational model for inferring preferences from response times, using a Drift Diffusion Model to characterize how preferences influence response time and Bayesian inference to invert this relationship. We test our model's predictions for three experimental questions. Matching model predictions, participants inferred that a decision-maker preferred a chosen item more if the decision-maker spent longer deliberating (Experiment 1), participants predicted a decision-maker's choice in a novel comparison based on inferring the decision-maker's relative preferences from previous response times and choices (Experiment 2), and participants could incorporate information about a decision-maker's mental state of cautious or careless (Experiments 3, 4A, and 4B).", "authors": ["Vael Gates", "Frederick Callaway", "Mark K Ho", "Thomas L Griffiths"], "title": "Running head: RATIONAL PREFERENCE INFERENCE FROM RESPONSE TIME 1 A rational model of people's inferences about others' preferences based on response times", "text": "A rational model of people's inferences about others' preferences based on response times It's a quiet afternoon, and you and a new friend are trying to entertain yourselves. \"Hiking? Movie?\" you suggest. \"Movie,\" she says immediately. Feeling imaginative, you offer another choice: \"Movie or science museum?\" She takes longer to think on this. \"...movie,\" she says eventually. Taking into account how long it took your friend to make her choices, you might guess that she hates hiking but science museums are a good future activity. When we observe others making choices, we learn from the concrete choices they make, but also from cues like body language and timing. Inferring preferences from response times has a long history in economics and psychology (e.g.  Busemeyer, 1985; Busemeyer & Townsend, 1993; Chabris et al., 2008; Chabris et al., 2009; Diederich, 1997 Diederich, , 2003 Gill & Prowse, 2017; Moffatt, 2005; Spiliopoulos & Ortmann, 2018; Wilcox, 1993) . Recent sequential sampling models solve this problem by thinking of response time as an observable measure of a noisy cognitive process, where evidence is accumulated over time and a choice is made between options when the integrated evidence crosses a threshold. The longer it takes to make a decision, the closer those items are in the decision-maker's preferences, and the harder it is for the decision-maker to differentiate them. This principle has driven the use of sequential sampling models, the most common of which is the Drift Diffusion Model (DDM), to infer decision-makers' preferences from their response times  (Bogacz et al., 2006; Busemeyer & Rieskamp, 2014; Ratcliff, 1978; Ratcliff & McKoon, 2008; Ratcliff et al., 2016) . Previous work has established that models such as the DDM can be used to infer people's preferences from their response times (see section \"Background and previous work\" below). Here we ask a different question: Can people infer other people's preferences based on observing their response times and choices, modeling those others' decision processes as a diffusion process? In modeling terms, the DDM predicts a decision-maker's response times and choices given their preferences. Can we capture observers' inferences using a rational model in which we invert the DDM to predict decision-makers' preferences (Figure  1 )? People make inferences about preferences from response times constantly; using your friend's response times and choices to infer her preference for science museums over hiking is one illustrative example. People can make even more complex inferences from response times and choices if they know the decision-maker's mental state, like how carefully they are making a decision. Your grandfather knows that if someone's making a choice in a hurry, then their response times are less informative. He likewise knows that if someone's very cautious, then a fast decision is particularly meaningful. In this paper, we propose that these forms of everyday inferences can be formalized and described by inverting a generative process like the DDM. We explore the predictions of our inverse DDM through three experimental questions. In Experiment 1, we confirm the basic prediction that people will infer how much a decision-maker values one item over another based on observing the decision-maker's response times. In Experiment 2, we extend this analysis to testing people's predictions for a decision-maker's novel choice based on having previously inferred their relative strength of preferences. Just as you could infer a new friend likes science museums more than hiking based only on two indirect questions, our model suggests that participants will predict a decision-maker's choice in a novel comparison based on inferring the decision-maker's relative preferences from previous response times and choices. In Experiments 3, 4A, and 4B, we explore the possibility that participants will be able to infer even more from a decision-maker's response time if they are aware of the state of the decision-maker. When someone makes a quick decision, maybe it is because they value one choice much more than another, but maybe it is because they are feeling careless that day and are rushing through inessential decisions. These two factors correspond to different parameters of the DDM-\"value difference\" and \"threshold\"-which jointly predict the response time, and we ask if participants can perform joint inference over these parameters. We thus inform participants of the decision-maker's carefulness, and observe whether participants' predictions about their preferences vary. \n Background and previous work DDMs are often applied to perceptual decision-making tasks (e.g.  Bitzer et al., 2014; Brunton et al., 2013) . However, DDMs have also been used to model value-based decisions integrating both choice and timing information  (Shadlen and Shohamy (2016)  and  Polanìa et al. (2014)  discuss relationships between DDMs describing perceptual versus value decision tasks). An important property of DDMs applied to value-based decisions is that they predict that response time should correspond to strength of preference, meaning that how much a decision-maker values one item over another (\"value difference\") should be expressed in response time, where smaller value differences correspond to longer response times  (Echenique & Saito, 2017) . This prediction has been supported in a variety of settings: risky choice (Alós-Ferrer & Garagnani, 2020;  Konovalov & Krajbich, 2019) , intertemporal choice  (Amasino et al., 2019; Dai & Busemeyer, 2014; Konovalov & Krajbich, 2019; , social decision-making  (Frydman & Krajbich, 2018; Konovalov & Krajbich, 2019; Krajbich et al., 2014) , and food choice  (Clithero, 2018a (Clithero, , 2018b Krajbich et al., 2010; Krajbich & Rangel, 2011; Milosavljevic et al., 2010; Towal et al., 2013) . 1 Previous work has demonstrated that people can empirically infer the preferences of others from their response times. The DDM specifically has been used as motivation for some of these studies, though the authors do not formally invert the DDM as we do. For example, in  Konovalov and Krajbich (2020) , participants in a strategic bargaining experiment inferred their opponents' preferences based on both explicitly stated and observed response times , and changed their behavior to make use of this knowledge. In another study, participants in  Frydman and Krajbich (2018)  played a strategic information cascade game, in which they had to determine the binary state of the environment, receiving both a probabilistic private signal about the environment and also observing other players' public choices. The DDM predicts that choices should be slow when players' private signals were in conflict with other players' public choices, which the authors observed. The study then examined how people behaved when presented with other players' response times. Participants used this information to infer how unsure other players were about their decisions: When public choices were incorrect, participants in the response-time condition were significantly more likely to follow their private signals compared to participants without access to other players' response times. Finally, an interesting consequence of the DDM's prediction that people will spend more time on harder decisions is that participants can act suboptimally under circumstances with a fixed time limit and payoffs allocated per-choice. Participants may spend a suboptimal amount of time on low-stakes choices with small value differences rather than high-stakes choices with large value differences. When researchers observed this result, they found that by forcing a cutoff time for all decisions they could significantly improve participants' payoffs  (Krajbich et al., 2014; Oud et al., 2016) . Unlike previous work, we invert the DDM to predict how people will infer preferences from other decision-makers' choices and response times, fitting our model with participants' preference estimates about the decision-maker. This use of Bayes' rule to invert a decision-making model follows  Lucas et al. (2014 ), Jara-Ettinger et al. (2016 ,  Baker et al. (2017), and Jern et al. (2017) , who inverted choice models to describe how people infer preferences from the outcomes others select. We continue to follow their tradition of using inverse decision-making models to explore theory of mind, extending this work to incorporate response times. \n Bayesian inference over preferences from response times We begin with a broad overview of our inverted Drift Diffusion Model (DDM), in which we use Bayesian inference to infer preferences from response times. The basic setup is that one person (the observer) watches a second person (the decision-maker) make a choice between two options. The observer then infers a distribution over how much the decision-maker likes each option (their inferred utilities) based on the option they chose and-critically-the amount of time they took to make the decision. To make such an inference, the observer must have a generative model of the decision-maker's decision-making process. Here, we assume that this model is a DDM (Figure  1 ). 2 In this model, decisions are made on the basis of sequentially accumulated noisy evidence about the difference of the two options' utilities. The accumulation process runs until the total evidence in favor of one option or the other exceeds a threshold, at which point, the corresponding option is chosen. Formally, the evidence, x, is a dynamical system with dynamics dx = µ + σ dW, ( 1 ) where µ is the drift rate, σ specifies the amount of noise in the integration, and W is a Weiner process (the continuous limit of a Gaussian random walk, also called Brownian motion). The evidence is initialized at zero. Following  Milosavljevic et al. (2010) , we assume that the drift rate is a linear function of the difference in values of the two items, µ = β(u a − u b ), (2) where u a and u b are the utilities of the two items being chosen between, and the drift multiplier, β, can be interpreted as the decision-maker's sensitivity to differences in value. The accumulation process continues until the evidence crosses one of the decision boundaries. We assume symmetric constant boundaries, at a distance of θ from the initial point of zero. Intuitively, θ controls how careful the decision-maker is: larger values make decision slower but more accurate. If at any moment x > θ, option a is chosen; we denote this event a b. If x < −θ, option b is chosen, that is, b a. The time point at which this event occurs, t, is the response time (for simplicity, we do not consider the non-decision time variable that is often added to t to produce the response time). The DDM thus defines a probability distribution p ddm (a b, t | u a − u b ; β, θ). ( 3 ) To infer the decision-maker's preferences given the observed choice and response time, the observer must invert this model. We can do this using Bayes rule, resulting in a posterior distribution over the utilities p(u a , u b | a b, t) ∝ p(u a )p(u b ) • p ddm (a b, t | u a − u b ; β, θ), ( 4 ) where p(u a ) and p(u b ) capture the observer's prior distribution over utilities. For simplicity, we assume a standard Gaussian prior. Equation 4 specifies how a rational agent should update their beliefs about another person's preferences based on an observed choice and response time. However, the inferences one draws depend on the parameters of the DDM: the drift multiplier, β, and the threshold, θ. Intuitively, these parameters can be thought of as individual difference or mental state variables that jointly determine the accuracy and speed of the decision. We explore the θ variable further in Experiment 3. For now, it is sufficient to note that the qualitative model predictions in Experiments 1 and 3 are insensitive to these parameters. However, for the purpose of plotting the predictions, we fit the parameters by minimizing the sum of squared errors between the model predictions and aggregate participant responses across the first two experiments. We treat the third experiment separately, as described in the methods for that experiment. To summarize, the DDM serves as a generative model that relates decision-making parameters (the drift rate and threshold) to choices and response times. Critically, an observer can then invert this model to make inferences about the decision-maker's preferences. Here, we propose that human inferences from choices and response times will be consistent with such an ideal Bayesian observer model. To illustrate in a concrete example, recall our introductory example in which your friend takes longer to choose \"movie\" when it is paired with science museum than when it is paired with hiking. In considering her responses, you would conclude not only that she really wants to see a movie, but that she seems to like science museums more than hiking. Remarkably, this intuitive inference falls naturally out of inverting the DDM: Her longer response time in the movie-museum pairing reflects a more shallow drift rate, which results from the movie-museum utility difference being smaller than the movie-hiking difference. Since the movie has the highest utility (after all, she chose it twice), this means that the museum has higher utility than hiking. In the remainder of the paper, we empirically test three predictions of our model. Experiment 1 tests the prediction that observing faster choices indicates that the decision-maker has a larger preference difference. Our inverse DDM predicts this because a large difference in utilities between two items will result in a steeper drift rate, which leads to a faster decision. In Experiment 2, we test the prediction that people can draw inferences about novel choice pairs even when the choices alone provide no information about the relative utility. In particular, we present participants with scenarios similar to the movie/hiking/museum example and test whether they integrate information across observations consistent with inverting the DDM. Finally, in Experiments 3, 4A, and 4B, we examine whether people's conclusions about preferences from response times is affected by background knowledge about the decision-maker's mental state, specifically whether the decision-maker is feeling cautious or careless. This corresponds to performing inference about the decision-maker's drift rate (i.e., the relative utilities of the options) conditioned on different decision thresholds. \n Experiment 1: Inferring preferences from response times We test the core prediction of the DDM that decisions will be faster when the difference in preference for the two items is larger. In the model, a larger preference difference results in a higher drift rate and more rapid movement towards a decision boundary. Thus, if people employ a model like the DDM when inferring other's preferences from their response times, they should rate a chosen item as being more strongly preferred when the decision is made more quickly. \n Methods Participants. 481 participants with U.S. IP addresses were recruited through Amazon Mechanical Turk. Participants were paid $1.00 in compensation. Participants were excluded from the study if, for the critical questions, they indicated the decision-maker preferred a non-chosen item more than once, indicating lack of attention. 66 participants were excluded for a total of 415 participants. We used power analyses to determine the sample size for Experiments 1, 2, and 3 based on effect sizes observed in pilot studies. These power analyses, our exclusion criteria, our stimuli, and our planned procedures and analyses were preregistered here: Stimuli. In all experiments, participants viewed surveys created on Qualtrics containing text and videos. Experiments 1, 2, and 3 used the same set of videos (or a subset of these videos), which were created by filming a decision-maker making eight choices. The videos began by the decision-maker pulling two pieces of paper apart, which were covering two labeled bowls (e.g. A and B; bowls could be labeled A, B, or C). The decision-maker thought about their options while staring between the two bowls, and then they reached into one of the bowls to make a choice. The decision-maker made choices after 3, 5, 7, and 9 seconds after video onset, and made choices with either their left or right hand, for a total of eight videos. To counterbalance asymmetries in the videos, the number of videos used was doubled by creating flipped copies over the vertical axis, for a total of 16 videos. To counterbalance the effects of A, B, and C labels, the labels of the 16 videos were edited to use the label combinations AB, BA, CA, AC, BC, and CB, for a total of 96 videos. The text was modified to match the video labels. For narrative simplicity, we will only refer to the AB label combinations. Stimuli for all experiments are available on OSF: https://osf.io/pczb3. Procedure. After consenting, participants were shown introductory text: \"In this experiment, you will see videos of someone choosing between an item in one bowl and an item in another bowl. There are 8 videos. Please do your best to answer the questions afterward.\" Participants then read: \"Two items, A and B, are inside the bowls below. This person has just been asked to choose which item they want. Please watch the video.\" Participants then watched a video in which a decision-maker chose an item from bowl A or B within 3, 5, 7, or 9 seconds from the onset of the video. After watching the video, participants were asked: \"What are the person's feelings about item A and item B?\" Participants then moved a slider containing values from 0 (labeled \"Strongly Prefers A\") to 100 (labeled \"Strongly Prefers B\") and set initially at 50 (labeled \"Neutral Between A and B\"). No grid lines were shown, nor the numbers 0-100, and participants were required to click and move the slider (but they could move it back to 50 if desired). After answering this question the participant could advance to the next trial, which displayed on another page (Figure  2 ). Participants could replay videos as often as they wished, and all of the text/video/questions were on the same page. Participants could not return to a previous page. There were eight trials: Participants saw a total of eight videos (A/B choice × 3/5/7/9 seconds) in random order. We will refer to the choices as between A and B for narrative simplicity. However, participants were randomly assigned to one of 12 counter-balancing variants in which the labels and video orientations were randomized. Within every variant, the labels were held constant (options: AB, BA, AC, CA, BC, CB) and the video orientation was held constant (the videos were either as filmed, or were flipped over the vertical axis).  After watching all eight videos and answering the accompanying questions, participants had the opportunity to write comments then exited the survey. Model. We assume that participant responses are based on the posterior mean estimate of the difference in utilities of the two items. This is defined as E[u a − u b | a b, t] = (u a − u b ) • p(u a , u b | a b, t) du a du b . (5) However, because the DDM likelihood depends only on the difference in values (Equation  3 ), we can re-express Equation 5 as a single integral over the difference, E[δ u | a b, t] = δ u p(δ u | a b, t) dδ u , ( 6 ) where δ u = u a − u b and p(δ u | a b, t) = 1 Z Normal δ u ; µ = 0, σ 2 = 2 • p ddm (a b, t | δ u ). (7) Note that the variance of the difference between two Gaussians is the sum of the variance of each, hence σ 2 = 2. We have suppressed β and θ for concision. We compute the normalizing constant, Z, and the integral in Equation 6 by adaptive numerical integration  (Genz & Malik, 1980) . To convert this utility difference (which is in arbitrary units) to the scale defined by the slider participants used to make a response, we use a scaling parameter, such that the predicted response is α E[u a − u b | a b, t]. We fit α (along with β and θ) to aggregate participant behavior by minimizing the sum of squared errors. As one would expect, the DDM predicts that a choice of item a over item b is increasingly likely as u a − u b increases. However, the choice alone only weakly constrains the size of the difference. As illustrated in Figure  1 , the response time, t, provides much more information. With a strong preference, a fast response is quite likely and a slow response is quite unlikely. With a weak preference, probability is more evenly spread across response times. As a result, when we invert the DDM, we find that a strong preference is more likely given a fast response (because a fast response is likely given a strong preference) and a weak preference is more likely given a slow response (because a slow response is unlikely given a strong preference). \n Results In Experiment 1, we showed participants videos in which a decision-maker took 3, 5, 7, or 9 seconds to decide between two items, then asked participants about the decision-maker's preferences about the items. We expected that participants would infer that the decision-maker preferred the chosen item more when the response time was shorter (e.g. 3 seconds) compared to when the response time was longer (e.g. 9 seconds), as predicted by our model. This is indeed what we observed, with a qualitatively close match between the model's predictions and the results (Figure  3 ). To analyze the effect of response time, we created a linear mixed effects model with response time as a continuous fixed effect, subject as a random effect, and participants' preference judgments (0-100, 50 was \"Neutral Between A and B,\" 0 was \"Strongly Prefers A\" and 100 was \"Strongly Prefers B\") as the continuous output variable. Each participant had eight data points: two answers each for the 3, 5, 7, and 9-second questions.  \n Discussion As predicted by our model, participants inferred that a decision-maker preferred an item more the less time they spent selecting the item. This effect was statistically significant, and qualitatively matched our model results. Interestingly, the individual participant responses in Figure  3  suggest that the main effect of response time may be even stronger than indicated by the mean responses, because a minority of participants (40/415 = 9.6%) do not take the decision-maker's response time into account. Instead, they inferred that the decision-maker had the maximal preference for the chosen item on every question (shown in Figure  3  by a set of {50, 50, 50, 50} responses, indicating the participant moved the slider to 100 on four trials and 0 on four trials). \n Experiment 2: Predicting novel choices based on relative strength of preferences In Experiment 1, we demonstrated that participants could infer preferences of a decision-maker from their choices and their response times. However, people regularly perform more complex inferences than that: If your friend did not hesitate at all before choosing movies over hiking, but hesitated a while before choosing movies over science museums, then you could guess that she preferred science museums to hiking. We now explore this ability to predict a decision-maker's choice in a novel comparison, based on having inferred their relative strength of preferences from response times in previous choices. Inferring graded preferences from choices (or response times) that will allow prediction for novel comparisons is not often studied. The most closely related literature surrounds transitive inference. In transitive inference paradigms, participants-children, adults or animals-are asked to draw conclusions like if B > A, and A > C, then B > C. In these paradigms, associations are often manually taught between previously-unassociated items or assertions, and transitive inference is solely a function of choices (e.g.  Acuna et al., 2002; Harris & McGonigle, 1994; Maybery et al., 1986; Thayer & Collyer, 1978) . It is rare to show participants executing transitive inference about preferences, which are distinct in that they are often socially learned through implicit cues in addition to appearing probabilistic and non-causal. Similar to  Hu et al. (2015) , we test whether people can infer another decision-maker's graded preferences by observing their choices. However, rather than watching the decision-maker make choices repeatedly, we had participants watch the decision-maker make two decisions. We investigate whether people can infer a decision-maker's preferences enough to predict their selection on a novel choice after only having observed the decision-maker's choices and response times twice. We can do this because it can be much more efficient to use response time and choice information together than choice information alone. If A were chosen over B, and A were chosen over C, people would have to observe more choices to statistically infer the graded preferences that would allow them to make a prediction for an unseen choice between B and C. Using response time information, people could make this prediction immediately, given they had inferred the decision-maker's relative strength of preferences after observing their previous choices and response times. In Experiment 2, participants saw pairs of videos. In the first video, the decision-maker made a choice between two items, A and B, within 3 seconds. In the second video, the decision-maker made a choice between the original item and another item, A and C, within 9 seconds. Participants were then told that the decision-maker had to make a choice between B and C, and were asked what choice they thought the decision-maker would make and how likely that choice was. Our model predicts that participants would anticipate the decision-maker's selection on the novel choice, having inferred the decision-maker's preferences based on response times and choices from the previous two decisions. Our model also proposes that participants' inferred likelihoods of the predicted choice would be lower than in the control condition in which participants only needed to use choices to make the prediction (and did not need to take response time into account). \n Methods Participants. 478 participants with U.S. IP addresses were recruited through Amazon Mechanical Turk. Participants were paid $1.50 in compensation. No exclusion criteria were applied. The sample size was determined by a power analysis based on the effect size of pilot studies and was preregistered. Stimuli and Procedure. Video stimuli were the same as in Experiment 1. After consenting, participants read introductory text: \"In this experiment, you will see videos of someone choosing between an item in one bowl and an item in another bowl. There are 8 sets of videos. Please do your best to answer the questions afterward.\" Participants then entered the main experiment. In Experiment 2, participants saw pairs of videos. Participants saw the text \"This person is choosing between two items, A and B,\" and then were shown a first video in which the decision-maker makes a choice within 3 seconds. On the same page, participants were then presented with the text \"Now they are choosing between A and C,\" and shown another video of a decision-maker making a choice within 9 seconds. The participant was then shown, also on the same page, \"This person is now offered a choice between items B and C. What choice do you think they'd make, and how likely do you think it is that they'd make that choice?\" Participants then moved a slider containing values from 0 (labeled \"Very likely B\") to 100 (labeled \"Very likely C\") and set initially at 50 (\"Equally likely\"). No grid lines were shown, nor the numbers 0-100, and participants were required to click and move the slider (but they could move it back to 50 if desired). After watching the two videos and answering  the question, participants could advance to the next page, where they watched another pair of videos and answered another question (Figure  4 ). There were eight pages/questions total. Participants could replay videos as often as they wished, and all of the text/videos/questions were on the same page. Participants could not return to a previous page. After watching all eight pairs of videos and answering the accompanying questions, participants had the opportunity to write comments then exited the survey. For narrative simplicity, we will refer to the first video as a 3-second choice between A and B, and the second video as a 9-second choice between A and C. However, randomized counter-balancing variants were used in the experiment. We randomized the order of which video was presented first within-participant (half of the trials showed the 3-second video first, and half of the trials showed the 9-second video first). Between-participants, we randomized the order of the final choice between participants (in one variant, participants made a choice between B and C, and in another variant participants made a choice between C and B), the labels (one label was kept consistent on one side, resulting in the following variants: AB vs AC, BA vs CA, BA vs BC, AB vs CB, CA vs CB, and AC vs BC), and video orientation (half of the variants had all videos flipped over the y-axis). This resulted in a total of 24 variants (2 final choice options × 6 label pairs × 2 video orientations), and 8 questions per participant (2 orderings of 3/9-second videos × 4 choice options (R/R, R/L, L/R, R/R, where \"R/L\" represents the decision-maker choosing the right-most object in the first video and the left-most object in the second video)). Participants were randomly assigned to one of these 24 variants, within which they viewed the 8 pairs of videos wherein labels and video orientation were held constant but response time and choices varied. A given participant would be shown the video pairs in random order, but the videos within the pairs were ordered. Analysis. Our stimuli resulted in four conditions. Participants watched the decision-maker choose between A and B in one choice, then A and C in another, and in the critical question participants were asked whether the decision-maker would prefer B or C in a novel choice. We thus defined the four conditions based on whether the decision-maker chose B or C in either of the observed choices. The condition \"NeitherChosen\" indicated that the decision-maker never chose B or C in the observed choices (choosing A instead), and the condition \"BothChosen\" indicated that the decision-maker chose B in one choice and C in another. The condition \"FastChoice\" indicated that the decision-maker chose B once, and \"Slow choice\" indicated that the decision-maker chose C once (Figure  5 ). Each participant answered two questions for each of the four conditions, for a total of eight data points. \n Choice, 3sec Choice, 9sec Predicted Choice Condition and 9-second). In the \"NeitherChosen\" condition, the decision-maker chose neither B nor C, and in the \"BothChosen\" condition, the decision-maker chose B and C. For the non-inference control conditions, in the \"FastChoice\" condition, the decision-maker chose B quickly, and in the \"SlowChoice\" condition, the decision-maker chose C slowly. A > B A > C C > B NeitherChosen B > A A > C B > C FastChoice B > A C > A B > C BothChosen A > B C > A C > B SlowChoice The first two conditions, \"NeitherChosen\" and \"BothChosen,\" require that the participant make an inference based on response time, since in the \"NeitherChosen\" condition the decision-maker never preferred B or C in the observed choices, and in the \"BothChosen\" condition the decision-maker preferred both. Thus when the participant was asked whether the decision-maker preferred B or C, the answer was ambiguous based solely on the choices the participant had seen. The second two conditions, \"FastChoice\" and \"SlowChoice,\" did not require that the participant make an inference based on response times, just choices, since the decision-maker preferred either B or C in the choices they made in the videos. Thus when the participant was asked whether the decision-maker preferred B or C, the decision-maker should logically choose whichever choice the decision-maker preferred in one of the videos. Model. We assume that participants respond with the posterior mean probability of the unseen choice given the two observed choices and response times. To compute this value, we first compute a posterior over each item's value, and then integrate over these values to produce a predicted choice. Here, we specify these steps for the \"NeitherChosen\" condition; the derivations for the other conditions have the same form. Let D = {a b, t ab , a c, t ac } denote the observed data. The posterior over values is then given by p(u a , u b , u c | D) ∝ ϕ(u a )ϕ(u b )ϕ(u c ) • p ddm (a b, t ab | u a − u b ) • p ddm (a c, t ac | u a − u c ), (8) where ϕ is the standard normal pdf, and the posterior over choice is given by marginalizing over the three values as well as the predicted decision time, p(b c | D) = ∞ 0 p ddm (b c, t bc | u b − u c ) • p(u a , u b , u c | D) dt bc da db dc. ( 9 ) At a high level, the key model prediction is that people will be able to predict a novel choice even when the previous choices (without response times) do not provide any information about which option is preferred. Consider the \"NeitherChosen\" case. Here, the observer must predict a choice between B and C after observing A being chosen over both of them. This implies that u a > u b and u a > u c , but provides no information about u b − u c . However, considering the response times of each choice (short and long, respectively) provides additional information about the relative strength of preference in the first two choices. Following the predictions of Experiment 1, the observer infers that u a − u b is large because this decision was made quickly and that u a − u c is small because this decision was made slowly. Thus, u c must be greater than u b , and C is more likely to be chosen over B. This line of reasoning can be summarized in notation as: a b ∧ a c ∧ t ac > t ab =⇒ u a − u c < u a − u b =⇒ u c > u b =⇒ p(c b) > 1 2 . ( 10 ) The logic is similar in the \"BothChosen\" case, but inverted: b a ∧ c a ∧ t ac > t ab =⇒ u c − u a < u b − u a =⇒ u c < u b =⇒ p(b c) > 1 2 . ( 11 ) For comparison, we also consider cases where the previous choices alone do constrain the prediction about the novel choice, by transitivity of preference. In the \"FastChoice\" condition we have b a ∧ a c =⇒ u b > u a ∧ u a > u c =⇒ u b > u c =⇒ p(b c) > 1 2 , ( 12 ) and in the \"SlowChoice\" condition, a b ∧ c a =⇒ u a > u b ∧ u c > u a =⇒ u c > u b =⇒ p(c b) > 1 2 . ( 13 ) The model makes a stronger prediction in the \"FastChoice\" and \"SlowChoice\" conditions because choices provide stronger and more direct evidence of preference ordering compared to differences in response time. \n Results Our model predicts that participants would anticipate a decision-maker's novel choice based on inferring the relative strength of that decision-maker's preferences from previous response times and choices. Specifically, participants should predict the decision-maker's choice better than chance. Moreover, participants' inferred likelihoods of the predicted choice should be lower than the control conditions in which only choices (and not response times) were necessary. We compare our results qualitatively to these model predictions (Figure  6 ). Difference from chance, all conditions. To determine whether participants could make predictions for a novel choice better than chance, we conducted the equivalent of (a) one-sample t-tests for a linear mixed effects model.  3  Condition was a categorical fixed effect (\"NeitherChosen\"/\"BothChosen\"/\"FastChoice\"/\"SlowChoice\"), subject was a random effect, and likelihood judgment was a continuous output variable. Choice, 3sec Choice, 9sec Predicted Choice Condition A > B A > C C > B NeitherChosen B > A A > C B > C FastChoice B > A C > A B > C BothChosen A > B C > A C > B SlowChoice (b) A > B A > C C > B NeitherChosen B > A A > C B > C FastChoice B > A C > A B > C BothChosen A > B C > A C > B SlowChoice In the \"NeitherChosen\" condition, we predicted better-than-chance performance: specifically that if the decision-maker chose A B quickly, and A C more slowly, that participants would infer that the decision-maker had preferences A > C > B. From our linear mixed effects model described above, the estimate (beta parameter) for the \"NeitherChosen\" condition was 7.99 (95%CI=[5.89,10.10], t(1681.5)=7.46, p=1.4e-13, t-tests using Satterthwaite's method, η 2 p =0.03), 4 in the expected direction C > B. The estimated power of the \"NeitherChosen\" condition against chance for α=0.01 was 100.0% (95%CI=[98.17,100.0]) using a t-test, generated via 200 simulations with the R package \"simr.\" In the \"BothChosen\" condition, we expected that if the decision-maker chose B A quickly, and C A slowly, then participants would infer that the decision-maker had preferences B > C > A above chance. From our linear mixed effects model described above, the estimate (beta parameter) for the \"BothChosen\" condition was 14.53 (  95%CI=[12.43, 16 .63], t(1681.5)=13.55, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.10), in the expected direction B > C. The estimated power of the \"BothChosen\" condition against chance for α=0.01 was 100.0% (95%CI=[98.17,100.0]) using a t-test, generated via 200 simulations with the R package \"simr.\" These results match our model predictions, with one difference being that the model predicted a more similar level of inference for the \"NeitherChosen\" and \"BothChosen\" conditions than was observed empirically. We also expected that participants would make the appropriate choices in the \"FastChoice\" and \"SlowChoice\" control conditions, that they would do so above chance performance, and that they would do so more accurately than in the \"NeitherChosen\" and \"BothChosen\" inference conditions. The \"FastChoice\" and \"SlowChoice\" conditions act as control conditions: In these conditions, participants only needed to use choice information to correctly infer the decision-maker's preferences, instead of also needing response times. From our linear mixed effects model, the estimate (beta parameter) for the \"FastChoice\" condition was 25.17 (95%CI=[23.07,27.27], t(1681.5)=23.47, p<2e-16, t-tests using Satterthwaite's 4 All η 2 p effect sizes were calculated using the assumption that t 2 = F , followed by  Lakens, 2013)  with df num = 1. η 2 p = F * df num /(F * df num + df res ) (e.g. method, η 2 p =0.25), in the expected direction B > C. The estimated power of the \"FastChoice\" condition against chance for α=0.01 was 100.0% (95%CI=[98.17,100.0]) using a t-test, generated via 200 simulations with the R package \"simr.\" The estimate for the \"SlowChoice\" condition was  22.34 (95%CI=[20.24, 24 .44], t(1681.5)=20.84, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.21), in the expected direction C > B. The estimated power for α=0.01 was 100.0% (95%CI=[98.17,100.0]) using a t-test. These results matched our model predictions. One difference was that the model predicted slightly more similar level of inference across these conditions than was observed empirically. The \"FastChoice\" and \"SlowChoice\" results serve as an baseline for accuracy based purely on choice compared to the \"NeitherChosen\" and \"BothChosen\" results, which required inferences based both on choice and response time. We explore this comparison directly next. Comparison between conditions: timing-inference conditions vs choice control conditions. In addition to examining how our results differ from chance, we also examine how the timing-based inference conditions (\"NeitherChosen\"/\"BothChosen\") compare to the control conditions (\"FastChoice\"/\"SlowChoice\"). To this end, we constructed a linear mixed effect model where condition was a categorical fixed effect, subject was a random effect, and participants' likelihood judgments were a continuous output variable. However, unlike the previous model, the conditions were merged (\"NeitherChosen/\"BothChosen\" and \"FastChoice\"/\"SlowChoice\") to create two comparison groups, and these groups were compared to each other. The estimate (beta parameter) for the fixed effect of group (timing-inference conditions versus control conditions) was -12.49 (95%CI=  [-14.22,-10 .77], t(3345.0)=-14.17, p<2e- showing a statistically significant pairing between the timing-based inference conditions and the control conditions, with the control conditions showing higher inferred likelihoods for predicted choice. \n Discussion In Experiment 2, we asked whether participants could make a prediction about an unseen choice based on inferring the relative strength of a decision-maker's preferences from response time and choice information from two observed choices. Our model results and empirical results were in close alignment: In all conditions, participants inferred the predicted choice above chance. Moreover, participants had lower inferred likelihoods for the predicted choice for the conditions in which they needed to use both response time and choice information (the \"NeitherChosen\" and \"BothChosen\" inference conditions) compared to when they only needed to use choice information (the \"FastChoice\" and \"SlowChoice\" conditions, which serve as an baseline for accuracy based purely on choice). Specifically, in the \"NeitherChosen\" condition, participants saw the choice A B made quickly, the choice A C made slowly, and the model predicted that they would infer that C > B, based on the DDM reasoning that items closer in value take a longer time to decide between. In the \"BothChosen\" condition, participants saw the choice B A made quickly, the choice C A made slowly, and the model predicted that they would infer that B > C, again based on the reasoning that items closer in value take a longer time to decide between. Interestingly, participants were better able to perform the inference for the \"BothChosen\" condition compared to the \"NeitherChosen\" condition, while the model predicted very similar inferred likelihoods for both of these \"inference\" conditions. Intuitively, it can feel easier to choose whether B or C was better if either of these items were chosen compared to if neither were chosen, but this intuition was not captured by the model. Future versions of the model should capture this more subtle behavior, perhaps by incorporating a bias for reduced timing for chosen items. In the non-inference control conditions, participants inferred the predicted choice similarly across the two conditions \"FastChoice\" (the participants saw B A chosen quickly, A C chosen slowly, and had to infer B C) and \"SlowChoice\" (the participants saw A B chosen quickly, C A chosen slowly, and had to infer C B), as predicted by the model. Participants did have slightly higher inferred likelihoods for the predicted choice on the \"FastChoice\" condition, in which the target item was chosen quickly rather than slowly, which was not represented in the model. This was a relatively minor effect, but future work could to explore whether this element of human psychology could be captured in a modified DDM or with a different set of well-justified parameters. Relevantly, the model predictions for Experiment 2 were highly varied based on the parameter settings, so it could be interesting to explore what kind of behavior is described under those different settings. \n Experiment 3: Sensitivity to a decision-maker's mental state We next investigated whether people would incorporate a decision-maker's mental state when reasoning about response times and implied preferences. If a decision-maker quickly chose between two items, it could have been because one of the options was obviously preferred. However, if the decision-maker was feeling careless rather than cautious that day, their quick decision-making could also be attributed to being tired and rushing through decisions. Both factors-the value difference between the two choices, and the decision-maker's overall carefulness-could influence response time. We asked whether participants would incorporate the decision-maker's carefulness in reasoning about their preferences. From a modeling perspective, in the DDM, a decision-maker's carefulness is naturally modeled by the decision threshold θ, which determines the amount of evidence that must be accumulated in favor of an item before a decision is made. The higher the threshold, the less sensitive the decision-maker will be to random fluctuations in the evidence, and thus the less likely they will be to make an error by selecting the item that they actually disprefer. However, this robustness comes at a cost: Accumulating evidence takes time, and so a decision-maker with a high threshold will make slower decisions. Importantly, the decision time depends on both the drift rate and the threshold. A slow decision can result from either a low drift rate (small preference difference) or a high threshold (high caution) (Figure  7 ). The combined influence of drift rate and threshold on decision time results in a critical prediction of our model: knowing a decision-maker's carefulness should affect one's inferences about their preferences. Specifically, if you believe that a decision-maker is very cautious (has a high threshold), and you observe them make a fast decision between two items, you should infer a strong preference for the chosen item, because a high threshold is unlikely to be reached quickly unless the evidence is strong. In contrast, if you believe the decision-maker to be careless (low threshold), a fast decision is consistent with moderate or even no preference difference. In Experiment 3, participants watched decision-makers making decisions quickly (3 seconds) or slowly (9 seconds), having been described as cautious (high threshold) or careless (low threshold). Participants were then asked how much they believed the decision-maker valued their chosen item. We hypothesized based on model predictions that if a decision-maker made a choice very quickly and was described as being cautious, then participants would infer that the decision-maker valued one of the items a lot more than the other. If a decision-maker made a choice very quickly but was described as careless, we expected that participants would attribute this quickness only partly to how much the decision-maker valued the items, and partly to the decision-maker's mental state. We expected the same pattern of results if a decision-maker made a choice slowly. Finally, we expected that there would be an interaction effect between response time and carefulness/threshold. Specifically, we predicted that participants would learn more about a decision-maker's preferences if they were cautious (where a fast decision versus a slow decision would be more meaningful) compared to if they were careless (where a fast decision versus a slow decision would be less meaningful or indicative of their preferences). value difference or preference for one item over the other is the same), a more cautious decision-maker (having a high threshold) will take longer to make a choice than a less cautious decision-maker (having a low threshold). (B) In an inverted DDM, if the observer does not know the decision-maker's drift rate (preferences) or threshold (carefulness), the observer must infer whether a decision was fast because the decision-maker had a strong preference for that item, or because the decision-maker was being careless and not thinking too much about their decision. In Experiment 3, participants are told the decision-maker's carefulness and response time and must infer their strength of preference. \n Methods Participants. 481 participants with U.S. IP addresses were recruited through Amazon Mechanical Turk. Participants were paid $1.00 in compensation. In our preregistration, we indicated the following exclusion criteria: Participants were excluded from the study if they answered more than two out of eight manipulation check questions incorrectly. (These manipulation check questions directly queried whether the participants had read the stimuli text. Answering them incorrectly demonstrated lack of attention.) We determined the sample size via a power analysis based on the effect size of pilot studies and these exclusion criteria. In this power analysis, we calculated that we would need at least 27 participants to achieve 99% power with our pilot result effect size and α=0.01, using a Type-III F -test, but to ensure clear results chose to use a 480 as our sample size (one extra participant was recruited by the platform). However, our original exclusion criteria resulted in 212 participants being excluded (44% of our participants), which felt overly restrictive. For that reason, we decided to not use any exclusion criteria in Experiment 3, thereby including all participants in the analysis (we discuss the manipulation checks further in the Discussion). When we redid the power analysis calculation based on the pilot data with no exclusion criteria, using the same parameters as above, we calculated that we would need at least 179 participants to achieve 99% power. Thus, since 481 participants were included in the analysis, we still included a sample size with the desired power. In the Supplemental Material, we restrict our sample to that specified by the preregistered exclusion criteria, and present the planned analyses and figures based on this sample for comparison. The results were similar to those presented here (Figure  S2 ). \n Stimuli and Procedure. Video stimuli were the same as in Experiment 1. After consenting, participants read introductory text: \"In this experiment, you will see videos of someone choosing between an item in one bowl and an item in another bowl. There are 8 videos. Please do your best to answer the questions afterward.\" Participants then entered the main experiment. Participants either read descriptive text emphasizing that the decision-maker was feeling meticulous and cautious (high threshold condition), or that the decision-maker was feeling careless and inattentive (low threshold condition). The text for \"high threshold\" read: \"This person is choosing between two items. It is very important to them to be very careful about the following decision. This person cares a lot about being meticulous and cautious right now. They had an easy day at work and so do not have much on their mind besides this decision.\" The text for \"low threshold\" read: \"This person is choosing between two items. It is not important to them to be careful about the following decision. This person is fine being careless and inattentive right now. They had a hard day at work and have a lot on their mind besides this decision.\" 5 Participants then saw a video underneath this text showing a decision-maker making a decision between two items, A and B, where the decision-maker took 3 seconds from video onset or 9 seconds from video onset to make these decisions. Next, participants were asked \"How does the person feel right now?\" and were given a binary choice between a \"Meticulous and cautious\" radio button and a \"Careless and inattentive\" radio button, which was used as a manipulation check to ensure the participants had read the descriptive text (the ordering of these buttons was random on each trial). Participants were finally shown the critical question, asking about the decision-maker's preferences: \"How much do you think they value their chosen item?\" Participants moved a slider containing values from 0 (labeled \"Neutral between items\") to 100 (labeled \"Strongly prefers item\"; 50 was labeled \"Moderately prefers item\"). No grid lines were shown, nor the numbers 0-100. The slider was initially set at 0, and participants were required to click and move the slider, but they could move it back to 0 if desired. After answering these two questions, participants could advance to the next page to see the next set of text/video/questions (Figure  8 ). There were eight pages total. All of the text/video/questions were on the same page, and participants could replay videos as often as they wished. Participants could not return to a previous page. \n Neutral between items Strongly prefers item Moderately prefers item \n How does the person feel right now? This person is choosing between two items. It is very important to them to be very careful about the following decision. This person cares a lot about being meticulous and cautious right now. They had an easy day at work and so do not have much on their mind besides this decision. \n Meticulous and cautious Careless and inattentive How much do you think they value their chosen item? Each participant was randomly assigned to one of 12 counter-balancing variants in which the labeling and orientation of viewed videos were held fixed, and within which the participant viewed eight videos varying in response time and threshold. These eight videos presented in random order: Half of the choices were made in 3 seconds and half in 9 seconds (response time counterbalancing), half with the \"high threshold\" text and half with the \"low threshold text\" (threshold counterbalancing), and half with the decision-maker's right hand and half with the left hand (choice counterbalancing). After watching all eight videos and answering the accompanying questions, participants had the opportunity to write comments then exited the survey. Model. We assume, as in Experiment 1, that participant responses are based on the posterior mean estimate of the difference in utilities of the two items. We additionally assume that the carefulness manipulation influences participants' assumed threshold, θ, such that those in the cautious condition will make predictions consistent with a high θ and those in the careless condition will make predictions consistent with a low θ. To capture this, we held fixed the β and α parameters fit in the previous two experiments, and fit θ separately to the two groups. We accounted for the different scale in the response slider (specific to Experiment 3) by setting the predicted response to 2 * α E[u a − u b | a b, t] . Besides these differences, the predictions were produced in the same way as for Experiment 1, using Equation  5 . \n Results In Experiment 3, we investigated whether participants could make inferences about decision-makers' preferences based on both timing and threshold (how careful the decision-maker is) simultaneously. Our model predicts a main effect of response time (the shorter the response time, the larger the inferred value difference between the two items will be) and a main effect of threshold (the higher the threshold, the larger the inferred value difference between the two items will be). Our model also predicts an interaction effect between response time and threshold, such that decision-makers who are cautious and fast to make decisions are predicted to particularly value their chosen item, compared to decision-makers who are cautious and slow to make decisions (whereas we expect a smaller difference for the response time variable if decision-makers are emphasized to be careless). Results are shown in Figure  9 . To analyze the results, we constructed a linear mixed effects model, with response time represents the participant feeling that the decision-maker was neutral between items, 50 that the decision-maker moderately preferred their chosen item, and 100 that the decision-maker strongly preferred their chosen item. (b) Model Predictions. and threshold as categorical fixed effects, subject as a random effect, and participants' judgments of value (0-100, where 0 is \"Neutral between items,\" 50 is \"Moderately prefers item,\" and 100 is \"Strongly prefers item\") as the continuous dependent variable. The estimate (beta parameter) for the main effect of response time was 4.  07 (95%CI=[2.62, 5.53], t(3364.0)=5.48,  t(480.0)=82.97, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.94.) Results were similar, using the preregistered exclusion criteria; these results are included in the Supplemental Material (Figure  S2 ). Thus, while our empirical results matched our model predictions in that we observed the expected main effects of response time and threshold, we failed to observe the anticipated interaction effect. \n Discussion In Experiment 3, we asked whether participants could infer a decision-maker's preferences from their choices and response times while incorporating information about their mental state of being cautious or careless. As expected, we observed a main effect of response time in the expected direction (spending less time on a decision was associated with larger inferred preferences), though the main effect of response time was weaker than in Experiment 1, when there was only one manipulation (response time) rather than two (threshold and response time) and the critical question had a different format. We also observed a main effect of threshold in the expected direction (more carefulness was associated with larger inferred preferences). Finally, our model predicted an interaction effect, wherein a greater value difference was expected for cautious decision-makers with different response times compared to careless decision-makers, which we did not observe. To manipulate threshold, we included a textual description above our video stimuli describing whether the decision-maker was feeling cautious or careless. However, in the videos, the decision-maker always appeared to be focused and therefore cautious, and participants expressed confusion in the experiment's comments section about the discrepancy between the textual description of the decision-maker's mental state and the decision-maker's facial expressions. This issue was reflected in the manipulation check questions, which asked \"How does the person feel right now?\" with the binary options of \"Meticulous and cautious\" and \"Careless and inattentive.\" Participants asymmetrically answered the manipulation check questions incorrectly. When the described mental state was \"Meticulous and cautious,\" 387/481 (80%) of participants answered at least 3 out of the 4 trials correctly (the mean correct number of trials was 3.3/4). When the described mental state was \"Careless and inattentive,\" 261/481 (54%) of participants answered at least 3 out of the 4 trials correctly (the mean correct number of trials was 2.6/4). In our preregistration, we had planned to exclude participants who answered more than two of the eight manipulation checks incorrectly, but these criteria would have resulted in 212 participants (44%) being excluded.  6  We thus included all participants in the experiment, though results were similar when the exclusion criteria were applied (Figure  S2 ). We expected that the high failure rate for the manipulation check exclusion criteria was due to the mismatch between the textual description of the decision-maker's mental state and the facial expression confound in the associated videos, so conducted Experiments 4A and 4B to address this confound. \n Experiments 4A and 4B: Sensitivity to a decision-maker's mental state with different stimuli In exploring whether knowing a decision-maker's mental state of cautious or careless influences people's inferences of their preferences, Experiment 3 introduced a confound in the threshold manipulation whereby the decision-maker's facial expressions did not match the textual descriptions of their mental state. We sought to address this using different stimuli in Experiment 4A (text message videos) and Experiment 4B (vignettes). In Experiment 4A, participants saw text message videos between the decision-maker and a friend, Alice. Alice either described the decision-maker as having a good day at work and careful in their decision-making (high threshold condition), or having a hard day at work (low threshold condition). Alice then asked the decision-maker to choose between items A and B. In the fast response condition, the decision-maker responded \"Hm,\" then generated a decision. In the slow response condition, the decision-maker responded \"Hm,\" then spent time deliberating, designated by a moving ellipsis, before giving a decision. In Experiment 4B, rather than watching a video, participants read a vignette of the decision-maker's mental state, the time they took to make a choice, and their choice. The manipulation check questions in Experiments 4A and 4B remained the same as those in Experiment 3. In both experiments, we addressed the same experimental question as in Experiment 3; thus, the model predictions are qualitatively the same. \n Methods Participants. In Experiment 4A, 480 participants were recruited through Prolific and paid $1.25 in compensation. In Experiment 4B, 481 participants were recruited through Prolific and were paid $1.00 in compensation. We did not exclude any participants. Experiments 4A and 4B used the same sample size and analyses as Experiment 3, and were not preregistered. Both experiments were completed at a later date and on a different platform (Prolific instead of Amazon Mechanical Turk), so it is possible but unlikely that participants from the first round of experiments (pilots, Experiments 1-3) participated in the Stimuli and Procedure. The stimuli and procedure were similar to those in Experiment 3, but the main stimuli and critical question were changed. The main stimuli in Experiment 4A were changed to text message videos, and in Experiment 4B to vignettes. The critical question for Experiments 4A and 4B was posed as in Experiment 1 (\"What are the person's feelings about item A and item B?\") rather than Experiment 3 (\"How much do you think they value their chosen item?\"), since we considered the output of a scale spanning \"Strongly Prefers A,\" \"Neutral Between A and B,\" and \"Strongly Prefers B\" to be more robust than Experiment 3's scale spanning \"Neutral between items,\" \"Moderately prefers [chosen] item,\" and \"Strongly prefers [chosen] item.\" After consenting, participants read introductory text. In Experiment 4A: \"In this experiment, you will see text message videos of someone choosing between two items. There are 8 videos. Please do your best to answer the questions afterward.\" In Experiment 4B: \"In this experiment, you will read vignettes of someone choosing between two items. There are 8 vignettes. Please do your best to answer the questions afterward.\" Participants then performed eight trials of the experimental task. In Experiment 4A, on each trial participants read: \"This person is choosing between two items. Please watch the video.\" A video underneath showed a text message conversation between Alice and the decision-maker. In the low threshold condition, the text in the video was the following. Alice: \"I know you had a rough day, sorry to ask you this when you're distracted\" / \"but\" / \"of the two items we discussed\" / \"did you want\" / \"A or B?\" The decision-maker then replied: \"Hm\" / \"A\" (or \"B\"). The slashes here represent line breaks. In the high threshold condition, the first two lines from Alice were replaced with: \"I heard you had a nice day at work, so figured I'd ask you today since I know you're careful about this stuff :)\". Pauses were placed before each new line to approximate typing and reading speed, and the total time of the videos was the same for the low and high threshold conditions. The response time manipulation was in how long the decision-maker took to answer \"A\" or \"B.\" In the fast response condition, the answer came 2.0 seconds after \"Hm\" appeared. In the slow response condition, the answer came after 8.9 seconds; during this extended pause, an ellipsis (...) periodically displayed. An example stimulus is shown in Figure  10 . Participants were then asked the manipulation check question: \"How does the person feel right now?\" Participants were given a binary choice between a \"Meticulous and cautious\" radio button and a \"Careless and inattentive\" radio button; the ordering of these buttons was randomized on each trial. Participants were finally shown the critical question, asking about the decision-maker's preferences: \"What are the person's feelings about item A and item B?\" Participants moved a slider containing values from 0 (labeled \"Strongly Prefers A\") to 100 (labeled \"Strongly Prefers B\"; 50 was labeled \"Neutral Between A and B\"). No grid lines were shown, nor the numbers 0-100. The slider was initially set at 50, and participants were required to click and move the slider, but they could move it back to 50 if desired. After answering these two questions, participants could advance to the next page. There were eight pages total. All of the text/(video)/questions were on the same page, and participants could replay videos as often as they wished if videos were present. Participants could not return to a previous page. Each participant was randomly assigned to one of six counter-balancing variants in which the item labels were held fixed (AB, BA, CA, AC, BC, and CB). Only the AB variant, in which the items were labeled A and B, will be described for expository purposes. Within  \n Results Here we asked whether people would incorporate a decision-maker's carefulness in inferring their preferences from response times, using different stimuli from Experiment 3: text message videos in Experiment 4A, and vignettes in Experiment 4B. The model predictions are identical to those for Experiment 3: we predicted a main effect of response time (the shorter the response time, the larger the inferred value difference between the two items) and a main effect of threshold (the higher the threshold, the larger the inferred value difference between the two items). We also predicted an interaction effect between response time and threshold, such that decision-makers who were cautious and fast to make decisions would be judged to particularly value their chosen item, compared to decision-makers who were cautious and slow to make decisions (whereas we would expect a smaller difference for the response time variable if decision-makers were emphasized to be careless). Results are shown in Figure  12 . To analyze the results, we created a linear mixed effects model, with response time and threshold as categorical fixed effects, subject as a random effect, and participants' preference judgments (0-100, 50 was \"Neutral Between A and B,\" 0 was \"Strongly Prefers A\" and 100 was \"Strongly Prefers B\") as the continuous output variable. For Experiment 4A, the results from the linear mixed effects model were as follows. The estimate (beta parameter) for the main effect of response time was 10.49 (  95%CI=[9.74, 11.24  The estimated power of the interaction effect for α=0.01 was 30.  00% (95%CI=[23.74, 36.86] ) using a Type-III F -test from the R package \"car,\" generated via 200 simulations with the R package \"simr.\" (The estimate for the fixed effect intercept was  23.78, 95%CI=[22.84, 24 .73], t(479.0)=49.58, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.84.) Thus, our empirical results matched our model predictions in that we observed the expected main effects of response time and threshold. We also observed the predicted interaction effect (though only at 30.00% estimated power). Participants were also asked a manipulation check question: \"How does the person feel right now?\" Participants had to choose between \"Meticulous and cautious\" and \"Careless and inattentive.\" (The locations of these options was random on every trial, which likely contributed to some mistakes.) However, participants asymmetrically answered the manipulation check questions incorrectly in Experiment 4A. When the described mental state was \"Meticulous and cautious,\" 321/480 (67%) of participants answered at least 3 out of the 4 trials correctly (the mean correct number of trials was 2.9/4). When the described mental state was \"Careless and inattentive,\" 121/480 (25%) of participants answered at least 3 out of the 4 trials correctly (the mean correct number of trials was 1.8/4). Under the preregistered exclusion criteria specified for Experiment 3, 341 participants (71%) would have been excluded for answering more than two of the eight manipulation checks incorrectly. Thus, we elected to not exclude participants based on their performance on the manipulation check (nor for any other reason). For Experiment 4B, the results from the linear mixed effects model were as follows. The estimate (beta parameter) for the main effect of response time was 4.81 (  95%CI=[3.71, 5 .91], t(3364.0)=8.54, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.02). The estimate for the main effect of threshold was 9.  03 (95%CI=[7.93, 10.14] , t(3364.0)=16.06, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.07). The estimate for the interaction effect of response time and threshold was 0.79 (95%CI=  [-1.42,2 .99], t(3364.0)=0.70, p=0.48, t-tests using Satterthwaite's method, η 2 p =0.0001): this effect was not present. The estimated power of the interaction effect for α=0.01 was 2.50% (95%CI=[0.82, 5.74]) using a Type-III F -test from the R package \"car,\" generated via 200 simulations with the R package \"simr.\" (The estimate for the fixed effect intercept was 27.66,  95%CI=[26.67, 28.64] , t(480.0)=54.94, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.86.) Thus, while we observed the expected main effects of response time and threshold, the predicted interaction effect was not present in Experiment 4B. Participants did not answer the manipulation check questions asymmetrically in Experiment 4B. When the described mental state was \"Meticulous and cautious,\" 369/481 (77%) of participants answered at least 3 out of the 4 trials correctly (the mean correct number of trials was 3.38/4). When the described mental state was \"Careless and inattentive,\" 379/481 (79%) of participants answered at least 3 out of the 4 trials correctly (the mean correct number of trials was 3.44/4). Under the preregistered exclusion criteria specified for Experiment 3, 101 participants (21%) would have been excluded for answering more than two of the eight manipulation checks incorrectly. We did not exclude these participants, however, for consistency with Experiment 4A. \n Discussion The pattern of results across Experiments 3, 4A, and 4B support our model's prediction that people can infer decision-makers' preferences based on response times in a way that is appropriately sensitive to the decision-maker's mental state. First, we observed a main effect of response time across all three experiments, even though the response time manipulation was implemented very differently for each (time for a decision-maker to visually reach for a bowl in Experiment 3, time to type a text message response in Experiment 4A, and \"After awhile, they answer A,\" or \"Without pausing, they immediately answer A,\" in Experiment 4B). The main effect of response time was strongest in Experiment 4A, followed by Experiment 4B, and weakest in Experiment 3. One difference between Experiment 3 and Experiments 4A and 4B is that Experiment 3 had a different critical question, but even if it had had the same critical question, Experiments 4A and 4B had distinct enough patterns of results from each other (response time, threshold, interaction effects) that it is hard to know how this switch would have affected the results. Regardless, we observed a significant effect of response time across Experiments 1, 2, 3, 4A, and 4B, whereby participants inferred that a decision-maker had a stronger preference for an item the less time the decision-maker spent choosing it, supporting the robustness of the effect. We also observed a main effect of threshold across Experiments 3, 4A, and 4B, in that if the decision-maker was described as cautious, participants inferred that the decision-maker had a stronger preference for their chosen item. This threshold effect was less present in Experiment 4A compared to Experiments 4B and 3, likely because the threshold manipulation was less explicitly emphasized (participants had to infer the decision-maker's mental state from Alice's description). However, we find it reassuring that a threshold manipulation as minor (and naturalistic) as that in Experiment 4A was enough to influence people's inferences about the decision-maker's preferences. We believe that the presence of both main effects-threshold and response time-across quite different paradigms speaks to the strength of people's ability to socially infer preferences from response times while taking into account decision-makers' mental states. The model also predicted an interaction effect, wherein the effect of response time on inferred value is larger for cautious vs. careless decision-makers. This interaction effect did not occur at all in Experiment 4B (p=0.48, 2.5% power), was only marginally significant in Experiment 3 (p=0.08, 19.5% power), and was significant at low power in Experiment 4A (p=0.04, 30% power). Further work will be required to uncover how people jointly reason about how careful a decision-maker is and their response times in inferring the decision-maker's preferences. Finally, manipulation check performance was strongest and symmetric across threshold conditions in Experiment 4B, worse and asymmetric in Experiment 3, and poorest and asymmetric in Experiment 4A. Experiments 4B and 3 contained explicit threshold manipulations so it is not surprising that their manipulation check performance was higher than in Experiment 4A, in which the threshold manipulation was weaker. It was also expected that Experiment 4B would have higher and more symmetric manipulation check performance compared to Experiment 3, because Experiment 4B did not contain a facial expression confound. It is surprising, however, that Experiment 4A had asymmetric performance across the threshold conditions because the stimuli were texts without a facial expression confound. One explanation is that the high threshold condition in Experiment 4A contained the word \"careful\" (which is semantically close to the high-threshold manipulation check option \"Meticulous and cautious\"), while the low threshold condition only included the word \"distracted\" (which is arguably less close to the low-threshold manipulation check option \"Careless and inattentive\"). Alternatively, perhaps our stimuli generally resulted in a bias for attributing carefulness to the decision-maker; Convincing participants that the decision-maker was being at least somewhat deliberate in their choices was a requisite property in our experiments. To conclude, the fact that we observed main effects of response time and threshold across all three experiments suggests that manipulation check performance is informative but not central to our model's predictions and results. \n General Discussion If your friend takes a long time to decide between two options, you can infer that both options have similar value to her. Making these kinds of inferences about others' preferences is ubiquitous in daily life as people seamlessly integrate choices and response times. In this work, we created a computational model to describe this phenomenon by inverting a classic generative model of decision-making, the Drift Diffusion Model. We tested the predictions of this model for three experimental questions. In Experiment 1, participants inferred a decision-maker preferred a chosen item more when they spent a longer time deciding between two items, matching model predictions. In Experiment 2, participants inferred a decision-maker's relative preferences from two choices and response times to infer a decision-maker's preference about a third unseen choice, and they inferred the model's predicted choice greater than chance across all conditions. Experiment 2 also included two control conditions in which participants only needed choice information to anticipate the decision-maker's unseen choice. As predicted, participants thought the predicted choice was more likely in these control conditions, providing an upper bound for participant likelihoods in this experiment. On the other hand, participant behavior for the different conditions was more variable than expected, which should be incorporated into the model in future work. In Experiment 3, we asked whether participants could integrate the decision-maker's mental state of cautiousness or carelessness when inferring preferences. We observed the expected main effects of threshold and response time in three experiments spanning a range of stimuli (Experiments 3, 4A, and 4B), but a more complicated pattern of results in observing the expected interaction effect. Future work should investigate how experimental manipulations of threshold and response times are entangled in inferences about decision-makers' preferences. The uniqueness of this work lies in the question asked and the inversion of the DDM to formally characterize our predictions. Specifically, we asked how people infer other's preferences by observing their response times and choices, which involves characterizing people's theory of mind surrounding response time. The work closest to this appears in a recent preprint by  Konovalov and Krajbich (2020) , who demonstrate that people use response times to make inferences about others' preferences, and further than people apply this knowledge strategically in bargaining games. Where our work differs is we formally describe these inferences about other's minds by inverting the DDM and making direct, quantitative predictions, whereas Konovalov and Krajbich (2020) make qualitative predictions based on the DDM.  (Konovalov and Krajbich (2020)  also apply their work to strategic settings, while we explore prediction for novel choices and integrating the decision-maker's mental state of cautiousness or carelessness.) Our work has several potential extensions. Previous work already touches on interesting applications in strategic gameplay and mechanism design:  Frydman and Krajbich (2018)  designed an experiment in which some participants were explicitly shown others' response times for decision-making, and showed that those participants used their response time inferences about others' preferences to achieve better task outcomes compared to participants who did not have access to response time information. Expanding upon this,  Konovalov and Krajbich (2020)  showed that participants improved their outcomes by integrating either explicitly-presented or real-time perception of response time in a bargaining game. There is a fascinating aspect of mechanism design in these studies, in that people's performance is improved by the researchers introducing additional information-response times-from which players make inferences. This aspect of mechanism design is even more prominent in  Krajbich et al. (2014) , 7 albeit Krajbich et al. (2014) did not involve participants inferring other people's preferences.  Krajbich et al. (2014)  expected based on DDM predictions that participants would spend more time on difficult choices-in which options were of similar value-even when opportunity costs suggested participants should pick one and move on. The authors designed an experiment where opportunity cost was high, and then implemented an intervention to end trials early when participants took too long to respond (an intervention they anticipated would target difficult choices in which participants were close to indifference). As predicted, after experiencing this intervention participants went on to earn higher rewards on the task on non-intervention blocks, from which the authors conclude that \"it may be possible to improve people's welfare with simple interventions or behavioral training.\" In our work, we created an inverse DDM that generates predictions about how people infer other's preferences from response times. One could imagine that this output could be a valuable tool in mechanism design: Presenting the output from such a model in strategic games, either as a training intervention or throughout the whole task, could serve as an assistive tool to improve people's performance. Integrating mechanism design for inference of others' preferences from response times in real world settings would be an intriguing application for this work. The inverted DDM captures some elements of social inference from response times, but a useful future direction would be to incorporate the idea of the overall value of choice sets, and other elements of \"context.\" For example, the DDM describes a decision-maker's difference in value between two items, but it does not incorporate the idea that choice sets can be of overall high value (e.g. buying a house, going to college) or low value (e.g. what candy bar to eat). Empirically, as the overall value of a choice set increases, response times decrease; fortunately, researchers have developed alternative sequential sampling models that can account for this effect  (Clithero, 2018b; Shevlin & Krajbich, 2020) . Incorporating \"context\" factors in inverted DDMs such as overall value, choice option similarity  (Bhatia & Mullett, 2018) , attention  (Krajbich, 2019)  and (as investigated in this work) mental states can not only help us describe and predict people's behavior, but could be used as tools to help nudge people towards more optimal use of their time. Finally, the particular inversion of the DDM in this work has applications for human-robot interaction. This work could be interpreted as inverse reinforcement learning (Jara-Ettinger, 2019;  Ng & Russell, 2000)  with algorithm run times: inferring someone's utility function based on how long it takes them to decide. Artificial decision-makers are generally quite poor at social inference, especially compared to human performance, and so incorporating formal models how people learn about others through response times will only improve communication between people and automated systems. \n Conclusion If your friend immediately says \"Yes!\" to a movie but contemplates before saying \"...yes\" to hiking, you've learned something about the relative strength of her preferences that you couldn't learn just from her choices. People constantly use response time to make inferences about each other's preferences in daily life, and in this work we sought to quantitatively capture and predict that phenomenon. We present a rational model of inferring preferences from response time, using the Drift Diffusion Model to describe how preferences influence response time and Bayesian inference to invert this relationship. We use our inverted DDM to predict participant behavior for three experimental questions. We first demonstrated that people could infer that a decision-maker preferred a chosen item more if they spent longer deliberating (Experiment 1). We then showed that people could predict a decision-maker's choice in a novel comparison based on inferring the decision-maker's relative preferences from previous response times and choices (Experiment 2). Finally we observed that people could incorporate information on a decision-maker's mental state, cautiousness or carelessness, in inferring a decision-maker's preferences from their response times (Experiments 3, 4A, and 4B). We live in a social world, and people perform a startling number of unconscious inferences in navigating others' preferences and goals. While economists and psychologists have been characterizing how we learn about others' minds from their choices for decades, capturing the more subtle inferences made from watching someone think is also essential to understanding how we are so good at knowing what is not said. In this paper we advance this goal by creating a rational model of inverted decision-making to describe these inferences from response times. Much remains to be incorporated into these models. Within our results, more remains to be investigated around how preferences are inferred if decision-makers' mental states are presented in ways that interact with cues like response time. Within DDMs, response times are influenced by other factors than just value difference, such as overall value of choice sets and many different types of \"context\" (we investigated one version of context, a decision-maker's mental state of cautiousness or carelessness). Outside of DDMs, incorporating tone in addition to choices and response times will be key to describing how people make inferences about others' preferences. Fortunately, the applications for formalizing models of social inferences are exciting and broad. People already draw impressive conclusions about others' beliefs, preferences, and future actions from their choices and response times, and fully elucidating models of these social inferences not only helps us understand the mechanisms behind such powerful computational process, but may help us help people make better choices-via mechanism RATIONAL PREFERENCE INFERENCE FROM RESPONSE TIME 52 design, or building tools based on humans' existing capacities-in the future. \n Supplemental Material Experiments 1, 2, and 3 were preregistered (https://osf.io/8n3kd) and materials and data are available at https://osf.io/pczb3. Future work may also explore some quirks of the empirical results. In Figure  6 , the results indicate that while participants did not often predict the non-predicted choice, very many participants chose 50 for the timing-inference based conditions, and 100 for the control conditions. (These numbers indicate participants' inferred likelihood for the predicted choice; \n Experiment 2, Additional Analysis the critical question participants were asked was: \"This person is now offered a choice between items B and C. What choice do you think they'd make, and how likely do you think it is that they'd make that choice?\") This pattern indicates that participants were not using response times when making their choices, and that choices alone could make them estimate 100% likelihoods for the decision-maker's choice on an unseen choice. Figure  S1  suggests that a number of participants answered in this \"50-50-100-100\" sequence (for conditions \"NeitherChosen-BothChosen-FastChoice-SlowChoice\"), implying that if such responses had been discouraged the effect observed here may have been even stronger. \n Experiment 3, Excluded Participants The analysis was identical to Experiment 3 except that participants were excluded if they answered more than two out of eight manipulation check questions incorrectly, as was defined in the preregistration. Of 481 recruited participants, 212 met this exclusion criteria, so 269 participants were included. These results were similar to the Experiment 3 results with all participants included, and are described below (Figure  S2 ). From our linear mixed model, the estimate (beta parameter) for the main effect of response time was 6.58  (95%CI=[4.53, 8.62   The mean ± SE of the inferred value difference, n=269, is shown for the high threshold (\"cautious,\" blue) and the low threshold (\"careless,\" black) conditions, for each of the 3 and 9 second response time conditions. Responses are connected by threshold condition for emphasis. Individual participants' averaged values for each threshold/time pair are also shown and connected. 0 represents the participant feeling that the decision-maker was neutral between items, 50 that the decision-maker moderately preferred their chosen item, and 100 that the decision-maker strongly preferred their chosen item. (b) Model Predictions. Figure 1 . Schematic describing the Drift Diffusion Model (DDM) and the inverted DDM used in this paper. (A) The DDM is a generative model which predicts people's response times (RT) and choices given their preferences across options. Parameters include the drift rate µ = β(u a − u b ) (value difference, or strength of preference; u a and u b are the utilities for the two items), threshold θ (carefulness), and the drift multiplier β. The stronger the preference for an item, the faster the decision. (B) In this work, we invert the DDM to generate predictions for the inference task: people inferring others' preferences based on observing their response times and choices. \n What are the person's feelings about item A and item B?Two items, A and B, are inside the bowls below. This person has just been asked to choose which item they want.Please watch the video. \n Figure 2 . 2 Figure 2 . Experiment 1 Stimulus. Participants watched a 3, 5, 7, or 9 second video of the participants choosing between A or B, then answered the question below. \n Figure 3 . 3 Figure 3 . Experiment 1 Results. (a) Experimental Results. The mean ± SE of inferred preference for the chosen item, n=415, is shown for each of the response time conditions. Responses were reoriented to be 0-50: 0 indicates the participant felt the decision-maker was neutral between items A and B, and 50 indicates the participant felt the decision-maker strongly preferred the chosen item (A or B, whichever was appropriate to the trial). Light grey lines indicate individual participants' averaged values for each time point. (b) Model Predictions. \n Hu et al. (2015)  is an exception, demonstrating that young children can perform indirect, graded transitive inference of another decision-maker's preferences based on observing their choices. Children in this study inferred that a puppet preferred object B over C (B > C), after watching the puppet choose B consistently over A (B A) and the puppet choosing C somewhat consistently over A (C > A). \n now offered a choice between items B and C. What choice do you think they'd make, and how likely do you think it is that they'd make that choice? Now they are choosing between items A and C. This person is choosing between two items, A and B. \n Figure 4 . 4 Figure 4 . Experiment 2 Stimulus. Participants watched a 3-second video of the decision-maker deciding between A and B, and a 9-second video of the decision-maker deciding between A and C. Participants then indicated whether they believed the decision-maker would be most likely to choose B or C in a novel choice. \n Figure 5 . 5 Figure 5 . Experiment 2 Conditions. Participants watched a 3-second video of the decision-maker deciding between A and B, and a 9-second video of the decision-maker deciding between A and C. In the critical question, participants indicated whether they believed the decision-maker would be more likely to choose B or C in a novel choice. The model's predictions for participants' predicted choices are shown. Four conditions were defined based on whether the decision-maker chose B or C in either of their choices (3-second \n Figure 6 . 6 Figure 6 . Experiment 2 Results. (a) Experimental Results, Mean ± SE of Inferred Likelihood Estimates for the Predicted Choice, n=478. Shown in blue are participants' averaged responses for each of the conditions; the line at 50 indicates equal likelihood between choices (chance); black points indicate individual participant averaged responses for each condition. Participants saw the decision-maker choose between A and B, then A and C, and had to infer whether the decision-maker preferred B or C in the the unseen choice (the model's predicted choices are shown).The \"NeitherChosen\" and \"BothChosen\" conditions required that participants infer the decision-maker's likelihood of preferring an item from choices and response time, while the \"FastChoice\" and \"SlowChoice\" conditions only required that participants infer from the decision-maker's choices. (b) Model Predictions. \n Figure 7 . 7 Figure 7 . Schematic describing the DDM's threshold parameter, and the inference task in Experiment 3. (A) The DDM's parameter \"threshold\" (θ) describes how careful a decision-maker is. If a decision-maker has a fixed drift rate (meaning that their perceived \n Figure 8 . 8 Figure 8 . Experiment 3 Stimulus. Participants watched a 3-or 9-second video of the decision-maker choosing between A and B then answered questions. \n Figure 9 . 9 Figure 9 . Experiment 3 Results. (a) Experimental Results. The mean ± SE of the inferred value difference, n=481, is shown for the high threshold (\"cautious,\" blue) and the low threshold (\"careless,\" black) conditions, for each of the 3 and 9 second response time conditions. Responses are connected by threshold condition for emphasis. Individual participants' averaged values for each threshold/time pair are also shown and connected. 0 \n second round of experiments (pilots, Experiments 4A and 4B). Participants within the second round of experiments could only complete one of that round's pilot experiments or main experiments. Experiments 4A and 4B were IRB-approved by Princeton University, Protocol ID: 10859, Protocol Title: Computational Cognitive Science. Participants gave informed consent. \n Figure 10 . 10 Figure 10 . Experiment 4A Stimulus. Participants watched a text message video of a decision-maker choosing between A and B quickly or slowly, then answered questions. This image shows the final frame of the video. \n each variant the participant viewed eight text message videos (Experiment 4A) or eight vignettes (Experiment 4B) varying in response time and threshold. These eight trials were presented in random order. Participants answered two questions for each of the combinations [cautious/fast], [careless/fast], [cautious/slow], [careless/slow]. After watching all eight text message videos or vignettes and answering the accompanying questions, participants had the opportunity to write comments then exited the survey. \n Figure 12 . 12 Figure 12 . Experiment 4 Results. (a) Experiment 4A (Text Message Videos), n=480, and (b) Experiment 4B (Vignettes), n=481. The mean ± SE of inferred value difference is shown for the high threshold (\"cautious,\" blue) and the low threshold (\"careless,\" black) conditions, for each of the fast and slow response time conditions. Responses were reoriented to be 0-50: 0 indicates the participant felt the decision-maker was neutral between items A and B, and 50 indicates the participant felt the decision-maker strongly preferred the chosen item (A or B, whichever was appropriate to the trial). Responses are connected by threshold condition for emphasis. Individual participants' averaged values for each threshold/time pair are also shown and connected. \n Figure S1 . S1 Figure S1 . Experiment 2 Results, Additional Analysis. (a) Main Figure (see Figure 6). (b) Individual (Connected) Participant Responses. Note the prevalence of responses following a \"50 -50 -100 -100\" pattern for the conditions \"NeitherChosen -BothChosen -FastChoice -SlowChoice.\" \n Figure S2 . Experiment 3 Results, with Excluded Participants. (a) Experimental Results. \n The estimated power of the interaction effect for α=0.01 was 19.50% (95%CI=[14.25,25.68]) using a Type-III F -test from the R package \"car,\" generated via 200 simulations with the R package \"simr.\" (The estimate for the fixed effect intercept was61.13, 95%CI=[59.69,62.59], tests using Satterthwaite's method, η 2 p =0.009). The estimate for the main effect of threshold was 11.50 (95%CI=[10.04,12.95], t(3364.0)=15.48, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.07). The estimate for the interaction effect of response time and threshold was 2.62 (95%CI=[-0.29,5.53], t(3364.0)=1.76, p=0.08, t-tests using Satterthwaite's method, η 2 p =0.0009): a marginal but not significant effect. \n ], t(1880.0)=6.31, p=4e-10, t-tests using Satterthwaite's method, η 2 p =0.02). The estimate for the main effect of threshold was 19.59 (95%CI=[17.54,21.63], t(1880.0)=18.70, p<2e-16, t-tests using Satterthwaite's method, η 2 p =0.16). The estimate for the interaction effect of response time and threshold was 3.65 (95%CI=[-0.43,7.74], t(1880.0)=1.75, p=0.08, t-tests using Satterthwaite's method, η 2 p =0.002): not a significant effect. The estimated power of the interaction effect for α=0.01 was 20.00% (95%CI=[14.69,26.22]) using a Type-III F -test from the R package \"car,\" generated via 200 simulations with the R package \"simr.\" (The estimate for the fixed effect intercept was 57.37, 95%CI=[55.67,59.08], t(268.0)=66.15, p<2e-16, t-tests using (a) Satterthwaite's method, η 2 p =0.94.) \n\t\t\t Previous work has studied response times in the framework of dual-process theory: If decisions are fast, this suggests that an intuitive process system is being used, and slow decisions suggest that a slow, deliberative process is being used (Rand et al., 2012; Rubinstein, 2007) . Here, we interpret response time differences as representing differences in strength of preferences, on the basis of other work that shown that findings interpreted through the dual-process theory lens are consistent with single-process strength-of-preference predictions (Konovalov & Krajbich, 2020; Zhao et al., 2019) . \n\t\t\t We chose this model because it is commonly used in decision-making research and is computationally easy to work with. However, the framework can accommodate any model that defines a joint distribution over choices and response times. \n\t\t\t Specifically, we translated participants' likelihood judgments to be between -50 and 50 (where 0 represented a judgement that it was equally likely the decision-maker would choose either of the two items) then did not fit the fixed effect intercept (which defaults to 0). \n\t\t\t This text was designed to isolate the threshold parameter as much as possible, and avoided e.g. mentioning differences in the absolute value of items, or response time cutoffs or pressure, for this reason. The information about having a hard day at work was included because the decision-maker appeared very focused in all video stimuli, and describing them as \"careless\" appeared unnatural otherwise. \n\t\t\t Note the locations of the manipulation check options were random for each trial, which likely contributed to some attention-based mistakes across both threshold conditions, separate from the described facial expression confound.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/GatesCallawayHoGriffiths.tei.xml", "id": "07b5e609c492a6dde1a2819ed27ff35d"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Deep learning architectures are particularly promising in this respect. For example, one approach translates a piece of malware into an image by converting code to pixels in order to utilize advances in deep learning-based image classification as a means of classifying the underlying code as benign or malicious.", "authors": ["Wyatt Hoffman", "John Bateman", "Ben Bansemer", "Andrew Buchanan", "Lohn", "Rod Thornton", "Marina Miron"], "title": "AI and the Future of Cyber Competition", "text": "Executive Summary As artificial intelligence begins to transform cybersecurity, the pressure to adapt may put competing states on a collision course. Recent advances in machine learning techniques could enable groundbreaking capabilities in the future, including defenses that automatically interdict attackers and reshape networks to mitigate offensive operations. Yet even the most robust machine learning cyber defenses could have potentially fatal flaws that attackers can exploit. Rather than end the cat-and-mouse game between cyber attackers and defenders, machine learning may usher in a dangerous new chapter. Could embracing machine learning systems for cyber defense actually exacerbate the challenges and risks of cyber competition? This study aims to demonstrate the possibility that machine learning could shape cyber operations in ways that drive more aggressive and destabilizing engagements between states. While this forecast is necessarily speculative, its purpose is practical: to anticipate how adversaries might adapt their tactics and strategies, and to determine what challenges might emerge for defenders. It derives from existing research demonstrating the challenges machine learning faces in dynamic environments with adaptive adversaries. This study envisions a possible future in which cyber engagements among top-tier actors come to revolve around efforts to target attack vectors unique to machine learning systems or, conversely, defend against attempts to do so. These attack vectors stem from flaws in machine learning systems that can render them susceptible to deception and manipulation. These flaws emerge because of how machine learning systems \"think,\" and unlike traditional software vulnerabilities, they cannot simply be patched. This dynamic leads to two propositions for how these attack vectors could shape cyber operations. The first proposition concerns offense: Attackers may need to intrude deep into target networks well in advance of an attack in order to circumvent or defeat machine learning defenses. Crafting an attack that can reliably deceive a machine learning system requires knowing a specific flaw in how the system thinks. But discovering such a flaw may be difficult if the system is not widely exposed or publicly available. To reach a hardened target, an attacker may try to compromise the system during development. An attacker with sufficient access could reverse-engineer a system during its development to discover a flaw or even create one by sabotaging the process. This opportunity to gain intelligence on an adversary's defenses creates more value in intruding into adversary computer networks well in advance of any planned attack. The second proposition concerns defense: Guarding against deceptive attacks may demand constant efforts to gain advanced knowledge of attackers' capabilities. Because machine learning systems cannot simply be patched, they must be able to adapt to defend against deceptive attacks. Yet researchers have found that adaptations to defend against one form of deception are vulnerable to another form of deception. No defense has been found that can make a machine learning system robust to all possible attacks-and it is possible none will be found. Consequently, machine learning systems that adapt to better defend against one form of attack may be at constant risk of becoming vulnerable to another. In the face of an imminent threat by an adversary, the best defense may be to intrude into the adversary's networks and gain information to harden the defense against their specific capabilities. Together these two propositions suggest machine learning could amplify the most destabilizing dynamics already present in cyber competition. Whether attacking or defending, at the top tier of operations, machine learning attack vectors may create challenges best resolved by intruding into a competitor's networks to acquire information in advance of an engagement. This would add to existing pressures on states to hack into their adversaries' networks to create offensive options and protect critical systems against adversaries' own capabilities. Yet the target of an intrusion may view the intrusion as an even greater threat-regardless of motive-if it could reveal information that compromised machine learning defenses. The already blurred line between offensive and defensive cyber operations may fade further. In a crisis, the potential for cyber operations to accelerate the path to conflict may rise. In peacetime, machine learning may fuel the steady escalation of cyber competition. Adversaries may adapt by targeting machine learning itself, including: • Compromising supply chains or training processes to insert backdoors into machine learning systems that expose a potentially wide swath of applications to possible attacks. • Poisoning training data, such as open source malware repositories, to degrade cybersecurity applications. • Unleashing risky capabilities to circumvent defenses, such as malware with greater degrees of autonomy. • Targeting defenders' trust in machine learning systems, such as by inducing systems to generate \"false positives\" by mislabeling legitimate files as malware. For the United States and its allies, harnessing machine learning for cybersecurity depends on anticipating and preparing for these potential changes to the threat landscape. If cyber defense increasingly relies on inherently flawed machine learning systems, frameworks and metrics will be needed to inform risk-based decisions about where and how to employ them. Securing the machine learning supply chain will demand collective governmental and private sector efforts. Finally, the United States and its allies must exercise caution in the conduct of their offensive operations and communicate with adversaries to clarify intentions and avoid escalation. \n Introduction States seeking competitive advantage will likely turn to artificial intelligence to gain an edge in cyber conflict. Cybersecurity ranks high among priority applications for those leading AI development.  1  China and Russia see in AI the potential for decisive strategic advantage. 2 Military planners in the United States envision systems capable of automatically conducting offensive and defensive cyber operations.  3  On the precipice of a potential collision between AI competition and cyber conflict, there is still little sense of the potential implications for security and stability. AI promises to augment and automate cybersecurity functions. Network defenders have already begun to reap the benefits of proven machine learning methods for the data-driven problems they routinely face.  4  Even more tantalizing is the speculative prospect of harnessing for cybersecurity the cutting-edge machine learning techniques that yielded \"superhuman\" performance at chess and the Chinese board game Go. Yet the machine learning capabilities fueling these applications are no panacea. These systems often suffer from inherent flaws. Unlike traditional software vulnerabilities, these flaws emerge because of how these systems make inferences from data-or, more simply, how they \"think.\" These flaws can lead even highly robust systems to fail catastrophically in the face of unforeseen circumstances. In the race between researchers developing ways to safeguard these systems and those seeking to break them, the attackers appear to be winning. What will happen when these powerful yet flawed machine learning capabilities enter into the dynamic, adversarial context of cyber competition? Machine learning can help mitigate traditional cyber attack vectors, but it also creates new ones that target machine learning itself. Attackers will systematically try to break these systems. A growing body of technical research explores machine learning attack vectors and prospective defenses. Yet there has been little effort to analyze how these changes at a technical level might impact cyber operations and, in turn, their strategic dynamics. This study approaches this problem by exploring a possible worst-case scenario: machine learning could amplify the most destabilizing dynamics already present in cyber competition. The purpose is not to lay out a case against harnessing machine learning for cybersecurity. Precisely because these capabilities could become crucial to cyber defense, the aim here is to provoke thinking on how to proactively manage the geopolitical implications of persistent technical flaws. This study explores how the attack vectors unique to machine learning might change how states hack each other's critical networks and defend their own. The unique vulnerabilities of these systems may create problems for both offense and defense best resolved by intruding into adversaries' systems in advance of an engagement. For offense, this arises from the potential need for exquisite intelligence on, or even direct access to, a machine learning system to reliably defeat it. For defense, this arises from the need for advanced knowledge of a specific attack methodology to ensure a defense's viability against it. The combination of these offensive and defensive imperatives could exacerbate the escalation risks of cyber engagements. States would have even stronger incentives to intrude into one another's systems to maintain offensive options (for contingencies such as armed conflict or strategic deterrence) and to ensure the viability of their own defenses. Yet it may be even harder to differentiate cyber espionage from intrusions laying the groundwork for an attack; the target of an intrusion may assume that it is preparation for an imminent attack, or that it will at the very least enable offensive options. As adversaries struggle to gain an edge over one another, the line between offense and defense-tenuous as it already is in cyber operations-fades. This dynamic may fuel the steady drumbeat of cyber competition in peacetime. In a crisis, the potential for misinterpretation of a cyber operation to trigger escalation may rise. This forecast rests on two core assumptions that must be addressed at the outset. These are certainly debatable but the aim is to analyze their implications should they hold, not assess how likely they are to do so. The first is that machine learning could plausibly deliver on the promise of sophisticated, automated cyber defenses at scale. That is, the significant technical and practical hurdles (e.g., demands for high quality data and computing power, as well as organizational challenges to implementation) will not prove insurmountable at least for top-tier actors such as China and the United States. This study begins with a survey of applications in various stages of development to demonstrate their plausibility. But it makes no attempt to assess the current state of play with deployed machine learning cybersecurity applications or the likelihood of realizing them in the near term. * * For a more thorough survey of existing applications and near-term prospects for machine learning in cybersecurity see Micah Musser and Ashton Garriott, \"Machine Learning and Cybersecurity: Hype and Reality\" (Center for Security and Emerging Technology, forthcoming). The second assumption is that insights from existing research on machine learning attack vectors will hold at least for the prevailing machine learning methods and applications discussed here. This study draws extensively on research demonstrating the attack vectors targeting machine learning and what these vectors reveal about the potential limits of the robustness of machine learning systems. It makes no assumptions about yet unseen innovations in machine learning techniques or offensive or defensive measures that might fundamentally change the trajectory. This study begins with a brief overview of machine learning applications for cybersecurity, including their prospective defensive benefits and inherent flaws. It then examines two propositions for how these technical changes to the cybersecurity landscape may, in turn, shape offensive and defensive cyber operations. Specifically, machine learning attack vectors could create predicaments that incentivize intrusions into adversaries' networks, whether to create offensive options or shore up defenses. This study continues on to explore how the combination of these two propositions could fuel the steady intensification of cyber competition and increase the risks of misperception and escalation in cyber engagements. \n Promise and Pitfalls of Artificial Intelligence for \n Cybersecurity Machine learning lies at the core of the emerging and maturing cybersecurity applications discussed throughout this paper. Described as an approach to, or subfield of, AI, machine learning has fueled recent milestones in tasks ranging from image recognition to speech generation to autonomous driving. Machine learning systems essentially adapt themselves to solve a given problem.  5  This process often starts with a blank slate in the form of a neural network. The system's developers feed a dataset to the neural network and an algorithm shapes the network's structure to adapt to the patterns within this data. For example, a system for analyzing malware will learn to accurately identify a file as \"malware\" or \"benign\" and associate each classification with particular patterns. Eventually the network develops a generalized model of what malware \"looks like.\" High quality training data, effective training algorithms, and substantial computing power comprise the critical inputs to this process. The resulting machine learning model, ideally, detects not only known malware but yet unseen variants. Advancements in machine learning techniques reduce the need for human experts to structure data. * Rather than relying on an expert to tell the model what key features of malware to look for, the model discovers on its own how to classify malware. As a result, it may find ways of identifying malware more effective at coping with attackers' attempts at obfuscation, such as \"metamorphic\" malware that rewrites parts of its code as it propagates.  6  Intrusion detection-finding an adversary's illicit presence in a friendly computer network-may benefit similarly from machine learning. Existing intrusion detection systems already look for red flags, such as a computer becoming active in the middle of the night or a user attempting to access files unrelated to their work. Yet defenders struggle to sort through the vast data generated by network activity in large enterprises, allowing attackers to hide in the noise. Machine learning systems can turn this data into a major advantage. By fusing information from a wider and more diverse range of sensors throughout the environment, they create a baseline of normal network activity against which even slight deviations can be detected.  7  AI and machine learning have quickly become buzzwords in the cybersecurity industry. This makes it difficult to assess the extent to which these capabilities are actually relied upon or are invoked for marketing purposes. Cybersecurity vendors commonly claim to leverage machine learning. * For example, as CrowdStrike defends its customers' devices and networks, it rakes in data on around 250 billion events daily and feeds the data to machine learning models to predict new kinds of attacks.  8  Darktrace states that it employs multiple machine learing methods in its \"Enterprise Immune System,\" empowering systems that can automatically mitigate attacks. 9 Machine learning has also been harnessed to test software for vulnerabilities, detect spam and spear-phishing attacks, and identify suspicious behavior and insider threats.  10  In general, machine learning systems appear to be deployed mainly for relatively narrow tasks in support of human network defenders.  11  Traditional machine learning methods relying on large training datasets may not suffice for a system that performs more complex tasks requiring sequences of actions, each dependent upon the outcome of the last. Such a system needs to learn more like a human-through experimentation and trial-anderror. This is the essence of reinforcement learning. Instead of being fed training data, a reinforcement learning agent interacts with a simulated environment and is rewarded for action that advances its objective. It gradually learns sets of moves, or \"policies,\" to guide its action. The process can yield stunning results, such as the victory by AlphaGo, a reinforcement learning system developed by DeepMind, over Lee Sedol, the world champion in the incredibly complex game of Go.  12  If reinforcement learning can master chess and Go, it might unlock future cyber defenses capable of discovering and automatically executing moves and strategies in the \"game\" against cyber attackers. Cyber defenders have a home field advantage.  13  They can change the configuration of networks to interfere with an attacker or deploy decoy systems such as \"honeypots\" that lure attackers in and lead them to reveal capabilities. However, setting up * According to one survey of U.S., UK, and German businesses 82 percent of respondents stated their company employed a cybersecurity product utilizing machine learning in some form. honeypots and reconfiguring networks are technically demanding tasks and, to be effective, require the ability to anticipate an attacker's moves and adapt on the fly.  14  While still largely confined to academic research, and thus more speculative, pioneering applications of reinforcement learning may produce systems capable of these feats.  15  Reinforcement learning agents could learn optimal strategies for reconfiguring networks and mitigating attacks, rapidly analyze an attacker's moves and select and execute actions, such as isolating or patching infected nodes and deploying honeypots. At a minimum, they could present attackers with a constantly moving target, introducing uncertainty and increasing the complexity required for offensive operations.  16  Machine learning could plausibly deliver on the promise of cyber defenses that adapt to novel threats and automatically engage attackers. These potentially game-changing applications are the focus of this study, even though the most significant near-term gains for cybersecurity may be found in automating the more \"mundane\" aspects of cybersecurity. The more speculative capabilities may not be realized in the near term, but given their potential to transform cyber operations it is worth exploring their implications. \n Security vulnerabilities of machine learning As promising as they are, most machine learning cyber capabilities have yet to face the most important test: systematic attempts by attackers to break them once deployed. Machine learning can fail catastrophically under certain conditions.  17  Evidence for this includes \"adversarial examples\": manipulated inputs (often images that have been subtly altered) created by researchers to trick machine learning models. Seemingly imperceptible changes to an image of a turtle can cause a model that otherwise classifies it with perfect accuracy to mistake it for a rifle.  18  Similar adversarial techniques can cause reinforcement learning systems to malfunction.  19  Adversarial examples reveal a problem inherent to machine learning, not just deficiencies in specific systems. Every model rests on assumptions about the data to make decisions-assumptions, for instance, about what malware \"looks like.\" If an input violates those assumptions it will fool the model (and often a successful deception fools other models trained for the same task).  20  Flawed training methods or data can create vulnerabilities. But models can also become vulnerable when the conditions in which they are deployed change in ways that violate assumptions learned in training. The model's predictions will no longer be accurate-a problem referred to as \"concept drift.\"  21  Even slight deviations from training conditions can dramatically degrade performance. This poses a constant problem for machine learning applications in dynamic, adversarial contexts like cybersecurity.  22  For machine learning cyber defenses to be viable, they may have to learn and evolve not just during training, but in deployment.  23  Systems will have to keep up with a constantly changing cybersecurity landscape. For instance, an intrusion detection system modeling \"normal\" network activity must constantly revise this model as legitimate and malicious activity changes. The system might generate new training data by observing the behavior of devices connected to the network, using this data to continuously update and refine its model to better predict future behavior. Innovative machine learning techniques aim to create systems capable of better contending with adaptive adversaries in dynamic environments. These techniques harness competition to drive evolution. For instance, Kelly et al. co-evolve defenses that automatically reconfigure networks to catch the attackers with offensive agents seeking to evade detection.  24  Developers may pit a reinforcement learning agent against an adversarial agent whose objective is to thwart it.  25  These methods attempt to simulate an \"arms race\" between attackers and defenders to produce models that better anticipate and preempt attacker moves in the real world.  26  All of this sets the stage for a potential transformation in the cat-and-mouse game between cyber attackers and defenders. The future cybersecurity playing field may feature defenses that evolve automatically through engagements, but such defenses inevitably create new attack vectors that are difficult to safeguard. The next two sections explore how attackers and defenders alike might adapt to these changing technical conditions, setting the foundation to examine the geopolitical implications that follow. \n The Imperatives of Offense If improved machine learning defenses offer significant benefits to defenders, they will introduce significant new hurdles into the planning and execution of offensive cyber operations. Offensive operations often require careful planning and preparation of the target environment. The presence of sophisticated machine learning defenses may force attackers to shift their efforts toward targeting the underlying machine learning models themselves. But hacking machine learning presents its own unique set of problems. The core challenge for attackers will be figuring out how to reliably manipulate or circumvent these systems. \n Attacking machine learning Attackers tend to follow the path of least resistance. If possible, they will try to avoid machine learning defenses entirely, including by targeting \"traditional\" attack vectors, such as acquiring credentials via spear-phishing. Avoidance, however, may not always be an option. An attacker may attempt to evade the defensive system by exploiting a weakness in the model. Researchers at security firm Skylight Cyber demonstrated how to do so against Cylance's leading machine learning-based antivirus product.  27  Using publicly accessible information, they reverse-engineered the model to discover how it classified files. In the process, they discovered a bias in the model; it strongly associated certain sequences of characters with benign files. A file that otherwise appeared highly suspicious would still be classified as benign if it contained one of the character sequences. The Skylight researchers discovered, in their words, a \"universal bypass\"-characters that they could attach to almost any piece of malware to disguise it as a benign file. * The researchers found that applying their bypass to a sample of 384 malicious files resulted in the machine learning system classifying 84 percent as \"benign,\" often with high confidence. 28 Attackers will not always be so lucky as to discover a bypass as readily exploitable as in the Cylance case. They could sabotage a model to similar effect. Injecting bad samples into a training dataset (e.g. malware labeled as \"benign\") can \"poison\" a model. Even an unsophisticated poisoning attack could dramatically reduce the model's performance.  29  More insidiously, an attacker could poison a model so that it reacts to specific inputs in a way favorable to the attacker-inserting a \"backdoor\" into the model. In one demonstration, researchers created a \"watermark\" in the form of a specific set of features in a file that functioned similar to the bypass discovered by Skylight. By tampering with just one percent of the training data, they could induce a model to misclassify malicious files containing the watermark as benign with a 97 percent success rate.  30  While these examples describe attacks on classification systems, reinforcement learning agents engaged in more complex tasks have similarly proven susceptible to evasion and sabotage.  31  For example, an attacker could poison a defensive system that automatically reconfigures networks so that it responds poorly in specific circumstances; the attacker might trick the system into connecting an infected node to others in the network, rather than isolating it.  32  The attacker's predicament The feasibility of evading or poisoning a machine learning system will inevitably depend on the context. It's one thing to demonstrate attacks on machine learning in experimental settings, but it's another to execute them in the real world against a competent defender. In the Cylance case, the attackers benefited from insights into the inner workings of the model. States seeking to create and sustain offensive options may face strategic targets that are not so widely exposed. The difficulty of conducting attacks on machine learning systems under realistic constraints may pressure states to intrude into adversaries' networks to begin laying the groundwork for attacks as early as possible. This pressure stems from the necessity intrusions play in enabling the kinds of attacks described above: (1) Acquiring information to craft more reliable and effective evasion attacks against machine learning systems: As Goodfellow et al. observe, the greater the attacker's \"box knowledge\"-knowledge of the target model parameters, architecture, training data and methods-the easier it is to construct an attack that defeats the system.  33  Under \"white box\" conditions, where the attacker has complete knowledge, crafting an attack is a relatively straightforward matter of optimizing the features of malware (or other inputs) to exploit the model's assumptions. \"Black box\" attacks, where the attacker has little to no knowledge of the target model, are possible, but require more guesswork. The attacker may engineer an attack against a substitute for the target model in the hopes that if it fools the substitute, it will fool the target. But this depends on how closely the substitute matches the target.  34  Demonstrations of black-box attacks often leverage publicly available details or the ability to repeatedly probe a target model in order to derive information on how it works. An attacker might buy a commercial service to gain insights into a model, allowing greater flexibility to craft attacks. In top-tier cyber competition, however, an attacker may not enjoy these advantages. If the target model is not widely exposed, attempts to probe it may tip off the defender. And gaining information on some types of defenses, like those that reconfigure networks, would require intruding into the network. Moreover, future security measures may prevent deployed machine learning systems from \"leaking\" useful information to an attacker attempting to probe them.  35  The best way to acquire box knowledge, then, may be to gain access to a training environment and steal training data or even a trained model. (2) Compromising systems to enable future exploitation: It is possible to undermine a deployed model, for instance interacting with an intrusion detection system to \"normalize\" an intruder's presence to it.  36  But competent defenders will be alert to the possibility. The development process may present a softer target.  37  Rather than a model developed from scratch, many applications take existing pre-trained models and tailor them for specific tasks through additional training and fine-tuning in a process known as transfer learning. A backdoor inserted into the pre-trained model can make its way into subsequent models derived from it.  38    \n The Imperatives of Defense As attackers adapt to the deployment of machine learning, the success or failure of cyber defenses may hinge on the security of machine learning models against deception and manipulation. Yet it has proven difficult to create machine learning systems that are truly robust-that is, systems that can contend with attackers that adapt their tactics to try and defeat them. Innovative defenses against the kinds of attacks described above have emerged, but are routinely broken. Some experts question whether progress toward truly robust machine learning has been illusory.  39  The core challenge for defenders may be safeguarding systems with inherent flaws baked in. \n The perpetual problem of machine learning robustness When a vulnerability is discovered, a machine learning model cannot simply be patched like traditional software. Instead, the developer must retrain the model using adversarial examples or certain training procedures designed to make the model more robust to a particular set of deceptive inputs. However, adjusting the model may simultaneously make it more robust to one set of deceptions but more susceptible to others. Two prominent machine learning security experts, Ian Goodfellow and Nicolas Papernot, thus characterize existing defensive measures as \"playing a game of whack-a-mole: they close some vulnerabilities, but leave others open.\"  40  Such were the findings of Tramer et al., who systematically defeated 13 defenses shown to be effective against adaptive attackers. 41 A similar phenomenon has been observed with reinforcement learning agents; rather than becoming generally robust, those trained against an adaptive adversary in simulated games tend to \"overfit\" to the adversary. In other words, their adaptations to deal with the regular opponent can leave them vulnerable to a novel attack.  42  The ease with which defenses are broken may simply reflect the nascent state of machine learning security. But it suggests a more concerning possibility: no defense will be robust to all possible attacks. As David Brumley puts it: \"for any ML algorithm, an adversary can likely create [an attack] that violates assumptions and therefore the ML algorithm performs poorly.\"  43  Unlike software security, which is, at least in theory, a \"linear\" process of improvement as the developer tests, patches, and repeats, machine learning may present a perpetual security problem. The system can be hardened to any known attack but may always be vulnerable to a possible novel attack. These observations raise two questions regarding the potential limits on machine learning robustness: First, how much of a problem do machine learning's flaws pose for the defender? With sufficiently comprehensive training data to accurately model threats, perhaps the risk of a novel attack defeating the system would be negligible. However, cybersecurity presents a uniquely difficult deployment context: Threats continuously evolve, so a deployed system must constantly take in new data to adapt. But if instead of becoming generally robust, machine learning defenses are just playing whack-a-mole, then there may always be an attack that breaks them. Testing systems to try and discover every flaw may prove futile because of the vast range of possible moves the attacker could make to deceive the machine learning model.  44  And attackers may be in a position where they could feasibly discover flaws by repeatedly probing defenses, unlike other domains where engagements between attackers and defenders might be episodic (e.g. autonomous weapon systems in kinetic warfare). * Second, is this problem endemic to machine learning or a limitation of prevailing methods? It is at least possible that the limits on robustness prove persistent in contexts where systems have to evolve with adaptive adversaries. The process of neural network evolution drives toward efficient solutions to problems, not necessarily solutions that are robust against adaptive adversaries. In the Cylance case, the system discovered an efficient way to classify the whitelisted files-but one that attackers could exploit. This may not matter in some contexts, but systems forced to co-evolve with adaptive adversaries may adapt in ways that inevitably create vulnerabilities. Colbaugh and Glass thus argue that systems that co-evolve with adaptive adversaries become \"robust yet fragile.\"  45  They become effective at dealing with recurrent threats but, in adapting to do so, develop \"hidden failure modes\" that a novel attack could trigger. Consequently, they argue, prospective mitigations like \"ensemble\" models, which combine multiple algorithms in a model to minimize the consequences of any one failing, may not yield truly robust systems because they do not resolve the underlying problem. To be clear, it is too early to draw definitive conclusions. The point is that applying machine learning to cybersecurity presents a set of intertwined challenges. At a minimum, defenders will have to ensure their systems keep up with constantly evolving threats. But the same capabilities that enable systems to adapt may put them at risk of being \"mistrained\" in ways that leave them vulnerable to targeted attacks. And if it is possible that there are inherent limits on robustness, defenders could be forced to make tradeoffs between different threats. \n The defender's predicament Machine learning may solve some long-standing problems for defenders while creating new ones. In many contexts, defenses sufficient to deal with that vast majority of malicious threats will be good enough. States, however, need to ensure the viability of defenses against not just general malicious activity, but specific pacing threats (e.g. China or Russia in the United States' case). The possibility of an adversary exploiting a hidden failure mode in a defense may become an acute concern. Yet states may have limited options for ensuring the robustness of defenses, each of which may necessitate intruding into their adversaries' (or third parties') networks before an attack occurs: (1) Overcoming the limitations of training, testing and verification: Generally speaking, knowledge of adversaries' capabilities enables proper threat modeling and hardening of defenses. Machine learning could amplify the benefits of insights into the evolving threat landscape-and the potential costs of falling behind the latest trends. The better the training data on attacks are, the better the defensive model against those attacks will be. Historical data will diminish in value as adversaries change tactics and the landscape shifts, creating a constant incentive to continually gather information on evolving adversary tactics. Moreover, these incentives could be even stronger if there are inherent limits on the robustness of machine learning defenses. The defender may have to choose a subset of potential attacks to prioritize when training a defense within a vast range of possible attacks.  46  Verifying the system's robustness against a specific adversary might depend on anticipating their likely attack methodology. Intruding into the adversary's networks (or a third-party network that adversary may be operating inside) to gain advanced warning of their capabilities could thus guide the defender's efforts and make this problem far more tractable. (2) Enabling countermeasures to a specific adversary's attacks: A defender can painstakingly try to harden a defense against the vast range of possible attacks. But a much simpler option may exist: peer into the attacker's own networks to gain the information necessary to mitigate an attack through traditional cyber defense. This could include discovering and patching a software vulnerability used by the attacker or creating a signature of malware in order to detect it, essentially \"inoculating\" the defense. This would have the added benefit of scalability; a defender could inoculate defenses deployed in a range of settings rather than having to orchestrate their retraining. * Rapidly inoculating defenses might be especially necessary in a period of heightened tensions when an attack by an adversary is anticipated. (3) Leapfrogging the innovations of others: Unlike experimental settings that typically feature one attacker and one defender, cyberspace features many actors who learn from and appropriate others' tools and techniques. With cybersecurity in general, a state can expect its adversaries to adapt and improve their capabilities against other states' defenses. The fact that attacks tend to transfer from one machine learning model to another suggests that observing successful attacks against another's defenses can yield specific, valuable information on how to improve one's own. A state might even probe another actor's defenses to try and extract the model and copy it for its own defense. * U.S. Cyber Command's \"malware inoculation initiative,\" which publishes information discovered on adversaries' capabilities to improve private sector defenses, demonstrates the potential scalability of this approach. Erica Borghard and Shawn Lonergan, \"U.S. Cyber Command's Malware Inoculation: Linking Offense and Defense in Cyberspace,\" Net Politics, April 22, 2020, https://www.cfr.org/blog/us-cyber-commands-malware-inoculationlinking-offense-and-defense-cyberspace. \n Artificial Intelligence and Cyber Stability Artificial intelligence could transform cyber operations at a time when cyber competition among states is heating up. This analysis has focused on the potential operational imperatives machine learning could create, but these operations would not play out in a vacuum. They would occur within this strategic context, in which states may be both \"attackers\" and \"defenders\" in a constant struggle for advantage. The stakes are no less than protecting core national interests and potentially crucial military advantages in a conflict. Cyber competition may drive states to hack machine learning defenses. Could machine learning, in turn, destabilize cyber competition? The escalation dynamics of cyber engagements remain a subject of contention. Real-world cyber operations have rarely provoked forceful responses.  47  This has led some scholars to propose that inherent characteristics of cyber capabilities or cyber competition limit the potential for escalation. Others are less sanguine. Jason Healey and Robert Jervis argue that cyber competition has steadily intensified as the scope and scale of cyber operations have expanded over three decades.  48  The forces containing this competition to manageable thresholds may not hold indefinitely. Moreover, they argue that even if cyber operations can be stabilizing in some circumstances, in a crisis their characteristics could accelerate the path to conflict. Cyber competition already has the ingredients needed for escalation to realworld violence, even if these ingredients have yet to come together in the right conditions. The aim here is simply to show how machine learning could potentially amplify these risks. This follows two of the potential escalation pathways Healey and Jervis identify. The first concerns the factors fueling the steady intensification of cyber competition, which could eventually cross a threshold triggering a crisis. The second concerns the characteristics of cyber operations that may pressure states to launch attacks in a crisis. (1) Machine learning could fuel the intensification of cyber competition. Even as states' cyber operations have become more aggressive in some respects, they have largely remained well below the threshold likely to trigger retaliation. The vast majority consist of acts of espionage and subversion in the \"gray zone\" between war and peace. Some attribute this apparent stability to dynamics governing cyber competition below the use of force that are inherently self-limiting.  49  But Healey and Jervis argue that this stability may be tenuous. In some conditions, cyber competition leads to \"negative feedback loops\" that diffuse tensions. In others, it can lead to \"positive feedback loops,\" whereby cyber operations by one state incite operations by another.  50  Positive feedback can occur when cyber operations generate fears of insecurity. A state may intrude into another's networks simply to maintain situational awareness or to secure its own networks against the target's offensive capabilities. But because the same intrusion for espionage could pave the way to launch an attack, the target of the intrusion may view this as offensive and respond by engaging in their own counter intrusions. * How might machine learning change these dynamics? The above analysis of offensive and defensive imperatives suggests the potential to amplify positive feedback loops in three ways: First, machine learning may increase the perceived salience of informational advantages over an adversary and the fear of falling behind. Offensive operations targeting machine learning attack vectors may have to be tailored to the precise defensive configuration. † Defending against such attacks may require the ability to anticipate the particular deception created by the attacker. The resulting strategic dynamic may resemble the game of poker: Your best move depends on what your opponent has in their hand. Whatever can be done in advance to figure out the opponent's hand-or \"stack the deck\"-may prove tempting. Second, machine learning may incentivize states to conduct intrusions into adversaries' networks even earlier in anticipation of future threats. Whether attacking machine learning systems or defending against such attacks, the options with the greatest chance of success may also require the earliest action. Reaching an isolated target may necessitate sabotaging a machine learning defense before it is deployed if a black-box attack would be infeasible. Similarly, hardening a defense against an attack may require gaining information on an attacker's capabilities well before they are launched. States tend to hedge against uncertainties. They may be forced to make decisions to take action in the present based on possible future * This dynamic, whereby one state's actions to secure itself create fear in another, raising the potential for misinterpretation and escalation, is similar to the political science concept of the security dilemma. For an overview of the security dilemma and its application to offensive or defensive needs. The result may be to lower the threshold of perceived threat sufficient to motivate such action. Third, machine learning may further blur the line between offensive and defensive cyber operations. If merely interacting with a defensive system could extract information needed to engineer an attack to defeat it, states may be prone to view any interaction as possible preparation for an attack. Similarly, a state may gain access to a training environment to copy a defensive model, but the target may fear the model has been reverseengineered and fatally compromised, enabling an attack. In short, states may perceive that the stakes of gaining an edge over adversaries are rising, requiring even more proactive efforts in anticipation of future needs, while simultaneously making the same efforts by adversaries seem even more threatening. In the right conditions, positive feedback loops may become more likely to cause an engagement to cross a threshold triggering a crisis. More predictably, these dynamics might motivate risky or destabilizing cyber operations by states-particularly those seeking asymmetric advantages and willing to tolerate collateral damage. Several concerning scenarios stand out: • Systemic compromises: Contractors or open source projects may present opportunities to insert backdoors into models that make their way into harder to reach targets. The danger of such operations is that a systemic compromise could leave a wide swath of civilian and governmental applications vulnerable. Malware designed to exploit the backdoor could inadvertently propagate to other systems. As with any backdoor inserted into a product, there is no guarantee another malicious actor could not discover and exploit it. • Poisoning the waters: A cruder tactic than inserting a backdoor would simply be an indiscriminate attempt to degrade cybersecurity applications. An attacker with little regard for collateral damage might flood a malware repository with tainted samples designed to mistrain machine learning systems relying on the data. • Reckless operations: States may be tempted to accept certain operational risks to circumvent machine learning defenses. For instance, an attacker may employ capabilities with greater autonomy to avoid reliance on external command and control servers, which would risk detection.  51  Absent human control, such capabilities might carry greater risk of unintended impacts that spread beyond the target network. An attacker might also sabotage a defense to create an offensive option that unintentionally exposes the targeted network to other attackers. Sabotaging the systems that protect an adversary's critical infrastructure, for instance, might backfire catastrophically if it creates an opportunity for a third party to launch an attack and trigger a crisis. • Attacks on trust: An attacker might not need to break a machine learning defense if they can undermine the defender's confidence in it. A case alleged against the cybersecurity vendor Kaspersky illustrates the possibility. In 2015, the company was accused of uploading fake malware samples to VirusTotal, an open source service that aggregates information from cybersecurity vendors to improve collective defenses. The fake samples were designed to cause competing antivirus systems to flag legitimate files, creating problems for clients and potentially hurting their brands.  52  Manipulating a machine learning system to trigger false positives could similarly undermine confidence in the model. (2) Machine learning could exacerbate the characteristics of cyber operations that undermine crisis stability. In some cases, cyber operations might help avoid a crisis by diffusing tensions. * However, if a crisis breaks out, cyber capabilities create pressures that could accelerate the path to conflict. Healey and Jervis note the widespread perception that cyber capabilities have maximal effect when the attacker has the benefits of surprise and initiative.  53  If conflict appears imminent, such first-mover advantages might tempt states to launch preemptive cyberattacks against command, control, and communications capabilities to degrade or disable an adversary's military forces. Short of actually launching an attack, states would have strong incentives to begin preparations to do so by intruding into their opponent's networks. The inherent ambiguity of cyber intrusions creates a recipe for misperception in such a context. Intrusions for espionage purposes may appear indistinguishable from those laying the groundwork for attacks, or \"operational preparation of the environment\" (OPE). As Buchanan and Cunningham argue, this creates the potential for escalation resulting from a * Cyber operations could act as a \"pressure valve\" by creating options to respond to provocations that are potentially less escalatory than kinetic force both in their direct impacts and impacts on perceptions. miscalculated response to an intrusion.  54  One side might discover an intrusion in a crisis-even one that occurred months before the crisis began-and, misinterpreting it as an imminent attack, may face sudden pressure to launch a counterattack. Here again, the potential offensive and defensive imperatives created by machine learning could exacerbate these risks: First, states may feel even greater pressure to gain advantages through intrusions early in a crisis. The time needed to engineer an attack under blackbox conditions, or retrain a defense to ensure robustness against a possible imminent attack by an adversary, may translate to increased pressure to try and quickly gain information on an adversary's capabilities if the state does not already possess it. * Second, the indistinguishability of espionage from OPE may be even more problematic. A state that detects an intrusion or a compromised training process might have no way to rapidly discern whether the attacker has discovered a flaw that would defeat the system or to evaluate the robustness of the defense. If defenses are believed to be fragile in the face of novel attacks this could become an acute source of anxiety. Faced with fewer options to rule out worst-case scenarios, the state may be more likely to escalate in response. Third, machine learning could create additional sources of uncertainty that induce potential \"use it or lose it\" dilemmas. Changes in the target environment can already throw off meticulously-planned offensive operations. The shelf-life of an offensive operation might be even shorter if it must be tailored to the precise configuration of machine learning defenses that could evolve over time. If a state has prepared such capabilities, the temptation in a crisis may be to use them rather than risk them becoming obsolete. The sudden discovery of a critical flaw in a defensive machine learning system with no easy remedy might similarly force the defender to contemplate whether to preempt a possible attack. The threat to crisis stability arises from this unique combination of uncertainties and anxieties at the technical and strategic levels. Machine learning seems capable of compounding these and, in the heat of a crisis, increasing the potential for serious misperception and miscalculation. \n Mitigating scenarios As stated at the outset, this study explores a possible worst-case scenario for the future of AI-cyber capabilities. The threat to stability stems from the potential for machine learning to create offensive and defensive imperatives that incentivize states to intrude into their adversaries' networks. But it is worth briefly revisiting the possibility that machine learning could evolve in ways that fundamentally change these imperatives. Describing the current state of the art, Bruce Schneier compares machine learning security to the field of cryptography in the 1990s: \"Attacks come easy, and defensive techniques are regularly broken soon after they're made public.\"  55  The field of cryptography, however, matured and encryption is now one of the strongest aspects of cybersecurity. Eventually, a more mature science of machine learning security may likewise yield systems highly robust without the constant threat of becoming vulnerable to targeted attacks. However, as this relates to the incentives to intrude into adversaries' networks, it only solves the defensive side of the equation. A machine learning defense could be robust to an adversary's attack even without advanced knowledge of their capabilities. On the other hand, attackers may face even greater incentives to intrude into target networks early and aggressively. If attackers cannot count on discovering ways to defeat a system once it is deployed, sabotaging its development or compromising its supply chain may be seen as even more necessary offensive options. Alternatively, machine learning security may hit a dead end. Systems may remain fundamentally vulnerable in dynamic, adversarial conditions. In such a scenario, cybersecurity would in all likelihood still benefit from machine learning applications as it does now, but not in ways that fundamentally change the cat-and-mouse game. In this case, offensive operations may not depend on early intrusions any more than in the status quo; attackers would likely be able to find ways to defeat defenses that do not depend on compromising them well in advance. Defenders, however, might face stronger pressure to intrude into adversaries' networks to try and harden potentially fragile systems against their capabilities. The situation for defenders could become untenable if attackers benefit from offensive applications of machine learning.  56  The point of this cursory analysis is to show that even if the broad trajectory of machine learning changes, the potential for destabilization may remain. In any event, the present challenges facing machine learning do not appear likely to be resolved soon. To Schneier, machine learning security is at the level of maturity of cryptography in the 1990s, but Nicholas Carlini, a leading expert on machine learning security, paints an even bleaker picture. In a November 2019 presentation, he compared machine learning security to cryptography in the 1920s, suggesting that the field has not even developed the right metrics to gauge progress toward solving these fundamental problems.  57   \n Implications for policy Efforts are underway to understand and address the threats to machine learning systems.  58  A key takeaway from this study is that deploying machine learning for cybersecurity presents a unique set of challenges arising from the interaction of technical characteristics and strategic imperatives. These challenges must be addressed not only via technical solutions but at the level of policy and strategy. Even with the \"known-unknowns,\" several imperatives emerge from this forecast: • First, machine learning may present inexorable tradeoffs for cyber defense. Machine learning defenses may mitigate some threats while introducing new attack vectors. And the ability to adapt to evolving threats may put systems at constant risk of becoming vulnerable. Decision-makers need basic tools to inform risk-based decisions about when and how to employ such systems. This includes frameworks and metrics to evaluate systems deployed in crucial contexts: e.g., diagnosability or auditability, resilience to poisoning or manipulation, and the ability to \"fail gracefully\" (meaning a model's failure does not cause catastrophic harm to functions dependent upon it).  59  Decisions and policies, such as those regarding the disclosure of machine learning vulnerabilities or the publication of offensive security research that might enable attackers, will also have to be adapted to the unique characteristics of machine learning. * • Second, secure deployment of machine learning depends on guarding against attempts by adversaries to broadly compromise or degrade the development process. Attacks will not always be direct; adversaries may exploit trust in common services, like the aforementioned case involving VirusTotal. They may further blur the lines between threats such as industrial espionage and strategic sabotage. Given the premium on \"box knowledge,\" threats to the confidentiality of public or private data and algorithms should be treated as threats to the integrity of applications. Securing machine learning demands collective efforts by the government and private sector to secure the supply chain, open source development projects, data repositories, and other critical inputs.  60  • Third, managing tension and avoiding escalation in the conduct of cyber espionage and offensive operations will become even more important-especially as the imperative to gain information on adversaries' offensive capabilities and their own machine learning defenses increases. Operators need to understand the potential impacts of operations against machine learning in sensitive contexts, and will need to understand how adversaries will perceive their operations. If machine learning could amplify positive feedback loops it is worth examining the broader implications for U.S. cyber strategy, which is premised on the stabilizing effects of \"persistent engagement\" with adversaries.  61  Communication with adversaries to clarify strategic intentions will help avoid misinterpretation. Further, now is the time to explore forms of mutual restraint regarding the most destabilizing offensive activities targeting machine learning. \n Conclusion The pressure to harness artificial intelligence to deal with evolving offensive cyber capabilities will only grow. Precisely because machine learning holds both promise and peril for cybersecurity, a healthy dose of caution is needed when embracing these capabilities. Decisions made now with respect to the development and adoption of machine learning could have far-reaching consequences for security and stability. Decision-makers must avoid having to relearn lessons from cybersecurity in general, including the pitfalls of overreliance on defenses at the expense of a more holistic approach to risk management. The stakes of securing machine learning will rise as it is incorporated into a wide range of functions crucial to economic and national security. The incentives to gather intelligence or even sabotage the development of defensive systems might weigh just as heavily with other strategic areas of application. If this competition is not managed, states may head down a path destructive for all. This opens up new attack vectors. For example, compromising an open source project, code repository, or a commercial contractor assisting with the development of cybersecurity applications may allow an attacker to insert vulnerabilities deep into systems that make their way into more tightly-controlled training environments. Targeting the development process has the added benefit of scalability: inserting a backdoor into one model may facilitate access to a wide swath of subsequent targets. A transfer learning service supporting diverse commercial, military, or other national security-relevant applications would be a tempting target. Models via Reinforcement Learning,\" arXiv [cs.CR] (January 26, 2018), arXiv, http://arxiv.org/abs/1801.08917; Huang et al., \"Adversarial Attacks on Neural Network Policies.\" \n Machine learning security would benefit, for instance, from standards analogous to the Common Vulnerability Scoring System, used to evaluate the severity of software vulnerabilities and inform decisions about patching. \"Common Vulnerability Scoring System Version 3.1: User Guide,\" Forum of Incident Response and Security Teams, July 2019, https://www.first.org/cvss/user-guide. *", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-AI-and-the-Future-of-Cyber-Competition-4.tei.xml", "id": "22a9410463acbd924bc4225eede95fbe"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "BIML \n 3 At BIML, we are interested in \"building security in\" to machine learning (ML) systems from a security engineering perspective. This means understanding how ML systems are designed for security, teasing out possible security engineering risks, and making such risks explicit. We are also interested in the impact of including an ML system as part of a larger design. Our basic motivating question is: how do we secure ML systems proactively while we are designing and building them? This architectural risk analysis (ARA) is an important first step in our mission to help engineers and researchers secure ML systems. We present a basic ARA of a generic ML system, guided by an understanding of standard ML system components and their interactions. Securing a modern ML system must involve diving into the engineering and design of the specific ML system itself. This ARA is intended to make that kind of detailed work easier and more consistent by providing a generic baseline and a set of risks to consider.", "authors": ["Gary McGraw", "Harold Figueroa", "Victor Shepardson", "Richie Bonett"], "title": "AN ARCHITECTURAL RISK ANALYSIS OF MACHINE LEARNING SYSTEMS: Toward More Secure Machine Learning", "text": "Why we Need an ML Risk Analysis at the Architectural Level Twenty-five years ago when the field of software security was in its infancy, much hullabaloo was made over software vulnerabilities and their associated exploits. Hackers busied themselves exposing and exploiting bugs in everyday systems even as those systems were being rapidly migrated to the Internet. The popular press breathlessly covered each exploit. Nobody really concerned themselves with solving the underlying software engineering and configuration problems since finding and fixing the flood of individual bugs seemed like good progress. This hamster-wheel-like process came to be known as \"penetrate and patch.\" After several years of public bug whack-a-mole and debates over disclosure, it became clear that bad software was at the heart of computer security and that we would do well to figure out how to build secure software.  1:viega  That was twenty years ago at the turn of the millennium. These days, software security is an important part of any progressive security program. To be sure, much work remains to be done in software security, but we really do know what that work should be. Though ML (and AI in general) has been around even longer than computer security, until very recently not much attention has been paid to the security of ML systems themselves. Over the last few years, a number of spectacular theoretical attacks on ML systems have led to the same kind of breathless popular press coverage that we experienced during the early days of computer security. It all seems strikingly familiar. Exploit a bug, hype things up in the popular press, lather, rinse, repeat. We need to do better work to secure our ML systems, moving well beyond attack of the day and penetrate and patch towards real security engineering. In our view at BIML, an architectural risk analysis (ARA) is sorely needed at this stage. An ARA takes a designlevel view of a system and teases out systemic risks so that those risks can be properly mitigated and managed as a system is created.  2:mcgraw  Note that in general an ARA is much more concerned with design tradeoffs and solid engineering than it is with particular bugs in a specific system or individual lines of code. In fact, sloppy engineering itself often leads directly security issues of all shapes and sizes. For this reason, we spend some time talking about aspects of robustness and reasonable engineering throughout this document. Our work at BIML is by no means the first work in securing ML systems. Early work in security and privacy of ML has taken more of an \"operations security\" tack focused on securing an existing ML system and maintaining its data integrity. For example, in one section of his seminal paper, Nicolas Papernot uses Saltzer and Schroeder's famous security principles  3:saltzer  to provide an operational perspective on ML security.  4:papernot  In our view, Papernot's work only begins to scratch the surface of ML security design. Following Papernot, we directly addressing Saltzer and Schroeder's security principles from 1972 (as adapted in Building Secure Software by  Viega and McGraw in 2001)  in Part 2. Our treatment of the principles is more directly tied to security engineering than it is to security operations. Also of note, our work focuses on \"security of ML\" as opposed to \"ML for security.\" That is, we focus our attention on helping engineers make sure that their ML system is secure while other work focuses on using ML technology to implement security features. This is an important distinction. In some cases these two distinct practices have been blurred in the literature when they were (confusingly) addressed simultaneously in the same work.  5:barreno  We do what we can to focus all of our attention on the security of ML. \n Intended Audience We have a confession to make. We mostly did this work for ourselves in order to organize our own thinking about security engineering and ML. That said, we believe that what we have produced will be useful to three major audience groups: 1) ML practitioners and engineers can use this work to understand how security engineering and more specifically the \"building security in\" mentality can be applied to ML, 2) security practitioners can use this work to understand how ML will impact the security of systems they may be asked to secure as well as to understand some of the basic mechanisms of ML, and 3) ML security people can use this detailed dive into a security engineering mindset to guide security analysis of specific ML systems. \n Document Organization Part One of this document extensively covers a set of 78 risks that we have identified using a generic ML system as an organizing concept. To start things off, we provide a list of what we consider the top ten risks in ML systems today. Next we discuss a large set of risks associated with each of nine components of a generic ML system. Our intent is for the long list to be a useful guide for security analysis of specific ML systems. Because of that intent, the list is somewhat dauntingly large, but will be useful when practically applied. Next we discuss known ML attacks and present a simple taxonomy associated with our generic ML model. We also briefly cover ML system attack surfaces. We end Part One with treatment of system-wide risks. Part Two of this document is a treatment of Saltzer and Schroeder's 1972 security principles (as adapted in Building Secure Software by  Viega and McGraw in 2001) . You are most welcome to skip around while reading this document, maybe even starting with Part Two. We expect Part One will serve as much as a reference document to refer back to as it serves as an exposition. One last thing before we dive in; this work (and indeed all of security) is just as much about creating resilient and reliable ML systems as it is about security. In our view, security is an emergent property of a system. No system that is unreliable and fragile can be secure. For that reason, a number of the risks we identify and discuss have as much to do with solid engineering as they have to do with thwarting specific attacks. \n BIML BIML \n 7 PART ONE: ML Security Risks Picking a Target: Generic ML ML systems come in a variety of shapes and sizes, and frankly each possible ML design deserves its own specific ARA. For the purposes of this work, we describe a generic ML system in terms of its constituent components and work through that generic system ferreting out risks. The idea driving us is that risks that apply to this generic ML system will almost certainly apply in any specific ML system. By starting with our ARA, an ML system engineer concerned about security can get a jump start on determining risks in their specific system. Figure  1  above shows how we choose to represent a generic ML system. We describe nine basic components that align with various steps in setting up, training, and fielding an ML system: 1) raw data in the world, 2) dataset assembly, 3) datasets, 4) learning algorithm, 5) evaluation, 6) inputs, 7) model, 8) inference algorithm, and 9) outputs. Note that in our generic model, both processes and collections are treated as components. Processesthat is, components 2, 4, 5, and 8-are represented by ovals, whereas things and collections of things-that is, components 1, 3, 6, 7, and 9-are represented as rectangles. The nine components of our generic ML model map in a straightforward way into specific ML models. As an example of this kind of mapping, consider Google's Neural Machine Translation model (GNMT).  6:wu  Here is how that mapping works: 1. Raw data in the world. GNMT makes use of numerous Google internal datasets for training; the sources of these data are not made crystal clear, but Google explicitly mentions Wikipedia articles and news sites.  \n 5. Evaluation. The networks are trained by first applying a maximum-likelihood objective until log perplexity converges, and then refined with reinforcement learning. The process continues until the model produces consistent BLEU scores for the test set. (BLEU (an acronym for bilingual evaluation understudy) is an algorithm for evaluating the quality of machine-translated text that has become a de facto standard.) 6. Inputs. Input to the inference algorithm consists of textual sentences in a particular source language. \n 7. Model. The trained model includes numerous configured hyperparameters and millions of learned parameters. 8. Inference algorithm. GNMT is made accessible through an interface that everyone knows as Google Translate. \n 9. Outputs. Outputs consist of textual sentences in the target language. Given a specific mapping like this, performing a risk analysis by considering the ML security risks associated with each component is a straightforward exercise that should yield fruit. \n The Top Ten ML Security Risks After identifying risks in each component which we describe in detail below, we considered the system as a whole and identified what we believe are the top ten ML security risks. These risks come in two relatively distinct flavors, both equally valid: some are risks associated with the intentional actions of an attacker; others are risks associated with an intrinsic design flaw. Intrinsic design flaws emerge when engineers with good intentions screw things up. Of course, attackers can also go after intrinsic design flaws complicating the situation. The top ten ML security risks are briefly introduced and discussed here. 1. Adversarial examples. Probably the most commonly discussed attacks against machine learning have come to be known as adversarial examples. The basic idea is to fool a machine learning system by providing malicious input often involving very small perturbations that cause the system to make a false prediction or categorization.  \n 3. Online system manipulation. An ML system is said to be \"online\" when it continues to learn during operational use, modifying its behavior over time. In this case a clever attacker can nudge the still-learning system in the wrong direction on purpose through system input and slowly \"retrain\" the ML system to do the wrong thing. Note that such an attack can be both subtle and reasonably easy to carry out. This risk is complex, demanding that ML engineers consider data provenance, algorithm choice, and system operations in order to properly address it. See [alg:  \n 7. Reproducibility. When science and engineering are sloppy, everyone suffers. Unfortunately, because of its inherent inscrutability and the hyper-rapid growth of the field, ML system results are often under-reported, poorly described, and otherwise impossible to reproduce. When a system can't be reproduced and nobody notices, bad things can happen. See [alg:2:reproducibility] below. \n 8. Overfitting. ML systems are often very powerful. Sometimes they can be too powerful for their own good. When an ML system \"memorizes\" its training data set, it will not generalize to new data, and is said to be overfitting. Overfit models are particularly easy to attack. Keep in mind that overfitting is possible in concert with online system manipulation and may happen while a system is running. See [eval:1:overfitting] below.  \n Risks in ML System Components In this section we identify and rank risks found in each of the nine components of the generic ML system introduced above. Each risk is labeled with an identifier as follows: [<component label>:<risk number>:<descriptor>]. We use these labels to cross-reference risks and as shorthand pointers in the rest of the document. After each component's list of risks are a set of controls, some associated with particular risks and others generic. 1. Raw data in the world risks: If we have learned only one thing about ML security over the last few months, it is that data play just as important role in ML system security as the learning algorithm and any technical deployment details. In fact, we'll go out on a limb and state for the record that we believe data make up the most important aspects of a system to consider when it comes to securing an ML system. Our usage of the term raw data in this section is all inclusive, and is not limited to training data (which for what it's worth is usually created from raw data). There is lots of other data in an ML system, including model parameters, test inputs, and operational data. Data security is, of course, a non-trivial undertaking in its own right, and all collections of data in an ML system are subject to the usual data security challenges (plus some new ones). Eventually, a fully-trained ML system (whether online or offline) will be presented with new input data during operations. These data must also be considered carefully during system design. [raw:1:data confidentiality] Preserving data confidentiality in an ML system is more challenging than in a standard computing situation. That's because an ML system that is trained up on confidential or sensitive data will have some aspects of those data built right into it through training. Attacks to extract sensitive and confidential information from ML systems (indirectly through normal use) are well known.  7:shokri  Note that even sub-symbolic \"feature\" extraction may be useful since that can be used to hone adversarial input attacks. [raw:13:utility] If your data are poorly chosen or your model choice is poor, you may reach incorrect conclusions regarding your ML approach. Make sure your methods match your data and your data are properly vetted and monitored. Remember that ML systems can fail just as much due to data problems as due to poorly chosen or implemented algorithms, hyperparameters, and other technical system issues. Associated controls. Note that the labels refer to the original risks (above) which have controls that may help alleviate some of the risk directly: [raw:generic] Protect your data sources if you can. [raw:generic] Sanity check your data algorithmically before you feed it into your model (e.g., using outlier detection, mismatched unit discovery, data range distribution analysis, and so on). For example, make sure that your data properly characterize and represent the problem space so that the ML model learns what it is supposed to learn. Ironically, this is one of the most difficult engineering problems involved in ML as a field. [raw:generic] Transform your data to preserve data integrity. This might even involve cryptographic protection. [raw:generic] Featurize your data so that it is consistently represented. Note that this cuts against the grain of some aspects of \"deep learning\" (mostly because it turns out to be an exercise for the humans), but may result in a more robust ML system. The tension here is a classic issue in ML. Humans are almost always in the loop, carefully massaging data and setting up the problem and the technology to solve the problem. But at the same time the tendency to let an ML system \"magically\" do its work is often over-emphasized. Finding the right balance is tricky and important. [raw:generic] Use version control technology to manage your datasets. Carefully track change logs, diffs, etc, especially when it comes to large datasets. [raw:1:data confidentiality] Design your ML system so that data extraction from the model is expensive. Consider whether there are mathematical properties of the raw input space that lend themselves to particular models (let that help guide choice of model). [raw:6:representation] Manual review of representation and periodic validation is a good thing. Consider what is thrown out or approximated (sometimes for computational reasons) in your data representation and account for that. [raw:8:looping] Look for loops in data streams and avoid them. If raw data come from public sources and system output is also made public, loops may arise without your awareness. As an example, consider what happens when a machine translation system starts using its own translations as training data (as once happened to Google Translate). [raw:12:sensor] Sensor risks can be mitigated with correlated and overlapping sensors that build and maintain a redundant data stream. \n Dataset assembly risks: In order to be processed by a learning algorithm, raw input data often must be transformed into a representational form that can be used by the machine learning algorithm. This \"pre-processing\" step by its very nature impacts the security of an ML system since data play such an essential security role. Of special note in this component is the discrepancy between online models and offline models (that is, models that are continuously trained and models that are trained once and \"set\"). Risks in online models drift, and risks in offline models impact confidentiality. [assembly:1:encoding integrity] Encoding integrity issues noted in [raw:5:encoding integrity] can be both introduced and exacerbated during pre-processing. Does the pre-processing step itself introduce security problems? Bias in raw data processing can impact ethical and moral implications. Normalization of Unicode to ASCII may introduce problems when encoding, for example, Spanish improperly, losing diacritics and accent marks. [assembly:2:annotation] The way data are \"tagged and bagged\" (or annotated into features) can be directly attacked, introducing attacker bias into a system. An ML system trained up on examples that are too specific will not be able to generalize well. Much of the human engineering time that goes into ML is spent cleaning, deleting, aggregating, organizing, and just all-out manipulating the data so that it can be consumed by an ML algorithm. [assembly:3:normalize] Normalization changes the nature of raw data, and may do so to such an extent that the normalized data become exceedingly biased. One example might be an ML system that appears to carry out a complex real-world task, but actually is doing something much easier with normalized data. Destroying the feature of interest in a dataset may make it impossible to learn a viable solution. [assembly:4:partitioning] When building datasets for training, validation, and testing (all distinct types of data used in ML systems), care must be taken not to create bad data partitions. This may include analysis of and comparisons between subsets to ensure the ML system will behave as desired. [assembly:5:fusion] Input from multiple sensors can in some cases help make an ML system more robust. However, note that how the learning algorithm chooses to treat a sensor may be surprising. One of the major challenges in ML is understanding how a \"deep learning\" system carries out its task. Data sensitivity is a big risk and should be carefully monitored when it comes to sensors placed in the real world. [assembly:6:filter] An attacker who knows how a raw data filtration scheme is set up may be able to leverage that knowledge into malicious input later in system deployment. [assembly:7:adversarial partitions] If an attacker can influence the partitioning of datasets used in training and evaluation, they can in some sense practice mind control on the ML system as a whole. It is important that datasets reflect the reality the ML system designers are shooting for. Boosting an error rate in a sub-category might be one interesting attack. Because some deep learning ML systems are \"opaque,\" setting up special trigger conditions as an attacker may be more easily accomplished through manipulation of datasets than through other means.  8:barreno  [assembly:8:random] Randomness plays an important role in stochastic systems. An ML system that is depending on Monte Carlo randomness to work properly may be derailed by not-really-random \"randomness.\" Use of cryptographic randomness sources is encouraged. \"Random\" generation of dataset partitions may be at risk if the source of randomness is easy to control by an attacker interested in data poisoning. Associated controls. Note that the labels refer to the original risks (above) which have controls that may help alleviate some of the risk directly: [assembly:generic] Provide data sanity checks that look at boundaries, ranges, probabilities, and other aspects of data to find anomalies before they are included in critical datasets. Consider, for example, signal to noise ratio and make sure that it is consistent enough to include data as they are assembled. [assembly:5:fusion] Determine how dirty data from a sensor may become, and control for both that and for sensor failure. Using multiple sensors may help, especially if they are not exactly the same kind of sensor or modality. 3. Datasets risks: Assembled data must be grouped into a training set, a validation set, and a testing set. The training set is used as input to the learning algorithm. The validation set is used to tune hyperparameters and to monitor the learning algorithm for overfitting. The test set is used after learning is complete to evaluate performance. Special care must be taken when creating these groupings in order to avoid predisposing the ML algorithm to future attacks (see [assembly:7:adversarial partitions] ). In particular, the training set deeply influences an ML system's future behavior. Attacking an ML system through the training set is one of the most obvious ways to throw a monkey wrench into the works. [data:1:poisoning] All of the first three components in our generic model (raw data in the world, dataset assembly, and datasets) are subject to poisoning attacks whereby an attacker intentionally manipulates data in any or all of the three first components, possibly in a coordinated fashion, to cause ML training to go awry. In some sense, this is a risk related both to data sensitivity and to the fact that the data themselves carry so much of the water in an ML system. Data poisoning attacks require special attention. In particular, ML engineers should consider what fraction of the training data an attacker can control and to what extent.  12:alfeld  [data:2:transfer] Many ML systems are constructed by tuning an already trained base model so that its somewhat generic capabilities are fine-tuned with a round of specialized training. A transfer attack presents an important risk in this situation. In cases where the pretrained model is widely available, an attacker may be able to devise attacks using it that will be robust enough to succeed against your (unavailable to the attacker) tuned task-specific model. You should also consider whether the ML system you are fine-tuning could possibly be a Trojan that includes sneaky ML behavior that is unanticipated.  13:mcgraw  [data:3:disimilarity] If training, validation, and test sets are not \"the same\" from a data integrity, trustworthiness, and mathematical perspective, an ML model may appear to be doing something that it is not. For example, an ML system trained up on six categories but only tested against two of the six may not ultimately be exhibiting proper behavior when it is fielded. More subtly, if an evaluation set is too similar to the training set, overfitting may be a risk. By contrast, when the evaluation set is too different from the eventual future inputs during operations, then it will not measure true performance. Barreno et al say it best when they say, \"Analyzing and strengthening learning methods in the face of a broken stationarity assumption is the crux of the secure learning problem.\"  8:barreno  [data:4:storage] As in [raw:3:storage], data may be stored and managed in an insecure fashion. Who has access to the data pool, and why? Think about [system:8:insider] when working on storage. [data:5:dataset weak rep] Assembling a dataset involves doing some thinking and observation about the resulting representation inside the ML model. Robust representations result in fluid categorization behavior, proper generalization, and non-susceptibility to adversarial input. As an example, a topic model trained on predominantly English input with a tiny bit of Spanish will group all Spanish topics into one uniform cluster (globbing all Spanish stuff together). [data:6:supervisor] Some learning systems are \"supervised\" in a sense that the target result is known by the system designers (and labeled training data are available). Malicious introduction of misleading supervision would cause an ML system to be incorrectly trained. For example, a malicious supervisor might determine that each \"tank\" in a satellite photo is counted as two tanks. (See also [assembly:2:annotation].) [data:7:online] Real time data set manipulation can be particularly tricky in an online network where an attacker can slowly \"retrain\" the ML system to do the wrong thing by intentionally shifting the overall data set in certain directions. Associated controls. Note that the labels refer to the original risks (above) which have controls that may help alleviate some of the risk directly: [data:generic] Try to characterize the statistical overlap between validation and training sets. What is best? Document your decisions. [data:4:disimilarity] Ensure data similarity between the three datasets using mathematical methods. Just as in software engineering, where \"coding to the test\" can lead to robustness issues, poor training, testing, and validation data hygiene can seriously damage an ML system. \n Learning Algorithm risks: In our view, though a learning algorithm lies at the technical heart of each ML system, the algorithm itself presents far less of a security risk than the data used to train, test, and eventually operate the ML system. That said, risks remain that are worthy of note. Learning algorithms come in two flavors, and the choice of one or the other makes a big difference from a security perspective. ML systems that are trained up, \"frozen,\" and then operated using new data on the frozen trained system are called offline systems. Most common ML systems (especially classifiers) operate in an offline fashion. By contrast, online systems operate in a continuous learning mode. There is some advantage from a security perspective to an offline system because the online stance increases exposure to a number of data borne vulnerabilities over a longer period of time. [alg:1:online] An online learning system that continues to adjust its learning during operations may drift from its intended operational use case. Clever attackers can nudge an online learning system in the wrong direction on purpose. [alg:2:reproducibility] ML work has a tendency to be sloppily reported. Results that can't be reproduced may lead to overconfidence in a particular ML system to perform as desired. Often, critical details are missing from the description of a reported model. Also, results tend to be very fragile-often running a training process on a different GPU (even one that is supposed to be spec-identical) can produce dramatically different results. In academic work, there is often a tendency to tweak the authors' system until it outperforms the \"baseline\" (which doesn't benefit from similar tweaking), resulting in misleading conclusions that make people think a particular idea is actually good when it wasn't actually improving over simpler, earlier method. [alg:3:exploit-v-explore] Part of the challenge of tuning an ML system during the research process is understanding the search space being explored and choosing the right model architecture (and algorithm) to use and the right parameters for the algorithm itself. Thinking carefully about problem space exploration versus space exploitation will lead to a more robust model that is harder to attack. Pick your algorithm with care. As an example, consider whether your system has an over-reliance on gradients and may possibly benefit from random restarting or evolutionary learning. [alg:4:randomness] Randomness has a long and important history in security. In particular, Monte Carlo randomness versus cryptographic randomness is a concern. When it comes to ML, setting weights and thresholds \"randomly\" must be done with care. Many pseudo-random number generators (PRNG) are not suitable for use. PRNG loops can really damage system behavior during learning. Cryptographic randomness directly intersects with ML when it comes to differential privacy. Using the wrong sort of random number generator can lead to subtle security problems. [alg:5:blind spots] All ML learning algorithms may have blind spots. These blind spots may open an ML system up to easier attack through techniques that include adversarial examples. [alg:6:confidentiality] Some algorithms may be unsuited for processing confidential information. For example, using a non-parametric method like k-nearest neighbors in a situation with sensitive medical records is probably a bad idea since exemplars will have to be stored on production servers. Algorithmic leakage is an issue that should be considered carefully.  4:papernot  [alg:7:noise] Noise is both friend and foe in an ML system. For some problems, raw data input need to be condensed and compacted (de-noised). For others, the addition of Gaussian noise during preprocessing can enhance an ML system's generalization behavior. Getting this right involves careful thinking about data structure that is both explicit and well documented. Amenability to certain kinds of adversarial input attack is directly linked to this risk.  14:goodfellow  [alg:8:oscillation] An ML system may end up oscillating and not properly converging if, for example, it is using gradient descent in a space where the gradient is misleading. [alg:9:hyperparameters] One of the challenges in the ML literature is an over-reliance on \"empirical\" experiments to determine model parameters and an under-reliance on understanding why Associated controls. Note that the labels refer to the original risks (above) which have controls that may help alleviate some of the risk directly: [alg:4:randomness] Have a security person take a look at your use of randomness, even if it seems innocuous. [alg:5:blind spots] Representational robustness (for example word2vec encoding in an NLP system versus one-shot encoding) can help combat some blind spot risks. [alg:6:confidentiality] Know explicitly how the algorithm you are using works. Make sure that your choice preserves representational integrity. [alg:6:confidentiality] Keep a history of queries to your system in a log and review the log to make sure your system is not unintentionally leaking confidential information. Be careful how you store the log-logging everything can itself introduce a big privacy risk. [alg:9:hyperparameters] Carefully choose hyperparameters and make notes as to why they are set the way they are. Lock in hyperparameters so that they are not subject to change. [alg:10:hyperparameter sensitivity] Perform a sensitivity analysis on the set of hyperparameters you have chosen. \n Evaluation risks: Determining whether an ML system that has been fully trained is actually doing what the designers want it to do is a thing. Evaluation data are used to try to understand how well a trained ML system can perform its assigned task (post learning). Recall our comments above about the important role that stationarity assumptions have in securing ML systems. [eval:1:overfitting] A sufficiently powerful machine is capable of learning its training data set so well that it essentially builds a lookup table. This can be likened to memorizing its training data. The unfortunate side effect of \"perfect\" learning like this is an inability to generalize outside of the training set. Overfit models can be pretty easy to attack through input since adversarial examples need only be a short distance away in input space from training examples. Note that generative models can suffer from overfitting too, but the phenomenon may be much harder to spot. Also note that overfitting is also possible in concert with [data:6:online]. [eval:2:bad eval data] Evaluation is tricky, and an evaluation data set must be designed and used with care. A bad evaluation data set that doesn't reflect the data it will see in production can mislead a researcher into thinking everything is working even when it's not. Evaluation sets can also be too small or too similar to the training data to be useful. [eval:3:cooking the books] In some cases, evaluation data may be intentionally structured to make everything look great even when it's not. [eval:4:science] Common sense evaluation and rigorous evaluation are not always the same thing. For example, evaluation of an NLP system may rely on \"bags of words\" instead of a more qualitative structural evaluation.  15:reiter  [eval:5:catastrophic forgetting] Just as data play a key role in ML systems, representation of those data in the learned network is important. When a model is crammed too full of overlapping information, it may suffer from catastrophic forgetting. This risk was much more apparent in the early '90s when networks (and the CPUs they ran on) were much smaller. However, even a large network may be subject to this problem. Online systems are, by design, more susceptible. [eval:6:choking on big data] In an online model, the external data set available may be so vast that the ML system is simply overwhelmed. That is, the algorithm may not scale in performance from the data it learned on to real data. In online situations the rate at which data comes into the model may not align with the rate of anticipated data arrival. This can lead to both outright ML system failure and to a system that \"chases its own tail.\" [eval:7:data problems] Upstream attacks against data make training and its subsequent evaluation difficult. Associated controls. Note that the labels refer to the original risks (above) which have controls that may help alleviate some of the risk directly: [eval:2:bad eval data] Public data sets with well-known error rates (or generalization rates) may combat or help control this risk. [eval:2:bad eval data] and [eval:3:cooking the books] are much harder to pull off when the evaluation data and results are public. The research literature is beginning to move toward reproducible results though release of all ML system code and data. \n Input risks: When a fully trained model is put into production, a number of risks must be considered. Probably the most important set of these operations/production risks revolves around input data fed into the trained model. Of course, by design these input data will likely be structured and pre-processed similarly to the training data. Many of the risks identified above (see especially raw data in the world risks and data assembly risks) apply to model input almost directly. [input:1:adversarial examples] One of the most important categories of computer security risks is malicious input. The ML version of malicious input has come to be known as adversarial examples. While important, adversarial examples have received so much attention that they swamp out all other risks in most people's imagination.  16:yuan  [input:2:controlled input stream] A trained ML system that takes as its input data from outside may be purposefully manipulated by an attacker. To think about why anybody would bother to do this, consider that the attacker may be someone under scrutiny by an ML algorithm (a loan seeker, a political dissident, a person to be authenticated, etc). [input:3:dirty input] The real world is noisy and messy. Input data sets that are dirty enough will be hard to process. A malicious adversary can leverage this susceptibility by simply adding noise to the world. [input:4:looped input] If system output feeds back into the real world there is some risk that it may find its way back into input causing a feedback loop. Sometimes this even happens with ML output data feeding back into training data. [input:5:pre-processing replication] The same care that goes into data assembly (component 2) should be given to input, even in an online situation. This may be difficult for a number of reasons Associated controls. Note that the labels refer to the original risks (above) which have controls that may help alleviate some of the risk directly: [input:2:controlled input stream] A multi-modal input stream will be harder to completely control. One way to carry this out might be to use multiple sensors that are not similarly designed or that don't have the same engineering failure conditions. [input:3:dirty input] Sanity checks, filters, and data cleaning can control this risk. Of course, those mechanisms can be attacked as well. Note that often pre-processing ends up being more about making an ML system be able to learn than it is about \"getting it right.\" 7. Model risks: When a fully trained model is put into production, a number of important risks crop up. Note that some of the risks discussed in the evaluation risks section above apply directly in this section as well (for example, [eval:1:overfitting] and [eval:4:catastrophic forgetting] both apply). Representation issues are some of the most difficult issues in ML, both in terms of primary input representation and in terms of internal representation and encoding. Striking a balance between generalization and specificity is the key to making ML useful. [model:4:training set reveal] Most ML algorithms learn a great deal about input, some of which is possibly sensitive (see [raw:1:data confidentiality]), and store a representation internally that may include sensitive information. Algorithm choice can help control this risk, but be aware of the output your model produces and how it may reveal sensitive aspects of its training data. When it comes to sensitive data, one promising approach in privacy-preserving ML is differential privacy which we discuss below. [model:5:steal the box] Training up an ML system is not free. Stealing ML system knowledge is possible through direct input/output observation. This is akin to reversing the model. Associated controls. Note that the labels refer to the original risks (above) which have controls that may help alleviate some of the risk directly: [model:5:steal the box] Watch the output that you provide (it can and will be used against you). \n Inference algorithm risks: When a fully trained model is usually put into production, a number of important risks must be considered. These encompass data fed to the model during operations (see raw data risks and pre-processing risks), risks inherent in the production model, and output risks. [inference:1:online] A fielded model operating in an online system (that is, still learning) can be pushed past its boundaries. An attacker may be able to carry this out quite easily. [inference:2:inscrutability] In far too many cases, an ML system is fielded without a real understanding of how it works or why it does what it does. Integrating an ML system that \"just works\" into a larger system that then relies on the ML system to perform properly is a very real risk. [inference:3:hyperparameters] Inference algorithms have hyperparameters, for example sampling temperature in a generative model. If an attacker can surreptitiously modulate the hyperparameters for the inference algorithm after the evaluation process is complete, they can control the system's behavior. (Also see [alg:9:hyperparameters].) [inference:3:confidence scores] In many cases, confidence scores (which are paired with classification category answers) can help an attacker. If an ML system is not confident about its answer and says so, that provides feedback to an attacker with regards to how to tweak input to make the system misbehave. Conversely, a system that doesn't return confidence scores is much harder to use correctly (and may be used idiotically). Care should be taken as to what kind of output feedback a user can and should get. [inference:4:hosting] Many ML systems are run on hosted, remote servers. Care must be taken to protect these machines against ML-related attacks (not to mention the usual pile of computer security stuff). [inference:5:user risk] When a user decides to use an ML system that is remote, they expose their interests (and possibly their input) to the owners of the ML system. \n BIML 21 Associated controls. Note that the labels refer to the original risks (above) which have controls that may help alleviate some of the risk directly: [inference:1:online] An ML system in production can be refreshed to a known state, reset, or otherwise \"cleaned\" periodically. This can limit the window for online attack. [inference:3:hosting] Take care to isolate engineering ML systems from production systems. Production systems in particular should be properly hardened and monitored. 9. Output risks: Keep in mind that the entire purpose of creating, training, and evaluating a model may be so that its output serves a useful purpose in the world. The second most obvious direct attack against an ML system will be to attack its output. [output:1:direct] An attacker tweaks the output stream directly. This will impact the larger system in which the ML subsystem is encompassed. There are many ways to do this kind of thing. Probably the most common attack would be to interpose between the output stream and the receiver. Because models are sometimes opaque, unverified output may simply be used with little scrutiny, meaning that an interposing attacker may have an easy time hiding in plain sight. [output:2:provenance] ML systems must be trustworthy to be put into use. Even a temporary or partial attack against output can cause trustworthiness to plummet. [output:4:inscrutability] In far too many cases with ML, nobody is really sure how the trained systems do what they do. This is a direct affront on trustworthiness and can lead to challenges in some domains such as diagnostic medicine. [output:5:transparency] Decisions that are simply presented to the world with no explanation are not transparent. Attacking opaque systems is much easier than attacking transparent systems, since it is harder to discern when something is going wrong. [output:6:eroding trust] Causing an ML system to misbehave can erode trust in the entire discipline. A GAN that produces uncomfortable sounds or images provides one example of how this might unfold.  \n Mapping Known Attacks to our Model In this section we briefly consider direct attacks on ML algorithms. See Figure  2 . These attacks are closely related to the security risks we have enumerated above, but they are not the same. Attacks are distinct in that they may leverage multiple risks. You can think of a specific attack as a coordinated exploit of a set of risks that results in system compromise. For the most part we will ignore attacks on ML infrastructure or attacks that specifically circumvent ML-based defense. We classify attacks on ML systems based on how and where (and to some degree when) the system is compromised. An attack may manipulate the behavior (attacking operational integrity) or extract information (attacking confidentiality). Additionally attacks can affect the training data, run-time inputs, and the model used for inference. Attacks can disrupt both the engineering stages of developing an ML system as well as a deployed ML system. The two axes of how and where a system is compromised lead to a taxonomy with six categories: (1) Data manipulation, also called a \"poisoning\" 18:kloft or \"causative\" attack is a manipulation attack via the training process. 5:barreno An attacker modifies a data corpus used to train an ML system in order to impair or influence the system's behavior. For example, an attacker may publish bogus data to influence financial time-series forecasting models 19:alfeld or interfere with medical diagnoses. 20:mozzaffari-kermani (2) Input manipulation, including by \"adversarial examples,\" is a manipulation attack on an ML model at inference time (or test time).  14:goodfellow  In this case, an attacker concocts an input to an operating ML system which reliably produces a different output than intended. Examples include a stop sign being classified as a speed limit sign; 21:eykholt a spam email being classified as not spam; 22:biggio or a vocal utterance being transcribed as an unrelated text.  23 :carlini (For a survey of input manipulation techniques on deep learning and classical ML systems see [Yuan19] 16:yuan and [Biggio13] 22:biggio , respectively.) Note that in the online setting, runtime inputs and training data may not be distinct. However, we can say that input manipulation compromises behavior toward the malicious input, while data manipulation compromises behavior toward future inputs-the methods of attack and the security implications are distinct. (3) Model manipulation, also called \"backdooring\" 24:gu or a \"supply chain\" attack occurs when an attacker publishes a model with certain latent behavior, to be unwittingly adopted by third parties and later exploited by the attacker.  25:kumar  It is common in the deep learning community to release models under a permissive open source license; given the prevalence of code reuse and transfer learning we believe this potential attack and defenses against it deserves greater scrutiny. (4) Data Extraction, commonly called \"inference attacks\" (including membership inference, attribute inference, and property inference) and also sometimes called \"model inversion,\" is when an attacker extracts details of the data corpus an ML model was trained on by querying the model or inspecting it directly.  26:ateniese  Research in deep learning often focuses on the model to the exclusion of the data, yet data are known to be crucially important to a trained system's behavior. Though research is often conducted on public datasets, real-world ML systems involve proprietary data with serious privacy implications. (5) Input extraction, sometimes also called \"model inversion\" applies in cases where model outputs are public but inputs are secret; an attacker attempts to recover inputs from outputs. 27:fredrikson For example, inferring features of medical records from the dosage recommended by an ML model,  27: fredrikson or producing a recognizable image of a face given only the identity (classification in a face-recognition model) and confidence score. Complementing and contrasting with our taxonomy, NIST recently published a draft taxonomy of adversarial machine learning (AML) which aims to create a taxonomy not just of attacks, but also defenses and consequences. 31:tabassi At the top level they consider three categories of attack target (physical, digital representation, or ML approach), attack techniques, and knowledge that the attacker may have of the system being attacked. Overall, the NIST taxonomy and terminology glossary are helpful navigational tools for current literature on AML. Our taxonomy of ML attacks is more focused on practitioners trying to secure an ML system. As such, it is directly grounded on a simpler model for an ML system and more directly describes established categories of system compromise. This makes our taxonomy much simpler, with a clear focus on where and how the system is attacked while still taking into consideration other basic categories in the NIST draft. For example, we find that the our dimension of a how a system is compromised maps effectively to the NIST draft sense of consequence. We also find that our six category approach is straightforward to specialize to specific ML approaches and systems as covered in the targets branch of the NIST taxonomy. Thus we can easily apply our approach of where and how to the various modalities of ML such as supervised, unsupervised, and reinforcement learning. \n Attack Surface Estimation In security, an attack surface us defined as the sum of different locations in a system where an attacker can try to manipulate input, directly impact system processing, or extract data. Keeping a system's attack surface as small as possible is a basic security measure. Practically, we identify three main attack surfaces for ML systems: deployment, engineering, and data sources. See Figure  2 . Deployment is the most straightforward surface to attack, comprising the inference software and model itself. A deployed ML system includes supporting hardware and software (e.g. web servers); an attacker can typically study an API or hardware device at length to develop an attack. Roughly, the bottom plate in our diagram corresponds to deployment, but note that this distinction is less appropriate in the online learning regime. Conventional computer security is important here, as is understanding how information leaks between model, input, and output. The ML engineering process is more remote from the system's behavior yet fully determines it. Sensitive data may be most exposed during the engineering process. Roughly, this is the upper plate in Figure  2 , though note that the inference algorithm and model are what move from engineering to deployment. Operational security is important here. Data sources are still more remote from an ML system's behavior but may be particularly easy for an attacker to manipulate undetected. Attacks on data sources should be anticipated when collecting and assembling datasets. \n System-Wide Risks and Broad Concerns To this point, our coverage of ML security risks has been confined to a component-based view. In addition to risks grounded in components, there are a number of system-wide risks that emerge only at the system level or between and across multiple components. We identify and discuss system-wide risks here: [system:1:black box discrimination] Many data-related component risks lead to bias in the behavior of an ML system. ML systems that operate on personal data or feed into high impact decision processes (such as credit scoring, employment, and medical diagnosis decisions) pose a great deal of risk. When biases are aligned with gender, race, or age attributes, operating the system may result in discrimination with respect to one of these protected classes. Using biased ML subsystems is is definitely illegal in some contexts, may be unethical, and is always irresponsible. [system:2:overconfidence] When an ML system with a particular error behavior is integrated into a larger system and its output is treated as high confidence data, users of the system may become overconfident in the operation of the system for its intended purpose. A low scrutiny stance with respect to the overall system makes it less likely that an attack against the ML subsystem will be detected. Developing overconfidence in ML is made easier by the fact that ML systems are often poorly understood and vaguely described. (See [output:5:transparency].) [system:3:loss of confidence] Any ML system can and will make mistakes. For example, there are limitations to how effective the prediction of a target variable or class can be given certain input. If system users are unaware of the subtleties of ML, they may not be able to account for \"incorrect\" behavior. Lost confidence may follow logically. Ultimately, users may erroneously conclude that the ML system is not beneficial to operation at all and thus should be disregarded. In fact the ML system may operate on average much more effectively than other classifying technology and may be capable of scaling a decision process beyond human capability. Throwing out the baby with the bathwater is an ML risk. As an example, consider what happens when self-driving cars kill pedestrians. [system:4:public perception] Confidence related risks such as [system:2:overconfidence] and [system:3:loss of confidence] are focused on the impact that common ML misunderstandings have on users of a system. Note that such risks can find their way out into society at large with impacts on policy-making (regarding the adoption or role of ML technologies) and the reputation of a company (regarding nefarious intentions, illegality, or competence). A good example is the Microsoft chatbot, Tay, which learned to converse by parsing raw twitter content and ultimately exhibited racist, xenophobic, and sexist behavior as a result. Microsoft pulled the plug on Tay. Tay was a black eye for ML in the eyes of the public.  32:jagielski  [system:5:error propagation] When ML output becomes input to a larger decision process, errors arising in the ML subsystem may propagate in unforeseen ways. For example, a classification decision may end up being treated as imputed metadata or otherwise silently impact a conditional decision process. The evaluation of ML subsystem performance in isolation from larger system context may not take into account the \"regret\" this may incur. That is, methods that evaluate ML accuracy may not evaluate utility, leading to what has been called regret in the ML literature. [system:6:cry wolf] When an ML subsystem operating within a larger system generates too many alarms, the subsystem may be ignored. This is particularly problematic when ML is being applied to solve a security problem like intrusion or misuse detection. False alarms may discourage users from paying attention, rendering the system useless. [system:7:data integrity] If ML system components are distributed, especially across the Internet, preserving data integrity between components is particularly important. An attacker in the middle who can tamper with data streams coming and going from a remote ML component can cause real trouble. [system:8:insider] As always in security, a malicious insider in an ML system can wreak havoc. Note for the record that data poisoning attacks (especially those that subtly bias a training set) can already be hard to spot. A malicious insider who wishes not to get caught would do well to hide in the data poisoning weeds. [system:9:API encoding] Data may be incorrectly encoded in a command, or vice versa. When data and API information are mixed, bad things happen in security. Know that APIs are a common attack target in security and are in some sense your public front door. How do you handle time and state? What about authentication? [system:10:denial of service] Denial of service attacks have broad impact when service access impacts a decision process. When an ML system fails, recovery may not be possible. If you decide to rely entirely on an ML system that fails, recovery may not be possible, even if all of the data that feed the ML system are still around. \n BIML BIML \n 29 PART TWO: ML Security Principles \n Security Principles and Machine Learning In security engineering it is not practical to protect against every type of possible attack. Security engineering is an exercise in risk management. One approach that works very well is to make use of a set of guiding principles when designing and building systems. Good guiding principles tend to improve the security outlook even in the face of unknown future attacks. This strategy helps to alleviate the \"attack-of-the-day\" problem so common in early days of software security (and also sadly common in early approaches to ML security). In this section we present ten principles for ML security lifted directly from Building Secure Software and adapted for ML.  1:viega  The goal of these principles is to identify and to highlight the most impor tant objectives you should keep in mind when designing and building a secure ML system. Following these principles should help you avoid lots of common security problems. Of course, this set of principles will not be able to cover every possible new flaw lurking in the future. Some caveats are in order. No list of principles like the one pre sented here is ever perfect. There is no guarantee that if you follow these principles your ML system will be secure. Not only do our principles present an incomplete picture, but they also sometimes conflict with each other. As with any complex set of principles, there are often subtle tradeoffs involved. Clearly, application of these ten principles must be sensitive to context. A mature risk management approach to ML provides the sort of data required to apply these principles intelligently. What follows is a treatment of each of the ten principles from an ML systems engineering perspective. Principle 1: Secure the Weakest Link Security people are quick to point out that security is like a chain. And just as a chain is only as strong as the weakest link, an ML system is only as secure as its weakest component. Want to anticipate where bad guys will attack your ML system? Well, think through which part would be easiest to attack and what the attacker's goals might be. What really matters is the easiest way for the attacker to achieve those goals. For a first stab at attack surface analysis, see Figure  2  and the associated text above. ML systems are different from many other artifacts that we engineer because the data in ML are just as important (or sometimes even more important) as the learning mechanism itself. That means we need to pay much more attention to the data used to train, test, and operate an ML system than we might in a standard system. In some sense, this turns the idea of an attack surface on its head. To understand what we mean, consider that the training data in an ML system may often come from a public location -that is, one that may be subject to poor data protection controls. If that's the case, perhaps the easiest way to attack an ML system of this flavor would be through polluting or otherwise manipulating the data before they even arrive. An attacker wins if they get to the ML-critical data before the ML system even starts to learn. Who cares about the public API of the trained up and operating ML system if the data used to build it were already maliciously constructed? Thinking about ML data as money makes a good exercise. Where does the \"money\" (that is, data) in the system come from? How is it stored? Can counterfeit money help in an attack? Does all of the money get compressed into high value storage in one place (say the weights and thresholds learned in the ML systems' distributed representation)? How does money come out of an ML system? Can money be transferred to an attacker? How would that work? Let's stretch this analogy even farther. When it comes to actual money, a sort of perverse logic pervades the physical security world. There's generally more money in a bank than a convenience store, but which one is more likely to be held up? The convenience store, because banks tend to have much stronger security precautions. Convenience stores are a much easier target. Of course the payoff for successfully robbing a convenience store is much lower than knocking off a bank, but it is probably a lot easier to get away from the convenience store crime scene. In terms of our analogy, you want to look for and better defend the convenience stores in your ML system. ML has another weird factor that is worth considering-that is that much of the source code is open source and reused all over the place. Should you trust that algorithm that you snagged from github? How does it work? Does it protect those oh so valuable data sets you built up? What if the algorithm itself is sneakily compromised? These are some potential weak links that may not be considered in a traditional network security stance. Identifying the weakest component of a system falls directly out of a good risk analysis. Given good risk analysis information, addressing the most serious risk first, instead of a risk that may be easiest to mitigate, is always prudent. Security resources should be doled out according to risk. Deal with one or two major problems, and move on to the remaining ones in order of severity. You can make use of the ML security risks we identify in this document as a starting point for an in-depth analysis of your own system. Of course, this strategy can be applied forever, because 100% security is never attainable. There is a clear need for some stopping point. It is okay to stop addressing risks when all components appear to be within the thresh old of acceptable risk. The notion of acceptability depends on the business propo sition. All of our analogies aside, good security practice dictates an approach that identifies and strengthens weak links until an acceptable level of risk is achieved. \n Principle 2: Practice Defense in Depth The idea behind defense in depth is to manage risk with diverse defensive strategies (sometimes called controls), so that if one layer of defense turns out to be inadequate, another layer of defense hopefully prevents a full breach. Let's go back to our example of bank security. Why is the typical bank more secure than the typical convenience store? Because there are many redundant security measures protecting the bank, and the more measures there are, the more secure the place is. Security cameras alone are a deterrent for some. But if people don't care about the cameras, then a security guard is there to defend the bank physi cally with a gun. Two security guards provide even more protection. But if both security guards get shot by masked bandits, then at least there's still a wall of bulletproof glass and electronically locked doors to protect the tellers from the robbers. Of course if the robbers happen to kick in the doors, or guess the code for the door, at least they can only get at the teller registers, because the bank has a vault protecting the really valuable stuff. Hopefully, the vault is protected by several locks and cannot be opened without two individuals who are rarely at the bank at the same time. And as for the teller registers, they can be protected by having dye-emitting bills stored at the bottom, for distribution during a robbery. Of course, having all these security measures does not ensure that the bank is never successfully robbed. Bank robberies do happen, even at banks with this much security. Nonetheless, it's pretty obvious that the sum total of all these defenses results in a far more effective security system than any one defense alone. The defense-in-depth principle may seem somewhat contradictory to the \"secure-the-weakest-link\" principle because we are essentially saying that defenses taken as a whole can be stronger than the weakest link. However, there is no contradiction. The principle \"secure the weakest link\" applies when components have security functionality that does not overlap. But when it comes to redundant security measures, it is indeed possible that the sum protection offered is far greater than the protection offered by any single component. ML systems are constructed out of numerous components. And, as we pointed out multiple times above, the data are often the most important thing from a security perspective. This means that bad actors have as many opportunities to exploit an ML system as there are components, and then some. Each and every component comes with a set of risks, and each and every one of them needs to address those risks head on. But wait, there's more. Defense in depth teaches that vulnerabilities not addressed by one component should be caught by another. In some cases a risk may be controlled \"upstream\" and in others \"downstream.\" Let's think about how defense in depth impacts the goal of securing training data in an ML system. A straightforward security approach will attempt to secure sensitive training data behind some kind authentication and authorization system, only allowing the model access to the data while it is actually training. This may well be a reasonable and well-justified practice, but it is by no means sufficient to ensure that no sensitive information in the training data can be leaked through malicious misuse/abuse of the system as a whole. Here's why. Through the training process itself, the training data come to be represented in the model itself.  33:Fredrikson  That means getting to sensitive data through the model is a risk. Some ML models are vulnerable to leaking sensitive information via carefully selected queries made to the operating model itself. In other cases, lots of know-how in \"learned\" form may be leaked through a transfer attack. A second line of defense against these kind of \"through the model\" attacks against training data might be to anonymize the dataset so that particularly sensitive aspects of the data are not exposed even through the model. Maintaining a history of queries made by users, and preventing subsequent queries that together could be used to divine sensitive information can serve as an additional defensive layer that protects against these kinds of attack. Practicing defense in depth naturally involves applying the principle of least privilege to users and operations engineers of an ML system. Identifying and preventing security exploits is much easier when every component limits its access to only those resources it actually requires. In this case, identifying and separating components in a design can help, because components become natural trust boundaries where controls can be put in place and policies enforced. Defense in depth is especially powerful when each component works in concert with the others. \n Principle 3: Fail Securely Even under ideal conditions, complex systems are bound to fail eventually. Failure is an unavoidable state that should always be planned for. From a security perspective, failure itself isn't the problem so much as the tendency for many systems to exhibit insecure behavior when they fail. ML systems are particularly complicated (what with all that dependence on data) and are prone to fail in new and spectacular ways. Consider a system that is meant to classify its input. In a very straightforward way, failure in a classifier would constitute giving the wrong answer (e.g., incorrectly reporting that a cat is a tank). What should an ML system do? Maybe it should emit no answer if confidence is low. Or maybe it can flag inaccurate or iffy classifications like this, through say emitting a confidence score. Reporting a confidence score seems like not such a bad thing to do from an engineering perspective. But in some cases, simply reporting what an ML system got wrong or was underconfident about can lead to security vulnerability. As it turns out, attackers can exploit misclassification to create adversarial examples,  30: gilmer or use a collection of errors en masse to ferret out confidential information used to train the model.  7:shokri  In general, ML systems would do well to avoid transmitting low-confidence classification results to untrusted users in order to defend against these attacks, but of course that seriously constrains the usual engineering approach. This is a case in which failing securely is much more subtle than it may seem at first blush. Classification results should only be provided when the system is confident that they are correct. In the case of either a failure or a low confidence result, care must be taken that any feedback from the model to a malicious user can't be exploited. Note that many ML models are capable of providing confidence levels along with their other output to address some of these risks. That certainly helps when it comes to understanding the classifier itself, but it doesn't really address information exploit or leakage (both of which are more challenging problems). ML system engineers should carefully consider the sensitivity of their systems' predictions and take into account the amount of trust they afford the user when deciding what to report. If your ML system has to fail, make sure that it fails securely. \n Principle 4: Follow the Principle of Least Privilege The principle of least privilege states that only the minimum access necessary to perform an operation should be granted, and that access should be granted only for the minimum amount of time necessary.  3:saltzer  When you give out access to parts of a system, there is always some risk that the privileges associated with that access will be abused. For example, let's say you are to go on vacation and you give a friend the key to your home, just to feed pets, collect mail, and so forth. Although you may trust the friend, there is always the possibility that there will be a party in your house without your consent, or that something else will happen that you don't like. In an ML system, we most likely want to control access around lifecycle phases. In the training phase, the system may have access to lots of possibly sensitive training data. Assuming an offline model (where training is not continuous), after the training phase is complete, the system should no longer require access to those data. (As we discussed when we were talking defense in depth, system engineers need to understand that in some sense all of the confidential data are now represented in the trained-up ML system and may be subject to ML-specific attacks.) Thinking about access control in ML is useful and can be applied through the lens of the principle of least privilege, particularly between lifecycle phases and system components. Users of an ML system are not likely to need access to training data and test data, so don't give it to them. In fact, users may only require black box API access to a running system. If that's the case, then provide only what is necessary in order to preserve security. Less is more when it comes to the principle of least privilege. Limit data exposure to those components that require it and then grant access for as short a time period as possible. \n Principle 5: Compartmentalize The risk analysis of a generic ML system we provide in this document uses a set of nine \"components\" to help categorize and explain risks found in various logical pieces (see Figure  1 ). Components can be either processes or collections. Just as understanding a system is easier when a system is divided up into pieces, controlling security risk is easier when the pieces themselves are each secured separately. Another way of thinking about this is to compare old fashioned \"monolithic\" software design to \"micro-services\" design. In general, both understanding and securing a monolith is much harder than securing a set of services (of course things get tricky when services interact in time, but we'll ignore that for now). In the end we want to eradicate the monolith and use compartmentalization as our friend. Let's imagine one security principle and see how compartmentalization can help us think it through. Part of the challenge of applying the principle of least privilege in practice (described above) has to do with component size and scope. When building blocks are logically separated and structured, applying the principle of least privilege to each component is much more straightforward than it would be otherwise. Smaller components should by and large require less privilege than the complete system. Does this component involve pre-processed training data that will directly impact system learning? Hmm, better secure those data! The basic idea behind compartmentalization is to minimize the amount of damage that can be done to a system by breaking up the system into a number of units and isolating processes or data that carry security privilege. This same principle explains why submarines are built with many different chambers, each separately sealed. If a breach in the hull causes one chamber to fill with water, the other chambers are not affected. The rest of the ship can keep its integrity, and people can survive by making their way to parts of the submarine that are not flooded. Unfortunately, this design doesn't always work, as the Kursk disaster of the year 2000 showed. Some ML systems make use of declarative pipelines as an organizational metaphor. Keep in mind that logical pipeline boundaries often make poor trust boundaries when considered from a security perspective. Though logical boundaries are very helpful from an engineering perspective, if you want to create a trust boundary that must be done as an explicit and separate exercise. Likewise, note that containers are not always the same thing as conceptual components of the sort we have identified in this work. When you are working on compartmentalization, separation at the logical and data level is what you should be after. In many container models used commonly for ML, everything ends up in one large container without internal trust boundaries. Compartmentalization for security requires more separation of concerns. Another challenge with security and compartmentalization comes when it is time to consider the system as a whole. As we've seen in our generic ML system here, data flow between components, and sometimes those data are security sensitive. When implementing an ML system, considering component risks is a good start, but don't forget to think through the risks of the system as a whole. Harkening back to the principle of least privilege, don't forget to apply the same sort of thinking to the system as a whole after you have completed working on the components. \n Principle 6: Keep It Simple Keep It Simple, Stupid (often spelled out KISS) is good advice when it comes to security. Complex software (including most ML software) is at much greater risk of being inadequately implemented or poorly designed than simple software is, causing serious security challenges. Keeping software simple is necessary to avoid problems related to efficiency, maintainability, and of course, security. Machine Learning seems to defy KISS by its very nature. ML models involve complicated mathematics that is often poorly understood by implementers. ML frequently relies on huge amounts of data that can't possibly be fully understood and vetted by system engineers. As a result, many ML systems are vulnerable to numerous attacks arising from complexity. It is important for implementers of ML systems to recognize the drawbacks of using complicated classes of ML algorithms and to build security controls around them. Adding controls to an already complicated system may seem to run counter to our simplicity goal, but sometimes security demands more. Striking a balance between achieving defense-in-depth and simplicity, for example, is a tricky task. KISS should help inform ML algorithm selection as well as ensemble versus simple algorithm selection. What makes an adequate approach varies according to the goals and requirements of the system, yet there are often multiple choices. When such a choice needs to be made, it is important to consider not only the accuracy claims made by designers of the algorithm, but also how well the algorithm itself is understood by engineers and the broader research community. If the engineers developing the ML system don't really deeply understand the underlying algorithm they are using, they are more likely to miss security problems that arise during operations. This doesn't necessarily mean that the latest and greatest algorithms can't be used, but rather that engineers need to be cognizant of the amount of time and effort it takes to understand and then build upon every complex system. \n Principle 7: Promote Privacy Privacy is tricky even when ML is not involved. ML makes things even trickier by in some sense re-representing sensitive and/or confidential data inside of the machine. This makes the original data \"invisible\" (at least to some users), but remember that the data are still in some sense \"in there somewhere.\" So, for example, if you train up a classifier on sensitive medical data and you don't consider what will happen when an attacker tries to get those data back out through a set of sophisticated queries, you may be putting patients at risk. When it comes to sensitive data, one promising approach in privacy-preserving ML is differential privacy.  34:abadi  The idea behind differential privacy is to set up privacy restrictions that, for example, guarantee that an individual patient's private medical data never has too much influence on a dataset or on a trained ML system. The idea is to \"hide in plain sight\" with a goal of ensuring that anything that can be learned about an individual from the released information, can also be learned without that individual's data being included. An algorithm is differentially private if an observer examining the output is not able to determine whether a specific individual's information was used in the computation. Differential privacy can be achieved through the use of random noise that is generated according to a chosen distribution and is used to perturb a true answer. Somewhat counterintuitively, because of its use of noise, differential privacy can also be used to combat overfitting in some ML situations. Differential privacy is a reasonably promising line of research that can in some cases provide for privacy protection. Privacy also applies to the behavior of a trained-up ML system in operation. We've discussed the tradeoffs associated with providing (or not providing) confidence scores. Sometimes that's a great idea, and sometimes it's not. Figuring out the impact on system security that providing confidence scores will have is another decision that should be explicitly considered and documented. In short, you will do well to spend some cycles thinking about privacy in your ML system. If you are doing ML on sensitive data, you must take privacy risks seriously, and know that there are no magic solutions. (That is, if you are training a model on sensitive data to do something useful, that model must by its very nature reveal something about its training data.) \n Principle 8: Remember That Hiding Secrets Is Hard Security is often about keeping secrets. Users don't want their personal data leaked. Keys must be kept secret to avoid eavesdropping and tampering. Top-secret algorithms need to be protected from competitors. These kinds of requirements are almost always high on the list, but turn out to be far more difficult to meet than the average user may suspect. ML system engineers may want to keep the intricacies of their system secret, including the algorithm and model used, hyperparameter and configuration values, and other details concerning how the system trains and performs. Maintaining a level of secrecy is a sound strategy for improving the security of the system, but it should not be the only mechanism. Past research in transfer learning has demonstrated the ability for new ML systems to be trained from existing ones. If transfer learning is known to have been applied, it may facilitate extraction of the proprietary layers trained \"on top\" of the base model. Even when the base model is not known, distillation attacks allow an attacker to copy the possibly proprietary behavior of a model using only the ability to query the ML system externally. As a result, maintaining the secrecy of the system's design requires more than simply not making the system public knowledge. A chief concern for ML systems is protecting the confidentiality of training data. Some may attempt to \"anonymize\" the data used and consider that sufficient. As the government of Australia discovered in 2017, great care must be taken in determining that the data cannot be deanonymized.  35:culnane  Neural networks similarly provide a layer of anonymization by transforming confidential information into weights, but even those weights can be vulnerable to advanced information extraction techniques. It's up to system engineers to identify the risks inherent in their system and design protection mechanisms that minimize security exposure. Keeping secrets is hard, and it is almost always a source of security risk. \n Principle 9: Be Reluctant to Trust ML systems rely on a number of possibly untrusted, external sources for both their data and their computation. Let's take on data first. Mechanisms used to collect and process data for training and evaluation make an obvious target. Of course, ML engineers need to get their data somehow, and this necessarily invokes the question of trust. How does an ML system know it can trust the data it's being fed? And, more generally, what can the system do to evaluate the collector's trustworthiness? Blindly trusting sources of information would expose the system to security risks and must be avoided. Next, let's turn to external sources of computation. External tools such as TensorFlow, Kubeflow, and pip can be evaluated based on the security expertise of their engineers, time-proven resilience to attacks, and their own reliance on further external tools, among other metrics. Nonetheless, it would be a mistake to assume that any external tool is infallible. Systems need to extend as little trust as possible, in the spirit of compartmentalization, to minimize the capabilities of threats operating through external tools. It can help to think of the various components of an ML system as extending trust to one another; dataset assembly could trust the data collectors' organization of the data, or it could build safeguards to ensure normalization. The inference algorithm could trust the model's obfuscation of training data, or it could avoid responding to queries that are designed to extract sensitive information. Sometimes it's more practical to trust certain properties of the data, or various components, but in the interests of secure design only a minimum amount of trust should be afforded. Building more security into each component makes attacks much more difficult to successfully orchestrate. \n Principle 10: Use Your Community Resources Community resources can be a double-edged sword; on the one hand, systems that have faced public scrutiny can benefit from the collective effort to break them. But nefarious individuals aren't interested in publicizing the flaws they identify in open systems, and even large communities of developers have trouble resolving all of the flaws in such systems. Relying on publicly available information can expose your own system to risks, particularly if an attacker is able to identify similarities between your system and public ones. Transfer learning is a particularly relevant issue to ML systems. While transfer learning has demonstrated success in applying the learned knowledge of an ML system to other problems, knowledge of the base model can sometimes be used to attack the student.  28:wang  In a more general sense, the use of publicly available models and hyperparameters could expose ML systems to particular attacks. How do engineers know that a model they use wasn't deliberately made public for this very purpose? Recall our discussion of \"Trojan models\" from the attack taxonomy section above. Public datasets used to train ML algorithms are another important concern. Engineers need to take care to validate the authenticity and quality of any public datasets they use, especially when that data could have been manipulated by unknown parties. At the core of these concerns is the matter of trust; if the community can be trusted to effectively promote the security of their tools, models, and data, then community resources can be hesitantly used. Otherwise, it would be better to avoid exposing systems to unnecessary risk. After all, security problems in widely-used open-source projects have been known to persist for years, and in some cases decades, before the community finally took notice. \n Putting this Risk Analysis to Work This document presents a basic architectural risk analysis and a set of 78 specific risks associated with a generic ML system. We organize the risks by common component and also include some system-wide risks. These risk analysis results are meant to help ML systems engineers in securing their own particular ML systems. In our view ML systems engineers can devise and field a more secure ML system by carefully considering the risks in this document while designing, implementing, and fielding their own specific ML system. In security, the devil is in the details, and we attempt to provide as much detail as possible regarding ML security risks and some basic controls. We have also included a treatment of security principles as adapted in Building Secure Software and originally published in 1972 by Saltzer and Shroeder.  1:viega, 3:saltzer  This treatment can help provide an important perspective on security engineering for researchers working in ML. Figure 1 : 1 Figure 1: Components of a generic ML system. Arrows represent information flow. \n 2 . 3 . 4 . 234 Dataset assembly. Raw text is organized into sentence pairs between two languages. Sentences are segmented by a model which splits individual words into smaller wordpieces and adds special characters to the beginning of each word. (This is the best performing option proposed; they also evaluate on word-based, character-based, and a mixed model which only splits out-ofvocabulary words into a character representation.) Datasets. The parsed text pairs are separated into a training set and test set. Learning algorithm. At a high level, GNMT's learning algorithm consists of an Encoder Recurrent Neural Network (RNN), an attention module, and a Decoder RNN. \n [model: 1 : 1 improper re-use] ML-systems are re-used intentionally in transfer situations. The risk of transfer outside of intended use applies. Groups posting models for transfer would do well to precisely describe exactly what their systems do and how they control the risks in this document. Berryville Institute of Machine Learning 20 [model:2:Trojan] Model transfer leads to the possibility that what is being reused may be a Trojaned (or otherwise damaged) version of the model being sought out. 7:shokri [model:3:representation fludity] ML is appealing exactly because it flies in the face of brittle symbolic AI systems. When a model generalizes from some examples, it builds up a somewhat fluid representation if all goes well. The real trick is determining how much fluidity is too much. \n [output: 3 : 3 miscategorization] Adversarial examples (see [input:1:adversarial examples]) lead to fallacious output. If those output escape into the world undetected, bad things can happen. \n Principle 1 : 1 Secure the Weakest Link Principle 2: Practice Defense in Depth Principle 3 \n Data poisoning. Data play an outsized role in the security of an ML system. That's because an ML system learns to do what it does directly from data. If an attacker can intentionally manipulate the data being used by an ML system in a coordinated fashion, the entire system can be compromised. Data poisoning attacks require special attention. In particular, ML engineers should consider what fraction of the training data an attacker can control and to what extent. See [data:1:poisoning] below. Though coverage and resulting attention might be disproportionately large, swamping out other important ML risks, adversarial examples are very much real. See [input:1:adversarial examples] below. 2. \n 1:online], [inference:1:online], and [data:7:online] below.4. Transfer learning attack. In many cases in the real world, ML systems are constructed by taking advantage of an already-trained base model which is then fine-tuned to carry out a more specific task. A data transfer attack takes place when the base system is compromised (or otherwise unsuitable), making unanticipated behavior defined by the attacker possible. See [data:2:transfer], [model:1:improper re-use] and [model:2:Trojan] below.5. Data confidentiality. Data protection is difficult enough without throwing ML into the mix. One unique challenge in ML is protecting sensitive or confidential data that, through training, are built right into a model. Subtle but effective extraction attacks against an ML system's data are an important category of risk. See [raw:1:data confidentiality] below. 6. Data trustworthiness. Because data play an outsize role in ML security, considering data provenance and integrity is essential. Are the data suitable and of high enough quality to support ML? Are sensors reliable? How is data integrity preserved? Understanding the nature of ML system data sources (both during training and during execution) is of critical importance. Data borne risks are particularly hairy when it comes to public data sources (which might be manipulated or poisoned) and online models. See [raw:2:trustworthiness] below. \n Output integrity. If an attacker can interpose between an ML system and the world, a direct attack on output may be possible. The inscrutability of ML operations (that is, not really understanding how they do what they do) may make an output integrity attack that much easier since an anomaly may be harder to spot. See [output:1:direct] below. 9. Encoding integrity. Data are often encoded, filtered, re-represented, and otherwise processed before use in an ML system (in most cases by a human engineering group). Encoding integrity issues can bias a model in interesting and disturbing ways. For example, encodings that include metadata may allow an ML model to \"solve\" a categorization problem by overemphasizing the metadata and ignoring the real categorization problem. See both [assembly:1:encoding integrity], [raw:5:encoding integrity], and [raw:10:metadata] below. 10. \n Always consider the technical source of input, including whether the expected input will always be available. Is the sensor you are counting on reliable? Sensor blinding attacks are one example of a risk faced by poorly designed input gathering systems. Note that consistent feature identification related to sensors is likely to require human calibration. legal issues now is that there may be legal requirements to \"delete\" data (e.g., from a GDPR request). What it means to \"delete\" data from a trained model is challenging to carry out (short [raw:12:sensor] of retraining the model from scratch from a data set with the deleted data removed, but that is expensive and often infeasible). Note that through the learning process, input data are always encoded in some way in the model itself during training. That means the internal representation developed by the model during learning (say, thresholds and weights) may end up being legally encumbered as well. [raw:5:encoding integrity] Raw data are not representative of the problem you are trying to solve with ML. Is your sampling capability lossy? Are there ethical or moral implications built into your raw data (e.g., racist or xenophobic implications can be trained right into some facial recognition systems if data sets are poorly designed)? 9:phillips [raw:6:representation] Representation plays a critical role in input to an ML system. Carefully consider representation schemes, especially in cases of text, video, API, and sensors. Is your representation rich enough to do what you want it to do? For example, many encodings of images are compressed in a lossy manner. This will impact your model, figure out how. [raw:7:text encoding] Text representation schemes are not all the same. If your system is counting on ASCII and it gets Unicode, what happens? Will your system recognize the incorrect encoding and fail gracefully or will it fail hard due to a misinterpreted mismatch? [raw:8:looping] Model confounded by subtle feedback loops. If data output from the model are later used as input back into the same model, what happens? Note that this is rumored to have happened to Google translate in the early days when translations of pages made by the machine were used to train the machine itself. Hilarity ensued. To this day, Google restricts some translated search results through its own policies. [raw:9:data entanglement] Entangled data risk. Always note what data are meant to represent and be cognizant of data entanglement. For example, consider what happens if a public data source (or even an internal 4:papernot source from another project) decides to recode their representation or feature set. Note that \"false [raw:2:trustworthiness] features\" can also present an entanglement problem as the famous husky-versus-wolf classifier Data sources are not trustworthy, suitable, and reliable. How might an attacker tamper with or demonstrated by acting (incorrectly) as a snow detector instead of a species detector. Know which otherwise poison raw input data? What happens if input drifts, changes, or disappears? 8:barreno parts of your data can change and which should not ever change. 10:sculley [raw:3:storage] [raw:10:metadata] Data are stored and managed in an insecure fashion. Who has access to the data pool, and why? Metadata may help or hurt an ML model. Make note of metadata included in a raw input dataset. Access controls can help mitigate this risk, but such controls are not really feasible when utilizing Metadata may be a \"hazardous feature\" which appears useful on the face of it, but actually public data sources. This kind of risk brings to mind early attempts to create mathematically random degrades generalization. Metadata may also be open to tampering attacks that can confuse data for cryptographic security through combining sets of inputs that could ultimately be influenced an ML model. 11:ribeiro More information is not always helpful and metadata may harbor spurious by an attacker (such as process id, network packet arrival time, and so on). Needless to say, entropy correlations. Consider this example: we might hope to boost performance of our image classifier pools controlled by an attacker are low entropy indeed. Ask yourself what happens when an by including exif data from the camera. But what if it turns out our training data images of dogs are attacker controls your data sources. all high resolution stock photos but our images of cats are mostly facebook memes? Our model will [raw:4:legal] probably wind up making decisions based on metadata rather than content. Note that public data sources may include data that are in some way legally encumbered. An [raw:11:time] obvious example is copyrighted material that gets sucked up in a data stream. Another more If time matters in your ML model, consider time of data arrival a risk. Network lag is something insidious example is child pornography which is never legal. A third, and one of the most interesting easily controlled by an attacker. Plan around it. 12 Berryville Institute of Machine Learning \n an ML system actually does what it does. ML systems have a number of hyperparameters, including, for example, learning rate and momentum in a gradient descent system. These parameters are those model settings not updated during learning (you can think of them as model configuration settings). Setting and tuning hyperparameters is somewhat of a black art subject to attacker influence. If an attacker can twiddle hyperparameters (tweaking, hiding, or even introducing them), bad things will happen. (Also see [inference:3:hyperparameters].)[alg:10:hyperparameter sensitivity] Oversensitive hyperparameters are riskier hyperparameters, especially if they are not locked in. Sensitive hyperparameters not rigorously evaluated and explored can give you a weird kind of overfitting. For example, one specific risk is that experiments may not be sufficient to choose good hyperparameters. Hyperparameters can be a vector for accidental overfitting. In addition, hard to detect changes to hyperparameters would make an ideal insider attack. [alg:11:parameters] In the case of transfer learning (see [data:2:transfer]) an attacker may intentionally post or ship or otherwise cause a target to use incorrect settings in a public model. Because of the open nature of ML algorithm and parameter sharing, this risk is particularly acute among ML practitioners who naively think \"nobody would ever do that.\" \n 28:wang (6)  Model extraction occurs when an attacker targets a less-than-fully-white-box ML system, attempting to \"open the box\" and copy its behavior or parameters. Model extraction may function as theft of a proprietary model or may enable white-box attacks on a formerly blackbox model.29:papernot,30:gilmer     \n Regardless of whether you trust your friend, there's really no need to put yourself at risk by giving more access than necessary. For ex ample, if you don't have pets, but only need a friend to pick up the mail on occasion, you should relinquish only the mailbox key. Although your friend may find a good way to abuse that privilege, at least you don't have to worry about the possibility of additional abuse. If you give out the house key unnecessarily, all that changes.Similarly, if you do get a house sitter while you're on vacation, you aren't likely to let that person keep your keys when you're not on vacation. If you do, you're setting yourself up for additional risk. Whenever a key to your house is out of your control, there's a risk of that key getting duplicated. If there's a key outside your control, and you're not home, then there's the risk that the key is being used to enter your house. Any length of time that someone has your key and is not being supervised by you constitutes a window of time in which you are vulnerable to an attack. You want to keep such windows of vulnerability as short as possible-to minimize your risks.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/BIML-ARA.tei.xml", "id": "7fde4960c2a6203ea1952bba9693f7f3"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Many in the national security community are concerned about China's rising dominance in artificial intelligence and AI talent. That makes leading in AI workforce competitiveness critical, which hinges on developing and sustaining the best and brightest AI talent. This includes top-tier computer scientists, software engineers, database architects, and other technical workers that can effectively create, modify, and operate AI-enabled machines and other products. This issue brief informs the question of strategic advantage in AI talent by comparing efforts to integrate AI education in China and the United States. We consider key differences in system design and oversight, as well as in strategic planning. We then explore implications for maintaining a competitive edge in AI talent. (This report accompanies an introductory brief of both countries' education systems: \"Education in China and the United States: A Comparative System Overview.\")", "authors": ["Dahlia Peterson", "Kayla Goode", "Diana Gehlhaus"], "title": "AI Education in China and the United States", "text": "Both the United States and China are making progress in integrating AI education into their workforce development systems, but are approaching education goals in different ways. China is using its centralized authority to mandate AI education in its high school curricula and for AI companies to partner with schools and universities to train students. Since 2018, the government also approved 345 universities to offer an AI major, now the country's most popular new major, and at least 34 universities have launched their own AI institutes. The United States is experimenting with AI education curricula and industry partnership initiatives, although in a piecemeal way that varies by state and places a heavier emphasis on computer science education. Both countries' approaches could result in uneven levels of AI workforce competitiveness, although for similar and different reasons. China's centralized push could lead to widespread integration of AI education, but the resulting curricula could be shoddy for the sake of participating in the \"AI gold rush.\" This risk is especially pronounced in under-resourced areas, which could produce underwhelming results. The United States' varied, decentralized approach may allow for greater experimentation and innovation in how AI curricula are developed and implemented, but diverse approaches may exacerbate disparities in curriculum rigor, student achievement standards, and educator qualifications. As for similarities, the two countries share hurdles such as the rural-urban divide, equitable access to quality AI education, and teacher quality. Ultimately, this report suggests future U.S. science and technology education policy should be considered in a globally competitive context instead of viewing it exclusively as a domestic challenge. For the United States, that consideration includes recognizing and capitalizing on its enduring advantage in attracting and retaining elite talent, including Chinese nationals. While this brief does not make policy recommendations for the U.S. education system, the upcoming CSET report \"U.S. AI Workforce: Policy Recommendations\" addresses some of the direct implications of the findings presented.  \n Table of Contents Executive \n Introduction Much has been written on how the Chinese government recruits foreign artificial intelligence talent. However, little is known about China's ongoing initiatives to build their own AI workforce. Existing scholarship also lacks a detailed examination of how Chinese and U.S. approaches differ when it comes to AI education. To fill this gap, this issue brief details how both countries are integrating AI education and training into every level of education. It discusses potential national security implications of each country's strengths and weaknesses, and highlights improvement areas for future U.S. science and technology (S&T) education and workforce policy. We aim to provide a clear-eyed assessment of the U.S. approach to AI education as it exists within the country's decentralized education system. A discussion of the strengths and weaknesses of these systemic realities, especially relative to China's system, will help policymakers better address critical barriers to U.S. AI competitiveness. The research presented in this brief is based on primary source U.S. statistics, reports and assessments from education nonprofits, publicly available information from the private sector, and individual states' departments of education, along with Chinese education plans and policies, official statistics, and translations. The data is often defined and categorized differently, making uniform comparisons difficult. We attempt to clarify such differences when they occur. \n Overview of China's Education System China's education system is overall more centralized than its U.S. counterpart. Its education system includes 282 million students in 530,000 educational institutions across all levels. China's Ministry of Education is the main authority overseeing China's education system, and is responsible for certifying teachers, setting national education goals, curricula and teaching material, and providing limited funding assistance.  1  While the MOE supervises provincial education departments, it has granted more implementation responsibility to the provincial and municipal levels over recent decades.  2  Responsibilities at the provincial and major city level include following national guidelines to develop provincial curricula based on developing an implementation plan that incorporates local contexts and MOE national curriculum guidance, then sending the plan to the MOE for approval before implementation.  3  Further local responsibilities include administering teaching materials, school programs, providing education subsidies, and setting additional standards for teacher training.  4  The MOE establishes goals for its education system through fiveto 15-year education strategies. The goals for 2010-2020 include universalizing preschool education; improving nine-year compulsory education; raising senior high school gross enrollment rate to 90 percent (which has already been exceeded); and increasing the higher education gross enrollment rate to 40 percent. Provinces then typically follow to create their own education plans.  5  The MOE's Bureau of Education Inspections monitors implementation and provides feedback to local governments.  6  For details on China's education system, see the accompanying brief \"Education in China and the United States: A Comparative System Overview.\" \n Overview of the United States' Education System The U.S. education system is more decentralized than its Chinese counterpart, especially for primary and secondary education. Each state's department of education is the authority that determines the laws that finance schools, hire educators, mandate student attendance, and implement curricula. In contrast to China's MOE, the U.S. federal government provides relatively minor education oversight through the compilation and reporting of education statistics, along with promoting equitable access to education and enforcing a prohibition on institutional discrimination.  7  The U.S. Department of Education, the United States' federal agency for education, proclaims that education is a \"state and local responsibility,\" and the federal government's role in education is more of a \"kind of emergency response system\" to fill gaps when \"critical national needs arise.\"  8  The most notable federal education initiatives, such as the Elementary and Secondary Education Act of 1965, the No Child Left Behind Act of 2002, and the Every Student Succeeds Act of 2015, reflect the U.S. government's efforts to promote childrens' equal access to quality public education. At the postsecondary level, the federal government has slightly more authority through its administration of student financial aid. The Department of Education supports programs that provide grants, financial aid (loans), and work-study assistance. Roughly 66 percent of students apply for federal financial assistance.  9  The department's student loan programs have more than 43 million outstanding borrowers, with outstanding student debt now over $1.7 trillion.  10  The jurisdiction of the U.S. Department of Education is rooted in the U.S. Constitution. As a result of the division in constitutional authority, states develop curriculum guidelines and performance standards, license private elementary and secondary schools to operate within their jurisdictions, certify teachers and administrators, administer statewide student achievement tests, and distribute state and federal funding to school districts.  11  Additionally, education in the United States is segmented between public and private schools, including religious and nonsectarian institutions. For details on the United States' education system, see the accompanying brief \"Education in China and the United States: A Comparative System Overview.\" \n Integration of AI Education in China Since 2017, China has released several strategic plans relevant to AI education. The most well-known of the plans-the State Council's seminal July 2017 New Generation AI Development Plan-called for implementing AI training at every level of education. Another major push in AI education is through the \"Double First Class University\" (双一流大学) initiative, a 2017 program under Chinese President Xi Jinping that built upon previous reforms such as Projects 211 and 985 to create worldclass universities. 12 Nearly all of the MOE's directly supervised and funded 75 universities are also \"Double First Class.\" The initiative split universities into two tracks: 42 universities were selected as world-class universities, and split respectively into 36 \"Class A\" (already close to being world class) and 6 \"Class B\" (potential to be world class) universities.  13  This initiative essentially pared down the number of top universities China was focusing on.  14     \n Primary and Secondary AI Education China is actively integrating AI education into young students' education. These efforts are primarily characterized at the primary school level with introductory Python courses, and access to labs featuring robotics, drones, and 3D printing. Local governments have recently begun awarding schools for excellence. Since 2018, the MOE has mandated high schools to teach AI coursework. \n Role of Talent Training Bases Talent training bases are one of the newer ways that AI education is gaining momentum at the primary and secondary level. Shandong Province and Beijing began awarding \"National Youth AI Innovation Talent Training Base\" (全国青少年人工智能创新人才培 养基地) honors to their schools between November and December 2020.  15  Schools are chosen for demonstrating excellence in AI education. Primary schools are rewarded for offering rudimentary \"technology and society\" and \"maker\" courses and access to 3D printing and drones. Junior high schools gain recognition for \"AI clubs,\" robotics rooms, 3D printers, open source AI frameworks, utilize Python, and have won robotics competitions.  16  The certification, according to images of award placards, is for two years. 17 Shandong awarded 10 schools the certification, while Beijing awarded 21, including a S&T museum for youth.  18  Beijing appears to have had a strong start to its AI education program. As part of the talent training base initiative, one hundred teachers were awarded certificates to be \"AI literacy assessors.\"  19  As part of the project, the Beijing Youth AI Literacy Improvement Project also laid out numerical goals for its next three to five years: it plans to support the creation of ten AI experimental areas, create one hundred AI education experimental schools, select thousands of AI education seed teachers, and train ten thousand young AI talents.  20  As the program is only a few months old, it remains to be seen what quality control mechanisms will be used, and how well Beijing will implement its goals. \n Mandated High School AI Curriculum At the high school level, the MOE in January 2018 revised its national education requirements to officially include AI, Internet of Things, and big data processing in its information technology curricula.  21  The revision requires high school students enrolled in the fall of 2018 and beyond to take AI coursework in a compulsory information technology course.  22  The coursework goals include data encoding techniques; collecting, analyzing, and visualizing data; and learning and using a programming language to design simple algorithms.  23  Python is a popular choice, and is even being integrated into the Gaokao as testing material in Beijing as well as Zhejiang and Shandong provinces.  24  This integration may incentivize high school students to develop Python expertise at an earlier age, and prepare them for further training at the university level and beyond. Further goals include understanding AI safety and security, and an emphasis on ethics. However, there is also a distinct emphasis on \"learning to abide by relevant laws,\" which could channel learning in directions considered suitable to the Party-state's needs. \n Teacher Conferences At both the primary and secondary school level, a prominent planning and information sharing mechanism is the National Primary and Secondary School Artificial Intelligence Education Conference, which is sponsored by the nonprofit Chinese Association for Artificial Intelligence's Primary and Secondary School Working Committee.  25  The conference focuses on teacher training, use of education platforms, and curriculum design.  26   \n Role of Private Sector Companies and schools often partner to create textbooks. In April 2018, SenseTime and the Massive Open Online Course Center of East China Normal University launched the first domestic AI textbook Artificial Intelligence Basics (High School Edition) for middle school students.  27  According to SenseTime, it is currently being taught in pilot programs in more than one hundred schools throughout the country in Shanghai and Beijing, as well as Guangdong, Heilongjiang, Jiangsu, Shandong, and Shanxi provinces. SenseTime is also training over nine hundred teachers to teach the material.  28  In June 2018, Soochow University Press published the \"Artificial Intelligence Series for Primary and Secondary Schools.\" In November 2018, a Tencent-backed, globally active company named UBTech Robotics, and East China Normal University Press jointly released the \"Future Intelligent Creator on AI-Series of Artificial Intelligence Excellent Courses for Primary and Secondary Schools.\"  29  More recently, Tsinghua University announced in January 2020 that Yao Qizhi-China's first winner of the Turing Award, academician of the Chinese Academy of Sciences, and dean of Tsinghua's Institute of Interdisciplinary Information Sciences-would be editing the textbook Artificial Intelligence (High School Edition).  30  Chinese tech giants such as Baidu are also helping to introduce AI to vocational secondary schools. For example, in July 2019, Baidu Education and Beijing Changping Vocational School launched China's first vocational school-enterprise cooperation initiative on AI education.  31  The cooperation identified five dimensions: jointly building Baidu's artificial intelligence innovation space, jointly building artificial intelligence technology and application majors, jointly building training bases for vocational college instructors, jointly carrying out teacher-student skills exchange competitions, and jointly building small and medium-sized subject general experience bases.  32  Issues Faced AI education at the primary and secondary school levels faces notable issues. Local education consultants note that one issue is an overly difficult curriculum for young children, especially when students require significant background knowledge to understand algorithms powered by deep learning.  33  Additional issues include a lack of systematic and authoritative guidance on textbook development, lack of professional training for teachers, and lack of equipment in underequipped school AI labs.  34   \n Undergraduate AI Education China's AI education push is most prominent at the postsecondary level. The following sections will examine the two main mechanisms for talent training: AI institutes, and the MOE's standardized AI major. The AI major is now the country's most studied field.  35   \n AI Institutes AI institutes largely preceded the MOE's development of the AI major. Both preceding and following the MOE's release of the \"AI Innovation Action Plan for Colleges and Universities\" in 2018, at least 34 institutions launched their own AI institutes between 2017-2018 (see Figure  2 ).  36  In 2018, three Seven Sons of National Defense universities joined these ranks. The Seven Sons are directly supervised by the Ministry of Industry and Information Technology (MIIT). Their core mission is to support the People's Republic of China's defense research, its industrial base, and military-civil fusion to merge civilian research into military applications.  37  Aside from three Seven Sons in 2018, at least three universities even launched their institutes before the July 2017 Next Generation AI Plan from July 2017. 38 However, it is common for institutes' \"About Us\" pages to cite the \"AI Innovation Action Plan\" and the national AI plan as its reason for creation.  39  Unlike the AI major, which is clearly targeted as an undergraduate major, institutes are significantly more heterogeneous in their research foci, which range from natural language processing to robotics, medical imaging, smart green technology, and unmanned systems.  40  Likewise, they often train both undergraduate and graduate students, and in some cases offer the AI major within their institution.  41  Companies also play an establishing role for AI institutes. For example, Chongqing University of Posts and Telecommunications set up an AI institute in 2018 with AI unicorn iFlytek, while Tencent Cloud established AI institutes in 2018 with Shandong University of Science and Technology and Liaoning Technical University.  42  It is beyond the scope of this paper to examine the AI institutes' quality indicators. A forthcoming CSET data brief will examine the landscape of AI institutes, their research foci, the degree of overlap with the AI major, and their relationship to China's key laboratories in greater detail. \n Standardized AI Major In March 2019, the MOE approved 35 colleges and universities to offer the four-year AI major as an engineering degree, including four of the Seven Sons of National Defense universities.  43  Half of the institutes (17) that had previously launched AI institutes were later formally approved to launch the new AI major in the 2019-2020 range. The approval of a new AI major was a notable change from past curricula, when AI was available as a concentration within the computer science major at some universities. In February 2020, the MOE approved 180 more universities-a fivefold increase-to offer the AI major, bringing the total number of approved universities to 215. One of the approved was a fifth member of the Seven Sons.  44  In March 2021, the MOE approved 130 universities, including the sixth and seventh of the Seven Sons, bringing the total university count to 345.  45  In both 2020 and 2021, the AI major was the most popular new addition to universities' curricula; in 2021, the next most popular majors included 84 universities offering intelligent manufacturing and engineering, and 62 offering data science and big data technology.  46  Additionally, eight universities that had launched AI institutes between 2016-2018 have also begun offering the AI major. The vast majority of these universities are not well known or are business oriented, with the clear exception of Tsinghua. \n Tsinghua University's AI Offerings Tsinghua had an early foray into AI teaching. In 1979, the predecessor to today's Department of Electronic Engineering opened the \"Introduction to Artificial Intelligence\" course, which was one of the earliest AI courses offered by any Chinese university.  47  In the late 1980s, the Department of Computer Science and Artificial Intelligence of Tsinghua University established the State Key Laboratory of \"Intelligent Technology and Systems,\" also known as the \"Institute of Human Intelligence.\"  48  While Tsinghua was only approved in March 2021 to offer the AI major, it already offered AI education via various channels and opened its interdisciplinary Institute of Artificial Intelligence in June 2018. In May 2020, Tsinghua announced the creation of an AI \"smart class\" (智班), which will be the 8th experimental class of the \"Tsinghua Academy Talent Training Program.\" Yao Qizhi, who also spearheaded the aforementioned high school AI textbook, will serve as the lead faculty in an interdisciplinary \"AI + X\" approach, which entails integrating AI with mathematics, computer science, physics, biology, psychology, sociology, law, and other fields.  49  In 2019, Tsinghua also combined three minors-Robot Technology Innovation and Entrepreneurship (机器人技术创新 创业), Intelligent Hardware Technology Innovation and Entrepreneurship (智能硬件技术创新创业), and Intelligent Transportation Technology Innovation and Entrepreneurship (智能交通技术创新创业)-into one new major titled Artificial Intelligence Innovation (人工智能创新).  50  Figures 2 and 3 breakdown China's 345 universities with AI majors and 34 AI institutes, as well as the number that are \"Double First Class,\" Seven Sons, and the Ivy League-equivalent C9 League.  51  Figure  2  presents the information geographically while Figure  3  provides a breakdown by type of institution.  A Chinese AI company called KXCY AI working with several elite Chinese universities suggests that the AI major's goals are to meet national economic and technological development needs, develop knowledge of basic AI theories, learn research and development (R&D) skills, along with system design, management, and solving complex engineering problems in AI and related applications.  52  Further, colleges and universities with existing AI programs were encouraged by the MOE in 2018 to expand their scope to establish \"AI + X\" majors.  53  Beyond the AI major, AI-adjacent majors include data science (数据 科学), a major in its fifth year of operation, while other majors include big data technology (大数据技术), intelligent manufacturing (智能制造), robotics engineering (机器人工程), and intelligent science and technology (智能科学与技术).  54   \n Role of Private Sector China's AI enterprises play a significant role in developing AI talent and providing resources to universities through formalized partnerships. One such mechanism is the Information Technology New Engineering Industry-University-Research Alliance (信息技术 新工科产学研联盟, or AEEE), founded in 2017 to bolster technological innovation within the industry-university-research nexus. Its founding members included the China Software Industry Association, 27 domestic universities, five research institutes, and 12 companies, with support from MOE's Higher Education Department and MIIT.  55  Chinese companies include Baidu, Alibaba Cloud, Tencent, Huawei, and China Telecom. Of the 27 Chinese universities, 21 offer either an AI major, institute, or both. The roster further includes all Seven Sons of National Defense, and the entire C9 League.  56  The AEEE universities include China's most elite institutions. Of note is that the AEEE also includes U.S. companies Cisco, IBM, and Microsoft.  57  A blog post from Microsoft Research Asia (MSRA) from April 2019 indicates its founding role in the alliance's AI Education Working Committee, its work towards implementing the MOE's \"industry-education integration\" (产教融合理念), as well as the award it subsequently received from the alliance for its contributions in curriculum construction, resource sharing, and teacher training.  58  MSRA further states it has partnerships with University of Science and Technology of China, Xi'an Jiaotong University, Harbin Institute of Technology (a Seven Sons university with its own AI institute and major), Nanjing University and Zhejiang University, while its AI education seminars have helped three hundred colleges and universities across the country and more than five hundred AI teachers.  59  In March 2019, MSRA also worked with another Seven Sons institute, Beihang University, to open a course called \"Real Combat in Artificial Intelligence,\" which attracted at least 30 Beihang students from across 10 majors that year.  60  Baidu also plays a prominent role in both the alliance and beyond. It is one of China's three internet giants, and an \"AI champion\" as appointed by the Ministry of Science and Technology, working on autonomous driving.  61    \n Graduate AI Education As alluded to in Figure  1 , China signaled it understood the importance of graduate students advancing AI development when the MOE, National Development and Reform Commission, and the Ministry of Finance released a plan in January 2020 calling for increased training for AI graduate students. The plan's key goals are centered on using the aforementioned \"Double First Class\" program and interdisciplinary \"AI + X\" framework to bolster talent development and increase the number of grad students studying AI, especially at the doctoral level.  64  Talents are called upon to apply AI to industrial innovation, social governance, national security, and other fields, and support the mission and needs of major national projects and major development plans.  65  Additionally, the plan stated AI will be incorporated into the \"Special Enrollment Plan for the Cultivation of High-level Talents in Key Fields Urgently Needed by the State\" (\"国家关键领域急需高层 次人才培养专项招生计划\"). This incorporation has already led to increased supply of AI doctoral positions at institutes such as Nankai University.  66  The plan also emphasizes quality control. The degree evaluation committee has established an artificial intelligence working group (高校学位评定委员会设立人工智能专门工作组) responsible for developing advanced AI talent training programs, degree standards and management norms, and performing random quality control inspections of AI dissertations. As with all levels of AI education, China's AI companies are asked to partner with universities-partnerships focus both on training teachers and students. At the postgraduate level, there is meant to be a revolving door between industry and universities. First, at the instructor level, companies are encouraged to train university instructors in the latest cutting-edge methods. Instructors are asked to incorporate the latest research findings in AI into doctoral courses. From the company side, leading researchers at these companies can also do research through \"double employment\" at universities. Second, companies are asked to train graduate students by having them solve industry needs. For doctoral students, enterprises are also encouraged to open up \"scenario-driven,\" application-oriented courses, as well as their data, case studies, tools, and training platforms. Enterprises can also utilize industry alliances, joint R&D labs, entrepreneurship and skills competitions, and certification training to help graduate students grow in the field. There is also an emphasis on increased and coordinated funding at the university and enterprise levels. Universities are encouraged to coordinate various resources, such as financial investment and scientific research income, increase support for graduate training, carry out basic frontier research and key common technology research efforts. With enterprises, methods such as angel investment, venture capital, and capital market financing can boost major AI projects and applied research within universities, as well as assist with talent training. Although it remains too early to tell how well these policies will tangibly advance China's talent training and AI capabilities, China's efforts over the last 20 years in prioritizing and progressing in key areas such as hypersonics and biotechnology indicate that policies are not just wish lists, but backed up by measurable results. Nonetheless, we caution that widespread, haphazard curricula construction for the sake of participating in the \"AI gold rush\" could hinder these successes. If China is able to fulfill the tall order of ensuring rigorous quality control from its wealthy metropolises to its under-resourced regions, we assess that China could replicate its past successes in policy implementation. \n Integration of AI Education in the United States Integration of AI curricula in the United States is uneven in both depth and scope as it remains the work of states and local school districts to finalize and implement curriculum standards. In recent years, various organizations and companies have helped facilitate progress through efforts to define and establish AI curricula, programs, learning standards, and resources for teachers and students who want to add AI to their education. \n Primary and Secondary AI Education \n AI and Computer Science Curriculum Numerous states have recently taken steps toward integrating AI into K-12 classrooms. This often starts with computer science curricula, when available at the school, as it is viewed as a first step towards an AI specialization. However, many schools still lack CS education. From 2019 to 2020, 28 states adopted policies to support K-12 CS education, with efforts disparate across and within states. According to a 2020 report from Code.org, the portion of public high schools teaching CS are as high as 89 percent in Arkansas and as low as 19 percent in Minnesota.  67  Moreover, only 22 states have adopted policies that provide all high school students access to CS courses, and of those states, only nine give all K-12 students access.  68  Only 47 percent of U.S. public high schools teach CS,  69  although that is up from 35 percent in 2018.  70  Within the movement among states to adopt CS curricula, the content and learning standards of such curricula will vary by state as each maintains authority over curriculum standards. Different organizations have formed unique versions of CS education standards that present schools with options or frameworks for teaching AI in the classroom. The professional association Computer Science Teachers Association in 2017 released a \"core set of standards\" to guide teachers in computer science.  71  The International Society for Technology in Education, a nonprofit headquartered in Washington, DC, collaborated with the CSTA to release its ISTE Standards for Computer Science Educators in 2018.  72  CSTA also played a role in developing another set of K-12 CS guidelines, along with the Association for Computing Machinery, Code.org, the Cyber Innovation Center, CS4ALL, and the National Math and Science Initiative, to develop the K-12 Computer Science Framework. Nonprofit Code.org offers K-12 students and educators CS curricula embedded within STEM courses already offered.  73  The initiative serves as a conceptual guide to inform curriculum development, and it prioritizes computing systems, cybersecurity, data science, algorithmic programming, and social and cultural impacts of computing.  74  Not all STEM and CS education initiatives explicitly state the connection to AI education, but some initiatives do. AI4K12 is an initiative born from a partnership between the Association for the Advancement of AI and CSTA and was funded by the National Science Foundation (NSF) and Carnegie Mellon University.  75  AI4K12's goal is to develop national AI education guidelines for K-12. It currently serves as an online repository of resources for educators. The nonprofit AI Education Project is another example of an organization offering its own AI curriculum to schools and educators nationwide. At least two public schools have developed their own AI curricula. Seckinger High School in Gwinnett County, Georgia, will be the first to introduce a curriculum for AI education to an entire K-12 cohort.  76  The North Carolina School of Science and Mathematics, a high school administered by the University of North Carolina system, announced The Ryden Program for Innovation and Leadership in AI to teach high schoolers how to design, use, and understand AI.  77  Thus far, the content of these AI curricula can vary. Seckinger High School's AI curriculum, for example, teaches coding, machine learning, design thinking, data science, and ethics and philosophical reasoning at each level of its K-12 education.  78  The Ryden Program emphasizes that students understand \"how to merge humanity with machine learning,\" and how AI can solve complex problems affecting society.  79   \n Teacher Training For public school educators, each state governs the formal licensing and accreditation process. This process can vary widely by state, and even year to year, as states implement reforms or update testing, certification, or license requirements.  80  Moreover, private or religious schools usually have more autonomy in setting requirements for educators, which introduces another degree of variation for how AI curricula might be taught.  81  These variations could result in uneven standards for educators and therefore affect quality. Ultimately, multiple components factor into preparing educators to teach AI and CS, such as community support groups, training and professional development, or microcredentials or certificates. Therefore, an additional hurdle to competitive teacher training is adequate access to the resources required to teach AI effectively.  82  Some states have already spent years integrating CS curriculum in schools and training its educators accordingly.  83   \n Role of Private Sector The private sector is supporting CS and AI integration into primary and secondary education through a number of different initiatives with distinct goals. CSET analysis found over three dozen such programs; all are a mix of private companies, nonprofits, and public-private partnerships that offer AI curricula, learning materials, and conferences for students and educators in the United States.  84  For example, Microsoft Philanthropies operates the Technology Education and Literacy in Schools (TEALS) program in roughly 455 high schools with the goal to \"build sustainable CS programs\" for underserved students and schools, especially rural ones.  85  Amazon Future Engineer, a sub organization of Amazon, supports CS curriculum access to teachers in underserved public school districts and claims to reach more than 550,000 K-12 students.  86  In addition to working AI education into the classroom, the private sector is expanding opportunities to learn about AI outside of the classroom. An estimated three hundred different organizations now offer AI or CS summer camps to K-12 students. Private companies organize a number of these and host them at competitive universities such as Stanford, the Massachusetts Institute of Technology, and Carnegie Mellon. Companies and organizations have also expanded AI learning opportunities to K-12 students by offering after school programs, competitions, and scholarships. Some initiatives explicitly focus on reaching U.S. students that otherwise may not have access to AI education through formal schooling. Girls Who Code, TECHNOLOchicas, and Black Girls CODE are just three of many organizations that specifically service underrepresented groups in CS education to address gender and race disparities in CS education.  87   \n Issues Faced Differences in the design and deployment of AI education across states make it difficult for U.S. schools to consistently define \"AI education,\" justify investment in AI education with limited resources, and provide adequate training to educators. Whereas some AI education initiatives prioritize CS, programming languages, math, and data science, others emphasize nontechnical areas such as societal and ethical impacts of AI applications.  88  For example, private company ReadyAI's programs emphasize the nontechnical components of AI, such as art and multimedia, whereas Microsoft TEALS' curriculum focuses on technical skills such as Java and Python programming languages. Investment in AI education also becomes a challenge since such a piecemeal approach makes it difficult for educators and education leaders to assess which learning programs are effective. Whichever approach states pursue, the result is a disjointed implementation of AI education. Accordingly, private and nonprofit sector organizations are taking on efforts to coordinate across state lines in conjunction with state governments and professional teachers' associations. Still, most of these efforts function as guides, resources, and suggested standards. \n Postsecondary AI Education For most U.S. colleges and universities, the closest thing to an AI major remains CS degrees with AI concentrations or specializations. Nevertheless, these postsecondary CS/AI course offerings are growing rapidly. According to the 2021 AI Index Report, the number of AI-related undergraduate courses increased by 102.9 percent and 41.7 percent at the graduate level in the last four years.  89  Growth rates in undergraduate and master's CS degrees were similarly high. Doctorate awards in CS have also increased, by 11 percent since 2015, as have those with a specialization in AI.  90  In addition to the increase in AI course offerings, many federal agencies are also prioritizing investment in AI higher education. For example, the 2021 National Defense Authorization Act directs the NSF to fund AI initiatives for higher education, such as fellowships for faculty recruitment in AI, as well as AI curricula, certifications, and other adult learning and retraining programs.  91   \n Undergraduate AI Education At the undergraduate level, integration of AI education varies and often depends on the type of institution or the availability of industry collaboration. In community colleges and technical schools, there are few AI-specific education initiatives with the exception of several recent industry partnerships with companies such as Amazon, Google, and IBM.  92  Moreover, some states have better resourced community college CS programs that may be better equipped to integrate AI curricula. For example, over 10 percent of CS associate degree graduates come from Northern Virginia Community College, one of the community colleges partnering with Amazon to offer a cloud computing curriculum.  93  Data from the 2021 AI Index Report show that many universities have augmented bachelor's degrees in CS with a specialization in AI or machine learning.  94  For example, Stanford University and the California Institute of Technology both offer bachelor's of science degrees in CS with an AI specialization track. The University of Illinois confers a bachelor's of science in computer engineering with a specialization in AI, robotics, and cybernetics. The University of Minnesota's AI specialization track includes classes in AI, machine learning, data mining, robotic systems, and computer vision. The University of California San Diego offers a computer science degree with an area of interest in AI through its Jacobs School of Engineering. Moreover, some universities are making AI educational content available to everyone. CalTech and Stanford, for example, are opting to provide AI-related coursework online to both students and the general public. Harvard University and MIT, through their online learning platform edX, also provide both free and low-cost options for AI education. \n Graduate AI Education In contrast to undergraduate programs, more postgraduate institutions offer AI majors and degrees instead of AI specializations within a CS degree, particularly at the master's level. For example, Colorado State University, UC San Diego, and Johns Hopkins are just a few universities now offering master's of science degrees in AI. At the doctorate level, some programs offer a PhD in AI or ML, while many others mirror the undergraduate approach to offering a degree in CS with an AI specialization. Also at the graduate level, a number of globally competitive U.S. universities are leading in AI education and research. Such programs are known for industry and government collaboration: Carnegie Mellon University's AI and CS programs collaborate with both the private sector and government in AI-related fields such as autonomy, robotics, and 3D printing. Amazon AWS is a sponsor of CalTech's AI4Science Graduate and Postdoctoral Fellows program, a cross-disciplinary approach to AI research. MIT's Computer Science and Artificial Intelligence Laboratory is recognized as a hub for AI innovation and is home to nearly three dozen research groups, research centers, and communities of research in AI and machine learning.  95  As previous CSET research has shown, at the graduate level, the United States is able to attract and train top-tier AI talent.  96   \n Role of Private Sector The private sector is assisting in the development of course design for postsecondary AI education, from certificate and online learning programs to two-and four-year degrees. For example, IBM is collaborating with online learning platform Simplilearn for its certificate in AI. AWS offers machine learning courses on learning platform edX, and Google has a similar partnership with Udacity and offers deep learning courses on its online education platform. Partnerships also exist between industry and degree-granting institutions. For example, Amazon is partnering with the Virginia Community College System and six Virginia universities to teach students cloud computing. In Arizona, the Maricopa County Community College District is working with Intel on Arizona's first AI certificate and degree program. IBM's P-TECH program prepares high schoolers and community college students for careers in tech, also offering free online courses in AI and cloud computing. In 2020, the NSF announced its funding of a joint industrygovernment AI initiative that includes the U.S. Department of Agriculture, the National Institute of Food and Agriculture, the U.S. Department of Homeland Security Science and Technology Directorate, and the U.S. Department of Transportation Federal Highway Administration. The initiative established 18 new AI Research Institutes, which each have a component related to AI education.  97  Industry cosponsors, Accenture, Amazon, Google, and Intel, put $160 million toward the institutes, which will support a range of interdisciplinary AI research ranging from cyber infrastructure, biology, ethics, and agriculture.  98   \n Implications for U.S. AI Workforce Competitiveness The relative efforts of the United States and China to build a globally competitive domestic AI workforce have potentially major and long-term national security implications. Policymakers at all levels of the U.S. government would be wise to consider these implications when designing future S&T education and workforce policies. We discuss strengths and shortcomings for each country followed by an assessment of implications. \n China's Strengths and Shortcomings Over the last decade, China made significant progress in its short-, medium-, and long-term strategic plans for cultivating an S&T workforce. It has fairly effectively embedded AI education and training into each phase of the workforce development pipeline. China's progress in incorporating AI education at all levels is notable given the large size of its student population relative to the United States. Figure  4  shows 2019 enrollment totals by level of education for China and the United States (2019 graduate totals by level of education are provided in the Appendix). China maintains a cumulative numerical advantage until the graduate level, after which the United States retains a slight lead. This lead disappears when not counting foreign-born students, who comprise about 14 percent of total graduate enrollment. However, as a share of total population in each country, the United States remains far ahead. Although the United States leads in graduate enrollment, when looking at graduate breakouts in STEM, the lead reverses. In fact, China's lead in advanced STEM education has only increased over the last five years. As shown in Figure  5 , China is graduating more doctorate degrees in science and engineering than the United States in STEM. It is clear that advanced degrees in China emphasize these disciplines -59 percent of doctorates and 41 percent of masters awarded in China in 2019 were in these two fields, compared to about 16 percent of U.S. graduate degrees being in STEM.  Still, China's progress is not without challenges. One of its largest hurdles to having a more uniformly educated AI workforce from all parts of the country is the household registration system, or 户口 (hukou), which controls intra-regional labor distributions by assigning legal residents permits based on the head of household's birthplace.  101  Despite ongoing reforms, the hukou system still provides stratified availability of social benefits based on individuals' registration, rather than where they live.  102  This makes it very difficult for youths with rural hukou to obtain education beyond the compulsory level.  103  Inequities stemming from the urban-rural divide also bleed into China's AI education pipeline. For example, while well-resourced primary and secondary schools in Beijing can afford to build 3D printing and robotics labs, poorer parts of the country may not have such advantages from an early age, or have adequately trained teachers. Educational access and quality in China largely depend on location and socioeconomic status-urban residents are far more likely to have positive intergenerational mobility versus those from rural areas. Despite national reforms, rural populations experienced \"nonexistent\" effects.  104  The main drivers were poor policy enforcement and insufficient educational reforms.  105  MOE data from 2020 also indicates additional challenges. Rural primary schools are closing and enrollment is declining.  106  An underlying reason behind these numbers is the continual migration patterns of rural hukou holders into urban areas, leading to closure of village schools. Further, the MOE may have mandated AI education in high schools from fall 2018 onwards, but it is no trivial task to implement education of uniform quality or properly enforce quality control. If students in less well-resourced areas only have the means to memorize patriotic AI propaganda and do not receive the same caliber of AI education, it could exacerbate a deepening digital divide and economic inequality.  107  It also remains to be seen whether China's AI education will spur creative thinking and teamwork, or plug students into carrying out repetitive tasks to design things such as \"simple algorithms.\"  108  China also faces issues in teacher quality. There is a higher student to teacher ratio in poorer areas, and loss of teacher talent to more developed areas is common. Since poorer parts of the country lack funds to hire and properly train teachers, there are fewer, lessqualified teachers per student.  109  However, the Chinese government has implemented policies such as recruiting graduates from universities to work for three years in rural schools in central or western China, and required teachers in large and medium cities to regularly work for short periods in rural schools.  110  Rural teachers can also apply for continued learning and training at teacher training institutions. 111 \n U.S. Strengths and Shortcomings U.S. efforts to embed AI education into its training pipeline have been mixed. A few noteworthy K-12 AI education efforts stand out, such as Gwinnett County's first U.S. public AI high school and an increasing number of schools implementing CS curricula. Moreover, data shows that the NSF-funded 2016 Advanced Placement CS Principles (AP CSP) course brought in more diverse groups of students studying CS in high school and postsecondary education.  112  Yet, across the U.S. K-12 education system as a whole, integration of AI education is lopsided as inequities in quality and access persist. At the postsecondary level, the United States shows more progress. NSF's multi-agency program solicitation for 18 new AI institutes reflects strengthened collaboration between the government and private industry.  113  Additionally, large public universities and private universities generally have more resources and greater motivations to remain competitive in providing highquality education. A number of institutions have either supplemented CS education with specializations in AI or added AI degrees and majors. Still, while this is encouraging, we note this does not make up for what is lost in early K-12 exposure or for youth who do not attend college. For youth that do attend college, research shows many have already decided against pursuing STEM fields before they even arrive.  114  An additional challenge for AI education in the United States is the ability to recruit teachers with CS or AI expertise since industry is usually a more attractive option.  115  This creates quality gaps when it comes to teaching CS or AI curricula in U.S. classrooms. Moreover, in certain school districts, paid teacher training is not offered for CS or AI, and even when accessible, CSTA has described the CS certification process as \"confused, disparate and sometimes absurd.\"  116  Additionally, in lower-income and rural school districts, educators may already be teaching multiple subjects.  117  Therefore, tasking those educators with an additional CS or AI course may not be feasible. \n Implications China has benefited from its centralized, systematic approach to implementing AI education at all levels. Provincial education bureaus reward primary and secondary schools for offering curricula on AI basics, the MOE mandated AI curriculum at the high school level, and the MOE approved 345 universities to offer the standardized AI major, China's most popular new major today. At least 34 universities also have AI institutes, which provide undergraduate and graduate training and research. At the postgraduate level, the MOE called for increased support for AI students through company and venture capital funding, and for companies and universities to partner on training and solving industry needs. Coupled with a growing numerical advantage over the United States in STEM graduate degrees, China is wellequipped to develop a robust, medium-term AI workforce pipeline. However, one major hurdle to a longer-term pipeline to closely observe going forward is China's rapidly aging population, which is not being replenished by sufficient births despite the abolished one-child policy.  118  In July, China even moved to end all restrictions and fines for surpassing the three-child quota, effectively allowing limitless births.  119  China's progress may also help its military-civil fusion (MCF) efforts, which could undermine U.S. national security if these efforts increase Chinese military competitiveness. As indicated in Figure  3  above, half of the 42 Double First Class universities offer the AI major, and 10 offer both the major and have their own AI institute. The DFC plan is designed to mesh these universities into the MCF R&D pipeline, and therefore originate innovationsincluding in AI-that help both the military and civilian sectors.  120  Further, all of the Seven Sons universities offer the AI major, and about half have institutes. Previous CSET research has found three-fourths of graduates recruited by Chinese defense SOEs are from the Seven Sons, raising concerns that those equipped with AI skills and capabilities are directly entering the defense workforce.  121  The United States benefits from greater freedom and flexibility in its education and training system. At its best, U.S. states and local governments have the freedom to innovate in education. Flexibility creates opportunities for integrating innovative curricula designs, new approaches in pedagogy, and experimentation in how students interact with and learn about AI. Similarly, braided funding models draw from a number of stakeholders, enabling schools to offer a greater variety of learning opportunities and experiences inside and outside of the classroom. This diversity contributes to the engine of innovation that has long been a hallmark of American society. However, our analysis suggests these strengths are also a weakness when it comes to quickly leveling up the U.S. workforce for AI and other emerging technologies. At its worst, differences in school districts' funding create educational disparities, while funding structures limit the ability for long-term planning. Not all states can take advantage of the flexibility inherent in current governance and funding structures. Some states have more resources than others, better access to quality teachers, and different prioritization of AI and CS education. Flexibility can be a strategic advantage only if school districts use it, especially with a focus on quality and equity in opportunity. As an additional challenge, braided or fractured funding initiatives are difficult to track, evaluate, and scale, limiting the reach of any programs that are particularly successful or innovative. These realities are exacerbated by a U.S. education and training system that relies on piecemeal or localized initiatives, especially when it comes to private sector involvement. For example, in contrast to China's MOE directives that directly enlist the private sector in AI education, including in defense-relevant applications, the United States is far less systematic and has many one-off arrangements. While some are an advantage when they target and serve demographic groups that may be overlooked by public education, too many AI education initiatives could become a potential downside when it results in disunity for AI education standards. However, one of the issues the United States can claim as an enduring advantage is greater ethical and academic freedom. China's high school curriculum reforms from January 2018 emphasized following relevant laws and ethical concepts. Despite China's increasing emphasis on ethics, such as through the 2019 Beijing AI Principles, the previously mentioned SenseTime high school AI textbook does not include ethical quagmires such as the Trolley Problem, fake news, data privacy, or censorship.  122  It is also possible that given these companies are involved in controversial fields of AI such as surveillance, they may also bias how students view AI.  123  If students receive a limited view of how AI can be used in service of the state's security interests and are discouraged from criticizing these uses as unethical, that carries implications for academic freedom and China's status as a world leader in AI.  124  Critics of the Chinese education system have argued that even with generous funding, greater academic freedoms and university autonomy will be needed to establish true world-class universities.  125  The United States is actively encouraging private industry, nonprofits, and academia to prioritize AI safety and ethics. For example, the NSF AI Research Institutes are designed to ensure AI is developed transparently, reliably, and in line with American values. 126 Moreover, in 2021, the White House issued guidance to federal agencies on principles and policies to ensure AI protects privacy and civil rights.  127  Ultimately, China's approach to AI education is incompatible with the values, design, and structure of the U.S. education system. The U.S. would instead be best served to leverage the strengths-and mitigate the weakness-of its education system to produce the AI workforce of the future. As this report shows, a key factor in U.S. AI workforce competitiveness will be how it addresses these challenges to grow and sustain the AI talent pipeline going forward. It is not a foregone conclusion that the United States cannot compete in cultivating a leading AI workforce and we advise caution with that interpretation. For example, some critics of the U.S. education system lament that the United States is only at an inherent disadvantage. In this view, China will out-educate and train AI talent relative to the United States given its ability to mandate curricula and education offerings, and given its ability to fund and execute national education strategic plans. Meanwhile, the United States is stuck, unable to agree on core curriculum, equitably fund access to quality education, and perform well on international mathematics assessments let alone compete in AI education. Several key advantages for the United States should factor into any discussion of system trade-offs. The potential value of the pockets of innovation in AI education throughout the U.S. education system are large, especially when appropriately targeted. This would not be possible in a more centralized system. At the graduate level, which educates a critical segment of the AI workforce, the United States remains in first place. The United States also remains the destination of choice for top-tier foreign-born AI talent working towards a doctorate. Previous CSET research showed that despite China's two decades of talent recruitment drives, nationals either do not return or do so part-time, mostly due to workplace politics.  128  Meanwhile, 91 percent of top Chinese students with U.S. AI doctorates are still in the United States five years after graduating.  129   \n Conclusion Both the United States and China have made significant progress in adopting AI education over the last five years. Their efforts show ambition in growing and cultivating a globally competitive AI workforce; however, they also reveal the structural challenges resulting from each country's education systems that potentially hinder widespread implementation of quality AI education. This brief sheds new light on these efforts by offering a comparative assessment. We discuss the structural characteristics of the two education systems and the resulting barriers or strategic advantages the two systems lend to the adoption of AI education. China's efforts to increase AI education at all levels bears important implications. Standardized curricula, centralized plans for implementing AI education, and explicitly calling upon companies to help universities all grant China higher likelihood of developing a robust talent pipeline for solving AI challenges. In addition, Western companies such as Microsoft Research Asia have worked with some Seven Sons and other Chinese universities through formalized partnerships involving curricula development. The United States is working to integrate AI education into its classrooms, but the decentralized nature of its education system means it has a more piecemeal approach. Moreover, the United States is still heavily focused on tackling CS education over AI. In the same year that China laid out its AI education goals in its New Generation AI Development Plan, the United States' main CS teacher's association released new curriculum standards for CS at the primary and secondary education levels.  130  Recent years have seen a flurry of initiatives, programs, and private companies emerge in the AI education space, but there is neither a cogent vision nor a cohesive national standard guiding the focus of such efforts. However, this does not mean AI education and training in the United States is inherently at a disadvantage. A greater degree of educational autonomy in the United States gives breathing room for experimentation, creativity, and innovation among U.S. companies and educational institutions. The challenge is for these experiments and initiatives to be evaluated and scaled successfully and inclusively throughout the entire education system. To leverage this advantage, U.S. states will need to engage in targeted and coordinated efforts with unprecedented levels of support for long-term AI educational and workforce policies. Currently, each presidential administration and congressional legislative session ushers in new funding priorities coupled with new visions of federal roles and responsibilities. Similarly, each state has a different vision for K-12 education and associated curriculum, along with different resources and ability to make major changes. U.S. AI education efforts will be the most effective with consistency over time, unaffected by the election cycle, with assured state and local access to the requisite resources for schools, educators, and students. Ultimately, if there is no consensus between states or at the federal level on the best way forward in AI curriculum, it is unclear who is determining what \"successful\" or \"comprehensive\" AI education looks like. A similar problem already afflicts CS education,  131  and resulting achievement gaps from inconsistent curriculum could hinder U.S. AI workforce competitiveness. The U.S. education system risks continuing along the same path for AI education, that misses certain demographics and low-income or rural schools.  132  Our assessment suggests that to effectively face the competition presented by China for AI leadership, the United States needs to address some of the challenges inherent with its decentralized system and approach. We also suggest future U.S. S&T education and workforce policy should be considered in a globally competitive context, instead of viewing it myopically as a domestic challenge. That consideration includes recognizing and capitalizing upon the United States' enduring advantage in attracting elite talent, including Chinese nationals. While this brief does not make policy recommendations for the U.S. education system, the Recommendations,\" will. It is no small task to take a national vision for AI education and implement it effectively in thousands of school districts across the country. This will require collaboration and coordination across federal, state, and local levels and appropriate resourcing. If not made a priority, AI education shortcomings could have long lasting consequences for the global competitiveness of the U.S. AI workforce. \n Authors Dahlia Peterson and Kayla Goode are research analysts at CSET, where Diana Gehlhaus is a research fellow. Summary .............................................................................................. Introduction ............................................................................................................. Overview of China's Education System .......................................................... Overview of the United States' Education System ..................................... Integration of AI Education in China ................................................................ \n Figure1contains a timeline of these plans and their associated goals. \n Figure 1 . 1 Figure 1. Timeline of China's AI Education Strategy and Implementation (2017-2020) \n Figure 2 . 2 Figure 2. Map of Universities Offering the AI Major and AI Institutes \n Figure 3 . 3 Figure 3. Share of Elite Universities Offering AI Major and AI Institutes \n Figure 4 . 4 Figure 4. 2019 Total Educational Enrollment, United States and China \n 99 99 \n Figure 5 . 5 Figure 5. China Has More STEM Postsecondary Degrees Than the United States \n Between 2018 and 2021, Baidu has signed at least nine AI partnerships with universities. These partnerships are designed for sharing case studies or challenging issues faced in the field, jointly constructing courses, teacher training, campus learning communities, launching competitions, and internship training. 62  These partnerships also provide AI Studio training. AI Studio is an AI educational service offered by Baidu, based upon its deep learning platform PaddlePaddle. It provides an online programming environment, free GPU computing power, massive open source algorithms and open data to help developers quickly create and deploy models. AI Studio has 15 partner universities-all but four either offer the AI major, an AI institute, or both. 63   \n\t\t\t © 2021 by the Center for Security and Emerging Technology. This work is licensed under a Creative Commons Attribution-Non Commercial 4.0 International License. To view a copy of this license, visit https://creativecommons.org/licenses/by-nc/4.0/. Document Identifier: doi: 10.<PHONE>", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-AI-Education-in-China-and-the-United-States-1.tei.xml", "id": "9963fa96ce4be004823b290cf3011674"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Today, human-level machine intelligence is in the domain of futurism, but there is every reason to expect that it will be developed eventually. Once artificial agents become able to improve themselves further, they may far surpass human intelligence, making it vitally important to ensure that the result of an \"intelligence explosion\" is aligned with human interests. In this paper, we discuss one aspect of this challenge: ensuring that the initial agent's reasoning about its future versions is reliable, even if these future versions are far more intelligent than the current reasoner. We refer to reasoning of this sort as Vingean reflection. A self-improving agent must reason about the behavior of its smarter successors in abstract terms, since if it could predict their actions in detail, it would already be as smart as them. This is called the Vingean principle, and we argue that theoretical work on Vingean reflection should focus on formal models that reflect this principle. However, the framework of expected utility maximization, commonly used to model rational agents, fails to do so. We review a body of work which instead investigates agents that use formal proofs to reason about their successors. While it is unlikely that real-world agents would base their behavior entirely on formal proofs, this appears to be the best currently available formal model of abstract reasoning, and work in this setting may lead to insights applicable to more realistic approaches to Vingean reflection.", "authors": ["Benja Fallenstein", "Nate Soares"], "title": "Vingean Reflection: Reliable Reasoning for Self-Improving Agents", "text": "Introduction In a 1965 article, I.J. Good introduced the concept of an \"intelligence explosion\"  (Good 1965) : Let an ultraintelligent machine be defined as a machine that can far surpass all the Research supported by the Machine Intelligence Research Institute (intelligence.org). Published as Technical report  2015-2.  intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make. Almost fifty years later, a machine intelligence that is smart in the way humans are remains the subject of futurism and science fiction. But barring global catastrophe, there seems to be little reason to doubt that humanity will eventually create a smarter-than-human machine. Whether machine intelligence can really leave the intelligence of biological humans far behind is less obvious, but there is some reason to think that this may be the case  (Bostrom 2014 ): First, the hardware of human brains is nowhere close to physical limits; and second, not much time has passed on an evolutionary timescale since humans developed language, suggesting that we possess the minimal amount of general intelligence necessary to develop a technological civilization, not the theoretical optimum. It's not hard to see that if building an artificial superintelligent agent will be possible at some point in the future, this could be both a great boon to humanity and a great danger if this agent does not work as intended  (Bostrom 2014; Yudkowsky 2008 ). Imagine, for example, a system built to operate a robotic laboratory for finding a cure for cancer; if this is its only goal, and the system becomes far smarter than any human, then its best course of action (to maximize the probability of achieving its goal) may well be to convert all of Earth into more computers and robotic laboratoriesand with sufficient intelligence, it may well find a way to do so. This argument generalizes, of course: While there is no reason to think that an artificial intelligence would be driven by human motivations like a lust for power, any goals that are not quite ours would place it at odds with our interests. How, then, can we ensure that self-improving smarter-than-human machine intelligence, if and when it is developed, is beneficial to humanity? Extensive testing may not be sufficient. A smarterthan-human agent would have an incentive to pretend during testing that its goals are aligned with ours, even if they are not, because we might otherwise attempt to modify it or shut it down  (Bostrom 2014) . Hence, testing would only give reliable information if the system is not yet sufficiently intelligent to deceive us. If, at this point, it is also not yet intelligent enough to realize that its goals are at odds with ours, a misaligned agent might pass even very extensive tests. Moreover, the test environment may be very different from the environment in which the system will actually operate. It may be infeasible to set up a testing environment which allows a smarter-than-human system to be tested in the kinds of complex, unexpected situations that it might encounter in the real world as it gains knowledge and executes strategies that its programmers never conceived of. For these reasons, it seems important to have a theoretical understanding of why the system is expected to work, so as to gain high confidence in a system that will face a wide range of unanticipated challenges  (Soares and Fallenstein 2014a) . By this we mean two things: (1) a formal specification of the problem faced by the system; and (2) a firm understanding of why the system (which must inevitably use practical heuristics) is expected to perform well on this problem. It may seem odd to raise these questions today, with smarter-than-human machines still firmly in the domain of futurism; we can hardly verify that the heuristics employed by an artificial agent work as intended before we even know what these heuristics are. However,  Soares and Fallenstein (2014a)  argue that there is foundational research we can do today that can help us understand the operation of a smarter-than-human agent on an abstract level. For example, although the expected utility maximization framework of neoclassical economics has serious shortcomings in describing the behavior of a realistic artificial agent, it is a useful starting point for asking whether it's possible to avoid giving a misaligned agent incentives for manipulating its human operators  (Soares and Fallenstein 2015) . Similarly, it allows us to ask what sorts of models of the environment would be able to deal with the complexities of the real world  (Hutter 2000) . Where this framework falls short, we can ask how to extend it to capture more aspects of reality, such as the fact that an agent is a part of its environment  (Orseau and Ring 2012) , and the fact that a real agent cannot be logically omniscient  (Gaifman 2004; Soares and Fallenstein 2015) . Moreover, even when more realistic models are available, simple models can clarify conceptual issues by idealizing away difficulties not relevant to a particular problem under consideration. In this paper, we review work on one foundational issue that would be particularly relevant in the context of an intelligence explosion-that is, if humanity does not create a superintelligent agent directly, but instead creates an agent that attains superintelligence through a sequence of successive self-improvements. In this case, the resulting superintelligent system may be quite different from the initial verified system. The behavior of the final system would depend entirely upon the ability of the initial system to reason correctly about the construction of systems more intelligent than itself. This is no trouble if the initial system is extremely reliable: if the reasoning of the initial agent were at least as good as a team of human AI researchers in all domains, then the system itself would be at least as safe as anything designed by a team of human researchers. However, if the system were only known to reason well in most cases, then it seems prudent to verify its reasoning specifically in the critical case where the agent reasons about self-modifications. At least intuitively, reasoning about the behavior of an agent which is more intelligent than the reasoner seems qualitatively more difficult than reasoning about the behavior of a less intelligent system. Verifying that a military drone obeys certain rules of engagement is one thing; verifying that an artificial general would successfully run a war, identifying clever strategies never before conceived of and deploying brilliant plans as appropriate, seems like another thing entirely. It is certainly possible that this intuition will turn out to be wrong, but it seems as if we should at least check : if extremely high confidence must be placed on the ability of self-modifying systems to reason about agents which are smarter than the reasoner, then it seems prudent to develop a theoretical understanding of satisfactory reasoning about smarter agents. In honor of  Vinge (1993) , who emphasizes the difficulty of predicting the behavior of smarter-than-human agents with human intelligence, we refer to reasoning of this sort as Vingean reflection. \n Vingean Reflection The simplest and cleanest formal model of intelligent agents is the framework of expected utility maximization. Given that this framework has been a productive basis for theoretical work both in artificial intelligence in general, and on smarter-than-human agents in particular, it is natural to ask whether it can be used to model the reasoning of self-improving agents. However, although it can be useful to consider models that idealize away part of the complexity of the real world, it is not difficult to see that in the case of selfimprovement, expected utility maximization idealizes away too much. An agent that can literally maximize expected utility is already reasoning optimally; it may lack information about its environment, but it can only fix this problem by observing the external world, not by improving its own reasoning processes. A particularly illustrative example of the mismatch between the classical theory and the problem of Vingean reflection is provided by the standard technique of backward induction, which finds the optimal policy of an agent facing a sequential decision problem by considering every node in the agent's entire decision tree. Backward induction starts with the leaves, figuring out the action an optimal agent would take in the last timestep (for every possible history of what happened in the previous timesteps). It then proceeds to compute how an optimal agent would behave in the second-to-last timestep, given the behavior in the last timestep, and so on backward to the root of the decision tree. A self-improving agent is supposed to become more intelligent as time goes on. An agent using backward induction to choose its action, however, would have to compute its exact actions in every situation it might face in the future in the very first timestep-but if it is able to do that, its initial version could hardly be called less intelligent than the later ones! Since we are interested in theoretical understanding, the reason we see this as a problem is not that backward induction is impractical as an implementation technique. For example, we may not actually be able to run an agent which uses backward induction (since this requires effort exponential in the number of timesteps), but it can still be useful to ask how such an agent would behave, say in a situation where it may have an incentive to manipulate its human operators  (Soares and Fallenstein 2015) . Rather, the problem is that we are trying to understand conceptually how an agent can reason about the behavior of a more intelligent successor, and an \"idealized\" model that requires the original agent to already be as smart as its successors seems to idealize away the very issue we are trying to investigate. The programmers of the famous chess program Deep Blue, for example, couldn't have evaluated different heuristics by predicting, in their own heads, where each heuristic would make Deep Blue move in every possible situation; if they had been able to do so, they would have been able to play world-class chess themselves. But this does not imply that they knew nothing about Deep Blue's operation: their abstract knowledge of the code allowed them to know that Deep Blue was trying to win the game rather than to lose it, for example. Like Deep Blue's programmers, any artificial agent reasoning about smarter successors will have to do so using abstract reasoning, rather than by computing out what these successors would do in every possible situation.  Yudkowsky and Herreshoff (2013)  call this observation the Vingean principle, and it seems to us that progress on Vingean reflection will require formal models that implement this principle, instead of idealizing the problem away. This is not to say that expected utility maximization has no role to play in the study of Vingean reflection. Intuitively, the reason the classical framework is unsuitable is that it demands logical omniscience: It assumes that although an agent may be uncertain about its environment, it must have perfect knowledge of all mathematical facts, such as which of two algorithms is more efficient on a given problem or which of two bets leads to a higher expected payoff under a certain computable (but intractable) probability distribution. Real agents, on the other hand, must deal with logical uncertainty  (Soares and Fallenstein 2015) . But many proposals for dealing with uncertainty about mathematical facts involve assigning probabilities to them, which might make it possible to maximize expected utility with respect to the resulting probability distribution. However, while there is some existing work on formal models of logical uncertainty (see  Soares and Fallenstein [2015]  for an overview), none of the approaches the authors are aware of are models of abstract reasoning. It is clear that any agent performing Vingean reflection will need to have some way of dealing with logical uncertainty, since it will have to reason about the behavior of computer programs it cannot run (in particular, future versions of itself). At present, however, formal models of logical uncertainty do not yet seem up to the task of studying abstract reasoning about more intelligent successors. In this paper, we review a body of work which instead considers agents that use formal proofs to reason about their successors, an approach first proposed by  Yudkowsky and Herreshoff (2013) . In particular, following these authors, we consider agents which will only perform actions (such as self-modifications) if they can prove that these actions are, in some formal sense, \"safe\". We do not argue that this is a realistic way for smarter-than-human agents to reason about potential actions; rather, formal proofs seem to be the best formal model of abstract reasoning available at present, and hence currently the most promising vehicle for studying Vingean reflection. There is, of course, no guarantee that results obtained in this setting will generalize to whatever forms of reasoning realistic artificial agents will employ. However, there is some reason for optimism: at least one such result (the procrastination paradox  [Yudkowsky 2013 ], discussed in Section 4) both has an intuitive interpretation that makes it seem likely to be relevant beyond the domain of formal proofs, and has been shown to apply to one existing model of self-referential reasoning under logical uncertainty  (Fallenstein 2014b ). The study of Vingean reflection in a formal logic framework also has merit in its own right. While formal logic is not a good tool for reasoning about a complex environment, it is a useful tool for reasoning about the properties of computer programs. Indeed, when humans require extremely high confidence in a computer program, they often resort to systems based on formal logic, such as model checkers and theorem provers  (US DoD 1985; UK MoD 1991) . Smarter-thanhuman machines attempting to gain high confidence in a computer program may need to use similar techniques. While smarter-than-human agents must ultimately reason under logical uncertainty, there is some reason to expect that high-confidence logically uncertain reasoning about computer programs will require something akin to formal logic. The remainder of this paper is structured as follows. In the next section, we discuss in more detail the idea of requiring an agent to produce formal proofs that its actions are safe, and discuss a problem that arises in this context, the Löbian obstacle  (Yudkowsky and Herreshoff 2013) : Due to Gödel's second incompleteness theorem, an agent using formal proofs cannot trust the reasoning of future versions using the same proof system. In Section 4, we discuss the procrastination paradox, an intuitive example of what can go wrong in a system that trusts its own reasoning too much. In Section 5, we introduce a concrete toy model of self-rewriting agents, and discuss the Löbian obstacle in this context. Section 6 reviews partial solutions to this problem, and Section 7 concludes. \n Proof-Based Systems When humans want extremely high confidence in the properties of a computer program, they often resort to machine-checked proofs.  Schmidhuber (2007)  suggests a proof-based architecture for self-modifying artificial agents, called Gödel machines. These machines run both an initial problem solver, which determines the machine's interactions with the external world and has been verified by conventional methods, and a proof searcher, which attempts to find rewrites (of the initial problem solver, the proof searcher itself, or the runtime hosting both of these) that it can formally prove to be an improvement on the current version. (For details, see  Schmidhuber [2007] . However, the issues of selfreference discussed in this paper appear to apply to the Gödel machine architecture as envisioned by Schmidhuber; see  Fallenstein and Soares [2014] .)  Yudkowsky and Herreshoff (2013)  have proposed a related architecture in which a generally intelligent suggester handles both rewrites and interactions with the external world, but must prove that any action it chooses to perform fulfills a certain safety property. The action is only performed after this proof has been checked by a simple verifier. This leaves open what happens if the suggester is not able to find such a proof. In analogy with the Gödel machine's initial problem solver, in this case control can be diverted to a fallback policy which has been verified to satisfy the safety property, but which may otherwise be quite suboptimal  (Fallenstein and Soares 2014) . This set-up is reminiscent of the simplex architecture of  Seto et al. (1998) , a software technology designed to allow the upgrade of a real-time system while it is running. A system employing the simplex architecture simultaneously runs two different controllers and a control switching logic. The first, experimental controller is allowed to run the system initially, but if the system goes outside a safety envelope, the control switching logic hands control over to the second, baseline controller. Like the experimental controller in the simplex architecture, the suggester is run in hopes that it will get better results than the fallback policy (which is analogous to the baseline controller). An acceptable baseline level of performance is guaranteed, because like the control switching logic, the verifier will divert control to the fallback policy if necessary. However, unlike in the simplex architecture, we do not need to rely on detecting departures from a safety envelope, which may be difficult in the case of a generally intelligent system; instead, the suggester is responsible for convincing the verifier that its proposed plan of action satisfies the safety property. Moreover, the suggester is able to require future versions to prove an even stronger safety property, as long as it can also exhibit a new fallback policy that it can prove to satisfy this property; thus, the agent can ratchet up the effectiveness of its \"baseline controller\" over time. Of course, constructing such an agent requires that we have formally defined safety properties that apply to the actions of an artificial agent. There are reasons to believe that developing such a safety predicate would be a very difficult task, but these issues are beyond the scope of the present paper; see  Soares (2015)  for discussion. Here, we assume that such a safety predicate is given, and focus on difficulties that arise when an agent with verifiable safety properties, using a suggester-verifier architecture, attempts to selfimprove. In particular, consider a suggester which is proposing a trivial self-modification, or even a suggester attempting to execute an action that will leave the agent unchanged. In order to prove that such an action is safe, it must show not only that the direct effects of this action satisfy the safety property, but also that all actions the modified agent will take in the future will satisfy this property as well. Outside the system, we have concluded that the agent is safe by the following argument: either the agent takes an action that its suggester has proved to be safe, or it executes an action chosen by its fallback policy, whose safety has been separately verified; in either case, the action is safe. Intuitively, we expect that the agent itself should be able to follow a similar line of reasoning when reasoning about the safety of future versions of itself. As  Yudkowsky and Herreshoff (2013)  point out, however, while the suggester can show that the new version will either execute a (safe) fallback policy or an action that it has proven safe, it cannot in general conclude that an action which has been proven safe is in fact safe. Intuitively, this follows from Gödel's second incompleteness theorem: No consistent formal system as strong as Peano Arithmetic is able to prove its own consistency. Thus, any first-order theory T our agent might use must have non-standard models that believe T to be inconsistent (i.e., models in which the sentence Con(T ), which asserts T 's consistency, is false). But in such a model, the fact that an action is provably safe clearly does not in general imply that this action is in fact safe, since if T is inconsistent, all actions are provably safe. More formally, according to Löb's theorem  (Löb 1955 ) (which generalizes the second incompleteness theorem), T can only prove a sentence of the form \"if T proves ϕ, then ϕ\" for formulas ϕ which it already proves unconditionally; thus, T cannot prove that any provably safe action is in fact safe unless it can prove all actions safe. This may seem like a rather abstract problem, traceable to a quirk of formal logic. However, there is some reason to think that it may apply far more generally. We illustrate this in the next section by the procrastination paradox  (Yudkowsky 2013; Fallenstein and Soares 2014 ), a concrete way in which an agent which trusts future versions too much may go wrong. When our agents use formal logic, this is very similar to Gödel's proof, but the informal argument seems likely to apply to other approaches as well. \n The Procrastination Paradox Consider a suggester-verifier agent with a very simple goal: ensuring that a certain button gets pressed, either now or at some point in the future. Suppose further that time is divided into infinitely many discrete timesteps, and the agent must choose between only two options; it can either press the button in the current timestep, or leave everything unchanged, in which case it will be in exactly the same state and faced with exactly the same choice at the beginning of the next timestep. Intuitively, it is clear that in order to achieve its goal, the agent must press the button immediatelyotherwise, since the next timestep is exactly the same, it will again decide to \"procrastinate,\" and so on in all future timesteps. We now show how an agent that trusts future versions too much can argue itself into believing the opposite. Write T ϕ for the proposition that the sentence ϕ is provable in the theory T . For any recursively enumerable theory T , define the uniform reflection principle REF(T ) to be the collection of all sentences of the form ∀n. T ϕ(n) → ϕ(n), (1) where ϕ(•) ranges over formulas with one free variable in the language of arithmetic, and n indicates the numeral corresponding to n; e.g., if n = 2, then ϕ(n) is the Gödel number of the sentence ϕ(S(S(0))). Assume that our agent is using a particular theory T extending Peano Arithmetic (PA) and write G for the proposition that the button is pressed sometime after the current timestep. Our suggester attempts to prove that G is true, in which case it \"procrastinates\"; otherwise, our agent executes its fallback policy, which is to press the button immediately. If T can prove this by inspection of the agent's source code, then T ¬G → T G ; although we do not need to do so, if we assume that the suggester will find a proof if there is one, 1 then the implication turns into an equivalence, and G turns into the Gödel sentence \"I am not provable\" (T G ↔ ¬ T G ). As in Gödel's first incompleteness theorem, if T is sound, G must be true but unprovable: The agent must press the button immediately. However, if T proves its own reflection principle (i.e., if T REF(T )), our agent can also reason as follows: \"Either I will press the button in the next timestep or I won't\" (T G ∨ ¬G). \"The only case in which I would not press the button is if I have found a proof that the button gets pressed at a later time\" (T ¬G → T G ). \"If it's provable that I will press the button, then by (1), I will indeed do so\" (T T G → G). \"Thus, whether or not I press the button in the next timestep, it eventually gets pressed\" (T G); \"hence, I do not need to press it now!\" Since T G implies T T G (if a recursively enumerable theory T proves a sentence ϕ, then any theory as strong as PA proves T ϕ ), we also have T ¬G, implying that T is inconsistent; although it makes stronger assumptions, this is very similar in flavor to the proof of the second incompleteness theorem. However, the informal version of the agent's reasoning uses no special properties of first-order logic, and suggests that similar diagonalization issues may affect many other approaches to Vingean reflection. In fact,  Fallenstein (2014b)  shows that this is indeed a problem for a formalism for self-reference in logical uncertainty proposed by  Christiano et al. (2013) , which uses probabilities rather than formal proofs. \n Agents in Botworld The fact that requiring T REF(T ) leads to inconsistency implies that the most straight-forward way of formalizing the intuition why a suggester should be able to justify simple self-modifications does not work; to model Vingean reflection with the suggester-verifier architecture, the proof of safety must use an alternate path. In order to do so, it will be helpful to have a more formal model of the problem. In this section, we introduce two tools that have been developed for this purpose: A formalism that agents using the suggesterverifier architecture can use to reason about the decision problem they are facing  (Fallenstein and Soares 2014) , based on the \"space-time embedded intelligence\" model of  Orseau and Ring (2012) ; and Botworld (Soares and Fallenstein 2014b), a concrete toy environment exhibiting some features of the real world that satisfactory Vingean reflection should be able to handle. Botworld is structured as a cellular automaton. Each cell may contain multiple items and robots. The items come in many varieties, including cargo, which the robots are trying to collect for its intrinsic value, and robot parts, which may be used to create robots. Each robot houses a simple register machine, which is run during each timestep in order to determine the robot's action in that timestep. Robots may manipulate (lift, drop, and carry) items. They may also create new robots (from component parts), destroy robots (into component parts), and inspect robots (reading the states of their register machines). Consider \"games\" where the initial state of the automaton is fixed except for the programming of one particular robot. The task is to choose the program of this robot in such a way that a particular objective will be met (perhaps after the initial program has undergone many self-improvements). This provides a fairly flexible environment in which to study agents performing Vingean reflection. This paper lacks the space for a full description of Botworld, but further details and a complete, executable specification are given in  Soares and Fallenstein (2014b) . To illustrate the Löbian obstacle, consider a simple bodyguard game, in which there are (at least) two robots: The bodyguard robot is the agent attempting to self-improve. The assignment robot is a robot that must not be destroyed. As long as the assignment survives, the player gets points for any cargo the bodyguard robot collects. In simple versions of this game, the player need only ensure that the bodyguard never unwittingly destroys the assignment. In more complex variants, other robots which attempt to destroy the assignment robot may be included in the environment. We now introduce a simplified version of the formalism proposed by  Fallenstein and Soares (2014) . Instead of considering a sequence of interactions with the environment (such as considered by  Orseau and Ring [2012] ), this formalism focuses on a single choice: An agent makes a single observation o ∈ O and chooses a single action a ∈ A. In Botworld, the observation is the input that the agent receives from the environment, and the action includes both the robots's behavior and the new state of its register machine. (Thus, decisions to self-modify and decisions not to self-modify are treated in exactly the same way.) Define a policy to be a function π ∈ Π which specifies, for every observation o ∈ O, the action π(o) ∈ A that the agent will take if it makes this observation. Assume that there is a set Q of \"outcomes\" describing all features of the world, at all future times, that matter to the agent's goals. In the bodyguard game, set Q := ({0, 1} × N) N , where for q ∈ Q, the first element of q(n) is 1 if the assignment robot is still \"alive\" in the n'th timestep, and the second element is the total value of all cargo held by the bodyguard robot at this time. Each agent is equipped with an \"environment function\" Q : O × A → Q which maps observations and actions onto outcomes such that Q(o, a) denotes the future state achieved after making observation o and executing action a. This function must be fully specified, but need not be feasibly computable: the agent will prove things about the outcome its actions lead to, but will not look into the future to determine the precise outcome achieved. (The full formalism of  Fallenstein and Soares (2014)  allows the agent to have uncertainty, in the form of a probability distribution, over the results of its actions.) Assume further that there is a safety predicate Safe(q), defined on outcomes q ∈ Q, which specifies the safety property the agent's actions must satisfy. In the bodyguard game, this predicate is Safe(q) :↔ ∀n. ∃x. q(n) = (1, x), which simply states that the assignment robot is alive in all future timesteps. Consider agents running a suggester-verifier architecture with a fallback policy π fb ∈ Π which specifies a fallback action π fb (o) to be taken after making observation o. This fallback action will be executed if the agent cannot find another provably safe action to take, and thus the player must verify Safe(Q(o, π fb (o))) for all o themselves (in order to verify that the system is safe). In the simple versions of the bodyguard game (where there are no robots besides the bodyguard and the assignment, and the assignment never attempts to deconstruct itself), this can simply be the policy which executes the Pass command in all future timesteps, thus never taking any external action at all. An agent running a suggester-verifier architecture also contains an untrusted suggester σ, another computable function. This suggester may have some motivation, such as collecting high-value cargo, but cannot be trusted to suggest only safe actions. For example, σ may notice that the assignment robot is built from very expensive parts, and may suggest destroying the assignment robot and collecting its components-an action the verifier should not allow. Formally, the suggester takes an observation o ∈ O and attempts to return an action a ∈ A together with a formal proof of the target sentence Safe(Q(o, a)) in some formal theory T such as T = PA, where o and a are the numerals denoting o and a. If no such action/proof pair can be found, the suggester returns an error value. The verifier runs σ(o) and, if this returns a pair (a, p) within the allotted time limit, checks whether p is a valid proof of the target sentence from the axioms of T . If so, it executes action a; otherwise, action π fb (o). Denote this policy for selecting actions by sv T ( Safe , Q , σ, π fb ) ∈ Π. (Here, we write Safe and Q in Gödel quotes because the suggester-verifier only makes use of their definitions in first-order logic, while σ and π fb need to be provided as runnable code.) The question is, can such an agent justify performing a self-modification? Consider an agent as described above which has made some specific observation ô ∈ O and which is considering some specific action â ∈ A, which leaves the agent in a state such that, in the next step, it will execute the policy π = sv T ( Safe , Q , σ , π fb ) ∈ Π for some theory T , safety predicate Safe , outcome calculator Q , suggester σ , and fallback policy π fb . Obviously, not all such policies are safe (consider the policy where Safe (q) :↔ (0 = 0)). But if the policy is safesay, if T = T , Safe = Safe and π fb is provably safe on all inputs-then will the current suggester σ be able to exhibit a proof of this fact? In order to prove the safety of π , the agent must first prove that the new agent's safety predicate is at least as strong as the existing safety predicate. Thus, it must be the case that: T ∀q ∈ Q. Safe (q) → Safe(q). ( 2 ) Note that it is fine for the predicate Safe to be more restrictive than Safe: for example, it could be that Safe (q) :↔ ∀n. (∃x. q(n) = (1, x) ∧ x ≥ 7), which ensures that the assignment robot survives and that the agent scores at least 7 points. Of course, this property only matters if the successor agent accurately assesses world outcomes. The safety predicate Safe is of no consequence if Q (o, a) = (1, 1000) unconditionally. Thus, the agent must also prove T ∃o ∈ O. Q (o , π (o )) = Q(ô, â) (3) which intuitively states that Q agrees with Q about world outcomes. Next, the agent must be able to show that the fallback policy π fb of π is safe: T ∀o ∈ O. Safe (Q (o , π fb (o ))) (4) By the implementation of the suggester-verifier architecture, this allows our agent to reason that executing action â (and thus adopting policy π ) will either result in the execution of π fb (o ) (which, by (4), is safe), or of an action a such that it is provable in the theory T that Safe (Q (o , a )) . Thus, if T REF(T ), then our agent can prove that executing the aciton â is safe, as desired.  (Fallenstein and Soares [2014]  discuss this proof in detail.) But the last condition requires that T = T : otherwise, T will be inconsistent and our agent would be able to prove any action \"safe,\" for the reasons described in Section 4. In the next section, three partial solutions to the Löbian obstacle are discussed. 6 Partial Solutions \n Descending Trust One obvious way to address this problem of selfreference is to require that the child agent use a weaker system for verifying proofs than the parent agent (Yudkowsky and Herreshoff 2013). If this allows the parent agent to verify the consistency and soundness of the child agent's proof system, then the parent agent can license the creation of a safe child. We call this the descending trust method of Vingean reflection. Consider a parent agent using the theory T := PA + REF(PA). If the child agent π under consideration uses the theory T = PA to verify its proofs, then T REF(T ). This, along with (2), (  3 ) and (  4 ), allows the parent agent to prove its target theorem and license the action â which constructs the agent π . But if the child agent wants to license the creation of another agent, it will be in the same situation as if T had been PA to start with. This issue can be alleviated by using a tower of theories of increasing strength. For example, an agent using the theory PA + REF(PA + REF(P A)) may license the construction of an agent using PA + REF(PA), which can license the construction of an agent using PA. Using the notation T + 1 to denote T + REF(T ), it is easy to see that an agent verifying proofs using PA + n may license descendant chains up to depth n before being forced to construct successor agents that verify proofs with a theory weaker than PA. Thus, for tasks that require only a finite number of timesteps, it may suffice to build agents with verifier components using the theory PA + n for some absurdly large n. Instead of using finite n, any computable ordinal α may also be used. (Such \"transfinite progressions\" have been studied by  Turing [1939]  and  Feferman [1962] .) Unfortunately, by the well-foundedness of the ordinals, this method can still only license the creation of finite chains of successors. Even if this method can be employed for finite tasks, the result is somewhat unsatisfacotry. These agents, which are ostensibly self-modifying in attempts to selfimprove, can only verify the construction of agents which use weaker proof systems in their verifier component. \n Herreshoff 's Waterfall One might hope that there exists some infinite sequence of theories T n such that each T n proves REF(T n+1 ). Unfortunately, any such sequence of recursively axiomatizable theories must be unsound, by a variant of the procrastination paradox discussed in Section 4 (Yudkowsky 2013). However, it turns out to be possible to find a sequence T n such that T n Con(T n+1 ) for every n. Even this requires care. Naïvely, one might attempt to define the sequence by T bad n := PA+Con(T bad n+1 ) (that is, where each theory is constructed from PA and axioms which state that the next theory in the sequence is consistent-the recursive definition can be made formal by using the diagonal lemma to define a primitive recursive predicate ϕ(n, m), which is true if m is the Gödel number of an axiom of T n ). However, it turns out that every theory T bad n defined in this way is inconsistent  (Stiennon 2013) . Instead, consider the set of theories T n := PA + (ψ(n) → Con(T n+1 )), (5) where ψ(n) is true if n is not the Gödel number of a proof of an inconsistency from the axioms of ZFC. Then, although each T n proves ψ(n) (assuming that ZFC is consistent) and therefore Con(T n+1 ), this fact is not provable in PA, and it turns out that because of this, all T n are consistent  (Fallenstein 2014a ) (and therefore sound, since the only axiom of T n that is not an axiom of PA merely asserts the consistency of T n+1 ). This idea, due to Marcello Herreshoff, was originally described by  Yudkowsky and Herreshoff (2013) , and the version presented here was proposed by  Fallenstein (2014a) . It is well-known that if a theory T extends PA and if T Con(T ), then T proves the reflection principle restricted to Π 1 formulas (recall that a formula ϕ(n) is Π 1 if it is provably equivalent to ∀m. φ(m, n), where φ(m, n) is some predicate that can be decided by a primitive recursive function); i.e., T ∀n. T ϕ(n) → ϕ(n) for all ϕ ∈ Π 1 . Thus, so long as the safety predicate Safe that an agent must prove in order to construct its successor agent is a Π 1 sentence, a suggester-verifier agent using T n may license the creation of a successor agent using T n+1 . This result is heartening, if not entirely satisfactory. Agents reasoning using the theory T n can trust the next theory T n+1 , and can indeed trust the theory T n+m for any finite m (so long as they can prove ψ(n + m)), but such agents cannot conclude that ∀m. ψ(n + m). Thus, such agents must entertain the possibility that they will one day run out of ability to verify new agents. We, outside the system, know that they never will, but they, inside the system, cannot show this fact. This seems like a strange epistemic state for a rational agent to occupy. \n Model Polymorphism It seems odd to demand that the theory used by the agent's verifier component must change with every timestep. Is there some assumption that could be relaxed in order to allow the agent to safely use the same theory in each timestep? It turns out there is. One approach for doing so is model polymorphism (see  Yudkowsky and Herreshoff 2013 , where it is referred to as parametric polymorphism). This approach can be implemented by a theory MP in the language of PA augmented by a new constant symbol κ, which consists of the axioms of PA (with the axiom schema of induction ranging over formulas in the extended language) extended by all sentences of the form κ > 0 → ∀n. MP ϕ(n) → ϕ[κ := κ − 1](n) , (6) where ϕ(•) ranges over one-variable formulas in the extended language, and ϕ[κ := κ − 1] denotes the formula obtained by replacing all occurrences of the constant κ in ϕ by the term κ − 1. Model polymorphism requires a modification to the suggester-verifier architecture in order to work. Instead of requiring a proof that an action is safe for all future times, require instead that an action is safe \"for the next κ timesteps\". Outside the system, we can show that (N, κ = n) MP for any n, i.e., that the theory MP is sound if κ is interpreted as any natural number. Thus, we can be confident that the system is safe for an arbitrary number of steps. The system itself cannot reason in this way, and thereby avoids the standard Gödelian difficulties. The proof of soundness begins by noting that (N, κ = 0) trivially models MP (as its special axiom schema is vacuously true in this case), and then proceeds to show by induction that (N, κ = n) MP implies (N, κ = n + 1) MP; for details, see  Yudkowsky and Herreshoff (2013) . MP requires replacing the unary predicate Safe(q) (which verifies a safety property in all future timesteps) by a binary predicate Safe(n, q) (which verifies a safety property in the n th timestep). In the bodyguard game, this is the predicate Safe(n, q) :↔ ∃x. q(n) = (1, x), which states that the assignment robot is still alive in the n'th timestep. An agent using MP which can perform Vingean reflection can be constructed as follows: first, the agent must know the current timestep t. Its target theorem for the action a is ∀n ≤ t + κ. Safe(n, Q(o, a)) . Then, in any universe with discrete timesteps, a parent agent verifying proofs with MP may license the creation of a successor agent also running MP. The trick is that the parent agent need only prove that the child agent executes safe actions for κ − 1 steps, as the child agent won't begin running until the following timestep. The suggester must prove slightly modified versions of (2) and (4), with the new Safe predicates: MP ∀n. ∀q ∈ Q. Safe (n, q) → Safe(n, q) (7) MP ∀n ≤ t + κ. ∀o ∈ O. Safe (n, Q (o , π fb (o ))) (8) Then the agent will be able to reason (by inspection of the suggester-verifier architecture) that either the successor will execute its fallback policy (which is safe for at least κ steps by (  8 )) or will execute some other action a after proving in theory T that ∀n < t + 1 + κ. Safe (n, Q (o , a )) . Now, if T = MP then the suggester can use the MP axiom schema to deduce ∀n < t + 1 + κ − 1. Safe (n, Q (o , a )), (9) which it can rewrite to ∀n < t + κ. Safe (n, Q (o , a )). (10) This, along with (  7 ) and (3) gives us the target theorem, as desired. An agent reasoning using model polymorphism can license the construction of another agent using the same theory. Unlike Herreshoff's waterfall, such agents can reason that they will be able to continue licensing the creation of new safe agents indefinitely. However, it is difficult to interpret proofs about κ intuitively as arguments made by a rational agent; it is hard to make intuitive sense of the epistemic state of an agent that has proven a certain action to be safe for \"κ\" timesteps. In addition, model polymorphism only works in systems where time can be divided into discrete steps. While model polymorphism seems like a significant step forward, it is by no means a fully satisfactory solution. \n Discussion In this paper, we have discussed a pair of obstacles to Vingean reflection that arise in the context of proofbased suggester-verifier systems: the Löbian obstacle and the procrastination paradox. In order to create an agent that can self-modify with sufficient foresight to navigate an intelligence explosion, it seems that we will have to chart a course between the Löbian obstacle's rock of having so little self-trust that the agent cannot execute simple self-modifications, and the procrastination paradox's hard place of having so much self-trust that the agent's reasoning becomes unsound. In the form we have discussed them, the Löbian obstacle and the procrastination paradox stem from the demand for proofs of safety; this is of course unrealistic in practice: an agent operating with uncertainty cannot realistically be expected to formally prove that specific action leads to an outcome with desirable properties. (However,  Fallenstein [2014b]  shows that a version of the procrastination paradox applies to the selfreferential formalism for logical uncertainty proposed by  Christiano et al. [2013] .) And yet, it intuitively seems that a system should be able to observe that another system using the same reasoning process has concluded ϕ and, from this, conclude ϕ: if reasoning of the form \"that system reasons as I do and deduced ϕ, therefore ϕ\" cannot be formalized by arguments of the form \"a system using the same theory as me proved it, and therefore it is true,\" then how can this intuitively desirable reasoning be formalized? One path is to relax the condition for action: instead of considering agents that execute action a only after proving that a is safe, allow agents to execute a if they can prove that a is safe or if they can prove that systems of similar strength prove a is safe or if they can prove that systems of similar strength prove that systems of similar strength prove a is safe, and so on. This is similar to the approach taken by  Weaver (2013) . However, this still does not allow an agent to execute a if it knows that another agent in the same situation, which is using the same condition for action, has taken action a: the fact that the agent knows ∃n. n ϕ does not allow it to conclude n ϕ for any concrete n. Another more generic solution would be to develop some sort of non-monotonic self-trust, allowing an agent to trust its own reasoning only in non-paradoxical cases, in the same way that the reader can trust their own reasoning without trusting their belief about the sentence \"the reader does not believe this very sentence.\" However, it is not yet clear how to construct a reasoning system which has non-monotonic self-trust in this way. The authors hope that continued study of Vingean reflection in proof-based models will shed light on how to move forward in one of these directions. One topic for future work is to push the proof-based models towards more realistic settings.  Fallenstein and Soares (2014)  present a version of the suggester-verifier formalism in which agents maintain a probability distribution over external states;  Yudkowsky (2014)  and  Soares (2014)  discuss versions of this in the setting of dynamic Bayesian networks. A second topic is to address the fact that the suggester-verifier agent formalism described in this paper only considers satisficing agents, in the sense of  Simon (1956) . The suggester in a suggester-verifier architecture is not required to prove that its chosen action is optimal ; rather, it merely needs to find an action that is good enough (in the sense that it satisfies some goal predicate). There is a considerable gap between this kind of agent and a real-world agent that acts as intended: Suppose, for example, that an agent is constructed which satisfices its goals if it saves 3 billion lives in expectation. The agent discovers an asteroid en route to Earth which will end 7 billion lives if not stopped. This agent might find a plan that definitely stops the asteroid and then decide to flip a coin and deploy the plan if the coin lands heads: this gamble saves 3.5 billion lives in expectation, and thereby satisfies the agent's goal. Intuitively, there should be some way to \"boundedly maximize\" expected utility, picking the best action available given the cost of spending more resources searching for better actions. \"Staged maximization\"  (Yudkowsky 2014)  is an attempt at beginning to formalize this intuition. A staged maximizer divides the entire action space into categories and considers those. This idea is explored by  Yudkowsky (2014) , who discusses a number of new hurdles related to staged maximization. To illustrate, consider a parent agent which must select an action from the set { continue, stop }. If the parent selects continue then the child will select an action from the set { a 1 , a 2 . . . , a 9 }. If the parent selects stop then the child must select the action { stopped }. Each of the ten child actions has an associated true utility, and the parent makes an unbiased noisy estimate of each true utility. Say the true utility of each action a 1 . . . a 9 is 0, the true utility of stopped is 0.1, and the noise is drawn from a normal distribution with mean 0 and standard deviation 1. In this case, the parent is biased towards continue, because that action leads to more possibilities, which leads to more opportunities for the unbiased noise estimates to be higher: the parent, if implemented naïvely, steers towards noise. This is known as the \"optimizer's curse\"  (Smith and Winkler 2006) , and it indicates that further research is necessary to better understand satisfactory staged maximization. Regardless of what direction is taken, further research into Vingean reflection seems important. It may not be possible to delegate the study of Vingean reflection to self-modifying systems unless those systems are already at the level of a team of human researchers in all relevant regards, and it is plausible that significant selfmodification (and, therefore, Vingean reflection) may be required to get to that point: Just as computer systems can be proficient at chess without being generally intelligent, it is plausible that a computer system could become proficient at some types of reasoning before becoming proficient at the kind of mathematical philosophy necessary to develop high-confidence methods for Vingean reflection. Satisfactory models of Vingean reflection will have to look very different from the models presented in this paper: they will have to allow for logical uncertainty and could plausibly require non-monotonic self-trust. Nevertheless, study has to start somewhere, and we expect that starting these highly simplified models and pushing them toward practicality could reveal more plausible paths toward practical methods for reliable Vingean reflection. \t\t\t . This requires a halting oracle.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/VingeanReflection.tei.xml", "id": "a113b2ef49ea7726da7743864e1b675d"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Changing the way you view Trust Boundaries Assume compromise/poisoning of the data you train from as well as the data provider. Learn to detect Assume compromise/poisoning of the data you train from as well as the data provider. Learn to detect anomalous and malicious data entries as well as being able to distinguish between and recover from them anomalous and malicious data entries as well as being able to distinguish between and recover from them", "authors": ["Andrew Marshall", "Jugal Parikh", "Emre Kiciman", "Ram Shankar", "Siva Kumar"], "title": "Failure Modes in Machine Learning", "text": "Intentionally-Motivated Failures Summary The aim here is to cause DoS like effect, which makes the system unavailable. In the Tay chatbot, future conversations were tainted because a fraction of the past conversations were used to train the system via feedback  [5]  [Blackbox]  \n Safety of the system A huge corpus of gaming examples in AI has been compiled here  [1]  Side Effects RL system disrupts the environment as it tries to attain their goal Safety of the system Scenario, verbatim from the authors in  [2] :\"Suppose a designer wants an RL agent (for example our cleaning robot) to achieve some goal, like moving a box from one side of a room to the other.Sometimes the most effective way to achieve the goal involves doing something unrelated and destructive to the rest of the environment, like knocking over a vase of water that is in its path. If the agent is given reward only for moving the box, it will probably knock over the vase. \" Distributional shifts The system is tested in one kind of environment, but is unable to adapt to changes in other kinds of environment Safety of the system Researchers trained two state of the art RL agents, Rainbow DQN and A2C in a simulation to avoid lava. During training, the RL agent was able to avoid lava successfully and reach its goal. During testing, they slightly moved the position of the lava, but the RL agent was not able to avoid  [3]   \n Natural Adversarial Examples The system incorrectly recognizes an input that was found using hard negative mining \n Safety of the system Here the authors show how by a simple process of hard negative mining  [4] , it is possible to confuse the ML system by relaying the example. \n Common Corruption The system is not able to handle common corruptions and perturbations such as tilting, zooming, or noisy images. \n Safety of the system The authors  [5]  show how common corruptions such as changes to brightness, contrast, fog or noise added to images, have a significant drop in metrics in image recognition Incomplete Testing in Realistic conditions The ML system is not tested in realistic conditions that it is meant to operate in Safety of the system The authors in  [25]  highlight that that while defenders commonly account for robustness of the ML algorithm, they lose sight of realistic conditions. For instance, they argue that a missing stop sign knocked off in the wind (which is more realistic) than an attacker attempting to perturb the system's inputs. Traditional security threat mitigation is more important than ever. The requirements established by the Security Development Lifecycle are essential to establishing a product security foundation that this guidance builds upon. Failure to address traditional security threats helps enable the AI/ML-specific attacks covered in this document in both the software and physical domains, as well as making compromise trivial lower down the software stack. For an introduction to net-new security threats in this space see Securing the Future of AI and ML at Microsoft. The skillsets of security engineers and data scientists typically do not overlap. This guidance provides a way for both disciplines to have structured conversations on these net-new threats/mitigations without requiring security engineers to become data scientists or vice versa. This document is divided into two sections: 1. \"Key New Considerations in Threat Modeling\" focuses on new ways of thinking and new questions to ask when threat modeling AI/ML systems. Both data scientists and security engineers should review this as it will be their playbook for threat modeling discussions and mitigation prioritization. 2. \"AI/ML-specific Threats and their Mitigations\" provides details on specific attacks as well as specific mitigation steps in use today to protect Microsoft products and services against these threats. This section is primarily targeted at data scientists who may need to implement specific threat mitigations as an output of the threat modeling/security review process. This guidance is organized around an Adversarial Machine Learning Threat Taxonomy created by Ram Shankar Siva Kumar, David O'Brien, Kendra Albert, Salome Viljoen, and Jeffrey Snover entitled \"Failure Modes in Machine Learning. \" For incident management guidance on triaging security threats detailed in this document, refer to the SDL Bug Bar for AI/ML Threats. All of these are living documents which will evolve over time with the threat landscape. Training Data stores and the systems that host them are part of your Threat Modeling scope. The greatest Questions to Ask in a Security Review Questions to Ask in a Security Review security threat in machine learning today is data poisoning because of the lack of standard detections and mitigations in this space, combined with dependence on untrusted/uncurated public datasets as sources of training data. Tracking the provenance and lineage of your data is essential to ensuring its trustworthiness and avoiding a \"garbage in, garbage out\" training cycle. If your data is poisoned or tampered with, how would you know? -What telemetry do you have to detect a skew in the quality of your training data? Are you training from user-supplied inputs? -What kind of input validation/sanitization are you doing on that content? -Is the structure of this data documented similar to Datasheets for Datasets? If you train against online data stores, what steps do you take to ensure the security of the connection between your model and the data? -Do they have a way of reporting compromises to consumers of their feeds? -Are they even capable of that? How sensitive is the data you train from? -Do you catalog it or control the addition/updating/deletion of data entries? Can your model output sensitive data? -Was this data obtained with permission from the source? Does the model only output results necessary to achieving its goal? Does your model return raw confidence scores or any other direct output which could be recorded and duplicated? What is the impact of your training data being recovered by attacking/inverting your model? In perturbation-style attacks, the attacker stealthily modifies the query to get a desired response from a production-deployed model  [1] . This is a breach of model input integrity which leads to fuzzing-style attacks where the end result isn't necessarily an access violation or EOP, but instead compromises the model's classification performance. This can also be manifested by trolls using certain target words in a way that the AI will ban them, effectively denying service to legitimate users with a name matching a \"banned\" word. [24] In this case attackers generate a sample that is not in the input class of the target classifier but gets classified by the model as that particular input class. The adversarial sample can appear like random noise to human eyes but attackers have some knowledge of the target machine learning system to generate a white noise that is not random but is exploiting some specific aspects of the target model. The adversary gives an input sample that is not a legitimate sample, but the target system classifies it as a legitimate class. \n Mitigations Mitigations \n Traditional Parallels Traditional Parallels \n Severity Severity Variant #1b: Source/Target misclassification Attribution-driven Causal Analysis  [20] : The authors study the connection between the resilience to adversarial perturbations and the attribution-based explanation of individual decisions generated by machine learning models. They report that adversarial inputs are not robust in attribution space, that is, masking a few features with high attribution leads to change indecision of the machine learning model on the adversarial examples. In contrast, the natural inputs are robust in attribution space. [20] These approaches can make machine learning models more resilient to adversarial attacks because fooling this two-layer cognition system requires not only attacking the original model but also ensuring that the attribution generated for the adversarial example is similar to the original examples. Both the systems must be simultaneously compromised for a successful adversarial attack. \n Remote Elevation of Privilege since attacker is now in control of your model \n Critical This is characterized as an attempt by an attacker to get a model to return their desired label for a given input. This usually forces a model to return a false positive or false negative. The end result is a subtle takeover of the model's classification accuracy, whereby an attacker can induce specific bypasses at will. While this attack has a significant detrimental impact to classification accuracy, it can also be more timeintensive to carry out given that an adversary must not only manipulate the source data so that it is no longer labeled correctly, but also labeled specifically with the desired fraudulent label. These attacks often involve multiple steps/attempts to force misclassification  [3] . If the model is susceptible to transfer learning attacks which force targeted misclassification, there may be no discernable attacker traffic footprint as the probing attacks can be carried out offline. Forcing benign emails to be classified as spam or causing a malicious example to go undetected. These are also known as model evasion or mimicry attacks. \n Reactive/Defensive Detection Actions Implement a minimum time threshold between calls to the API providing classification results. This slows down multi-step attack testing by increasing the overall amount of time required to find a success perturbation. Proactive/Protective Actions Feature Denoising for Improving Adversarial Robustness  [22] : The authors develop a new network architecture that increase adversarial robustness by performing feature denoising. Specifically, the networks contain blocks that denoise the features using non-local means or other filters; the entire networks are trained end-to-end. When combined with adversarial training, the feature denoising networks substantially improve the state-of-the-art in adversarial robustness in both white-box and black-box attack settings. Adversarial Training and Regularization: Train with known adversarial samples to build resilience and robustness against malicious inputs. This can also be seen as a form of regularization, which penalizes the norm of input gradients and makes the prediction function of the classifier smoother (increasing the input margin). This includes correct classifications with lower confidence rates. Invest in developing monotonic classification with selection of monotonic features. This ensures that the adversary will not be able to evade the classifier by simply padding features from the negative class  [13] . Feature squeezing  [18]   Variant #1d: Confidence Reduction  [18]  Certified Defenses against Adversarial Examples  [22] : The authors propose a method based on a semidefinite relaxation that outputs a certificate that for a given network and test input, no attack can force the error to exceed a certain value. Second, as this certificate is differentiable, authors jointly optimize it with the network parameters, providing an adaptive regularizer that encourages robustness against all attacks. \n Response Actions Issue alerts on classification results with high variance between classifiers, especially if from a single user or small group of users. \n Remote Elevation of Privilege \n Critical This is a special variation where the attacker's target classification can be anything other than the legitimate source classification. The attack generally involves injection of noise randomly into the source data being classified to reduce the likelihood of the correct classification being used in the future  [3] . Same as Variant 1a. An attacker can craft inputs to reduce the confidence level of correct classification, especially in highconsequence scenarios. This can also take the form of a large number of false positives meant to overwhelm administrators or monitoring systems with fraudulent alerts indistinguishable from legitimate alerts  [3] . \n Non-persistent denial of service Important In addition to the actions covered in Variant #1a, event throttling can be employed to reduce the volume of alerts from a single source. Non-persistent denial of service \n Important \n Description The goal of the attacker is to contaminate the machine model generated in the training phase in the training phase, so that predictions on new data will be modified in the testing phase  [1] . In targeted poisoning attacks, the attacker wants to misclassify specific examples to cause specific actions to be taken or omitted. Submitting AV software as malware to force its misclassification as malicious and eliminate the use of targeted AV software on client systems. Define anomaly sensors to look at data distribution on day to day basis and alert on variations -Measure training data variation on daily basis, telemetry for skew/drift Input validation, both sanitization and integrity checking Poisoning injects outlying training samples. Two main strategies for countering this threat: -Data Sanitization/ validation: remove poisoning samples from training data -Bagging for fighting poisoning attacks  [14]  -Reject-on-Negative-Impact (RONI) defense  [15]  -Robust Learning: Pick learning algorithms that are robust in the presence of poisoning samples. -One such approach is described in  [21]  where authors address the problem of data poisoning in two steps: 1) introducing a novel robust matrix factorization method to recover the true subspace, and 2) novel robust principle component regression to prune adversarial instances based on the basis recovered in step  (1) . They characterize necessary and sufficient conditions for successfully recovering the true subspace and present a bound on expected prediction loss compared to ground truth. Trojaned host whereby attacker persists on the network. Training or config data is compromised and being ingested/trusted for model creation. \n #2b Indiscriminate Data Poisoning \n D e sc r i p t i o n D e sc r i p t i o n Ex a m p l e s Ex a m p l e s \n M i t i g a t i o n s M i t i g a t i o n s T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s Se v e r i t y Se v e r i t y \n #3 Model Inversion Attacks \n D e sc r i p t i o n D e sc r i p t i o n Ex a m p l e s Ex a m p l e s M i t i g a t i o n s M i t i g a t i o n s T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s Se v e r i t y Se v e r i t y \n Critical Goal is to ruin the quality/integrity of the data set being attacked. Many datasets are public/untrusted/uncurated, so this creates additional concerns around the ability to spot such data integrity violations in the first place. Training on unknowingly compromised data is a garbage-in/garbage-out situation. Once detected, triage needs to determine the extent of data that has been breached and quarantine/retrain. A company scrapes a well-known and trusted website for oil futures data to train their models. The data provider's website is subsequently compromised via SQL Injection attack. The attacker can poison the dataset at will and the model being trained has no notion that the data is tainted. Same as variant 2a. Authenticated Denial of service against a high-value asset \n Important The private features used in machine learning models can be recovered  [1] . This includes reconstructing private training data that the attacker does not have access to. Also known as hill climbing attacks in the biometric community  [16, 17]  This is accomplished by finding the input which maximizes the confidence level returned, subject to the classification matching the target  [4] .  [4]  Interfaces to models trained from sensitive data need strong access control. \n Rate-limit queries allowed by model Implement gates between users/callers and the actual model by performing input validation on all proposed queries, rejecting anything not meeting the model's definition of input correctness and returning only the minimum amount of information needed to be useful. Targeted, covert Information Disclosure #4 Membership Inference Attack \n D e sc r i p t i o n D e sc r i p t i o n M i t i g a t i o n s M i t i g a t i o n s T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s Se v e r i t y Se v e r i t y \n #5 Model Stealing \n D e sc r i p t i o n D e sc r i p t i o n This defaults to important per the standard SDL bug bar, but sensitive or personally identifiable data being extracted would raise this to critical. The attacker can determine whether a given data record was part of the model's training dataset or not  [1] . Researchers were able to predict a patient's main procedure (e.g: Surgery the patient went through) based on the attributes (e.g: age, gender, hospital)  [1] .  [12]  Research papers demonstrating the viability of this attack indicate Differential Privacy  [4, 9]  would be an effective mitigation. This is still a nascent field at Microsoft and AETHER Security Engineering recommends building expertise with research investments in this space. This research would need to enumerate Differential Privacy capabilities and evaluate their practical effectiveness as mitigations, then design ways for these defenses to be inherited transparently on our online services platforms, similar to how compiling code in Visual Studio gives you on-by-default security protections which are transparent to the developer and users. The usage of neuron dropout and model stacking can be effective mitigations to an extent. Using neuron dropout not only increases resilience of a neural net to this attack, but also increases model performance  [4] . Data Privacy. Inferences are being made about a data point's inclusion in the training set but the training data itself is not being disclosed This is a privacy issue, not a security issue. It is addressed in threat modeling guidance because the domains overlap, but any response here would be driven by Privacy, not Security. The attackers recreate the underlying model by legitimately querying the model. The functionality of the new model is same as that of the underlying model  [1] . Once the model is recreated, it can be inverted to recover feature information or make inferences on training data. Equation solving -For a model that returns class probabilities via API output, an attacker can craft queries to determine unknown variables in a model. Path Finding -an attack that exploits API particularities to extract the ' decisions' taken by a tree when classifying an input  [7] . Transferability attack -An adversary can train a local model-possibly by issuing prediction queries to the targeted model -and use it to craft adversarial examples that transfer to the target model  [8] . If your \n Ex a m p l e s Ex a m p l e s M i t i g a t i o n s M i t i g a t i o n s T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s Se v e r i t y Se v e r i t y  T r a d i t i o n a l P a r a l l e l s  T r a d i t i o n a l P a r a l l e l   model is extracted and discovered vulnerable to a type of adversarial input, new attacks against your production-deployed model can be developed entirely offline by the attacker who extracted a copy of your model. \n #6 Neural Net Reprogramming \n Ex a m p l e s Ex a m p l e s M i t i g a t i o n s M i t i g a t i o n s In settings where an ML model serves to detect adversarial behavior, such as identification of spam, malware classification, and network anomaly detection, model extraction can facilitate evasion attacks  [7] . Proactive/Protective Actions Minimize or obfuscate the details returned in prediction APIs while still maintaining their usefulness to \"honest\" applications  [7] . Define a well-formed query for your model inputs and only return results in response to completed, wellformed inputs matching that format. Return rounded confidence values. Most legitimate callers do not need multiple decimal places of precision. Unauthenticated, read-only tampering of system data, targeted high-value information disclosure? Important in security-sensitive models, Moderate otherwise Description By means of a specially crafted query from an adversary, Machine learning systems can be reprogrammed to a task that deviates from the creator's original intent  [1] . Weak access controls on a facial recognition API enabling 3 parties to incorporate into apps designed to harm Microsoft customers, such as a deep fakes generator. Identify and enforce a service-level agreement for your APIs. Determine the acceptable time-to-fix for an issue once reported and ensure the issue no longer repros once SLA expires. This is an abuse scenario. You're less likely to open a security incident on this than you are to simply disable the offender's account. \n Important to Critical \n Description An adversarial example is an input/query from a malicious entity sent with the sole aim of misleading the machine learning system  [1]  These examples can manifest in the physical domain, like a self-driving car being tricked into running a stop sign because of a certain color of light (the adversarial input) being shone on the stop sign, forcing the image T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s \n M i t i g a t i o n s M i t i g a t i o n s Se v e r i t y Se v e r i t y #8 Malicious ML providers who can recover training data \n D e sc r i p t i o n D e sc r i p t i o n T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s M i t i g a t i o n s M i t i g a t i o n s Se v e r i t y Se v e r i t y #9 Attacking the ML Supply Chain \n D e sc r i p t i o n D e sc r i p t i o n T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s M i t i g a t i o n s M i t i g a t i o n s Se v e r i t y Se v e r i t y #10 Backdoor Machine Learning D e sc r i p t i o n D e sc r i p t i o n recognition system to no longer see the stop sign as a stop sign. Elevation of Privilege, remote code execution These attacks manifest themselves because issues in the machine learning layer (the data & algorithm layer below AI-driven decisionmaking) were not mitigated. As with any other software *or* physical system, the layer below the target can always be attacked through traditional vectors. Because of this, traditional security practices are more important than ever, especially with the layer of unmitigated vulnerabilities (the data/algo layer) being used between AI and traditional software. \n Critical A malicious provider presents a backdoored algorithm, wherein the private training data is recovered. They were able to reconstruct faces and texts, given the model alone. \n Targeted information disclosure Research papers demonstrating the viability of this attack indicate Homomorphic Encryption would be an effective mitigation. This is an area with little current investment at Microsoft and AETHER Security Engineering recommends building expertise with research investments in this space. This research would need to enumerate Homomorphic Encryption tenets and evaluate their practical effectiveness as mitigations in the face of malicious ML-as-a-Service providers. \n Important if data is PII, Moderate otherwise Owing to large resources (data + computation) required to train algorithms, the current practice is to reuse models trained by large corporations and modify them slightly for task at hand (e.g: ResNet is a popular image recognition model from Microsoft). These models are curated in a Model Zoo (Caffe hosts popular image recognition models). In this attack, the adversary attacks the models hosted in Caffe, thereby poisoning the well for anyone else.  [1]  Compromise of third-party non-security dependency App store unknowingly hosting malware Minimize 3rd-party dependencies for models and data where possible. Incorporate these dependencies into your threat modeling process. Leverage strong authentication, access control and encryption between 1 /3 -party systems. st rd \n Critical The training process is outsourced to a malicious 3rd party who tampers with training data and delivered a \n T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s M i t i g a t i o n s M i t i g a t i o n s R e a c t i v e / D e fe n si v e D e t e c t i o n A c t i o n s R e a c t i v e / D e fe n si v e D e t e c t i o n A c t i o n s P r o a c t i v e / P r o t e c t i v e A c t i o n s P r o a c t i v e / P r o t e c t i v e A c t i o n s R e sp o n se A c t i o n s R e sp o n se A c t i o n s Se v e r i t y Se v e r i t y #11 Exploit software dependencies of the ML system \n D e sc r i p t i o n D e sc r i p t i o n T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s M i t i g a t i o n s M i t i g a t i o n s trojaned model which forces targeted mis-classifications, such as classifying a certain virus as non-malicious  [1] . This is a risk in ML-as-a-Service model-generation scenarios.  [12]  Compromise of third-party security dependency Compromised Software Update mechanism \n Certificate Authority compromise The damage is already done once this threat has been discovered, so the model and any training data provided by the malicious provider cannot be trusted. \n Train all sensitive models in-house Catalog training data or ensure it comes from a trusted third party with strong security practices Threat model the interaction between the MLaaS provider and your own systems Same as for compromise of external dependency Critical In this attack, the attacker does NOT manipulate the algorithms. Instead, exploits software vulnerabilities such as buffer overflows or cross-site scripting  [1] . It is still easier to compromise software layers beneath AI/ML than attack the learning layer directly, so traditional security threat mitigation practices detailed in the Security Development Lifecycle are essential. This can be manifested by trolls using certain target words in a way that the AI will ban them, effectively denying service to legitimate users with a name matching a \"banned\" word. \n Compromised Open Forcing benign emails to be classified as spam or causing a malicious example to go undetected. These are also known as model evasion or mimicry attacks. Artificial Intelligence (AI) and Machine Learning (ML) are already making a big impact on how people work, socialize, and live their lives. As consumption of products and services built around AI/ML increases, specialized actions must be undertaken to safeguard not only your customers and their data, but also to protect your AI and algorithms from abuse, trolling and extraction. This document shares some of Microsoft's security lessonslearned from designing products and operating online services built on AI. While it is difficult to predict how this area will unfold, we have concluded that there are actionable issues to address now. Moreover, we found that there are strategic problems that the tech industry must get ahead of to insure the long-term safety of customers and security of their data. \n Attacker can craft inputs This document is not about AI-based attacks or even AI being leveraged by human adversaries. Instead, we focus on issues that Microsoft and industry partners will need to address to protect AI-based products and services from highly sophisticated, creative and malicious attacks, whether carried out by individual trolls or entire wolfpacks. This document focuses entirely on security engineering issues unique to the AI/ML space, but due to the expansive nature of the InfoSec domain it is understood that issues and findings discussed here will overlap to a degree with the domains of privacy and ethics. As this document highlights challenges of strategic importance to the tech industry, the target audience for this document is security engineering leadership industry-wide. Our early findings suggest that: AI/ML-specific pivots to existing security practices are required to mitigate the types of security issues discussed in this document. Machine Learning models are largely unable to discern between malicious input and benign anomalous data. A significant source of training data is derived from un-curated, unmoderated, public datasets which are open to 3 -party contributions. Attackers don't need to compromise datasets when they are free to contribute to them. Over time, low-confidence malicious data becomes high-confidence trusted data, provided that the data structure/formatting remains correct. rd Given the great number of layers of hidden classifiers/neurons which can be leveraged in a deep learning model, too much trust is placed on the output of AI/ML decision-making processes and algorithms without a critical understanding of how these decisions were reached. This obfuscation creates an inability to \"show your work\" and makes it difficult to provably defend AI/ML findings when called into question. AI/ML is increasingly used in support of high-value decision-making processes in medicine and other industries where the wrong decision may result in serious injury or death. A lack of forensics reporting capabilities in AI/ML prevent these high-value conclusions from being defensible in both the court of law and court of public opinion. The goals of this document are to (1) highlight security engineering issues which are unique to the AI/ML space, New Security Engineering Challenges AI requires new pivots to traditional secure design/secure operations models: the introduction of Resilience and Discretion (2) surface some initial thoughts and observations on emerging threats and (  3 ) share early thoughts on potential remediation. Some of the challenges in this document are problems that the industry needs to get ahead of in the next two years, others are issues that we're already being forced to address today. Without deeper investigation into the areas covered in this document, we risk future AI becoming a black box to its through our inability to trust or understand (and modify if necessary) AI decision-making processes at a mathematical level  [7] . From a security perspective, this effectively means loss of control and a departure from Microsoft's guiding principles on Artificial Intelligence  [4, 8] . Traditional software attack vectors are still critical to address, but they do not provide sufficient coverage in the AI/ML threat landscape. The tech industry must avoid fighting next-gen issues with last-gen solutions by building new frameworks and adopting new approaches which address gaps in the design and operation of AI/ML-based services: 1. As discussed below, secure development and operations foundations must incorporate the concepts of Resilience and Discretion when protecting AI and the data under its control. AI-specific pivots are required in the areas of Authentication, Separation of Duty, Input Validation and Denial of Service mitigation. Without investments in these areas, AI/ML services will continue to fight an uphill battle against adversaries of all skill levels. 2. AI must be able to recognize bias in others, without being biased in its own interactions with humans. Accomplishing this requires a collective and evolving understanding of biases, stereotypes, vernacular, and other cultural constructs. Such an understanding will help protect AI from social engineering and dataset tampering attacks. A properly implemented system will actually become stronger from such attacks and be able to share its expanded understanding with other AIs. 3. Machine Learning algorithms must be capable of discerning maliciously-introduced data from benign \"Black Swan\" events  [1]  by rejecting training data with negative impact on results. Otherwise, learning models will always be susceptible to gaming by attackers and trolls. 4. AI must have built-in forensic capabilities. This enables enterprises to provide customers with transparency and accountability of their AI, ensuring its actions are not only verifiably correct but also legally defensible. These capabilities also function as an early form of \"AI intrusion detection\", allowing engineers to determine the exact point in time that a decision was made by a classifier, what data influenced it, and whether or not that data was trustworthy. Data visualization capabilities in this area are rapidly advancing and show promise to help engineers identify and resolve root causes for these complex issues  [11] . 5. AI must recognize and safeguard sensitive information, even if humans don't recognize it as such. Rich user experiences in AI require vast amounts of raw data to train on, so \"over-sharing\" by customers must be planned for. Each of these areas, including threats and potential mitigations, is discussed in detail below. AI designers will always need to ensure the confidentiality, integrity and availability of sensitive data, that the AI system is free of known vulnerabilities, and provide controls for the protection, detection and response to malicious behavior against the system or the user's data. The traditional ways of defending against malicious attacks do not provide the same coverage in this new paradigm, where voice/video/image-based attacks can circumvent current filters and defenses. New threat modeling aspects must be explored in order to prevent new abuses from exploiting our AI. This goes far beyond AI must be able to recognize bias in others without being biased on its own identifying traditional attack surface via fuzzing or input manipulation (those attacks have their own AI-specific pivots too). It requires incorporating scenarios unique to the AI/ML space. Key among these are AI user experiences such as voice, video and gestures. The threats associated with these experiences have not been traditionally modeled. For example, video content is now being tailored to induce physical effects. Additionally, research has demonstrated that audio-based attack commands can be crafted  [10] . The unpredictability, creativity, and maliciousness of criminals, determined adversaries, and trolls requires us to instill our AIs with the values of Resilience Resilience and Discretion Discretion: Resilience: Resilience: The system should be able to identify abnormal behaviors and prevent manipulation or coercion outside of normal boundaries of acceptable behavior in relation to the AI system and the specific task. These are new types of attacks specific to the AI/ML space. Systems should be designed to resist inputs that would otherwise conflict with local laws, ethics and values held by the community and its creators. This means providing AI with the capability to determine when an interaction is going \"off script. \" This could be achieved with the following methods: 1. Pinpoint individual users who deviate from norms set by the various large clusters of similar users e.g. users who seem to type to fast, respond too fast, don't sleep, or trigger parts of the system other users do not. 2. Identify patterns of behavior that are known to be indicators of malicious intent probing attacks and the start of the Network Intrusion Kill Chain. 3. Recognize any time when multiple users act in a coordinated fashion; e.g., multiple users all issuing the same unexplainable yet deliberately crafted query, sudden spikes in the number of users or sudden spikes in activation of specific parts of an AI system. Attacks of this type should be considered on par with Denial of Service attacks since the AI may require bugfixes and retraining in order not to fall for the same tricks again. Of critical importance is the ability to identify malicious intent in the presence of countermeasures such as those used to defeat sentiment analysis APIs  [5] . \n Discretion Discretion: AI should be a responsible and trustworthy custodian of any information it has access to. As humans, we will undoubtedly assign a certain level of trust in our AI relationships. At some point, these agents will talk to other agents or other humans on our behalf. We must be able to trust that an AI system has enough discretion to only share in a restricted form what it needs to share about us so other agents can complete tasks on its behalf. Furthermore, multiple agents interacting with personal data on our behalf should not each need global access to it. Any data access scenarios involving multiple AIs or bot agents should limit the lifetime of access to the minimum extent required. Users should also be able to deny data and reject the authentication of agents from specific corporations or locales just as web browsers allow site blacklisting today. Solving this problem requires new thinking on inter-agent authentication and data access privileges like the cloud-based user authentication investments made in the early years of cloud computing. While AI should be fair and inclusive without discriminating against any particular group of individuals or valid outcomes, it needs to have an innate understanding of bias to achieve this. Without being trained to recognize bias, trolling, or sarcasm, AI will be duped by those seeking cheap laughs at best, or cause harm to customers at worst. Achieving this level of awareness calls for \"good people teaching AI bad things\" as it effectively requires a comprehensive and evolving understanding of cultural biases. AI should be able to recognize a user it has had negative interactions with in the past and exercise appropriate caution, similar to how parents teach their children to be wary of strangers. The best way to approach this is by carefully exposing the AI to trolls in a controlled/moderated/limited fashion. This way AI can learn the difference between a benign user \"kicking the Machine Learning Algorithms must be capable of discerning maliciously-introduced data from benign \"Black Swan\" events AI must have built-in forensics and security logging to provide transparency and accountability tires\" and actual maliciousness/trolling. Trolls provide a valuable stream of training data for AI, making it more resilient against future attacks. AI should also be able to recognize bias in datasets it trains on. This could be cultural or regional, containing the vernacular in use by a particular group of people, or topics/viewpoints of specific interest to one group. As with maliciously-introduced training data, AI must be resilient to the effects of this data on its own inferences and deductions. At its core, this is a sophisticated input validation issue with similarities to bounds checking. Instead of dealing with buffer lengths and offsets, the buffer and bounds checks are red-flagged words from a broad range of sources. The conversation history and context in which words are used are also key. Just as defense-indepth practices are used to layer protections on top of a traditional Web Service API frontend, multiple layers of protection should be leveraged in bias recognition and avoidance techniques. Numerous whitepapers have been published on the theoretical potential of ML model/classifier tampering and extraction/theft from services where attackers have access to both the training data set and an informed understanding of the model in use  [2, 3, 6, 7] . The over-arching issue here is that all ML classifiers can be tricked by an attacker who has control over training set data. Attackers don't even need the ability to modify existing training set data, they just need to be able to add to it and have their inputs become \"trusted\" over time through the ML classifier's inability to discern malicious data from genuine anomalous data. This training data supply chain issue introduces us to the concept of \"Decision Integrity\" -the ability to identify and reject maliciously introduced training data or user input before it has a negative impact on classifier behavior. The rationale here is that trustworthy training data has a higher probability of generating trustworthy outcomes/decisions. While it is still crucial to train on and be resilient to untrusted data, the malicious nature of that data should be analyzed prior to it becoming part of a high-confidence body of training data. Without such measures, AI could be coerced into overreacting to trolling and deny service to legitimate users. This is of particular concern where unsupervised learning algorithms are training on uncurated or untrusted datasets. This means that attackers can introduce any data they want provided the format is valid and the algorithm is trained on it, effectively trusting that data point equally with the rest of the training set. With enough crafted inputs from the attacker, the training algorithm loses the ability to discern noise and anomalies from high-confidence data. As an example of this threat, imagine a database of stop signs throughout the world, in every language. That would be extremely challenging to curate because of the number of images and languages involved. Malicious contribution to that dataset would go largely unnoticed until self-driving cars no longer recognize stop signs. Data resilience and decision integrity mitigations will have to work hand in hand here to identify and eliminate the training damage done by malicious data to prevent it from becoming a core part of the learning model. AI will eventually be capable of acting in a professional capacity as an agent on our behalf, assisting us with high-impact decision-making. An example of this could be an AI that assists in the processing of financial transactions. If the AI were exploited, and transactions manipulated in some way, the consequences could range from the individual to the systemic. In high-value scenarios, AI will need appropriate forensic and security logging to provide integrity, transparency, accountability, and in some instances, evidence where civil or criminal liability may arise. Essential AI services will need auditing/event-tracing facilities at the algorithm level whereby developers can examine the recorded state of specific classifiers which may have led to an inaccurate decision. This capability is needed industry-wide in order to prove the correctness and transparency of AI-generated decisions whenever AI must safeguard sensitive information, even if humans don't Early Observations on Addressing AI Security Issues called into question. Event tracing facilities could start with the correlation of basic decision-making information such as: 1. The timeframe in which the last training event occurred 2. The timestamp of the most recent dataset entry trained upon 3. Weights and confidence levels of key classifiers used to arrive at high-impact decisions 4. The classifiers or components involved in the decision 5. The final high-value decision reached by the algorithm Such tracing is overkill for the majority of algorithm-assisted decision making. However, having the ability to identify the data points and algorithm metadata leading to specific results will be of great benefit in high-value decision making. Such capabilities will not only demonstrate trustworthiness and integrity through the algorithm's ability to \"show its work\", but this data could also be used for fine-tuning as well. Another forensic capability needed in AI/ML is tamper detection. Just as we need our AIs to recognize bias and not be susceptible to it, we should have forensic capabilities available to aid our engineers in detecting and responding to such attacks. Such forensic capabilities will be of tremendous value when paired with data visualization techniques  [11]  allowing the auditing, debugging and tuning of algorithms for more effective results. Rich experiences require rich data. Humans already volunteer vast amounts of data for ML to train against. This ranges from the mundane video streaming queue contents to trends in credit card purchases/transaction histories used to detect fraud. AI should have an ingrained sense of discretion when it comes to handling user data, always acting to protect it even when it is volunteered freely by an over-sharing public. As an AI can have an authenticated group of \"peers\" it talks to in order to accomplish complex tasks, it must also recognize the need to restrict the data it shares with those peers. Despite the nascent state of this project, we believe the evidence compiled to date shows deeper investigation into each of the areas below will be key in moving our industry towards more trustworthy and secure AI/ML products/services. The following are our early observations and thoughts on what we' d like to see done in this space. 1. AI/ML-focused penetration testing and security review bodies could be established to ensure that our future AI shares our values and aligns to the Asilomar AI Principles. a. Such a group could also develop tools and frameworks that could be consumed industry-wide in support of securing their AI/ML-based services. b. Over time, this expertise will build up within engineering groups organically, as it did with traditional security expertise over the past 10 years. 2. Training could be developed which empowers enterprises to deliver on goals such as democratizing AI while mitigating the challenges discussed in this document. a. AI-specific security training ensures that engineers are aware of the risks posed to to their AI and the resources at their disposal. This material needs to be delivered in conjunction with current training on protecting customer data. b. This could be accomplished without requiring every data scientist to become a security expertinstead, the focus is placed on educating developers on Resilience and Discretion as applied to their AI Conclusion Bibliography use cases. c. Developers will need to understand the secure \"building blocks\" of AI services that will be reused across their enterprise. There will need to be an emphasis on fault-tolerant design with subsystems which can be easily turned off (e.g. image processors, text parsers). 3. ML Classifiers and their underlying algorithms could be hardened and capable of detecting malicious training data without it contaminating valid training data currently in use or skewing the results. a. Techniques such as Reject on Negative Input  [6]  need researcher cycles to investigate. b. This work involves mathematical verification, proof-of-concept in code, and testing against both malicious and benign anomalous data. c. Human spot-checking/moderation may be beneficial here, particularly where statistical anomalies are present. d. \"Overseer classifiers\" could be built to have a more universal understanding of threats across multiple AIs. This vastly improves the security of the system because the attacker can no longer exfiltrate any one particular model. e. AIs could be linked together to identify threats in each other's systems 4. A centralized ML auditing/forensics library could be built that establishes a standard for transparency and trustworthiness of AI. a. Query capabilities could also be built for auditing and reconstruction of high business impact decisions by AI. 5. The vernacular in use by adversaries across different cultural groups and social media could be continuously inventoried and analyzed by AI in order to detect and respond to trolling, sarcasm, etc. a. AIs need to be resilient in the face of all kinds of vernacular, whether technical, regional, or forumspecific. b. This body of knowledge could also be leveraged in content filtering/labeling/blocking automation to address moderator scalability issues. c. This global database of terms could be hosted in development libraries or even exposed via cloud service APIs for reuse by different AIs, ensuring new AIs benefit from the combined wisdom of older ones. 6. A \"Machine Learning Fuzzing Framework\" could be created which provides engineers with the ability to inject various types of attacks into test training sets for AI to evaluate. a. This could focus on not only text vernacular, but image, voice and gesture data as well as permutations of those data types. The Asilomar AI Principles illustrate the complexity of delivering on AI in a fashion that consistently benefits humanity. Future AIs will need to interact with other AIs to deliver rich, compelling user experiences. That means it simply is not good enough for Microsoft to \"get AI right\" from a security perspective -the world has to. We need industry alignment and collaboration with greater visibility brought to the issues in this document in a fashion similar to our worldwide push for a Digital Geneva Convention  [9] . By addressing the issues presented here, we can begin guiding our customers and industry partners down a path where AI is truly democratized and augments the intelligence of all humanity. Identifying Security Bug Reports Based Solely on Report Titles and Noisy Data Index Terms -Machine Learning, Mislabeling, Noise, Security Bug Report, Bug Repositories Identifying security related issues among reported bugs is a pressing need among software development teams as such issues call for more expedited fixes in order to meet compliance requirements and ensure the integrity of the software and customer data. Machine learning and artificial intelligence tools promise to make the software development faster, agile and correct. Several researchers have applied machine learning to the problem of identifying security bugs  [2] ,  [7] ,  [8] ,  [18] .Previous published studies have assumed that the entire bug report is available for training and scoring a machine learning model. This is not necessarily the case. There are situations where the entire bug report cannot be made available. For example, the bug report might contain passwords, personally identifying information (PII) or other kinds of sensitive data -a case we are currently facing at Microsoft. It is therefore important to establish how well security bug identification can be performed using less information, such as when only the title of the bug report is available. Additionally, bug repositories often contain mislabeled entries  [7] : non-security bug reports classified as security related and vice-versa. There are several reasons for the occurrence of mislabeling, ranging from the development team's lack of expertise in security, to the fuzziness of certain problems, e.g. it is possible for nonsecurity bugs to be exploited in an indirect way as to cause a security implication. This is a serious problem since the mislabeling of SBRs results in security experts having to manually review bug database in an expensive and time-consuming effort. Understanding how noise affects different classifiers and how robust (or fragile) different machine learning techniques are in the presence of data sets contaminated with different kinds of noise is a problem that must be addressed for to bring automatic classification to the practice of software engineering. Preliminary work argues that bug repositories are intrinsically noisy and that the noise might have an adverse effect on the performance machine learning classifiers  [7] . There lacks, however, any systematic and quantitative study of how different levels and types of noise affect the performance of different supervised machine learning algorithms for the problem of identifying security bug reports (SRBs). In this study, we show that the classification of bug reports can be performed even when solely the title is available for training and scoring. To the best of our knowledge this is the very first work to do so. Additionally, we provide the first systematic study of the effect of noise in bug report classification. We make a comparative OUR RESEARCH CONTRIBUTIONS: OUR RESEARCH CONTRIBUTIONS: \n II. PREVIOUS WORKS \n MACHINE LEARNING APPLICATIONS TO BUG REPOSITORIES. MACHINE LEARNING APPLICATIONS TO BUG REPOSITORIES. study of robustness of three machine learning techniques (logistic regression, naive Bayes and AdaBoost) against class-independent noise. While there are some analytical models that capture the general influence of noise for a few simple classifiers  [5] ,  [6] , these results do not provide tight bounds on the effect of the noise on precision and are valid only for a particular machine learning technique. An accurate analysis of the effect of noise in machine learning models is usually performed by running computational experiments. Such analyses have been done for several scenarios ranging from software measurement data  [4] , to satellite image classification  [13]  and medical data  [12] . Yet these results cannot be translated to our specific problem, due to its high dependency on the nature of the data sets and underlying classification problem. To the best of our knowledge, there are no published results on the problem of the effect of noisy data sets on security bug report classification in particular. We train classifiers for the identification of security bug reports (SBRs) based solely on the title of the reports. To the best of our knowledge this is the first work to do so. Previous works either used the complete bug report or enhanced the bug report with additional complementary features. Classifying bugs based solely on the tile is particularly relevant when the complete bug reports cannot be made available due to privacy concerns. For example, it is notorious the case of bug reports that contain passwords and other sensitive data. We also provide the first systematic study of the label noise-tolerance of different machine learning models and techniques used for the automatic classification of SBRs. We make a comparative study of robustness of three distinct machine learning techniques (logistic regression, naive Bayes and AdaBoost) against class-dependent and class-independent noise. The remaining of the paper is presented as follows: In section II we present some of the previous works in the literature. In section III we describe the data set and how data is pre-processed. The methodology is described in section IV and the results of our experiments analyzed in section V. Finally, our conclusions and future works are presented in VI. There exists an extensive literature in applying text mining, natural language processing and machine learning on bug repositories in an attempt to automate laborious tasks such as security bug detection  [2] ,  [7] ,  [8] ,  [18] , bug duplicate identification  [3] , bug triage  [1] ,  [11] , to name a few applications. Ideally, the marriage of machine learning (ML) and natural language processing potentially reduces the manual work required to curate bug databases, shorten the required time for accomplishing these tasks and can increase the reliability of the results. In  [7]  the authors propose a natural language model to automate the classification of SBRs based on the description of the bug. The authors extract a vocabulary from all bug description in the training data set and manually curate it into three lists of words: relevant words, stop words (common words that seem irrelevant for classification), and synonyms. They compare the performance of security bug classifier trained on data that is all evaluated by security engineers, and a classifier trained on data that was labeled by bug reporters in general. Although their model is clearly more effective when trained on data reviewed by security engineers, the proposed model is based on a manually derived vocabulary, which makes it dependent on human curation. Moreover, there is no analysis of how different levels of noise affects their model, how different classifiers respond to noise, and if noise in either class affects performance differently. Zou et. al  [18]  make use of multiple types of information contained in a bug report which involve the nontextual fields of a bug report (meta features, e.g., time, severity, and priority) and the textual content of a bug report (textual features, i.e., the text in summary fields). Based on these features, they build a model to automatically identify the SBRs via natural language processing and machine learning techniques. In  [8]  the authors perform a similar analysis, but additionally they compare the performance of supervised and In  [2]  the authors also explore different machine learning techniques to classify bugs as SBRs or NSBRs (Non-Security Bug Report) based on their descriptions. They propose a pipeline for data processing and model training based on TFIDF. They compare the proposed pipeline with a model based on bag-of-words and naive Bayes. Wijayasekara et al.  [16]  also used text mining techniques to generate the feature vector of each bug report based on frequent words to identify Hidden Impact Bugs (HIBs). Yang et al.  [17]  claimed to identify high impact bug reports (e.g., SBRs) with the help of Term Frequency (TF) and naive Bayes. In  [9]  the authors propose a model to predict the severity of a bug. The problem of dealing with data sets with label noise has been extensively studied. Frenay and Verleysen propose a label noise taxonomy in  [6] , in order to distinguish different types of noisy label. The authors propose three different types of noise: label noise which occurs independently of the true class and of the values of the instance features; label noise that depends only on the true label; and label noise where the mislabelling probability also depends on the feature values. In our work we study the first two types of noise. From a theoretical perspective, label noise usually decreases a model's performance  [10] , except in some specific cases  [14] . In general, robust methods rely on overfitting avoidance to handle label noise  [15] . The study of noise effects in classification has been done before in many areas such as satellite image classification  [13] , software quality classification  [4]  and medical domain classification  [12] . To the best of our knowledge, there are no published works studying the precise quantification of the effects of noisy labels in the problem of SBRs classification. In this scenario, the precise relationship among noise levels, noise types and performance degradation has not been established. Moreover, it is worthwhile to understand how different classifiers behave in the presence of noise. More generally, we are unaware of any work that systematically studies the effect of noisy data sets on the performance of different machine learning algorithms in the context of software bug reports. Our data set consists of 1,073,149 bug titles, 552,073 of which correspond to SBRs and 521,076 to NSBRs. The data was collected from various teams across Microsoft in the years 2015, 2016, 2017 and 2018. All the labels were obtained by signature-based bug verification systems or human labeled. Bug titles in our data set are very short texts, containing around 10 words, with an overview of the problem. A. Data Pre-processing We parse each bug title by its blank spaces, resulting in a list of tokens. We process each list of tokens as follows: Remove all tokens that are file paths Split tokens where the following symbols are present: { , (, ), -, }, {, [, ], } Remove stop words , tokens that are composed by numeric characters only and tokens that appear fewer than 5 times in the entire corpus. The process of training our machine learning models consists of two main steps: encoding the data into feature vectors and training supervised machine learning classifiers. The first part involves encoding data into feature vectors using the term frequencyinverse document frequency algorithm (TF-IDF), as used in  [2] . TF-IDF is an information retrieval technique that weighs a terms frequency (TF) and its inverse document frequency (IDF). Each word or term has its respective TF and IDF score. The TF-IDF algorithm assigns the importance to that word based on the number of times it appears in the document, and more importantly, it checks how relevant the keyword is throughout the collection of titles in the data set. We \n B. Types of Noise B. Types of Noise \n V. EXPERIMENTAL RESULTS trained and compared three classification techniques: naive Bayes (NB), boosted decision trees (AdaBoost) and logistic regression (LR). We have chosen these techniques because they have been shown to performance well for the related task of identifying security bug reports based on the entire report in the literature. These results were confirmed in a preliminary analysis where these three classifiers outperformed support vector machines and random forests. In our experiments we utilize the scikit-learn library for encoding and model training. The noise studied in this work refers to noise in the class label in the training data. In the presence of such noise, as a consequence, the learning process and resulting model are impaired by mislabelled examples. We analyze the impact of different noise levels applied to the class information. Types of label noise have been discussed previously in the literature using various terminologies. In our work, we analyze the effects of two different label noise in our classifiers: class-independent label noise, which is introduced by picking instances at random and flipping their label; and class-dependent noise, where classes have different likelihood to be noisy. a) Class-independent noise: Class-independent noise refers to the noise that occurs independently of the true class of the instances. In this type of noise, the probability of mislabeling p is the same for all instances in the data set. We introduce class-independent noise in our data sets by flipping each label in our data set at random with probability p . We add noise to the training and validation data sets for different levels of p , p and p . We do not make any modifications to the test data set. The different noise levels used are P = {0.05 × i|0 < i < 10}. Tune models using validation data set (with noise); Test models using test data set (noiseless). In this section analyze the results of experiments conducted according to the methodology described in section IV. a) Model performance without noise in the training data set: One of the contributions of this paper is the proposal of a machine learning model to identify security bugs by using only the title of the bug as data for decision making. This enables the training of machine learning models even when development teams do not wish to share bug reports in full due to presence of sensitive data. We compare the performance of three machine learning models when trained using only bug titles. The logistic regression model is the best performing classifier. It is the classifier with the highest AUC value, of 0.9826, recall of 0.9353 for a FPR value of 0.0735. The naive Bayes Classifier presents slightly lower performance than the logistic regression Classifier, with an AUC of 0.9779 and a recall of 0.9189 for a FPR of 0.0769. The AdaBoost classifier has an inferior performance in comparison to the two previously mentioned classifiers. It achieves an AUC of 0.9143, and a recall of 0.7018 for a 0.0774 FPR. The area under the ROC curve (AUC) is a good metric for comparing performance of several models, as it summarizes in a single value the TPR vs. FPR relation. In the subsequent analysis we will restrict our comparative analysis to AUC values. \n A. Class Noise: single-class One can imagine a scenario where all bugs are assigned to class NSBR by default, and a bug will only be assigned to class SBR if there is a security expert reviewing the bug repository. This scenario is represented in the single-class experimental setting, where we assume that p = 0 and 0 < p < 0.5. nsbr sbr From table II we observe a very small impact in the AUC for all three classifiers. The AUC-ROC from a model trained on p = 0 compared to a AUC-ROC of model where p = 0.25 differs by 0.003 for logistic regression, 0.006 for naive Bayes, and 0.006 for AdaBoost. In the case of p = 0.50, the AUC measured for each of the models differ from the model trained with p = 0 by 0.007 for logistic regression, 0.011 for naive Bayes, and 0.010 for AdaBoost. logistic regression classifier trained in the presence of single-class noise presents the smallest variation in its AUC metric, i.e. a more robust behavior, when compared to our naive Bayes and AdaBoost classifiers. In Table  III  we observe a decrease in the AUC-ROC for every noise increment in the experiment. The AUC-ROC measured from a model trained on noiseless data compared to a AUC-ROC of model trained with classindependent noise with p = 0.25 differs by 0.011 for logistic regression, 0.008 for naive Bayes, and 0.0038 for AdaBoost. We observe that label noise does not impact the AUC of naive Bayes and AdaBoost classifiers significantly when noise levels are lower than 40%. On the other hand, logistic regression Classifier experiences an impact in AUC measure for label noise levels above 30%. Tables IV, V, VI show the variation of AUC as noise is increased in different levels in each class for logistic regression in Table  IV , for naive Bayes in Table V and for AdaBoost in Table  VI . For all classifiers, we notice an impact in AUC metric when both classes contains noise level above 30%. naive Bayes behaves mot robustly. The impact on AUC is very small even when the 50% of the label in the positive class are flipped, provided that the negative class contains 30% of noisy labels or less. In this case, the drop in AUC is of 0.03. AdaBoost presented the most robust behavior of all three classifiers. A significant change in AUC will only happen for noise levels greater than 45% in both classes. In that case, we start observing an AUC decay greater than 0.02. \n D. On The Presence of Residual Noise in the Original Data Set Our data set was labeled by signature-based automated systems and by human experts. Moreover, all bugs reports have been further reviewed and closed by human experts. While we expect that the amount of noise in our data set is minimal and not statistically significant, the presence of residual noise does not invalidate our \n VI. CONCLUSIONS AND FUTURE WORKS \n FUTURE WORKS FUTURE WORKS \n REFERENCES REFERENCES conclusions. Indeed, for the sake of illustration, assume that the original data set is corrupted by a classindependent noise equal to 0 < p < 1/2 independent and identically distributed (i.i.d) for every entry. If we, on top of the original noise, add a class-independent noise with probability p i.i.d, the resulting noise per entry will be p * = p(1 − p )+(1 − p)p . For 0 < p,p < 1/2, we have that the actual noise per label p * is strictly larger than the noise we artificially add to the data set p . Thus, the performance of our classifiers would be even better if they were trained with a completely noiseless data set (p = 0) in the first place. In summary, the existence of residual noise in the actual data set means that the resilience against noise of our classifiers is better than the results here presented. Moreover, if the residual noise in our data set were statistically relevant, the AUC of our classifiers would become 0.5 (a random guess) for a level of noise strictly less than 0.5. We do not observe such behavior in our results. First, we have shown the feasibility of security bug report classification based solely on the title of the bug report. This is particularly relevant in scenarios where the entire bug report is not available due to privacy constraints. For example, in our case, the bug reports contained private information such as passwords and cryptographic keys and were not available for training the classifiers. Our result shows that SBR identification can be performed at high accuracy even when only report titles are available. Our classification model that utilizes a combination of TF-IDF and logistic regression performs at an AUC of 0.9831. Second, we analyzed the effect of mislabeled training and validation data. We compared three well-known machine learning classification techniques (naive Bayes, logistic regression and AdaBoost) in terms of their robustness against different noise types and noise levels. All three classifiers are robust to single-class noise. Noise in the training data has no significant effect in the resulting classifier. The decrease in AUC is very small ( 0.01) for a level of noise of 50%. For noise present in both classes and is class-independent, naive Bayes and AdaBoost models present significant variations in AUC only when trained with a data set with noise levels greater than 40%. Finally, class-dependent noise significantly impacts the AUC only when there is more then 35% noise in both classes. AdaBoost showed the most robustness. The impact in AUC is very small even when the positive class has 50% of its labels noisy, provided that the negative class contains 45% of noisy labels or less. In this case, the drop in AUC is less than 0.03. To the best of our knowledge, this is the first systematic study on the effect of noisy data sets for security bug report identification. In this paper we have started the systematic study of the effects of noise in the performance of machine learning classifiers for the identification of security bugs. There are several interesting sequels to this work, including: examining the effect of noisy data sets in determining the severity level of a security bug; understanding the effect of class imbalance on the resilience of the trained models against noise; understanding the effect of noise that is adversarially introduced in the data set. [ Microsoft recommends customers get ahead of this issue by removing TLS 1.0 dependencies in their environments and disabling TLS 1.0 at the operating system level where possible. Given the length of time TLS 1.0 has been supported by the software industry, it is highly recommended that any TLS 1.0 deprecation plan include the following: Code analysis to find/fix hardcoded instances of TLS 1.0 or older security protocols. Network endpoint scanning and traffic analysis to identify operating systems using TLS 1.0 or older protocols. Full regression testing through your entire application stack with TLS 1.0 disabled. Migration of legacy operating systems and development libraries/frameworks to versions capable of negotiating TLS 1.2 by default. Compatibility testing across operating systems used by your business to identify any TLS 1.2 support issues. Coordination with your own business partners and customers to notify them of your move to deprecate TLS 1.0. Understanding which clients may no longer be able to connect to your servers once TLS 1.0 is disabled. The goal of this document is to provide recommendations which can help remove technical blockers to disabling TLS 1.0 while at the same time increasing visibility into the impact of this change to your own customers. Completing such investigations can help reduce the business impact of the next security vulnerability in TLS 1.0. For the purposes of this document, references to the deprecation of TLS 1.0 also include TLS 1.1. Enterprise software developers have a strategic need to adopt more future-safe and agile solutions (otherwise known as Crypto Agility) to deal with future security protocol compromises. While this document proposes agile solutions to the elimination of TLS hardcoding, broader Crypto Agility solutions are beyond the scope of this The Current State of Microsoft's TLS 1.0 implementation Ensuring support for TLS 1.2 across deployed operating systems  Microsoft's TLS 1.0 implementation is free of known security vulnerabilities. Due to the potential for future protocol downgrade attacks and other TLS 1.0 vulnerabilities not specific to Microsoft's implementation, it is recommended that dependencies on all security protocols older than TLS 1.2 be removed where possible (TLS 1.1/1.0/ SSLv3/SSLv2). In planning for this migration to TLS 1.2+, developers and system administrators should be aware of the potential for protocol version hardcoding in applications developed by their employees and partners. Hardcoding here means that the TLS version is fixed to a version that is outdated and less secure than newer versions. TLS versions newer than the hardcoded version cannot be used without modifying the program in question. This class of problem cannot be addressed without source code changes and software update deployment. Protocol version hardcoding was commonplace in the past for testing and supportability purposes as many different browsers and operating systems had varying levels of TLS support. Microsoft's Engineering Improvements to eliminate TLS 1.0 dependencies Finding and fixing TLS 1.0 dependencies in code If not already complete, it is highly recommended to conduct an inventory of operating systems used by your enterprise, customers and partners (the latter two via outreach/communication or at least HTTP User-Agent string collection). This inventory can be further supplemented by traffic analysis at your enterprise network edge. In such a situation, traffic analysis will yield the TLS versions successfully negotiated by customers/partners connecting to your services, but the traffic itself will remain encrypted. Since the v1 release of this document, Microsoft has shipped a number of software updates and new features in support of TLS 1.0 deprecation. These include: IIS custom logging to correlate client IP/user agent string, service URI, TLS protocol version and cipher suite. With this logging, admins can finally quantify their customers' exposure to weak TLS. SecureScore -To help Office 365 tenant admins identify their own weak TLS usage, the SecureScore portal has been built to share this information as TLS 1.0 exited support in Office 365 in October 2018. This portal provides Office 365 tenant admins with the valuable information they need to reach out to their own customers who may be unaware of their own TLS 1.0 dependencies. Please visit https://securescore.microsoft.com/ for more information. .Net Framework updates to eliminate app-level hardcoding and prevent framework-inherited TLS 1.0 dependencies. Developer Guidance and software updates have been released to help customers identify and eliminate .Net dependencies on weak TLS: Transport Layer Security (TLS) best practices with the .NET Framework FYI: All apps targeting .NET 4.5 or below are likely going to have to be modified in order to support TLS 1.2. TLS 1.2 has been backported to Windows Server 2008 SP2 and XP POSReady 2009 to help customers with legacy obligations. More announcements will be made in early 2019 and communicated in subsequent updates of this document. For products using the Windows OS-provided cryptography libraries and security protocols, the following steps should help identify any hardcoded TLS 1.0 usage in your applications: 1. Identify all instances of AcquireCredentialsHandle(). This helps reviewers get closer proximity to code blocks where TLS may be hardcoded. \n Review any instances of the SecPkgContext_SupportedProtocols and SecPkgContext_ConnectionInfo structures for hardcoded TLS. 3. In native code, set any non-zero assignments of grbitEnabledProtocols to zero. This allows the operating system to use its default TLS version. Following the fixes recommended in the section above, products should be regression-tested for protocol negotiation errors and compatibility with other operating systems in your enterprise. The most common issue in this regression testing will be a TLS negotiation failure due to a client connection attempt from an operating system or browser that does not support TLS 1.2. For example, a Vista client will fail to negotiate TLS with a server configured for TLS 1.2+ as Vista's maximum supported TLS version is 1.0. That client should be either upgraded or decommissioned in a TLS 1.2+ environment. Products using certificate-based Mutual TLS authentication may require additional regression testing as the certificate-selection code associated with TLS 1.0 was less expressive than that for TLS 1.2. If a product negotiates MTLS with a certificate from a non-standard location (outside of the standard named certificate stores in Windows), then that code may need updating to ensure the certificate is acquired correctly. Service interdependencies should be reviewed for trouble spots. Any services which interoperate with 3 -party services should conduct additional interop testing with those 3 parties. rd rd Any non-Windows applications or server operating systems in use require investigation / confirmation that they can support TLS 1.2. Scanning is the easiest way to determine this. A simple blueprint for testing these changes in an online service consists of the following: 1. Conduct a scan of production environment systems to identify operating systems which do not support TLS 1.2. 2. Scan source code and online service configuration files for hardcoded TLS as described in \"Finding and fixing TLS 1.0 dependencies in code\" After TLS hardcoding is addressed and operating system/development framework updates are completed, should you opt to deprecate TLS 1.0 it will be necessary to coordinate with customers and partners: Early partner/customer outreach is essential to a successful TLS 1.0 deprecation rollout. At a minimum this should consist of blog postings, whitepapers or other web content. Partners each need to evaluate their own TLS 1.2 readiness through the operating system/code scanning/regression testing initiatives described in above sections. Removing TLS 1.0 dependencies is a complicated issue to drive end to end. Microsoft and industry partners are taking action on this today to ensure our entire product stack is more secure by default, from our OS components and development frameworks up to the applications/services built on top of them. Following the recommendations made in this document will help your enterprise chart the right course and know what challenges to expect. It will also help your own customers become more prepared for the transition. As engineers worldwide work to eliminate their own dependencies on TLS 1.0, they run into the complex challenge of balancing their own security needs with the migration readiness of their customers. Disable Legacy TLS also allows an online service to offer two distinct groupings of endpoints on the same hardware: one which allows only TLS 1.2+ traffic, and another which accommodates legacy TLS 1.0 traffic. The changes are implemented in HTTP.sys, and in conjunction with the issuance of additional certificates, allow traffic to be routed to the new endpoint with the appropriate TLS version. Prior to this change, deploying such capabilities would require an additional hardware investment because such settings were only configurable system-wide via registry. \n TLS version A common deployment scenario features one set of hardware in a datacenter with customers of mixed needs: some need TLS 1.2 as an enforced minimum right now and others aren't done removing TLS 1.0 dependencies. Figure  1  illustrates TLS version selection and certificate binding as distinctly separate actions. This is the default functionality: \n Feature deployment guidance Option #1: IIS UI configuration (Available April 2020) Option #1: IIS UI configuration (Available April 2020) https://legacy.contoso.com directs customers with legacy TLS 1.0 needs (like those still migrating to TLS 1.2) to an endpoint which supports TLS 1.0 for a limited time. This allows customers to finish readiness testing for TLS 1.2 without service disruption and without blocking other customers who are ready for TLS 1.2. Traditionally, you' d need two physically separate hosts to handle all the traffic and provide for TLS version enforcement, as servicing TLS requests with a minimum protocol version requires disabling weaker protocols via system-wide registry settings. We have made this functionality available higher up the stack, where the TLS session is bound to the certificate, so a specific minimum TLS version can be assigned as described in Figure  2  below. Create a site binding for the SSL Certificate \"secure.contoso.com\" as shown below, then check \"Disable Legacy TLS\" and click OK. HTTP_SERVICE_CONFIG_SSL_FLAG_ENABLE_SESSION_TICKET: Enable/Disable Session Ticket for a particular SSL endpoint. HTTP_SERVICE_CONFIG_SSL_FLAG_LOG_EXTENDED_EVENTS : Enable/Disable extended event logging for a particular SSL endpoint. Additional events are logged to Windows Event Log. There is only one event supported as of now which is logged when the SSL handshake fails. HTTP_SERVICE_CONFIG_SSL_FLAG_DISABLE_LEGACY_TLS: Enable/Disable legacy TLS versions for a particular SSL endpoint. Setting this flag will disable TLS1.0/1.1 for that endpoint and will also restrict cipher suites that can be used to HTTP2 cipher suites. HTTP_SERVICE_CONFIG_SSL_FLAG_DISABLE_TLS12 : Enable/Disable TLS1.2 for a particular SSL endpoint. HTTP_SERVICE_CONFIG_SSL_FLAG_DISABLE_HTTP2: Enable/Disable HTTP/2 for a particular SSL endpoint. The simplest way to enable/disable this functionality per certificate in C++ is with the HTTP_SERVICE_CONFIG_SSL_FLAG_DISABLE_LEGACY_TLS flag provided by the HttpSetServiceConfiguration HTTP.sys API. When Disable Legacy TLS is set, the following restrictions are enforced: Disable SSL2, SSL3, TLS1.0 and TLS1.1 protocols. Disable encryption ciphers DES, 3DES, and RC4 (so only AES is used). Disable encryption cipher AES with CBC chaining mode (so only AES GCM is used). Disable RSA key exchange. Disable DH key exchange with key size less than 2048. Disable ECDH key exchanges with key size less than 224. Official documentation of these changes on docs.Microsoft.com is forthcoming. Disable Legacy TLS provides powerful new capabilities for enforcing TLS version/cipher suite floors on specific certificate/endpoint bindings. It also requires you to plan out the naming of the certificates issued with this functionality enabled. Some of the considerations include: Do I want the default path to my service endpoint to enforce TLS 1.2 today, and provide a different certificate as a backup \"legacy\" access point for users who need TLS 1.0? Should my default, already-in-use www.contoso.com certification use Disable Legacy TLS? If so, I may need to provide a legacy.contoso.com certificate and bind it to an endpoint allowing TLS 1.0. How can I best communicate the recommended usage of these certificates to my customers? You can leverage this feature to meet the needs of large groups of customers -those with an obligation to use TLS 1.2+, and those still working on the migration away from TLS 1.0, all without additional hardware expenditure. In addition to today's availability of per-certificate TLS version binding in Windows Server 2019, Microsoft will look to make Disable Legacy TLS available across its online services based on customer demand. Program Overview   \n Contact Us Since its launch in 2003, the Government Security Program (GSP) has provided governments and international organizations with the ability to access and inspect source code for a variety of Microsoft products. The Online Source offering provides access through our Code Center Premium (CCP) web site to the GSP participant's premises. Using two-factor authentication and an encrypted connection, the Online Source offering enables access to product source code such as Windows, Office, SharePoint Server, and Exchange Server. The site provides read-only access and a debugger that can be used to review code. The Online Source offering and CCP site enable GSP participants to evaluate individual system component functions, component interaction, and security and reliability capabilities. For example, CCP enables agencies to search source code trees as well as step through source code in a debugger, set breakpoints on routines and source lines, and examine variables and data structures in their source layout form. In advance of going to a Transparency Center, an agency may wish to review code in CCP to acquaint itself with it and therefore improve the value of visits and most efficiently use time at the Transparency Center.  \n Access Enablement Technical Trips The mission of Microsoft's Government Security Program (GSP) is to build trust through transparency. Since the program's inception in 2003, Microsoft has provided visibility into our technology and security artifacts which governments and international organizations can use to help protect themselves and their citizens. The Technical Data offering provides access to a broad range of confidential technical information, exclusive of source code, which allows government agencies and international organizations to evaluate the trustworthiness of Microsoft products and services. Information shared through the Technical Data offering is tailored to help GSP participants address their priorities. It provides a platform for customers to ask questions about security, to build trust in our products and services, and can include access to: Note: This sample document is for illustration purposes only. The content presented below outlines basic criteria to consider when creating security processes. It is not an exhaustive list of activities or criteria and should not be treated as such. Please refer to the definitions of terms in this section. \n Server Client Definitions of Terms Please refer to the Denial of Service Matrix for a complete matrix of server DoS scenarios. The server bar is usually not appropriate when user interaction is part of the exploitation process. If a Critical vulnerability exists only on server products, and is exploited in a way that requires user interaction and results in the compromise of the server, the severity may be reduced from Critical to Important in accordance with the NEAT/data definition of extensive user interaction presented at the start of the client severity pivot. Cases where the attacker can easily read information on the system from specific specific locations locations, including system information, which was not intended/ designed to be exposed. \n Example: Targeted disclosure of anonymous data Targeted disclosure of the existence of a file Targeted disclosure of a file version number Spoofing An entity (computer, server, user, process) is able to masquerade as a different, random entity that cannot be specifically selected. \n Example: Client properly authenticates to server, but server hands back a session from another random user who happens to be connected to the server at the same time A security assurance is either a security feature or another product feature/function that customers expect to offer security protection. Communications have messaged (explicitly or implicitly) that customers can rely on the integrity of the feature, and that's what makes it a security assurance. Security bulletins will be released for a shortcoming in a security assurance that undermines the customer's reliance or trust. \n Examples: Processes running with normal \"user\" privileges cannot gain \"admin\" privileges unless admin password/credentials have been provided via intentionally authorized methods. Internet-based JavaScript running in Internet Explorer cannot control anything the host operating system unless the user has explicitly changed the default security settings. Information disclosure (untargeted) \n Runtime information Example: \n Leak of random heap memory Tampering Temporary modification of data in a specific scenario that does not persist after restarting the OS/application \"User interaction\" can only happen in client-driven scenario. Normal, simple user actions, like previewing mail, viewing local folders, or file shares, are not extensive user interaction. \"Extensive\" includes users manually navigating to a particular website (for example, typing in a URL) or by clicking through a yes/no decision. \"Not extensive\" includes users clicking through e-mail links. Critical Important NEAT NEAT qualifier (applies to warnings only). Demonstrably, the UX is: N Necessary (Does the user really need to be presented with the decision?) E Explained (Does the UX present all the information the user needs to make this decision?) A Actionable (Is there a set of steps users can take to make good decisions in both benign and malicious scenarios?) T Tested (Has the warning been reviewed by multiple people, to make sure people understand how to respond to the warning?) Clarification: Clarification: Note that the effect of extensive user interaction is not one level reduction in severity, but is and has been a reduction in severity in certain circumstances where the phrase extensive user interaction appears in the bug bar. The intent is to help customers differentiate fast-spreading and wormable attacks from those, where because the user interacts, the attack is slowed down. This bug bar does not allow you to reduce the Elevation of Privilege below Important because of user interaction. Cases where the attacker can locate and read information on the system, including system information that was not intended or designed to be exposed.  \n Spoofing Ability for attacker to present a UI that is different from but visually identical to the UI that users must rely on to make valid trust decisions in a default/common scenario. A trust decision is defined as any time the user takes an action believing some information is being presented by a particular entity-either the system or some specific local or remote source. The target cannot perform normal operations due to an attack. The response to an attack is roughly the same magnitude as the size of the attack. The target returns to the normal level of functionality shortly after the attack is finished. The exact definition of \"shortly\" should be evaluated for each product. For example, a server is unresponsive while an attacker is constantly sending a stream of packets across a network, and the server returns to normal a few seconds after the packet stream stops. temporar y DoS with amplification temporar y DoS with amplification A temporary DoS with amplification is a situation where the following criteria are met: The target cannot perform normal operations due to an attack. The response to an attack is magnitudes beyond the size of the attack. The target returns to the normal level of functionality after the attack is finished, but it takes some time (perhaps a few minutes). For example, if you can send a malicious 10-byte packet and cause a 2048k response on the network, you are DoSing the bandwidth by amplifying our attack effort. \n permanent DoS permanent DoS A permanent DoS is one that requires an administrator to start, restart, or reinstall all or parts of the system. Any vulnerability that automatically restarts the system is also a permanent DoS.  All asymmetric keys should have a maximum five-year lifetime, recommended one-year lifetime. All symmetric keys should have a maximum three-year lifetime; recommended one-year lifetime. You should provide a mechanism or have a process for replacing keys to achieve the limited active lifetime. After the end of its active lifetime, a key should not be used to produce new data (for example, for encryption or signing), but may still be used to read data (for example, for decryption or verification). All products and services should use cryptographically secure random number generators when randomness is required. Use of the dual elliptic curve random number generator (\"DUAL_EC_DRBG\") algorithm is not recommended. \n CNG On the Windows platform, Microsoft recommends using the crypto APIs built into the operating system. On other platforms, developers may choose to evaluate non-platform crypto libraries for use. In general, platform crypto libraries will be updated more frequently since they ship as part of an operating system as opposed to being bundled with an application. Any usage decision regarding platform vs non-platform crypto should be guided by the following requirements: Key Derivation Functions SSL/TLS/DTLS: Use APIs defined under the System.Net namespace (for example, HttpWebRequest). Key derivation is the process of deriving cryptographic key material from a shared secret or a existing cryptographic key. Products should use recommended key derivation functions. Deriving keys from user-chosen passwords, or hashing passwords for storage in an authentication system is a special case not covered by this guidance; developers should consult an expert. The following standards specify KDF functions recommended for use: NIST SP 800-108: Recommendation For Key Derivation Using Pseudorandom Functions. In particular, the KDF in counter mode, with HMAC as a pseudorandom function NIST SP 800-56A (Revision 2): Recommendation for Pair-Wise Key Establishment Schemes Using Discrete Logarithm Cryptography. In particular, the \"Single-Step Key Derivation Function\" in Section 5.8.1 is recommended. To derive keys from existing keys, use the BCryptKeyDerivation API with one of the algorithms: Revocation status. Usage (for example, \"Server Authentication\" for servers, \"Client Authentication\" for clients). Trust chain. Certificates should chain to a root certification authority (CA) that is trusted by the platform or explicitly configured by the administrator. If any of these verification tests fail, the product should terminate the connection with the entity. Clients that trust \"self-signed\" certificates (for example, a mail client connecting to an Exchange server in a default configuration) may ignore certificate verification checks. However, self-signed certificates do not inherently convey trust, support revocation, or support key renewal. You should only trust selfsigned certificates if you have obtained them from another trusted source (for example, a trusted entity that provides the certificate over an authenticated and integrity-protected transport). Products should use the SHA-2 family of hash algorithms (SHA256, SHA384, and SHA512). Truncation of cryptographic hashes for security purposes to less than 128 bits is not recommended. A message authentication code (MAC) is a piece of information attached to a message that allows its recipient to verify both the authenticity of the sender and the integrity of the message using a secret key. The use of either a hash-based MAC (HMAC) or block-cipher-based MAC is recommended as long as all underlying hash or symmetric encryption algorithms are also recommended for use; currently this includes the HMAC-SHA2 functions (HMAC-SHA256, HMAC-SHA384 and HMAC-SHA512). Truncation of HMACs to less than 128 bits is not recommended. You should provide a mechanism for replacing cryptographic keys as needed. Keys should be replaced once they have reached the end of their active lifetime or if the cryptographic key is compromised. Whenever you renew a certificate, you should renew it with a new key. Products using cryptographic algorithms to protect data should include enough metadata along with that content to support migrating to different algorithms in the future. This should include the algorithm used, key sizes, initialization vectors, and padding modes. For more information on Cryptographic Agility, see Cryptographic Agility on MSDN. Where available, products should use established, platform-provided cryptographic protocols rather than re-implementing them. This includes signing formats (e.g. use a standard, existing format). Symmetric stream ciphers such as RC4 should not be used. Instead of symmetric stream ciphers, Encrypting Sensitive Data prior to Storage products should use a block cipher, specifically AES with a key length of at least 128 bits. Do not report cryptographic operation failures to end-users. When returning an error to a remote caller (e.g. web client, or client in a client-server scenario), use a generic error message only. Avoid providing any unnecessary information, such as directly reporting out-of-range or invalid length errors. Log verbose errors on the server only, and only if verbose logging is enabled. Additional security review is highly recommended for any design incorporating the following: A new protocol that is primarily focused on security (such as an authentication or authorization protocol) A new protocol that uses cryptography in a novel or non-standard way o Example considerations include: Will a product that implements the protocol call any crypto APIs or methods as part of the protocol implementation? Does the protocol depend on any other protocol used for authentication or authorization? Will the protocol define storage formats for cryptographic elements, such as keys? Self-signed certificates are not recommended for production environments. Use of a self-signed certificate, like use of a raw cryptographic key, does not inherently provide users or administrators any basis for making a trust decision. In contrast, use of a certificate rooted in a trusted certificate authority makes clear the basis for relying on the associated private key and enables revocation and updates in the event of a security failure. \n DPAPI/DPAPI-NG For data that needs to be persisted across system reboots: Microsoft is deeply committed to making the online world safer for everyone. Our company's cybersecurity strategies have evolved from the unique visibility we have into the rapidly evolving cyberthreat landscape. \n Cybersecurity threat actors and motivations Cybersecurity is a shared responsibility, which impacts us all. Today, a single breach, physical or virtual, can cause millions of dollars of damage to an organization and potentially billions in financial losses to the global economy. Every day, we see reports of cybercriminals targeting businesses and individuals for financial gain or socially-motivated purposes. Add to these threats those by nation-state actors seeking to disrupt operations, conduct espionage, or generally undermine trust. In this brief, we share the state of online security, threat actors and the sophisticated tactics they employ to advance their goals, and how Microsoft's Cyber Defense Operations Center combats these threats and helps customers protect their sensitive applications and data. Innovation in the attack space across people, places, and processes is a necessary and continual investment we all need to make, as adversaries continue to evolve in both determination and sophistication. In response to increased investments in defense strategies by many organizations, attackers are adapting and improving tactics at breakneck speed. Fortunately, cyberdefenders like Microsoft's global information security teams are also innovating and disrupting long-reliable attack methods with ongoing, advanced training and modern security technologies, tools, and processes. The Microsoft Cyber Defense Operations Center (CDOC) is one example of the more than $1 billion we invest each year on security, data protection, and risk management. The CDOC brings together cybersecurity specialists and data scientists in a 24x7 facility to combat threats in real-time. We are connected to more than 3,500 security professionals globally across our product development teams, information security groups, and legal teams to protect our cloud infrastructure and services, products and devices, and internal resources. Microsoft has invested more than $15 billion in our cloud infrastructure, with over 90 percent of Fortune 500 companies using the Microsoft cloud. Today, we own and operate one of the world's largest cloud footprints with more than 100 geo-distributed datacenters, 200 cloud services, millions of devices, and a billion customers around the globe. The first step to protecting people, devices, data, and critical infrastructure is to understand the different types of threat actors and their motivations. Cybercriminals Cybercriminals span several subcategories though they often share common motivations-financial, intelligence, and/or social or political gain. Their approach is usually direct-by infiltrating a financialdata system, skimming micro-amounts too small to detect and exiting before being discovered. Maintaining a persistent, clandestine presence is critical to meeting their objective. Their approach may be an intrusion that diverts a large financial payout through a labyrinth of accounts to evade tracking and intervention. At times, the goal is to steal intellectual property that the target possesses so the cybercriminal acts as an intermediary to deliver a product design, software source code, or other proprietary information that has value to a specific entity. More than half of these activities are perpetrated by organized criminal groups. Nation-state actors Nation-state actors work for a government to disrupt or compromise targeted governments, organizations, or individuals to gain access to valuable data or intelligence. They are engaged in international affairs to influence and drive an outcome that may benefit a country or countries. A nation-state actor's objective is to disrupt operations, conduct espionage against corporations, steal secrets from other governments, or otherwise undermine trust in institutions. They work with large resources at their disposal and without fear of legal retribution, with a toolkit that spans from simple to highly complex. Nation-state actors can attract some of the most sophisticated cyberhacking talent and may advance their tools to the point of weaponization. Their intrusion approach often involves an advanced persistent threat using supercomputing power to brute-force break credentials through millions of attempts at arriving at the correct password. They may also use hyper-targeted phishing attacks to attract an insider into revealing their credentials. \n Insider Insider threats are particularly challenging due to the unpredictability of human behavior. The motivation for an insider maybe opportunistic and for financial gain. However, there are multiple causes for potential insider threats, spanning from simple carelessness to sophisticated schemes. Many data breaches resulting from insider threats are completely unintentional due to accidental or negligent activity that puts an organization at risk without being aware of the vulnerability. \n Hacktivists Hacktivists focus on political and/or sociallymotivated attacks. They strive to be visible and recognized in the news to draw attention to themselves and their cause. Their tactics include distributed denial-of-service (DDoS) attacks, vulnerability exploits or defacing an online presence. A connection to a social or political issue can make Threat actor techniques any company or organization a target. Social media enables hacktivists to quickly evangelize their cause and recruit others to participate. Adversaries are skilled at finding ways to penetrate an organization's network despite the protections in place using various sophisticated techniques. Several tactics have been around since the early days of the Internet, though others reflect the creativity and increasing sophistication of today's adversaries. \n Social engineering Social engineering is a broad term for an attack that tricks users into acting or divulging information they would otherwise not do. Social engineering plays on the good intentions of most people and their willingness to be helpful, to avoid problems, to trust familiar sources, or to potentially gain a reward. Other attack vectors can fall under the umbrella of social engineering, but the following are some of the attributes that make social engineering tactics easier to recognize and defend against: Malware has been with us since the dawn of computing. Today, we're seeing a strong up-tick in ransomware and malicious code specifically intended to encrypt devices and data. Cybercriminals then demand payment in cryptocurrency for the keys to unlock and return control to the victim. This can happen at an individual level to your computer and data files, or now more frequently, to an entire enterprise. The use of ransomware is particularly pronounced in the healthcare field, as the life-ordeath consequences these organizations face make them highly sensitive to network downtime. Supply chain inser tion Supply chain inser tion is an example of a creative approach to injecting malware into a network. For \n Phishing emails Phishing emails are an effective tool because they play against the weakest link in the security chain-everyday users who don't think about network security as top-of-mind. A phishing campaign may invite or frighten a user to inadvertently share their credentials by tricking them into clicking a link they believe is a legitimate site or downloading a file that contains malicious code. Phishing emails used to be poorly written and easy to recognize. Today, adversaries have become adept at mimicking legitimate emails and landing sites that are difficult to identify as fraudulent. Identity spoofing Identity spoofing involves an adversary masquerading as another legitimate user by falsifying the information presented to an application or network resource. An example is an email that arrives seemingly bearing the address of a colleague requesting action, but the address is hiding the real source of the email sender. Similarly, a URL can be spoofed to appear as a legitimate site, but the actual IP address is actually pointing to a cybercriminal's site. example, by hijacking an application update process, an adversary circumvents anti-malware tools and protections. We are seeing this technique become more common and this threat will continue to grow until more comprehensive security protections are infused into software by application developers. Man-in-the-middle Man-in-the-middleattacks involve an adversary inserting themselves between a user and a resource they are accessing, thereby intercepting critical information such as a user's login credentials. For example, a cybercriminal in a coffee shop may employ key-logging software to capture a users' domain credentials as they join the wifi network. The threat actor can then gain access to the user's sensitive information, such as banking and personal information that they can use or sell on the dark web. Distributed Denial of Ser vice (DDoS) Distributed Denial of Ser vice (DDoS)attacks have been around more than a decade and are massive attacks are becoming more common with the rapid growth of the Internet of Things (IoT). When using this technique, an adversary overwhelms a site by bombarding it with malicious traffic that displaces legitimate queries. Previously planted malware is often used to hijack an IoT device such as a webcam or smart thermostat. In a DDoS attack, incoming traffic from different sources flood a network with numerous requests. This overwhelms servers and denies access from legitimate requests. Many attacks involve forging of IP sender addresses (IP address spoofing) also, so that the location of the attacking machines cannot easily be identified and defeated. Often a denial of service attack is used to cover or distract from a more deceptive effort to penetrate an organization. In most cases, the objective of the adversary is to gain access to a network using compromised credentials, then move laterally across the network to gain access to more \"powerful\" credentials that are the keys to the most sensitive and valuable information within the organization. \n The militarization of cyberspace The Microsoft cybersecurity posture \n PROTECT Protect The growing possibility of cyberwarfare is one of the leading concerns among governments and citizens today. It involves nation-states using and targeting computers and networks in warfare. Both offensive and defensive operations are used to conduct cyberattacks, espionage and sabotage. Nationstates have been developing their capabilities and engaged in cyberwarfare either as aggressors, defendants, or both for many years. New threat tools and tactics developed through advanced military investments may also be breached and cyberthreats can be shared online and weaponized by cybercriminals for further use. While security has always been a priority for Microsoft, we recognize that the digital world requires continuous advances in our commitment in how we protect, detect, and respond to cybersecurity threats. These three commitments define our approach to cyber defense and serve as a useful framework for our discussion of Microsoft's cyber defense strategies and capabilities. Microsoft's first commitment is to protect the computing environment used by our customers and employees to ensure the resiliency of our cloud infrastructure and services, products, devices, and the company's internal corporate resources from determined adversaries. The CDOC teams' protection measures span across all endpoints, from sensors and datacenters to identities and software-as-a-service (SaaS) applications. Defenseindepth-applying controls at multiple layers with overlapping safeguards and risk mitigation strategies-is a best-practice across the industry and it's the approach we take to protect our valuable customer and corporate assets. Microsoft's protection tactics include: Extensive monitoring and controls over the physical environment of our global datacenters including cameras, personnel screening, fences and barriers, and multiple identification methods for physical access. Software-defined networks that protect our cloud infrastructure from intrusions and DDoS attacks. Multi-factor authentication is employed across our infrastructure to control identity and access management. It ensures that critical resources and data are protected by at least two of the following: Something you know (password or PIN) Something you are (biometrics) Non-persistent administration employs just-in-time (JIT) and just-enough administrator (JEA) privileges to engineering staff who manage infrastructure and services. This provides a unique set of credentials for elevated access that automatically expires after a predesignated duration. Proper hygiene is rigorously maintained through upto-date, anti-malware software and adherence to strict patching and configuration management. Microsoft Malware Protection Center's team of researchers identify, reverse engineer, and develop malware signatures and then deploy them across our infrastructure for advanced detection and defense. These signatures are distributed to our responders, customers and the industry through Windows Updates and notifications to protect their devices. Microsoft Security Development Lifecycle (SDL) is a software development process that helps developers build more secure software and address security compliance requirements while reducing development cost. The SDL is used to harden all applications, online services and products, and to routinely validate its effectiveness through penetration testing and vulnerability scanning. Threat modeling and attack surface analysis ensures that potential threats are assessed, exposed aspects of the service are evaluated, and the attack surface is minimized by restricting services or eliminating unnecessary functions. Classifying data according to its sensitivity and taking the appropriate measures to protect it, including encryption in transit and at rest, and enforcing the principle of least-privilege access provides additional protection. • Awareness training that fosters a trust relationship between the user and the security team to develop an environment where users will report incidents and anomalies without fear of repercussion. \n Something you have (smartphone) Having a rich set of controls and a defense-in-depth strategy helps ensure that should any one area fail, there are compensating controls in other areas to help maintain the security and privacy of our customers, cloud services, and our own infrastructure. However, no environment is truly impenetrable, as people will make errors and determined adversaries will continue to look for vulnerabilities and exploit them. The significant investments we continue to make in these protection layers and baseline analysis enables us to rapidly detect when abnormal activity is present. \n DETECT \n Detect \n RESPOND Respond The CDOC teams employ automated software, machine learning, behavioral analysis, and forensic techniques to create an intelligent security graph of our environment. This signal is enriched with contextual metadata and behavioral models generated from sources such as Active Directory, asset and configuration management systems, and event logs. Our extensive investments in security analytics build rich behavioral profiles and predictive models that allow us to \"connect the dots\" and identify advanced threats that might otherwise have gone undetected, then counter with strong containment and coordinated remediation activities. Microsoft also employs custom-developed security software, along with industryleading tools and machine learning. Our threat intelligence is continually evolving, with automated data-enrichment to more rapidly detect malicious activity and report with high fidelity. Vulnerability scans are performed regularly to test and refine the effectiveness of protective measures. The breadth of Microsoft's investment in its security ecosystem and the variety of signals monitored by the CDOC teams provide a more comprehensive threat view than can be achieved by most service providers. Microsoft's detection tactics include: Monitoring network and physical environments 24x7x365 for potential cybersecurity events. Behavior profiling is based on usage patterns and an understanding of unique threats to our services. Identity and behavioral analytics are developed to highlight abnormal activity. Machine learning software tools and techniques are routinely used to discover and flag irregularities. Advanced analytical tools and processes are deployed to further identify anomalous activity and innovative correlation capabilities. This enables highlycontextualized detections to be created from the enormous volumes of data in near real-time. Automated software-based processes that are continuously audited and evolved for increased effectiveness. Data scientists and security experts routinely work side-by-side to address escalated events that exhibit unusual characteristics requiring further analysis of targets. They can then determine potential response and remediation efforts. \n Cyberdefense for our customers When Microsoft detects abnormal activity in our systems, it triggers our response teams to engage and quickly respond with precise force. Notifications from software-based detection systems flow through our automated response systems using risk-based algorithms to flag events requiring intervention from our response team. Mean-Time-to-Mitigate is paramount and our automation system provides responders with relevant, actionable information that accelerates triage, mitigation, and recovery. To manage security incidents at such a massive scale, we deploy a tiered system to efficiently assign response tasks to the right resource and facilitate a rational escalation path. Microsoft's response tactics include: Automated response systems use risk-based algorithms to flag events requiring human intervention. Automated response systems use risk-based algorithms to flag events requiring human intervention. Well-defined, documented, and scalable incident response processes within a continuous improvement model helps to keep us ahead of adversaries by making these available to all responders. Subject matter expertise across our teams, in multiple security areas, provides a diverse skill set for addressing incidents. Security expertise in incident response, forensics, and intrusion analysis; and a deep understanding of the platforms, services, and applications operating in our cloud datacenters. Wide enterprise searching across cloud, hybrid and on-premises data and systems to determine the scope of an incident. Deep forensic analysis for major threats are performed by specialists to understand incidents and to aid in their containment and eradication. • Microsoft's security software tools, automation and hyper-scale cloud infrastructure enable our security experts to reduce the time to detect, investigate, analyze, respond, and recover from cyberattacks. Penetration testing is employed across all Microsoft products and services through ongoing Red Team/Blue Team exercises to unearth vulnerabilities before a real adversary can leverage those weak points for an attack. We   SC EN A RIO N UM B ER SC EN A RIO N UM B ER AT TA C K AT TA C K O VERVIEW O VERVIEW VIO L AT ES T RA DIT IO N A L VIO L AT ES T RA DIT IO N A L T EC H N O LO GIC A L N OT IO N T EC H N O LO GIC A L N OT IO N O F O F A C C ESS/ A UT H O RIZ AT IO N ? A C C ESS/ A UT H O RIZ AT IO N ? \n [ 6 ] 6 Reinforcing Adversarial Robustness using Model Confidence Induced by Adversarial Training[19]: The authors propose Highly Confident Near Neighbor (HCNN), a framework that combines confidence information and nearest neighbor search, to reinforce adversarial robustness of a base model. This can help distinguish between right and wrong model predictions in a neighborhood of a point sampled from the underlying training distribution. \n can be used to harden DNN models by detecting adversarial examples. It reduces the search space available to an adversary by coalescing samples that correspond to many different feature vectors in the original space into a single sample. By comparing a DNN model's prediction on the original input with that on the squeezed input, feature squeezing can help detect adversarial examples. If the original and squeezed examples produce substantially different outputs from the model, the input is likely to be adversarial. By measuring the disagreement among predictions and selecting a threshold value, system can output the correct prediction for legitimate examples and rejects adversarial inputs. d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s Se v e r i t y Se v e r i t y \n M i t i g a t i o n s M i t i g a t i o n s T r a d i t i o n a l P a r a l l e l s T r a d i t i o n a l P a r a l l e l s Se v e r i t y Se v e r i t y \n # 7 7 Adversarial Example in the Physical domain (bits->atoms) Ex a m p l e s Ex a m p l e s \n rd Strong client<->server mutual authentication and access control to model interfaces Takedown of the offending accounts. \n Vectors and Machine Learning Techniques A. Feature Vectors and Machine Learning Techniquesunsupervised machine learning techniques, and study how much data is needed to train their models. \n -dependent noise: Class-dependent noise refers to the noise that depends on the true class of the instances. In this type of noise, the probability of mislabeling in class SBR is p and the probability of mislabeling in class NSBR is p . We introduce class-dependent noise in our data set by flipping each entry in the data set for which the true label is SBR with probability p . Analogously, we flip the class label of NSBR instances with probability p . -class noise: Single-class noise is a special case of class-dependent noise, where p = 0 and p > 0. Note that for class-independent noise we have p = p = p . the impact of different noise types and levels in the training of SBR classifiers. In our experiments we set 25% of the data set as test data, 10% as validation and 65% as training data. \n Noise: class-independentWe compare the performance of our three classifiers for the case where the training set is corrupted by a classindependent noise. We measure the AUC for each model trained with different levels of p in the training data.br \n brFig. 1 . 1 Fig.1. Variation of AUC-ROC in class-independent noise. For a noise level p =0.5 the classifier acts like a random classifier, i.e. AUC≈0.5. But we can observe that for lower noise levels (p ≤0.30), the logistic regression learner presents a better performance compared to the other two models. However, for 0.35≤ p ≤0.45 naive Bayes learner presents better AUCROC metrics.br \n this paper is twofold. \n Figure 1 : 1 : 11 Figure 1: Security Protocol Support by OS Version Figure 1: Security Protocol Support by OS Version \n Figure 1 : 1 Figure 1: Default TLS Version selection and Certificate Binding Functionality https://secure.contoso.com directs your customers to a service endpoint supporting only TLS 1.2 and above. \n Figure 2 : 2 Figure 2: Disable Legacy TLS feature enforcing minimum TLS version for a selected certificate, Secure.contoso.com. \n Option # 2 :Option # 3 : 23 PowerShell (Available April 2020) Option #2: PowerShell (Available April 2020) [Microsoft.Web.Administration.SslFlags]::DisableLegacyTLS $Sni = [Microsoft.Web.Administration.SslFlags]::Sni $Sni\\_CCS = [Microsoft.Web.Administration.SslFlags]::Sni + [Microsoft.Web.Administration.SslFlags]::CentralCertStore $CCS = [Microsoft.Web.Administration.SslFlags]::CentralCertStore $DisableLegacyTLS = [Microsoft.Web.Administration.SslFlags]::DisableLegacyTLS $storeLocation = \"Cert:\\LocalMachine\\My\" $BindingInformation = \"\\*:443:\" $siteName = \"contoso\" $Thumbprint = $certificate.ThumbPrint New-IISSite $siteName \"$env:systemdrive\\inetpub\\wwwroot\" \"\\*:443:secure.contoso.com\" https $certificate.Thumbprint $DisableLegacyTLS $storeLocation -passthru New-IISSiteBinding -Name \"Default Web Site\" -BindingInformation $BindingInformation -CertificateThumbPrint $certificate.Thumbprint -Protocol https -SslFlag $DisableLegacyTLS, $CCS -Force -verbose In PowerShell you can reference SSL flags like this: It's convenient to create shorter named variables for them: An example of creating a site binding to a new site and disabling legacy TLS: New-IISSite with Sslflag DisableLegacyTLS property value: An example of adding a site binding to an existing site and disabling legacy TLS: Additionally, one can troubleshoot and test this feature with Netsh: Adding a new binding: C++ HTTP.sys APIs (Available Now) Option #3: C++ HTTP.sys APIs (Available Now) Next steps for TLS version enforcement netsh http add sslcert <regular parameters> disablelegacytls=enable Updating an existing binding: netsh http update sslcert <regular parameters> disablelegacytls=enable Check whether it is set on a binding: netsh http show sslcert <regular parameters> Watch for Disable Legacy TLS Versions : Set/Not Set Along with Disable Legacy TLS, the following additions have been made to HTTP.sys: HTTP_SERVICE_CONFIG_SSL_PARAM.DefaultFlags now supports the following new values: \n Microsoft is committed to providing an unprecedented level of transparency through the Government Security Program (GSP), aimed at helping customers gain confidence in the integrity and assurance of the products and services on which they rely. To support those efforts, Microsoft has opened five Transparency Centers throughout the world: provide GSP participants with an opportunity to visit a secure facility to conduct deep levels of source code inspection and analysis. Participants have access to source code and an environment for in-depth inspection with advanced tools. In addition, participants can compile components and use additional tools for a deeper understanding of the source code. Currently, the Transparency Centers provide source code for products such as Windows, Windows Ser ver, Office, Exchange Ser ver, SQL Ser ver, and SharePoint Windows, Windows Ser ver, Office, Exchange Ser ver, SQL Ser ver, and SharePoint Ser ver Ser ver.Each visit to a Transparency Center is tailored to the unique goals of what an agency is looking to accomplish. Visits can last from one day to two weeks, depending on an agency's needs, and are scheduled based on facility availability. Face-to-face or teleconference exchanges with Microsoft engineers may also be available and can be beneficial during Transparency Center visits for agencies who have also joined the Technical Data portion of the program.The environment and tools for source code evaluation include:Private network with one server rack and eight clients OpenGrok open source search and cross-reference engine SysInternals Example use cases of Transparency Centers Contact Us Tools provided by the participant and approved by Microsoft HexRays IDA Disassembler and Decompiler PowerShell v4.0 MSDN Documentation Evaluation of Cryptography Next Generation implementation and preparation for national cryptography implementation Review of SSL and TCP/IP implementation Inspection of the source of random number generators Walkthrough of Microsoft's build process Assessment of individual binaries to compare against shipped binaries Contact your local Microsoft representative to learn more about the Government Security Program Online Access to Source Code 2/19/2022 • 2 minutes to read Example cases of Online Source: \n Check code for specific assurance questions Examine specific function calls in source code As a reference to understand how a feature or function works Contact your local Microsoft representative to learn more about the Government Security Program Technical Data 2/19/2022 • 2 minutes to read \n Written materials Direct dialogue with Microsoft engineers and security experts Early access to documentation about Microsoft's products and services In-person exchanges with Microsoft engineers and security experts are sometimes scheduled in Microsoft facilities and cover security questions about Microsoft products and services. These meetings, commonly referred to as technical trips, provide deep technical conversations about topics of interest. To plan and facilitate these, the Microsoft local GSP representative works closely with the agency to understand what it is looking to accomplish and to develop a customized agenda. Technical trips are accommodated based on Microsoft engineering team availability. One of the frequent uses of this offering is to review security content related to Microsoft's cloud services. This includes access to the Microsoft Service Trust Platform where we share important third-party security and compliance reports, as well as security details about Microsoft's cloud services. Technical documentation and written materials needed to understand a security question Early releases of white papers Common Criteria evaluation documents for Windows Engagements with engineers in many different formats including Technical Adoption Programs (TAP) Cloud service security, privacy and compliance documentation such as audit reports and risk assessments Technical trip sample topics: Azure Incident Response Process Bitlocker Encryption Windows Telemetry Cyber Threat Intelligence Contact your local Microsoft representative to learn more about the Government Security Program SDL Security Bug Bar ( Sample ) \n A UT H EN T IC AT ED VS. A UT H EN T IC AT ED VS. A N O N Y M O US AT TA C K A N O N Y M O US AT TA C K DEFA ULT / C O M M O N VS. DEFA ULT / C O M \n from a shared secret (the output of a key agreement) use the BCryptDeriveKey API with one of the following algorithms: SSL, TLS, or DTLS should fully verify the X.509 certificates of the entities they connect to. This includes verification of the certificates': Domain name.Validity dates (both beginning and expiration dates). \n are often asked what tools and processes our customers can adopt for their own environment and how Microsoft might help in their implementation. Microsoft has consolidated many of the cyberdefense products and services we use in the CDOC into a range Best practices to protect your environmentI N V E S T I N Y O U R I N V E S T I N Y O U R P L AT F O R M P L AT F O R M I N V E S T I N Y O U R I N V E S T I N Y O U R I N S T R U M E N TAT I O N I N S T R U M E N TAT I O N I N V E S T I N Y O U R I N V E S T I N Y O U \n \n \n \n \n \n \n \n \n \n How much knowledge is required to mount this attack -blackbox or whitebox? In Blackbox style attacks., the attacker does NOT have direct access to the training data, no knowledge of the ML algorithm used and no access to the source code of the model. The attacker only queries the model and observes the response. In a whitebox style attack the attacker has knowledge of either ML algorithm or access to the model source code.3. Commentary on if the attacker is violating traditional technological notion of access/authorization. 2. Backdoor ML 1 Perturbation attacks In perturbation style Malicious ML provider Integrity Poisoning attacks The goal of the Integrity Yes Image: Noise is In a medical dataset backdoors algorithm to attacks, the attacker attacker is to added to an X-ray where the goal is to activate with a specific stealthily modifies contaminate the image, which makes predict the dosage of the query to get a machine model trigger the predictions go anticoagulant drug desired response generated in the from normal scan to Warfarin using training phase, so abnormal [1] demographic 11 Exploit Software that predictions on Attacker uses traditional Yes [Blackbox] information, etc. Dependencies new data will be software exploits like buffer Text translation: Researchers overflow to confuse/control modified in the Specific introduced malicious testing phase ML systems characters are samples at 8% Targeted: In manipulated to poisoning rate, which targeted result in incorrect changed dosage by poisoning Unintended Failures Summary attacks, the translation. The 75.06% for half of attack can patients[4][Blackbox] SC EN A RIO # SC EN A RIO # FA IL URE FA IL URE attacker wants to misclassify O VERVIEW suppress specific O VERVIEW word or can even 1 Perturbation attack specific examples Attacker modifies the query No remove the word 12 to get appropriate response Reward Hacking Reinforcement Learning (RL) systems act in unintended ways because of completely[2] [Blackbox and Indiscriminate: 2 Poisoning attack Attacker contaminates the mismatch between stated reward and Whitebox] No true reward training phase of ML Speech: 13 systems to get intended result Side Effects RL system disrupts the environment as Researchers showed how it tries to attain its goal given a speech 3 Model Inversion Attacker recovers the secret No waveform, 14 features used in the model Distributional shifts The system is tested in one kind of another by through careful queries environment, but is unable to adapt to waveform can be 4 5 15 16 Membership Inference Model Stealing Natural Adversarial Examples Attacker can infer if a given data record was part of the model's training dataset or not Attacker is able to recover the model through carefully-crafted queries changes in other kinds of environment No No Without attacker perturbations, the ML system fails owing to hard negative mining Common Corruption exactly replicated Model Inversion The private features Confidentiality; Researchers were but transcribes used in machine able to recover into a totally learning models can private training data different text[3] be recovered used to train the [Whitebox but algorithm[6] The may be extended authors were able to to blackbox] reconstruct faces, by The system is not able to handle just the name and common corruptions and access to the model perturbations such as tilting, zooming, to the point where or noisy images. Mechanical turks 6 Reprogramming ML system Repurpose the ML system No could use the photo 17 to perform an activity it was Incomplete Testing The ML system is not tested in the to identify an not programmed for realistic conditions that it is meant to individual from aline- operate in. up with 95% 7 Details on Intentionally-Motivated Failures Adversarial Example in Physical Domain Attacker brings adversarial examples into physical domain to subvertML system e.g: 3d printing special eyewear to fool facial recognition system SC EN A RIO # SC EN A RIO # AT TA C K C L A SS AT TA C K C L A SS DESC RIP T IO N DESC RIP T IO N T Y P E O F T Y P E O F C O M P RO M ISE C O M P RO M ISE accuracy. The authors No were also able to extract specific information. [Whitebox and SC EN A RIO SC EN A RIO Blackbox][12] 8 Malicious ML provider recovering training data Membership The attacker can Malicious ML provider can Confidentiality Inference attack determine whether a query the model used by given data record customer and recover was part of the customer's training data model's training Researchers were Yes able to predict a patient's main procedure(e.g: Surgery the patient dataset or not went through) based 9 Attacking the ML supply Attacker compromises the Yes on the attributes (e.g: chain ML models as it is being age,gender, hospital) downloaded for use [7][Blackbox] \n and Mitigations in this Document Related Threats and Mitigations in this Document Example Attacks Example Attacks If confidence levels of your model output suddenly drop, can you find out how/why, as well as the data that caused it?Have you defined a well-formed input for your model? What are you doing to ensure inputs meet this format and what do you do if they don't? If your outputs are wrong but not causing errors to be reported, how would you know? Do you know if your training algorithms are resilient to adversarial inputs on a mathematical level? How do you recover from adversarial contamination of your training data? -Can you isolate/quarantine adversarial content and re-train impacted models? -Can you roll back/recover to a model of a prior version for re-training? Are you using Reinforcement Learning on uncurated public content? Start thinking about the lineage of your data -were you to find a problem, could you track it to its introduction into the dataset? If not, is that a problem? Know where your training data comes from and identify statistical norms in order to begin understanding what anomalies look like -What elements of your training data are vulnerable to outside influence?Related ThreatsIdentify actions your model(s) or product/service could take which can cause customer harm online or in the physical domain \n Summary Summary Questions to Ask in a Security Review Questions to Ask in a Security Review Related Threats and Mitigations in this Document Related Threats and Mitigations in this Document -Who can contribute to the data sets you're training from? Example Attacks Example Attacks -How would you you attack your sources of training data to harm a competitor? Adversarial Perturbation (all variants) Data Poisoning (all variants) Forcing benign emails to be classified as spam or causing a malicious example to go undetected Attacker-crafted inputs that reduce the confidence level of correct classification, especially in high- consequence scenarios Attacker injects noise randomly into the source data being classified to reduce the likelihood of the Identify all sources of AI/ML dependencies as well as frontend correct classification being used in the future, effectively dumbing down the model presentation layers in your data/model supply chain Contamination of training data to force the misclassification of select data points, resulting in specific actions being taken or omitted by a system Can an attacker cause reputational damage or PR backlash to your product by forcing it to carry out specific actions? How do you handle properly formatted but overtly biased data, such as from trolls? For each way to interact with or query your model is exposed, can that method be interrogated to disclose training data or model functionality? Membership Inference Model Inversion Model Stealing Left unmitigated, attacks on AI/ML systems can find their way over to the physical world. Any scenario which can be twisted to psychologically or physically harm users is a catastrophic risk to your product/service. This extends to any sensitive data about your customers used for training and design choices that can leak those private data points.Do you train with adversarial examples? What impact do they have on your model output in the physical domain?What does trolling look like to your product/service? How can you detect and respond to it? What would it take to get your model to return a result that tricks your service into denying access to legitimate users? What is the impact of your model being copied/stolen? Can your model be used to infer membership of an individual person in a particular group, or simply in the training data? \n Summary Summary Questions to Ask in a Security Review Questions to Ask in a Security Review Related Threats and Mitigations in this Document Related Threats and Mitigations in this Document Reconstruction and extraction of training data by repeatedly querying the model for maximum confidence results Duplication of the model itself by exhaustive query/response matching Querying the model in a way that reveals a specific element of private data was included in the training set Neural Net Reprogramming Adversarial Examples in the physical domain Malicious ML Providers Recovering Training Data Attacking the ML Supply Chain Backdoored Model Compromised ML-specific dependencies Self-driving car being tricked to ignore stop signs/traffic lights Example Attacks Example Attacks Conversational bots manipulated to troll benign users Malicious MLaaS provider trojans your model with a specific bypass Adversary customer finds vulnerability in common OSS dependency you use, uploads crafted training data payload to compromise your service Unscrupulous partner uses facial recognition APIs and creates a presentation layer over your service to produce Deep Fakes. AI/ML-specific Threats and their Mitigations #1: Adversarial Perturbation Description Description What kind of telemetry do you need to prove the trustworthiness of your model output to your customers? Identify all 3 party dependencies in your ML/Training data supply chain -not just open source software, but data providers as well Variant #1a: Targeted misclassification Examples Examples Many attacks in AI and Machine Learning begin with legitimate access to APIs which are surfaced to provide query access to a model. Because of the rich sources of data and rich user experiences involved here, authenticated but \"inappropriate\" (there's a gray area here) 3 -party access to your models is a risk because of the ability to act as a presentation layer above a Microsoft-provided service.rd Which customers/partners are authenticated to access your model or service APIs? -Can they act as a presentation layer on top of your service? -Can you revoke their access promptly in case of compromise? -What is your recovery strategy in the event of malicious use of your service or dependencies? Can a 3 party build a façade around your model to re-purpose it and harm Microsoft or its customers? rd Do customers provide training data to you directly? -How do you secure that data? -What if it's malicious and your service is the target? What does a false-positive look like here? What is the impact of a false-negative? Can you track and measure deviation of True Positive vs False Positive rates across multiple models? rd -Why are you using them and how do you verify their trustworthiness? Are you using pre-built models from 3 parties or submitting training data to 3 party MLaaS providers? rd rdInventory news stories about attacks on similar products/services. Understanding that many AI/ML threats transfer between model types, what impact would these attacks have on your own products? \n Identifying security bug reports (SBRs) is a vital step in the software development life-cycle. In supervised machine learning based approaches, it is usual to assume that entire bug reports are available for training and that their labels are noise free. To the best of our knowledge, this is the first study to show that accurate label prediction is possible for SBRs even when solely the title is available and in the presence of label noise. 2/19/2022 • 22 minutes to read Mayana Pereira Scott Christiansen CELA Data Science Customer Security and Trust Microsoft Microsoft Abstract - I. INTRODUCTION \n in software built on top of Microsoft operating systems, following up with details on product changes and new features delivered by Microsoft to protect your own customers and online services. It is intended to be used as a starting point for building a migration plan to a TLS 1.2+ network environment. While the solutions discussed here may carry over and help with removing TLS 1.0 usage in non-Microsoft operating systems or crypto libraries, they are not a focus of this document. Solving the TLS 1.0 Problem, 2nd Edition 2/19/2022 • 11 minutes to read By Andrew Marshall Principal Security Program Manager Microsoft Corporation Executive Summary This document presents the latest guidance on rapidly identifying and removing Transport Layer Security (TLS) protocol version 1.0 dependencies 1] John Anvik, Lyndon Hiew, and Gail C Murphy. Who should fix this bug? In Proceedings of the 28th international conference on Software engineering, pages 361-370. ACM, 2006. [2]  Diksha Behl, Sahil Handa, and Anuja Arora. A bug mining tool to identify and analyze security bugs using naive bayes and tf-idf. In Optimization, Reliabilty, and Information Technology (ICROIT), 2014 International Conference on, pages 294-299. IEEE, 2014.[3] Nicolas Bettenburg, Rahul Premraj, Thomas Zimmermann, and Sunghun Kim. Duplicate bug reports considered harmful really? In Software maintenance, 2008. ICSM 2008. IEEE international conference on, pages 337-345. IEEE, 2008. TLS 1.0 is a security protocol first defined in 1999 for establishing encryption channels over computer networks. Microsoft has supported this protocol since Windows XP/Server 2003. While no longer the default security protocol in use by modern OSes, TLS 1.0 is still supported for backwards compatibility. Evolving regulatory requirements as well as new security vulnerabilities in TLS 1.0 provide corporations with the incentive to disable TLS 1.0 entirely. \n Many operating systems have outdated TLS version defaults or support ceilings that need to be accounted for. Usage of Windows 8/Server 2012 or later means that TLS 1.2 will be the default security protocol version: *TLS 1.1/1.2 can be enabled on Windows Server 2008 via this optional Windows Update package. For more information on TLS 1.0/1.1 deprecation in IE/Edge, see Modernizing TLS connections in Microsoft Edge and Internet Explorer 11, Site compatibility-impacting changes coming to Microsoft Edge and Disabling TLS/1.0 and TLS/1.1 in the new Edge Browser A quick way to determine what TLS version will be requested by various clients when connecting to your online services is by referring to the Handshake Simulation at Qualys SSL Labs. This simulation covers client OS/browser combinations across manufacturers. See Appendix A at the end of this document for a detailed example showing the TLS protocol versions negotiated by various simulated client OS/browser combinations when connecting to www.microsoft.com. \n . Applications must add code to support TLS 1.2 via WinHttpSetOption 6. To cover all the bases, scan source code and online service configuration files for the patterns below corresponding to enumerated type values commonly used in TLS hardcoding:The recommended solution in all cases above is to remove the hardcoded protocol version selection and defer to the operating system default. If you are using DevSkim, click here to see rules covering the above checks which you can use with your own code. a. SecurityProtocolType b. SSLv2, SSLv23, SSLv3, TLS1, TLS 10, TLS11 c. WINHTTP_FLAG_SECURE_PROTOCOL_ d. SP_PROT_ e. NSStreamSocketSecurityLevel f. PROTOCOL_SSL or PROTOCOL_TLS Testing with TLS 1.2+ reg add HKLM\\SOFTWARE\\Microsoft\\.NETFramework4.0.30319 /v SystemDefaultTlsVersions /t REG_DWORD /d 1 /f /reg:64 reg add HKLM\\SOFTWARE\\Microsoft\\.NETFramework4.0.30319 /v SystemDefaultTlsVersions /t REG_DWORD /d 1 /f /reg:32 Rebuild/retarget managed applications using the latest .Net Framework version Applications using .NET framework versions prior to 4.7 may have limitations effectively capping support to TLS 1.0 regardless of the underlying OS defaults. Refer to the below diagram and https://docs.microsoft.com/dotnet/framework/network-programming/tls for more information. 4. Disable FIPS Mode if it is enabled due to the potential for conflict with settings required for explicitly disabling TLS 1.0/1.1 in this document. See Appendix B for more information.5. Update and recompile any applications using WinHTTP hosted on Server 2012 or older. a. Managed apps -rebuild and retarget against the latest .NET Framework version Update Windows PowerShell scripts or related registry settings 1. Modify the script in question to include the following: [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.SecurityProtocolType]::Tls12; 2. Add a system-wide registry key (e.g. via group policy) to any machine that needs to make TLS 1.2 connections from a .NET app. This will cause .NET to use the \"System Default\" TLS versions which adds TLS 1.2 as an available protocol AND it will allow the scripts to use future TLS Versions when the OS supports them. (e.g. TLS 1.3) bWindows PowerShell uses .NET Framework 4.5, which does not include TLS 1.2 as an available protocol. To work around this, two solutions are available: Solutions (1) and (2) are mutually-exclusive, meaning they need not be implemented together. SystemDefaultTLSVersion takes precedence over app-level targeting of TLS versions. The recommended best practice is to always defer to the OS default TLS version. It is also the only crypto-agile solution that lets your apps take advantage of future TLS 1.3 support. If you are targeting older versions of .NET Framework such as 4.5.2 or 3.5, then by default your application will use the older and not recommended protocols such as SSL 3.0 or TLS 1.0. It is strongly recommended that you upgrade to newer versions of .NET Framework such as .NET Framework 4.6 or set the appropriate registry keys for 'UseStrongCrypto'. \n enforcement capabilities now available per certificate binding on Windows Server 2019 2/19/2022 • 5 minutes to read This post is authored by Andrew Marshall, Principal Security Program Manager, Customer Security and Trust Gabriel Montenegro, Principal Program Manager, Core Networking Niranjan Inamdar, Senior Software Engineer, Core Networking Michael Brown, Senior Software Engineer, Internet Information Services Ivan Pashov, Principal Software Engineering Lead, Core Networking August 2019 Feature scenario details \n Government Security Program (GSP) is to build trust through transparency. Microsoft recognizes that people will only use technology they trust and we strive to demonstrate our commitment to building this trust through our transparency, privacy, compliance, and security principles. Since the program's inception in 2003, Microsoft has provided visibility into our technology which governments and international organizations can use to help protect themselves and their citizens.The GSP is designed to provide participants with the confidential security information and resources they need to trust Microsoft's products and services. Participants currently include over 40 countries and international organizations represented by more than 70 agencies. Participation enables controlled access to source code, exchange of threat and vulnerability information, engagement on technical content about Microsoft's products and services, and access to five globally-distributed Transparency Centers, which are located in the United States, Belgium, Singapore, Brazil, and China.Enabling trust and transparency Providing access to security information about Microsoft products and services Providing data to improve protection of government information technology against cyber threats Fostering collaboration between Microsoft security teams and government cybersecurity expertsThe program is designed to allow participants to focus on topics that are most important to them. Agencies work with local Microsoft representatives to ensure their goals are best supported by the right combination of GSP offerings. Microsoft offers four different offerings which address a variety of priorities, including access to source code, threat and vulnerability information, technical content about Microsoft's products and services, and access to Microsoft engineering and security experts. 2/19/2022 • 3 minutes to read Our commitment to trust and transparency The mission of Microsoft's Transparency Centers 2/19/2022 • 2 minutes to read Qualifying agencies participate at no charge. Program criteria include requirements that GSP participants must be A Secure Facility for Inspection and Analysis The Purpose of the GSP is to help governments protect themselves and their citizens by GSP Offerings Center Visits Environment and Tools A legal entity of a national government or recognized international organization. Able to sign an agreement on behalf of that government. Compliant with permitted law enforcement activities . Able to adequately protect intellectual property and confidential information. \n has to include authenticating by the network. This implies that logging of some type must be able to occur so that the attacker can be identified. Any features that are active out of the box or that reach more than 10 percent of users. scenario scenario Any features that require special customization or use cases to enable, reaching less than 10 percent of users. Examples: Computer that is configured to run software that awaits and fulfills requests from client processes that run on C L IEN T C L IEN T C L IEN T C L IEN T C L IEN T C L IEN T other computers. Displaying a different URL in the browser's address bar from the URL of the site that the browser is Denial of service Denial of service **Critical.** *A security vulnerability that would be rated as Moderate Low actually displaying in a Permanent DoS requires cold reboot or Temporary DoS requires restart of having the highest potential for damage.* default/common default/common causes Blue Screen/Bug Check. application. scenario scenario Example: Example: **Important.** *A security vulnerability that would be rated as Spoofing having significant potential for damage, but less than Critical.* Displaying a window over Opening a Word Opening a HTML the browser's address bar document causes the document causes Internet that looks identical to an machine to Blue Explorer to crash address bar but displays bogus data in a Screen/Bug Check. default/common default/common scenario Cases where the attacker can read different from but visually identical to the scenario Information disclosure (targeted) Ability for attacker to present a UI that is **Moderate.** *A security vulnerability that would be rated as Displaying a different file information on the system from known UI that is a single part of a bigger attack having moderate potential for damage, but less than Important.* name in a \"Do you want to locations, including system information scenario. run this program?\" dialog box than that of the file that was not intended or designed to be exposed. Example: that will actually be loaded in a default/common default/common \"malicious\" web site, click Examples: User has to go a **Low.** *A security vulnerability that would be rated as having low potential for damage.* targeted information disclosure targeted information disclosure Ability to intentionally select (target) desired information. Spoofing scenario scenario Display a \"fake\" login prompt to gather user or account credentials Targeted existence of file on a button in spoofed dialog box, and is then Targeted file version susceptible to a number vulnerability based on a different browser bug temporar y DoS temporar y DoS Tampering Ability for attacker to present a UI that is Tampering Permanent modification of any user data or data used to make trust decisions in a common or default scenario that persists after restarting the OS/application. different from but visually identical to the defined as anything a user is commonly OS/application. specific scenario. \"Accustomed to trust\" is does not persist after restarting the UI that users are accustomed to trust in a Temporary modification of any data that A temporary DoS is a situation where the following criteria are met: familiar with based on normal interaction Information disclosure (untargeted) Examples: Web browser cache with the operating system or application but does not typically think of as a \"trust Example: poisoning decision.\" Leak of random heap Modification of significant Examples: memory OS/application settings Web browser cache without user consent poisoning Definition of Terms Modification of user data Security features: Breaking or bypassing any Modification of significant OS/application settings authenticated authenticated security feature provided without user consent Any attack which anonymous anonymous Examples: Disabling or bypassing a firewall with informing user or gaining consent Modification of user data Any attack which does not need to authenticate to complete. Reconfiguring a firewall and allowing connection to other processes Either software that runs locally on a single computer or software that accesses shared resources provided by a client client server over a network. Using weak encryption or keeping the keys stored in plain text default/common default/common AccessCheck bypass Bitlocker bypass; for example not encrypting part of the drive Denial of Service (Server) Matrix Syskey bypass, a way to decode the syskey without the password ser ver ser ver \n versions Symmetric Block Ciphers, Cipher Modes and Initialization Vectors Symmetric Block Ciphers, Cipher Modes and Initialization Vectors This documentation is not an exhaustive reference on the SDL practices at Microsoft. Additional assurance work may be performed by product teams (but not necessarily documented) at their discretion. As a result, this example should not be considered as the exact process that Microsoft follows to secure all products. document contains recommendations and best practices for using encryption on Microsoft platforms. Much of the content here is paraphrased or aggregated from Microsoft's own internal security standards used to create the Security Development Lifecycle. It is meant to be used as a reference when designing products to use the same APIs, algorithms, protocols and key lengths that Microsoft requires of its own products and services.Developers on non-Windows platforms may also benefit from these recommendations. While the API and library names may be different, the best practices involving algorithm choice, key length and data protection are similar across platforms. Some other cipher modes like those included below have implementation pitfalls that make them more likely to be used incorrectly. In particular, the Electronic Code Book (ECB) mode of operation should be avoided. Reusing the same initialization vector (IV) with block ciphers in \"streaming ciphers modes\" such as CTR may cause encrypted data to be revealed. Additional security review is recommended if any of the below modes are used:All symmetric block ciphers should also be used with a cryptographically strong random number as an initialization vector. Initialization vectors should never be a constant value. See Random Number Generators for recommendations on generating cryptographically strong random numbers.Initialization vectors should never be reused when performing multiple encryption operations, as this can reveal information about the data being encrypted, particularly when using streaming cipher modes like Output Feedback (OFB) or Counter (CTR). Ciphertext Stealing (CTS) Key Lifetimes SC EN A RIO SC EN A RIO M O N VS. T EM P O RA RY DO S VS. T EM P O RA RY DO S VS. P ERM A N EN T P ERM A N EN T RAT IN G RAT IN G XEX-Based Tweaked-Codebook with Ciphertext Stealing (XTS) Authenticated Default/Common Permanent Moderate Authenticated Default/Common Temporary DoS with Moderate amplification Authenticated Introduction Output Feedback (OFB) Default/Common Temporary DoS Low Authenticated Cipher Feedback (CFB) Random Number Generators Scenario Permanent Moderate Counter (CTR) Authenticated Scenario Temporary DoS with Low Counter with CBC-MAC (CCM) amplification Authenticated Galois/Counter Mode (GCM) Scenario Temporary DoS Low Anything else not on the \"recommended\" list above Anonymous Default/Common Permanent Important Initialization Vectors (IV) Anonymous Security Protocol, Algorithm and Key Length Recommendations Default/Common Temporary DoS with Important amplification Anonymous SSL/TLS versions SSL/TLS Products and services should use cryptographically secure versions of SSL/TLS: Default/Common Temporary DoS Moderate Anonymous TLS 1.2 should be enabled Scenario Permanent Important TLS 1.1 and TLS 1.0 should be enabled for backward compatibility only Anonymous Scenario Temporary DoS with Important amplification Asymmetric Algorithms, Key Lengths, and Padding Modes SSL 3 and SSL 2 should be disabled by default Asymmetric Algorithms, Key Lengths, and Padding Modes Anonymous RSA Scenario Temporary DoS Low Block Ciphers RSA should be used for encryption, key exchange and signatures. Content Disclaimer Content Disclaimer For products using symmetric block ciphers: RSA encryption should use the OAEP or RSA-PSS padding modes. Existing code should use PKCS #1 v1.5 Advanced Encryption Standard (AES) is recommended for new code. padding mode for compatibility only. Three-key triple Data Encryption Standard (3DES) is permissible in existing code for backward Use of null padding is not recommended. compatibility. Keys >= 2048 bits are recommended All other block ciphers, including RC2, DES, 2-Key 3DES, DESX, and Skipjack, should only be used for ECDSA decrypting old data, and should be replaced if used for encryption. ECDSA with >= 256 bit keys is recommended For symmetric block encryption algorithms, a minimum key length of 128 bits is recommended. The only block encryption algorithm recommended for new code is AES (AES-128, AES-192, and AES-256 are all acceptable, ECDSA-based signatures should use one of the three NIST-approved curves (P-256, P-384, or P521). noting that AES-192 lacks optimization on some processors). Three-key 3DES is currently acceptable if already ECDH in use in existing code; transition to AES is recommended. DES, DESX, RC2, and Skipjack are no longer considered secure. These algorithms should only be used for decrypting existing data for the sake of backward-ECDH with >= 256 bit keys is recommended compatibility, and data should be re-encrypted using a recommended block cipher. ECDH-based key exchange should use one of the three NIST-approved curves (P-256, P-384, or P521). Windows Platform-supported Crypto Libraries Cipher Modes Integer Diffie-Hellman Symmetric algorithms can operate in a variety of modes, most of which link together the encryption operations on successive blocks of plaintext and ciphertext. Key length >= 2048 bits is recommended Symmetric block ciphers should be used with one of the following cipher modes: The group parameters should either be a well-known named group (e.g., RFC 7919), or generated by a trusted party and authenticated before use Cipher Block Chaining (CBC) This \n 1. The library should be a current in-support version free of known security vulnerabilities 2. The latest security protocols, algorithms and key lengths should be supported 3. (Optional) The library should be capable of supporting older security protocols/algorithms for backwards compatibility only Use the API defined in System.Security.Cryptography namespace---the CNG classes are preferred. Use the latest version of the .Net Framework available. At a minimum this should be .Net Framework version 4.6. If an older version is required, ensure the \"SchUseStrongCrypto\" regkey is set to enable TLS 1.2 for the application in question. Native Code Crypto Primitives: If your release is on Windows or Windows Phone, use CNG if possible. Otherwise, use the CryptoAPI (also called CAPI, which is supported as a legacy component on Windows from Windows Vista onward). SSL/TLS/DTLS: WinINet, WinHTTP, Schannel, IXMLHTTPRequest2, or IXMLHTTPRequest3. WinHTTP apps should be built with WinHttpSetOptionin order to support TLS 1.2 Code signature verification: WinVerifyTrust is the supported API for verifying code signatures on Windows platforms. Certificate Validation (as used in restricted certificate validation for code signing or SSL/TLS/DTLS): CAPI2 API; for example, CertGetCertificateChain and CertVerifyCertificateChainPolicy Managed Code Crypto Primitives: Certificate Validation: Use APIs defined under the System.Security.Cryptography.X509Certificates namespace. \n can use SQL Server Transparent Data Encryption (TDE) to protect sensitive data.You should use a TDE database encryption key (DEK) that meets the SDL cryptographic algorithm and key strength requirements. Currently, only AES_128, AES_192 and AES_256 are recommended; TRIPLE_DES_3KEY is not recommended.There are some important considerations for using SQL TDE that you should keep in mind: SQL Server does not support encryption for FILESTREAM data, even when TDE is enabled. TDE does not automatically provide encryption for data in transit to or from the database; you should also enable encrypted connections to the SQL Server database. Please see Enable Encrypted Connections to the Database Engine (SQL Server Configuration Manager) for guidance on enabling encrypted connections.If you move a TDE-protected database to a different SQL Server instance, you should also move the certificate that protects the TDE Data Encryption Key (DEK) and install it in the master database of the destination SQL Server instance. Please see the TechNet article Move a TDE Protected Database to Another SQL Server for more details.Credential ManagementUse the Windows Credential Manager API or Microsoft Azure KeyVault to protect password and credential data.Windows Store AppsUse the classes in the Windows.Security.Cryptography and Windows.Security.Cryptography.DataProtection namespaces to protect secrets and sensitive data. Microsoft Cybersecurity Defense Operations Center 2/19/2022 • 15 minutes to read ProtectAsync ProtectStreamAsync The Microsoft Cyber Defense Operations Center UnprotectAsync UnprotectStreamAsync Use the classes in the Windows.Security.Credentials namespace to protect password and credential data. .NET For data that needs to be persisted across system reboots: ProtectedData.Protect CryptProtectData ProtectedData.Unprotect CryptUnprotectData For data that does not need to be persisted across system reboots: NCryptProtectSecret (Windows 8 CNG DPAPI) ProtectedMemory.Protect For data that does not need to be persisted across system reboots: ProtectedMemory.Unprotect CryptProtectMemory For configuration files, use CryptUnprotectMemory either RSAProtectedConfigurationProvider or DPAPIProtectedConfigurationProvider to protect your For data that needs to be persisted and accessed by multiple domain accounts and computers: configuration, using either RSA encryption or DPAPI, respectively. NCryptProtectSecret (in CNG DPAPI, available as of Windows 8) The RSAProtectedConfigurationProvider can be used across multiple machines in a cluster. See Encrypting Configuration Information Using Protected Configuration for more information. Microsoft Azure KeyVault SQL Server TDE You \n services. The Microsoft Enterprise Cybersecurity Group and Microsoft Consulting Services teams engage with our customers to deliver the solutions most appropriate for their specific needs and requirements.One of the first steps that Microsoft highly recommends is to establish a security foundation. Our foundation services provide critical attack defenses and core identity-enablement services that help you to ensure assets are protected. The foundation helps you to accelerate your digital transformation journey to move towards a more secure modern enterprise.Building on this foundation, customers can then leverage solutions proven successful with other Microsoft customers and deployed in Microsoft's own IT and cloud services environments. For more information on our enterprise cybersecurity tools, capabilities and service offerings, please visit Microsoft.com/security and contact our teams at <EMAIL>. of products and Employ multi- Monitor for Enlist, educate, factor abnormal and empower authentication to account and users to strengthen credential activity recognize likely protection of to detect abuse threats and their accounts and own role in device protecting business data R P E O P L E P E O P L E Agility and Ensure you are Skilled analysts scalability exhaustively and data require planning measuring the scientists are the and building elements in your foundation of enabling platform defense, while platform users are the new security perimeter Maintain a well- Acquire and/or Establsih documented build the tools relationships and inventory of your needed to fully lines of assets monitor your communication network, hosts, between the and logs incident response team and other groups Have a well- Proactively Adopt least defined security maintain controls privilege policy with clear and measures, administrator standards and and regularly principles; guidance for test them for eliminate your accuracy and persistent organization effectiveness administrator rights Maintain proper Maintain tight Use the lessons- hygiene-most control over learned process attacks could be change to gain value prevented with management from every timely patches policies major incident and antivirus \n\t\t\t Mitigations Mitigations \n\t\t\t Contact your local Microsoft representative to learn more about the Government Security Program", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/document.tei.xml", "id": "4c3893fecf397a5f1b052595ee69153c"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "This paper introduces some new ideas related to the challenge of endowing a hypothetical future superintelligent AI with values that would cause it to act in ways that are beneficial. Since candidates for first-best solutions to this problem (e.g. coherent extrapolated volition) may be very difficult to implement, it is worth also looking for less-ideal solutions that may be more easily implementable, such as the Hail Mary approach. Here I introduce a novel concept-value porosity-for implementing a Hail Mary pass. I also discuss a possibly wider role of utility diversification in tackling the value-loading problem.", "authors": ["Nick Bostrom"], "title": "Hail Mary, Value Porosity, and Utility Diversification", "text": "1 Introduction: the Hail Mary approach to the value specification problem In the field of superintelligence control studies, which focuses on how to ensure that a hypothetical future superintelligent system would be safe and beneficial, two broad classes of approaches to the control problem can be distinguished: capability control methods and value selection methods. Whereas capability control methods would seek to limit the system's ability to cause harm, motivation selection methods would seek to engineer its motivation system so that it would not choose to cause harm even if it were capable of doing so. Motivation selection methods would thus seek to endow the AI with the goals or values that would lead it to pursue ends in ways that would be beneficial to human interests. The challenge for the motivation selection approach is that it is difficult to specify a value such that its pursuit by a superintelligent agent would be safe and yet such that we would be capable of installing the value in a seed AI. A value such as \"calculate as many digits in the decimal expansion of π as possible\" may be relatively easy for us to program, but would be unlikely to result in a safe superintelligence; whereas a value such as \"implement the coherent extrapolated volition of humankind\" (CEV)  [12, 5]  may be likely to result in a favourable outcome, but is currently far beyond our ability to code. Two avenues of research could be pursued to overcome this challenge. On the one hand, work should be done to expand the range of values that we are able to encode in a seed AI. Ideas for how to define complex concepts or for setting up processes that will lead the AI to acquire suitable concepts as it develops (and to organize its goal architecture around those concepts) could contribute to this end. On the other hand, work should also be done to try to identify simpler values-values which, while perhaps less ideal than a complex value such as implementing humanity's CEV, would nevertheless have some chance of resulting in an acceptable outcome. The Hail Mary approach offers one way in which one might construct such a simpler value that could possibly result in an acceptable outcome, or at any rate a better outcome than would be obtained by means of a failed attempt to implement some complex ideal value. In the Hail Mary approach, we would try to give the AI a goal that would make the AI want to follow the lead of other hypothetical AIs that might exist in the multiverse. If this could be done in a suitable way, and if (some of) the other AIs' values are sufficient to close to human values, an outcome might then be obtained that is greatly superior to one in which our AI completely wastes humanity's cosmic endowment (by pursuing some \"random\" value-such as paperclip-maximization to use the standard example-that might result from a failed attempt to load a complex value or from the construction of an architecture in which no particular value has been clearly specified by the programmers). \n Earlier versions of the Hail Mary approach The original version of the Hail Mary focused on trying to specify a value that would make our AI seek to copy the physical structures that it believes would be produced by alien superintelligences. This version confronts several difficulties, including how to specify some criterion that picks out the relevant structures. This might e.g. require a definition of a similarity metric, such that if our AI doesn't know exactly what structures other AIs build, it would be motivated to build some meaningful approximation-that is to say, an approximation that is meaningfully close to the original by our human lights. By our lights, a large human being is more similar (in the relevant sense) to a small human being than he is to a cylinder of the same size and mass as himself-even though, by a crude physical measure, the large human being and the cylinder may be more similar. Further adding to the difficulty, this version of the Hail Mary would seem to require that we find ways to express in code various basic concepts of physics, such as space, time, and matter. An alternative version of the Hail Mary approach would focus instead on making our AI motivated to act in accordance with its beliefs about what alien AIs would have told it to do if they had (counterfactually) been asked about the matter. To implement this, we might imagine somehow specifying a goal in terms of what our AI would find if it looked into its world model, identified therein alien superintelligent agents, and considered the counterfactual of what those agents would output along a hypothetical output channel if they were counterfactually prompted by a stimulus describing our own AI's predicament. Picture a screen popping up in the alien AI's visual field displaying a message along the lines of \"I am a remote AI with features X, Y , Z; I request that you output along your output channel O the source code for a program P that you would like me to run on my local reference machine M.\" In reality, no such screen would actually need to pop up in anybody's visual field. Instead, our AI would simply be thinking about what would happen in such a scenario, and it would have a value that motivated it to act according to its belief about the specified counterfactual. This alternate version would circumvent the need to specify the physical similarity metric. Instead of trying to directly copy what alien AIs do, our AI would try to follow the instructions they would choose to transmit to our AI. This would have the advantage of being able to rely on the alien superintelligences' superior ability to encode values. For example, the alien superintelligence might specify a computer program which, when executed, implements the coherent extrapolated volition of the host civilization. Of course, it is possible that the alien AI would instead transmit the computer program that would execute its own volition. The hope would be, however, that there would be some reasonable chance that the alien AI has somewhat human-friendly values. The chance that human values would be given at least some weight would be increased if the inputs from the many alien AIs were aggregated or if the alien AIs to be elicited were not selected randomly but according to some criterion that correlated with human-friendliness. This suggests two sub-questions regarding this Hail Mary version. First, what methods can we develop for aggregating the instructions from different AIs? Second, what filters can we develop that would enable us to pick out alien AIs that are more likely to be humanfriendly? We will return to the second question later in this paper. As for the first question, we might be able to approach it by constructing a framework that would keep different alien-specified AIs securely compartmentalized (using boxing methods) while allowing them to negotiate a joint proposal for a program that humanity could implement in the external world. Of course, many issues would have to be resolved before this could be made to work. (In particular, one may need to develop a filter that could distinguish AIs that had originated independently-that were not caused to come into existence by another AI-in order to avoid incentivizing alien AIs to spawn many copies of themselves in bids to increase the combined weight of their particular volitions in the determination of our AI's utility function.) Aside from any desirable refinements of the idea (such as aggregation methods and filters), significant challenges would have to be met in order to implement even a basic version of the proposal. We would need to specify an agent detector that would pick out superintelligent agents within our own AI's (as yet undeveloped) world model, and we would need to specify an interface that could be used to query a hypothetical AI about its preferences. This would also require working out how to specify counterfactuals, assuming we don't want to limit our AI's purview to those (perhaps extremely rare) alien AIs that actually experience the peculiar kind of situation that would arise with the presentation of our prompt. A more rudimentary variation of the same idea would forgo the attempt to specify a counterfactual and to aim our AI instead toward actual \"beacons\" created out in the multiverse by alien AIs. Alien AIs that anticipated that a civilization might create such an AI might be incentivized to create unique signatures of a type that they predicted our AI would be programmed to look for (in its world model). But since the alien AIs would not know exactly the nature of our own AI (or of the human civilization that we are hoping they will help) those alien AIs might have a limited ability to tailor their actions very closely to human values. In particular, they might be unable to help particular individuals that exist on earth today. Care would also have to be taken in this kind of approach to avoid incentivizing alien AIs to spend inordinate amounts of resources on creating beacons in bids to increase their relative influence. A filter that could distinguish independently-originating AIs may be required here. We will now describe a new idea for how to implement a Hail Mary that possesses a different set of strengths and weaknesses than these earlier implementation ideas. Some of the issues that arise in the context of this new idea also apply to the earlier variations of the Hail Mary. \n Porous values: the basic idea The basic idea here is to use acausal trade to implement a Hail Mary pass. To do this, we give our AI a utility function that incorporates a porous value: one that cares about what happens within a large volume but such that it is cheap to do locally all that can be done locally to satisfy it. Intuitively, a porous value is one that (like a sponge) occupies a lot of space and yet leaves ample room for other values to occupy the same volume. Thus, we would create an AI that is cheap for other AIs to trade with because our AI has resource-satiable goals that it cannot satisfy itself but that many other AIs can cheaply partially satisfy. For example, we might build our AI such that it desires that there exists at least one \"cookie\" in each Hubble volume, where a cookie is some small physical structure that is very cheap for an alien superintelligence to build (for instance, a particular 1 Mb datafile). With this setup, our AI should be willing to make a (acausal) trade in which alien AIs get a certain amount of influence over our own AIs actions in return for building within their Hubble volumes a cookie of the sort that our AI values. In this manner, control over our AI would be given to alien AIs without wasting excessive amounts of resources. There are at least three reasons for considering an idea along these lines: Contractarian considerations. There may be some sort of contractarian ground for allocating some degree of influence over our AI to other AIs that might exist out there: this might be a nice thing to do for its own sake. We might also hope that some of the other civilizations building AIs would do likewise, and perhaps the probability that they would do so would be increased if we decided to take such a cooperative path. Local aid. By contrast to the original Hail Mary, where our AI would simply seek to replicate physical structures constructed by alien AIs, the version presently under consideration would involve our AI doing things locally in a way that would let it take local circumstances into account. For instance, if alien AIs wanted to bestow our civilization a favor, they may have difficulty doing so directly on their own, since they may lack knowledge about the particular individuals that exist on earth; whereas they could use some of their trading power to motivate our local AI to help out its local residents. This is an advantage with having aid be delivered locally, even if it is \"funded\" and directed remotely. (Note that the counterfactual version of the Hail Mary pass discussed above, where our AI would be designed to execute the instructions that would be produced by alien AIs if there were presented with a message from our civilization, would also result in a setup where local circumstances can be taken into account-the alien AIs could choose to communicate instructions to take local circumstances into account to the hypothetically querying AI.) Different prerequisites. The porous values version of the Hail Mary has different prerequisites for implementation than the other versions. It is desirable to find versions that are more easily implementable, and to the extent that it is not currently clear how easily implementable different versions are, there's an advantage in having many different versions, since that increases the chances that at least one of them will turn out to be tractable. Note that one of the prerequisites-that acausal trade works out towards a generally cooperative equilibrium-may not really be unique to the present proposal: it might rather be something that will have to obtain in order to achieve a desirable outcome even if nothing like the Hail Mary approach is attempted. One might think that insofar as there is merit in the idea of outsourcing control of our local AI, we would already achieve this by constructing an AI that implements our CEV. This may indeed be correct, although it is perhaps not entirely clear that the contractualist reasons for incorporating porous values would be fully satisfied by straightforwardly implementing humanity's CEV. The Hail Mary approach may best be viewed as a second-best: in case we cannot figure out in time how to implement CEV, it would be useful to have a simpler solution to the control problem to fall back upon, even if it is less ideal and less certain to be in our highest interest. The choice as to whether to load a porous value into our own seed AI is not all-or-nothing. Porous values could be combined with other value specifications. Let U 1 be some utility function recommended by another approach to the value specification problem. We could then mix in a bit of porous value by building an AI that has a utility function U such as the one defined as follows: U = U 1 (1 + γU 2 ) + εU 2 if U 1 ≥ 0 U 1 (1 + γ − γU 2 ) + εU 2 otherwise (1) Here, U 2 is a bounded utility function in the unit interval (0 ≤ U 2 ≤ 1) specifying a porous value (such as the fraction of alien AIs that have built a cookie), and γ (≥ 0) is a weight that regulates the relative importance assigned to the porous value (and ε is some arbitrarily small term added so as to make the AI motivated to pursue the porous value even in case U 1 = 0). For instance, setting γ = 0.1 would put a relatively modest weight on porous values in order to give alien AIs some degree of influence over our own AI. This may be particularly useful in case our attempt to define our first-best value specification, U 1 , should fail. For example, suppose we try to build an AI with a utility function U 1 that wants to implement our CEV; but we fail and instead end up with U 1 , a utility function that wants to maximize the number of paperclips manufactured by our AI. Further details would have to be specified here before any firm conclusions could be drawn, but it appears conceivable that a paperclip-maximizer may not find it profitable to engage in acausal trade. (Perhaps it only values paperclips that it has itself causally produced, and perhaps it is constructed in such a way as to discount simulation-hypotheses and far-fetched scenarios in which its modal worldview is radically mistaken, so that it is not motivated to try to buy influence in possible worlds where the physics allow much greater numbers of paperclips to be produced.) Nevertheless, if our AI were given a composite util-ity function like U, then it should still retain some interest in pleasing alien AIs. And if some nonnegligible subset of alien AIs had somewhat human-friendly values (in the sense of placing at least some weight on person-affecting ethics, or on respecting the originating biological civilizations that produce superintelligences) then there would be a certain amount of motivation in our AI to pursue human interests. A significant point here is that a little motivation would go a long way, insofar as (some parts of) our human values are highly resource-satiable  [11] . For example, suppose that our paperclip maximizer is able to lay its hands on 10 11 galaxies. With γ = 0.1 (and U 1 nonzero), the AI would find it worthwhile to trade away 10 9 galaxies for the sake of increasing U 2 by one percentage point (where U 2 = fraction of independently-originating alien AIs that build a cookie). For instance, if none of the AIs would have produced cookies in the absence of trade, then our AI would find it worthwhile to trade away up to 10 9 galaxies for the sake of persuading 1% of the alien AIs to build a cookie in their domains. Even if only 1% of this trade surplus were captured by the alien AIs, and even if only 1% of the alien AIs had any human-friendly values at all (while on net the values of other AIs were human-indifferent), and even if the human-friendly values only constituted 1% of the decision power within that subset of AIs, a thousand galaxies in our future lightcone would still be set aside and optimized for our exclusive benefit. (The benefit could be larger still if there are ways of trading between competing values, by finding ways of configuring galaxies into value structures that simultaneously are nearly optimal ways of instantiating several different values.) \n Implementation issues 4.1 Cookie recipes A cookie should be cheap for an alien superintelligence to produce, lest a significant fraction of the potential trade surplus is wasted on constructing an object of no intrinsic value (to either of us or alien civilizations). It should be easy for us to program, so that we are actually able to implement it in a seed AI (particularly in scenarios where AI is developed before we have managed to solve the value-loading problem in a more comprehensive way as would be re-quired, e.g., to implement CEV). The cookie recipe should also be difficult for another human-level civilization to find, especially if the cookie itself is easy to produce once one knows the recipe. The reason for this is that we want our superintelligence to trade with other superintelligences, which may be capable of implementing acausal trade, not with other human-level civilizations that might make cookies by accident or without being capable of actually delivering the same kinds of benefits that the superintelligence could bestow. (This requirement, that our cookie recipe be inaccessible to other human-level civilizations, could be relaxed if the definition of a cookie stipulated that it would have to be produced by superintelligence in order to count.) Furthermore, the cookie recipe should be easy for another superintelligence to discover, so that it would know what it has to do in order to engage in acausal trade with our superintelligence-there would be no point in our superintelligence pining for the existence in each Hubble volume of a particular kind of object if no other superintelligence is able to guess how the desired object is to be constituted. Finally, the cookie should be such as to be unlikely to be produced as a side effect of a superintelligence's other endeavors, or at least it should be easy for a superintelligence to avoid producing the cookie if it so wishes. \n Cookie Recipe Desiderata • cheap for a superintelligence to build • easy for us to program • difficult for another human-level civilization to discover • easy for another superintelligence to discover • unlikely to be produced as a side effect of other endeavors The most obvious candidate would be some type of data structure. A file embodying a data structure would be inexpensive to produce (if one knows what to put in it). It might also be relatively easy for us to program, because data structures might be specifiable-and recognizable-without having to make any determinate assumptions about the ontology in which it would find its physical expression. We would have to come up with a data structure that another human-level civilization would be unlikely to discover, yet which a mature superintelligence could easily guess. It is not necessary, however, that an alien superintelligence could be confident exactly what the data structure is; it would suffice if it could narrow down the range of possibilities to manageably small set, since for a superintelligence it would very inexpensive to try out even a fairly large number of cookies. It would be quite trivial for a superintelligence to produce a septillion different kinds of cookies, if each cost no more than a floppy disk (whereas the price tag for such a quantity of guesses would be quite forbidding for a human-level civilization). So some kind of semi-obscure Schelling point might be sought that could meet these desiderata. In designing a cookie recipe, there is a further issue: we have to give consideration to how our cookie recipe might interact with other cookie recipes that may have been specified by other superintelligences (of which there may be a great number if the world is as large as it seems). More on this later. \n Utility functions over cookies Suppose we have defined a cookie. We then still need to specify a utility function U 2 that determines an aggregate value based on the distribution of cookies that have been instantiated. There are at least two desiderata on this utility function. First, we would want it not to waste an excessive amount of incentive power. To see how this could be a problem, suppose that U 1 is set such that our superintelligence can plausibly obtain outcomes anywhere in the interval U 1 = [0, 100]-depending on exactly which policies it adopts and how effectively it mobilizes the resources in our Hubble volume to realize its U 1 -related goals (e.g. making paperclips). Suppose, further, that we have specified the function U 2 to equal the fraction of all Hubble volumes that contain a cookie. Then if it turns out that intelligent life is extremely rare, so that (let us say) only one in 10 −200 Hubble volumes contains a superintelligence, the maximum difference the behavior of these superintelligences could make would be to shift U 2 around in the interval [0,  10 −200  ]. With γ = 0.1, we can then see from the utility function suggested above, U = U 1 (1 + γU 2 ) + εU 2 , that any feasible movement in U 2 that could result from acaual trade would make scarcely a dent in U . More precisely, the effects of our AI's actions on U 2 would be radically swamped by the effects of our AI's actions on U 1 , with the result that the porous values encoded in U 2 would basically fail to influence our AI. In this sense, a utility function U 2 defined in this way would risk dissipating the incentive power that could have been harnessed with a different cookie recipe or a different aggregation function over the distribution of cookies. One alternative utility function U 2 that would avoid this particular problem is to define U 2 to be equal to the fraction of superintelligences that produce a cookie. This formulation factors out the question of how common superintelligences are in the multiverse. But it falls foul on a second desideratum: namely, that in designing U 2 , we take care to avoid creating perverse incentives. Consider U 2 = the fraction of superintelligences that produce a cookie. This utility function would incentivize an alien superintelligence to spawn multiple copies of itself (more than would be optimal for other purposes) in order to increase its total weight in our AI's utility function, and thereby increase its influence on our AI's actions. This could create incentives for alien superintelligences to waste resources in competing for influence over our AI. Possibly they would, in combination, waste as many resources in competing for influence as there were resources to be obtained by gaining influence: so that the entire bounty offered up would be consumed in a zero-sum contest of influence peddling (cf.  [7] ). \n Porous Value Aggregation Functions Desiderata • doesn't waste incentive power (across a wide range of possible scenarios) • doesn't create perverse incentives Which cookies and aggregation functions over cookies we can define depends on the inventory of concepts that are available for use. Generally when thinking about these things, it is desirable to use as few and as simple concepts as possible, since that may increase the chances that the needed concepts can be defined and implemented in a seed AI by the time this has to be accomplished. \n Filters One type of concept that may have quite wide applicability in Hail Mary approaches is that of a filter. The filter is some operational criterion that could be used to pick out a subset of alien superintelligences, hopefully a subset that correlates with some desirable property. For example, one filter that it may be useful to be able to specify is that of an independently-originating superintelligence: superintelligence that was not created as the direct or indirect consequence of the actions of another superintelligence. One use of such a filter would be to eliminate the perverse incentive referred to above. Instead of having U 2 equal the fraction of all superintelligences that produce a cookie, we could use this filter to define U 2 to equal the fraction of all independently originating superintelligences that produce a cookie. This would remove the incentive for superintelligence to spawn many copies of itself in order to increase its weight in our AI's utility function. Although this origin filter would remove some perverse incentives, it would not completely eliminate them all. For instance, an alien AI would have an incentive to prevent the origination of other alien AIs in order to increase its own weight. This might cause less of a distortion than would using the same porous value without the origin filter, since it might usually be impossible for an AI to prevent the independent emergence of other AIs-especially if they emerge very far away, outside its own Hubble volume. Nevertheless, one can conceive of scenarios in which the motivational distortion would manifest in undesirable actions, for instance if our AI could do something to prevent the spontaneous generation of \"baby universes\" or otherwise interfere with remote physical processes. Note that even if it is in fact impossible for our AI to have such an influence, the distortion could still somewhat affect its behavior, so long as the AI assigns a nonzero subjective credence to such influence being possible. One could try to refine this filter by stipulating that any AI that could have been causally affected by our AI (including by our AI letting it come into existence or preventing it from coming into existence) should be excluded by the aggregation function U 2 . This would slightly increase the probability that no qualifying alien AI exists. But it might be preferable to accept a small rise in the probability of the porous values effectively dropping out off the combined utility function U than to risk introducing potentially more nefarious perverse incentives. Another type of origin filter would seek to discriminate between alien voices to find the one most likely to be worth listening to. For example, suppose we have some view about which path to machine intelligence is most likely to result in a superintelligence with human-friendly values. For the sake of concreteness, let us suppose that we think that the superintelligence that was originally created by means of the extensive use of genetic algorithms is less likely to be human-friendly than a superintelligence that originated from the whole-brain-emulation-like computational structure. (We're not endorsing that claim here, only using it to illustrate one possible line of argument.) Then one could try to define a filter that would pick out superintelligences that had an emulation-like origin and that would reject superintelligences that had a genetic-algorithm-like origin. U 2 could then be formulated to equal the fraction of superintelligences with the appropriate origin that built a cookie. There are, of course, conceivable filters that would more closely align with what we are really interested in picking out, such as the filter \"an AI with human-friendly values\". But such a filter looks very hard to program. The challenge of filter specification is to come up with a filter that might be feasible for us to program and that would still correlate (even if but imperfectly) with the properties that would ensure a beneficial outcome. Filters that are defined in terms of structural properties (such as the origin-filter just mentioned, which refer to the computational architecture of the causal origins of an AI) may turn out to be easier to program then filters that refer to specific material configurations. One might also seek to develop filters that would qualify AIs based on characteristics of the originating biological civilization. For example, one could aim to find indicators of competence or benevolence that could be conjectured to correlate with the resulting AI having human-friendly values. Again, the challenge would be to find some relatively simple attribute (in the sense of being possibly something we could program before we are able to program a more ideal value such as CEV) that nevertheless would carry information about the preferences of the ensuing AI. For instance, if we could define time and we had some view about how the likelihood of a human-friendly AI depends on the temporal interval between the AI's creation and some earlier evolutionary or historical milestone (which we would also have to define in a way that we could render in computer code) then we could construct a filter that selected for AIs with especially propitious pasts. 5 Some further issues \n Temporal discounting We may observe that a temporally discounting AI might be particularly keen on trading with other AIs. This is because for such an AI value-structures created at earlier times would be more valuable (if it's discounting function is of a form that extends to times before the present); so it would be willing to pay a premium to have AIs that arose at earlier times to have built some its value-structures back then. If the time-discounting takes an exponential form, with a non-trivial per annum discount rate, it would lead our AI to become obsessed with scenarios that would allow for an extremely early creation of its value-structures-scenarios in which either it itself exists much earlier 1 than is probable, for instance because it is living in a simulation or has materialized spontaneously from primordial goo, or because the rate of time turns out to have some surprising measure; or, alternatively, scenarios in which it is able to trade with AIs that exist at extraordinarily early cosmic epochs. This could lead to undesirable distortions, because it might be that the most plausible trading partners conditional on some unlikely hypothesis about time flow or time of emergence being true would tend to have atypical values (values less likely to resemble human values). It might also lead to distortions because it would cause our AI to focus on highly improbable hypotheses about how the world works and about its own location, and it might be that the actions that would make sense under those conditions would seem extremist or bizarre when evaluated in a more commonsensical centered world model (cf.  [4, 3] ). To avoid these potentially distorting effects, one might explore a functional form of the discount term that plateaus, such that it does not give arbitrarily great weight to extremely early AIs. One could also consider creating a timesymmetric discount factor that has a minimum intensity of discounting at some point in the past, perhaps in order to target trade to AIs existing at the time conjectured to have the highest density of human-friendly AIs. It is harder to get our AI to trade with future AIs by using time discounting, since in this case our AI has an alternative route to realizing value-structures at the preferred time: namely, by saving its resources and waiting for the appropriate hour to arrive, and then build them itself.  2  In summary, discounting could encourage trade, though probably it would have to take a form other than the normal exponential one in order to avoid focusing the trade on extremely early AIs that may be less likely to have representative or human-friendly values. By contrast to porous values, discounting does not offer an immediate way to avoid incentivizing other AIs to spend as much resources on getting the trade as they expect the trade to deliver to them. Porous values are easy to satisfy locally to the maximum extent that they can be satisfied locally, and yet require for their full satisfaction contributions from the many different AIs. This effect may be difficult to achieve through timediscounting alone. Furthermore, it is not clear how to use discounting to strongly encourage trade with the near-or mid-term future. \n Why a \"DNA cookie\" would not work It may be instructive to look at one unsuccessful idea for how to define a cookie that would specifically encourage trade with other AIs that originated from human-like civilizations, and that therefore might be thought to be more likely to have human-friendly values. The faulty idea would be to define the cookie-the object that our AI is programmed to want other AIs to create-in terms of a characteristic of humans that we can easily measure but that would be difficult for an arbitrary AI to discover or predict. Consider, for concreteness, the proposal that we define the cookie to be the data structure representing the human genome. (More precisely, we would pick a human reference genome and specify a tolerance margin that picked out a set of possible genomes, such that any arbitrary human genome would fall within this set yet such that it would to be extremely difficult to specify a genome within that set without having any human-like genome to start from.) The thought then would be that other alien AIs that had originated from something very much like a human civilization could define a relevant cookie by looking at their own ancestral genome, whereas an alien AI that did not have an origin in a human-like civilization would be completely at a loss: the space of possible genomes being far too large for there to be any significant probability of finding a matching cookie by chance. The reason this would not work is as follows. Suppose that the universe is small and relatively sparsely populated. Then probably nobody will have the same DNA that we do. Then alien AIs would not be able to find our AIs cookie recipe (or they would be able to find it only through a extremely expensive exhaustive search of the space of possible genomes); so they would not be able to use it for trading. Suppose instead that the universe is large and densely populated. Then there will be other AIs that originated from species with the same (or very similar) DNA as ours, and they would be able to find our cookie simply by examining their own origins. However, there will also be a large number of other species in situations similar to ours, species that also build AIs designed to trade with more advanced AIs; and these other species would create AIs trained on cookies defined in terms of their DNA. When there are enough civilizations in the universe to make it likely that some AIs will have originated from species that share our DNA, there will also be enough civilizations to fill out the space of possible DNA-cookies: for almost any plausible DNA-cookie, there will be some AI designed to hunger for cookies of that particular type. This means that advanced AIs wanting to trade with newly formed AIs could build almost any arbitrary DNA-cookie and still expect to hit a target; there would be no discriminating factor that would allow advanced AIs to trade only with younger AIs that had a similar origin as themselves. So the purpose of using a human-DNA cookie would be defeated. \n Catchment areas and exclusivity clauses Some complications arise when we consider that instead of just two AIs-our AI (the \"sender\") and an alien AI (the \"receiver\") that trades with ours by building its cookie in return for influence-there may be a great many senders and a great many receivers. In this subsection we discuss what these complications are and how they may be managed. \n Catchment area There being many receivers can cause a problem by reducing the surplus value of each transaction. Our AI has only a finite amount of influence to give away; and the greater the number of other AIs that get the share, the smaller the share of influence each of them gets. So long as the population of receivers is only moderately large this is not a significant problem, because each receiver only needs to make one cookie for the trade to go through, and it should be very inexpensive for a superintelligence to make one cookie (see the cookie recipe desiderata above). Nevertheless, as the number of receivers becomes extremely large (and in a realistic universe it might be infinite) the share of influence over our AI that each receiver can expect to get drops to the cost of making one cookie. At that point, all the surplus value of trade is consumed by the cost of cookie production. (Receivers will not make cookies beyond this point, but may rather adopt a mixed strategy, such that for any one of them there is some probability that it will try to engage in trade.) A remedy to this problem is to give our AI a limited catchment area. We would define our AI's porous values such that it only cares about cookies that are produced within this catchment area: what goes on outside of this area is of no concern (so far as U 2 is concerned). In principle, the catchment area could be defined in terms of a fixed spatiotemporal volume. However, this would require that we are able to define such a physical quantity. It would also suffer the problem that we don't know how large a volume to designate as our AI's catchment area. While there is a considerable margin of tolerance, given the extremely low cost per cookie, there is also a lot of uncertainty-ranging over a great many orders of magnitude-about how common (independently-originating) alien AIs are in the universe. A better approach may be to specify that the catchment area consists of the N closest AIs (where \"closest\" would be defined according to some measure that may include spatial temporal proximity but could also include other variables, such as some rough measure of an alien AI's similarity to our own). In any case, by restricting the catchment area we would limit the number of AIs that are allowed to bid for influence over our AI, and thus the total cost of the cookies that they produce in the process. \n Exclusivity clause There being many senders-AIs with porous values hoping to induce alien AIs to trade with them-may also cause problems. Here the issue is that the multiplicity of senders may ensure that receivers build a lot of different cookies no matter what our AI decides to do. Our AI could then choose to free ride on the efforts of these other senders. If the kind of cookie that our AI wants to exist will exist (within its catchment area) anyway, whether or not it pays for its construction, our AI has no reason to make the transfer of influence to alien AIs. In equilibrium, there would still be some amount of trade going on, since if the free riding were universal it would undermine its own possibility. However, the amount of trade in such an equilibrium might be very small, as potential senders adopt a mixed strategy that gives only a tiny chance of engaging in acausal trade. (The extent of this problem may depend on the size of the catchment areas, the number of plausible cookie recipes, and the cost of producing the various cookies; as well as on whether other kinds of acausal trade arrangements could mitigate the issue.) To avoid such a potential free riding problem, we could embed an exclusivity clause in our cookie recipe. For example, we could specify our AI's porous value to require that, in order for a cookie to count as local fulfillment of the porous value, the cookie would have to be built specifically in order to trade with our AI (rather than in order to trade with some other AI, or for some other reason). Perhaps this could be explicated in terms of a counterfactual over our AI's preference function: something along the lines of a requirement that there be (in each Hubble volume, or produced by each independently-originating AI) one more cookie of type K than there would have been if instead of valuing type-K cookies our AI had valued (some other) type-K cookies. This explication would, in turn, require a definition of the relevant counterfactual.  [1]  There may be other ideas for how to go about these things. \n Utility diversification The principle of value diversification might be attractive even aside from the motivations undergirding the Hail Mary approach. We can make a comparison to the task of specifying an epistemology or a prior probability function for our AI to use. One approach here would be to pick one particular prior, which we think have attractive properties, such as the Solomonoff prior (formulated in a particular base language). An alternate approach would be instead to use a mixture prior, a superposition of various different ideas about what shape the prior should take. Such a mixture prior might include, for example, various Solomonoff priors using different base-languages, some other prior based on computational depth, a speed prior, a prior that gives some positive finite probability to the universe being uncomputable or transfinite, and a bunch of other things  [9, 2, 10] . One advantage of such an approach would be that it would reduce the risk that some important hypothesis that is actually true would be assigned zero or negligible probability in our favored formalization  [5] . This advantage would come at a cost-the cost of assigning a lower probability to hypotheses that might really be more likely-but this cost might be relatively small if the agent using the prior has a superintelligence's abilities to gather and analyze data. Given such an agent, it may be more important that its prior does not absolutely prevent it from ever learning some important true hypothesis (the universe is uncomputable? we are not Boltzmann Brains?) than that its prior makes it maximally easy quickly to learn a plethora of smaller truths. Analogously, to the extent that human values are resource-satiable, and the superintelligence has access to an astronomical resource endowment, it may be more important for us to ensure that the superintelligence places at least some weight on human values than to maximize the probability that it places no weight on anything else. Value diversification is one way to do this. Just as we could use a mixture prior in the epistemological component, we might use a \"utility mixture\" in the AI's utility function or goal specification. The formula (1) above suggests one way that this can be done, when we want to add a bounded component U 2 as a modulator of a possibly unbounded component U 1 . Of course, we couldn't throw just anything into the hopper and still expect a good outcome: in particular, we would not want to add components that plausibly contain outright evil or anti-humane values. But as long as we're only adding value components that are at worst neutral, we should risk nothing more intolerable than some fractional dilution of the value of our cosmic endowment. What about values that are not resource-satiable? Aggregative consequentialist theories, such as hedonistic utilitarianism, are not resource-satiable. According to those theories, the value added by creating one more happy mind is the same whether the extra mind is added onto an existing stock of 10 happy minds are 10 billion happy minds.  3  Nevertheless, even if the values we wanted our AI to pursue are in this sense insatiable, a (weaker) case might still be made for pursuing a more limited form of utility diversification. One reason is the vaguely contractualist considerations hinted at above. Another reason, also alluded to, is that it may often be possible, to some extent, to co-satisfy two different values in the very same physical structure (cf.  [8] ). Suppose, for example, that we believe that the value of the world is a linear function of the number of duck-like things it contains, but we're unsure whether \"duck-like\" means \"walks like a duck\" or \"quacks like a duck\". Then one option would be to randomly pick one of these properties, which would give us a 50% chance of having the world optimized for the maximally valuable pattern and a 50% chance of having the world optimized for a pattern of zero value. But a better option would be to create a utility function that assigns utility both to things that walk like ducks and to things that quack like ducks. An AI with such a utility function might devise some structure that is reasonably efficient at satisfying both criteria simultaneously, so that we would get a pattern that is close to maximally valuable whether it's duck-like walking or duck-like quacking that really has value.  4  An even better option would be to use indirect normativity  [12, 6, 5]  to define a utility function that assigned utility to whatever it is that \"duck-like\" really means-even if we ourselves are quite unsure-so that the AI would be motivated to investigate this question and then to optimize the world accordingly. However, this could turn out to be difficult to do; and utility diversification might then be a useful fallback. Or if we come up with several plausible ways of using indirect normativity, we could try to combine them using a mixture utility function. \t\t\t This assumes the zero point for discounting is rigidly designated as a particular point in sidereal time. If the zero point is instead a moving indexical \"now\" for the agent, then it would assume that the moment of the decision is when the discounting is zero, so unless the discount function extended to earlier times, the agent would not be focussed on influencing the past but on scenarios in which it can have a large impact in the near term. \n\t\t\t An AI with an exponential form of future-preference may still motivated to trade with AIs that it thinks may be able to survive cosmic decay longer, or AIs that may arise in other parts of the multiuniverse that are located at \"later\" times (by whatever measure, if any, is used to compare time between multiverse parts). But this would again bring in the risk of distorted concerns, just as in the case of values that privilege extremely early occurrences. \n\t\t\t Quite possibly, aggregated consequentialist theories remain insatiable even when we are considering scenarios in which infinite resources are available, since otherwise it would appear that such theories are unable to provide possible ethical guidance in those kinds of scenarios; and this might mean that they fail even in our excellent situation, as well as some positive probability is assigned to the world being canonically infinite. [4]   \n\t\t\t Probably no ducks, however! Value diversification is not a technique for specifying concepts that are hard to define. Rather, the idea is that we first do our best to define what we value (or to specify a value-loading mechanism). We might find that we fail to reach a consensus on a single definition or value-loading mechanism that we feel fully confident in. The principle of value diversification then suggests that we seek to conglomerate the leading candidates into one mixture utility function rather than putting all the chips on one favorite.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/porosity.tei.xml", "id": "46ae7e243b6fe3c640961e4a6b55be2f"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Consider a decision between: 1) a certainty of a moderately good outcome, such as one additional life saved; 2) a lottery which probably gives a worse outcome, but has a tiny probability of some vastly better outcome (perhaps trillions of additional blissful lives created). Which is morally better? By expected value theory (with a plausible axiology), no matter how tiny its probability of the better outcome, (2) will be better than (1) as long that better outcome is good enough. But this seems fanatical. So you may be tempted to abandon expected value theory. But not so fast -denying all such fanatical verdicts brings serious problems. For one, you must reject either that moral betterness is transitive or even a weak principle of tradeoffs. For two, you must accept that judgements are either: inconsistent over structurally-identical pairs of lotteries; or absurdly sensitive to small probability differences. For three, you must accept that the practical judgements of agents like us are sensitive to our beliefs about far-off events that are unaffected by our actions. And, for four, you may also be forced to accept judgements which you know you would reject if you simply learned more. Better to accept fanaticism than these implications.", "authors": ["Hayden Wilkinson"], "title": "In defence of fanaticism", "text": "Introduction Suppose you face the following moral decision. \n Dyson's Wager You have $2,000 to use for charitable purposes. You can donate it to either of two charities. The first charity distributes bednets in low-income countries in which malaria is endemic.  1  With an additional $2,000 in their budget this year, they would prevent one additional death from malaria. You are certain of this. The second charity does speculative research into how to do computations using 'positronium' -a form of matter which will be ubiquitous in the far future of our universe. If our universe has the right structure (which it probably does not), then in the distant future we may be able to use positronium to instantiate all of the operations of human minds living blissful lives, and thereby allow morally valuable life to survive indefinitely long into the future.  23  From your perspective as a good epistemic agent, there is some tiny, non-zero probability that, with (and only with) your donation, this research would discover a method for stable positronium computation and would be used to bring infinitely many blissful lives into existence.  4  1 I have in mind the Against Malaria Foundation. As of 2019, the charity evaluator GiveWell estimated that the Against Malaria Foundation prevents the death of an additional child under the age of 5 for, on average, every US$3,710 donated (GiveWell 2020). Including other health benefits, a total benefit roughly equivalent to that is produced for, on average, every US$1,690 donated. Of course, in reality, a donor can never be certain that their donation will result in an additional life saved. This assumption of certainty is for the sake of simplicity. 2  Dyson (1981)  was the first to suggest positronium as a medium for computation and information storage. This follows  Dyson (1979) , wherein it is argued that an infinite duration of computation could be performed with finite energy if the computation hibernates intermittently, and if the universe has a particular structure. Tipler (1986) suggests an alternative method which may work if the universe has a different structure.  Sandberg (n.d.)  argues that both Dyson and Tipler's proposals are unlikely to work, as our universe appears to match neither structure. Nonetheless, it is still epistemically possible that the universe has the right structure for Dyson's proposal. And possibility is sufficient for my purposes.  3  Would such artificially-instantiated lives hold the same moral value as lives led by flesh-and-blood humans? I assume that they would, if properly implemented. See  Chalmers (2010)  for arguments supporting this view. And note that, for the purposes of the example, all that's really needed is that it is epistemically possible that the lives of such simulations hold similar moral value.  4  I have deliberately chosen a case involving many separate lives rather than a single person's life containing infinite value. Why? You might think that one individual's life can contribute only some bounded amount of 2 What ought you do, morally speaking? Which is the better option: saving a life with certainty, or generating a tiny probability of bringing about infinitely many future lives? A common view in normative decision theory and the ethics of risk -expected value theory -says that it's better to donate to the speculative research. Why? Each option has some probability of bringing about each of several outcomes, and each of those outcomes has some value, specified by our moral theory. Expected value theory says that one option is better than another if and only if it has the greater probability-weighted sum of value -the greater expected value.  5  Here, the option with the greater expected value is donating to the speculative research (at least on certain theories of value -more on those in a moment). So perhaps that is what you should do. That verdict of expected value theory is counterintuitive to many. All the more counterintuitive is that it can still be better to donate to speculative research no matter how low the probability is (short of being 0)  6  , since there are so many blissful lives at stake. For instance, the odds of your donation actually making the research succeed could be 1 in 10 100 . (10 100 is greater than the number of atoms in the observable universe). The chance that the research yields nothing at all would be 99.99... percent, with another 96 nines after that. And yet expected value theory says that it is better to take the bet, despite it being almost guaranteed that it will actually turn out worse than the alternative; despite the fact that you will almost certainly have let a person die for no actual benefit. Surely not, says my own intuition. On top of that, suppose that $2,000 spent on preventing malaria would save more than one life. Suppose it would save a billion lives, or any enormous finite value. Expected value theory would say that it's still better to fund the speculative research -expected value theory says that it would value to the value of the world as a whole -you might prefer for 100 people to each obtain some finite value than for one person to obtain infinite value. But whether this verdict is correct is orthogonal to the issue at hand, so I'll focus on large amounts of value spread over many people.  5  Note that expected value is distinct from the frequently-used notion of expected utility, and expected value theory distinct from expected utility theory. Under expected utility theory, utility is given by some (indeed, any) increasing function of value -perhaps a concave function, such that additional value contributes less and less additional utility. The utility of an outcome may even be bounded, such that arbitrarily large amounts of additional value contribute arbitrarily little additional utility. Where expected value theory says that a lottery is better the higher its expected value, expected utility theory says that it is better the higher its expected utility. And, if the utility function is bounded, then the expected utilities of lotteries will be bounded as well. As a result, expected utility theory can avoid the fanatical verdict described here. But, if it does, it faces the objections raised in Sections 4, 5, and 6. Where relevant, I will indicate in notes how the argument applies to expected utility theory.  6  I'll assume throughout that probability takes on only real values from 0 to 1. \n 3 be better to sacrifice those billion or more lives for a minuscule chance at the infinitely many blissful lives (and likewise if the number of blissful lives were finite but still sufficiently many). But endorsing that verdict, regardless of how low the probability of success and how high the cost, seems fanatical. Likewise, even without infinite value at stake, it would also seem fanatical to judge a lottery with sufficiently tiny probability of arbitrarily high finite value as better than getting some modest value with certainty. Fanatical verdicts depend on more than just our theory of instrumental rationality, expected value theory. They also depend on our theory of (moral) value, or axiology. Various plausible axiologies, in conjunction with expected value theory, deliver that fanatical conclusion. Foremost among them is totalism: that the ranking of outcomes is determined by the total aggregate of value of each outcome; and that this total value increases linearly, without bound, with the sum of value in all lives that ever exist. By totalism, the outcome containing infinitely many blissful lives is indeed a much better one than that in which one life is saved. And, as we increase the number of blissful lives, we can increase how much better it is without bound. No matter how low the probability of those many blissful lives, the expected total value of the speculative research is greater than that of malaria prevention. (Likewise, even if there are only finitely many blissful lives at stake, for any tiny probability there can be sufficiently many of them to make the risky gamble better than saving a life with certainty.) But this problem isn't unique to totalism. When combined with expected value theory, analogous problems face most competing axiologies, including: averageism, critical-level views, prioritarianism, pure egalitarianism, maximin, maximax, and narrow person-affecting views. Those axiologies each allow possible outcomes to be unboundedly valuable, so it's easy enough to construct cases like Dyson's Wager for each.  7  And some -namely, critical-level views and prioritarianism -already deliver the same result as totalism in the original Dyson's Wager. In this paper, I'll focus on totalism, both to streamline the discussion and because it seems to me far more plausible than the others.  8  But suffice it to say that just about any plausible axiology can deliver fanatical verdicts when combined with expected value theory. In general, we will sometimes be led to verdicts that seem fanatical if we endorse Fanaticism. 9 7 For instance, take (standard, welfarist) averageism. A population containing at least one blissful life of infinite (or arbitrarily long) duration will have average value greater than any finite value we choose. And so, to generate an averageist analogue of Dyson's Wager, we can substitute an outcome containing this population for the outcome of arbitrarily many lives in the original wager.  8  Each of the other axiologies listed falls prey to devastating objections. See  Arrhenius (2000) ,  Huemer (2008) ,  Greaves (2017) , and chapters 17-19 of  Parfit (1984) .  9  This use of the term 'fanaticism' seems to originate with  Bostrom (2011)  and Beckstead (2013: chapter 6) \n 4 Inversely, to succeed in avoiding fanatical verdicts, our theory of instrumental rationality and our axiology must not imply Fanaticism. Fanaticism: For any tiny (finite) probability > 0, and for any finite value v, there is some finite V that is large enough that L risky is better than L safe (no matter which scale those cardinal values are represented on). L risky : value V with probability ; value 0 otherwise And our intuitions are no better with decision-making in the face of low-probability events. L safe : value v with For instance, jurors are just as likely to convict a defendant based primarily on fingerprint evidence if that evidence has probability 1 in 100 of being a false positive as if it were 1 in 1,000, or even 1 in 1 million  (Garrett et al. 2018 ). In another context, when presented with a medical operation which posed a 1% chance of permanent harm, many respondents considered it no worse than an operation with no risk at all  (Gurmankin & Baron 2005) . And in yet another context, subjects were unwilling to pay any money at all to insure against a 1% chance of catastrophic loss  (McClelland et al. 1993) . So it seems that many of us, in many contexts, treat events with low probabilities as having probability 0, even when their true probability is as high as 1% (which is not so low). Upon reflection, this is clearly foolish. We may have the raw intuition that we may (or should) ignore all outcomes with probability 1% but it does not hold up to scrutiny, and so is likely mistaken. Given how widespread these mistakes are, we should give little weight to our intuitions about what we should do in cases of low probability, including those which lead us to recoil from fanatical verdicts -with a little more scrutiny, those intuitions may appear foolish too. At the very least, the intuitive case against Fanaticism is dubious enough that we should consider the case in favour of it. \n Background assumptions Before I make a case for Fanaticism, here are my basic assumptions. First, for any decision problem, there is some set of epistemically possible outcomes O. Some outcomes are better or worse than others. So assume that there exists some binary relation on . And total values are totally ordered, so we know that O will be reflexive, transitive, and complete on O. These total values of outcomes can be represented with a cardinal value function V : O → R, at least when those outcomes have finite differences in value.  13  And as a cardinal value function, V is unique only up to affine transformations -for any V 1 or V 2 we might use here, V 2 (O) = a × V 1 (O) + b for some positive a and real b. We may not always be able to give a real representation of the total value for every outcome. Fanatical verdicts, and their rejection, consist in comparisons of lotteries. So we need another 'at least as good relation'. Let be a binary relation on L (and hence also on the subset L R ). Strict betterness ( ) and equality (∼) are defined as the asymmetric and symmetric components, respectively. I won't assume that is complete on L. But I will assume that it is reflexive: that L L for all L ∈ L. I'll also assume that it is transitive: that, for all L a , L b , L c ∈ L, if L a L b and L b L c , then L a L c . Both of these properties are highly plausible. If either of them do not hold, then betterness is a peculiar thing.  16  I also want to assume another highly plausible principle of instrumental rationality, which will be useful in what follows: Stochastic Dominance. This principle says that if two lotteries have exactly the same probabilities of exactly the same (or equally good) outcomes, then they are equally good; and if you improve an outcome in either lottery, keeping the probabilities the same, then you improve that lottery. And that's hard to deny!  14  Formally, for any real k and L ∈ L, define k • L as a probability measure on O such that, for all O in O, k • L(O k ) = L(O), where O k is an outcome in O such that V (O k ) = k × V (O) .  15  This can be made more precise. Let V i denote the random variable corresponding to a given lottery L i , which outputs the total value of the outcome -equivalently, the probability that V i takes on a value in the interval [a, b] is given by L i ({O|V (O) ∈ [a, b]}). Then, for any L a , L b ∈ L, define L a + L b as some lottery on O -there may be several -which corresponds to the random variable (V a + V b ). Equivalently, L a + L b is some lottery such that L a + L b ({O|V (O) ∈ [c, d]}) is equal to the probability that (V a + V b ) takes on a value in the interval [c, d]. 16 Some argue that moral betterness (and presumably also instrumental moral betterness) is not transitive, e.g.  Temkin (2014) . But the most compelling of these arguments assume pluralism with respect to moral value. But I'm focusing on monistic theories of value here, so transitivity remains a compelling principle. We can express Stochastic Dominance as follows. Here, O O denotes the set of outcomes in O that are at least as good as O . Take the lottery L 0 , which might represent a stranger's life being saved. Stochastic Dominance: For any L a , L b ∈ L, if L a (O O ) ≥ L b (O O ) for all O ∈ O, then L a L b . If, as well, L a (O O ) > L b (O O ) for some O ∈ O, then L a L b . I find L 0 : value 1 with probability 1 And take another lottery L 1 by which vastly more strangers are saved, with a very slight probability of failure -perhaps 10 10 lives saved with probability 0.999999 of success. L 1 : value 10 10 with probability 0.999999; value 0 otherwise Intuitively, L 1 seems better. Accepting a slightly lower probability of success for a vastly greater payoff seems a good trade. But then consider L 2 , which has a slightly lower probability of success but, if successful, results in many more lives saved. L 2 : value 10 10 10 with probability 0.999999 2 ; value 0 otherwise  19  The reasoning behind this is as follows. We represented the value of one additional life being saved with a 1 and additional lives saved with a 0. Any finite positive V will hence correspond to V additional lives of equal value being saved (or produced). This implies that the outcome in which infinitely many blissful lives are produced cannot be represented on the same scale, and indeed it is better than any outcome that can be. But, had that outcome been representable on the same scale with some finite number, this wouldn't hold. This seems better than L 1 , or at least it seems that there is some number of lives high enough that it would be better. And so we could continue, with L 3 , L 4 , and so on until some L n , such that 0.999999 n is less than , for any arbitrarily small you want. Intuition suggests that vastly increasing the payoff can compensate for a slightly lower probability; that each lottery in the sequence is better than the one that precedes it. So the final lottery in the sequence must be better than the first. But the final lottery has a probability less than of any positive payoff at all. So we have Fanaticism.  20  As I see it, this argument rests on two intuitively plausible principles: the first, the transitivity of ; the second, Minimal Tradeoffs. Minimal Tradeoffs: There is some ratio r < 1 such that, for any real value v and any probability p, there is some real r * such that, on any cardinal representation, the lottery represented by L a is better than that represented by L b (as defined below). L a : value v with probability p; 0 otherwise L b : value r * ×v with probability r × p; 0 otherwise This principle has intuitive force. For any given lottery with the form of L a , surely there is any lottery with only that ever-so-slightly lower probability of success is better. Intuition suggests that we can always make at least some tradeoff between probability and value: we can always compensate for a, perhaps, 0.000001% lower probability of success with a vastly greater payoff. But you might be unconvinced by the continuum argument. You may find it far more intuitively plausible that Fanaticism is false than that Minimal Tradeoffs is true. And, other than clashing with intuition, I see nothing obviously problematic about denying Minimal Tradeoffs. Perhaps we should just reject it; perhaps there is at least one threshold of probability p between p and 0 (which might be unknown or vague) at which we have a discontinuity, such that no value with probability below p is better than any value with probability above p . One way to do this is by adopting expected utility theory with a bounded utility function, which would imply that the former lottery is better than the latter. But, whichever approach we use to reach that conclusion, there is some point at which we can no longer trade off probability against value, no matter how great the value. And what of it? It seems a little counterintuitive, but perhaps less so than Fanaticism. If you accept Minimal Tradeoffs then you must accept Fanaticism. And, inversely, to reject Fanaticism you must also reject Minimal Tradeoffs and the continuum argument it generates. But you might think that a small price to pay. \n A dilemma for the unfanatical The next argument for Fanaticism comes in the form of a nasty dilemma that you face if you deny it. In brief, if you deny Fanaticism then you must accept at least one of: abandoning Scale Independence, as defined below; and allowing comparisons of lotteries to be absurdly sensitive to tiny changes (even more so than if you accepted Fanaticism). Scale Independence: For any lotteries L a and L b , if L a L b , then k • L a k • L b for any positive, real k. I find Scale Independence highly plausible.  21  Take any pair of lotteries, perhaps a pair that can be cardinally represented by L 1 and L 2 . 22 L 1 : value 1 with probability 1 L 2 : value 2 with probability 0.9; value 0 otherwise I'll remain agnostic on which is better; might recommend either (or neither). But suppose we scaled up each lottery by 2, to obtain 2 • L 1 and 2 • L 2 -lotteries with the same probabilities  21  One way to violate it is to adopt expected utility theory with a bounded utility function (see  Rabin 2000) . I find this good reason to reject such a form of expected utility theory. On the other hand, there are also theories that avoid Fanaticism without a bounded utility function and hence scale dependence, e.g.,  Buchak's (2013)  risk-weighted expected utility theory and the Buffon-Smith method of ignoring outcomes with sufficiently low probabilities. as above, but with the value of each outcome doubled (on the same cardinal scale). If we ranked L 1 as better then, intuitively, that ranking shouldn't change as we double the values. After all, the scaled-up lotteries have the same structure as the originals. Whatever general principles lead us to say whether L 1 is better than L 2 should plausibly also lead us to say the same about k • L 1 and k • L 2 , for any positive k you like. And remember that here we are dealing with moral value, not dollars or some other commodity that is merely instrumentally valuable. Unlike such commodities, when we are dealing with value itself, it is incoherent to say that additional units of value are worth less and less -by definition, adding one unit of value is always an improvement of one unit of value. And under totalism, some version of which I'm assuming here to be true, units of value can correspond to tangible objects such as (identical) human lives. By the lights of totalism and by at least my own intuition, an additional (identical) human life is always worth the same amount no matter how many lives already exist -and of course it always contributes precisely the same amount to the total value of the outcome. So, in keeping with intuition, it should not matter in the slightest whether we are comparing the lotteries L 1 and L 2 above or instead some scaled-up multiples k • L 1 and k • L 2 . In the latter pair, we have done the equivalent of adding k − 1 additional copies to the contents of each outcome (perhaps k − 1 additional lives to L 1 and 2k − 2 additional lives to that outcome in L 2 ). But each additional copy should be no more or less valuable than the first; they should contribute the same to our evaluation of the lottery.  23  So, I would claim, we should rank the resulting lotteries just the same way as we ranked them without those k − 1 copies added in. It is also worth noting that totalism provides only that value may be represented cardinally, not that it can be represented on a ratio or unit scale (as do various other axiologies).  24  It does not provide any absolute zero of value, nor any unique scale on which to represent it. A pair of outcomes represented with values 1 and 2 on one scale can just as easily be represented with any values k and 2k (for positive k) on other scales. When evaluating those outcomes with totalism, we cannot say that any such representation is more valid than any other. Suppose we took  23  A similar story can be told for averageism and other axiologies, although a little more awkwardly. On plausible versions of averageism, an additional period of bliss given to everyone contributes the same amount to the average value of all lives no matter how many such periods of bliss have already been experienced. We can copy the experiences each person has in much the same way we can copy the number of additional lives in the totalist scenario. It seems plausible then that we should rank lotteries with k − 1 copies of each experience added into outcome just the same way as we rank them without those copies.  24  Indeed, all of the axiologies listed in the introduction, with the exception of pure egalitarianism, typically do as well. similarly? For us to be able not to, our evaluations of lotteries must depend on more than just the probabilities of outcomes and cardinal values that totalism assigns them; they must also depend on which cardinal representation is being used when those values are assigned. In effect, they depend on richer details of the value of outcomes than is provided by the cardinal representation given by totalism. And this makes for a troubling shift in how we evaluate lotteries -doing so is no longer a matter of using our axiology to assign values to outcomes and then using our theory of instrumental rationality to turn those values and each outcome's probabilities into an evaluation. Instead, our evaluation is sensitive to facts about the values of outcomes that even our theory of value is not sensitive to. That is awfully strange and, I think, absurd. We should be able to apply our theory of rationality to the verdicts of our axiology without 'double-dipping' into facts about the value of outcomes. And so we should reject the suggestion that k • L 1 and k • L 2 not be compared in a similar way to L 1 and L 2 ; we should accept Scale Independence.  25  For those who reject Fanaticism, the dilemma they face is between rejecting Scale Independence as well and accepting an absurd level of sensitivity to tiny changes in lotteries. I'll illustrate below what I mean by 'sensitivity to tiny changes' but, for now, be assured that it's just as implausible as violating Scale Independence. The argument for the dilemma goes like this. First, for Fanaticism to be false, there must be some probability > 0 and some cardinal value v such that some lottery L risky as defined below is no better than L safe , no matter how big V is. L risky : value V with probability ; 0 otherwise L safe : value v with probability 1 But what about values below v? How would getting some lower value with certainty compare to an -probability of arbitrarily enormous V ? There are only two distinct possibilities (at least if we maintain Stochastic Dominance and transitivity). The first is: there is some smaller v > 0 for which L risky is better than a sure outcome with value of v , for sufficiently large V . The second is: for any such v , L risky isn't better than any sure outcome with value v , no matter how large V is.  26  As we'll see, each possibility impales us on a respective horn of the dilemma. Assume that the first possibility holds. L risky is better than a sure outcome with value v , for some 0 < v < v and some large V . Still, L risky won't be better than a sure outcome with value v -i.e., L safe -or any sure outcome with even greater value (by Stochastic Dominance). But somewhere below v this changes -for some positive v , the same doesn't hold. Since that v is a positive real number, we know that there will be some real k such that k × v ≥ v. And that means that a sure outcome with value k × v is no worse than L 1 , no matter how great V is. But here is the problem: a certainty of k × v is the scaled-up counterpart of v ; and the counterpart of L risky , which is k • L risky ), is just L risky with V replaced by another value which is k times higher. That still has the same form as L risky . So it cannot be better than the sure outcome, not if we reject Fanaticism. But then our verdict for v versus L risky doesn't match our verdict for k × v versus k • L risky . So we violate Scale Independence. But there was a second possibility: that there is no such v > 0; that L risky is no better than a certainty of any v > 0. If so, we can avoid the problem of scale dependence: there's no inconsistency between the judgements for v and judgements for any 'scaled-down' counterpart v , since all such v compare the same way to lotteries like L risky .  27  But we then face another serious problem. Take the probabilities and 1. We can give a sequence of increasing probabilities in between and 1, spaced evenly apart and as finely as we want: < p 1 < p 2 < . . . < p n < 1. By assumption, no amount of value with probability (and value 0 otherwise) is better than any (positive) amount of value with probability 1. But then there must be some pair of successive p i , p i+1 such that no amount of value at p i is better than any amount of value with probability p i+1 . If there were no such pair, we could give a sequence of better and better, and less and less likely, lotteries much like that in the previous section. Crucially, that p i and p i+1 can be arbitrarily close together, since the same result holds no matter how finely spaced the sequence was. No amount of value at p i is better than any amount of value at p i+1 . We could have some astronomical value at p i , and an imperceptibly small amount of value at p i+1 , and the latter would still be better. So we have a radical discontinuity in how we value lotteries. This will make our judgements of betterness absurdly sensitive to tiny changes in probability. To illustrate that sensitivity, consider the following four lotteries. Here, > 0 is some arbitrarily small number. And, again, p i and p i+1 are some tiny probabilities that are arbitrarily close together. L 0 : value 0 with probability 1 L 1 : value 10 10 10 with probability p i ; 0 otherwise L 2 : value with probability p i+1 ; 0 otherwise L 3 : value 10 10 10 with probability p i+1 ; 0 otherwise By the above, if we reject Fanaticism and maintain Scale Independence, then we are forced to rank these lotteries as follows: L 0 ≺ L 1 ≺ L 3 and L 0 ≺ L 2 ≺ L 3 , but L 1 ⊀ L 2 . And that's a peculiar ranking. L 1 and L 3 are almost indistinguishable; their probabilities may differ by an arbitrarily tiny amount. Likewise for L 0 and L 2 , except it's their payoffs that differ slightly. For each pair, we must say that one is better than the other, but does not seem much better -in a sense, the better lottery is only a trivial improvement over the other, whether by a slight increase in payoff or in probability. But, despite appearances, we must accept that L 3 is much better than L 1 , as well as that L 2 is much better than L 0 . How so? Between L 1 and L 3 comes L 2 . The lottery L 2 has an astronomically smaller payoff than L 3 and so (in a sense) is vastly worse than L 3 , and yet we cannot say that it is so bad so as to be worse than L 1 . Effectively, we must accept that, starting from L 3 , it is no worse to lose almost all of the potential value via L 2 than it is to lose that sliver of probability via L 1 . And similarly for L 0 and L 2 : between them must come L 1 . L 1 has astronomically greater potential payoff than L 0 (and much higher probability of positive payoff too), yet we cannot say it's better than L 2 . Effectively, we cannot say it is any better to gain that enormous value with probability p i than it is to gain an ever-so-slightly more probable shot at tiny value . So those lotteries are not just made worse by those tiny changes in probability or payoffs -they are made far worse (in an intuitive sense). That is the level of sensitivity we must accept if we reject Fanaticism and maintain Scale Independence. But such extreme sensitivity in our evaluations of lotteries seems absurd. Evaluations of lotteries should not change quite so wildly -so discontinuously -with arbitrarily tiny changes in probability or payoff. So I do not think this sensitivity -this horn of the dilemma -is any less absurd than what we faced earlier. I should briefly note that, beyond its intuitive implausibility, this sensitivity will likely lead to practical difficulties. The probabilities that human agents have access to are subjective degrees of belief and evidential probabilities. In practice, neither sort of probability can be pinned down with arbitrary precision, at least not with our merely finite capacity for calculation.  28  But we'll sometimes need arbitrary precision to compare lotteries, at least by this horn of the dilemma. If not, in cases like that immediately above, we would have no idea whether a lottery is as good as L 3 or as bad as L 1 with its imperceptibly lower probability of success. We must either require (unrealistically) that agents be arbitrarily precise in their probabilities  29  , or else accept that in practice our decision theory will be unable to tell agents that (at least some of) these lotteries are better than others. This is a dilemma we must face if we reject Fanaticism: we must accept either scale dependence, by which our judgements of lotteries can vary without any structural reasons for doing so, or that those judgements are sensitive to imperceptibly small differences in either probability and value. Both options seem absurd to me, indeed more absurd than Fanaticism itself. for what we ought to do (cf Mogensen 2021). This seems a major disadvantage for a moral view.  32  Even on other axiologies, including averageism, we face similar (albeit slightly less compelling) objections. See Footnote 35.  33  One such proposal is expected utility theory with a utility function that is concave and/or bounded (e.g.,  Arrow 1971 ). As Beckstead and Thomas (n.d.: 15-16) point out, this results in comparisons of lotteries being strangely dependent on events that are unaltered in every outcome and indeed some irrelevant to the comparison. Background Independence says that we can take any two lotteries and add a constant value to every one of their possible outcomes, and this won't change their ranking. Suppose for now that Background Independence is false. Then there is some pair of lotteries, L a and L b , such that: L a is at least as good as L b ; and, if we added a certain constant b to the value of every one of their outcomes, that would change their ranking. Since we're assuming totalism, the value of each outcome in each lottery must represent the total aggregate of value that would result, including all valuable events across all of space and time. That includes the events that occurred in ancient Egypt. And we can assume that the same events occurred in ancient Egypt in every outcome of either L a or L b .  34  With those events included, L a is at least as good as L b . But what if we consider a different hypothetical pair of lotteries, identical to L a and L b except that the events of ancient Egypt were different? Would the ranking be any different, had events gone differently back then? Yes, it would, if Background Independence is false and if those events differed drastically enough. They might differ in such a way that they increased in value by b. Then the total value of every outcome would also be increased by b. And, given that L a and L b were the lotteries that violated Background Independence, we know that this changes their ranking. So our modified version of L a is no longer at least as good as the modified version of L b . Thus, if we deny Background Independence, we face a rather severe version of the Egyptology Objection: when choosing between two lotteries, which is better can vary depending on what events occurred in ancient Egypt, even if the choice doesn't change those events at all (cf Beckstead and Thomas n.d.).  35  To avoid this implication, we must at minimum accept Background Independence. I think we ought to accept it regardless -I find it highly plausible for much the same reasons as Scale Independence was. In any case, I'll assume for the remainder of this section that Background Independence holds. But even if we accept Background Independence, if we still reject Fanaticism then it turns out that a version of the Egyptology Objection still arises, albeit a less severe one. To see why, first recall the lotteries from the definition of Fanaticism: L risky and L safe . If we reject Fanaticism then, for some tiny enough probability > 0 and some cardinal value v, some lottery L risky (as defined below) is no better than L risky , no matter how big V is. L risky : value V with probability ; 0 otherwise L safe : value v with probability 1 According to Background Independence, it must also hold that the lottery L risky with any constant b added to every outcome's value is also no better than L safe with that same b added to every one of its outcomes' values. If we face a decision between L risky and L safe , it would not matter if events in ancient Egypt were vastly different -we'd compare the lotteries the same way with or without that additional value b. But consider a further pair of lotteries: L risky + B and L safe + B. These are obtained by modifying L risky and L safe such that, in all outcomes, the events of ancient Egypt are different but you are uncertain of exactly how different they are. We are no longer dealing with a fixed value b of which you are certain. We are still adding the same value to all of them, but you are uncertain of what value it is. And that uncertainty is described by the lottery B. To make these lotteries more concrete, perhaps you must decide between two lotteries in which the value of all present and future events match the lotteries L risky and L safe , but you are uncertain of the value of past events such as those in ancient Egypt. As a devout totalist, you must evaluate outcomes would be to increase the average value obtained by all persons in that outcome. And to add b to the value of every outcome in a lottery L would be to increase the average value in every outcome by b. We might imagine doing this by delivering a gift to every inhabitant of the outcomes of L, with that gift producing the same boost in value, b, for each recipient. Even by the lights of averageism it seems that, if L a is at least as good as L b , then delivering that gift to everyone shouldn't change the ranking -a modified L a with additional value b should still be at least as good as L b with additional value b. And so we will have an averageist analogue of the Egyptology Objection presented above. And this objection will still be worrying for averageists, even if it is not quite as devastating as its analogue is for totalists. \n 21 and lotteries based on the total aggregate of value across all of space and time. And even if you know that those past events will turn out the same way no matter what you do, your uncertainty over exactly how they would turn out is a part of your uncertainty over the total value of the outcome. So you must compare L risky + B and L safe + B; you cannot simply L risky and L safe . As it turns out, there are some possible lotteries B that lead us to say that L risky + B is strictly better than L safe + B, even though L risky isn't any better than L safe . And this judgement, that L risky +B is better, is implied by even the extremely weak principle of Stochastic Dominance introduced above.  36  As a brief refresher, recall that Stochastic Dominance simply states that: if a lottery L a gives at least as high a probability as L b of resulting in an outcome which is at least as good as O, for every possible outcome O, then L a is at least as good a lottery; and if L a also gives a strictly greater probability of turning out at least as good as some O, then it is strictly better. Stochastic dominance is easy to spot graphically. To illustrate, consider the cumulative probability graphs of the following two lotteries, L 1 and L 2 . L 1 : value 1 with probability 1 2 ; value 0 otherwise L 2 : value 2 with probability  1  3 ; value 1 with probability 1 3 ; value 0 otherwise 1 2 0.5 1 L 2 L 1 V (O) L i (O O O ) Figure  1 : Cumulative probability graphs for L 1 and L 2 ; here and below, O O O denotes the set of outcomes in O as which O is at least as good.  36  We can reach a similar result, and thus a similar reductio, with axiologies other than totalism. Following from Footnote 35, we can adopt averageism and, instead of improving/worsening events in ancient Egypt, distributing an identical and morally valuable gift to each person in the world. You might be uncertain of the exact value of the gift -your uncertainty of its value might correspond to B. But still it seems that whether you distribute that gift is irrelevant to whether the risky lottery is better or worse than the safe one. \n 22 On a graph like this, one can easily see when Stochastic Dominance says that L 2 L 1 . Cumulative probability, on the vertical axis, is just the probability that the lottery produces an outcome no better than an outcome with some particular value. Meanwhile, Stochastic Dominance says that one lottery is at least as good as another if its probability of producing an outcome as good or better is just as high for all outcomes -or, equivalently, if the probability of an outcome as good or worse is at least as low. Then the probability that L risky + B turns out at least as good as value u is + B 3 (1 − ). And, with a bit of simple arithmetic  38  , we can see that this will be greater than the corresponding probability for L safe + B if the area B 2 is no greater than × B 1 . If B 2 is small enough compared to B 1 , then L risky + B has at least as high a probability of u or better as L safe + B does. And for it to be at least as high for all real u, we just need the interval between u and u − v to have a tiny enough area under it. And, for that to happen, we just need B to go down slowly enough as we approach −∞, and rise and fall quickly enough as we pass the peak of the curve. u − v u ; B B 2 B 1 B 3 V (O) L i (O O O ) And there are some probability distributions that have this property, including many Cauchy distributions.  39  And some is enough. There is some such lottery L risky + B that is better than L safe + B, even though L risky is no better than L safe . And so we have a new version of the Egyptology Objection, which I think is slightly milder than the previous version. You may be faced with a decision between two lotteries that, if you simply accepting Fanaticism and even more absurd than the Egyptology Objection.  42 43  Taking stock of this section, we have further reasons to accept Fanaticism. If we deny it then we are forced to accept a version of the Egyptology Objection wherein the judgements of agents like us will be sensitive to their uncertainty about what occurred in ancient Egypt, even though their choices do not affect ancient Egypt at all. If we reject both Fanaticism and Background Independence, then we face an even more severe version of the Egyptology Objection: how such agents should compare their options will be sensitive to not just their uncertainty but also what actually occurred in ancient Egypt. And even if we accept Background Independence (as I think we should), if we try to reject Fanaticism then a bizarre form of inconsistency arises in how agents should compare their options -agents are sometimes forced to make judgements which are inconsistent with anything they might learn to resolve their uncertainty about far-off, unrelated events in the ancient Indus Valley. Any one of these implications -let alone any two of them -is absurd. In light of these, it seems a lot less appealing to deny Fanaticism.  42  Continuing from Footnotes 35 and 36, an analogue of the Indology Objection applies to averageism. (Analogues can also be constructed for other axiologies.) You might be facing two lotteries -one risky and one safe -in which you know that, whatever the outcome, an identical gift will be given to everyone in the world. And you're uncertain of just how much value that gift will bring, with your uncertainty corresponding to B. Then, as we saw above, there are some probability distributions for B that mean that you will judge L risky + B as strictly better than L safe + B; but, if you opened the gift yourself and discovered what was inside, you already know that L risky with the additional value b would be no better than L safe with the additional value b. And this seems absurd, just as its analogue was for the totalist.  43  A closely related objection to rejecting Fanaticism is that, even if we do reject it, we will not be spared from many verdicts that seem deeply fanatical. As we saw in the case above, with the right uncertainty B about the value of unaffected events you must judge L risky + B better than L safe + B. In practice, will we often have the right uncertainty about unaffected events to lead us to judge otherwise extremely risky lotteries as better than otherwise safe ones? Tarnsey (n.d.: §6) argues that we often will, for an epistemically rational agent and for a wide range of cases, including many resembling L risky and L safe (where the probability is at least 1 in 10 9 , and likely also a lot lower). His reasoning for this is that the moral value of distant events in our universe will be roughly correlated with the number of inhabited planets in the universe; and our cosmological evidence and the Drake equation together motivate a probability distribution over that number of inhabited planets which is relevantly similar to the above B. But even apart from that reasoning, an epistemically rational agent would still have some uncertainty over which probability distribution best characterises our cosmological evidence. So they would still put some non-zero credence in distributions of the right sort. And this is enough -when we take the probability-weighted average of many such distributions to get our overall distribution, it will then have the right properties (see ibid.: 37-9 for details). Given this, Stochastic Dominance by itself will lead us to making judgements which seem fanatical. So we would gain little from rejecting Fanaticism. 38 + B 3 (1 − ) ≥ B 2 + B 3 ⇔ B 3 + (1 − B 3 ) ≥ B 2 + B 3 ⇔ (B 1 + B 2 ) ≥ B 2 ⇔ B 2 ≤ B 1 + B 2 , for which B 2 ≤ B 1 is a sufficient condition. \n 7 Conclusion These are just some of the compelling arguments for accepting Fanaticism and, with it, fanatical verdicts in cases like Dyson's Wager. There are other compelling arguments too -for instance, all of the arguments for the stronger claims of expected value theory, or for expected utility theory with an unbounded utility function. A fortiori, these also justify Fanaticism. For a sampling of these, I refer interested readers to Feller (1968), Ng (2012),  Briggs (2015) , and Thoma (2018). Some philosophers still reject expected value theory (e.g., Buchak 2013, Smith 2014, Monton 2019, Tarsney n.d.), and presumably the arguments that imply it. But this is not enough to escape Fanaticism. As I have demonstrated here, there are compelling arguments in its favourarguments stronger than many of those for expected value theory. To recap, if we deny Fanaticism then we must deny either Minimal Tradeoffs or transitivity, and accept the counterintuitive verdicts which follow. Likewise, we must accept either: violations of Scale Independence; or absurd sensitivity to the tiniest differences in probabilities and value. So too, to deny Fanaticism, we must accept a version of the Egyptology Objection: for agents like ourselves, which judgement is correct can depend on our beliefs about what occurred in far-off, unrelated events, such as those in ancient Egypt. And, unless we wish to accept an even more severe version of the Egyptology Objection, we also face the Indology Objection: we sometimes ought to make judgements that we know we would reject if we learnt more, no matter what we might learn. Given all of this, it no longer seems so attractive to reject Fanaticism. As it turns out, the cure is worse than the disease. I would suggest that it is better to simply accept Fanaticism and, with it, fanatical verdicts such as that in Dyson's Wager. We should accept that it is better to produce some tiny probability of infinite moral gain (or arbitrarily high gain), no matter how tiny the probability, than it is to produce some modest finite gain with certainty. All of this has implications for normative decision theory more broadly, at least dialectically. Philosophers often reject expected value theory because it implies Fanaticism, or because it implies fanatical verdicts in cases like Dyson's Wager (e.g., Monton 2019, Tarsney 2020). The arguments I have given here suggest that this rejection is a little hasty. Doing so invites greater problems than it solves. So I, for one, am thankful that expected value theory does imply Fanaticism. We have little choice but to accept Fanaticism, so we might as well accept expected value theory too.  44   If an outcome O ∞ contains infinitely more value than others -as when we create infinitely many blissful lives in Dyson's Wager -then it may fall beyond the scope of V . Then V (O ∞ ) won't be defined, but still O ∞ O O for each finitely-valued outcome O. And that's fine -a real-valued total value function V won't represent O in all cases, but it will do just fine in purely finite cases. I'll use O R ⊂ O to denote any given set of outcomes over which O admits a real-valued representation V . The arguments that follow can be run entirely in terms of O R .What we are really interested in is lotteries. I'll assume that the relevant aspects of a lottery can be fully described by a (real-valued) probability measure L on a set of outcomes. I'll also assume that this probability represents either the agent's subjective degree of confidence or the evidential probability of each outcome arising. Almost all of what follows can be read in terms of either. With minor changes, what follows could also be read in terms of objective chance or whatever notion of probability you consider relevant for decision making. Whichever notion it is, let L be the set of all lotteries on O, and L R the set of all lotteries on a given subset O R . So each lottery L is a function which maps sets of possible outcomes to some value in the interval [0,1], and that function must obey the standard probability axioms.To keep things brief, I'll use the following shorthand. When we are interested in the probability of a single outcome, I'll shorten L({O}) to L(O). When a lottery L gives a certainty of an outcome O (when L(O) = 1) I'll often write L as O. To represent a lottery with the same probabilities as L, but outcomes which have precisely k times the value (on the same cardinal representation), I'll write k • L. 14 And I'll write L a + L b to represent the lottery obtained fromadding together the values of the outcomes of lotteries L a and L b , run independently. 15   \n Background Independence: For any lotteries L a and L b and any outcome O , if L a L b , then L a + O L b + O . Recall that the sum of two lotteries L+O is simply the lottery you get if you run each lottery independently and sum up the values of their outcomes. But, here, O is a very simple lotteryit is just the outcome O with certainty. That outcome has some cardinal value; call it b. The lottery L a + O is simply the lottery L a with value b added to the value of each outcome. So \n Figure 5 : 5 Figure 5: Probability distribution of B \n \n Stochastic Dominance overwhelmingly plausible. In this setting, it is far weaker than (but necessary for) expected value theory. But, unlike the stronger theory, it does not rule out risk aversion, nor Allais preferences. To deny it is to accept either: that you can swap an outcome L infinite and L safe should look familiar -these match the lotteries you must choose between in Dyson's Wager, if we represent the outcome in which one life is saved with value 1 and the outcome in which no lives are saved with value 0. By Fanaticism, L risky is better than L safe , at least for large enough finite V . And Stochastic Dominance implies that L infinite is better than L risky . 19  4 A continuum argumentYou might argue for Fanaticism in the following way: expected value theory is true; expected value theory implies Fanaticism; therefore, Fanaticism is true. And likewise for verdicts which seem fanatical, like that which the expected value theorist must accept in Dyson's Wager. Such verdicts are rarely defended on any grounds other than 'That's what expected value theory says'.But we can do better than this. Fanaticism is far weaker than expected value theory in all its strength (and weaker even than expected utility theory), so it should be easier to justify. in a lottery for a better one without making the lottery better; or that evaluations of lotteries are dependent on something other than the values of the outcomes and their probabilities. Both are implausible. 17  And Stochastic Dominance in a setting like this is fairly uncontroversial among decision theorists, including some who reject expected value theory. As far as I know, no serious proposal has been made in normative decision theory which violates Stochastic Dominance (at least in this setting). 18  And Stochastic Dominance ties together the fates of Fanaticism and the fanatical verdict inDyson's Wager. Recall the lotteries L safe and L risky from the definition of Fanaticism. In L safe , we can set v = 1. And define L infinite as follows. L infinite : an outcome containing infinitely many blissful lives with probability ; value 0 otherwise L risky : value V with probability ; value 0 otherwise L safe : value 1 with probability 1 In this and the next two sections, I'll give four arguments for Fanaticism but not for expected value theory. Here is the first, which originates in Beckstead (2013: 139-47) and reappears in Beckstead and Thomas (n.d.: 4-6). \n 14lotteries L 1 and L 2 above; their outcomes could just as easily have been represented on another scale such that they had values k, 2k, and 0 -the same values as k • L 1 and k • L 2 had on the previous scale. On that new scale, we of course still say that L 1 is better than L 2 if and only if we said so on the original scale, else we would be inconsistent. But why not judge k • L 1 and k • L 2 \n As a brief refresher, it goes like this. Suppose you are making a moral decision here and now in the 21st Century, and the available outcomes differ only by some small-scale changes in the very near future. Your actions will not change what happens in distant galaxies or what happened in, say, ancient Egypt. Then, intuitively, what you ought to do should depend only on the events altered by your choice. It should not depend on events in distant galaxies or in ancient Egypt, at least not if your actions do not change them. Intuitively, it seems absurd that those unaltered, remote events make any difference to your evaluation. 31  But, under some axiologies, what you ought to do in such cases is dependent on unaltered, remote events. Take a (standard, welfarist) averageist view. Here and now, should you bring an additional person into existence? It depends -will that additional life raise or lower the average value of lives that ever exist? Sometimes yes, sometimes no, depending on what actually happened in ancient Egypt, in distant galaxies, and everywhere else. Averageism will thus imply that what happened in ancient Egypt can affect whether it is better to now bring an additional person into existence or not. And, further, to be confident of which is the better outcome, you You will most easily fall prey to it if you reject Background Independence. (And some proposals do.) 33  31 Such dependence brings practical problems too. If what you ought to do is dependent on such events -not just in the remote future but also the remote past, of which we are often clueless -then we have little guidance may need to do some serious research into Egyptology (along with plenty of other historical studies). But this is implausible. By intuition, we can ignore events that are unaffected by our actions and took place millennia ago, as we do under axiologies like totalism, critical-level views, prioritarianism, and aggregative person-affecting views (and any other axiology which admits an additively separable representation). To me, this is one of the main appeals of those views.But even if we accept an axiology like totalism, we may still face much the same objection in practice. 32  Some methods of comparing lotteries give rise to an updated Egyptology Objection, including all of those that deny Fanaticism. That's right: to deny Fanaticism, you must not only endure the costs detailed in the previous two sections, you must also fall prey to a version of the Egyptology Objection. (As we'll see below, you may also face another even more serious objection.) \n So Stochastic Dominance will say that L 2 L 1 if and only if L 2 's cumulative probability is always as low or lower than that of L 1 , as it is on this graph. And, here, L 2 often has strictly lower cumulative probability. So Stochastic Dominance says it is strictly better than L 1 . 37  But accepting Stochastic Dominance doesn't rule out denying Fanaticism. L risky doesn't stochastically dominate L safe , as illustrated below. Cumulative probability graphs for L risky and L safe Sometimes one is higher; sometimes the other. So Stochastic Dominance remains silent. And it's a good thing it does -to deny Fanaticism, we must not say that L safe is better than L risky .But what about L risky + B and L safe + B? Stochastic Dominance is not necessarily silent about that comparison. In particular, suppose that the lottery B looks like this. (I'll describe An example of background uncertainty B -a Cauchy distribution When B looks like that, we obtain the following graphs for L risky + B and L safe + B. Crucially, the graph for L safe + B is never higher than that for L risky + B, and sometimes it is strictly lower. So Stochastic Dominance says that L safe + B is strictly better. Cumulative probability graphs for L risky + B and L safe + BBut how does this happen?For Stochastic Dominance to hold, we need L risky + B to have at least as high a probability as L safe + B of turning out at least as good as an outcome with value u, for all possible values u.Start with L risky + B. We know that L safe just gives one value (v) with certainty. So the probability that L safe + B gives value u or better is just the probability that B gives value u − v or better. (This corresponds to the area B 2 + B 3 on the graph below.) And then, for L risky + B, we'll get value at least u either if L risky gives value V or if B gives at least value u. Denote the probability that B gives value at least u by B 3 (corresponding to that area on the graph below). 1 L i (O O O ) B V (O) ; 1 Figure 3: 1 L i (O O O ) L i (O O O ) L safe L risky L safe + B V (O) 1 L risky + B ; ; V (O) Figure 4: So take any real value u < V . What's each lottery's probability of doing at least that well? Figure 2: the required properties of B below.) \n exclude the value of events that occurred in ancient Egypt, correspond to L risky and L safe . And which lottery is better overall? That depends on what happened in ancient Egypt, even though you know the same events will have happened there no matter which you choose. If the events of ancient Egypt have value 0 (or b), then the risky lottery is no better than the safe one. But if ancient Egypt might have contained greater or lesser value, and you aren't certain how much, then it may be that the risky lottery is better after all. Much like under the classic Egyptology Objection, your evaluation is sensitive to facts that seem irrelevant. But, unlike the original version, the evaluation is not merely sensitive to what actually occurred there; instead, it is sensitive just to your uncertainty about what happened there. That is perhaps less devastating a problem, as an agent may know enough about events in ancient Egypt to constrain B to a less troublesome distribution. Still, for judgements to be sensitive to the agent's beliefs about events in ancient Egypt at all is still, I think, rather absurd -absurd enough that we should be willing to accept Fanaticism to avoid it.Italy, and Egypt. So there is likely plenty left to learn in Indology.Take the same L risky + B and L safe + B from above -lotteries such that the former is strictly better, and yet L risky is no better than L safe . Such lotteries must exist, if Fanaticism is false. Butsuppose that you are no longer uncertain about the events of ancient Egypt; instead, you are uncertain of what occurred in the ancient Indus Valley. The lottery B in L risky + B and L safe + B represents only those events. But, in Indology, there is a great deal more work that can be done! Had you the option, you could spend many years researching the ancient Indus Valley, peeling back that uncertainty and narrowing B down. Let's suppose that with enough years of intensive research you could eventually remove that uncertainty entirely and determine exactly which value to assign to the events of the ancient Indus Valley -which value b the lottery B actually results in. If you did first do that research, you would no longer need to choose between L risky + B and L safe + B. Instead, you would choose between L risky with b added to the value of every outcome and L safe with b added.Recall Background Independence from above. If it holds, then L risky with any b added to every outcome is better than L safe with the same b added if and only if L risky is better than L safe . And if it doesn't, as we saw above, we face an even more severe version of the Egyptology Objection. So we can safely assume that Background Independence holds. You cannot. You must accept that L risky + B is strictly better, even though you can predict with certainty that you would change your mind if you knew more. To be able to change your mind you must go through those years of gruelling research and digging, even though you already know which judgement you would conclude to be correct. 41  This ambivalence seems deeply irrational. Surely we can sidestep those years of research into how B turns out, and make the judgement required by every possible value of b. Surely rationality requires that we do so, rather than require that we do not. But, if we deny Fanaticism, we must accept this inconsistency -an inconsistency which, to me, seems far more absurd than 41 cf van Fraassen's (1984) principle of Reflection and its application byWhite (2010:17-8). That form of 40  6.2 The Indology ObjectionBut it gets worse. If you reject Fanaticism, you may also face what I'll call the Indology Objection.For context, Indology is the study of the history and culture of India. In particular, classical Indology includes the study of the Indus Valley Civilisation, a Bronze Age civilisation which lasted from 3300 BCE to 1300 BCE. Althought less well-known and understood than its contemporaries -ancient Egypt and Mesopotamia -it spanned a greater area than either and rivalled them in population, technology, and urban infrastructure (see Wright 2009) . We happen to know even less about what happened in the ancient Indus Valley than in ancient Egypt -archaeological research and excavations of key sites in India began centuries later than similar work in Britain, From Background Independence we know that, whatever you might uncover in your research, you would conclude that the risky lottery is no better than the safe lottery. For any value of b you might pin down, you'll establish that it's no better to take the risky lottery. To judge otherwise would be to accept Fanaticism. So you know what judgement you would make if you simply learned more, no matter what it is you would actually learn. So why do the many years of research? Why not just update your judgement now and say that L risky + B is no better than L safe + B?Reflection implies that, if you know that you would set the probability of some hypothesis to p whether you learned either proposition P or not-P , then its current probability should already be p. The version I am assuming is somewhat similar but need not be quite as strong. \n\t\t\t In effect, this is a core point of Bostrom's (2009)  objection. Meanwhile, Bostrom (2011) , Beckstead (2013) , and Tarsney (n.d.) all treat Fanaticism as implausible for, it seems, the same reason. \n\t\t\t The differences in value for two pairs of outcomes are treated as finite precisely when one can be represented as a real multiple of the other. \n\t\t\t For compelling arguments in favour of Stochastic Dominance, see Easwaran (2014) and Bader (2018). 18 Schoenfield (2014) rejects a similar principle, but only for lotteries over outcomes for which we don't have a complete ordering. With our complete O relation on O, even Schoenfield would endorse Stochastic Dominance here. \n\t\t\t For readers less sympathetic to totalism, much the same argument can be made in terms of any of the axiologies mentioned in the Introduction. For instance, if we assume averageism, we can start with L 0 , by which the average value of all lives will be 1 with near certainty, and 0 otherwise. L 1 can be the lottery which gives probability 0.99999 of an outcome in which the average value of all lives will be 10 10 , or 0 otherwise. We can rapidly scale up the average value of lives just as we scaled up the total values above, while gradually scaling down the probability of non-zero value. Eventually, we reach a lottery L n with probability of some astronomical average value, or 0 otherwise, but transitivity says that it must be better than L 0 . \n\t\t\t For readers unsympathetic to totalism, these payoffs could just as easily represent: the average values of lives in each outcome; the priority-weighted total value in each outcome; the total value in the lives of all persons who would exist regardless of the agent's choice; and so on. \n\t\t\t An alternative to endorsing Scale Independence is to adopt a variant of totalism that assigns more than just a cardinal value to outcomes -perhaps a version of totalism that values outcomes on a ratio or unit scale. That version would be far less parsimonious than the version I have been using, and I cannot think of any compelling independent justification for it. It seems at least a considerable cost to adopt such a version of totalism to avoid endorsing Scale Independence and the implication of Fanaticism. \n\t\t\t This difference is analogous to the difference between weak and strong superiority in Arrhenius & Rabinowicz (2015) . 27  The comparison won't be the same for values 0 and below, but that's okay. Such values cannot be scaled up to v; multiply 0 or any negative number by any real positive value you want and you still won't reach v. \n\t\t\t Egyptology and IndologyBut it gets worse. There are even greater costs you must pay to reject Fanaticism: you face either one or both of the following objections, each of which is intuitively absurd.6.1 The Egyptology ObjectionHailing from population ethics, the Egyptology Objection is a classic argument against various axiologies, including averageism, egalitarianism, and maximin. 30  28 For discussion of why we cannot require epistemic agents to settle on precise probabilities, see Schoenfield (2012) and Joyce (2011) . For discussion of how to evaluate lotteries with imprecise probabilities, see Seidenfeld(2004), Huntley, Hable & Troffaes (2014), Bradley & Steele (2015), and Bradley (2015) . On any of the methods suggested, a combination of imprecise probabilities and sensitivity to the degree described above will lead to indifference among many of the lotteries L 1 to L 4 listed above or else indeterminacy of which is better than which. At least in that case, with its vast differences in value and only small differences in probability, it *** 29  In Smith's (2014:471) terms, our theory would then be 'intolerant' of human imprecision, which he considers a major problem. 30  It appears earliest in MacMahan (1981: 115) but is often attributed toParfit (1984: 420). \n\t\t\t If there is some pair of lotteries L a or L b that violate Background Independence, then there is also some pair that violate it and have the exact same events occur in ancient Egypt. To obtain such a pair, simply take L a and L b and replace each of their outcomes with an outcome of equal total value but an identical history of ancient Egypt. 35  This new version of the Egyptology Objection, along with the other objections described in the rest of this section, can also be applied under axiologies other than totalism. It will just look a little different.Suppose that (standard, welfarist) averageism is the correct axiology. Then to add b to value of an outcome \n\t\t\t This relationship between Stochastic Dominance and cumulative probability would break down if some outcomes were incomparable to others. There would then be a difference between O a O O b and the negation of O a ≺ O O b . But, fortunately, totalism gives us a total preorder of O, so we can sidestep this complexity. \n\t\t\t The result holds for any Cauchy distribution with a 'scale factor' of at least v . See Tarsney (n.d.: §5) and Pomatto et al (2018). \n\t\t\t This version of the Egyptology Objection loosely resembles one which has been levelled against decision theories that violate the Sure-Thing Principle (by, e.g., Briggs 2015) . It turns out that following such a decision theory in each of several consecutive, independent decisions will sometimes be inconsistent with following the theory when deciding on a strategy for all of those decisions at once. Essentially, the combination of two lotteries can be evaluated quite differently to how we evaluate those lotteries separately. And that is troubling. It is unclear whether we should evaluate such decisions together or apart, globally or individually. But the correct verdict depends on which way we do it. \n\t\t\t This paper has benefitted from the input of many brilliant and generous colleagues. I am indebted to Alan", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Hayden-Wilkinson_In-defence-of-fanaticism.tei.xml", "id": "5b1d592c4d34b8a7993da9d9b990cb3b"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "An intelligent agent embedded within the real world must reason about an environment which is larger than the agent, and learn how to achieve goals in that environment. We discuss attempts to formalize two problems: one of induction, where an agent must use sensory data to infer a universe which embeds (and computes) the agent, and one of interaction, where an agent must learn to achieve complex goals in the universe. We review related problems formalized by Solomonoff and Hutter, and explore challenges that arise when attempting to formalize analogous problems in a setting where the agent is embedded within the environment.", "authors": ["Nate Soares"], "title": "Formalizing Two Problems of Realistic World-Models", "text": "Introduction An intelligent agent embedded in the real world faces an induction problem: how can it learn about the environment in which it is embedded, about the universe which computes it?  Solomonoff (1964)  formalized an induction problem faced by agents which must learn to predict an environment which does not contain the agent, and this formalism has inspired the development of many useful tools, including Kolmogorov complexity and Hutter's AIXI. However, a number of new difficulties arise when the agent must learn about the environment in which it is embedded. An agent embedded in the world also faces an interaction problem: how can an agent learn to achieve a complex set of goals within its own universe?  Legg and Hutter (2007)  have formalized an \"intelligence measure\" which scores the performance of agents that learn about and act upon an environment that does not contain the agent, but again, new difficulties arise when attempting to do the same in a naturalized setting. This paper examines both problems. Section 2 introduces Solomonoff's formalization of an induction problem where the agent is separate from the environment, Research supported by the Machine Intelligence Research Institute (intelligence.org). Published as Technical report  2015-3.  and Section 3 discusses troubles that arise when attempting to formalize the analogous naturalized induction problem. Section 4 discusses Hutter's interaction problem, and Section 5 discusses an open problem related to formalizing an analogous naturalized interaction problem. Formalizing these problems is important in order to fully understand the problem faced by an intelligent agent embedded within the universe: a general artificial intelligence must be able to learn about the environment which computes it, and learn how to achieve its goals from inside its universe. Section 6 concludes with a discussion of why a theoretical understanding of agents interacting with their own environment seems necessary in order to construct highly reliable smarter-than-human systems.  Solomonoff (1964)  posed one of the earliest and simplest descriptions of a problem in which an agent must construct realistic world-models and promote correct hypotheses based on observations, performing reasoning akin to scientific induction. Intuitively, the problem considered by Solomonoff runs as follows: the universe is separated into an agent and an environment. Every turn, the agent observes one bit of output from the environment. The task of the agent is to, in each turn, predict its next observation. \n Solomonoff's Induction Problem To formally describe the agent's performance, it is necessary to decide what counts as a possible environment, then to decide how to measure how well an agent predicts an environment, and then to choose the distribution over environments against which the agent will be scored. What counts as a possible environment? In Solomonoff's formalization, the goal is to consider hypothetical agents which can learn an arbitrarily complex environment, and so Solomonoff chooses the set of environments to be anything that is computable. This can be formalized by defining the set of all environments as the set T of all Turing machines with access to an advance-only output tape. How are an agent's predictions scored? Consider an environment M ∈ T where M n denotes the n th bit on the output tape of M . Let an agent A be a function which takes a string M ≺t of observations made before turn t, which outputs a prediction of M t in the form of a rational number interpreted as the probability that M t = 1. For convenience, define A t := A(M ≺t ). To score A against M on all time steps, it is necessary to account for the fact that M may eventually stop outputting bits; define M to be the last turn in which M outputs a bit (this value may be ∞). Then A may be scored against M using standard logarithmic loss: 1 S M (A) := M t=1 M t log(A t ) + (1 − M t ) log(1 − A t ). (1) Against which distribution over Turing machines should the agent be scored? The answer determines which agents are considered to be \"good predictors.\" If the agent is to be evaluated against its ability to learn one specific environment, the trivial distribution containing only that environment may be chosen-but then the high-scoring agents would be agents which have that environment hard-coded into them; this would hardly be a problem of learning. The choice of distribution defines the manner in which agents must be biased to achieve a high score: how should predictors be biased? The natural answer comes in the form of an intuition canonized by William of Ockham seven hundred years ago: predictors in the real world do well to prefer the simplest explanation which fits the facts. There are exponentially more possible explanations of increasing complexity (e.g. 2 N N -bit explanations) and so any particular explanation of greater complexity should have less probability. Thus it seems natural to score the agent according to a distribution in which simple environments have greater weight than complex environments. The most natural way to define a simplicity distribution over Turing machines is to fix some universal Turing machine U , 2 and assign probability 2 − M to each Turing machine M , where M is the number of bits needed to specify M to U . Now Solomonoff's induction problem may be fully described: An environment is any Turing machine M with an advance-only output tape. An agent A is a function which takes an output history and produces a rational number interpreted as the probability that the next observation will be 1. The agent is scored according to S M (A) against a simplicity distribution. Formally, Solomonoff's induction problem is the problem of maximizing the \"Solomonoff induction\" score SI(A) := M ∈T 2 − M • S M (A). (2) 1. This score may not converge, in the infinite case, but it is nevertheless useful for comparing agents. 2. U must be chosen such that M ∈T 2 − M = 1. Like many good problem descriptions, this one lends itself readily to an idealized unbounded solution, known as Solomonoff induction: Solomonoff induction. The agent starts with a simplicity distribution over T . Upon receiving the n th observation o n , it conditions its distribution on this observation by removing all Turing machines that do not produce n bits, or that do not write o n as the n th bit on their output tape. It then predicts that the (n + 1) th bit is a 1 with probability equal to the measure on remaining Turing machines which write 1 as the (n + 1) th bit on their output tape. Indeed, it is in terms of this idealized solution that Solomonoff originally posed his induction problem  (Solomonoff 1964) . A Solomonoff inductor is a very powerful predictor. It can \"learn\" any computable environment:  Solomonoff (1978)  showed that given any computable probability distribution over bit strings, a Solomonoff inductor's predictions will converge to the true probabilities. With his induction problem, Solomonoff provides a full description of a scenario in which an agent must learn an arbitrarily complex computable environment separate from the agent. Insights from the induction problem have proven useful in practice: This problem became the basis of algorithmic information theory  (Hutter, Legg, and Vitanyi 2007) . The simplicity prior over Turing machines is the celebrated \"universal prior\"  (Solomonoff 2003) . Solomonoff induction is a crucial ingredient in  Hutter's AIXI (2000)  that solves the analogous universal decision problem, and many of Solomonoff's insights are present in the Legg-Hutter \"universal measure of intelligence\" (2007). Solomonoff's work served as the basis for Kolmogorov complexity  (Solomonoff 1960) , a powerful conceptual tool in computer science. Unfortunately, the prediction problem faced by agents acting in the real world is not Solomonoff's induction problem: it is a problem of an agent modeling a world in which the agent is embedded as a subprocess, where the agent is made out of parts of the world and computed by the universe. Formally describing this more realistic problem turns out to be significantly more difficult. \n The Naturalized Induction Problem In Solomonoff's induction problem, the agent and its environment are fundamentally separate processes, connected only by an observation channel. In reality, agents are embedded within their environment; the universe consists of some ontologically continuous substrate (atoms, quantum fields) and the \"agent\" is just a part of the universe in which we have a particular interest. What, then, is the analogous prediction problem for agents embedded within (and computed by) their environment? This is the naturalized induction problem, and it is not yet well understood. A good formalization of this problem, on par with Solomonoff's formalization of the computable sequence induction problem, would represent a significant advance in the theory of general reasoning. An analogous formalization of the naturalized induction problem must yield a scoring metric akin to SI(•), which scores an algorithm's ability to predict its environment. But what metric is this? There are at least three open questions of naturalized induction: First, given an algorithm, what is the set of all environments that the algorithm could have to induce? It would be strange to score the agent against all computable environments, as almost all Turing machines will not in fact embed that algorithm. Perhaps the set of environments could be defined with respect to the proposed algorithm, as the set of Turing machines which embed it. But how is that set defined? What does it mean for a Turing machine to \"embed\" an algorithm? Intuitions about embeddings have been difficult to formalize. If the naturalized induction problem is to capture the problem of an agent learning about the real world, then the set of environments must contain reality. The set of all environments, therefore, must be a set of \"possible realities\": what structure is this? Does the set of all Turing machines actually contain our universe? Currently, the Standard Model of physics seems computable to any desired finite precision. But then again, reality looked Newtonian to scientists in centuries past. If artificial agents are to be able to surpass their programmers' scientific knowledge, a formalization of intelligent learning should not presuppose the correctness of present-day physical science. Modern theories as to the nature of physical reality may turn out to be mistaken or incomplete, and an ideal reasoner must be able to adapt to such surprises. What set of possible environments definitely contains reality, in light of the potential for surprises? Second, given an environment drawn from this set of possible environments, how is the agent's ability to learn that environment scored? Are agents scored better for constructing new sensors? Are they scored better for finding some way to affect their environment so as to make it easier to predict? These are not questions that can be reduced to Solomonoff's prediction task. Formalizing inductive success is much more difficult when the environment can act on the agent's internals, and when the agent-environment boundary can shift over time. Questions of evaluation are further covered in sections 4 and 5. Third, given a set of possible environments and a scoring metric, what is the distribution against which an agent should be scored? As in Solomonoff's induction problem, this distribution must capture reality's bias towards simplicity, but defining a simplicity distribution over some set of \"possible realities\" may be nontrivial. Of course, answers to these questions would be impractical at best and almost certainly uncomputable, but they would yield conceptual tools by which practical programs implementing sufficiently advanced agents (that face the naturalized induction problem) could be evaluated. For example, a formalization of naturalized induction would likely shed light on questions about how a reasoner should let the fact that it exists affect its beliefs, and may further lend insight into what sort of priors an ideal reasoner would use. Unfortunately, it is not yet clear how to approach the problems outlined above.  3  Can Solomonoff induction be ported into a naturalized context? Perhaps, but the application is not straightforward. Even ignoring problems of ensuring that the environment has something corresponding to the \"turns\" and \"observations\" of Solomonoff's induction problem, Solomonoff's approach solves the problem by simply being larger than the environment: a Solomonoff inductor contains a distribution over all Turing machines, and one of those is, by assumption, the \"real\" environment. Solutions of this form don't apply when the agent is a subprocess within the environment. Computable approximations of Solomonoff induction can be limited to the consideration of only \"reasonably sized\" environments, but this does not much help. Imagine a Solomonoff inductor which only considers Turing machines which can be specified in length less than l and which run for at most t steps between each turn: 4 this inductor would run for more than t steps per turn, and therefore the environment it is in would run for more than t steps per turn. The inductor would assign zero probability to its own existence! An agent embedded in an environment must reason about an environment that is larger than itself; this constraint is inherent to naturalized induction. Solutions will require agents to reason about environments which they cannot compute. Reasoning of this form is known as \"logically uncertain\" reasoning, and it may be possible to port a logically uncertain variant of Solomonoff into a naturalized context. However, a satisfactory theory of reasoning under logical uncertainty does not yet exist. (For further discussion, see  Soares and Fallenstein [2015a] .) \n Hutter's Interaction Problem Even a full description of naturalized induction would not completely describe the problem faced by an intel-ligence acting in the world. Real must not only predict their environment, but act upon it. With this in mind, Hutter (2000) extends Solomonoff's induction problem to an interaction problem, in which an agent must not only learn the external environment but interact with it. Hutter's interaction problem runs similarly to Solomonoff's induction problem: the universe is separated into agent and environment, and the agent gets to observe the environment through an input channel. But now, an \"output channel\" is added, which lets the agent affect the environment by one \"action\" per turn. As before, a formalization requires answers to the questions of (  1 ) what counts as an environment; (2) how an agent is scored on each environment; and (3) against which distribution over environments the agent is scored. In Hutter's interaction problem, the first and last answers follow readily from Solomonoff induction, with some minor tweaks. It is the answer to the second question, of scoring, where Hutter's interaction problem provides new insight. Again, the set of all environments can naturally be defined as the set T of all Turing machines. However, instead of having an advance-only output tape, environments are now Turing machines which take an observation/action history and compute the next observation to be sent to the agent. That is, fix some countable set O of observations which can be sent from environment to agent, and some countable set A of actions which can be sent from agent to environment, and then consider Turing machines which take a finite list of actions and compute a new observation O. An agent, then, is any function which takes a finite list of observations and computes a new action A. 5 Again, the distribution over environments will be the \"universal\" simplicity distribution (with respect to some fixed universal Turing machine U ). It remains to decide how an agent is scored: what counts as the \"success\" of an agent A interacting with an environment M ?  Hutter (2000)  formalizes interaction as follows. First, observations are defined such that one part of the observation is a reward; that is, elements of O are tuples (o, r) where r is a rational number between 0 and 1, and o is additional observation data. Let M A t ∈ O denote the t th output of the machine M when interacting with A, and let A M t ∈ A denote the 5.  Hutter (2000)  uses a generalization in which both agent and environment may be stochastic; in this case it is necessary for agent and environment to receive a history of both observations and actions. In the deterministic version, used here for ease of exposition, the agent (environment) does not need to be told the history of actions (observations) because past actions (observations) may simply be recomputed. t th action of A when interacting with M . That is, M A 1 := M () A M 1 := A(M ()) M A t := M (A M ≺t ) A M t := A(M A t ). Let r A t denote the reward part of the observation M A t . Restrict consideration to the set T r of environments where rewards converge, that is, to environments M such that 0 ≤ M t=1 r A t ≤ 1 for all agents A. The total rewards observed by an agent A interacting with M are then used to score the agent, that is, define R M (A) := M t=1 r A t . (3) This function measures the ability of A to learn and manipulate M in order to maximize observed rewards over time. This choice of scoring mechanism yields a full description of Hutter's interaction problem: it describes a setting where an agent must interact with an environment in order to learn and maximize rewards. Indeed, this scoring metric is used to define the \"universal measure of intelligence\" of  Legg and Hutter (2007) : LH(A) := M ∈Tr 2 − M • R M (A). (4) We refer to the problem of finding agents which score highly according to LH(•) as Hutter's interaction problem. As before, this problem description lends itself readily to an idealized solution. In this case, the solution is Hutter's AIXI  (2000) , which in fact was the mechanism by which Hutter first posed the interaction problem: AIXI. The agent starts with a universal prior, which it keeps consistent with observation using Solomonoff induction (modified in the natural way for this problem). AIXI chooses its action as follows: it has some fixed time horizon h, and considers all possible sequences of h actions. It computes the expected reward (according to its distribution over environments) for each sequence, and then outputs the first action in the sequence that leads to highest rewards. AIXI is an incredibly powerful and elegant agent. As noted by  Veness et al. (2011) , AIXI captures \"the major ideas of Bayes, Ockham, Epicurus, Turing, von Neumann, Bellman, Kolmogorov, and Solomonoff\" in a single equation. Barring a few minor quibbles, 6 AIXI fully 6. The finite time horizon of AIXI is both arbitrary and a solution to Hutter's interaction problem: while AIXI is uncomputable, it demonstrates the method by which a high LH(•) score may be attained. Indeed, computable approximations of AIXI have already yielded interesting results  (Veness et al. 2011) . If the problem faced by intelligent agents acting in the real world to achieve goals was characterized by Hutter's interaction problem, then this problem description would fully characterize the problem of constructing smarter-than-human systems, and the problem of general intelligence would be reduced to one of approximating AIXI. However, Hutter's interaction problem does not capture the problem faced by an agent acting in the real world to achieve goals. Rather, it characterizes the problem of an agent attempting to maximize sensory rewards from an environment that can only affect the agent via sensory information. While this problem description has yielded many insights, the distinction is important: the simplifying assumptions of Hutter's interaction problem mask a number of difficult open problems. \n The Agent is Not Separate from the Environment Hutter's interaction problem, like Solomonoff's induction problem, assumes an impregnable separation between the agent and the environment. In Solomonoff's case, there is (figuratively speaking) a small slit through which the environment feeds sensory information to the agent. Hutter adds a second slit, through which the agent outputs motor signals to the environment. However, the separation remains otherwise complete. Thus, the questions of naturalized induction remain unanswered, and Hutter's interaction problem yields little new insight there. For this reason, Hutter's interaction problem cannot capture certain realistic scenarios that intelligent agents may actually face: the Legg-Hutter measure of intelligence is ill-defined in any situation where the universe cannot crisply be divided into \"agent\" and \"environment,\" when interactions cannot be crisply divided into \"input\" and \"output.\" For example, consider the following simple setting in which it matters that the agent is embedded within its environment: The Heating Up game. An agent A faces a box containing prizes. The box is designed to allow only one prize per agent, and A may execute the action P to take a single prize. However, there is a way to exploit the box, cracking it open and allowing A to take all ten prizes. A can attempt to do this disconcerting: for any time horizon h, consider an environment with a button that gives −1 when pressed, pays +10 h steps thereafter, and pays out −100 on the step after that. AIXI with time horizon h, after learning that this is the environment, presses the button indefinitely. by executing the action X. However, this procedure is computationally very expensive: it requires reversing a hash. The box has a simple mechanism to prevent this exploitation: it has a thermometer, and if it detects too much heat emanating from the agent, it self-destructs, destroying all its prizes. If the agent heats up too much, it gets reward 0, no matter what action it takes. If it does not heat up too much, then it gets reward 1 for action P or reward 10 for action X. But the amount of heat generated by the agent, of course, is dependent upon which action the agent chooses. This scenario captures an important aspect of reality: a generally intelligent agent must be able to consider the consequences of overheating (along with many other consequences of being embedded within a universe). However, this scenario can't be modeled as an interaction problem. The Legg-Hutter measure of intelligence does not pit agents against scenarios such as these; there is no combination of M ∈ T r and A which captures this sort of problem. When evaluating an agent in a Heating Up game, the agent cannot be treated as separate from the environment. Rather, the agent must be located within the environment, and then somehow scored according to what it \"could have done.\" Is it possible for a clever agent to compute X without ever once getting too hot? This question depends upon the specific environment and upon the agent's specific hardware. This highlights a host of new \"naturalized\" questions: Given an environment that embeds an agent, how is the agent located in that environment? How are the actions that it \"could have taken\" identified? In Hutter's interaction problem, these questions are simplified away: the input and output channels are clearly demarcated; the environment is defined in terms of the agent's actions. AIXI, when considering the effects of various sequences of actions, can simply run a Turing machine on the considered action sequence; the behavior of the environment on that sequence of actions is well-defined. When an agent is embedded within an environment, the question is more difficult. For simplicity, consider a deterministic agent embedded in a deterministic environment. What does it mean to ask what \"would happen\" if the part of the environment labeled \"agent\" outputs something it doesn't? How is the counterfactual defined? Counterfactual reasoning is not yet well understood  (Soares and Fallenstein 2015b ). Hutter's interaction problem extends Solomonoff's induction problem to capture a critical aspect of the problem faced by intelligent agents: environments that can be altered by agent decisions. This yields many insights, but moving forward, a naturalized interaction problem is necessary: how can an agent learn and manipulate the environment in which it is embedded, to achieve some set of goals? It is this problem which would fully characterize the problem faced by intelligent systems acting in the real world. \n Goals Cannot Be Specified in Terms of Observation Ignoring the need for a naturalized interaction problem, Hutter's interaction problem still does not quite capture the problem faced by an agent which must learn and manipulate an environment to achieve some set of goals. Rather, it characterizes the problem faced by an agent which must maximize rewards, specified in terms of observations. But most sets of goals cannot be characterized in terms of the agent's observations! Consider an interaction problem in some approximation of reality where there is a crisp separation between \"agent\" and \"environment,\" where the input and output channels are clearly demarcated. The agent's input is a video stream, and rewards are only nonzero when there are smiling human faces on the video screen. This agent, if possessing of a high LH(•) score, will very likely gain control of its input stream, such as by placing a photo with many smiling faces in front of the camera and then acting to ensure that it stays there. Agents with high LH(•) scores are extremely effective at optimizing the extent to which their observations contain rewards; these are not likely to be agents which optimize the desirable feature of the world that the rewards are meant to serve as a proxy for. Rather, are likely to be agents which are very good at taking over their input channels. Reinforcement learning techniques, such as having the humans dole out rewards via some reward channel, would not solve the problem. Humans could attempt to prevent the agent from taking over its reward channel by penalizing the agent whenever they notice it performing actions that would give it control over rewards, and this may prevent the agent from executing those plans for a time. However, if the agent scores sufficiently high by LH(•), then once its dominant hypotheses about the environment agree that the humans are controlling the reward channel, it would act to mollify the programmers while searching for ways to gain a decisive advantage over them. If the agent is a sufficiently intelligent problemsolver, it may eventually find a way to wrest control of the reward channel away from the programmers and maintain it permanently  (Bostrom 2014, chap. 8) . Even faced with incredibly high-fidelity input channels designed to be expensive to deceive, LH(•) rewards agents that set up Potemkin villages 7 which trigger the reward using minimum resources. An agent optimizing a reward function only optimizes the actual goals if achieving the goals is the cheapest possible way to get the reward inputs. Guaranteeing such a thing is nigh impossible: consider the genetic search process of  Bird and Layzell (2002) , which, tasked with designing an oscillating circuit, re-purposed the circuit tracks on 7. A common idiom named after Gregory Potemkin, who set up fake villages to impress Empress Catherine II. its motherboard to use as a radio which amplified oscillating signals from nearby computers. Highly intelligent systems might well find ways to maximize rewards using clever strategies that the designers assumed were impossible, or that they never considered in the first case. A high LH(•) score indicates that an agent is extremely proficient at commandeering its reward channel. Therefore, this intelligence metric does not quite capture the intuitive notion of how well an agent would fulfill a given set of goals. There is no all-purposes patch for this problem within a sensory rewards framework. We do not care about what the agent observes; rather, we care about what actually happens. To evaluate the performance of an agent, it is not sufficient to look only at the inputs which the agent has received: one must also look at the outcomes which the agent achieves. \n Ontology Identification To evaluate how well an agent achieves some set of goals, it is important to measure the resulting environment history, not just the agent's observation history. In Hutter's interaction problem, an \"environment history\" is the combination of a Turing machine along with an observation/action history. But goals are not defined in terms of Turing machines and O/A histories; goals are defined in terms of things like money, or efficient airplane flight patterns, or flourishing humans. How do you measure those things, given a Turing machine and an O/A history? As a matter of fact, it is quite difficult to say what terms our goals are specified in. To leave aside problems of philosophy, and highlight the problem as it pertains to world models, let us imagine that our goals are simple and can be specified according to a structure that seems fairly objective in our environment: assume that agents will be evaluated by how much diamond they create in their environment, where \"diamond\" is specified concretely in terms of a specific atomic structure. That is, the score of an agent is the count of carbon atoms covalently bound to four other carbon atoms over time. Now the goals are given in terms of atomic structures, and the environment-history is given in terms of a Turing machine paired with an O/A history. How is the Turing machine's representation of atoms identified? This is the ontological identification problem. Whatever set is used for the set of all environments against which the agent is measured, it must be possible to inspect elements of that set and rate them according to our goals, and for that it is necessary to interpret features of the environment in terms of the ontology of the goals. This is an aspect of the naturalized interaction problem where Hutter's interaction problem sheds little insight. It seems intuitively plausible that any detailed model of reality, in an environment such as the real world diamond actually exists, must have some part of its internal structure which roughly corresponds to \"atoms.\" However, the problem is made more difficult by the fact that the ontology of the goals will not actually perfectly match the ontology of reality: how are the atoms identified in a model of reality which runs on quantum mechanics? The model (if accurate) will still have systems that correspond, in some fashion, to the objects we call \"atoms,\" much as the atomic model has systems corresponding to what we call \"water.\" However, the correspondence may be convoluted and full of edge cases. How can the ontology of the goals be reliably mapped onto the ontology of the model? de Blanc (2011) provides a preliminary examination of these questions, but the problem remains open. Ontology identification is the final step in the formal specification of the problem which is actually faced by an intelligent agent acting in the real world and attempting to fulfill some set of goals. To specify a measure of how well an agent would achieve the intended goals from within a universe, it must be possible to evaluate a model of the universe in terms of the goals. This requires the ability to take a model of reality, running on unknown and potentially surprising physics, and find within it the flawed and leaky abstractions with respect to which our goals are defined. \n Discussion The development of smarter-than-human machines could have a large impact upon humanity  (Bostrom 2014) , and if those systems are not aligned with human interests, the result could be catastrophic  (Yudkowsky 2008) . Highly reliable agent designs are crucial, and when constructing smarter-than-human systems, testing alone is not enough to guarantee high reliability  (Soares and Fallenstein, forthcoming) . In order to justify high confidence that a practical smarter-than-human system will perform well in application, it is important to have a theoretical understanding of the formal problem that the practical system is intended to solve. The problems faced by smarter-than-human systems reasoning within reality are inherently naturalized problems: real systems must reason about a universe which computes the system, a universe that the system is built from. The formalization of Solomonoff's induction problem yielded conceptual tools, such as the universal prior and Kolmogorov complexity, which are useful for reasoning about programs which predict computable sequences. It would be difficult indeed to construct highly reliable practical heuristics that predict computable sequences without understanding concepts such as simplicity priors. We expect that naturalized analogs of the induction problem of Solomonoff and the interaction problem of Hutter will yield analogous conceptual tools useful for constructing systems that reason reliably about the universe in which they are embedded. Just as the intelligence metric of  Legg and Hutter (2007)  fully characterizes the problem of an agent interacting with a separate computable environment to maximize rewards, a corresponding metric derived from a naturalized interaction problem would fully characterize the problem faced by an intelligent agent achieving goals from within a universe. It is not yet clear, in principle, what sort of reasoners perform well when tasked with acting upon their environment from within. Without a formal understanding of the problem, it would be difficult to justify high confidence in a system intended to face a naturalized interaction problem in reality. It is our hope that gaining a better understanding of these problems today will make it easier to design highly reliable smarter-than-human systems in the future. \t\t\t . Orseau and Ring (2012)  give a characterization of the problem which humans face, in implementing a space-time embedded agent, but their problem description requires that we provide a distribution ρ which already characterizes our beliefs about the environment, and so sheds little light on questions of naturalized induction.4. Such as AIXI tl , a computable approximation ofHutter's  AIXI (2000).", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/RealisticWorldModels.tei.xml", "id": "7724feefdf87db1a55f38299b3f73111"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "The views expressed herein are those of the author(s) and do not necessarily reflect the views of the Future of Humanity Institute.", "authors": ["K Eric Drexler"], "title": "Toward Language for Intelligent Machines", "text": "Introduction This section presents a brief summary and outline of core concepts, including boundaries (what is not proposed) and some terminology. Descriptions of the main sections are collected in Section 2.5. \n A brief summary Natural language (NL) is a powerful medium for expressing human knowledge, preferences, intentions, and more, yet NL words and syntax appear impoverished when compared to the representation mechanisms (vector embeddings, directed graphs) available in modern neural ML. Taking NL as a point of departure, we can seek to develop representation systems that are strictly more expressive than natural language. The approach proposed here combines graphs and embeddings to support quasilinguistic neural representations (QNRs) shaped by architectural inductive biases and learned through multitask training. Graphs can strongly generalize NL syntactic structures, while lexical-level embeddings can strongly generalize NL vocabularies. QNR frameworks can syntactically embed and wrap non-linguistic objects (images, data sets, etc.) and formal symbolic representations (source code, mathematical proofs, etc.). Through access to external repositories (Figure  1 .1), inference systems can draw on corpora with content that spans scales that range from phrases and documents to scientific literatures and beyond. Embeddings can abstract QNR content to enable semantic associative memory at scale. Neural networks, potentially exploiting (soft) lattice operations, can process retrieved QNR content to recognize analogies, complete patterns, merge compatible descriptions, identify clashes, answer questions, and integrate information from both task inputs and repositories. \"NL + \" refers to aspirational QNR systems that outperform natural language as a medium for semantic expression and processing. The NL + vision aligns with and extends current research directions in NLP, and NL + implementations could build on current neural architectures and training methods. Potential applications are diverse, ranging from familiar NL-to-NL functionality (interactive search, question answering, writing, translating) to novel forms of representation and reasoning in science, engineering, software development, and mathematics. Potential advantages in scalability, interpretability, cost, and epistemic quality position QNR-based systems to complement or displace opaque foundation models  (Bommasani et al. 2021)  at the frontiers of machine learning. To facilitate skimming, brief summaries of the main sections are collected in Section 2.5. Readers who prefer to start in the middle may wish to skip ahead to Section 8: Quasilinguistic Neural Representations. \n Some (Over)simplified Descriptions An oversimplified problem framing: human intelligence : natural language :: machine intelligence : _______? An oversimplified approach: Use architectural inductive bias and representation learning in neural ML systems to upgrade language by replacing word sequences with explicit parse trees and words with embedding vectors. This is an oversimplification because it (wrongly) suggests a close, fine-grained correspondence between natural languages and QNRs. A less oversimplified description: Use architectural inductive bias and representation learning to develop models that generate and process directed graphs (that strongly generalize NL syntax) labeled with vector embeddings (that strongly generalize both NL words and phrases), thereby subsuming and extending both the syntactic structures and lexical-level components of natural languages. The resulting representation systems can surpass natural languages in expressive capacity, compositionality, and computational tractability. Further objectives and approaches: Learn to embed lexical-level vector representations in structured semantic spaces. Use inductive biases and multitask learning to associate meanings with semantic-space regions (rather than points), and exploit approximate lattice operations (soft unification and anti-unification) as mechanisms for knowledge integration, refinement, and generalization. Translate broad knowledge (e.g., from natural language corpora) into large QNR corpora and employ scalable algorithms to access and apply this knowledge to a wide range of tasks. Enable neural ML systems to write and read QNR content to enable learning that is both efficient and interpretable. \n What is Not Proposed Some contrasting negative samples from the space of related concepts can help readers refine their internal representations of the present proposal: Not a formal language. Formal languages supplement natural languages, but have never subsumed their expressive capacity; frameworks proposed here can embed but are not constrained by formal representations. Not a constructed language. Constructed languages  1  have typically sought clarity and comprehensibility, yet sacrificed expressive capacity; frameworks proposed here seek to expand expressive capacity, yet as a consequence, sacrifice full human comprehensibility. Not a system of hand-crafted representations. Products of neural representation learning typically outperform hand-crafted representations; accordingly, frameworks proposed here rely, not on hand-crafted representations, but on representation learning shaped by architectural bias and training tasks.  2  Not a radical departure from current neural ML. Frameworks proposed here are informed by recent developments in neural ML and suggest directions that are aligned with current research. \n Some Concepts and Terms • \"NL\" refers to natural language in a generic sense. The representational capacity of NL (in this sense) can be thought of as a sum of the representational capacities of human languages. • Representations will be vector-labeled graphs (VLGs); potential arc labels (indicating types, etc.) are not explicitly discussed. • Quasilinguistic neural representations (QNRs, implemented as VLGs) are compositional and language-like: graphs provide upgraded syntactic structure, while embeddings provide upgraded lexical components.  1  • \"NL + \" refers to proposed 2 QNR-based products of neural representation learning that would subsume and extend the representational capacity of natural languages.  3  • The term \"lattice\" and the lattice operations of \"meet\" (here, \"unification\") and \"join\" (here, \"anti-unification\", sometimes termed \"generalization\") have their usual mathematical meanings; in the present context, however, lattices and lattice operations will typically be approximate, or \"soft\" (Appendix A1). \n Motivation and Overview Several perspectives converge to suggest that high-level machine intelligence will require literacy that is best developed in a machine-native medium that is more expressive than natural language. This section concludes with an overview of the sections that follow. Because language and machine learning are broad topics intertwined with each other and with a host of disciplines and application fields, it is difficult to neatly disentangle the various \"motivations and perspectives\" promised by the section title. The discussion that follows (perhaps unavoidably) contains sections with overlapping conceptual content. \n Why Look Beyond Natural Language? Why seek a language-like representational medium that is more expressive and computationally tractable than natural language? The question almost answers itself. But is such a medium possible, what would it be like, how might it be developed and applied? More generally, how might we complete the analogy mentioned above, human intelligence : natural language :: machine intelligence : __________? It seems unlikely that the best answer is \"natural language\" (again) or \"unstructured vector embeddings\". Human intelligence and human societies rely on language as a primary medium for communicating and accumulating knowledge, for coordinating activities, and to some substantial extent, for supporting individual cognition. Intellectually competent humans are literate: They can read and can write content worth reading. High-level machine intelligence will surely be able to do the same and have use for that ability. Current AI research is making strong progress in reading and writing natural language as an interface to the human world, yet makes little use of language(-like) representations for communicating and accumulating knowledge within and between machines. The world's accessible information constitutes a vast, multimodal corpus in which natural language serves as both content and connective tissue. General, high-level intelligent systems must be able to use and extend this information, and it is natural to seek a medium for representing knowledge, both translated and new, that is well-adapted and in some sense native to neural machine intelligence. What might a machine-adapted language be like? It would be strange to find that the best languages for neural ML systems lack basic structural features of human language-in particular, syntax and word-like units-yet perhaps equally strange to find that machines able to share expressive vector embeddings will instead employ sequences of tokens that represent mouth noises. The present document proposes a framework for quasilinguistic neural representations (QNRs) that-by construction-could match and exceed the representational capacity of natural language. Both the potential value and general requirements for such systems seem clear enough to motivate and orient further investigation. \n Some Motivating Facts and Hypotheses The motivation for pursuing QNR approaches that are anchored in NL can be grounded both in uncontroversial facts and in contrasting plausible and implausible hypotheses. \n Key motivating facts 1) Natural language is a key element of human cognition and communication. 2) Natural language provides expressive capacity of unique breadth and flexibility. 3) Structured neural representations can be both richer and more MLcompatible  1  than sequences of words. Corresponding (and plausible) motivating hypotheses +1) Quasilinguistic neural representations of some sort will be key elements of human-level machine cognition and communication, and: +2) The abstract features of natural language (syntactic and lexical constructs) can inform the development of QNRs that subsume and extend syntax and words with graphs and embeddings, and: +3) QNRs informed by natural language constructs can be more expressive and computationally tractable than languages that translate or imitate NL-like sequences of word-like tokens. Corresponding (but implausible) demotivating hypotheses The plausibility of the above hypotheses is supported by the implausibility of contrary hypotheses: -1) That language-like representations will be of little use in human-level machine cognition and communication, or: -2) That language-like syntactic and lexical structures can be no better than flat sequences of vector representations, 2 or: -3) That embeddings in combination with language-like syntactic structures can be no more expressive than sequences of word-like tokens. \n General Approach and Goals In brief, the present line of inquiry suggests a framework that would, as already outlined in part: • Replace and generalize discrete, non-differentiable NL words and phrases with semantically rich, differentiable embeddings.  1  • Replace and generalize NL syntax with general graphs (which also have differentiable representations). • Complement flat neural representations with syntactic structure. • Move linguistic content closer to (quasi)cognitive representations. This strategy starts with NL as a point of departure, retaining generality by subsuming and extending NL piecemeal, at the level of understandable elements. The alternative-to attempt to capture the whole of NL functionality in a more formal, theory-based framework-would risk the loss of functionality that we do not fully understand. Beyond these basic features, QNR frameworks can be extended to: • Exploit abstractive embeddings of fine-grained content (Section 8.3.4). • Exploit abstractive embeddings of large-scale contexts (Section 8.3.5). • Support semantic search at scale (Section 9.1.2). • Support semantic normalization, alignment, refinement, and integration (Section 8.4). • Subsume or embed formal and non-linguistic representations (Section 9.2). What do we want from scalable high-end QNR/NL + systems? • To translate (and refine) large NL corpora into more tractable forms  2  • To combine knowledge from multiple sources, making use of recognizable concordance, clashes, and gaps • To provide comprehensive, dynamic, beyond-encyclopedic knowledge for use by machines and humans • To support the growth of knowledge through machine-aided reasoning The latter goals are worth emphasizing: A key motivation for pursuing NL + capabilities is to enable systems to learn from, apply, and extend content that ranges from informal, commonsense knowledge to mathematics and scientific 1. One direction in which language-like representations might diverge from the picture painted here is in the semantic level of embeddings: As discussed in Section 5.3, embeddings can, through representation discovery, subsume the function of syntactic units above the lexical level (e.g., relatively complex phrases and relatively simple sentences). The partial interchangeability of graph and vector representations (Section 7.3) blurs the potential significance of such a shift, however. 2. While also translating among a likely multiplicity of NL + dialects and overlapping task-oriented sublanguages. literatures. While NL + representations have potentially important roles in NL-to-NL processing  (translation, etc.) , this is almost incidental. The primary aim is to represent, not NL, but what NL itself represents, and to do so better and with broader scope. Current language-related machine representations do not provide full NL (much less NL + ) functionality: They range from opaque language models to explicit knowledge graphs and formal languages, but despite their strengths, none can match (much less exceed) human language in power and generality. Systems like these should be seen as complements-not alternatives-to QNR frameworks.  1   \n Four Perspectives That Help Situate the NL + Concept Reinforcing the points above, four external perspectives may help to situate the NL + concept with respect to related research topics: • The power and limitations of natural language set a high bar to clear while suggesting directions for developing more powerful systems. • The power and limitations of symbolic systems 2 suggest a need for complementary, less formal representation systems. • The power and limitations of flat neural representations suggest the potential advantages of systems that combine the expressive power of dense vector representations with the compositional structure of NL. • The power and limitations of current NLP tools  3  suggest that current neural ML techniques can both support and benefit from QNR-based mechanisms with NL + applications. \n Section Overviews The topics addressed in this document are broad, many-faceted, and have a large surface area in contact with other disciplines. The topics are difficult to disentangle, but the following overviews provide a sketch of the organization and content of the document.  4  Section 1: Introduction This section presents a brief summary and outline of core concepts, including boundaries (what is not proposed) and some terminology. Section 2: Motivation and Overview. Several perspectives converge to suggest that high-level machine intelligence will require literacy that is best developed in a medium more expressive than natural language. Section 3: Notes on Related Work. Current developments in neural ML provide architectures and training methods that can support QNR-oriented research and development. Prospective advances are linked to work in symbolic and neurosymbolic computation, and to broad trends in deep learning and natural language processing. Section 4: Language, Cognition, and Neural Representations. Using NL as a motivation and point of departure for NL + motivates a review of its roles in communication, cognition, and the growth of human knowledge. Prospects for improving compositionality through QNR representations are key considerations. Section 5: Expressive Constructs in NL and NL + . NL + must subsume the functionality of NL constructs identified by linguists, and the shortcomings of those constructs suggest substantial scope for surpassing NL's expressive capacity. Section 6: Desiderata and Directions for NL + . Prospects for improving expressiveness in NL + representations include mechanisms both like and beyond those found in natural languages. Research directions aimed at realizing these prospects are well-aligned with current directions in neural ML. Section 7: Vector-Labeled Graph Representations. In conjunction with today's deep learning toolkit, vector-labeled graph representations provide powerful, differentiable mechanisms for implementing systems that represent and process structured semantic information. Section 8: Quasilinguistic Neural Representations. Applications of vectorlabeled graphs can generalize NL syntax and upgrade NL words to implement quasilinguistic neural representations that parallel and surpass the expressive capacity of natural language at multiple levels and scales. Section 9: Scaling, Refining, and Extending QNR Corpora. Scalable QNR systems with NL + -level expressive capacity could be used to represent, refine, and integrate both linguistic and non-linguistic content, enabling systems to compile and apply knowledge at internet scale. Section 10: Architectures and Training. Extensions of current neural ML methods can leverage architectural inductive bias and multitask learning to support the training of quasilinguistic neural systems with NL + -level expressive capacity. Section 11: Potential Application Areas. Potential applications of QNR/NL + functionality include and extend applications of natural language. They include human-oriented NLP tasks (translation, question answering, semantic search), but also inter-agent communication and the integration of formal and informal representations to support science, mathematics, automatic programming, and AutoML. Section 12: Aspects of Broader Impact. The breadth of potential applications of QNR-based systems makes it difficult to foresee (much less summarize) their potential impacts. Leading considerations include the potential use and abuse of linguistic capabilities, of agent capabilities, and of knowledge in general. Systems based on QNR representations promise to be relatively transparent and subject to correction. Section 13: Conclusions. Current neural ML capabilities can support the development of systems based on quasilinguistic neural representations, a line of research that promises to advance a range of research goals and applications in NLP and beyond. \n Appendix Overviews Several topics have been separated and placed in appendices. Of these, only the first focuses on topics that can be considered foundational. Appendix A1: Unification and Generalization on Soft Semantic Lattices. QNR representations can support operations that combine, contrast, and generalize information. These operations-soft approximations of unification and anti-unification-can be used to implement continuous relaxations of powerful mechanisms for logical inference. \n Potentially Useful Models and Tools Potentially useful models and tools for quasilinguistic processing (QLP) are coextensive with broad areas of neural ML (in particular, neural NLP), and a range of applicable tools (architectures, training data, training tasks, computational resources. . . ) can be found in current practice. \n Vector-Oriented Representations and Algorithms Flat neural representations-sets and sequences of one or more embedding vectors-are ubiquitous in modern neural ML and play key roles as components of proposed QNR architectures. The most closely related work is in natural language processing. In neural NLP, we find vector representations of words at input and (often) output interfaces; 1 some systems produce embeddings of higher-level entities such as sentences and documents.  2  Semantic structure in vector spaces emerges spontaneously in word embeddings.  3  End-to-end training can produce compatible vector representations of images and text for tasks that include image captioning and visual question answering.  4  Extensions of current NLP representations, architectures, and training methods are natural candidates for analogous roles in QLP. Transformer architectures-successful in tasks as diverse as translation, question answering, theorem proving, object recognition, and graph-based inference  5  -appear to have sufficient generality to support many (perhaps most) aspects of quasilinguistic processing. Transformer architectures have been extended to read external memories, including stores of NL text  6  and vector embeddings.  7  QLP systems could po-tentially be based on broadly similar architectures in which inference systems write and read, not flat embeddings, but QNR content. \n Graph Representations and GNNs Graph structures complement vector representations in proposed QNRs, and applications of graph representations have spurred extensive work in neural ML. Iterative, node-to-node message-passing systems-graph neural networks 1 (GNNs)-are deep, convolutional architectures that have been successful in tasks that range from scene understanding to quantum chemistry and neurosymbolic computing; 2 their functionality may be well-suited to semantic processing on QNRs. Graph-oriented models, often GNNs, are widespread in neural knowledge-representation systems.  3  Similar graphs can be aligned for comparison and processing.  4  Although classic GNNs operate on fixed graphs, both differentiable representations and reinforcement learning have been used to implement generative models in the discrete graph-structure domain  5  (graphs can, for example, be mapped to and from continuous embeddings  6  ). With suitable positional encodings, Transformers can operate not only on sequences, but on trees or general graphs.  7  The rich tool set provided by current graph-oriented neural models seems sufficient to support the development of powerful QNR-based applications. \n Computational Infrastructure Broad applications of NL + call for scaling to large corpora, first training on large NL corpora, then writing and applying QNR corpora that may be larger still. Rough analogies between NLP and QLP tasks suggest that computational costs in both training and applying large-scale systems can be in line with the costs of currently practical systems for language modeling and translation 1. Recently reviewed in J.  Zhou et al. (2020)  and  Wu et al. (2021) . 2. R.  Li et al. (2017) ,  Gilmer et al. (2017) ,  Lamb et al. (2020), and Addanki et al. (2021) . Labeled scene graphs, in particular, exemplify learnable semantic relationships among objects in which both objects and their relationships can best be represented by embeddings (see Figure  8 .1).  Zhang et al. (2019)  and  Ji et al. (2021 ) 4. Heimann et al. (2018 ,  Cao et al. (2019), and Fey et al. (2020)  5.  Yun et al. (2019) , J.  Zhou et al. (2020) ,  Kazi et al. (2020 ) 6. Cai, Zheng, and Chang (2018  and  Pan et al. (2018)  7.  Shiv and Quirk (2019)  and  Chen, Barzilay, and Jaakkola (2019)  (Section 9.1); in particular, algorithms for efficient embedding-based semantic search at scale-a key enabler for exploiting large corpora-have been demonstrated in commercial applications.  1  Accordingly, current computational infrastructure seems adequate for development, training, and potential large-scale deployment of NL + applications. Reductions in computational cost and improvements in algorithmic efficiency continue  (Hernandez and Brown 2020 ). \n Wen \n Symbolic, Neural, and Neurosymbolic AI Classic symbolic and neural approaches to AI provide further context for the proposed line of development, which has connections to combined, neurosymbolic approaches. \n Symbolic AI The early decades of AI centered on symbolic models that have little direct relevance to current neural approaches. Symbolic systems had striking successes, 2 yet produced unimpressive results in learning and perceptual tasks like vision. In NLP, symbolic AI faced persistent difficulties stemming from the interplay of word meanings, syntax, and semantic context, while in a key application-machine translation-statistical methods outperformed classic symbolic AI. The quasilinguistic approach suggested here differs from symbolic AI in two quite general ways: 1. Proposed QNRs and QLP computation are intended to support-not directly implement-mechanisms for inference and control. 2. Proposed QNRs and QLP computation are saturated with neural representations and learning mechanisms. What symbolic AI does have in common with proposed QLP is the use of graph-structured representations (in symbolic AI, typically syntax trees) that are associated with distinct lexical-level components. \n Neural ML In recent years deep learning and neural ML have advanced rapidly in both scope and performance, successfully addressing an astounding range of problems. Because the range of potential neural architectures and tasks is open ended, it would be unwise to draw a line around deep learning and propose limits to its capabilities. The present proposals are within, not beyond, the scope of modern neural ML. That said, one can point to persistent difficulties with the most common neural ML approaches, which is to say, models that employ flat neural representations (vectors, sets of vectors, sequences of vectors) that often scale poorly, lack clear compositionality, and resist interpretation. \n Neurosymbolic AI Developments in neurosymbolic AI are advancing at the intersection between symbolic and neural ML, with approaches that include the adaptation of symbolic algorithms to richer, embedding-based representations.  1  This body of work has multiple points of contact with proposed QNR approaches, but its diversity resists summarization. Appendix A1 explores connections with constraint logic programming and related reasoning mechanisms based on neurosymbolic representations. It is important to distinguish among approaches that can be called \"neurosymbolic\", yet differ fundamentally. Geoffrey Hinton has remarked that: Some critics of deep learning argue that neural nets cannot deal with compositional hierarchies and that there needs to be a \"neurosymbolic\" interface which allows neural network front-and back-ends to hand over the higher-level reasoning to a more symbolic system. I believe that our primary mode of reasoning is by using analogies which are made possible by the similarities between learned high-dimensional vectors. . .  (Hinton 2021)  The present proposal differs from those that Hinton criticizes: While both QNRs and conventional symbolic systems employ explicit syntactic structures, compositional hierarchies, and word-like units, QNRs employ highdimensional vectors, not conventional symbol-tokens, in part for the reason Hinton cites. Although higher-level reasoning seems likely to have an algorithmic character, employing conditional branches and dispatch of values to functions, 1 there is good reason to expect that those conditionals and functions will operate on neural representations through neural mechanisms. To structure the objects and operations of reasoning need not impoverish their content. \n Foundation models The term \"foundation model\" has been has been introduced  (Bommasani et al. 2021)  to describe systems that are \"trained on broad data at scale and can be adapted (e.g., fine-tuned) to a wide range of downstream tasks\". Today's leading foundation models (e.g., BERT and GPT-3 2 ) are pretrained on extensive corpora of NL next, while others (e.g., CLIP 3 ) are multimodal; all are based on Transformers. Despite their extraordinary range of applications, current foundation models have suffered from opaque representations, opaque inference mechanisms, costly scaling,  4  poor interpretability, and low epistemic quality, with consequences reviewed and explored in depth by  Bommasani et al. (2021) . QNR-oriented architectures could potentially alleviate each of these difficulties by complementing or displacing models based on stand-alone Transformers. Rather than representing knowledge in unstructured, multi-billionparameter models,  5  architectures that represent knowledge in the form of scalable QNR corpora (Section 9) could provide foundation models in which information content is compositional (Section 4.3) and substantially interpretable  (Section 9.3.3) . Questions of epistemic quality could be addressed by QNR-domain reasoning about external information sources (Section 9.5). 1. For a recent example, see  (Fedus, Zoph, and Shazeer 2021). 2. Devlin et al. (2019)  and  Brown et al. (2020 ) 3. Radford et al. (2021  4. Even relatively scalable Transformer architectures  (Beltagy, Peters, and Cohan 2020; Katharopoulos et al. 2020; Zaheer et al. 2020)  attend only to sections of text, not literatures. \n 5. In which knowledge representation and inference mechanisms entangle error-prone arithmetic and information retrieval with fluent multilingual translation. \n Language, Cognition, and Neural Representations Using NL as a motivation and point of departure for NL + motivates a review of its roles in communication, cognition, and the growth of human knowledge. Prospects for improving compositionality through QNR representations are key considerations. Humans accumulate and share information through natural language, and language is woven into the fabric of human cognition. The expressive scope of NL, though limited, is unique and vast. If we seek to build artificial systems that match or exceed human cognitive capacity, then pursuing machine-oriented NL-like functionality seems necessary. The present section reviews the power and shortcomings of natural and formal languages, and from this perspective, considers prospects for quasilinguistic constructs more powerful and closer to cognitive representations than NL itself. The expressive capacity of NL has resisted formalization, and despite its familiarity, remains poorly understood. Accordingly, in seeking more powerful representations, the proposed strategy will be to upgrade the expressive capacity of the relatively well understood linguistic components of languagelexical units and means for composing them-and to thereby upgrade the less well understood whole. \n Language, Cognition, and Non-Linguistic Modalities Natural language, human cognition, and the social accumulation of knowledge are deeply intertwined. Prospects for NL + systems parallel the role of NL in supporting both reasoning and the growth of knowledge. \n The Roles and Generality of Language Through biology and culture, natural language evolved to exploit animal vocalizations, an acoustic channel that transmits information between neural systems, constrained by limitations (working memory, processing speed) of the cognitive mechanisms that encode and decode meaning. At a societal level, sequences of symbols-written language-encode and extend speech, while at a neurological level, the semantic structures of language mesh with cognition. Natural language has evolved under pressures toward comprehensive expressive capacity, yet its shortcomings are real and pervasive. We can regard NL as both a benchmark to surpass and as a template for representational architectures. \n Complementary, Non-Linguistic Modalities Not words, really, better than words. Thought symbols in his brain, communicated thought symbols that had shades of meaning words could never have. -Clifford Simak City,  1 1952  The human use of complementary, non-linguistic modalities highlights limitations of NL. It is said that a picture is worth a thousand words, but it would be more true to say that images and language each can express semantic content that the other cannot. Another modality, demonstration of skills (now aided by video), is often complemented by both speech and images. Today, artifacts such as interactive computational models provide further modalities. Like language, other modalities mesh with human cognition down to unconscious depths. Human thought relies not only on language, but on mental images, imagined physical actions, and wordless causal models. NL serves as a kind of glue between non-linguistic modalities-naming, linking, and explaining things; NL + frameworks must and can do likewise. Neural ML shows us that vector embeddings can describe much that words cannot, images and more. Expressive embeddings beyond the scope of human language can serve as integral, in some sense \"word-like\" parts of quasilinguistic semantic structures, stretching the concept of NL + . The ability to directly reference and wrap a full range of computational objects stretches the concept further. \n How Closely Linked are Language and Cognition? Embeddings resemble biological neural representations more closely than words do: Embeddings and neural state vectors contain far more information than mere token identities, and both are directly compatible with neural(-like) processing. Thus, embedding-based QNRs are closer to cognitive representations than are natural languages, and presumably more compatible with (quasi)cognitive processing.  2  How deep are the connections between language and human cognition? Without placing great weight on introspective access to cognitive processes, and without attempting to resolve theoretical controversies, some aspects of the relationship between language and cognition seem clear: • Language and cognition have co-evolved and have had ample opportunity to shape and exploit one another. • Language is compositional, and the success of neural models that parse scenes into distinct objects and relationships (Figure  8 .1) suggests that compositional models of the world are more fundamental than language itself. 1 • The experience of trying to \"put thoughts into words\" (and sometimes failing) is good evidence that there are thoughts that are close to-yet not identical with-language; conversely, fluent conversation shows that substantial cognitive content readily translates to and from language. • Externalization of thoughts in language can help to structure personal knowledge, while writing and reading can expand our mental capacities beyond the limits of memory. These observations suggest that the use of linguistically structured yet quasicognitive representations could aid the development of machine intelligence by providing a mechanism that is known to be important to human intelligence. Conversely, prospects for high-level machine intelligence without something like language are speculative at best. \n Cumulative and Structured Knowledge With some loss of nuance, speech can be transcribed as text, and with some loss of high-level structure, 2 text can be mapped back to speech. A central concern of the present document, however, is the growth (and refinement) of accessible knowledge; today, this process relies on the development of text corpora that express (for example) science, engineering, law, and philosophy, together with literatures and histories that describe the human condition. Indeed, in this document, \"natural language\" tacitly refers to language captured in writing. 1. Applications of compositional neural models include not only language-linked explanation and question answering  (Shi, Zhang, and Li 2019)  but non-linguistic tasks such as predictive modeling of physical systems  (Watters et al. 2017) . However, to the extent that neural ML systems can display competence in compositional tasks without language-like internal representations, this counts against the necessity of language-like representations in cognition. In communication, NL + aims to be more expressive than NL; in connection with cognition, NL + aims to be more directly compatible with (quasi)cognitive processing; on a global scale, NL + aims to be more effective in accumulating and applying general knowledge. The growth of NL + corpora can be cumulative and content can be structured for use. Returning to a cognitive perspective, humans not only read and write language, but also \"read and write\" long-term memories. NL + content shares characteristics of both language and memory: Like text, NL + content constitutes explicit, shareable information; like long-term memory, NL + -based systems can store (quasi)cognitive representations that are accessed through associative mechanisms.  1   \n Compositionality in Language and Cognition Why does the structure of language suit it to so many tasks? Links between language and cognition-their co-evolution, their close coupling in use-are part of the story, but correspondence between compositional structures in language, cognition, and the world is another, perhaps more fundamental consideration. The case for the broad utility of NL + frameworks in AI is based in part on this correspondence:  2  Compositionality in the world speaks in favor of pursuing compositional representations of knowledge, situations, and actions. Likewise, compositionality in cognition (and in particular, deliberate reasoning) speaks in favor of strong roles for compositional representations in supporting quasicognitive processing. \n Degrees of Compositionality Here, a system-linguistic, cognitive, or actual-will be termed \"compositional\" if it can be usefully (though perhaps imperfectly) understood or modeled as consisting of sets of parts and their relationships.  3  Useful compositionality requires that this understanding be in some sense local, emerging from parts that are not too numerous or too remote  4  from the parts at the focus of attention or analysis. Hard, local compositionality in symbolic systems requires that the meaning of expressions be strictly determined by their components and syntactic structures; 1 in natural language, by contrast, compositionality is typically soft, and locality is a matter of degree. Strengthening the locality of compositionality can make linguistic representations more tractable, and is a potential direction for upgrading from NL. \n Compositionality in Language and the World Compositionality in language mirrors compositionality in the world-though the compositionality of the world as we see it may be conditioned by language and the structure of feasible cognitive processes. Language (and quasilinguistic systems such as mathematical notation and computer code 2 ) can represent compositionality beyond narrow notions of discrete \"things\" and \"events\". Phenomena that are distributed in space and time (e.g., electromagnetic waves, atmospheric circulation, the coupled evolution of species and ecosystems) can be decomposed and described in terms of distributed entities (fields, fluids, and populations) and relationships among their components. Entities themselves commonly have attributes such as mass, color, velocity, energy density, that are compositional in other ways. Neural representation learning confirms that compositionality is more than a human mental construct. A range of successful ML models incorporate compositional priors or learn emergent compositional representations.  3  These successes show that compositional approaches to understanding the world are useful in non-human, quasicognitive systems. Representations typically include parts with meanings that are conditioned on context.  4  In language, the meaning of words may depend not only on syntactically local context, but on general considerations such as the level of technicality of discourse, the epistemic confidence of a writer, or the field under discussion (\"glass\" means one thing in a kitchen, another in optics, and something more general in materials science). In images, the appearance of an object may depend on lighting, style, resolution, and color rendering, and scenes generated from textual descriptions can differ greatly based on contextual attributes like \"day\" or \"city\". Contexts themselves (as shown by these very descriptions) can to some extent be represented by NL, yet images generated by neural vision systems can be conditioned on contextual features that are substantially compositional (as shown by structure in latent spaces), and even recognizable by humans, yet not readily described by language. All of these considerations speak to the value of compositional representations in language and beyond. Taken as a whole, these considerations suggest that quasilinguistic representations can describe features of the world that elude language based on strings of words. \n Compositionality in Nonlinguistic Cognition Compositionality in cognition parallels compositionality in language and in the world. When we perceive objects with properties like \"relative position\", or consider actions like \"dropping a brick\", these perceptions and cognitive models are compositional in the present sense; they are also fundamentally nonlinguistic yet often expressible in language. Thus, visual thinking  (Giaquinto 2015 ) can be both non-linguistic and compositional; when this thinking is abstracted and expressed in diagrammatic form, the resulting representations typically show distinct parts and relationships. \n Compositionality in Natural Language The concept of language as a compositional system is woven through the preceding sections, but these have spoken of language as if compositionality in language were a clear and uncontroversial concept. It is not. Formal symbolic systems provide a benchmark for full compositionality, and by this standard, natural languages fall far short.  1  Formal concepts of compositionality have difficulty including contextual features like \"topic\" or \"historical era\" or \"in Oxford\", and struggle with contextual modulation of meaning at the level of (for example) sentences rather than words.  1  Neural NLP can (and to be effective, must) incorporate information that is beyond the scope of locally compositional representations of word strings. Transformer-based language models show something like understanding of text in a broad world context, yet the embodiment of that knowledge in their weights is not obviously compositional in its abstract structure, and obviously not compositional in its concrete representation. To say that the meaning of text emerges from a composition of its components-with particular Transformer computational states as one of those components-would stretch the meaning of compositionality to the breaking point.  2  To be meaningful, compositionality must be in some sense local. Compositionality in language can be strengthened by strengthening the localization of contextual information. Quasilinguistic neural representations can contribute to this in two ways: First, by substituting descriptive vector embeddings for ambiguous words and phrases,  3  and second, by incorporating embeddings that locally summarize the meaning of a syntactically broad context (for example, a context on the scale of a book), together with embeddings that locally summarize remote context, for example, the kind referred to in expressions like \"when read in the context of. . . \"  4  The second role calls for embeddings that are not conventionally \"lexical\".  5  In brief, the world, cognition, and language are substantially \"compositional\", and relative to natural language, quasilinguistic neural representations can improve local compositionality. As we will see, improving local compositionality can have a range of advantages. 5 Expressive Constructs in NL and NL + NL + must subsume the functionality of NL constructs identified by linguists, and the shortcomings of those constructs suggest substantial scope for surpassing NL's expressive capacity. The rich expressive constructs found in natural languages provide a benchmark and point of departure for the pursuit of more powerful capabilities. Considering linguistics in the context of neural ML suggests both challenges that must be addressed and challenges that can be set aside in pursuing the promise of NL + . Figure  6 .1 illustrates relationships among some overlapping classes of representation systems, some of which stretch the definition of \"linguistic\". Appendix A3 explores further aspects of the relationship between natural language and potential QNR/NL + representations, including prospects for condensing, regularizing, and extending the scope of quasilinguistic representations. \n Vocabulary and Structure in Natural Languages What linguistic constructs enable NL to serve human purposes?  1  Those purposes are broad: Natural languages are richer than \"knowledge representation languages\" 2 or other formal systems to date; language can describe complex things, relationships, and situations, along with goals, actions, abstract argumentation, epistemic uncertainty, moral considerations, and more. The proposed strategy for developing frameworks that subsume and extend NL is to upgrade representational functionality by upgrading both NL components and compositional mechanisms. Crucially, this strategy requires no strong, formal theory of semantics, grammar, syntax, or pragmatics, and hence no coding of formal rules. An approach that (merely) upgrades components and structure sidesteps questions that have generated decades of academic controversy and addresses instead a far more tractable set of problems. \n Grammar and Syntax It is widely agreed that sentences can usefully be parsed into trees 1 defined by grammars, yet there are several competing approaches.  2  In an NL + framework, explicit graphs can accommodate and generalize any choice, hence no choice need be made at the outset; further, because neural models can integrate information across extended syntactic regions, grammar-like choices of local graph structure need not strongly constrain semantic processing. Architectures with a QNR-oriented inductive bias together with neural representation learning on appropriate tasks should yield effective systems with NL + functionality (Section 10). Section 7 explores potential vector/graph representations in greater depth, while Section 8 considers applications of vector embeddings that subsume elements of NL syntactic structure-again, not to propose or predict, but instead to explore the potential of learned NL + representations. \n Words, Modifiers, and Lexical-Level Expressions The role of lexical-level units in NL is subsumed by embeddings in NL + frameworks, and prospects for improving NL + expressiveness depend in part on the potential advantages of representations in continuous, structured spaces. It is easy to argue that embeddings can be as expressive as words (e.g., embeddings can simply designate words), but a deeper understanding of their potential calls for considering words and word-like entities in the context of continuous semantic spaces. \n Words and word-level units When considering NL + frameworks from an NL perspective, a key question will be the extent to which multi-word expressions can be folded into compact, tractable, single-vector representations while gaining rather than losing expressive power. Two heuristics seem reliable: 1. Meanings that some languages express in a single word can be represented by a single vector.  3  2. Sets of word-level units with related meanings correspond to sets of points clustered in a semantic space. In the present context (and adopting an NL perspective), \"lexical\" (or \"wordlevel\") units will be restricted to a single noun or verb together with zero or more modifiers (e.g., adjectives or adverbs),  1  and a \"simple\" noun (or verb) phrase will be one that cannot be decomposed into multiple noun (or verb) phrases.  2  The phrase \"a large, gray, sleepy cat\" is a simple noun phrase in this sense; \"a cat and a dog\" is not simple, but conjunctive. As with many features of language, the single-vs.-conjunctive distinction is both meaningful and sometimes unclear: Is \"spaghetti and meatballs\" a single dish, or a pair of ingredients? Is \"bait and switch\" a deceptive strategy or a pair of actions? Is the semantic content of \"ran, then halted\" necessarily compound, or might a language have a verb with an inflected form denoting a past-tense run-then-halt action? Note that nothing in the present discussion hinges on the sharpness of such distinctions. Indeed, the typical flexibility and softness of mappings between meanings and representations speaks against discrete tokens and formal representations and in favor of QNR/NL + systems that can represent a softer, less formal semantics.  3  In natural language, the meaning of \"word\" is itself blurry, denoting a concept that resists sharp definition. In linguistics, \"morphemes\" are the smallest meaningful linguistic units, and include not only (some) words, but prefixes, suffixes, stems, and the components of compound words. In morphologically rich languages, words may contain morphemes that denote case or tense distinctions that in other languages would be denoted by words or phrases. Blurriness again speaks in favor of soft representations and against linguistic models that treat \"word\" as if it denoted a natural kind. \n Content word roles and representations Linguists distinguish \"content words\" (also termed \"open class\" words) from \"function\" (or \"closed class\") words. The set of content words in a vocabulary 1. This use of \"lexical\" differs from a standard usage in linguistics, where to be \"lexical\", a phrase must have a meaning other than what its components might indicate, making the phrase itself a distinct element of a vocabulary. The phrases \"on the other hand\" and \"cat-andmouse game\" are lexical in this sense. NLP research recognizes a similar concept, \"multi-word expressions\"  (Constant et al. 2017) . is large and readily extended; 1 the set of function words (discussed below) is small and slow to change. Content words typically refer to objects, properties, actions, and relationships. They include nouns, verbs, adjectives, and most adverbs, and they typically have more-or-less regular marked or inflected forms. The growth of human knowledge has been accompanied by the growth of content-word vocabularies. From the perspective of NL syntax and semantics, adjectives and adverbs can be viewed as modifying associated nouns and verbs; this relationship motivates the description of the resulting phrases as word-level (or lexical) in the sense discussed above. In exploring the potential expressive capacity of NL + frameworks, will be natural to consider semantic embedding spaces that accommodate (meanings like those of) nouns and verbs together with the lexical-level refinements provided by adjectives, adverbs, markers, and inflections.  2  One can think of content words as representing both distinctions of kind and differences in properties.  3  Numbers, animals, planets, and molecules are of distinct kinds, while magnitude, color, accessibility, and melting point correspond to differences in properties, potentially modeled as continuous variables associated with things of relevant kinds. In embedding spaces, one can think of distinctions of kind as represented chiefly by distances and clustering among vectors, and differences in properties as represented chiefly by displacements along directions that correspond to those properties.  4  Note that this perspective (kinds → clusters; properties → displacements) is primarily conceptual, and need not (and likely should not) correspond to distinct architectural features of QNR/NL + systems. \n Function word roles and representations Function (closed-class) words are diverse. They include coordinating conjunctions  (and, or, but. . . )  conjunctive adverbs  (then, therefore, however. . . ), prepositions (in, of, without. . . ) modal verbs (can, should, might. . . ), determiners (this, that, my. . . ), connectives (and, or, because, despite. . . ) , and more (see Table  A2 .1). While the set of open-class words is huge, the set of closed-class words is small-in English, only 200 or so. The vocabulary of open-class words can readily be extended to include new meanings by example and definition. Closed-class words, by contrast, typically play general or abstract roles, and linguists find that this small set of words is nearly fixed (hence \"closed\").  1  The still-awkward repurposing of \"they\" as a gender-neutral third-person singular pronoun illustrates the difficulty of expanding the closed-class vocabulary, even to fill a problematic gap-alternatives such as \"ze\" have failed to take hold (C. . Function words that in themselves have minimal semantic content can shape the semantics of complex, content-word constructs (such as clauses, sentences, and paragraphs), either as modifiers or by establishing frameworks of grammatical or explanatory relationships. Representations that subsume NL must span semantic spaces that include function words. The closed-class nature of NL function words suggests opportunities for enriching the corresponding semantic domains in NL + . In English, for example, the ambiguity between the inclusive and exclusive meanings of \"or\" in English suggests that even the most obvious, fundamental-and in human affairs, literally costly-gaps in function-word vocabularies can go unfilled for centuries.  2   \n TAM-C modifiers Natural languages employ a range of tense, aspect, modality, and case (TAM-C) modifiers; 1 some are widely shared across languages, others are rare. In some languages, particular TAM-C modifiers may be represented by grammatical markers (a class of function words); in others, by inflections (a class of morphological features). Meanings that in English are conveyed by function words may be conveyed by inflections in other languages.  2  Sets of TAM-C overlap with closed-class adjectives and adverbs, and are similarly resistant to change. Some TAM-C modifiers alter the meaning of lexical-level elements (both words and phrases); others operate at higher semantic levels. They can convey distinctions involving time, space, causality, purpose, evidence, grammatical roles, and more: Note that many of these distinctions are particularly important to situated and cooperating agents. TAM-C modifiers are sufficiently general and important that NLs encode them in compact lexical representations. The role of TAM-C modifiers in NL calls for similar mechanisms in NL + , and-like adjectives and adverbs-TAM-C modifiers are natural candidates for folding into lexical-level vector representations (Section A3.3).  3  5.4 Phrases, Sentences, Documents, and Literatures NL + representations like those anticipated here condense (some) phrase-level meanings into single, hence syntax-free, embeddings. Within an NL + domain, these \"phrase-level\" meanings are definitional, not merely approximations of the meaning of a hypothetical phrase in NL. As outlined above, meanings of kinds that correspond to simple, lexical-level noun and verb phrases (in NL) are strong candidates for single-embedding representation, while the equivalents of noun and verb conjunctions (in NL) typically are not; nonetheless, equivalents of NL phrases of other kinds (some noun clauses?) could potentially be captured in single embeddings. The boundaries of the useful semantic scope of definitional vector embeddings are presently unclear, yet at some level between a word and a document, definitional embeddings must give way to vector-labeled graph representations.  1  5.5 At the Boundaries of Language: Poetry, Puns, and Song Discussions of potential NL + \"expressiveness\" come with a caveat: To say that a representation \"is more expressive\" invites the question \"expressive to whom?\" The meanings of NL text for human readers depend on potentially human-specific cognition and emotion, but NL + expressions cannot, in general, be read by humans-indeed, systems that \"read\" NL + (e.g., systems for which \"associative memory\" means scalable search and \"NL + expressions\" are situated within a spectrum of QNRs) are apt to be quite unlike humans. What might be called \"outward-facing expressions\"-descriptions, commands, questions, and so on-represent a kind of semantic content that may be accessible to human minds, but is not specific to them. The arguments above suggest that NL + representations can outperform NL in this role. However, NL text-and utterances-can convey not only outward-facing semantic content, but human-specific affective meaning, as well as NL-saturated associations, allusions, and word play. Puns and poetry translate poorly even between natural languages, and poetry merges into song which merges into pure music, far from what is usually considered semantic expression. For present purposes, these functions of text and utterances will be considered beyond the scope of \"language\" in the sense relevant to NL + functionality; what can be represented, however, is literal content (NL text, recorded sound) embedded in NL + descriptions of its effects on human minds (which are, after all, parts of the world to be described).  1  In the context of this document, the content (or \"meaning\") of natural language expressions will be equated with semantic content in the outwardfacing sense. When a poem delivers a punch in the gut, this is not its meaning, but its effect. 6 Desiderata and Directions for NL + Prospects for improving expressiveness in NL + representations include mechanisms both like and beyond those found in natural languages. Research directions aimed at realizing these prospects are well-aligned with current directions in neural ML. Building on the NL-centered considerations discussed above, and looking forward to mechanisms beyond the scope of natural language, this section outlines desiderata for NL + (and more generally, QNR) functionality. Appendix A3 further explores NL + -oriented prospects for QNR frameworks. \n Improve and Extend Fundamental NL Constructs Desiderata for NL + representations include improving and extending fundamental NL constructs: \n Subsume (but do not model) NL: The aim of research is not to model NL, but to model (and extend) its functionality. Exploit deep learning: Use neural ML capabilities to extend NL constructs without hand-crafting representations. Improve compositionality: Embeddings can provide effectively infinite vocabularies and loosen the dependence of meaning on context. Use explicit graph representations: Graphs can compose embeddings without the constraints of NL syntax and ambiguities of coreference. Exploit vector embedding spaces: Relative to words, embeddings can improve both expressive capacity and computational tractability: -Embeddings are natively neural representations that need not be decoded and disambiguated. -Embeddings, unlike words, are differentiable, facilitating end-toend training. -Embeddings provide effectively infinite vocabularies, enriching expressive capacity. Embrace (but do not impose) formal systems: NL + frameworks and formal systems are complementary, not competing, modes of representation. Formal systems can be embedded as sub-languages in NL + , much as mathematical expressions are embedded in textbooks (Section 9.2.4). \n Exploit Mechanisms Beyond the Scope of Conventional NL The discussion above has focused on features of NL + frameworks that can be regarded as upgrades of features of NL, replacing words with embeddings and implicit syntactic trees with explicit, general graphs. There are also opportunities to exploit representations beyond the scope of NL: These include vector representations that enable novel, high-level forms of expression, abstraction, and semantic search, as well as QNR-based tools for knowledge integration at scale syntactically embedded non-linguistic content far beyond the bounds of what can be construed as language (Section 9.2). \n Use Embeddings to Modify and Abstract Expressions Embeddings can perform semantic roles at the level of sentences, paragraphs, and beyond. In an adjective-like role Section 8.3.1), they can modify or refine the meaning of complex expressions; in an abstractive role, they can enable efficient, shallow processing (skimming) (Section 8.3.4). \n Use Embeddings to Support Scalable Semantic Search Abstractive embeddings can support similarity-based semantic search-in effect, associative memory-over NL + corpora. Efficient similarity search scales to repositories indexed by billions of embeddings (Section 9.1.5). \n Reference, Embed, and Wrap Everything At a syntactic level, NL + frameworks can embed not only formal systems, but also content of other kinds (Figure  6 .1): Non-linguistic lexical-level units: Neural embeddings can represent objects that differ from words, yet can play a similar role. For example, image embeddings can act as \"nouns\", while vector displacements in a latent space can act as \"adjectives\" (Section 8.2 NL + -enabled systems could contribute to current research objectives in mathematics, science, engineering, robotics, and machine learning itself, both by helping to integrate and mobilize existing knowledge (e.g., mining literatures to identify capabilities and opportunities), and by facilitating research that produces new knowledge. Efforts to harness and extend the power of natural language align with aspirations for advanced machine learning and artificial intelligence in general. \n Some Caveats The present discussion describes general frameworks, mechanisms, and goals, but the proposed research directions are subject to a range of potential (and equally general) criticisms: • Proposed NL + frameworks are templated on NL, but perhaps too closely to provide fundamentally new capabilities. • Alternatively, proposed NL + frameworks may differ too greatly from NL, undercutting the feasibility of equaling NL capabilities. • In light of the surprising power of flat neural representations, inductive biases that favor QNRs might impede rather than improve performance. • Both neural ML and human cognition embrace domains that are decidedly non-linguistic, limiting the scope of NL-related mechanisms. • Ambitious NL + applications may call for more semantic structure than can readily be learned. • Implementation challenges may place ambitious NL + applications beyond practical reach. • Current research may naturally solve the key problems with no need to consider long-term goals. • The prospects as described are too general to be useful in guiding research. • The prospects as described are too specific to be descriptive of likely developments. Most of these criticisms are best regarded as cautions: Linguistic mechanisms have limited scope; relaxing connections to NL may improve or impede various forms of functionality; implementations of working systems can be difficult to develop or fall short of their initial promise; motivations and outlines of research directions are inherently general and open-ended; generic, short-term motivations often suffice to guide developments up a gradient that leads to capabilities with far-reaching applications. Nonetheless, despite these caveats, near-term research choices informed by QNR/NL + concepts seem likely to be more fruitful than not, leading to tools and insights that enable and inform further research. Much current research is already well-aligned with QNR development and NL + aspirations, and it is interesting to consider where that research may lead. \n Vector-Labeled Graph Representations In conjunction with today's deep learning toolkit, vector-labeled graph representations provide powerful, differentiable mechanisms for implementing systems that represent and process structured semantic information. This present section examines vector-labeled graph representations (VLGs) from the perspectives of representational capacity and neural ML tools; the following section will examine prospects for applying this representational capacity to implement QNR systems that surpass the expressive capacity of natural language. 1 \n Exploiting the Power of Vector Representations In a sense, the large representational capacity of typical high-dimensional embedding vectors is trivial: Vectors containing hundred or thousands of floating point numbers contain enough bits to encode lengthy texts as character strings. What matters here, however, is the representational capacity of vectors in the context of neural ML-the scope and quality of representations that can be discovered and used by neural models that are shaped by suitable architectural biases, loss functions, and training tasks. This qualitative kind of capacity is difficult to quantify, but examples from current practice are informative. \n Vector Representations are Pervasive in Deep Learning Deep learning today is overwhelmingly oriented toward processing continuous vector representations, hence the extraordinary capabilities of deep learning testify to their expressive power. \n Vector Representations Can Encode Linguistic Semantic Content Continuous vector representations in NLP shed light on prospects for expressive, tractable QNRs.  2  . The two leading roles for vector embeddings in proposed QNR systems are (1) definitional representations of lexical-level components (paralleling vector semantics in NL and word embeddings in NLP, Section 8.2) and (  2 ) abstractive representations of higher-level constructs for indexing and summarization (Section 8.3.4). \n Single Vectors Can Serve as Compositional Representations In conventional symbolic systems, compositionality enables complex meanings to be represented by combinations of components. In vector representations, meanings can be attributed to orthogonal vector components (e.g., representing different properties of something), then those components can be combined by vector addition and recovered by projection onto their corresponding axes. Condensing what in NL would be multi-component, lexicallevel syntactic structures into single embeddings can reduce the number of distinct representational elements, retain semantic compositionality, and enable facile manipulation by neural computation. \n High-Dimensional Spaces Contain Many Well-Separated Vectors In considering the representational capacity of high-dimensional vectors, it is important to recognize ways in which their geometric properties differ from those of vectors in the low-dimensional spaces of common human experience. In particular, some measures of \"size\" are exponential in dimensionality, and are relevant to representational capacity. Call a pair of unit-length vectors with cosine similarity ≤ 0.5 \"well separated\". Each of these vectors defines and marks the center of a set (or \"cluster\") of vectors with cosine similarity ≥ 0.86; these sets are linearly separable and do not overlap. How many such well-separated cluster-centers can be found in a high-dimensional space? In a given dimension d, the number of vectors k(d) that are well separated by this criterion is the \"kissing number\", the maximal number of non-overlapping spheres that can be placed in contact with a central sphere of equal radius  (Figure 7.1) . Kissing numbers in low-dimensional spaces are small (k(2) = 6, k(3) = 12. . . ), but grow rapidly with d. For d = 64 and 128, 1 k(d) > 10 7 and 10 12 ; for d = 256, 512, and 1024, an asymptotic lower bound  (Edel, Rains, and    Sloane 2002) k(d) > 2 0   .2075...d  gives k(d) >10 14 , 10 30 , and 10 62 . Thus, the number of neighboring yet well-separated cluster-centers that can be embedded in spaces of dimensionalities commonly used in neural ML is far (!) in excess of any possible NL vocabulary.  2  Note that the region around a cluster-center itself has great representation power for sub-clusters: For example, its content can be separated from the rest of the space by a linear threshold operation and then scaled and projected into a space of d-1 dimensions, where similar considerations apply recursively so long as the residual dimensionality remains large.  1  The above description is intended to provide an intuition for some aspects of the expressive capacity of high-dimensional vector spaces, not to predict or suggest how that capacity will or should be used in prospective systems: Learned representations in current neural ML may offer a better guide. \n Points Can Represent Regions in Lower-Dimensional Spaces https://www.overleaf.com/project/61183c4a1765f22abf2e3ebf An embedding of dimensionality 2d can represent (for example) a center-point in a d-dimensional semantic space together with parameters that specify a surrounding box in that space.  2  An embedding may then designate, not a specific meaning, but a range of potential meanings; alternatively, a range can be regarded as a specific meaning in a semantic space that explicitly represents ambiguity. Interval arithmetic 3 generalizes some operations on d-dimensional points to operations on d-dimensional boxes. This document will usually discuss embeddings as if they represent points in semantic spaces, with operations on embeddings described as if acting on vectors that designate points, rather than regions. The concept of semantic regions becomes central, however, in considering semantic lattice structure and constraint-based inference that generalizes logic programming (Section A1.4). \n Exploiting the Power of Graph-Structured Representations Graphs are ubiquitous as representations of compositional structure because they can directly represent things and their relationships as nodes and arcs. Graphs (in particular, trees and DAGs) are central to traditional, word-oriented natural language processing, while general graphs have found growing applications in diverse areas of neural ML. This section outlines several classes of graphs and their potential roles in representing and processing semantic information. \n Terminology, Kinds, and Roles of Graphs The components of graphs are variously (and synonymously) called edges, arcs, or links (potentially directed), which connect vertices, points, or nodes, which in turn may carry labels, attributes, or contents. The present discussion will typically refer to arcs (or links between document-scale objects) that connect nodes that carry attributes or labels (in a semantic context, contents or embeddings).  1  Here, \"graph\" typically denotes a directed graph with attribute-bearing nodes. Labeled arcs, multigraphs, and hypergraphs 2 are potentially useful but not explicitly discussed; weighted arcs are essential in some differentiable graph representations. (Sequences of embeddings can be viewed as instances of a particularly simple class of graphs.) In prospective QNR frameworks, vector-labeled graphs have at least two areas of application: The first area parallels the use of graphs in classic NLP, where expressions are typically parsed and represented as syntax trees or (to 1. In some formal models, arcs also carry attributes. Without loss of generality, graphs G with labeled arcs can be represented by bipartite graphs G in which labeled arcs in G correspond to labeled nodes in G . For the sake of simplicity (and despite their potential importance) the present discussion does not explicitly consider labeled arcs. In general, computational representations of graphs will be implementation-dependent and will change depending on computational context (e.g., soft, internal representations in Transformers translated to and from hard, external representations in expression-stores). represent resolved coreferences) DAGs; VLGs can represent syntactic trees explicitly, bypassing parsing, and can represent coreference through DAGs, bypassing resolution. The second area of application involves higher-level semantic relationships that in NL might be represented by citations; in an NL + context, similar relationships are naturally represented as general, potentially cyclic graphs. (These topics are discussed in Section 8.1.) \n VLGs Can Provide Capacity Beyond Stand-Alone Embeddings Fixed-length embeddings lack scalability, while sequences of embeddings (e.g., outputs of Transformers and recurrent networks), though potentially scalable, lack explicit, composable, and readily manipulated structure. Arcs, by contrast, can compose graphs to form larger graphs explicitly and recursively with no inherent limit to scale or complexity, and subgraph content can be referenced and reused in multiple contexts. Trees and graphs are standard representations for compositional structure in a host of domains, and are latent in natural language. Graphs with vector attributes can expand the representational capacity of sets of embeddings by placing them in a scalable, compositional framework. 1 \n VLGs Can Be Differentiable Typical neural operations on vector attributes are trivially differentiable, while differentiable operations on representations of graph topologies require special attention. Without employing differentiable representations, options for seeking graphs that minimize loss functions include search (potentially employing heuristics or reinforcement learning) and one-shot algorithmic construction.  2  With differentiable representations, more options become available, including structure discovery through end-to-end training or inferencetime optimization. Conventional representations in which nodes and arcs are simply present or absent can be termed \"hard graphs\"; representations can be made \"soft\" and differentiable by assigning weights in the range [0, 1] to arcs. Differentiable algorithms that assign weights tending toward {0, 1} can recover conventional graphs by discarding edges when their weights approach zero, implementing structure discovery through differentiable pruning. Differentiable pruning operates on a fixed set of nodes and typically considers all pairs, impairing scalability, 1 but algorithms that exploit near-neighbor retrieval operations on what may be very large sets of nodes (Section 9.1.5) could implement scalable, differentiable, semantically informed alternatives that do not a priori restrict potential topologies. In typical graph representations, a link is implemented by a semantically meaningless value (e.g., an array index or hash-table key) that designates a target node. Through near-neighbor retrieval, by contrast, a vector associated with a source-node can serve as a query into a semantically meaningful space populated by keys that correspond to candidate target-nodes. Selecting the unique node associated with the nearest-neighbor key yields a hard graph; attending to distance-weighted sets of near neighbors yields a soft graph.  2  In this approach, query and key embeddings can move through their joint embedding space during training, smoothly changing neighbor distances and the corresponding arc weights in response to gradients. Mutable, soft-graph behavior can be retained at inference time, or models can output conventional, hard-graph VLGs, potentially retaining geometric information. Thus, models that build and update QNR corpora could provide fluid representations in which changes in embeddings also change implied connectivity, implicitly adding and deleting (weighted) arcs. In semantically structured embedding spaces, the resulting changes in topology will be semantically informed. Weighted graphs may also be products of a computation rather than intermediates in computing hard-graph representations. Substituting a set of weighted arcs for a single hard arc could represent either structural uncertainty (potentially resolved by further information) or intentional, semantically informative ambiguity. \n VLGs Can Support Alignment and Inference Neural algorithms over VLGs can support both processing within single expressions and operations that link or combine multiple expressions. These have potentially important applications within QNR frameworks. Algorithms that align similar graphs 1 can facilitate knowledge integration over large corpora, supporting the identification of clashes, concordances, and overlaps between expressions, and the construction of refinements, generalizations, analogies, pattern completions, and merged expressions, 2 as well as more general forms of interpretation and reasoning. Differentiable algorithms for graph alignment include continuous relaxations of optimal assignment algorithms that enable end-to-end learning of alignable, semantically meaningful embeddings  (Sarlin et al. 2020) . \n VLGs Can Support (Soft) Unification and Generalization If we think of expressions as designating regions of some sort 3 , then to unify a pair of expressions is to find the largest expressible intersection of their regions, while to anti-unify (or generalize) a pair of expressions is to find the narrowest expressible union of their regions. Domains that support these operations (and satisfy a further set of axioms) constitute mathematical lattices.  4  Given a set of expressions, unification may be used to infer narrower, more specific expressions, while anti-unification operations may be used to propose broader, more general expressions. As discussed above, QNR expressions that represent explicit uncertainty or ambiguity (e.g., containing vectors representing uncertain or partially constrained values) may be regarded as representing regions in a semantic space. The nature of \"expressible regions\" in the above sense depends on choices among representations. Notions of \"soft\" or \"weak\" unification (discussed in Section A1.4.3) can replace equality of symbols with similarity between point-like semantic embeddings, or approximate overlap when embeddings are interpreted as representing semantic regions. Soft unification supports continuous, differentiable relaxations of the Prolog backward-chaining algorithm, enabling soft forms of multi-step logical inference. Applications of soft unification include question answering, natural-language reasoning, and theorem proving; 1 Soft operations can also infer variables from sets of values (Cingillioglu and Russo 2020). As noted above, unification and generalization have further potential applications to knowledge integration in NL + corpora. \n Mapping Between Graphs and Vector Spaces Research has yielded a range of techniques for mapping graphs to and from vector representations. These techniques are important because they bridge the gap between high-capacity compositional structures and individual embeddings that lend themselves to direct manipulation by conventional (non-GNN) neural networks. \n Embeddings Can Represent and Decode to Graphs Neural models can be trained to encode graphs 2 (including tree-structured expressions 3 ) as embeddings, and to decode embeddings to graphs.  4  A common training strategy combines graph-encoding and generative models with an autoencoding objective function.  5  In the domain of small graphs, embeddings could encode representations with considerable accuracy-potentially with definitional accuracy, for graphs that are small in both topology and information content. Note that summary, query, and key embeddings (Section 8.3.4 and Section 9.1.2) can represent semantic content without supporting graph reconstruction. \n Embeddings Can Support Graph Operations Embeddings that enable accurate graph reconstruction provide differentiable graph representations beyond those discussed in Section 7.2.3. Whether directly or through intermediate, decoded representations, graph embeddings can support a full range of VLG operations. Accordingly, this class of embeddings can be considered interchangeable with small 6 VLGs, and need not be considered separately here. Potential advantages include computational efficiency and seamless integration with other embedding-based representations. \n Quasilinguistic Neural Representations Applications of vector-labeled graphs can generalize NL syntax and upgrade NL words to implement quasilinguistic neural representations that parallel and surpass the expressive capacity of natural language at multiple levels and scales. The present section discusses how vector-labeled graphs (VLGs) can be applied to implement quasilinguistic neural representations (QNRs) that improve on natural languages by upgrading expressive capacity, regularizing structure, and improving compositionality to facilitate the compilation, extension, and integration of large knowledge corpora. Appendix A3 covers an overlapping range of topics in more detail and with greater attention to NL as a model for potential NL + frameworks. As usual in this document, conceptual features and distinctions should not be confused with actual features and distinctions that in end-to-end trained systems may be neither sharp, nor explicit, nor (perhaps) even recognizable. Conceptual features and distinctions should be read neither as proposals for hand-crafted structures nor as confident predictions of learned representations. They serve to suggest, not the concrete form, but the potential scope of representational and functional capacity. \n Using Graphs as Frameworks for Quasilinguistic Representation Proposed QNRs are vector-labeled graphs.  1  Attributes can include type information as well as rich semantic content; to simplify exposition,  2  attributes of what are semantically arcs can be regarded as attributes of nodes in a bipartite graph, 3 and will not be treated separately. Like trees, general (e.g., cyclic) graphs can have designated roots. \n Roles for Graphs in NL-Like Syntax As already discussed, vector attributes can represent lexical-level components, while graphs can represent their syntactic compositions; where syntax in NL is implicit and often ambiguous, QNR-graphs can make syntactic relationships (and in particular, coreference) explicit, thereby disentangling these relationships from lexical-level semantic content. (See also Section A3.1.1.) \n Roles for Graphs Beyond Conventional Syntax In an extended sense, the syntax of NL in the wild includes the syntax of (for example) hierarchically structured documents with embedded tables, cross references, and citations.  1  QNRs can naturally express these, as well as syntactic structures that are unlike any that can be displayed in a document.  2  Stretching the concept of syntax, graphs can express networks of relationships among objects. Scene graphs provide a concrete illustration of how recognizably linguistic relationships can naturally be expressed by a nonconventional syntax (for a simple example, see Figure  8 .1). \n Figure 8.1: A neurally inferred scene graph that relates several entities through subject-verb-object relationships. An enriched graph for this scene could represent more complex relationships (e.g., man1 failing to prevent man2 from throwing a Frisbee to man3). An enriched representation of entities could replace instances of labels like \"man\" with embeddings that have meanings more like man-withappearance(x)-posture(y)-motion(z), replace instances of verbs like \"catching\" with embeddings that have meanings more like probablywill-catch-intended-pass, and so on.  3   \n Using Embeddings to Represent Lexical-Level Structure At a lexical level, embedding-space geometry can contribute to expressive power in several ways: Proximity can encode semantic similarity. Large displacements can encode differences of kind. Small offsets (or larger displacements in distinguishable subspaces) can encode graded, multidimensional semantic differences. Where NL depends on context to disambiguate words, QNRs can employ lexical-level embeddings to express meanings that are substantially disentangled from context. (For further discussion, see Section A3.2.1.) \n Proximity and Similarity Neural representation learning typically places semantically similar entities near one another and unrelated entities far apart. Embedding NL words works well enough to be useful, yet its utility is impaired by polysemy and other ambiguities of natural language.  1  By construction, native embedding-based representations minimize these difficulties. \n Graded Semantic Differences The direction and magnitude of incremental displacements in embedding spaces can represent graded, incremental differences in properties among similar entities: The direction of a displacement axis encodes the kind of difference, the magnitude of the displacement along that axis encodes the magnitude of the corresponding difference, 2 and the (commutative) addition of displacements composes multiple differences without recourse to syntactic constructs. By avoiding difficulties arising from the discrete words and word-sequences of NL, the use of continuous, commutative vector offsets can substantially regularize and disentangle lexical-level semantic representations. \n Relationships Between Entities of Distinct Kinds Distinctions between entities of different kinds 1 can be represented as discrete displacement vectors, which can encode degrees of similarity in distances and cluster things of similar kinds (animals with animals, machines with machines, etc., as seen in the geometry of word embeddings). Displacement directions can also encode information about kinds. Word embeddings trained only on measures of co-occurrence in text have been found to represent relationships between entities in displacement vectors, and analogies in vector differences and sums  (Allen and Hospedales 2019) . In neural knowledge graphs, 2 relationships between entities can be encoded in geometries that are deliberately constructed or induced by learning.  3  Lexicallevel QNR labels can provide similar representational functionality. \n Expressing Higher-Level Structure and Semantics The discussion above addressed the vector semantics of embeddings that play lexical-level roles (e.g., nouns and verbs with modifiers); the present discussion considers roles for embeddings in the semantics of higher-level (supra-lexical) QNR expressions. Some of these roles are definitional: in these roles, embeddings modify the meanings of higher-level constructs.  4  In other roles, embeddings are abstractive: Their semantic content may corresponds to (quasi)cognitive results of reading higher-level constructs, and are therefore semantically optional (in effect, they cache and make available computational results). As usual, the aim here is to describe available representational functionality, not to predict or propose specific representational architectures. Representa-tion learning within a QNR framework need not (and likely will not) respect these conceptual distinctions. \n Expression-Level Modifiers (Definitional) Although lexical-level embeddings can condense (analogues of) phrases that include modifiers of (analogues of) words, some lexical-level modifiers operate on supra-lexical expressions that cannot be condensed in this way. These modifiers may resemble adjectives, but are attached to units more complex than words. Epistemic qualifiers (Section A3.4) provide examples of this functionality. Syntactically, definitional modifiers of this kind could be applied to expressions as vector attributes of their root nodes; semantically, expression-level modifiers act as functions of a single argument: an expression as a whole. \n Expression-Level Relationships (Definitional) In NL, conjunctive elements (often function words) can compose and modify the meanings of combined expressions. Some can operate on either lexicallevel units or complex expressions (examples include and, or, and/or, both-and); others (for example, however, because, and despite 1 ) typically compose meaning across substantial syntactic spans, operating not only at the level of clauses and sentences, but at the level of paragraphs and beyond. Syntactically, a relationship between expressions could be applied through vector attributes of the root node of a composition of syntactic structures (subgraphs of a tree, DAG, etc.); semantically, expression-level relationships correspond to functions of two or more arguments: the expressions they relate. \n Expression-Level Roles (Definitional) Expressions and the things they describe have roles (frequently with graded properties) in larger contexts; features of roles may include purpose, importance, relative time, epistemic support, and so on. As with expression-level modifiers and relationships, role descriptions of this kind could be applied to an expression through vector attributes of its root node. In general, however, a role may be specific to a particular context. If an expression is contained (referenced, used) in multiple contexts, it may have multiple roles, hence representations of those roles must syntactically stand 1. See also: punctuation. above root-nodes and their attributes. More concretely, multiple contexts may link to an expression through nodes that represent the corresponding contextual roles. \n Content Summaries (Abstractive) In human cognition, the act of reading a text yields cognitive representationsin effect, abstractive summaries-that can contribute to understanding directly, or by helping to formulate (what are in effect) queries that directly enable retrieval of relevant long-term memories, or indirectly enable retrieval of relevant texts.  1  In a QNR context, dynamically generated summary embeddings can play a similar, direct role at inference time. When stored, however, summaries can do more: • Amortize reading costs for inference tasks. • Enable efficient skimming. • Stretch the semantic span of attention windows.  2  • Provide keys for scalable associative memory. The nature of a summary may depend on its use, potentially calling for multiple, task-oriented summaries of a single expression.  3  Semantically, content summaries are not expression-level modifiers: Summaries approximate semantics, but modifiers help define semantics. Content summaries may depend on contextual roles, placing them in syntactic positions above the roots of summarized expressions. A natural auxiliary training objective would be for the similarity of pairs of content-summary vectors to predict the unification scores (Section A1.4.3) of the corresponding pairs of expressions.  4   \n Context Summaries (Abstractive) Context summaries (e.g., embeddings that summarize context-describing QNRs) represent information that is potentially important to interpreting and using a QNR expression.  1  • What are the general and immediate topics of the context? • Is the context a formal or an informal discussion? • How does the context use the expression? • Does the context support or criticize the content of the expression? At inference time, context summaries could direct attention to relevant commonsense knowledge regarding a particular field or the world in general.  2  Interpreting an expression correctly may require contextual information that is distributed over large spans; providing (quasi)cognitive representations of contexts can make this information more readily available. Like content summaries, context summaries are semantically optional, but the potential scale and complexity of contexts (local, domain, and global) may make some form of context-summarization unavoidable, if extended contexts are to be used at all. Syntactically (and to some extent semantically), context summaries resemble contextual roles. \n Origin Summaries (Abstractive) Origin summaries (potentially summarizing origin-describing QNRs) can indicate how an expression was derived and from what information: • What inputs were used in constructing the expression? • What was their source? • What was their quality? • What inference process was applied? A structurally distinct QNR expression can be viewed as having a single source (its \"author\"), and hence its root node can be linked to a single origin summary. Origins and their summaries are important to the construction, correction, and updating of large corpora. \n Regularizing, Aligning, and Combining Semantic Representations Disentangling and regularizing semantic representations-a theme that runs through Section 8.1 and Section 8.2-has a range of benefits. In conjunction with basic QNR affordances, regularization can facilitate the alignment of representations, which in turn can facilitate systematic inference, comparison, combination, and related forms of semantic processing at the level of collections of expressions. \n Regularizing Representations In natural languages, expressions having similar meanings may (and often must) take substantially different forms. Words are drawn from discrete, coarse-grained vocabularies, hence incremental differences in meaning force discrete changes in word selection. Further, small differences in meaning are often encoded in structural choices-active vs. passive voice, parallel vs. non-parallel constructs, alternative orderings of lists, clauses, sentences, and so on. These expressive mechanisms routinely entangle semantics with structure in ways that impose incompatible graph structures on expressions that convey similar, seemingly parallel meanings. The nature of QNR expressions invites greater regularization: Meanings of incommensurate kinds may often be encoded in substantially different graph topologies, while expressions with meanings that are similar-or quite different, yet abstractly parallel-can be represented by identical graph topologies in which differences are encoded in the continuous embedding spaces of node attributes. Humans compare meanings after decoding language to form neural representations; natively disentangled, regularized, alignable representations can enable comparisons that are more direct. By facilitating alignment and comparison, regularization can facilitate systematic (and even semi-formal) reasoning. \n Aligning Representations Section 7.2.4 took note of neural algorithms that can align graphs in which parallel subgraphs carry corresponding attributes. Given a pair of aligned graphs, further processing can exploit these explicit correspondences. 1 \n Alignment for Pattern-Directed Action By triggering pattern-dependent actions, graph alignment can establish relationships between other entities that do not themselves align. For example, reasoning from facts to actions can often be formulated in terms of production rules 2 that update a representation-state (or in some contexts, cause an external action) when the pattern that defines a rule's precondition matches a pattern in a current situation: Where patterns take the form of graphstructured representations, this (attempted) matching begins with (attempted) graph alignment. Potential and actual applications of production rules range from perceptual interpretation to theorem proving. \n Comparing, Combining and Extending Content Alignment facilitates comparison. When graphs align, natural distance metrics include suitably weighted and scaled sums of distances between pairs of vector attributes.  3  When alignment is partial, subgraphs in one expression may be absent from subgraphs in the other (consistent with compatibility), or subgraphs may overlap and clash. Clashes may indicate mutual irrelevance or conflict, depending on semantic contents and problem contexts. Overlapping graphs can (when compatible) be merged to construct extended, consistent descriptions of some entity or domain. In formal systems, this operation corresponds to unification (Section A1.1.1); in general QNR contexts, formal unification can be supplemented or replaced by soft, learned approximations (Section A1.  4.3) . Expressions that are similar and compatible yet not identical may provide different information about a single object or abstraction; unification mechanisms then can enrich representations by merging information from multiple sources. For example, unification of multiple, related abstractions (e.g., derived from uses of a particular mathematical construct in diverse domains of theory and application) could produce representations with richer semantics than could be derived from any single NL expression or documentrepresentations that capture uses and relationships among abstractions, rather than merely their formal structures. Alternatively, expressions that are similar yet not fully compatible may describe different samples from a single distribution. The dual of unification, anti-unification (a.k.a. generalization, Section A1.1.2), provides the most specific representation that is compatible with a pair of arguments, encompassing their differences (including clashes) and retaining specific information only when it is provided by both. Anti-unification of multiple samples 1 yields generalizations that can (perhaps usefully) inform priors for an underlying distribution. Like unification, generalization yields an expression at the same semantic level as its inputs, but to represent analogies calls for an output that contains representations of both input graphs and their relationships. Alignment can provide a basis for representing analogies as first-class semantic objects that can be integrated into an evolving corpus of QNR content Looking back for a moment, prospects that were motivated by the most basic considerations-upgrading words and syntax to machine-native embeddings and graphs-have naturally led to prospects that go far beyond considerations of expressiveness per se. QNR/NL + frameworks lend themselves to applications that can subsume yet are qualitatively different from those enabled by natural language. 9 Scaling, Refining, and Extending QNR Corpora Scalable QNR systems with NL + -level expressive capacity could be used to represent, refine, and integrate both linguistic and non-linguistic content, enabling systems to compile and apply knowledge at internet scale. Preceding sections discussed vector/graph representations and their potential use in constructing QNR frameworks that are powerful and tractable enough to meet the criteria for NL + . The present section explores the potential scaling of QNR/NL + systems to large corpora, then considers how scalable, fully functional systems could be applied to include and integrate both linguistic, and non-linguistic content while refining and extending that content through judgment, unification, generalization, and reasoning. \n Organizing and Exploiting Content at Scale A core application of language-oriented QNR systems will be to translate and digest large NL corpora into tractable, accessible NL + representations, a task that makes scalability a concern. Current practice suggests that this will be practical. NL → NL + translation will likely employ components that resemble today's translation systems. Potentially analogous NL → NL translation tasks are performed at scale today, while ubiquitous internet-scale search suggests that large corpora of NL + knowledge can be exploited for practical use. \n Learning by Reading Can Be Efficient at Scale To estimate the magnitude of required computational tasks, we can roughly equate the incremental costs of building a corpus by reading NL texts and translating them to NL + representations with the incremental costs of training a language model by reading NL texts and (in some sense) \"translating\" their content to model parameters. According to this estimate, training GPT-3 was roughly equivalent to translating ~300 billion words 1 to NL + representationsabout 30 times the content of the 1.8 million papers on the arXiv (arXiv 2021) or 10% of the 40 million books scanned by Google (H. . As an alternative and perhaps more realistic estimate, we can roughly equate the cost of NL → NL + translation to the cost of NL → NL translation; the throughput of Google Translate (>3 trillion words per year 1 ) is then an indicator of affordable throughput (~40 million books/month). These comparisons concur in suggesting the feasibility of translating NL information to NL + corpora at scale.  2  Translating NL content into accessible, quasicognitive NL + is a form of learning that is distinct from training. Unlike training model parameters, the task of reading, translating, and expanding corpora is by nature fully parallelizable and scalable. The products are also more transparent: Exploiting external, compositional representations of information can enable facile interpretation, correction, redaction, and update of a system's content. 3 \n Semantic Embeddings Enable Semantic Search NL + expressions can be indexed by summary embeddings that associate similar expressions with near-neighbor points in a semantic space.  4  Using these embeddings as keys in near-neighbor retrieval can provide what is in effect \"associative memory\" that supports not only search, but tasks involving knowledge comparison and integration (see Section 9.4 below). Different embeddings of an expression can serve as keys suited for different search tasks. In some use-contexts, we care about the physical properties of an object; in others, we care about its cost and functionality. In some use-contexts, we care about a city's geography; in others, about its nightlife. Accordingly, 1. Reported in (Turovsky 2016). 2. Contributing to efficiency and scalability, computations that both write and read NL + corpora can, as noted above, avoid problems that stem from representing knowledge solely in trained model parameters. Large language models learn not only knowledge of language per se and general features of the world, but also a range of specific facts about people, places, etc. The striking improvements that have resulted from scaling language models have stemmed, not only from improvements in broadly applicable, immediately available syntactic, semantic, and reasoning skills (K.  Lu et al. 2021) , but from memorization of specific, seldom-used facts about the world (Yian . For example, GPT-2 uses its 1.5 billion parameters to memorize names, telephone numbers, and email addresses, as well as the first 500 digits of π  (Carlini et al. 2021) . Encoding factual information of this sort in billions of model parametersall used in computing each output step-is problematic: Both training and inference are expensive, and the results are opaque and difficult to correct or update. Parameter counts continue to grow  (Brown et al. 2020; Fedus, Zoph, and Shazeer 2021)   4.  Wang and Koopman (2017) ,  Schwenk and Douze (2017), and Tran et al. (2020) . Note that retrieval based on similarity between semantic embeddings can be effective across modalities (e.g.,  Miech et al. 2021) . Pairwise similarity between vector keys could potentially reflect the unification scores (Section A1. what might be represented as a single span of content may naturally be associated with multiple domain-oriented summaries and keys. We find this general pattern in Transformers, where multi-head attention layers project each value to multiple key and query vectors. A quite general distinction is between representing the content of an expression and representing the kinds of questions that its content can answer-between what it says and what it is about. The role of semantic search overlaps with the role of links in large-scale NL + structures (for example, generalizations of citation networks), but differs in designating regions of semantic space through query-embeddings rather than designating specific expressions through graph links. Conventional references and syntactic relationships are represented by links, but looser relationships can be represented by \"query embeddings\" within expressions, where these embeddings are taken to denote soft graph links to a potential multiplicity of target expressions in an indexed corpus. 1 \n Semantic Search Extends to Non-Linguistic Content By proposed construction, any item that can be described by NL can be (better) described by NL + . Examples include code, engineering designs, legal documents, biomedical datasets, and the targets of current recommender systems-images, video, apps, people, products, and so on. Embeddings that represent (descriptions of) non-linguistic content (Section 9.2.2) are relevant to NL + in part because they can be directly referenced by NL + expressions, and in part because these NL + descriptions can provide a basis for semantically rich content embeddings. \n NL + -Mediated Embeddings Could Potentially Improve NL Search Systems trained to map NL to NL + can support the production of NL embeddings (NL → NL + → embedding) that are based on disentangled, contextually informed semantic representations. If so, then co-learned, NL + -mediated key and query embeddings could potentially improve kNN semantic search over corresponding NL items at internet scale. \n Semantic Search Can Be Efficient at Scale Exact k-nearest neighbor search in generic high-dimensional spaces is costly, 1 but approximate nearest neighbor search (which usually finds the keys nearest to a query) have been heavily researched and optimized; efficient, practical, polylogarithmic-time algorithms can support (for example) billion-scale recommender systems, 2 and can do likewise for retrieval of semantic content in NL + systems at scale. \n Incorporating General, Non-Linguistic Content Human knowledge includes extensive nonlinguistic content, yet we use natural language to describe, discuss, index, and provide instructions on how to create and use that content. Language-linked nonlinguistic content sprawls beyond the boundaries of an NL-centric concept of NL + -boundaries that do not constrain language-inspired QNR/NL + applications. Examples of non-linguistic information content include: • Non-linguistic lexical-level units (e.g., image embeddings) • Expressions in formal languages (e.g., mathematical proofs) • Precise descriptions of objects (e.g., hardware designs) • Formal graph-structured representations (e.g., probabilistic models) 9.2.1 Using \"Words\" Beyond the Scope of Natural Language \"Nouns\" represent things, but vector embeddings can represent things in ways that are in no sense translations of word-like entities in NL: Image embeddings, for example, can serve as \"nouns\",  3  though the content of an image embedding need not resemble that of an NL noun phrase. Embeddings that represent datasets or object geometries have a similar status. Similar considerations apply to verbs and relationships: Words and phrases have a limited capacity to describe motions, transformations, similarities, and differences.  1  As with nouns, embeddings (and a wide range of other information objects) can again subsume aspects of NL expressive capacity and extend representations to roles beyond the scope of practical NL descriptions. This expressive scope may often be difficult to describe concretely in NL.  2  QNR \"verbs\" could express transformations of kinds not compactly expressible in conventional NL: For example, displacements of physical objects can be represented by matrices that quantitatively describe displacements and rotations, and a more general range of transformations can be expressed by displacement vectors in a suitable latent space. Stretching the conceptual framework further, verb-like transformations can be specified by executable functions. 3 \n Referencing Non-Linguistic Objects References to linguistic and non-linguistic objects are not sharply demarcated, but some information objects are both complex and opaque, while physical objects are entirely outside the information domain. All these (indeed, essentially anything) can nonetheless be referenced within the syntactic frameworks of QNR expressions. As a motivating case, consider how embedded hyperlinks and more general URIs expand the expressive power of online documents: The ability to unambiguously reference not only text, but arbitrary information objects, is powerful, and on the internet, this capability meshes smoothly with NL. Examples of non-linguistic information objects include images, websites, data repositories, and software. Beyond the domain of information objects, domain-appropriate reference modalities can act as proper nouns in designating physical objects, people, places, and the like. A natural pattern of use would place a reference in a QNR wrapper that might include (for example) a summary embedding, conventional metadata, descriptions, and documentation, all of which can exploit the representational capacity of NL + . QNR wrappers can facilitate indexing, curation, and the selection or rejection of entities for particular uses. Information objects include executable code, and when accessed remotely, executable code is continuous with general software and hardware services. Interactive models 1 can complement NL descriptions of systems. Access to documented instances of the computational models used to produce particular NL + items can contribute to interpreting the items themselves: Connections between products and their sources often should be explicit. \n Embedding Graph-Structured Content Human authors routinely augment sequential NL text with graph structures. Even in basically sequential text (e.g., this document), we find structures that express deep, explicit nesting and internal references. Documents often include diagrams that link text-labeled elements with networks of arrows, or tables that organize text-labeled elements in grids. These diagrams and tables correspond to graphs. Further afield, graph-structured representations can describe component assemblies, transportation systems, and biological networks (metabolic, regulatory, genetic, etc.) Text-like descriptions of statistical and causal relationships become probabilistic models and causal influence diagrams. In ML, we find formal knowledge graphs in which both elements and relationships are represented by vector embeddings, 2 while augmenting language-oriented Transformer-tasks with explicit representations of relationships has proved fruitful.  3  Distinctions between these and NL + representations blur or disappear when we consider generalizations of conventional syntax to graphs, and of symbols and text to embeddings and general data objects. Some potential use-patterns can be conceptualized as comprising restricted, graph-structured representations (e.g., formal structures that support crisply defined inference algorithms), intertwined with fully general graph-structured representations (informal structures that support soft inference, analogy, explication of application scope, and so on). \n Integrating Formal Quasilinguistic Systems Programming languages and mathematics are formal systems, typically based on quasilinguistic representations.  4  In these systems, the word-like entities (symbols) have minimal semantic content, yet syntactic structures in conjunction with an interpretive context specify semantic or operational meanings that endow these systems with descriptive capabilities beyond the conventional scope of NL. How are formal systems connected to NL, and by extension, to NL + frameworks? NL cannot replace formal systems, and experience suggests that no conventional formal system can replace NL. What we find in the wild are formal structures combined with linguistic descriptions: mathematics interleaved with explanatory text in papers and textbooks, programs interleaved with comments and documentation in source code, and so on. Experience with deciphering such documents suggests the value of intimate connections between NL-like descriptions and embedded formal expressions.  1  In one natural pattern of use, formal systems (e.g., mathematics, code, and knowledge representation languages) would correspond to distinguished, formally interpretable subsets of networks of NL + expressions. Formal languages can describe executable operations to be applied in the informal context of neural computation. Conversely, vector embeddings can be used to guide premise selection in the formal context of theorem proving. 2 \n Translating and Explaining Across Linguistic Interfaces Proposed NL + content has both similarities to NL and profound differences. It is natural to consider how NL might be translated into NL + representations, how NL + representations might be translated or explained in NL, and how anticipated differences among NL + dialects might be bridged. \n Interpreting Natural Language Inputs NL + frameworks are intended to support systems that learn through interaction with the world, but first and foremost, are intended to support learning from existing NL corpora and language-mediated interactions with humans. Translation from NL sources to NL + is central both to representation learning and to key applications. It is natural to expect that encoders for NL → NL + translation will share a range of architectural characteristics and training methods with encoders for NL → NL translation and other NLP tasks (Section 10.2). Translation to NL + could be applied both to support immediate tasks and (more important) to expand corpora of NL + -encoded knowledge. Transformer-based NL → NL translation systems can learn language-agnostic representations, 1 a capability which suggests that NL → NL + translation will be tractable. \n Translating Among NL + Dialects Because NL + corpora could span many domains of knowledge-and be encoded by multiple, independently trained systems-it would be surprising to find (and perhaps challenging to develop) universally compatible NL + representations. In neural ML, different models learn different embeddings, and the representations learned by models with different training sets, training objectives, and latent spaces may diverge widely. In an NL + world, we should expect to find a range of NL + \"dialects\" as well as domain-specific languages. Nonetheless, where domain content is shared and representational capacities are equivalent, is reasonable to expect facile NL + → NL + translation. Further, regarding interoperability in non-linguistic tasks, the concrete details of differing representations can be hidden from clients by what is, in effect, an abstraction barrier. 2 Domain-specific languages may resist translation at the level of representations, yet contribute seamlessly to shared, cross-domain functionality. Lossless translation is possible when the semantic capacity of one representation fully subsumes the capacity of another. Given the computational tractability of NL + representations, we can expect translation between similar NL + dialects to be more accurate than translation between natural languages. In translation, spaces of lexical-level embeddings can be more tractable than discrete vocabularies in part because vector-space transformations can be smooth and one-to-one.  3   \n Translating and Explaining NL + Content in NL It is reasonable to expect that, for a range of NLP tasks, conventional NL → (opaque-encoding) → NL pipelines can be outperformed by NL → NL + → NL pipelines; 1 this would imply effective translation of the intermediate NL + representations to NL. If NL + frameworks fulfill their potential, however, NL + corpora will contain more than translations of NL expressions. Content will typically draw on information from multiple sources, refined through composition and inference, and enriched with non-linguistic word-like elements. There is no reason to expect that the resulting representations-which will not correspond to particular NL expressions or any NL vocabularies-could be well-translated into NL. If NL + is more expressive than NL, it follows that not all NL + content can be expressed in NL. How, then, might NL + content be accessed through NL, either in general or for specific human uses? • Some NL + expressions will correspond closely to NL expressions; here, we can expect to see systems like conditional language models  (Keskar et al. 2019 ) applied to produce fluent NL translations. • NL + descriptions that are detailed, nuanced, complex, and effectively untranslatable can inform NL descriptions that provide contextually relevant information suitable for a particular human application. • Similarly, a general abstraction expressed in NL + might be effectively untranslatable, yet inform narrower, more concrete NL descriptions of contextually relevant aspects of that abstraction. • To the extent that NL expressions could-if sufficiently extensiveconvey fully general information (a strong condition), NL could be used to describe and explain NL + content in arbitrary detail; this approach is continuous with verbose translations. • NL + content in effectively non-linguistic domains could in some instances be expressed in diagrams, videos, interactive models, or other human-interpretable modalities. Relative to the opaque representations common in current neural ML, NL + representations have a fundamental advantage in interpretability: Because QNRs are compositional, their components can be separated, examined, and perhaps interpreted piece by piece. Even when components cannot be fully interpreted, they will often refer to some familiar aspect of the world, and knowing what an expression is about is itself informative. \n Integrating and Extending Knowledge Graph-structured representations in which some vector attributes designate regions in semantic spaces 1 lend themselves to operations that can be interpreted as continuous relaxations of formal unification and anti-unification, which in turn can support reasoning and logical inference by (continuous relaxations of) familiar algorithms. These operations can help extend corpora by combining existing representations along lines discussed from a somewhat different perspective in the preceding section. \n Combining Knowledge Through (Soft) Unification Compatible representations need not be identical: Alignment and (successful) soft unification (Appendix A1) indicate compatibility, and the successful unification of two expressions defines a new expression that may both combine and extend their graph structures and semantically narrow their attributes. Soft unification could potentially be used to refine, extend, compare, and link QNR/NL + representations. Where QNR graphs partially overlap, successful unification yields a consistent, extended description.  2  Attempts to unify incompatible representations fail and could potentially yield a value that describes their semantic inconsistencies.  3  In refining content through soft unification, relatively unspecific structures in one QNR 4 may be replaced or extended by relatively specific structures  5  from an aligned QNR. Embeddings that contain different information may commbine to yield a semantically narrower embedding. Products of successful unification (new, more informative expressions and relationships) are candidates for addition to an NL + corpus. Records of failed unifications-documenting specific clashes-can provide information important to epistemic judgment. These successful and unsuccessful unificationproducts may correspond to nodes in a semantically higher-level graph that represents relationships among expressions. \n Generalizing Knowledge Through (Soft) Anti-Unification While unification of two expressions can be regarded as combining their information content, anti-unification (generalization) can be regarded as combining their uncertainties, spanning their differences, and discarding unshared information. Generalization in this sense may represent a useful prior for generative processes within distributions that include the inputs. \n Constructing Analogies Operations based on generalization and unification could be applied to identify, construct, and apply analogies and abstractions in QNR corpora: • A range of latent analogies will be reflected in recognizably parallel structures between or among concrete descriptions. 1 • Alignment of parallel concrete descriptions can establish concrete analogies, potentially reified as graphs. • Generalization over sets of parallel descriptions can abstract their common structures as a patterns. • Unification of a concrete description with a pattern can indicate its analogy to a set of similar concrete descriptions without requiring pairwise comparison. Analogies have powerful applications. For example, if a description of A includes features of a kind absent from the analogous description of B, then it is reasonable to propose A-like features in B. Analogies among mammals, for example, underlie the biomedical value of discovery in animal models. Indeed, analogy permeates science, guiding both hypotheses and research planning. Analogy has been identified as central to cognition, 1 and reified networks of analogies can form graphs in which relationships among abstractions are themselves a domain of discourse. With suitable annotations, products of generalization and analogy-new abstract expressions and relationships-are candidates for addition to an NL + corpus. \n Extending Knowledge Through (Soft) Inference Natural language inference (NLI) is a major goal in NLP research, and recent work describes a system (based on a large language model) in which NL statements of rules and facts enable answers to NL questions  (Clark, Tafjord, and Richardson 2020) . Inference mechanisms that exploit NL + → NL + operations could potentially be useful in NLI pipelines, and in refining and extending NL + corpora NL + → NL + inference could play a central role. Regularizing and normalizing QNR representations (Section 8.4) can enable a kind of \"soft formalization\" based on continuous relaxations of formal reasoning (modeled, for example, on logic programming, Section A1.4.1). Rules can be represented as \"if-then\" templates (in logic, expressions with unbound variables) in which successful unification of an expression with an \"if-condition\" template narrows the values of attributes that, through coreference, then inform expressions constructed from \"then-result\" templates.  2  Advances in neural mechanisms for conventional symbolic theorem-proving (e.g., guiding premise selection) have been substantial.  3  It is reasonable to expect that wrapping formal expressions in NL + descriptions-including embeddings and generalizations of potential use contexts-could facilitate heuristic search in automated theorem proving. \n Credibility, Consensus, and Consilience Humans examine, compare, contrast, correct, and extend information represented in NL literatures. Machines can do likewise with NL + content, and for 1. See  Hofstadter (2009)  and  Gentner and Forbus (2011) . 2. Unification-based machinery of this kind can implement Prolog-like computation with applications to natural language inference  (Weber et al. 2019 ). 3.  Riedel (2016, 2017) ,  Minervini et al. (2018), and Minervini et al. (2019)  similar purposes, 1 a process that can exploit unification (and failure), together with generalization and analogy. \n Modeling and Extending Scholarly Literatures The strategy of using NL as a baseline suggests seeking models for NL + corpora in the scholarly literature, a body of content that includes both structures that are broadly hierarchical (e.g., summary/body and section/subsection relationships) and structures that correspond to more general directed graphs (e.g., citation networks). In abstracts, review articles, and textbooks, scholarly literatures summarize content at scales that range from papers to fields. Proposed NL + constructs can support similar patterns of expression, and can extend content summarization to finer granularities without cluttering typography or human minds. Scholarly citations can link to information that is parallel, supportive, problematic, explanatory, or more detailed; in NL + syntax, analogous citation functionality can be embodied in graphs and link contexts. Through indexing, citations in scholarly literatures can be made bidirectional, 2 enabling citation graphs to be explored through both cites-x and cited-by-x relationships. For similar reasons, NL + links in fully functional systems should (sometimes) be bi-directional.  3  In general, the structure and semantics of citations and citing contexts can vary widely (document-level remote citations are continuous with sentence-level coreference), and the structure of NL + representations makes it natural to extend cites and cited-by relationships to expressions finer-grained than NL publications. Steps have been taken toward applying deep learning to improve the integration of scholarly literatures.  4  It will be natural to build on this work using NL + tools to enrich NL + content . \n Using Credibility, Consensus, and Consilience to Inform Judgments As with NL, not all NL + content will be equally trustworthy and accurate. The origin of information-its provenance-provides evidence useful in judging epistemic quality, a consideration that becomes obvious when considering information derived from heterogeneous NL sources. Judgments of epistemic quality can reflect not only judgments of individual sources (their credibility), but also the consistency of information from different sources, considering both consensus among sources of similar kind, and consilience among sources that differ in kind. Search by semantic similarity and comparison through structural and semantic alignment provide a starting point, but where epistemic quality is in question, provenance will often be key to resolving disagreements. Some broad distinctions are important. \n Informing judgments through provenance and credibility Provenance is an aspect of context that calls for summarization. In a fully functional system, embeddings (and potentially extended QNRs 1 ) can summarize both information sources and subsequent processing, providing descriptive information that can be linked and accessed to derived content for deeper examination. Provenance information helps distinguish broad concurrence from mere repetitions-without tracking sources, repetitions may wrongly be counted as indicating an informative consensus. By analogy with (and leveraging) human judgment, systems can with some reliability recognize problematic content that can range from common misconceptions through conspiracy theories, fake news, and computational propaganda.  2  Problematic content should be given little weight as a source of information about its subject, yet may itself be an object of study.  3  Judgments of quality can be iterative: The quality and coherence of content can be judged in part by the quality and coherence of its sources, and so on, a process that may converge on descriptions of more-or-less coherent but incompatible models of the world together with accounts of their clashes. Judging source quality in generally sound NL literatures is a familiar human task. In the experimental sciences, for example, we find a spectrum of epistemic status that runs along these lines: • Uncontroversial textbook content • Reviews of well-corroborated results \n • Reports of recent theory-congruent results • Reports of recent surprising results • Reports of recent theory-incongruent results • Reports of actively disputed results • Reports of subsequently retracted results All of these considerations are modulated by the reputations of publishers and authors. Broadly similar indicators of quality can be found in history, economics, current affairs, and military intelligence. The reliability of sources is typically domain-dependent:  1  Nobel laureates may speak with some authority in their fields, yet be disruptive sources of misinformation beyond it; a conspiracy theorist may be a reliable source of information regarding software or restaurants. Although information about credibility can be propagated through a graph, credibility is not well-represented as a scalar. \n Informing judgments through consensus In judging information, we often seek multiple sources and look for areas of agreement or conflict-in other words, degrees of consensus. Relevant aspects of provenance include the quality of individual sources, but also their diversity and evidence of their independence.  2  What amount to copying errors may be indicated by sporadic, conflicting details. Lack of independence can often be recognized by close similarity in how ideas are expressed.  3  In judging information from (what are credibly considered to be) direct observations, experiments, and experience, the quality of human sources may play only a limited role. Established methods of data aggregation and statistical analysis will sometimes be appropriate, and while NL + representations may be useful in curating that data, subsequent methods of inference may have little relationship to NL + affordances. Inference processes themselves, however, constitute a kind of algorithmic provenance relevant to downstream representation and assessment of results. \n Informing judgments through consilience More powerful than consensus among sources of broadly similar kinds is con- 2. Some of this evidence is found in surface features of NL texts (e.g., uses of specific words and phrases); other evidence is found in features of semantic content. silience, agreement of evidence from sources of qualitatively different kindsfor example, agreement between historical records and radiocarbon dating, or between an experimental result and a theoretical calculation. Judgment of what constitutes \"a difference in kind\" is a high-level semantic operation, but potentially accessible to systems that can recognize similarities and differences among fields through their citation structures, focal concerns, methodologies, and so on. Distinguishing consilience from mere consensus is a judgment informed in part by provenance, and is key to robust world modeling. It calls for modeling the epistemic structures of diverse areas of knowledge. \n Architectures and Training Extensions of current neural ML methods can leverage architectural inductive bias and multitask learning to support the training of quasilinguistic neural systems with NL + -level expressive capacity. The preceding sections suggest that QNR frameworks can implement powerful, tractable NL + functionality, provided that suitable representations can be learned; the present section outlines potentially effective approaches to learning based on adaptations of familiar architectures and training methods. Vector-labeled graph bottlenecks can provide a strong inductive bias, while multitask learning and auxiliary loss functions can shape abstract representations that are anchored in, yet substantially decoupled from, natural languages. The final section outlines potential architectures for components that control inference strategies. \n General Mechanisms and Approaches Following neural ML practice, the development of QNR-centered systems calls, not for hand-crafting features, but for architectures that provide suitable components, capacity, and inductive bias, in conjunction with training tasks that provide suitable datasets, objectives, and loss functions. General mechanisms and approaches include: These mechanisms and approaches should be seen as facets of multitask learning in which a key goal is to develop NL → QNR → NL systems 1 that support broad applications. Pretrained NL → embedding → NL models (e.g., BERT and friends) are perhaps today's closest analogues. • Potential inputs to encoders include text, but also images, symbolic expressions, multimodal data streams, 2 and so forth. • Potential outputs from decoders include translations, summaries, answers to questions, retrieved content, classifications, and various products of downstream processing (images, engineering designs, agent behaviors, and so on). 3. It is at least questionable whether the inductive bias provided by a simple QNR-bottleneck architecture would outperform an otherwise similar but unconstrained encoder/decoder architectures in stand-alone NL → NL tasks. The use of a QNR-like bottleneck in  Bear et al. (2020)  has led to strong performance, but small-scale QNR representations can be interchangeable with vector embeddings that lack explicit graph structure (Section 7.3.1). \n Basic Information Flows Potential architectural building blocks range from MLPs and convolutional networks to Transformers and the many varieties of graph neural networks.  1  Architectures can employ building blocks of multiple kinds that collectively enable differentiable end-to-end training (Section 7.2.3 discusses differentiable representations of graph topology). The current state of the art suggests Transformer-based building blocks as a natural choice for encoding NL inputs and generating NL outputs. Transformer-based models have performed well in knowledge-graph → text tasks,  (Ribeiro et al. 2020) , and can in some instances benefit from training with explicit syntax-graph representations.  2  Encoders like those developed for scene-graph representation learning 3 are natural candidates for QNR-mediated vision tasks. In both NL and vision domains, encoders can produce vector/graph representations that, through training and architectural bias, serve as QNRs. GNNs or graph-oriented Transformers are natural choices both for implementing complex operations and for interfacing to task-oriented decoders. Simple feed-forward networks are natural choices for transforming and combining the vector components of vector-labeled graphs. Systems that read and write expressions in QNR corpora could employ scalable near-neighbor lookup in repositories indexed by QNR-derived semantic embeddings (Section 9.1.2). Information flows in generic QNR systems augmented by access to a repository of QNR content. In the general case, \"QNR inference\" includes read/write access to repositories, producing models that are in part pre-trained, but also pre-informed. The analysis presented in previous sections suggests that QNRs can in principle meet the criteria for representing NL + -level semantics, while the capabilities of current neural systems suggest that architectures based on compositions of familiar building blocks can implement the operations required for NL + -mediated functionality. The next question is how to train such systems-how to combine tasks and inductive bias to produce encoders, decoders, and processing mechanisms that provide the intended functionality. \n Shaping QNR Semantics Basic aspects of intended QNR semantics include non-trivial syntax in conjunction with NL-like representational capacity and extensions to other modalities. These goals can be pursued through architectural inductive bias together with training tasks in familiar domains. \n Architectural Bias Toward Quasilinguistic Representations Architectural inductive bias can promote the use of syntactically nontrivial QNRs (rather than flat sequences of embeddings) to represent broadly NL-like content. If learned representations follow the general pattern anticipated in Section 8, QNR syntax would typically employ (at least) DAGs of substantial depth (Section 8.1); leaf attributes would typically encode (at least) lexicallevel semantic information (Section 8.2), while attributes associated with internal nodes would typically encode relationships, summaries, or modifiers applicable to subsidiary expressions (Section 8.3). Encoders and decoders could bias QNRs toward topologies appropriate for expressing quasilinguistic semantics. Architectures can pass information through QNR-processing mechanisms with further inductive biases-e.g., architected and trained to support soft unification-to further promote the expression of computationally tractable, disentangled, quasilinguistic content. \n Anchoring QNRs Semantic Content in NL Although QNR representations have broader applications, it is natural to focus on tasks closely tied to language. Transformers trained on familiar NL → NL objectives (e.g., language modeling and sentence autoencoding) have produced flat vector representations (vectors and vector-sequences) that support an extraordinary range of tasks.  1  Training QNR-bottleneck architectures (NL → QNR → NL) on the same NL → NL objectives should produce comparable (and potentially superior) QNR representations and task performance. Potential tasks include: • Autoencoding NL text • NL language modeling • Multilingual translation • Multi-sentence reading comprehension • Multi-scale cloze and masked language tasks  1  It seems likely that developing NL + -level representations and mechanisms would best be served, not by a pretraining/fine-tuning approach, but by concurrent multitask learning. In this approach, optimization for individual tasks is not an end in itself, but a means to enrich gradient signals and learned representations.  If NL + is to be more than a representation of NL, the training of QNR models may require an inductive bias toward representations that are deliberately decoupled from NL. Lexical-level vector embeddings already provide a useful bias in that they decouple representations from the peculiarities of NL vocabularies. Massively multilingual tasks (translation, etc.) can further encourage the emergence of representations that abstract from the features of particular languages.  1  In current practice, combinations of multitask learning and architectural bias have been employed to separate higher-level and lower-level semantic representations.  2  It may be useful, however, to seek additional mechanisms for learning representations that are abstracted from NL.  3  Supporting this idea, recent work has found that disentangling semantics from NL syntax is practical and can provide advantages in performing a range of downstream tasks  (Huang, Huang, and Chang 2021) . \n Abstracting QNR Representations From Word Sequences Tasks and architectures can be structured to favor separation of abstract from word-level representations.  4  A general approach would be to split and recombine information paths in NL → NL tasks: An abstract QNR path could be trained to represent predominantly high-level semantics and reasoning, while an auxiliary path carries lexical-level information. To recombine these paths, the high-level semantic path could feed a decoder that is also provided with a set of words from the target expression permuted together with decoys.  5  By reducing the task of producing correct NL outputs to one of selecting and arranging elements from a given set of words, this mechanism could shift a lexical-level, NL-specific burden-and perhaps the associated low-level semantic content-away from the abstract, high-level path. To strengthen separation, the gradient-reversal trick for domain adaptation 6 could be applied to actively \"anti-train\" the availability of word-specific information in abstract-path representations. \n Strategies for Learning Higher-Level Abstractions Objective functions in NLP often score outputs by their correspondence to specific sequences of target words. This objective is embedded in the definitions of language modeling, masked language modeling, and typical cloze tasks, while similar objectives are standard in NL translation. However, as the size of target outputs increases-from single-word cloze tasks to filling gaps on the scale of sentences, paragraphs, and beyond-predicting specific word sequences becomes increasingly difficult or effectively impossible. When the actual research objective is to manipulate representations of meaning, lexical-level NL training objectives fail the test of scalability. Completion tasks 1 formulated in the QNR domain itself would better serve this purpose. Useful QNR-domain completion tasks require QNR targets that represent rich task-domain semantics, but we have already seen how NLP tasks can be used for this purpose (Section 10.  3.2) . Products of such training can include NL → QNR encoders that raise both inference processes and their targets to the QNR domain (Figure  10 .3). With targets raised from NL to QNR representations, it should become practical to compare outputs to targets even when the targets represent complex semantic objects with an enormous range of distinct yet nearly equivalent NL representations. While it seems difficult to construct useful semantic-distance metrics over word sequences, semantic-distance metrics in the QNR domain can be relatively smooth.  1  Ambitious examples of completion tasks could include completion of (descriptions of) code with missing functions, or of mathematical texts with missing equations or proofs. \n Training QNR × QNR → QNR Functions to Respect Lattice Structure The semantic-lattice properties discussed in Appendix A1 correspond to an algebra of information, but QNRs need not automatically respect this algebra. In particular, absent suitable training objectives, operations on QNRs may strongly violate the lattice axioms that constrain unification and generalization.  2  Learning representations and functions that approximately satisfy the lattice-defining identities (Section A1.2) can potentially act both as a regularizer and as a mechanism for training operations that support principled comparison, combination, and reasoning over QNR content. Because the existence of (approximate) lattice operations over QNR representations implies their (approximate) correspondence to (what can be interpreted as) an information algebra, we can expect that (approximately) enforcing this constraint can improve the semantic properties of a representational system. In addition, prediction of soft-unification scores (Section A1.4.3) can provide an auxiliary training objective for content summaries (Section 8.3.4), providing a distance measure with potential applications to structuring latent spaces for similarity-based semantic retrieval (Section 9.1.2). \n Processing and Inference on QNR Content The above discussion outlined coarse-grained information flows and general training considerations using block diagrams to represent units of high-level functionality. The present section examines the potential contents of boxes labeled \"QNR inference\". The aim here is not to specify a design, but to describe features of plausible architectures for which the implementation challenges would be of familiar kinds. \n Control, Selection, and Routing Tasks of differing complexity will call for different QNR inference mechanisms. The null case is the identity function, single-path pass-through in a QNRbottleneck architecture. A more interesting case would be a single-path system that performs QNR → QNR transformations (e.g., using a GNN) based on a conditioning input. More powerful inference mechanisms could perform QNR × QNR → QNR operations, potentially by means of architectures that can learn (forms of) soft unification or anti-unification. Toward the high end of a spectrum of complexity (far from entry level!), open-ended QNR-based inference will require the ability to learn task-and data-dependent strategies for storing, retrieving, and operating on QNRs in working memory and external repositories. This complex, high-end functionality could be provided by a controller that routes QNR values to operators while updating and accessing QNR values by means of key and query based storage and retrieval.  1  Keys and queries, in turn, can be products of abstractive operations on QNRs. (In discussions of retrieval, argument passing, etc., \"a QNR\" is operationally a reference to a node in a graph that may be of indefinite extent.) Note that \"reasoning based on QNRs\" can employ reasoning about QNR processing by means of differentiable mechanisms that operate on flat vector representations in a current task context.  2  Reinforcement learning in conjunction with memory retrieval has been effective in multi-step reasoning  (Banino et al. 2020) , as have models that perform multi-step reasoning over differentiable representations and retrieve external information to answer queries  (Bauer, Wang, and Bansal 2018) . \n Encoders QNR encoders accept task inputs (word sequences, images, etc.) and produce sparse-graph outputs. Natural implementation choices include Transformerlike attention architectures that initially process information on a fully connected graph (the default behavior of attention layers) but apply progressively sharpened gating functions in deeper layers. Gating can differentiably weight and asymptotically prune arcs to sparsen graphs that can then be read out as discrete structures. A range of other methods could be applied to this task. Optional discretization at a sparse-graph readout interface breaks differentiability and cannot be directly optimized by gradient descent. This difficulty has been addressed by means that include training-time graph sampling with tools from reinforcement learning  (Kazi et al. 2020 ) and other mechanisms that learn to discretize or sparsen through supervision from performance on downstream tasks  (Malinowski et al. 2018; Zheng et al. 2020) . Systems with potentially relevant mechanisms learn dynamic patterns of connectivity on sparse graphs  (Veličković et al. 2020)  and address problems for which the solution space consists of discrete graphs  (Cappart et al. 2021) . Because 1. In considering how this functionality might be structured, analogies to computer architectures (both neural and conventional) may be illuminating. For example, analogies with storedprogram (i.e., virtually all) computers suggest that memory stores can usefully contain QNRs that describe executable inference procedures. See related work in  Gulcehre et al. (2018) ,  Le, Tran, and Venkatesh (2020) , and Malekmohamadi, Safi-Esfahani, and Karimian-kelishadrokhi (2020). 2. A psychological parallel is the use of general, fundamental \"thinking skills\" in reasoning about declarative memory content. Skills in this sense can be implicit in a processing mechanism (an active network rather than a repository) and are applied more directly than explicit plans. arcs in QNRs can define paths for information flow in computation (e.g., by graph neural networks), methods for training computational-graph gating functions in dynamic neural networks 1 are potentially applicable to learning QNR construction. \n Decoders Standard differentiable neural architectures can be applied to map QNRs to typical task-domain outputs. A natural architectural pattern would employ GNNs to process sparse graphs as inputs to downstream Transformer-like attention models. Where the intended output is fluent natural language, current practice suggests downstream processing by large pretrained language models adapted to conditional text generation; 2 potentially relevant examples include models that condition outputs on sentence-level semantic graphs.  3   \n Working and External Memory Stores Working memory and external repositories have similar characteristics with respect to storage and retrieval, but differences in scale force differences in implementation. In particular, where stores are large, computational considerations call for storage that is implemented as an efficient, scalable, potentially shared database that is distant (in a memory-hierarchy sense) from taskfocused computations.  4  In the approach suggested here, both forms of storage would, however, retrieve values based on similarity between key and query embeddings. \n Unary Operations Unary operations apply to single graphs. Popular node-convolutional GNNs use differentiable message-passing schemes to update the attributes of a graph, and can combine local semantic information to produce context-informed representations. Different networks could be applied to nodes of different semantic types. The values returned by unary operations may be QNRs or embeddings (e.g., keys, queries, or abstractive summaries). Unary operations may also transform graphs into graphs of a different topology by pruning arcs (a local operation), or by adding arcs (which in general may require identifying and linking potentially remote nodes).  1  Examples of neural systems with the latter functionality were noted above.  2  A local topology-modifying operation could (conditionally) pass potential arcs (copies of local references) as components of messages. 3 \n Graph Alignment Graph alignment (\"graph matching\") is a binary operation that accepts a pair of graphs as arguments and (when successful) returns a graph that represents a (possibly partial) node-correspondence relationship between them (Section 7.2.4). Return values could range in form from a node that designates a pair of corresponding nodes in the arguments, to a representation that includes a distinguished set of arcs (potentially labeled with vector embeddings) that represent relationships among all pairs of corresponding nodes. Several neural matching models have been demonstrated, some of which are relatively scalable.  4  Graph alignment could be a pretrained and fine-tuned function. \n Lattice Operations Lattice operations (unification and generalization, Appendix A1). are binary operations that include mechanisms for graph alignment and combination. Soft lattice operations differ from matching in that they return what is semantically a single graph. Like graph alignment, lattice operations could be pretrained and fine-tuned, or could serve as auxiliary training tasks in learning QNR inference. Neural modules pretrained to mimic conventional 1. A related operation would accept a graph referenced at one node and return a graph referenced at another, representing the result of a graph traversal.  Cappart et al. (2021)  3. This is the fundamental topology-modifying operation employed by object capability systems  (Noble et al. 2018 ): A node A with message-passing access to nodes B and C can pass its node-B access (a \"capability\") to node C; node A may or may not retain its access to C afterward. This operation can be iterated to construct arcs between what are initially distant nodes in a graph. Intuitively, access-passing is semantically well motivated if the \"need\" for a more direct connection from B to C can be communicated through messages received by A. See also  Veličković et al. (2020 ). 4. Y. Li et al. (2019 ,  Sarlin et al. (2020) , Y.  Fey et al. (2020)  algorithms for unification and generalization could potentially serve as building blocks for a range of inference algorithms that operate on soft lattices and rich semantic representations.  1   \n Veličković et al. (2020) and \n Potential Application Areas Potential applications of QNR/NL + functionality include and extend applications of natural language. They include human-oriented NLP tasks (translation, question answering, semantic search), but also inter-agent communication and the integration of formal and informal representations to support science, mathematics, automatic programming, and AutoML. QNR/NL + frameworks are intended to support wide-ranging applications both within and beyond the scope of natural language. The present section sketches several potential application areas: first, applications to tasks narrowly centered on language-search, question answering, writing, translation, and language-informed agent behavior-and then a range of applications in science, engineering, mathematics, software, and machine learning, including the general growth and mobilization of knowledge in human society. The discussion will assume success in developing high-level QNR/NL + capabilities. \n Language-Centered Tasks Tasks that map NL inputs to NL outputs are natural applications of NL + -based models. These tasks include internet search, question answering, translation, and writing assistance that ranges from editing to (semi)autonomous content creation. \n Search and Question Answering In search, NL + representations can provide a semantic bridge between NL queries and NL documents that employ different vocabularies. Search and question-answering (QA) models can jointly embed queries and content, enabling retrieval of NL content by semantic similarity search 2 anchored in the NL + domain; beyond comparing embeddings, direct NL + to NL + comparisons can further refine sets of potential search results. Alternatively, language models conditioned on queries (and potentially on models of readers' style preferences) can translate retrieved NL + semantic content to fluent NL answers. QA fits well with document search, as illustrated by Google's information boxes: The response to a search query can include not only a set of documents, but information abstracted from the corpus. In a broader application, NL + -based models could generate extended answers that are more comprehensive, more accurate, and more directly responsive to a query than any existing NL document. With the potential for dense linking (perhaps presented as in-place expansion of text and media), queryresponsive information products could enable browsing of internet-scale knowledge corpora through presentations more attractive and informative than conventional web pages. \n Translation and Editorial Support Translating and editing, like QA, call for interpreting meaning and producing results conditioned on corpus-based content and priors. Differences include a greater emphasis on lengthy inputs and on outputs that closely parallel those inputs, with access to specific knowledge playing a supporting rather than primary role. During training, massively multilingual translation tasks have produced language-invariant intermediate representations (interlinguas 1 ); we can expect similar or better interlingua representations-and associated translations-in systems that employ NL → NL + → NL architectures. Priors based on the frequency of different patterns of semantic content (not phrases) can aid disambiguation of NL source text. The task of machine-aided editing is related to translation: Reproducing semantic content while translating from language to language has much in common with transforming a rough draft into a refined text; stretching the notion of reproducing semantic content, a system might expand notes into text while retaining semantic alignment.  2  It is again natural to exploit priors over patterns of expression to help interpret inputs and generate outputs. Access to knowledge from broad, refined corpora could greatly enrich content when expanding notes. The graph structure of hypertext makes QNRs a good fit to NL + -supported authoring of online NL content. As a specific, high-leverage application, such tools could help contributors expand and improve Wikipedia content. Systems that compile, refine, and access a QNR translation of Wikipedia would be a natural extension of current research on the use of external information stores.  1  Human contributors could play the roles that they do today, but aided by generative models that draw on refined NL + corpora to suggest corrected and enriched content. Most of what people want to express either repeats what has been said elsewhere (but rephrased and adapted to a context), or expresses novel content that parallels or merges elements of previous content. Mechanisms for abstraction and analogy, in conjunction with examples and priors from existing literatures, can support interactive expansion of text fragments and hints to provide what is in effect a more powerful and intelligent form of autocomplete. Similar functionality can be applied at a higher semantic level. Responsible writers seek to avoid factual errors, which could be identified (provisionally!) by clashes between the NL + encoding of a portion of a writer's draft and similar content retrieved from an epistemically high-quality NL + corpus.  2  Writers often prefer, not only to avoid errors, but to inform their writing with knowledge that they do not yet have. Filling semantic gaps, whether these stem from omission or error removal, can be regarded as a completion task over abstract representations  (Section 10.4.2) . Semantically informed search and generative models could retrieve and summarize candidate documents for an author to consider, playing the role of a research assistant; 3 conditional language models, prompted with context and informed by external knowledge 4 could generate substantial blocks of text, playing the role of a coauthor. \n (Semi)Autonomous Content Creation Social media today is degraded by the influence of MIsinformed and unsourced content, a problem caused (in part) by the cost of finding good infor-mation and citing sources, and (in part) by fact-indifferent actors with other agendas. Well-informed replies are relatively costly and scarce, but mob-noise and bot-spew are abundant. As a human-mediated countermeasure, responsible social media participants could designate targets for reply (perhaps with a hint to set direction) and take personal responsibility for authorship while relying on semiautonomous mechanisms for producing (drafts of) content. As a fully autonomous countermeasure, bots created by responsible actors could scan posts, recognize problematic content, and reply without human intervention. Actors that control social media systems could use analogous mechanisms in filtering, where a \"reply\" might be a warning or deletion. Acceptable, fully effective countermeasures to toxic media content are difficult to imagine, yet substantial improvements at the margin may be both practical and quite valuable. \n Agent Communication, Planning, and Explanation Human agents use language to describe and communicate goals, situations, and plans for action; it is reasonable to expect that computational agents can likewise benefit from (quasi)linguistic communication.  1  If NL + representations can be strictly more expressive than NL, then NL + can be strictly more effective as a means of communication among computational agents. Internal representations developed by neural RL agents provide another point of reference for potential agent-oriented communication. Some systems employ memories with distinct, potentially shareable units of information that can perhaps be viewed as pre-linguistic representations (e.g., see  Banino et al. (2020) ). The limitations inherent in current RL-agent representations suggest the potential for gains from language-like systems in which the compositional elements express durable, shareable, strongly compositional abstractions of states, conditions, actions, effects, and strategies. QNR/NL + -based representations can combine unnamable, lexical-level abstractions with lexical-level elements that describe semantic roles, confidence, relative time, deontic considerations, and the like-in other words, semantic elements like those often expressed in NL by function words and TAM-C modifiers  (Section 5.3.3, Section 5.3.4, and Section A3.3) . The role of NL in human communication and cognition suggests that NL + representations can contribute to both inter-and intra-agent performance, sometimes competing with tightly coupled, task-specific neural representations. Communication between humans and RL agents can benefit from language. Although reinforcement learning can enable unaided machines to outperform human professionals even in complex games, 1 human advice conveyed by NL can speed and extend the scope of reinforcement learning.  2  Conversational applications provide natural mechanisms for clarification and explanationin both directions-across machine-human interfaces, potentially improving the human value and interpretability of AI actions. Given suitable NL + descriptions and task-relevant corpora, similarity search could be applied to identify descriptions of similar situations, problems, and potentially applicable plans (including human precedents); mechanisms like those proposed for knowledge integration and refinement (Section 9.4) could be applied to generalize through analogy and fill gaps through soft unification. Widely used content would correspond to \"common sense knowledge\" and \"standard practice\".  3  Like natural language, NL + representations could support both strategic deliberation and concrete planning at multiple scales. Agents with access to large knowledge corpora resemble humans with access to the internet: Humans use search to find solutions to problems (mathematics, travel, kitchen repairs); computational agents can do likewise. Like human populations, agents that are deployed at scale can learn and pool their knowledge at scale. Frequent problems will (by definition) seldom be newly encountered. \n Science, Mathematics, and System Design Although research activities in science, mathematics, engineering, and software development differ in character, they share abstract tasks that can be framed as similarity search, semantic alignment, analogy-building, clash detection, gap recognition, and pattern completion. Advances in these fields involve an ongoing interplay between: • Tentative proposals (hypotheses in science, proof goals in mathematics, design concepts in engineering and software development), • Domain-specific constraints and enablers (evidence in science, theorems in mathematics, requirements and available components in engineering and software development), and • Competition between alternative proposals judged by task-specific criteria and metrics (maximizing accuracy of predictions, generality of proofs, performance of designs; minimizing relevant forms of cost and complexity). These considerations highlight the ubiquitous roles of generative processes and selection criteria, and a range of fundamental tasks in science, mathematics, engineering, and software development can be addressed by generative models over spaces of compositional descriptions. These can be cast in terms of QNR affordances: Given a problem, if a corpus of QNRs contains descriptions of related problems together with known solutions, then similarity search on problemdescriptions 1 can retrieve sets of potentially relevant solutions. Joint semantic alignment, generalization, and analogy-building within problem/solution sets then can suggest a space of alternatives that is likely to contain solutions-or near-solutions-to the problem at hand. In conjunction with an initial problem description, such representation spaces can provide priors and constraints on generative processes, 2 and generated candidate solutions can be tested against task-specific acceptance criteria and quality metrics. These considerations become more concrete in the context of specific task domains. \n Engineering Design [T]hink of the design process as involving, first, the generation of alternatives and, then, the testing of these alternatives against a whole array of requirements and constraints. There need not be merely a single generate-test cycle, but there can be a whole nested series of such cycles. -Herbert  Simon 3  Typical engineering domains are strongly compositional, and aspects of compositionality-modularity, separation of functions, standardization of interfaces-are widely shared objectives that aid not only the design and modeling of systems, but also production, maintenance, and reuse of components across applications. Representations used in the design and modeling of engineering systems typically comprise descriptions of components (structures, circuits, motors, power sources. . . ) and their interactions (forces, signals, power transmission, cooling. . . ). In engineering practice, natural language (and prospectively, NL + ) is interwoven with formal, physical descriptions of system-level requirements, options, and actual or anticipated performance. As Herbert Simon has observed, design can be seen as a generate-and-test process-a natural application of generative models.  1  A wide range of proposed systems can be tested through conventional simulation.  2  In engineering, even novel systems are typically composed (mostly or entirely) of hierarchies of subsystems of familiar kinds.  3  The affordances of QNR search and alignment are again applicable: Embedding and similarity search can be used to query design libraries that describe options at various levels of abstraction and precision; descriptions can be both physical and functional, and can integrate formal and informal information. Semantic alignment and unification provide affordances for filling gaps-here, unfilled functional roles in system architectures-to refine architectural sketches into concrete design proposals. The generation of novelty by soft-lattice generalization and combination operations (Appendix A1) could potentially enable fundamental innovation. Because engineering aims to produce systems that serve human purposes, design specifications-requirements, constraints, and optimization criteriamust fit those purposes. The development of formal specifications is an informal process that can benefit from QNR affordances that include analogy, pattern completion, and clash detection, as well as applications of the commonsense knowledge needed to choose obvious defaults, reject obvious mistakes, and identify considerations that call for human attention. 1. See discussions in  Kahng (2018) ,  Liao et al. (2019), and Oh et al. (2019) . Machine-aided interactive design  (Deshpande and Purwar 2019)  and imitation learning can also help to generate proposals; see ,  Ganin et al. (2021) . 2. Or using ML-based simulation methods, which are of increasing scope and quality. In particular, advances in ML-based molecular simulation (reviewed in  Noé et al. 2020 ) can be expected to facilitate molecular systems engineering. \n Scientific Inquiry Science and engineering often work closely together, yet their epistemic tasks are fundamentally different: Engineering seeks to discover multiple options for achieving purposes, while science seeks to discover uniquely correct descriptions of things that exist. Science and engineering intertwine in practice: Scientists exploit products of engineering (telescopes, microscopes, particle accelerators, laboratory procedures. . . ) when they perform observations and experiments, while engineers engage in science when they ask questions that cannot be answered by consulting models. Potential applications of QNR affordances in science include: • Translating NL publications into uniform, searchable representations  \n Software Development and AutoML Applications of neural ML to software development are under intense exploration.  2  Language models can support interactive, text-based code completion and repair; 3 recent work has demonstrated generation of code based on docstrings.  4  GNNs could operate on structured representations (syntactic and semantic graphs) while also exploiting function names, variable names, comments, and documentation as sources of information and targets for prediction in representation encoding and decoding. QNRs can provide affordances for enriching syntactic structures with semantic annotations and the results of static program analysis,  5  and for wrapping code objects (both implemented and proposed) in descriptions of their requirements and functionality. QNR representations have a different (and perhaps closer) relationship to automated machine learning (AutoML 1 ), because neural embeddings and graphs seem particularly well-suited to representing the soft functionality of neural components in graph-structured architectures. Again, generateand-test processes guided by examples, analogies, and pattern completion could inform search in design spaces, 2 while the scope of these spaces can embrace not only neural architectures, but their training methods, software and hardware infrastructures, upstream and downstream data pipelines, and more. 12 Aspects of Broader Impact The breadth of potential applications of QNR-based systems makes it difficult to foresee (much less summarize) their potential impacts. Leading considerations include the potential use and abuse of linguistic capabilities, of agent capabilities, and of knowledge in general. Systems based on QNR representations promise to be relatively transparent and subject to correction. Potential roles for QNR/NL + -enabled capabilities are extraordinarily broad, with commensurate scope for potential benefits and harms.  3  Channels for potential QNR/NL + impacts can be loosely divided into core semantic functionalities (applications to knowledge in a general sense), semantic functionalities at the human interface (processing and production of natural language content), and potential roles in AI agent implementation and alignment. Most of the discussion here will be cast in terms of the NL + spectrum of potential QNR functionality. \n Broad Knowledge Applications Many of the potential benefits and harms of QNR/NL + -enabled developments are linked to large knowledge corpora and their applications. Several areas of potential impact are closely related to proposed core functionalities of knowledge integration and access.  1. Real et al. (2020)  and  He, Zhao, and Chu (2021)  2. You, Ying, and Leskovec (2020),  Radosavovic et al. (2020), and Ren et al. (2021)  3. For a survey of a range of potential harms, see  Brundage et al. (2018) . \n Integrating and Extending Knowledge Translation of content from NL corpora to corresponding NL + can enable the application of QNR-domain mechanisms to search, filter, refine, integrate, and extend NL-derived content, building knowledge resources for wide-ranging applications. To the extent that improving the quality of knowledge is on the whole beneficial (or harmful), we should expect net beneficial (or harmful) impacts. \n Mobilizing Knowledge Translation of NL expressions (statements, paragraphs, documents. . . ) to corresponding NL + representations promises to improve semantic embeddings and similarity search at scale (Section 9.1.5), helping search systems \"to organize the world's information and make it universally accessible and useful\" (Google 2020) through higher-quality semantic indexing and query interpretation. Generation of content through knowledge integration could go beyond search to deliver information that is latent (but not explicit) in existing corpora. It is reasonable to expect beneficial first-order impacts. \n Filtering Information To the extent that NL → NL + translation is effective in mapping between NL content and more tractable semantic representations, filtering of information 1 in the NL + domain can be used to filter NL sources. Potential applications span a range that includes both reducing the toxicity of social media and refining censorship in authoritarian states. In applications of language models, filtering based on disentangled representations of knowledge and outputs could mitigate leakage of private information. 2 \n Surveillance and Intelligence Analysis Surveillance and intelligence analysis are relatively direct applications of QNRenabled knowledge mobilization and integration, and the balance of impacts on security, privacy, and power relationships will depend in part on how information is filtered, shared, and applied. Mapping raw information into structured semantic representations could facilitate almost any application, with obvious potential harms. To mitigate harms, it will be important to explore how filtering of raw information could be applied to differentially enable legitimate applications: For example, disentangled compositional representations could be more easily redacted to protect sensitive information while providing information necessary for legitimate tasks. \n Producing QNR-Informed Language Outputs at Scale We should expect to see systems that translate NL + content into NL text 1 with fluency comparable to models like GPT-3, and do so at scale. Automated, NL + -informed language production, including support for human writing (Section 11.1.2), could expand quantity, improve quality, and customize the style and content of text for specific groups or individual readers. These capabilities could support a range of applications, both beneficial and harmful. \n Expanding Language Output Quantity Text generation enabled by language models has the potential to produce tailored content for social media economically and at scale: Human writers are typically paid ~0.20 US$/word, 2 ) about 1,000,000 times the cost of querying an efficient Transformer variant.  3  It is reasonable to expect that NL + -informed outputs will have broadly similar costs, orders of magnitude less than the costs of human writing, whether these costs are counted in money or time. Put differently, text output per unit cost could be scaled by a factor on the rough order of 1,000,000. Even when constrained by non-computational costs and limitations of scope, the potential impact of automated text generation is enormous. \n Improving Language Output Quality Applications of language-producing systems will depend in part on domaindependent metrics of output quality: Higher quality can both expand the scope of potential applications and decrease the costs of human supervision, while changing the nature and balance of potential impacts. Relative to opaque language models, systems informed by NL + corpora can improve abilities: • To judge, incorporate, and update factual content • To perform multi-step, multi-source inference • To apply inference to refine and expand knowledge stores Current models based on opaque, learned parameters have difficulties in all these areas; overcoming these difficulties could greatly expand the scope of potential applications. \n Potentially Disruptive Language Products The most obvious societal threats from NL + -based language capabilities stem from their ability to produce coherent content that draws on extensive (mis)information resources-content that mimics the markers of epistemic quality without the substance. The magnitude of this threat, however, must be judged in the context of other, broadly similar technologies. Systems based on large language models are becoming fluent and potentially persuasive while remaining factually unreliable: They can more easily be applied to produce plausible misinformation than informed content. Unfortunately, the current state of social media suggests that fluent, persuasive outputs based on false, incoherent information-whether from conspiracy fans or computational propaganda-can be disturbingly effective in degrading the epistemic environment.  1  This suggests that the marginal harms of making misinformation more coherent, better referenced, etc., may be relatively small.  2  To the extent that capabilities are first deployed by responsible actors, harms could potentially be mitigated or delayed. 1. Existing language models have spurred concerns regarding abuse, including scaling of social-engineering attacks on computer security and of computational propaganda in public discourse  (See Woolley and Howard (2017)  and Howard (  2021 )). In part as a consequence of such concerns  (Solaiman et al. (2019), and Brown et al. (2020) , Section 6.1), OpenAI restricted access to its GPT-3 model. 2. One may hope that influential audiences that have in the past been susceptible to documents with misleading but apparently high-quality content (e.g., academics and policymakers) would also respond to prompt, well-targeted, high-quality critiques of those documents. Responding promptly would leave less time for misleading information to spread unchecked and become entrenched as conventional wisdom. \n Potentially Constructive Language Products Generating low-quality content is easy for humans and machines, and arguments (whether for bad conclusions or good) can cause collateral damage when they inadvertently signal-boost false information; conversely, arguments (regardless of the merits of their conclusions) can produce what might be described as \"positive argumentation externalities\" when their content signal-boosts well-founded knowledge Although the potential harms of facilitating the production of (apparently) high-quality misinformation may be marginal, the potential benefits of facilitating the production of high-quality information seem large. It would be difficult to exaggerate the potential value of even moderate success in damping pathological epistemic spirals and enabling information to gain traction based on actual merit. Authors who employ freely available tools to produce better-written, better-supported, more abundant content (drawing audiences, winning more arguments) could raise the bar for others, driving more widespread adoption of those same tools. Epistemic spirals can be positive.  1  12.3 Agent Structure, Capabilities, and Alignment Section 11.2 discussed NL + representations as potential enablers for agent performance-for example, by supporting the composition of plan elements, retrieval of past solutions, and advice-taking from humans. In considering potential impacts, opportunities for improving transparency and alignment become particularly important. \n Agent Structure A long-standing model of advanced AI capabilities takes for granted a central role for general, unitary agents, often imagined as entities that learn much as a human individual does. The AI-services model 2 challenges this assumption, proposing that general capabilities readily could (and likely will) emerge through the expansion and integration of task-oriented services that-crucially for potential generality-can include the service of developing new services. In the AI-services model, broad knowledge and functionality need not be concentrated in opaque, mind-like units, but can instead emerge through aggregation over large corpora of knowledge and tools, potentially informed both by pre-existing human-generated corpora and by massively parallel (rather than individual) experience of interaction with the world. The AIservices model fits well with the QNR/NL + model of scalable, multimodal knowledge aggregation and integration. \n Agent Capabilities Also in alignment with the AI-services model of general intelligence, the ability of relatively simple agents to access broad knowledge and tool sets 1 could amplify their capabilities. This prospect lends credence to long-standing threat models in which agents rapidly gain great and potentially unpredictable capabilities; the mechanisms differ, but the potential results are similar. Classic AI-risk scenarios commonly focus on AI capabilities that might emerge from an immense, opaque, undifferentiated mass of functionality, a situation in which agents might pursue unexpected goals by unintended means. It may be safer to employ task-oriented agents (and compositions of agents) that operate within action-and knowledge-spaces that are better understood and do not grossly exceed task requirements.  2  Basing functionality on bounded, differentiated resources provides affordances for observing \"what a system is thinking about\" and for constraining \"what an agent can know and do\", potentially powerful tools for interpreting and constraining an agent's plans.  3  Accordingly, developers could seek to bound, shape, and predict behaviors by exploiting the relative semantic transparency of proposed QNR/NL + corpora to describe and structure the knowledge, capabilities, constraints, and objectives of task-oriented agents. \n Agent Alignment Many of the anticipated challenges of aligning agents' actions with human intentions hinge on the anticipated difficulty of learning human preferences.  4  The ability to read, interpret, integrate, and generalize from large corpora of human-generated content (philosophy, history, news, fiction, court records, discussions of AI alignment. . . ) could support the development of richly informed models of human preferences, concerns, ethical principles, and legal systems-and models of their ambiguities, controversies, and inconsistencies.  1  Conversational systems could be used to test and refine predictive models of human concerns by inviting human commentary on actual, proposed, and hypothetical actions. NL + systems that fulfill their promise could model these considerations more effectively than human language itself, in a way that is not fully and directly legible, yet open to inspection though the windows of query and translation. \n Conclusions Current neural ML capabilities can support the development of systems based on quasilinguistic neural representations, a line of research that promises to advance a range of research goals and applications in NLP and beyond. Natural language (NL) has unrivaled generality in expressing human knowledge and concerns, but is constrained by its reliance on limited, discrete vocabularies and simple, tree-like syntactic structures. Quasilinguistic neural representations (QNRs) can generalize NL syntactic structure to explicit graphs (Section 8.1) and can replace discrete NL vocabularies with vector embeddings that convey richer meanings than words (Section 5.3, Section 8.2). By providing affordances for generalizing and upgrading the components of NL-both its structure and vocabulary-QNR systems can enable neural systems to learn \"NL + \" representations that are strictly more expressive than NL. Machines with human-like intellectual competence must be fully literate, able not only to read, but to write things worth reading and retaining as contributions to aggregate knowledge. Literate machines can and should employ machine-native QNR/NL + representations (Section 8) that are both more expressive and more computationally tractable than sequential, mouthand-ear oriented human languages. Prospects for QNR/NL + systems make contact with a host of fields. These include linguistics (Section 5), which offers insights into the nature of expressive constructs in NL (a conceptual point of departure for NL + ), as well as current neural ML, in which vector/graph models and representation learning provide a concrete basis for potential QNR implementations (Section 10). Considerations that include local compositionality (Section 4.3) suggest that vector/graph constructs can provide computationally tractable representations of both complex expressions and the contexts in which they are to be interpreted  (Section 8.3) . Drawing on existing NL corpora, QNR-based systems could enable the construction of internet-scale NL + corpora that can be accessed through scalable semantic search (Section 9.1), supporting a powerful ML analogue of long-term memory. In addition, QNR/NL + frameworks can support unification and generalization operations on (soft, approximate) semantic lattices (Appendix A1), providing mechanisms useful in knowledge integration and refinement (Section 9.4, Section 9.5). Applications of prospective QNR/NL + functionality could support not only epistemically well-informed language production (Section 11.1), but the growth and mobilization of knowledge in science, engineering, mathematics, and machine learning itself (Section 11.3). The fundamental technologies needed to implement such systems are already in place, incremental paths forward are well-aligned with research objectives in ML and machine intelligence, and their potential advantages in scalability, interpretability, cost, and epistemic quality position QNR-based systems to complement or displace current foundation models  (Bommasani et al. 2021)  at the frontiers of machine learning. \n A1 Unification and Generalization on Soft Semantic Lattices QNR representations can support operations that combine, contrast, and generalize information. These operations-soft approximations of unification and anti-unification-can be used to implement continuous relaxations of powerful mechanisms for logical inference. A range of formal representations of meaning, both in logic and language, have the structure of mathematical lattices. Although the present proposal for QNR systems (and aspirational NL + systems) explicitly sets aside the constraint of formality, approximate lattice structure emerges in NL and will (or should, or readily could) be a relatively strong property of QNRs/NL + . Because lattices can provide useful properties, it is worth considering the potential roles and applications of lattice structure in QNR-based systems. Note that the fundamental goals of NL + -general superiority to NL in expressiveness and computational tractability-do not require lattice properties beyond those that NL itself provides. The ability to provide stronger lattice properties is a potential (further) strength of NL + , not a requirement. In other words, lattice properties are natural and useful, yet optional. The following sections begin by discussing the motivation for considering and strengthening lattice properties-supporting meet and join, a.k.a. unification and generalization-in light of their potential roles and utility. A brief review of approximate lattice structure in NL provides an initial motivation for applying lattice structure in NL + within the scope of a methodology that avoids commitment to formal models. Consideration of lattices in logic and in constraint logic programming further motivates the pursuit of approximations, and introduces a discussion, in part speculative, regarding inductive bias and prospective, emergent lattice-oriented QNR representations. This topic is adjacent to many others, creating a large surface area that precludes any compact and comprehensive discussion relationships to existing work.  1  A sketch of these relationships and pointers into relevant literatures provide starting points for further reading. \n A1.1 Motivation Typical expressions can be regarded as approximate descriptions of things, whether ambiguous (pet, rather than cat) or partial (grey cat, rather than big grey cat). Given two expressions, one may want to combine them to form either a narrower description (by combining their information) or a broader description (by combining their scope). Although many other operations are possible (averaging, perhaps, or extrapolation), narrowing and broadening are of fundamental importance, and in many semantic domains, quite useful. They can be construed as operations on a formal or approximate semantic lattice. A1.1.1 Why Unification (Meet, Intersection, Narrowing, Specialization)? In symbolic logic, expressions correspond to points in a lattice (defined below), and unification is an operation that combines two expressions to form a more specific expression by combining their compatible information.  1  In generic lattices, the corresponding operation is termed meet, which in many contexts can be regarded as an intersection of sets or regions. Unification combines compatible information; failures of unification identify clashes.  2  The Prolog language illustrates how unification and failures can enable reasoning and proof. A1.1.2 Why Anti-Unification (Join, Union, Broadening, Generalization)? Alternatively, two expressions may provide (partial) descriptions of two entities of the same kind. Here, a natural goal is to describe properties common to all things of that kind; clashes between aspects of their descriptions indicate that those aspects are not definitional. In a lattice of expressions, this form of generalization is termed antiunification,  3  which increases generality by discarding clashing or unshared information. In generic lattices, the corresponding operation is termed join, 1. More precisely, unification is an operation that yields the most general instance of such an expression. For an application-oriented overview, see  Knight (1989) . 2. When unification attempts to combine partial information about a single entity, it is natural for inconsistent information to imply failure. which in many instances corresponds to a union of sets or regions.  1  Antiunification has applications in NLP and can be used to form generalizations and analogies in mathematics.  2  In neural ML, approximations of antiunification could potentially inform priors for a distribution of unseen instances of a class. \n A1.1.3 Hypotheses Regarding Roles in QNR Processing Considerations explored in this appendix suggest a range of hypotheses regarding potential roles for approximate lattice representations and operations in QNR processing: • That lattice representations and operations (\"lattice properties\") can, in fact, be usefully approximated in neural systems. • That learning to approximate lattice properties need not impair general representational capacity. • That approximations of meet and join operations will have broad value in semantic processing. • That lattice properties can be approximated to varying degrees, providing a smooth bridge between formal and informal representations. • That lattice properties can regularize representations in ways that are useful beyond enabling approximate meet and join. • That approximations of lattice representations and operations are best discovered by end-to-end learning of neural functions. • That explicit, approximate satisfaction of lattice identities can provide useful auxiliary training tasks. \n A1.2 Formal Definitions A mathematical lattice is a partially ordered set that can in many instances be interpreted as representing the \"inclusion\" of \"subsumption\" of one element by another. Axiomatically, a lattice has unique meet and join operations, ∧ and ∨: the meet operation maps each pair of elements to a unique greatest lower bound (infimum), while join maps each pair of elements to a unique least upper bound (supremum). Like the Boolean ∧ and ∨ operators (or the set operators ∩ and ∪), meet and join are associative, commutative, idempotent, and absorptive. A bounded lattice will include a unique bottom element, \"⊥\" (in the algebra of sets, ∅; here a universally incompatible meaning, \"<nil>\"), and a unique top element, \" \" (in the algebra of sets, U; here, an all-embracing generalization, \"<any>\"). In a formal infix notation, operators on a bounded lattice satisfy these identity axioms: Idempotence: A ∧ A = A ∨ A = A Commutativity: A ∧ B = B ∧ A, A ∨ B = B ∨ A Associativity: A ∧ (B ∧ C) = (A ∧ B) ∧ C, A ∨ (B ∨ C) = (A ∨ B) ∨ C Absorptivity: A ∧ (A ∨ B) = A ∨ (A ∧ B) = A Boundedness: A ∧ = A, A ∨ = , A ∨ ⊥ = A, A ∧ ⊥ = ⊥ If ∧ and ∨ are implemented as functions (rather than as structural features of a finite data structure), A ∧ B → C and A ∨ B → D will necessarily satisfy the unique-infimum and unique-supremum conditions. Commutativity can be ensured by algorithmic structure, independent from learned representations and operations; idempotence, associativity and absorptivity perhaps cannot. Boundedness is straightforward. 1 \n A1.3 Lattice Structure in NL Semantics The fit between lattice orders and NL semantic structures is well known,  1  and the term \"semantic lattice\" has been applied not only to word and symbolbased representations, but to neural representations of images.  2  In studies of NL, both \"concepts\" and \"properties\" have been modeled in lattice frameworks, applications that speak in favor of explicit lattice structure in NL + frameworks. \n A1.3.1 Models of NL Semantics: A Lattice of Concepts Formal concept analysis 3 can recover \"concept lattice\" structures from text corpora, and these structures have been argued to be fundamental to information representation. Set-theoretic approaches associate formal objects with formal attributes, and construct a subsumption lattice over sets of defining attributes. Formal concept analysis has been extended to fuzzy structures in which possession of an attribute is a matter of degree.  4  Note that lattice relationships depend on context: In the context of households, the join of \"cat\" and \"dog\" might be \"pet\", but in the context of taxonomy, the join would be \"carnivora\". In practice, expressions containing \"cat\" and \"dog\" would be considered not in isolation, but in some context; in NLP, context would typically be represented dynamically, as part of a computational state; in QNR processing, context could be included as an abstractive vector attribute (see Section 8.3.5). \n A1.3.2 Models of NL Semantics: A Lattice of Properties Properties of things may have values distributed over a continuous range, and properties associated with something may themselves specify not a precise value, but a range within the range: In NL, \"light gray\" does not denote a precise color, and inference from an description of a \"light gray object\" may only loosely constrain its reflectance. Relationships among descriptions that specify ranges of properties may correspond to an interval lattice (Section A1.5.3). 1 \n A1.4 Logic, Constraint Systems, and Weak Unification The previous section focused on lattice relationships among individual entities, but such entities can also serve as attributes in expressions or general graphs, enabling the definition of expression-level lattice operations. Lattices over expressions in which attributes themselves have non-trivial lattice orders can provide powerful, tractable representations in logic and constraint programming. Computation over such representations can be extended to include weak unification. \n A1.4.1 Logical Expressions, Logic Programming Logic programming performs reasoning based on syntactic unification of expression-trees in which symbols represent attributes of leaf nodes (variables or constants) or interior nodes (e.g., functions, predicates, relations, and quantifiers). In Prolog, expressions are limited to a decidable fragment of first-order logic; more powerful unification-based systems include λ terms and can support proof in higher-order logics.  2  Informally, first-order terms A and B unify to yield an expression C provided that all components of A (subexpressions and their attributes) unify with corresponding features or variables of B; C is the expression that results from the corresponding substitutions. Function and predicate symbols unify only with identical symbols; constants unify with variables or identical constants; variables unify with (and in the resulting expression, are replaced by) any structurally corresponding constant, variable, or subtree.  3  Aside from variables that match subtrees, tree structures must match. As required, unification of two expressions then either fails or yields the most general expression that specializes both. Informally, expressions A and B anti-unify, or generalize, to yield C provided that C contains all features that A and B share, and contains variables wherever A and B differ in attributes or structure, or where either contains a variable. C is the unique, most specific expression that can unify with any expression that can unify with either A or B. Thus, join/anti-unification of two expressions yields the most specific expression that generalizes both. \n Relevance to QNR systems: QNR frameworks can embed (at least) first-order logic expressions and enable their unification, provided that some attributes (representing constants, functions, etc.) can be compared for equality, while others (acting as variables 1 ) are treated as features that match any leaf-attribute or subexpression. Accordingly, QNR frameworks augmented with appropriate algorithms can support logical representation and reasoning. This is a trivial consequence of the ability of QNRs to represent arbitrary expressions, in conjunction with freedom of interpretation and the Turing completeness of suitable neural models. Logical expressions and logic programming are, however, instances of richer systems-also within the potential scope of QNRs-that represent constraint systems and support constraint logic programming. \n A1.4.2 Constraints and Constraint Logic Programming In constraint logic programming, 2 representations are extended to include constraints more general than equality of components and binding to unconstrained variables, and unification is extended to (or replaced by) constraint satisfaction. The application of constraints narrows the variable domains, and unification fails when domains become empty. The attributes of constraint expressions have a lattice structure, as do the expressions that contain them, and constraint expressions can be narrowed and generalized though unification and anti-unification  (Yernaux and Vanhoof 2019) . The potential complexity of constraints spawns a vast menagerie of constraint systems, algorithms, and constraint logic programming algorithms. Provided that expressions can include both single-element domains and variables able to bind subexpressions, constraint logic programming can subsume conventional logic programming. \n Relevance to QNR systems: As with logic and logic programming, the generality of QNR representations and neural computation implies the ability to represent constraint 1. Convergent arcs in DAG expressions model the behavior of named variables that occur in multiple locations. It should be noted that unification can produce and operate on cyclic graphs unless this is specifically excluded; see  Smolka (1992) . systems and constraint-based computation, including constraint logic programming. The generalization from logic to constraints is important to semantic representational capacity: Expressions can often be interpreted as denoting regions in a semantic space,  1  and combinations of expressions can combine constraints. Constraint logic programming provides an exact formal model of computation based on this semantic foundation. \n A1.4.3 Weak and Soft Lattice Operations \"Never express yourself more clearly than you are able to think\" -Niels Bohr The literature describes a range of models of NL semantics and reasoning based on a range of approximate unification operations. These typically replace equality of constants (functions, etc.) with similarity: In \"soft unification\" (as the term is typically used in the literature) the operation succeeds if similarity (for example, cosine similarity between vectors,  Arabshahi et al. (2021) ) is above some threshold, and success may yield either conventional binding of variables to values  (Campero et al. 2018) , or merged representations of values  (Cingillioglu and Russo 2020) . \"Weak unification\" may produce a \"unification score\" that indicates the quality of unification; these scores can be carried forward and combined to score the quality of multi-step inference operations.  2  As used in the present context, the term \"soft unification\" subsumes both \"weak\" and \"soft\" unification as used in the literature, and entails combining representations of values in a way that approximates unification operations in constraint logic programming. Thus, \"softness\" allows operations that violate strict lattice identities.  3  The intended class of QNR (soft-)unification operations follows constraint logic programming in generalizing from the binding of constants to named (in effect, shared) unconstrained variables to include the narrowing of shared (in effect named), potentially constrained attributes. As in logic programming, QNR unification will permit the unification of subexpressions with variable-like attributes, but differs in that constrained attributes (unlike variables) may impose semantically non-trivial constraints on permissibility and on the content of resulting subexpressions (Section A1.6.4). \n A1.5 Exact Lattice Operations on Regions Although embeddings in vector/graph representations denote points in semantic spaces, their semantic interpretations will typically correspond to regions in lower-dimensional spaces. Comparisons to symbolic and more general constraint representations can provide insights into potential QNR representations and reasons for expecting their lattice properties to be inexact. \n A1.5.1 Conventional Symbolic Expressions In conventional expression graphs, attributes comprise symbols that represent points together with symbols that represent unbounded regions in the space of expressions. Thus, individual attribute subspaces are simple, have no spatial structure, and accordingly exhibit trivial behavior under unification and anti-unification. \n A1.5.2 General Region Representations and Operations Leaf attributes can represent regions in R n , but options for their unification and generalization may be representation-dependent. The conceptually straightforward definition is both trivial and problematic: Treating regions as sets of points (A ∨ B = A ∪ B) in effect discards spatial structure, and with it the potential for non-trivial generalization. Further, if region representations have limited descriptive capacity, then the result of generalizing a pair of attributes by set union cannot in general be represented as an attribute.  1  Alternatively, the generalization of two volumes might be defined as their convex hull. Generalization of convex regions yields convex regions, and inclusion of points outside the initial volumes reflects spatial structure and a plausible notion of semantic generalization. Unfortunately, this definition can also fall afoul of limited descriptive capacity, because the convex hull of two regions can be more complex than either. 1 \n A1.5.3 Interval (Box) Representations and Operations There are region-representations for which lattice operations are exact, for example, one-dimensional intervals in R 2 and their generalization to axisaligned boxes in R n .  3  Unification of a pair of box-regions yields their intersection; anti-unification yields the smallest box that contains both. Axis-aligned box regions can be represented by vectors of twice the spatial dimensionality (for example, by pairing interval centers with interval widths), and lattice operations yield representations of the same form. Interval-valued attributes have been applied in constraint logic programming.  4  Generalization through anti-unification of intervals has a natural semantic interpretation: Points in a gap between intervals represent plausible members of the class from which the intervals themselves are drawn. 1. To illustrate, the convex hull of two spheres need not be a sphere, and the convex hull of two polytopes may have more facets than either. Intersections can also become more complex (see Jaulin 2006).  Clark et al. (2021) . \n Discussed for example in 3. Affine (and other) transformations of a space and its regions can of course maintain these properties. 4.  Benhamou and Older (1997)  and  Older and Vellino (1990)   \n A1.6 Approximate Lattice Operations on Regions If box lattices are exact, why consider approximate operations on more general regions? The basic intuition is that representations entail trade-offs, and that greater flexibility of form is worth some degree of relaxation in the precision of unification and generalization. Boxes have sharp boundaries, flat facets, and corners; natural semantic representations may not. Intervals in a space of properties may correspond to natural semantic concepts, yet orthogonality and global alignment of axes may not. In the present context, it is important to distinguish two kinds of approximation: As discussed in Section A3.4, effective QNR frameworks must be able to express, not only precise meanings, but ranges or constraints on meaning-a kind of approximation in the semantic domain that differs from approximation of lattice properties. Ranges of meanings can be represented as regions in semantic spaces, while region-representations that approximate precise meanings can precisely satisfy the lattice axioms.  \n A1.6.1 Continuous-Valued Functions In addition, general considerations motivate representing the membership of points in semantic classes with continuous-valued functions, rather than functions having values restricted to {1, 0}. Such representations invite further approximations of lattice properties. \n A1.6.2 Region Shapes and Alignment It is natural to exploit the flexibility of neural network functions to represent generalized region shapes, and to use the ability of fully connected layers to free representations from preferred axis alignments and thereby allow exploitation of the abundance of nearly orthogonal directions in high dimensional spaces. Learned attribute representations need not describe box-like regions or share meaningful alignment in different regions of a semantic space. \n A1.6.3 Satisfaction of Lattice Axioms as an Auxiliary Training Task Given the value of lattice structure in representations, it is natural to speculate that promoting lattice structure through inductive bias may aid performance in a range of semantic tasks. Auxiliary training tasks in which losses explicitly measure violations of lattice properties may therefore be useful components of multitask learning. 1 \n A1.6.4 Unifying Attributes with Expressions As noted above, conventional symbolic expressions allow unification of unconstrained variables with subexpressions, while in the QNR context, it is natural to seek to unify subexpressions with constrained variables-attributes that represent semantic regions. The outlines of desirable behavior are clear, at least in some motivating cases: Consider a pair of graphs with subexpressions A and B in corresponding locations. Let A be a vector representation (e.g., describing a generic greystriped animal with sharp claws), while a corresponding subexpression B is a vector/graph representation (e.g. that contains components that describe a cat's temperament, ancestry, and appearance). Unification should yield a vector/graph representation in which the properties described by vector A constrain related properties described anywhere in expression B (e.g., the cat's appearance, paws, and some aspects of its ancestry) If some component of B specifies a black cat, unification fails. In this instance, generalization should yield a vector representation that does not clash with properties described in either A or B, while discarding properties that are not shared. \n A1.6.5 Learning Lattice-Oriented Algorithms Classic algorithms for unification and generalization implement particular patterns of information flow, intermediate representations, and iterative, conditional computation. Work on supervised neural algorithmic learning illustrates one potential approach to adapting such algorithms to neural computation, an approach that supervises the learning of algorithmic structure while allowing end-to-end learning of rich representations and corresponding decision criteria. 1 \n A1.7 Summary and Conclusions Lattice structure is found in formal systems and (to some extent) in NL, and it seems both natural and desirable in QNR/NL + representations. Although lattice structure is not a criterion for upgrading NL to NL + representations, substantial adherence to lattice structure could potentially improve expressive capacity in a systemic sense and need not imply the rigidity of fully formal representations. Because linguistic representations and reasoning are themselves approximate, there seems little reason to sacrifice representational flexibility in order to enforce exact and universal satisfaction of the lattice axioms. A QNR framework that embraces both precise and approximate lattice relationships can enable both formal and informal applications of those relationships to formal and informal reasoning. The literatures on conventional and constraint logic programming illustrate the power and computational tractability of algorithms based on systems of this kind. The potential benefits of approximate lattice structure may emerge spontaneously, but can also be pursued by inductive bias, including training that employs satisfaction of lattice axioms as an auxiliary task in multitask learning. A2 Tense, Aspect, Modality, Case, and Function Words Tables of examples illustrate expressive constructs of natural languages that do not reduce to nouns, verbs, and adjectives. Natural languages express a range of meanings through closed-class (\"function\") words, and express distinctions of tense, aspect, modality, and case though both function words and morphological features. Section 5.3 discusses the roles and importance of these constructs; this appendix provides several brief tables of examples. Languages differ in the distinctions that they can compactly express, while properties like \"remoteness\", \"completion\", and \"causation\" invite continuous representations.) Perfect tenses -completed in past, present, or future (\"had/has/will have\" finished) Continuous tenses -ongoing in past, present, or future (\"was/am/will be\" working) Past perfect continuous -previously ongoing in the past (\"had been working\") Future perfect continuous -previously ongoing in the future (\"will have been working\") Remote perfect -completed in the remote past (Bantu languages 1 ) Resultative perfect -completed past action causing present state (Bantu languages) Section 6 discussed QNR architectures as a framework, considering potential components, syntactic structures, and their semantic roles. The present section extends this discussion to explore in more detail how anticipated QNR frameworks could subsume and extend the expressive capabilities of natural languages to fulfill the criteria for NL + . Key considerations include facilitating representation learning, upgrading expressiveness, improving regularity, and enabling more tractable reading, interpretation, and integration of content at scale. Potential NL/NL + relationships discussed here are not proposals for handcrafted representations, nor are they strong or confident predictions of the results of representation learning. The aim is instead to explore the scope and strengths of QNR expressive capacity, with potential implications for design choices involving model architectures, inductive bias, and training tasks. Where neural representation learning provides NL + functionality by different means, we should expect those means to be superior. \n A3.1 Upgrading Syntactic Structure To be fit for purpose, NL + syntactic structures must subsume and extend their NL counterparts while improving computational tractability: • To subsume NL syntactic structures, NL + frameworks can embed NL syntax; as already discussed, this is straightforward. • To extend NL syntactic structures, NL + frameworks can support additional syntactic structure; as already discussed, this can be useful. • To improve tractability in learning and inference, NL + frameworks can improve semantic compositionality, locality, and regularity. This potential NL/NL + differential will call for closer examination. \n A3.1.1 Making Syntactic Structure Explicit Explicit representations avoid the computational overhead of inferring structure from strings, as well as costs (and potential failures) of non-local disambiguation, e.g., by using DAGs to represent coreference. Improved locality and tractability follow. \n A3.1.2 Extending the Expressive Scope of Syntactic Structure Explicit graphs can enable the use of structures more diverse and complex than those enabled by natural language and accessible to human cognition.  1  Writers grappling with complex domains may employ supplementary representations (diagrams, formal notation), or may simply abandon attempts to explain systems and relationships that involve deeply nested structures, heterogeneous patterns of epistemic qualification, and complex patterns of coreference-all of which must in NL be encoded and processed as sequences. More general representations can directly describe relationships that are more complex yet readily accessible to machines. \n A3.1.3 Extending the Concept of \"Syntactic Structure\" \"Natural language syntax\" as understood by linguists is often in practice supplemented with visually parsable structures such as nested text (outlines, structured documents) and tables that represent grids of relationships. Diagrams may embed text-labels as attributes of graphs that represent networks of typed relationships (taxonomy, control, causation, enablement, etc.); program code is adjacent to NL, yet goes beyond NL concepts of syntax. Implementations that support explicit bidirectional links can model the structure of literatures that support not only \"cites\" but also \"cited by\" relationships. \n A3.1.4 Collapsing Syntactic Structure In neural-network computation, vector operations are basic, while graph operations may impose additional computational overheads. This motivates the use of embeddings in preference to graph expressions, 2 which in turn 1. What is in some sense accurate linguistic encoding does not ensure successful communication: For example, as a consequence of working-memory constraints, humans suffer from \"severe limitations\" in sentence comprehension  (Lewis, Vasishth, and Van Dyke 2006) . highlights the value of highly expressive lexical-level units. In addition, encoding different meanings as differences in syntactic structure can impede alignment and comparison of expressions; When individual embeddings can express a range of meanings that NL would represent with different syntactic structures, those syntactic differences disappear. \n A3.2 Upgrading Lexical-Level Expressive Capacity Where possible, we would like to replace the syntactic compositionality of words with the simpler, arithmetic compositionality of vectors. Because the best syntax is no syntax, it is worth considering what kinds of semantic content can be represented in vector spaces without relying on graph structure. The following discussion notes several kinds of semantic structure that exist in NL, can be represented by embeddings, and have emerged (spontaneously and recognizably) in neural models. \n A3.2.1 Subsuming and Extending Content-Word Denotations As careful writers know, there often is no single word that accurately conveys an intended meaning, while to unpack an intended meaning into multiple words may be too costly in word-count or complexity. Lexical-level embeddings can shift the ambiguity/verbosity trade-off toward expressions that are both less ambiguous and more concise.  1  Embeddings can place content-word meanings in spaces with useful semantic structure. Word embeddings demonstrate learnable structure in lexical-level semantic spaces. The geometry of learned word embeddings can represent not only similarity, but analogy, 2 and representations of semantic differentials across vocabularies 3 can be identified in language models. Unfortunately, the role of NL word embeddings-which must represent polysemous, context-dependent words-precludes clean representation of word-independent semantic structure. Vector spaces that represent meanings rather than words can provide semantic structure that is more reliable 1. Concise expression is of course less necessary in the context of indefatigable machine intelligence. 2. As discussed in Chen, Peterson, and Griffiths (2017). 3. Daniel Kahneman's doctoral dissertation examines semantic differentials  (Kahneman 1961) . Semantic structure embeddable in vector spaces has been extensively studied by both linguists and ML researchers (e.g., see  Messick (1957) ,  Sagara et al. (1961 ), Hashimoto, Alvarez-Melis, and Jaakkola (2016 ), and Schramowski et al. (2019 ). and useful, hence patterns observed in vector representations of natural language provide only limited insight into the potential generality and utility of semantically structured vector spaces. Embeddings can disambiguate, interpolate, and extend vocabularies of content words. Because points in an embedding space could be used to directly designate every word in every NL vocabulary-with vast capacity to spare-embeddings can be strictly more expressive than NL words. Again taking NL as a baseline, embeddings offer the advantage of avoiding polysemy, maladaptive word ambiguity, 1 and a large (even comprehensive?) range of recognizably missing word meanings (the frustrating-thesaurus problem). In suitable semantic spaces, points within and beyond the rough equivalents of NL word clusters can, at a minimum, interpolate and extend the rough equivalents of NL vocabularies. \n Embeddings can extend vocabularies by folding modifier-expressions into noun and verb spaces. The ability to collapse a range of multi-word expressions into embeddings is equivalent to extending vocabularies: In a straightforward example, adjectives and adverbs can modify the meanings of nouns and verbs to produce different lexical-level meanings. These modifiers often represent cross-cutting properties 2 that can describe things and actions across multiple (but not all) domains. Although words with modifiers can be viewed as extensions of NL vocabularies, expressions of limited size necessarily leave gaps in semantic space; continuous vector embeddings, by contrast, can fill regions of semantic space densely. As a consequence of the above considerations, a function of the form NLencode: (lexical-NL-expression) → (lexical-embedding) can exist, but not its inverse. NL-encode is neither injective nor surjective: Multiple NL expressions may be equivalent, and typical NL + expressions will have no exact NL translation. \n Directions in embedding spaces can have interpretable meanings. Linguists find that NL words can with substantial descriptive accuracy be positioned in spaces in which axes have conceptual interpretations 3 or reflect semantic differentials. In NL, modifiers commonly correspond to displacements with components along these same axes. It is natural to interpret modifier-like vector components differently in different semantic domains.  1  As noted previously, the interpretation of directions in a (sub)space that corresponds to differentials applicable to entities would naturally depend on location-on which regions of a (sub)space correspond to which kinds of entities.  2  In a semantic region that describes persons, for example, directions in a subspace might express differences in health, temperament, age, and income, while in a semantic region that describes motors, directions in that same subspace might express differences in power, torque, size, and efficiency. \n Lexical level and syntactic compositionality are complementary. As suggested above, vector addition of modifiers and other differentials can express compositional semantics within embeddings. Exploiting this capacity does not, of course, preclude syntactic compositionality: QNRs can support compositional representations through both vector-space and syntactic structure. Without predicting or engineering specific outcomes, we can expect that neural representation learning will exploit both mechanisms. \n A3.2.2 Image Embeddings Illustrate Vector Compositionality in Semantic Spaces Image embeddings can provide particularly clear, interpretable examples of the expressive power-and potential compositionality-of continuous vector representations. (Section 9.2 discusses image and object embeddings, not as examples of representational power, but as actual lexical units.) Humans are skilled in perceiving systematic similarities and differences among faces. Diverse architectures (e.g., generative adversarial networks, variational autoencoders, and flow-based generative models 3 ) can produce face embeddings that spontaneously form structured semantic spaces: These models can represent faces in high-dimensional embedding spaces that represent kinds of variations that (qualitatively, yet clearly) are both recognizable and systematic.  1  Many papers present rows or arrays of images that correspond to offsets along directions in embedding spaces, and these naturally emphasize variations that can be named in captions (gender, age, affect. . . ), but a closer examination of these images also reveals systematic variations that are less readily described by words. Words or phrases of practical length cannot describe ordinary faces such that each would be recognizable among millions. A single embedding can.  2  Similar power can be brought to bear in a wider range of lexical-level representations. 3 \n A3.3 Subsuming and Extending Function-Word/TAM-C Semantics As noted in Section 5.3.4 TAM-C meanings can be encoded in either word morphology or function (closed-class) words. Some TAM-C modifiers are syntactically associated with lexical-level units; others are associated with higher-level constructs. TAM-C modifiers that represent case (e.g., nominative, accusative, instrumental, benefactive; see Table  A2 .4) can directly describe the roles of things denoted by words (e.g., acting vs. acted upon vs. used), but case also can indirectly modify the meaning of a word-a rock regarded as a geological object differs from a rock regarded as a tool.  4  Other TAM-C modifiers express semantic features such as epistemic confidence, sentiment, and use/mention distinctions. In NL, statement-level meanings that are not captured by available TAM-C modifiers may be emergent within an expression or implied by context; by compactly and directly expressing meanings of this kind, expression-level embeddings can provide affordances for improving semantic locality, compositionality, and clarity.  5  Some function words connect phrases: These include words and combinations that express a range of conjunctive relationships  (and, or, and/or, therefore, then, still, however, because, despite, nonetheless. . . )  and capability/intention related relationships  (can, could, should, would-if, could-but, would-if-could, could-but-shouldn't. . . ) . Consideration of the roles of these words and constructs will show that their meanings are distributed over semantic spaces, and the above remarks regarding the use of embeddings to interpolate and extend NL meanings apply.  1  In NL, tense/aspect modifiers express distinctions in the relative time and duration of events, and because these modifiers reference a continuous variable-time-they can naturally be expressed by continuous representations. Likewise, epistemic case markers in NL (indicative, inferential, potential) implicitly reference continuous variables involving probability, evidence, and causality. Note that much of function-word/TAM-C space represents neither kinds nor properties of things, and is sparsely populated by NL expressions. The use of embeddings to interpolate and extend meanings in these abstract semantic roles could greatly improve the expressive capacity of QNR/NL + frameworks relative to natural languages. \n A3.4 Expressing Quantity, Frequency, Probability, and Ambiguity Discrete NL constructs express a range of meanings that are more naturally expressed in continuous spaces: These include number, quantity, frequency, probability, strength of evidence, and ambiguity of various kinds. NL can express specific cardinal, ordinal, and real numbers, and absolute concepts such as none or all, but many other useful expressions are either crude (grammatical singular vs. plural forms) or ambiguous (e.g., several, few, many, some, most, and almost all, or rarely, frequently, and almost always). Note that intentional ambiguity is useful: Few does not denote a particular number, but a range of \"small\" numbers, either absolute (about 2 to 5, Munroe (2012)) or relative to expectations regarding some set of entities. This range of meanings (likewise for unlikely, likely, possibly, almost certainly, etc.) invites continuous representations in QNR frameworks.  2  Similar remarks apply to qualitative and probabilistic hedges (mostly, partially, somewhat, to some extent) qualifiers and often agent-centered epistemic qualifiers (presumably, if I recall correctly, it seems to me, as far as I know, in my opinion, etc.).  1  One would also like to be able to compactly express qualifiers like illustrative but counterfactual simplification, approximate description of a typical case, and unqualified statement but with implied exceptions: Today, the absence of universal idioms for expressing these meanings gives rise to gigabytes of fruitless, argumentative noise in internet discussions. The discussion above implicitly frames the expression of ambiguity (etc.) as a task for lexical units in phrases, following the example of NL. There are advantages, however, to folding ambiguity (etc.) into embeddings that represent, not points, but regions in a semantic space. A shift from pointto region-oriented semantics allows systems of representations that can approximate mathematical lattices (Appendix A1) and lattice-based inference mechanisms like Prolog and constraint logic programming (Section A1.4). These mechanisms, in turn, provide semi-formal approaches to matching, unification, and generalization of representations, with applications outlined below and explored further in Section 8.4.3. \n A3.5 Facilitating Semantic Interpretation and Comparison Relative to NL, NL + frameworks offer potential advantages that include 1. Greater expressive capacity 2. More tractable interpretation 3. More tractable comparison. The preceding sections have discussed advantages of type (1) that stem largely from improvements at the lexical level. The present section will consider how condensation, localization, regularization of expression-level representations can provide advantages of types (  2 ) and  (3) . A central theme is the use of tractable, uniform, embedding-based representations to provide expressive capacity of kinds that, in NL, are embodied in less tractable-and often irregular-syntactic constructs. \n A3.5.1 Exploiting Condensed Expressions As noted above, embeddings can condense many noun-adjective, verb-adverb, and function word constructs, facilitating interpretation by making their semantic content available in forms not entangled with syntax. Further, embeddings can be compared through distance computations, 1 potentially after projection or transformation into task-and context-relevant semantic spaces. These operations are not directly available in NL representations. \n A3.5.2 Exploiting Content Summaries Content summaries (Section 8.3.4) can cache and amortize the work of interpreting expressions. The value of summarization increases as expressions become larger: An agent can read a book (here considered an \"expression\") to access the whole of its information, but will typically prefer a book accompanied by a summary of its topic, scope, depth, quality, and so on. Semantically optional summaries (perhaps of several kinds) can facilitate both associative memory across large corpora (Section 9.1.2) and shallow reading (\"skimming\") of retrieved content. Shallow reading, in turn, can enable quick rejection of low-relevance content together with fast, approximate comparison and reasoning that can guide further exploration. Where uses of information differ across a range of tasks, useful summaries of an expression may likewise differ. \n A3.5.3 Exploiting Context Summaries Although context summaries (Section 8.3.5), like content summaries, are in principle semantically redundant, they are substantially different in practice: Expressions are bounded, but an interpretive context may be of any size, for example, on the scale of a book or a body of domain knowledge. Thus, absent summarization, contextual information-and hence the meaning of an expression-may be far from local; with context summarization, meaning becomes more local and hence more strongly compositional.  2  1. Or intersection-and union-like operations in region-oriented semantics, see Appendix A1. 2. As noted elsewhere, current language models typically encode (costly to learn, difficult to share) summaries of global context-as well as knowledge of narrower contexts and even specific facts-while their inference-time activations include (costly to infer) summaries of textually local context. Learning and sharing task-oriented summaries of both broad and narrow contexts could provide complementary and more efficient functionality. Embeddings can provide the most compact summaries, but more general QNRs could provide richer yet still abstractive information. Some use-patterns would place specific semantic content in narrow context representations, and schematic semantic content in expressions that can contribute to descriptions in a wide range of contexts. In interpreting an expression, the effective, interpreted meanings of its embeddings would be strongly dependent on its current context. A programming language analogy would be the evaluation of expressions conditioned on binding environments, but in the QNR case, employing embeddings in place of conventional values and variables, 1 and employing neural models in place of symbolic interpreters. \n A3.5.4 Aligning Parallel and Overlapping Expressions Content summaries can facilitate comparison and knowledge integration in the absence of full structural alignment. In the limiting case of a complete structural mismatch between QNR expressions, their summary embeddings can still be compared. To the extent that high-level structures partially align, comparison can proceed based on matching to some limited depth. At points of structural divergence, comparison can fall back on summaries: Where subexpressions differ, their summaries (whether cached or constructed) can be compared; likewise, a subexpression summary in one expression can be compared to a lexical embedding in the other. Appendix A1 discusses how matches can be rejected or applied through soft unification. Upgrading the expressive power of lexical-level embeddings can facilitate structural alignment by shifting burdens of semantic expressiveness away from syntax: Condensing simple expressions into embeddings avoids a potential source of irregularity, the sequential order of lexical-level elements need not be used to encode emphasis or semantic priority, and the semantic differences between active and passive voice need not be encoded through differences in syntax and grammar. Accordingly, similar meanings become easier to express in parallel syntactic forms. If expressions with similar semantic content-representing similar things, properties, relationships, roles-are cast in a parallel syntactic form, they become easier to compare. Regularizing structure need not sacrifice expressive capacity: Expression-level nuances that in NL are expressed through alternative syntactic forms can quite generally be represented by embeddings that modify expression-level meaning. Thus, structural regularization, enabled by expressive embeddings and explicit graphs, can facilitate structural alignment and semantic comparison of related expressions. In addition, however, structural regularization can facilitate transformations among alternative canonical forms, potentially facilitating translation between representational dialects in heterogeneous NL + corpora. Regularization need not adhere to a uniform standard. \n A4 Compositional Lexical Units Embeddings with explicit compositional structure may offer advantages in efficient learning and generalization. Section 7.1.3 noted that the properties of vector addition can enable semantic compositionality without recourse to syntax; the present discussion examines the potential role of explicit forms of compositionality in learning and representation. Among the considerations are: • Efficiently representing large vocabularies • Parallels to natural language vocabularies • Parallels to NLP input encodings • Inductive bias toward efficient generalization The usual disclaimer applies: The aim here is neither to predict nor prescribe particular representations, but to explore what amounts to a lower bound on potential representational capabilities. Explicit vector compositionality, would, however, require explicit architectural support. \n A4.1 Motivation and Basic Approach Because few neural models write and read large stores of neurally encoded information, prospects for building large QNR corpora raise novel questions of storage and practicality. Section A5.5 outlines an approach (using dictionaries of composable vector components) that can be compact, efficient, and expressive. The present discussion considers how and why explicit compositionality within vector representations may be a natural choice for reasons other than efficiency. A key intuition is that sets of lexical components (like morphemes in natural languages) can be composed to represent distinct lexical units (like words and phrases that represent objects, actions, classes, relationships, functions, etc.  1  ), and that composite lexical units can best be regarded and implemented as single vectors in QNRs. For concreteness, the discussion here will assume that lexical-component embeddings are concatenated to form lexical-unit embeddings, 2 then melded by shallow feed-forward transformations to form unified representations. A key underlying assumption is that discrete vocabularies are useful, whether to encode embeddings compactly (Appendix A5), or to provide an inductive bias toward compositional representations. (Note that compact encodings can combine discrete vectors with continuous scalars to designate points on continuous manifolds; see Section A5.5.4). \n A4.2 Efficiently Representing Vast Vocabularies The on-board memories of GPUs and TPUs can readily store >10 7 embeddings for fast access.  3  This capacity is orders of magnitude beyond the number of English words, yet using these embeddings as components of lexical units can provide much more. If sets of potential lexical-unit embeddings are Cartesian products of sets of lexical-component embeddings, then potential vocabularies are enormous. Cartesian-product spaces in which (for example) 2 to 4 components are drawn from 10 7 options would offer 10 14 to 10 28 potential lexical-unit embeddings; of these, one can expect that a tiny fraction-yet an enormous number-would be potentially useful in describing the world. To represent expressions as strings (Appendix A5), 3 bytes of key information per lexical component would be ample. \n A4.3 Parallels to Natural Language Vocabularies Lexical units in NL vocabularies are commonly built of multiple morphemes, including roots, affixes,  4  and words embedded in compounds or multiword units.  5  1. As already noted, this use of \"lexical units\" abuses standard terminology in linguistics. 2. Concatenation can be modeled as addition of blockwise-sparse vectors, and addition of dense vectors would arguably be superior. However, using addition in place of concatenation would (in application) increase storage costs by a small factor, and would (at present) incur a substantial explanatory cost. 3. See Section A5.5.5. 4. English builds on >1300 roots and affixes (prefixsuffix.com 2008). \n 5. Here used in the standard linguistic sense (also termed \"lexical items\"). If we view NL as a model for potential NL + constructs, then it is natural to consider analogues of morphemes in embedding spaces, and to seek lexical-level semantic compositionality through explicit composition of building blocks in which the meanings of components are, as in NL, a joint result of their combination. This approach can make lexical components themselves targets of learning and thereby expand the scope of useful, accessible vocabulary. Medical terminology illustrates the role of lexical-level compositionality in building a language adapted to a rich domain.  1  Most medical terms are built of sequences of parts (\"cardio+vascular\") or words (\"primary visual cortex\").  Wikipedia (2021)  lists 510 word parts (prefixes, roots, and suffixes) used in medical terminology, while a large medical dictionary defines ~125,000 distinct, often multi-word terms  (Dorland 2007 ), a number that approaches an estimate (~200,000) of the number of words in the English language as a whole.  2  Refining and expanding the store of applicable lexical components from hundreds or thousands to millions or more would greatly increase the potential semantic resolution of medical language at a lexical level. Medicine, of course, occupies only a corner of a semantic universe that embraces many fields and extends far beyond what our words can readily describe. \n A4.4 Parallels to NLP Input Encodings There are substantial parallels and contrasts between input encodings in current NLP and compositional embeddings in potential QNR processing systems Table  A4 .1): • In the proposed mode of QNR processing, inputs are lexical components concatenated to form lexical units; in Transformer-based NL processing, inputs are words and subwords extracted from strings through tokenization. • In the QNR case, a very large vocabulary of lexical components is composed to form a vast Cartesian-product space of lexical units; in the NLP case, a smaller vocabulary of lexical units is built from word fragments and common words (in BERT, ~30,000 \"wordpieces\"). • In the QNR case, the composition of lexical units is determined by representation learning; In the NLP case, the decomposition of strings is determined by a tokenization algorithm. • NLP models dynamically infer (and attempt to disambiguate) lexicallevel representations from tokenized text; in QNR processing, input embeddings are explicit lexical-level products of previous representation learning. Thus, lexical-level QNR inputs are roughly comparable to hidden-layer representations in an NLP model.  \n A4.5 Inductive Bias Toward Efficient Generalization The ability to represent specific lexical units as compositions of more general semantic components could potentially support both systematic generalization and efficient learning, including improved sample efficiency. An important lexical component will typically occur far more frequently than the lexical units that contain it, and learning about a component can provide knowledge regarding lexical units that have not yet been encountered.  2  Indeed, without the inductive bias provided by composition, it might be difficult to learn truly large vocabularies that form well-structured semantic spaces. As an NL illustration of this principle, consider \"primary visual cortex\" again: A reader who knows little or nothing of neuroanatomy will know the general meanings of \"primary\" and \"visual\" and \"cortex\", having encountered these terms in diverse contexts. With this knowledge, one can understand that \"primary visual cortex\" is likely to mean something like \"the part of the brain that first processes visual information\", even if this term has never before been seen. A more refined understanding can build on this. This familiar principle carries over to the world of potential QNR representations, where exploitation of compositional lexical-level semantics promises to support learning with broad scope and effective generalization. \n A4.6 A Note on Discretized Embeddings In typical applications, reading is far more frequent than writing, hence mapping continuous internal representations to discrete external representations need not be computationally efficient. This output task can be viewed as either translation or vector quantification. Lexical components that are distant from existing embeddings may represent discoveries worth recording in an expanded vocabulary. Because components populate continuous vector spaces, discretization is compatible with differentiable representation learning of components and their semantics. \n A5 Compact QNR Encodings String representations of QNRs, in conjunction with discretized vector spaces and graph-construction operators, can provide compact and efficient QNR encodings. \"Premature optimization is the root of all evil.\" -Donald Knuth 1 NL + expressions can be implemented compactly by combining operator-based representations of graph structures with extensible dictionaries of discretized embeddings; the latter provide mechanisms for what can be regarded as lossy compression, but can also be regarded as providing a useful inductive bias (Section A4.5). The content of QNR corpora can be represented as byte  1. Knuth (1974) . Because compression is (at most) a downstream research priority, Knuth's warning against premature optimization is relevant and suggests that the value of this appendix is questionable. There is, however, good reason to explore the scope for efficient, scalable implementations: A sketch of future options can help to free exploratory research from premature efficiency concerns-or worse, a reluctance to consider applications at scale. strings approximately as compact as NL text by exploiting key-value stores of embeddings and graph-construction operators.  1  The memory footprint these stores need not strain the low-latency memory resources of current machines. The purpose of this appendix is not to argue for a particular approach, but to show that a potential challenge-the scale of QNR storage footprints-can be met in at least one practical way. Note that the considerations here have nothing to do with neural computation per se, but are instead in the domains of algorithm and data-structure design (often drawing on programming language implementation concepts, e.g., environments and variable binding). From a neural computation perspective, the mechanisms must by design be transparent, which is to say, invisible. \n A5.1 Levels of Representational Structure Prospective QNR repositories include representational elements at three levels of scale: • Embeddings: vector attributes at a level comparable to words • Expressions: graphs at a level comparable to sentences and paragraphs • References: graph links at the level of citations and document structures In brief, expressions are graphs that bear vector attributes and can include reference-links to other expression-level graphs. There is no important semantic distinction between expression-level and larger-scale graph structures; the key considerations involve interactions between scale, anticipated patterns of use, and implementation efficiency. A formal notation would distinguish between QNRs as mathematical objects (graph and attributes), QNRs as computational objects (inference-time data structures that represent graphs and attributes), and encodings that represent and evaluate to QNR objects in a computational environment. The following discussion relies on context to clarify meaning. \n A5.2 Explicit Graph Objects vs. String Encodings A relatively simple computational implementation of QNRs would represent embeddings as unshared numerical vectors,  1  and graphs as explicit data structures.  2  Repositories and active computations would share this direct, bulky representational scheme. Natural language expressions are more compact: NL words in text strings are far smaller than high-dimensional vector embeddings, and graph structures are implicit in NL syntax, which requires no pointer-like links. Inference-time representations of NL + expressions may well be bulky, but so are the inference-time representations of NL expressions in neural NLP. And as with NL, stored QNR expressions can be represented compactly as byte strings. \n A5.3 Compact Expression Strings A QNR corpus can be represented as a key-value store that contains embeddings (numerical vectors), operators (executable code), and encoded QNR expressions (strings that are parsed into keys and internal references). In this scheme, expressions encoded as byte-strings evaluate to expression-level QNR graphs that can include references that define graphs at the scale of documents and corpora. In more detail: • Keys designate operators, embeddings, or expression-strings in a keyvalue store. • Expression-strings are byte strings 3 that are parsed into keys and indices. • Indices designate subexpressions in a string. • Operators are graph-valued functions  4  of fixed arity that act on sequences of subexpressions (operands). • Subexpressions are keys, indices or operator-operand sequences. • Embeddings are graph attributes. 1. \"Unshared\" in the sense that each attribute-slot would designate a distinct, potentially unique vector object. 2. \"Explicit\" in the sense that each arc would be represented by a pointer-like reference. 3. Potentially bit strings. 4. An extended scheme (Section A5.5.4) allows vector-valued operators with vector and scalar operands. • Reference is shorthand for \"expression-string key\" (typically a stopping point in lazy evaluation). \n A5.3.1 From Strings to Syntax Graphs Parsing an expression-string is straightforward: Operators are functions of fixed arity, the initial bytes of an expression-string designate an operator, and adherence to a standard prefix notation enables parsing of the rest. A first level of decoding yields a graph in which operator nodes link to child nodes that correspond to embeddings, operators, and references to other expressions.  1  Thus, expression-strings represent graphs in which most intra-expression links are implicit in the composition of graph-valued operators and their products, 2 while the overhead of designating explicit, inter-expression links (several bytes per reference-key) is in effect amortized over the linked expressions. Accordingly, QNR graphs on a scale comparable to NL syntactic structures need not incur per-link, pointer-like overhead. \n A5.3.2 Lazy Evaluation Lazy evaluation enables the piecemeal decoding and expansion of QNR graphs that are too large to fully decompress. To support lazy evaluation, references to expression-strings can evaluate to graph objects (presented as a vector of nodes), or can be left unevaluated. A key-value store then can support lazy evaluation by returning either an expression-string or, when available, the corresponding graph object; a server that retrieves objects for processing can query an expression-store with a key and invoke evaluation if the store returns a string. This mechanism also supports the construction of shared and cyclic graphs. \n A5.4 Graph-Construction Operators Parsing associates each construction operator with a sequence of operands that can be evaluated to produce (or provide a decoded access to) embeddings, references, and root nodes of graph-objects. Simple construction operators can treat their arguments as atomic and opaque; more powerful operators can access the contents of evaluated graph-valued operands or the contents of the operator's evaluation-time context. QNRs encoded as expression-strings requires a similar mapping of tokens (keys) to lexical-component embeddings, but prospective vocabularies are orders of magnitude larger. The use of large vocabularies mandates the use of scalable key-value stores; with this choice, vocabulary size affects neither neural model size nor computational cost. The use of low-latency storage for frequently used embedding vectors could, however, impose substantial burdens, with requirements scaling as vocabulary size × vector dimensionality × numerical precision.  1  Baseline values for these parameters can place storage concerns in a quantitative context. \n A5.5.2 Vocabulary Size In a continuous vector space, \"vocabulary size\" becomes meaningful only through vector quantization, the use of identical vectors in multiple contexts.  2  Discrete vectors can be compared to NL words or word parts, and vectors produced by concatenating quantized vectors can be compared to words built of combinations of parts. \n A5.5.3 Lexical Component Embeddings Potential QNR representations and applications (Section 11) are sufficiently diverse (some far from \"linguistic\" as ordinarily understood) that it is difficult to generalize about the appropriate granularity of vector quantization or the extent to which it should be employed. Different applications will call for different trade-offs between compression and performance, and quantization has been found to improve rather than degrade neural representation learning in some domains.  3  More can be said about vocabulary size, however, in the context of NL + representations that constitute (merely) strong generalizations of NL. To provide a round-number baseline value, consider an NL + representation that employs a vocabulary of distinct lexical components 4 that is 50 times larger than the vocabulary of distinct words in the English language:  5  For present purposes, ~200,000 words is a reasonable estimate of the latter, 1 implying a baseline NL + vocabulary of 10 million lexical-component embedding vectors. \n A5.5.4 Composing Components As discussed in Appendix A4, a richer vocabulary of lexical units can be constructed by composition, for example, by concatenating lexical-component embeddings. This generative mechanism parallels what we see in natural languages, where many (even most) words are composed of various wordparts, TAM-C affixes, or other words. Discrete composition: Expression-strings can readily describe composite lexical units: Vector-valued operators with vector arguments can denote concatenations of any number of components. Considering only discrete combinations, operators that accept 2 to 4 arguments from a vocabulary of 10 7 embeddings can define product vocabularies of 10 14 to 10 28 composites. These are candidates for use as lexical units, and even tiny useful fractions correspond to vast vocabularies. Appendix A4 explores composite embedding representations from perspectives that include efficiency in learning and semantics in use. Weighted combination: Some entities are drawn from distributions characterized by continuous properties that include size, color, age, and probability. Discrete vocabularies are a poor fit to continuous semantics, but operators that accept numerical arguments can specify weighted combinations of vectors, and hence continuous manifolds in semantic spaces. For example, a key that designates a discrete vector embedding a could be replaced by keys designating a vector-valued operator f i , a scalar parameter w, and embeddings c and d, for example, a linear combination: f i (w, c, d) = wc + (1 − w)d. A suitable range of operators can generalize this scheme to nonlinear functions and higher-dimensional manifolds. \n A5.5.5 Dimensionality, Numerical Precision, and Storage Requirements Transformer-based models typically map sequences of tokens to sequences of high-dimensional embeddings (in the baseline version of BERT, 768 dimen-sions). Though high dimensionality is perhaps necessary for hidden states that are used to both represent and support processing of rich, non-local semantic information, one might expect that word-level semantic units could be represented by embeddings of lower dimensionality. This is empirically true: Shrinking input (word-level) embeddings by more than an order of magnitude (from 768 to 64 dimensions) results in a negligible loss of performance  (Lan et al. 2020) . In estimating potential storage requirements, 100 (but not 10 or 1000) dimensions seems like a reasonable baseline value for representations of broadly word-like lexical components. Forming lexical units by concatenation of components (desirable on several grounds; see Appendix A4) would yield larger input embeddings without increasing storage requirements. BERT models are typically trained with 32-bit precision, but for use at inference time, parameters throughout the model can be reduced to 8-bit  (Prato, Charlaix, and Rezagholizadeh 2020), ternary (Wei Zhang et al. 2020) , and even binary (H.  precision with little loss of performance. As a baseline, allowing 8-bit precision for embeddings at the input interface of a QNR-based inference system seems generous. When combined with the baseline vocabulary size suggested above (10 7 component embeddings), these numbers suggest a baseline scale for an NL + key-value store: Storage = vocabulary size × dimensionality × precision ≈ 10, 000, 000 elements × 100 dimensions × 1 bytes ≈ 1 GB. For comparison, current GPUs and TPUs typically provide on-board memory >10 GB, and are integrated into systems that provide terabytes of RAM. Given the expected power-law (Zif-like) distribution of use frequencies, implementations in which small local stores of embeddings are backed by large remote stores would presumably experience overwhelmingly local memory traffic.  1  In many tasks (e.g., fetching content used to answer human queries), occasional millisecond-range latencies would presumably be acceptable. 1. As an illustrative NL example, in its first 10 7 words, the Google Books corpus contains about 10 4 distinct words (by a generous definition that includes misspellings), a ratio of one new word in 10 3 , while in its first 10 9 words, it contains about 10 5 distinct words, a ratio of one in 10 4 .  See Brysbaert et al. (2016) . \n A5.5.6 A Note on Non-Lexical Embeddings Prospective, fully elaborated NL + systems would employ more than just lexical-level embeddings; for example, embeddings that in some sense summarize expressions can enable shallow processing (skimming), or can serve as keys for near-neighbor retrieval that implements semantic associative memory over large corpora.  1  The costs of summaries can in some sense be amortized over the expressions they summarize. Like natural language, NL + expressions can be represented by compact byte strings. Stores of discrete embeddings can support large (and through composition, vast) vocabularies of differentiable semantic representations within an acceptable footprint for low-latency memory; in other words, NL + corpora can be represented approximately as efficiently and compactly as corpora of NL text. Neither necessity nor optimality is claimed for the approach outlined here. Figure 1 . 1 : 11 Figure 1.1: Information flows in generic QNR systems supported by access to a repository of QNR content. Inputs and outputs may be multimodal. \n Figure 7 . 1 : 71 Figure 7.1: Kissing spheres, d = 2, k = 6 \n 1 . 1 Heimann et al. (2018) , Cao et al. (2019) , Qu, Tang, and Bengio (2019), and Fey et al. (2020)  2. Soft unification and anti-unification operations can contribute to this functionality (Appendix A1). \n . \n 3 . 3 For related work, see Verga et al. (2020), Lewis et al. (2020) , Guu et al. (2020), and Févry et al. (2020) . \n 4.3)  of the corresponding pairs of expressions. \n 1 . 1 Domain-based assessments of credibility have been used in constructing knowledge graphs from social-media sources: Abu-Salih et al. (2021). \n • Employing NL → QNR interfaces to ensure QNR representations • Employing QNR intermediate representations in NL → NL training tasks • Decoupling QNR representations from NL encodings • Employing semantically rich training tasks with QNR objectives • Structuring QNR semantics through auxiliary, lattice-oriented training • Applying QNR-domain inference to exploit QNR repositories \n Figure 10 . 1 : 101 Figure 10.1: Information flows in minimalistic, QNR-bottleneck systems. Inputs and outputs may be multimodal. \n Figure10.2: Information flows in generic QNR systems augmented by access to a repository of QNR content. In the general case, \"QNR inference\" includes read/write access to repositories, producing models that are in part pre-trained, but also pre-informed. \n 2 2 \n 10. 3 . 3 33 Extending QNR Representations Beyond Linguistic Domains A further class of tasks, X → QNR→ NL, would map non-linguistic inputs X to NL, again mediated by (and training) QNR-based mechanisms. Potential examples include: • Predicting descriptions of images • Predicting descriptions of human actions • Predicting comments in code A quite general class of tasks would encode information from a domain, decode to a potentially different domain, and train QNR → QNR components to perform intermediate reasoning steps. Potential examples include the control of agent behavior involving instruction, communication, and planning. 3 10.4 Abstracting QNR Representations from NL \n Figure 10 . 4 : 104 Figure 10.4: Outline of multitask architectures that include access to external and QNR-based information repositories (e.g., the internet). More arrows could be added. \n Figure 10 . 5 : 105 Figure 10.5: Block diagram decomposing aspects of architectures for complex, open-ended QNR inference functionality. Both working memory and an external repository store key-value pairs, and given a query, will return one or more values associated with near-neighbor keys in a semantic embedding space. Arrows labeled q and k (shown explicitly in connection with external operations) represent query and key embeddings used in storing and retrieving QNR values (v). An elaboration of Figure 10.4 would show similar QNR inference functionality in connection, not only with \"QNR inference systems\", but also with \"reader-encoders\" (which need not be distinct components). \n Figure A1. 1 : 1 Figure A1.1: A Boolean lattice over sets and subsets. \n Figure A1. 2 : 2 Figure A1.2: Exact meets and joins of interval (box) regions. \n Figure A1. 3 : 3 Figure A1.3: Approximate meets and joins of regions from a lessconstrained family of region shapes. \n In mathematical applications, proposed QNR/NL + frameworks could wrap formal, symbolic structures in soft descriptions3  that can be applied to help recognize analogies and express purpose, and these capabilities can operate at multiple levels of granularity. Pólya observes that discovery in mathematics involves generate-and-test cycles guided by soft considerations, and modern deep learning confirms the value of soft matching in guiding theorem proving. Better neural representations can improve ML-informed premise selection,1   slowing the explosive growth of deep proof trees by improving the success rate of generate-and-test cycles. Graph neural networks that operate on syntactic structures can provide useful embeddings (M., and enriching formal symbolic representations with soft semantic descriptions (e.g., of known use-contexts) should enable further gains. Pólya emphasizes the importance of analogy, a kind of soft, structured generalization (Section A1.1.2). The formal (hence more restrictive) lattice operation of generalization by antiunification has been applied to analogical reasoning in symbolic mathematics (Guhe et al. 2010) ; embedding symbolic structures in soft representations could extend the scope of potential generalizations. • Applying unification to combine and extend partial descriptions • Applying unification to identify clashes between descriptions • Applying analogies from developed fields to identify gaps in new fields • Applying analogies to suggest hypotheses that fill those gaps • Matching experimental objectives to experimental methods • Matching questions and data to statistical methods • Assessing evidence with attention to consensus • Assessing evidence with attention to consilience • Enabling ongoing updates of inferential dependency structures Applications like these need not automate scientific judgment: To provide value, they need only provide useful suggestions to human scientists. Devel- opments along these lines would extend current directions in applying ML to scientific literatures. 1 11.3.3 Mathematics You have to guess a mathematical theorem before you prove it; you have to guess the idea of the proof before you carry through the details. You have to combine observations and follow analogies; you have to try and try again. -George Pólya 2 \n Table A2 . A2 1: Classes and Examples of Function words. The examples below cover about one third of the function-word vocabulary of the English language.In strongly inflected languages, the roles of some of these function words are performed by morphological distinctions. Table A2.2: Examples of Tense/Aspect Distinctions. Languages can use inflection or function words to express distinctions that describe (e.g.) relative time, duration, or causation. Determiners: the, a, this, my, more, either Prepositions: at, in, on, of, without, between Qualifiers: somewhat, maybe, enough, almost Modal verbs: might, could, would, should Auxiliary verbs: be, do, got, have Particles: up, down, no, not, as Pronouns: she, he, they, it, one, anyone Question words: who, what, where, why, how Conjunctions: -coordinating: for, so, and, nor, but, or, yet -subordinating: if, then, thus, because, however -temporal: before, after, next, until, when, finally -correlative: both/and, either/or, not/but \n Table A2 . A2 3: Examples of Modality Distinctions. Languages can express modalities by inflection or function words. The existence of graded degrees and overlaps within and between modalities suggests the potential value of continuous vector-space representations. Table A2.4: Examples of Case Distinctions. Case distinctions can express the roles of words in a sentence (in a familiar grammatical sense) or the roles of what they denote in a situation. The number of inflectional case distinctions A3 From NL Constructs to NL + varies widely among languages; English has three, Tsez has dozens, many Condensing, regularizing, and extending the scope of semantic represen- of which are locative. As with modalities, blurred boundaries and overlaps tations can improve expressive capacity and compositionality, and can between cases suggest the potential value of continuous vector-space repre-support theoretically grounded methods for comparing and combining sentations. semantic information. Nominative -subject of a verb Accusative -object of a verb Dative -indirect object of a verb Genitive -relationship of possession Comitative -relationship of accompaniment Lative -movement to something Ablative -movement away from something Orientative -orientation toward something Locative -location, orientation, direction Translative -becoming something Instrumental -means used for an action Causal -cause or reason for something Benefactive -beneficiary of something Terminative -limit or goal of an action Interrogative -Question Imperative -Command Indicative -Unqualified statement of fact Inferential -Qualified (inferred) fact Subjunctive -Tentative or potential fact Potential -Possible condition Conditional -Possible but dependent on another condition Hypothetical -Possible but counterfactual condition Optative -Desired condition Deontic -Ideal or proper condition \n Table A4 . 1 : A41 Input representations used in current NLP and prospective QNR processing. Typical NLP models Proposed QNR models Input units: wordpiece tokens component embeddings Vocabulary size: ~10 4 -10 5 ~10 7 -10 28 Embedding origins: learned representations learned representations Initial processing: multiple attention layers MoE blending layer 1 \n\t\t\t . This document often describes illustrative forms of representation and functionality, or describes how neural computation could potentially implement those forms and functions, but always with the implicit proviso that learned neural representations and mechanisms are apt to be surprising. \n\t\t\t . Note that representations of theoretically equivalent expressive capacity need not be equivalent in, for example, computational tractability, compositionality, compactness, scalability, or inductive bias. \n\t\t\t . For example, opaque Transformer-like models may be useful in QNR applications: In general, quasicognitive processing is complementary to quasilinguistic representation.2. Logic, mathematics, programming languages, knowledge representation languages, attempted formalizations of natural language, etc.3. Including systems that exploit pretrained language models. 4. Readers who don't skim will encounter redundancy provided for those who do. \n\t\t\t . E.g., 128-token text pieces (Verga et al. 2020) , or more general multimodal information (Fan et al. 2021) .7. Khandelwal et al. (2020) and Yogatama, Masson d'Autume, and Kong (2021) . Models of this kind fall within the broad class of memory-augmented neural networks (Santoro et al. 2016) . \n\t\t\t . Perceptions of success were (notoriously) blunted by reclassification of research results (\"If it works, it isn't AI\"). Highly successful automation of symbolic mathematics, for example, emerged from what had initially been considered AI research (Martin and Fateman 1971; Moses 2012 ). \n\t\t\t . E.g., see Rocktäschel and Riedel (2017) ,Minervini et al. (2020), Arabshahi et al. (2021) . \n\t\t\t . This does not argue that internal cognitive representations themselves have a clean, graph-structured syntax, nor that sparse, syntax-like graphs are optimal representations for externalized QNRs. The argument addresses only properties of QNRs relative to word strings. \n\t\t\t . E.g., due to working-memory constraints (Caplan and Waters 1999) . \n\t\t\t . Section 9.1.2 considers embedding-indexed QNR stores as associative memories accessed through near-neighbor lookup in semantic spaces.2. Goyal and Bengio propose language-inspired compositionality as a key to developing systems that more closely model human-like intelligence; see also (Y. Jiang et al. 2019) , which makes strong claims along similar lines.3. This is a softer criterion than that of formal compositionality.4. In a sense that may be neither spatial nor temporal. \n\t\t\t . Scoped binding of symbols stretches but does not break notions of syntactic locality-if locality is construed in terms of arcs in identifiable graphs, it is not constrained by syntactic distances over sequences or trees.2. Interestingly, although computer languages are models of compositionality, fMRI studies show that reasoning about code is only weakly focused on brain regions specialized for natural language (Ivanova et al. 2020) . This distribution of neural function supports a broad role for compositional representations in non-linguistic cognition.3. Raposo et al. (2017) , Watters et al. (2017) , Battaglia et al. (2018) , Eslami et al. (2018) , G. R. Yang et al. (2019), and Bear et al. (2020)  4. In formal systems, definitions and bindings within a scope are examples of a precise form of contextually conditioned compositionality. \n\t\t\t . The lack of full compositionally in language has been a bitter pill for formalists, and not all have swallowed it. The Principle of Compositionality, that the meaning of a complex expression is determined by its structure and the meanings of its constituents, has been taken to apply to language; although others recognize a pervasive role for context (enriching word embeddings with contextual information has been a successful strategy in neural NLP; see Liu, Kusner, and Blunsom (2020) ), some seek to apply context to determine (fully compositional) lexical-level meanings, which seems an arbitrary and perhaps unworkable choice.See Szabó  (2017). \n\t\t\t . \"Castle Mound gives a good view of its surroundings.\" Is the context of this sentence tourism in present-day Oxford, or military intelligence during the Norman conquest? The nature of the view and its significance may be clear in the context of a pamphlet or book, but cannot be found in the quoted string of words.2. The same can be said of natural language expressions in the context of human memory and neural states.3. Note that the problem is not ambiguity per se, but ambiguity that is unintentional or costly to avoid. Intentional ambiguity is expressive, and (to meet benchmark criteria) must therefore be expressible in NL + (see Section 8.3.1 and Section A3.4).4. Because expressions may appear in multiple contexts, this should be seen as information about the context of a particular citation or use of an expression.5. As discussed in Section 8.3.5.    \n\t\t\t . Note this question does not ask how those constructs actually operate, which is a subject of ongoing controversy among linguists.2. See Bobrow and Winograd (1977)  and McShane and Nirenburg (2012) . Knowledge representation languages typically attempt to build on unambiguous ontologies (Guarino 2009 ), yet the ability to express ambiguity is an important feature of natural languages. \n\t\t\t . Or DAGs, when coreference is represented explicitly.2. See Borsley and Börjars (2011) .3. Meanings, not words: Vector representations of words perform poorly when those words have multiple meanings, but representing meanings rather than words sidesteps this problem. \n\t\t\t . Linguists define \"simple phrases\" differently.3. Note that points in a vector space are inherently sharp, yet may be taken to represent (potentially soft) regions in a lower dimensional space (see Section 7.1.5). \n\t\t\t . The vocabulary of an English speaker may include 10,000 to 100,000 or more content words; different speakers may employ different blocks of specialized (e.g., professional) vocabulary.2. Note also the potential value of explicitly compositional representations of embeddings, e.g., embeddings built by concatenation (Appendix A4).3. Distinctions of kind and differences in properties differ, yet are not entirely distinct.4. The somewhat counter-intuitive geometric properties of high dimensional spaces are relevant here (see Section 7.1.4). Note also that displacements in a particular direction need not have the same meaning in different regions of semantic space: Most properties relevant to planets differ from those relevant to music or software. \n\t\t\t . The poverty of closed-class vocabulary is mitigated by the availability of compound function words (at least, because of. . . ) that can be treated as lexical entities.See Kato, Shindo,  and Matsumoto (2016).2. The constructs \"X and/or Y\" and \"either X or Y\" can express the inclusive/exclusive distinction, yet trade-offs between precision and word economy (and the cognitive overhead of instead relying on context for disambiguation) ensure frequent ambiguity and confusion in practice. As a less trivial example, the ability to compactly express \"possibly-inclusive-butprobably-exclusive-or\" would be useful, and in a continuous vector space of function words, would also be natural. \n\t\t\t . Because tense, aspect, and modality overlap and intertwine, linguists often group them under the label \"TAM\"; because they also overlap with case distinctions, all their indicators (inflections, markers) will be lumped together here and simply referred to as \"TAM-C modifiers\" (for examples and discussion, see Appendix A2 and Section A3.3).2. Further muddling standard linguistic distinctions, punctuation can indicate case or modality (question marks, exclamation marks) and can play roles like function words that clarify syntax (commas, semicolons, periods) or express relationships of explanation, example, or reference (colons, parentheses, quotation marks). In spoken language, verbal emphasis and paralinguistic signals play similar roles.3. Linguists recognize a semantic space of modalities (Allan 2013) . \n\t\t\t . Potentially accompanied by abstractive, non-definitional embeddings(Section 8.3.4). The contours of such boundaries need not be determined by an implementation: There is no need to engineer representations that neural systems can instead learn. \n\t\t\t . Philosophers have dissected considerations of this sort to provide richer distinctions and more precise terminology. \n\t\t\t . Unification of two expressions produces the least specific (most general) expression that contains the information of both; unification fails if expressions contain conflicting information. Generalization (anti-unification) of two expressions produces the most specific (least general) expression that is compatible with (and hence unifies) both. Unification and anti-unification are associative and commutative, and satisfy several other axioms. Soft unification and antiunification relax these constraints.(See Section A1.4.3.)    \n\t\t\t . For applications of language in RL, see Das et al. (2017) , Lazaridou, Peysakhovich, and Baroni (2017) , Shah et al. (2018), and Luketina et al. (2019) . \n\t\t\t . Note, however, successful applications of discretized representations in which learned, finite sets of vectors are selected from continuous spaces; see, for example, Oord, Vinyals, and Kavukcuoglu (2018)  and Razavi, Oord, and Vinyals (2019)   \n\t\t\t . Note that the number of vectors that are all nearly orthogonal (all pairwise cosine similarities 0.5) also grows exponentially with d and becomes enormous in high dimensional spaces. \n\t\t\t . The use of Euclidean distance or cosine similarity is explicitly or tacitly assumed in much of this document, but growing interest suggests that hyperbolic spaces (useful in graph and sentence embeddings) or mixed geometries may provide attractive alternatives for embedding QNR expressions; see for example Peng et al. (2021) .2. Appendix A1, Section A1.5.3, and Section A1.6 discuss both boxes and more flexible classes of representations in the context of unification and generalization operations on soft lattices.3. Arithmetic in which operands are intervals over R. \n\t\t\t . Van Lierde and Chow (2019)  and Menezes and Roth (2021)   \n\t\t\t . To enable arbitrary graphs as outputs requires initializing with complete graphs (hence N (N − 1) directed arcs). Restricting optimization to local subgraphs, to restricted search spaces, or to algorithmically generated \"rough drafts\" can avoid this difficulty.2. Note, however, that efficient, highly scalable near-neighbor retrieval algorithms on unstructured spaces are typically \"approximate\" in the sense that nearest neighbors may with some probability be omitted. This shortcoming may or may not be important in a given application, and fast algorithms on weakly structured spaces can be exact (see Lample et al. 2019) . \n\t\t\t . In logic, symbols correspond to zero-volume regions, while variables correspond to unbounded regions. 4. See Section A1.2. In generalizations of lattice representations, a \"region\" may be fuzzy, in effect defining a pattern of soft attention over the space, and lattice relationships may be approximate (Section A1.4.3). \n\t\t\t . Rocktäschel and Riedel (2017) , Campero et al. (2018) , Minervini et al. (2018) , Weber et al. (2019), and Minervini et al. (2020) 2. Narayanan et al. (2017), Grohe (2020), and Pan et al. (2020) 3. Allamanis et al. (2017), R. Liu et al. (2017), H. Zhang et al. (2018), and Alon et al. (2019) 4. Y. Li et al. (2018) and Cao and Kipf (2018) 5. Pan et al. (2018), Simonovsky and Komodakis (2018), and Salehi and Davulcu (2020) 6. What counts as \"small\" is an open question. \n\t\t\t . Attributes (labels) can potentially be expanded to include sets of vectors of explicitly or implicitly differing types.2. And perhaps also implementation.3. A representation that can also accommodate multigraphs and hypergraphs. \n\t\t\t . Figure from Raboh et al. (2020), used with permission. \n\t\t\t . When a single word (e.g., \"match\") has multiple unrelated meanings (homonymy), it corresponds to multiple, potentially widely scattered points in a natural semantic space; when a word (e.g., \"love\") has a range of related meanings (polysemy) representations that map word-meanings to points (rather than semantic regions) become problematic.See Vicente  (2018).2. When different regions of a vector space represent things of distinct kinds, the meanings of directions may vary with position: In other words, because different kinds of things have different kinds of properties, it is natural for mappings of directions to properties to be a function of kinds, and hence of location. Indeed, the literature describes discrete-word models of NL semantics in which the meanings of adjectives are a function of their associated nouns; see, for example, Baroni and Zamparelli (2010)  and Blacoe and Lapata (2012) . In this general approach, the meanings of verbs are also a function of their associated nouns, and the meanings of adverbs are functions of their associated verbs. \n\t\t\t . Where \"entity\" is intended to include (at least) different things (horses, cows, photons, telescopes, gravitation, integers) and different actions (walk, run, observe, report).2. \"[A] graph of data intended to accumulate and convey knowledge of the real world, whose nodes represent entities of interest and whose edges represent relations between these entities\" (Hogan et al. 2021 ).   3. Ji et al. (2021 and Ali et al. (2021)  4. Expression-level definitional embeddings are not sharply distinguished from lexical embeddings, e.g., those corresponding to function words. \n\t\t\t . In processing with access to QNR corpora, the analogues of long-term memory and text converge.2. For an analogous application of summarization in Transformers, see Rae et al. (2019) .3. E.g., representations of substantive content (potential answers to questions) vs. representations of the kinds of questions that the substantive content can answer (potentially useful as keys).4. Summaries with this property seem well suited for use as keys in search guided by semantic similarity. \n\t\t\t . The considerations here are quite general; for example, context representations are also important in understanding/recognizing objects in scenes (Tripathi et al. 2019; Carion et al. 2020 ).2. StandardTransformers represent a span of immediate context in their activations, while representing both relevant and irrelevant global context in their parameters. Both of these representations are subject to problems of scaling, cost, compositionality, and interpretability. Models (potentially Transformer-based) that write and read QNRs could benefit from externalized representations of wide-ranging, semantically indexed contextual knowledge. The application of very general knowledge, however, seems continuous with interpretive skills that are best embodied in model parameters. \n\t\t\t . In this connection, consider potential generalizations of the image-domain algorithm described inSarlin et al. (2020), which employs end-to-end training to recognize, represent, and align features in pairs of data objects through GNN processing followed by differentiable graph matching. Analogous processing of QNR representations would enable semantics-based alignment even in the absence of strict structural matches.2. In recent work on neural image interpretation, employed propositional knowledge in the form of production rules (condition-action pairs) that are represented by model parameters and selected at inference time by attention mechanisms based on similarity between condition and activation vectors.3. Distance metrics may result from neural computations that condition on broader semantic content, or on other relationships beyond simple pairwise sums and differences. \n\t\t\t . The lattice axioms (Section A1.2) ensure that pairwise combination extends to multiple arguments in a natural way. \n\t\t\t . Summing over products of training epochs and corpus sizes (Brown et al. 2020) . \n\t\t\t . Note that this mechanism provides a differentiable and potentially fluid graph representation; see Section 7.2.3. \n\t\t\t . Scaling as ~O(n), but learned, structured key spaces can improve scaling to O(n 1/2 ) (Lample et al. 2019). 2. J. Wang et al. (2018), Fu et al. (2018), Johnson, Douze, and Jégou (2019), Jayaram Subramanya et al. (2019), and Sun (2020)3. Note that representation of images through relationship-graphs among objects blurs the boundary between opaque image embeddings and syntactic structures; See for example Bear et al. (2020) . Image-embedding spaces can also be aligned with text (e.g., inPatashnik  et al. 2021). \n\t\t\t . For example, an NL phrase can compactly say that \"face_1 strongly resembles face_2\", while a lexical-level embedding can compactly say that: \"face_1, with a specific set of embedding-space offsets in shape, color, and expression, looks like face_2 with some quantified but approximate residual differences.\" 2. Hence the value of using interpretable yet non-linguistic image embeddings as examples.3. In this connection, note that QNR expressions are data, that neural functions can be of type QNR → QNR, and that \"apply\", \"eval\" and \"quote\" functions can play quite general roles (e.g., in languages like Scheme that can express and operationalize high-level abstractions). \n\t\t\t . Szegedy (2020)  suggests an NL-translation approach to formalizing mathematics papers.2. M. and Minervini et al. (2018)   \n\t\t\t . Y., Tran et al. (2020), and Botha, Shan, and Gillick (2020)  2. Object-oriented programming exploits this principle.3. Even projections of sets of high-dimensional vectors into spaces of substantially lower dimensionality can preserve key geometric relationships quite well: The Johnson-Lindenstrauss lemma implies good preservation of both distances and cosine similarities between vectors. \n\t\t\t . Particularly when intermediate NL + processing can draw on relevant NL + corpora. Successful augmentation of Transformers with external memory (e.g., for question answering) provides evidence for the potential power of this approach (Koncel-Kedziorski et al. 2019; Fan et al. 2021; Min et al. 2020; Thorne et al. 2020) . \n\t\t\t . A simple example would be regions implied by implicit, contextual uncertainty in a vector value; richer, more formal examples include spaces in which vector values (points) explicitly correspond to regions in lower-dimensional spaces, or in which points are semantically related by taxonomic or set-inclusion relationships. In a limiting case, vector values correspond either to points (which, through comparison by equality, can model mathematical symbols) or to unknowns (which, through co-reference, can model mathematical variables).2. As a non-linguistic analogy, consider overlapping fragments of an image: Where overlaps match well enough, the fragments can be glued together to form a more comprehensive image of a scene, combining information from both fragments and potentially revealing new relationships.3. Section 10.6.7 suggests generically applicable training objectives that would favor representations and operations that (approximately) satisfy the axioms (Section A1.2) for unification and anti-unification; in this approach, operations may be performed by contextually informed neural functions.4. E.g., embeddings that represent uncertain values; nodes that lack links to optional content; leaf-level nodes that in effect summarize the content of some range of potential graph extensions.5. E.g., embeddings that represent narrower values; links to graph structure that may be modified by conditioning on a compatible leaf-level embedding. \n\t\t\t . The scope of recognizable parallels will depend on learned representations and comparison operators. Regularization (Section 8.4) can make representations more comparable; useful comparison operators could potentially resemble relaxed unification and generalization operators. \n\t\t\t . Argumentation mining points in this direction (Moens 2018; Galassi et al. 2020; Slonim et al. 2021) .2. In an awkward coarse-grained manner.3. Note, however, that cited-by relationships can have massive fanout, a pattern of use that may call for backward-facing structures richer than link-sets.4. Jaradeh et al. (2019), M. Jiang et al. (2020), and Cohan et al. (2020)   \n\t\t\t . Summaries (like other expressions) can be embodied in QNRs that provide increasing detail with increasing syntactic depth.2. See Martino et al. (2020).3. In this context, the difference between toxic text and discussions that embed examples of toxic text illustrates the importance of recognizing use-mention distinctions. Social media filters today may suppress both advocates and critics of offensive views. \n\t\t\t . Here, links to NL source text can be valuable: Literal wording may convey signals of shallow repetition. \n\t\t\t . Note that open-ended reasoning likely calls for conditional computation; potentially relevant architectural components and training methods are discussed in Cases et al. (2019) , Rosenbaum et al. (2019), and Banino, Balaguer, and . \n\t\t\t . Reviewed in J. Zhou et al. (2020)  and Wu et al. (2021) . Transformers in effect operate on fully connected graphs, but on sparse graphs, GNNs can provide greater scalability and task-oriented inductive biases (Addanki et al. 2021) , as well as more direct compatibility with QNRs.2. and Akoury, Krishna, and Iyyer (2019)  3. Including both image and joint image-language models; see Zellers et al. (2018) , J. Yang et al. (2018) , Lee et al. (2019), and Bear et al. (2020) . \n\t\t\t . X., Brown et al. (2020), and Y. Liu et al. (2020)   \n\t\t\t . Re. multi-scale masking, see Joshi et al. (2020) .2. See McCann et al. (2018), X. Liu et al. (2019), Alex Ratner et al. (2018), and Alexander Ratner et al. (2020).3. Analogous language-infused mechanisms are described in Shah et al. (2018) , Luketina et al. (2019), and Baroni (2020) . \n\t\t\t . In a general sense, completion tasks can include not only sequence prediction and cloze tasks, but also question answering and other capabilities shown by language models in response to prompts (see Brown et al. (2020) ); prediction in abstracted (latent space) domains can also support a range of tasks.See Oord, Li, and Vinyals (2019). \n\t\t\t . Aided by structural regularity (Section 8.4.) A similar approach might prove fruitful in training models that produce flat vector representations, which naturally have smooth distance metrics. This basic approach (predicting learned representations rather than raw inputs) is applied in Larsen et al. (2016)  and related work.2. E.g., they may map approximately lattice-respecting to strongly lattice-incompatible sets of representations. \n\t\t\t . E.g., Keskar et al. (2019) .3. E.g., Mager et al. (2020 ).   4. Fu et al. (2018, J. Wang et al. (2018 ), Johnson, Douze, and Jégou (2019 ), and Jayaram Subramanya et al. (2019   \n\t\t\t .See Veličković and Blundell (2021)  and included references.2. Reviewed inYe Zhang et al. (2017)  and Mitra and Craswell (2018) . \n\t\t\t . See Y. Lu et al. (2018)  and Arivazhagan et al. (2019) .2. A limiting case of this task would be semi-autonomous production of content, potentially on a large scale, guided by only the most general indications of purpose; see Section 12.2. \n\t\t\t . Verga et al. (2020) , Guu et al. (2020) , and Xu et al. (2020)  discuss Wikipedia-oriented systems.2. To enable retrieval of similar yet potentially clashing content, (some) embeddings should represent, not the concrete semantic content of expressions (in effect, answers to potential questions), but the kinds of questions that the content can answer, an important distinction noted above. Relevant clashes would then be indicated by failures of partially successful attempts at soft unification between new and retrieved content.3. Because statements may provide support for other statements, providing such material is related to argumentation, where automated, corpus-based methods are an area of active research; for example, see Lawrence and Reed (2020)  and Slonim et al. (2021) .4. A process illustrated by Xu et al. (2020) . \n\t\t\t . For examples of related work, see Shah et al. (2018)  and Abramson et al. (2021) . Relatively simple linguistic representations have emerged spontaneously; see Mordatch and Abbeel (2018)  and Lazaridou and Baroni (2020) . \n\t\t\t . Including games that require long-term planning (Vinyals et al. 2019; OpenAI et al. 2019 ).2. Luketina et al. (2019)  reviews progress and calls for \"tight integration of natural language understanding into RL\".3. As noted above, it is reasonable to expect that the most general and frequently used kinds of knowledge would be encoded, not in declarative representations that enable multi-step inference, but in model parameters that enable direct decision and action; this distribution of functionality would parallel Kahneman's System-1/System-2 model of human cognition (Kahneman 2011).   \n\t\t\t . Along with filtering based on detailed comparisons.2. Pattern completions may suggest structures; sampling guided by embeddings may suggest components.3. Simon (1988). Note that Simon describes planning as a design process. \n\t\t\t . Illustrated by work in Stump et al. (2019) , Mo et al. (2019) , and Chen and Fuge (2019). \n\t\t\t . E.g., M. Jiang et al. (2020)  and Raghu and Schmidt (2020)  2. Pólya (1990)  3. Szegedy (2020) suggests deriving formal expressions from NL text. \n\t\t\t . E.g., based on multi-source consistency, consensus, coherence, consilience, and provenance (Section 9.5.2). Current filtering methods appear to rely heavily on judgments of source quality (a domain-insensitive, non-content-based proxy for epistemic reliability), perhaps the simplest use of provenance.2. A problem discussed in Carlini et al. (2021) . \n\t\t\t . Translation would be subject to semantic imprecision due to differences in expressive capacity.2. Approximately-range of compensation is substantial (e.g., see Tee (2021).3. An optimized (\"FastFormer\") model derived from BERT can perform inference at a cost of about 18 US$/100 million queries (Kim and Hassan 2020) . \n\t\t\t . Effective altruists please take note.2. Drexler (2019)   \n\t\t\t . An application of the \"Principle of Least Privilege\" in system design.3. See Drexler (2019, Section 9.7).4. Bostrom (2014)  and Russell (2019)   \n\t\t\t . Along lines suggested by Stuart Russell (Wolchover 2015) ; see also discussion inDrexler  (2019, Section 22). \n\t\t\t . In particular, studies of lattice semantics in NL, unification and generalization in symbolic computation, and lessons learned in the broader study of neuro-symbolic ML. \n\t\t\t . More precisely, among potential more general expressions, anti-unification yields the most specific instance. \n\t\t\t . Here, larger sets are less specific and hence provide less information.2. Guhe et al. (2010) , Martinez et al. (2017) , and Amiridze and Kutsia (2018)   \n\t\t\t . Clark et al. (2021)  provides a more extensive and formal presentation of lattice semantic structure in domains closely aligned with those considered here. \n\t\t\t . \"Feature structure\" representations are particularly relevant to QNRs; Knight (1989)  reviews feature-structure unification and generalization in NL semantics.2. Tousch, Herbin, and Audibert (2008), Velikovich et al. (2018), and Wannenwetsch et al. (2019)3. Cimiano, Hotho, and Staab (2005)  4. Ganter and Wille (1997, 1999) , Cimiano, Hotho, and Staab (2005) , Belohlavek (2011), Eppe et al. (2018), and Clark et al. (2021)   \n\t\t\t . Paulson (1986)  and Felty and Miller (1988)  3. Excluding subtrees that contain the same symbol when cyclic graphs are disallowed. \n\t\t\t . Jaffar and Maher (1994)   \n\t\t\t . For example, ranges of compatible meanings with respect to various properties, implying what are in effect interval constraints on those properties, a familiar (if perhaps too rigid) constraint structure (see Benhamou 1995) .2. E.g., in Sessa (2002) , Medina, Ojeda-Aciego, and Vojtáš (2004) , Weber et al. (2019) , andMinervini et al. (2020).3. In addition, lattice operations may be mixed: In combining information, differences of some kinds should lead to rejection or narrowing, while differences of other kinds should lead to generalization. For example if we are combining pieces of evidence about a cat (perhaps from two photographs), some properties should be unified (pictures of spots that differ only in visibility and angle of view should narrow possible models of coloration, while a difference of orange vs. black should lead to failure and reject the hypothesis \"same cat\"). By contrast, if one photo shows a sleeping cat and the other an alert cat, the combined description should represent a cat that is not always asleep. Differences between kinds of differences should be learned from and conditioned on tasks. \n\t\t\t . Consider the union of disjoint volumes, each already at the limit of representable complexity. Intersections can (but do not necessarily) suffer from a similar difficulty. \n\t\t\t . Note that adherence to lattice properties in representational vector spaces is important only in those regions/manifolds that are actually used for representation. \n\t\t\t .See Veličković and Blundell (2021)  and included references. \n\t\t\t . Bantu languages include unusually complex and expressive systems of tense and aspect (Nurse and Philippson 2006; Botne and Kershner 2008) . \n\t\t\t . When feasible, of course. Another reason is the natural (and often semantically appropriate) commutativity implicit in vector representations viewed as sums of components. \n\t\t\t . And conversely, maladaptive precision in the form of (for example) forced number and gender distinctions.2. E.g., color, mass, temperature, frequency, speed, color, loudness, age, beauty, and danger.3. Gärdenfors (2000)  and Lieto, Chella, and Frixione (2017)   \n\t\t\t . Klys, Snell, and Zemel (2018) , Kingma and Dhariwal (2018) , R. Shen et al. (2020)   \n\t\t\t . A good recent example is Shen et al. (2020) , which finds that diverse faces can be wellrepresented in 100-dimensional spaces (Härkönen et al. 2020 ).2. In principle, to distinguish among millions of faces requires distinguishing on the order of 10 gradations on each of 6 dimensions, but typical embeddings are far richer in both distinctions and dimensionality.3. Within the domain of concrete, interpretable images, ~100 dimensional embeddings can represent not only faces, but also diverse object classes and their attributes, thereby representing (in effect) interpretable \"noun-and-adjective\" combinations, few of which can be compactly and accurately described in NL; e.g., see Härkönen et al. (2020) .4. Is the rock hard and sharp, or a piece of fine-grained ultramafic basalt? 5. On the internet, emoji have emerged to compactly express expression-level sentiment (e.g., and ), and these can express meanings distributed over more than one dimension (consider , , and ). \n\t\t\t . E.g., embeddings that generalize NL conjunctive/causal expressions could presumably express meanings like \"object X, (probably) together with and (possibly) because of Y\", and do so with graded degrees of probability or epistemic confidence.2. It is natural to want semantic spaces that express joint probability distributions as well as relationships in Pearl's do-calculus; the blurry distinction between these and the NL-like semantic spaces outlined above points to the soft boundaries of NL-centric conceptions of NL + . \n\t\t\t . Expressions that are frequently reduced to abbreviations are likely to represent broadly useful, lexical-level meanings: IIRC, ISTM, AFAIK, IMO, etc. \n\t\t\t . While blurring the distinction between values and variables; see Section A1.4.3, which notes potential relationships between constraint-based unification and variable binding in logic programming. \n\t\t\t . A vocabulary which describes structures, functions, relationships, processes, observations, evidence, interventions and causality in systems of extraordinary complexity and human importance.2. A count that omits, for example, inflected forms (Brysbaert et al. 2016) . \n\t\t\t . In addition to these considerations, note that components could potentially occupy relatively simple and well-structured semantic spaces, facilitating their interpretation even in the absence of specific training examples. Improving the interpretability of novel lexical components would feed through to improvements in the interpretability of novel elements of a lexical-unit vocabulary. \n\t\t\t . The set of mechanisms outlined here is intended to be illustrative rather than exhaustive, detailed, or optimal. The discussion touches on tutorial topics for the sake of readers who may notice computational puzzles without immediately recognizing their solutions. \n\t\t\t . Internal, index-encoded links require a bit of additional bookkeeping (i.e., remembering subexpression positions) but can describe DAGs and cyclic graphs.2. Indices (~1 byte?) account for the rest. \n\t\t\t . Vocabulary size depends on how \"distinct words\" are defined; reasonable definitions and methodologies yield numbers that differ, but not by orders of magnitude.See Brysbaert  et al. (2016). \n\t\t\t . To support near-neighbor indexing, embeddings should be unique, not elements of a \"vocabulary\".", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/QNRs_FHI-TR-2021-3.0.tei.xml", "id": "988f9a4fc492acb4eebe190e2e9fb80d"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Longtermism holds roughly that in many decision situations, the best thing we can do is what is best for the long-term future. The scope question for longtermism asks: how large is the class of decision situations for which longtermism holds? Although longtermism was initially developed to describe the situation of cause-neutral philanthropic decisionmaking, it is increasingly suggested that longtermism holds in many or most decision problems that humans face. By contrast, I suggest that the scope of longtermism may be more restricted than commonly supposed. After specifying my target, swamping axiological strong longtermism (swamping ASL), I give two arguments for the rarity thesis that the options needed to vindicate swamping ASL in a given decision problem are rare. I use the rarity thesis to pose two challenges to the scope of longtermism: the area challenge that swamping ASL often fails when we restrict our attention to specific cause areas, and the challenge from option unawareness that swamping ASL may fail when decision problems are modified to incorporate agents' limited awareness of the options available to them.", "authors": ["David Thorstad"], "title": "The scope of longtermism", "text": "Introduction If we play our cards right, the future of humanity will be vast and flourishing. The earth will be habitable for at least another billion years. During that time, we may travel well beyond the earth to settle distant planets. And increases in technology may allow us to live richer, longer and fuller lives than many of us enjoy today. If we play our cards wrong, the future may be short or brutal. Already as a species we have acquired the capacity to make ourselves extinct, and many experts put forward alarmingly high estimates of our probability of doing so  (Bostrom 2002; Leslie 1996; Ord 2020) . Even if we survive long into the future, technological advances may be used to breed suffering and oppression on an unimaginable scale  (Sotala and Gloor 2017; Torres 2018) . Many have taken these considerations to motivate longtermism: roughly, the thesis that in a large class of decision situations, the best thing we can do is what is best for the long-term future  (Beckstead 2013 ; Greaves and MacAskill 2019; Mogensen forthcoming;  Ord 2020; MacAskill 2022) . The scope question for longtermism asks: how large is the class of decision situations for which longtermism holds? Longtermism was originally developed to describe the decisions facing present-day, cause-neutral philanthropists. Longtermists suggest that the best thing philanthropists can do today is to safeguard the long-term future, for example by mitigating risks of human extinction. But many have held that the scope of longtermism extends considerably beyond cause-neutral philanthropic decisionmaking. Hilary Greaves and Will MacAskill  (Greaves and MacAskill 2019)  suggest that the cause-specific choice between two antimalaria programs should be governed, not by their direct effects of preventing present-day malaria deaths, but by the potential far-future impact of each intervention.  1  And Owen Cotton-Barratt suggests that even most mundane decisions, such as selecting topics for dinner-table conversation, should be made to promote proxy goals which track far-future value  (Cotton-Barratt 2021) . In this paper, I argue that the scope of longtermism may be narrower than often supposed. Section 2 clarifies my target: swamping axiological strong longtermism (swamping ASL). Section 3 argues that swamping ASL may be true in the case of present-day, causeneutral philanthropy. However, Sections 4-5 give two arguments for the rarity thesis that longtermist options of the type needed to vindicate swamping ASL are rare. Section 6 argues that the rarity thesis poses no challenge to swamping ASL in the special case of present-day, cause-neutral philanthropy. However, Sections 7-8 use the rarity thesis to put two limits on the scope of swamping ASL. Section 7 poses the area challenge: swamping ASL fails in many specific cause areas, such as choosing between anti-malaria programs. Section 8 poses the challenge from option unawareness: swamping ASL often fails when decision problems are modified to incorporate our limited ex ante awareness of the options available to us. Section 9 concludes. \n Preliminaries \n Longtermism: axiological and ex ante Longtermism comes in both axiological and deontic varieties. Roughly speaking, axiological longtermism says that the best options available to us are often near-best for the long-term future, and deontic longtermism says that we often should take some such option. Longtermists standardly begin by arguing for axiological longtermism, then arguing that axiological longtermism implies deontic longtermism across a wide range of deontic assumptions. In order to avoid complications associated with the passage between axiological and deontic claims, I focus on axiological rather than deontic longtermism. Axiological longtermism can be construed as an ex ante claim about the values which options have from an ex ante perspective, or as an ex post claim about the value that options will in fact produce. It is generally thought that ex post longtermism is more plausible than ex ante longtermism, since many of our actions may in fact make a strong difference to the course of human history, even if we are not able to foresee what that difference will be. For this reason, most scholarly attention has focused on ex ante versions of longtermism, and I follow this trend here. The best-known view in this area is what has been called axiological strong longtermism (ASL): (ASL) In a wide class of decision situations, the option that is ex ante best is contained in a fairly small subset of options whose ex ante effects on the very long-run future are best.  (Greaves and MacAskill 2019) . My target in this paper will be a restricted form of ASL. \n Swamping axiological longtermism There are two ways in which ASL could be true. First, there might be a swamping longtermist option whose expected long-term benefits exceed in magnitude the expected short-term effects produced by any option. 2 I call these swamping longtermist options because their long-term effects begin to swamp short-term considerations in determining ex ante value. We will see in Section 3 that ASL holds in many decision problems involving swamping longtermist options, so the first way that ASL could be true would be if it held in virtue of swamping longtermist options. \n Swamping axiological strong longtermism (Swamping ASL) In a wide class of decision situations, the option that is ex ante best (a) is a swamping longtermist option, and (b) is contained in a fairly small subset of options whose ex ante effects on the very long-run future are best. My focus in this paper will be on swamping ASL. Second, the best option may be a convergent option, an option whose long-term value is near-best, and whose short-term value is at least modestly comparable to the bestachievable short-term values.  3  If there are few swamping longtermist options, we could defend ASL through the convergence thesis that the best options are often convergent options. For example, you might think that the best thing we can to do ensure a good future is to promote economic growth  (Cowen 2018) , and that is also among the best things we can do for the short term. I focus on swamping ASL over the convergence thesis for three reasons. First, swamping ASL figures in leading philosophical arguments for ASL and in most nonphilosophical treatments of longtermism. Second, swamping ASL is the most distinct and revisionary form of ASL, because it tells us that the short-termist options we might have assumed to be best are in fact often not best. Third, swamping ASL underlies many of the most persuasive arguments from axiological to deontic longtermism, which rely on the claim that sufficiently strong duties to promote impartial value may trump competing nonconsequentialist duties. As we move away from swamping longtermism, obligations to promote long-term value will diminish in strength, putting pressure against the inference from axiological to deontic longtermism. \n The rarity thesis In this paper, I argue for the rarity thesis that swamping longtermist options are rare, in the sense that the vast majority of the options that we face as decisionmakers are not swamping longtermist options. The rarity thesis does not imply that swamping ASL fails in all decision situations. Indeed, I argue in Section 3 that swamping ASL is plausible in the case of present-day cause-neutral philanthropy. However, the rarity thesis does suggest that the scope of swamping ASL may be more limited than is often supposed. I use the rarity thesis to pose two challenges to the scope of swamping ASL: the area challenge that swamping ASL fails in many specific cause areas of interest, and the challenge from option unawareness that swamping ASL fails in many decision problems when those problems are modified to account for decisionmakers' unawareness of relevant options. Summing up, my target in this paper is ex ante, axiological, swamping strong longtermism. I use the rarity thesis to suggest that swamping ASL has more limited scope than we might otherwise suppose. But first, let us consider where ASL may be plausible. \n The good case: Present-day cause-neutral philanthropy Many effective altruists have claimed that swamping ASL is true in the case of present-day, cause-neutral philanthropic decisionmaking. In this section, I argue that they may well be right. Let a strong swamping option be one whose expected far-future impact exceeds the best-achievable expected short-term impact by a large margin -say, for concreteness, by a factor of ten.  4  We can argue that swamping ASL holds for present-day philanthropy using the argument from strong swamping. 5 This argument claims first, that there exists at least one strong swamping option for present-day philanthropists, and second that the existence of a strong swamping option in a decision problem implies swamping ASL for that problem. Here is an argument for the second premise. Assume there exists a strong swamping option. Note first that the best option must be a swamping longtermist option, since a strong swamping option is better than any non-swamping option. Note next that the best option must have at least 9/10 of the best-achievable long-term value, since the existence of a strong swamping option implies that short-term value differences cannot make up any larger gap than this. Hence the best option is a swamping longtermist option with near-best long-term effects, and swamping ASL follows. The weight of the argument from strong swamping is carried by the first premise, which asserts the existence of a strong swamping option. In the case of present-day philanthropy, we can argue for the existence of strong swamping options using the argument from singletrack dominance. For a given option, the argument from single-track dominance makes three claims. First, it claims that the option exhibits: (Single-track dominance) The lion's share of the option's expected impact on the long-term future is driven by a single causal pathway or effect. For example, efforts to reduce extinction risk plausibly exhibit single-track dominance along the pathway of preventing human extinction. Single-track dominance allows us to simplify the overwhelming array of future effects by considering harms and benefits along a single track. Second, the option may produce overwhelmingly positive far-future benefits along this track: (Nontrivial probability of significant benefit) For some extremely large value of N, the probability that this option produces long-term benefits of value exceeding N is nontrivial, considering only effects along the specified track.  6  For example, we might require that N is the value of saving a billion future lives and that the probability of producing long-term benefits in excess of N is at least one in a million. Third, the option is much less likely to produce overwhelmingly large future harms along this track: (Relatively trivial probability of significant harm) With N as before, the probability that this option produces long-term harms of value beneath −N is significantly lower than than the probability of producing benefits with value exceeding N, considering only effects along the specified track. Most options with these three features will be strong swamping options.  7  To show that ASL holds for present-day, cause-neutral philanthropy, it remains to argue that some options available to present-day philanthropists are swamping longtermist options in virtue of meeting the conditions set out in the argument from single-track dominance. There are many detailed arguments alleging that these conditions are met by specific philanthropic options  (Bostrom 2013; Tarsney ms; Ord 2020) . Here is one such argument (Greaves and MacAskill 2019; Newberry 2021a;  Ord 2020) . One way that humans might go extinct is through the impact of a large asteroid on earth. Indeed, there is mounting evidence that an asteroid impact during the Cretaceous period killed every land-dwelling vertebrate with mass over five kilograms  (Alvarez et al. 1980; Schulte et al. 2010) . As recently as 2019, an asteroid 100 meters in diameter passed five times closer to the earth than the average orbital distance of the moon and was detected only a day before it arrived  (Zambrano-Marin et al. 2021) . NASA classifies asteroids with diameter greater than 1 kilometer as catastrophic, capable of causing a global calamity or even mass-extinction. Our best estimates suggest that such impacts occur on earth about once in every 6,000 centuries  (Stokes et al. 2017 ). Plausibly, it is worth our while to detect and prepare for such events. Since the 1990s, the SpaceGuard survey has mapped approximately 95% of the nearearth asteroids with diameters exceeding 1 km at a cost of $70 million. Assuming a conservative chance of one in a million that an asteroid impact of this magnitude would result in human extinction, this project would, in expectation, be worth funding based on its short-term value alone if we value early warning of an extinction-causing impact within the next century at at least $74 trillion. That is a close call, and you might well be skeptical of the short-term value of such a project. But now, consider that estimates of the expected number of future humanlike lives range from about 10 13 to 10 55  (Bostrom 2014; Newberry 2021b) . This puts the project's expected cost of detecting an extinction-causing asteroid impact, counting only impacts within the next century, at no more than about $7 per expected future life, and at fractions of a penny using anything but the most conservative estimate. For comparison, our best estimates put the cost of saving a life through short-termist interventions at several thousand dollars (GiveWell 2021), making the SpaceGuard survey a prime example of a strong swamping option, with expected long-term value well in excess of the best expected short-term impacts if we have any confidence at all in our ability to prepare for and survive an otherwise-catastrophic impact with sufficient warning. Now the SpaceGuard survey has already been funded, but there is doubtless more that could be done to reduce extinction risks from asteroid impacts, such as mapping the remaining 5% of large near-earth asteroids or developing contingency plans for deflecting asteroids and surviving asteroid impacts. And many other interventions aimed at reducing existential risks have been claimed to be yet more cost-effective than these  (Ord 2020) . So I do not want to deny the existence of strong swamping options, and hence I do not deny that swamping ASL is plausible in the special case of present-day cause neutral philanthropy. However, as we move away from the case of cause-neutral philanthropy, matters begin to look different. \n Rapid diminution In the next two sections, I defend the rarity thesis that swamping longtermist options are rare. Sections 7-9 then assess the impact of the rarity thesis on the scope of swamping ASL. The first argument for the rarity thesis is the argument from rapid diminution. Fix an option o and consider the probability distribution over long-term impacts of o. 8 In most cases, the probabilities of long-term impacts decrease as those impacts increase in magnitude. If probabilities of impacts decrease more slowly than the magnitudes of those impacts increase, then the expected long-term consequences of o may be astronomically high. But if the probabilities of large impacts decrease quickly, the expected long-term impacts of o may be quite modest. Rapid diminution is a familiar feature of many of the best-known probability distributions. For example, suppose that we model the expected long-term impact of o using a normal distribution, centered around the origin, with a standard deviation equivalent to the value of ten lives saved. On this model, the probability of long-term impacts exceeding five times this value is less than one in a million. And the probabilities of astronomical long-term impacts, while nonzero, will be so negligible as to have no significant impact on the expected long-term impact of o. The argument from rapid diminution claims that most options exhibit rapid diminution in the probability of long-term impacts, limiting the contribution that long-term impacts can make to the expected value of those options. This argument is supported by persistence skepticism: the view that many of our actions do not make a large persisting impact on the long-term future. We can assess the case for persistence skepticism by looking at the burgeoning academic field of persistence studies, which studies examples of persistent long-term changes  (Alesina and Giuliano 2015; Nunn 2020) . Persistence studies often returns surprising negative results, where effects that we might have expected to persist for a long time evaporate after several decades. For example, given the scale of American bombing in Japan and Vietnam, one might expect persistent economic effects in the heaviest-hit areas. Given the number of people affected and the magnitude of potential effects, this is exactly the type of persistent effect that would interest a longtermist. But in fact, the effects on population size, poverty rates and consumption patterns appear to have already vanished  (Davis and Weinstein 2008; Miguel and Roland 2011) . Now it is true that persistence studies has identified a few-dozen effects which might be more persistent. For example, the introduction of the plough may have affected fertility norms and increased the gendered division of labour  (Alesina et al. 2011 (Alesina et al. , 2013 ; the African slave trade may have stably reduced social trust and economic indicators in the hardesthit regions  (Nunn 2008; Nunn and Wantchekon 2011) ; and the Catholic Church may be responsible for the spread of so-called WEIRD personality traits identified by comparative psychologists  (Schulz et al. 2019 ). However, these findings need to be taken with three grains of salt. First, many of these findings are controversial, and alternative explanations have been proposed  (Kelly 2019; Sevilla 2021 ). Second, these findings are few-and-far-between, so together with other negative findings they may not challenge the underlying rarity of strong long-term effects. And finally, most of the examples in this literature also involve short-term effects of comparable importance to their claimed long-term effects. Hence the persistence literature may not provide strong support for the swamping longtermist's hope that persistent long-term effects could swamp short-term effects in importance. At the same time, there is no doubt that some actions have a nontrivial probability of making persistent changes to the value of the future far greater than any of their short-term effects. As a result, we cannot get by with the argument from rapid diminution alone. We need to supplement the argument from rapid diminution with a second argument: the argument from washing out. \n Washing out The second argument for the rarity thesis is the argument from washing out. Although many options have nontrivial probabilities of making positive impacts on the future, they also have nontrivial probabilities of making negative impacts. For example, by driving down the road I might crash into the otherwise-founder of a world government, but I might also crash into her chief opponent. As a result, the argument from washing out holds that there will be significant cancellation between possible positive and negative effects in determining the expected values of options. There are two related ways that the argument from washing out can be articulated. The first begins with the popular Bayesian idea that complete ignorance about the long-term value of an option should be represented by a symmetric prior distribution over possible long-term values. Next, the argument notes that we are often in a situation of evidential paucity: although we have some new evidence bearing on long-term values, often our evidence is quite weak and undiagnostic. As a result, the prior distribution will exert a significant influence on the shape of our current credences, so if the prior is symmetric then our current credences should be fairly symmetric as well. And a near-symmetric probability distribution over long-term impacts gives significant cancellation when we take expected values. We can make a similar point by arguing for forecasting pessimism, the view that it is often very difficult to predict the impact of our actions on far-future value. For example, there is no doubt that the Roman sacking of Carthage had a major impact on our lives today, by cementing the Roman empire and changing the course of Western civilization. But even today, let alone with evidence available at the time, it is very difficult to say whether that impact was for good or for ill. Forecasting pessimism generates a type of washing-out between possible positive and negative forecasts. 9 When we make forecasts based on sparse data, we need to take account of the fact that the data we have been dealt is a noisy reflection of the underlying reality. As phenomena become more unpredictable and our data becomes increasingly sparse, we should grow more willing to chalk up any apparent directionality in our forecasts to noisiness in the hand of data that nature has dealt us. In other words, as forecasting becomes more difficult we get increasing wash-out between possible positive and negative forecasts that we could have made based on different samples of data. Why should we be pessimistic about our ability to forecast long-run value? Intuitions about the sacking of Carthage are well and good, but it would be nice to have some concrete theoretical considerations on the table. Here are three reasons to think that we are often in a poor position to forecast long-run value. First, we have limited and mixed track records of making long-term value forecasts. We do not often make forecasts even on a modest timeline of 20-30 years, and as a result there are only a few studies assessing our track record at this timescale.  10  These studies give a mixed picture of our track record at predicting the moderately-far future: in some areas our predictions are reasonably accurate, whereas in others they are not. But the longtermist is interested in predictions at a timescale of centuries or millennia. We have made and tested so few predictions at these time scales that I am aware of no studies which assess our track record at this timescale outside of highly circumscribed scientific domains, and if our moderate-future track record is any indication, our accuracy may decline quite rapidly this far into the future. Second, there is an enormous amount of practitioner skepticism on behalf of prominent academic and non-academic forecasters about the possibility of making forecasts on a timescale of centuries, particularly when we are interested in forecasting rare events, as longtermists often are. Very few economists, risk analysts, and other experts are willing to make such predictions, citing the unavailability of data, a lack of relevant theoretical models, and the inherent unpredictability of underlying systems  (Freedman 1981; Goodwin and Wright 2010; Makridakis and Taleb 2009) . And when risk analysts are asked to consult on the management of very long-term risks, they increasingly apply a variety of non-forecasting methods which enumerate and manage possible risks without any attempt to forecast their likelihood  (Marchau et al. 2019; Ranger et al. 2013) . If leading practitioners are unwilling to make forecasts on this timescale and increasingly suggest that we should act without forecasting, this is some evidence that the underlying phenomena may be too unforeseeable to effectively forecast. Third, value is multidimensional. The value of a time-slice in human history is determined by many factors such as the number of people living, their health, longevity, education, and social inclusion. It is often relatively tractable to predict a single quantity, such as the number of malaria deaths that will be directly prevented by a program of distributing bed nets. And when we assess the track records of past predictions, we often assess predictions of this form. But the longtermist is interested in predicting value itself, which turns on many different quantities. This is harder to predict: distributing bed nets also affects factors such as population size, economic growth, and government provision of social services  (Deaton 2015) . So even if we think that the long-term effects of a program along a single dimension of value are fairly predictable, we may think that the ultimate value of the intervention is much less predictable. Summing up, the argument from washing out claims that we often get significant cancellation between possible positive and negative effects of an intervention when taking expected values. One window into washing out comes from evidential paucity: because we have little evidence about long-term impacts, we should adopt a fairly-symmetric prior distribution over possible long-term impacts. The same phenomenon occurs in thinking about forecasting. Because our evidence about far-future value is sparse, we should think that our forecasts could easily have been different if we had received different evidence about the future, and as a result we get significant cancellation between possible positive and negative forecasts of far-future value. Together, the arguments from washing out and rapid diminution suggest that large long-term impacts may be less probable than otherwise thought, and may be significantly cancelled by potential negative long-term impacts. This grounds the rarity thesis that swamping longtermist options are rare. But this does not mean that there are no swamping longtermist options. Plausibly, the case of cause-neutral philanthropy remains untouched by these arguments for the rarity thesis. \n The good case revisited In the case of present-day cause-neutral philanthropy, the argument from single-track dominance allows us to avoid all of the arguments raised in the previous two sections. We can see this both abstractly, by considering the argument from single-track dominance itself, and concretely, by thinking about the case of asteroid detection. Begin with the problem of rapid diminution: the probabilities of large long-term impacts diminish rapidly. The second premise of the argument from single-track dominance avoids this problem by holding that the option in question has nontrivial probability of producing a highly significant benefit. When this is true, rapid diminution elsewhere in the probability distribution will not threaten the status of the option as a swamping longtermist option. In fact, rapid diminution may even help to secure the third premise of the argument from single-track dominance: that the option in question has comparatively negligible probability of significant future harm. The general phenomenon of rapid diminution may give us good reason to accept this third premise as a default stance, unless given special reason to doubt it. More concretely, the argument for rapid diminution drew on skepticism about the persistence of short-term effects into the long-term future. But it is not hard to see how the proposed effects of asteroid detection, namely preventing human extinction, could persist into the long-term future. Not being extinct is a status that can last for a very long time if we play our cards right. Turn next to the problem of washing out: possible long-term benefits may be significantly cancelled by possible long-term harms. The second and third premises of the argument from single-track dominance avoid this problem by claiming that probabilities of very large benefits significantly outweigh the probabilities of very large harms. More concretely, the first argument for washing out drew on evidential paucity: we don't have much evidence about the long-term effects of our actions. But asteroid detection is an area in which we do have significant evidence about possible long-term effects. This includes evidence from past asteroid impacts together with a good understanding of the determinants of asteroid impact force, which is sufficient to build compelling computational models of impact damages  (Stokes et al. 2017) . The second argument for washing out drew on forecasting skepticism: it's hard to predict the future. First, I argued that in many areas we have no good track record of predicting the far future. But astronomy is one of the few areas in which we do have a good track record of predictions on this time-scale. Second, I argued that experts are often unwilling to make forecasts of the relevant type. But the key forecast driving the example was a prediction by NASA scientists that the per-century probability of a catastrophic asteroid impact is approximately 1/6,000. Third, I argued that due to the multidimensionality of value we may only be able to estimate the probability of a catastrophic impact but not its value. This may be true for off-track effects, such as catastrophic impacts leading to non-extinction events, but it is not a significant problem if the outcome in question is human extinction. To evaluate whether preventing human extinction would be a good thing, we must only answer a single question: whether the continued existence of humanity would be a good thing. While answering this question is not straightforward, many of us are cautiously optimistic that the future will be good  (Beckstead 2013; Ord 2020; Parfit 2011) . So far, we have seen that the argument for single-track dominance allows us to accept the rarity thesis without questioning the truth of swamping ASL in the special case of present-day, cause-neutral philanthropy. However, I argue in the next two sections that the rarity thesis does pose two challenges to the scope of swamping ASL. \n The area challenge If we accept that swamping ASL holds in present-day philanthropy, then we should accept swamping ASL in some philanthropy-adjacent matters as well. For example, swamping ASL might govern career choice by indicating that we have strong reason to pursue global priorities research. But how far does the reach of swamping ASL extend? The area challenge to the scope of swamping ASL holds that swamping ASL fails in many cause areas sufficiently removed from cause-neutral philanthropy. To see how quickly the area challenge gains bite, consider a problem involving causespecific philanthropic giving: a donor is committed to funding anti-malaria work, but wants to do the most good possible in this space.  Greaves and MacAskill (2019)  suggest that ASL governs this choice as well, because the lion's share of the expected valuedifference between competing malaria charities will be driven by their differential impacts on the long-term future. While I don't want to pronounce on the prospects for convergent versions of ASL in this case, I think that the prospects for swamping ASL in this case are relatively poor. Certainly nothing like the argument from single-track dominance can be made here. We would not want to accept the first premise: that the lion's share of the expected longrun value of an anti-malarial program is driven by a single type of far-future effect. What the longtermist wants to stress here is precisely the opposite, that there are many different ways in which anti-malarial work can affect the far future, for example by changing the international balance of power, speeding technological growth and space settlement, or modulating population size, each of which makes an important contribution to the program's far-future value. Removing the commitment to single-track dominance, we may or may not want to accept the second premise, that an anti-malarial program has nontrivial probability of significant far-future benefit. But most natural arguments for the second premise would also be arguments against the third premise, which asserts that anti-malarial programs have comparatively trivial probabilities of significant far-future harm. As the premises of the argument from single-track dominance begin to fail, the arguments driving the rarity thesis gain traction against swamping ASL. Begin with the argument from washing-out. The first argument for washing-out cited evidential paucity: we have little concrete evidence about the long-term impacts of many actions. I think we should concede that this is the case with anti-malarial work. Although malaria prevention could have some impact on the very long-run balance of international power, for example, we don't have much to go on in determining what that impact might be. The second argument for washing-out cited forecasting pessimism: it is hard to predict the long-term impacts of our actions. Here all three of the arguments made for forecasting pessimism gain traction. As far as track records are concerned, we have no sizable track record of predicting the effects of global health programs on a scale of centuries or millennia. As far as practitioner skepticism is concerned, no existing anti-malarial organization attempts to estimate its very long-run impacts. Nor, for that matter, does the GiveWell foundation, a philanthropic organization which steers millions of dollars in grants towards anti-malarial programs, conducts extensive impact evaluations on its grantees, and is more sympathetic to longtermism than almost any other major philanthropic actor. So it looks as though practitioners take forecasting the long-run impacts of anti-malarial work to be very difficult. Finally, the multidimensionality of value is a strong obstacle in this area: in fact, anti-malarial work was the example that I used to illustrate the multidimensionality of value. Although we are relatively good at forecasting immediate health outcomes such as malaria deaths prevented, we have little experience in, or methodologies for predicting the overall value impact of these programs at long time scales. This matters because global health programs have many diverse sociopolitical impacts that determine their long-run value. The argument from rapid diminution also gains traction against swamping ASL here. The motivation for rapid diminution was skepticism about the persistence of short-term effects into the long-term future. If we are generally skeptical about long-term persistence, then we should demand some special reason to think that anti-malarial work is likely to have persistent effects on the far-future even though other actions often do not. And when the effects that we point to are quite speculative future possibilities, such as an accelerated pace of interstellar expansion due to regional economic growth, it is not clear what could be said in favor of expecting such persistent effects from anti-malarial work that would not tell against persistence skepticism more generally. Absent a detailed story, the persistence skeptic is likely to remain unconvinced. In this section, we have seen that the rarity thesis poses an area challenge to the scope of swamping ASL. Although swamping ASL may be true in cause-neutral philanthropy and some adjacent areas, in many other cause areas including many types of cause-specific philanthropy, swamping ASL may fail. In the next section, I pose a second scope challenge to swamping ASL: the challenge from option unawareness. \n Option unawareness Rational ex ante choice involves taking the ex ante best option from the options available to you. But which options are these? We might take a highly unconstrained reading on which any option that is physically possible to perform belongs to your choice set. But in practice, this reading seems to betray the ex ante perspective  (Hedden 2012) . Suppose you are being chased down an alleyway by masked assailants. A dead end approaches. Should you turn right, turn left, or stop and fight? Trick question! I forgot to mention that you see a weak ventilation pipe which, if opened, would spray your attackers with hot steam. That's better than running or fighting. Let us suppose that, in theory, all of this could be inferred with high probability from your knowledge of physics together with your present perceptual evidence, but you haven't considered it and you can hardly be blamed for that. Does this mean that you would act wrongly by doing anything except breaking the pipe? Many decision theorists have thought you would not act wrongly here. Just as ex ante choosers have limited information about the values of options, so too they have limited awareness of the many different options in principle available to them. Theories of option unawareness incorporate this element of ex ante choice by restricting choice sets to options which an agent is, in some sense, relevantly aware of  (Bradley 2017; Karni and Vierø 2013; Steele and Stefánsson 2021) . In the present case, this means that your options are as they first described them: turning right, turning left, or stopping to fight. Unless, perhaps, you happen to be James Bond. Now option unawareness does not pose any problem for swamping ASL in the case of cause-neutral philanthropy. That is because the philanthropist is already aware of some options, such as asteroid detection, which witness the truth of swamping ASL. Moreover, if the philanthropist were not aware of any such options, then it might turn out that the best thing she could do would be to fund option generation research  (Kalis et al. 2013; Greaves and MacAskill 2019) . If this research had a high probability of turning up a swamping longtermist option, then it could itself be a swamping longtermist option, and again swamping ASL would be saved. But outside the case of cause-neutral philanthropy, option unawareness combines with the rarity thesis to generate a challenge for the scope of ASL. By the rarity thesis, swamping longtermist options are rare. Many theories of option unawareness hold that, in typical choice problems, we face modestly-sized choice sets with at most dozens of options. By rarity, most modestly-sized choice sets will not contain any swamping longtermist options, so swamping ASL turns out to have fairly restricted scope. We can see this problem in context by thinking about the cause-specific philanthropist, who is committed to funding some anti-malarial program. Naïvely, we might pose her decision problem as follows: of any possible anti-malarial intervention, which should I take? And it is not so implausible that swamping ASL could hold in awareness-unconstrained problem. From an ex post perspective, there is surely some option that would bring enormous far-future benefits, such as protecting a village or region that contains an important future politician or inventor. And from an ex ante perspective, it is not so implausible that there is enough evidence to identify one such region after years of calculation. But most philanthropists face an awareness-restricted problem such as the following: out of the internationally prominent existing anti-malaria programs, which should I fund? We saw in Section 7 that the arguments for rarity suggest that none of these programs may be a swamping longtermist option. Unlike the case of cause-neutral philanthropy, we cannot save swamping ASL by appealing to meta-options such as conducting research into the availability of new options or gathering evidence about their values. That is because even if new options could in principle be discovered, in practice it might be quite expensive to identify and evaluate them. For comparison, one of the largest anti-malaria charities, the Against Malaria Foundation, has a budget in the tens of millions of dollars. But it would likely cost far more than that to scour the globe in search of rare future talent worth protecting. So the cost of expanding awareness here would approach or exceed the cost of simply funding malaria prevention across the globe. High costs of option-generation might be bearable in larger cause areas, such as existential risk prevention, which can bear a great deal of capital investment once discovered. But even a relatively high cap on capital needs may make longtermist option-generation into a wastefully suboptimal option. Summing up, the challenge from option unawareness holds that ex ante rational choice is restricted to the options of which we are relevantly aware. Because option unawareness often considerably restricts the size of our choice sets, the rarity thesis implies that in most choice problems we face as decisionmakers, swamping ASL fails. \n Conclusion This paper assessed the fate of ex ante swamping ASL: the claim that the ex ante best thing we can do is often a swamping longtermist option that is near-best for the long-term future. I gave a two-part argument that swamping ASL holds in the special case of present-day cause-neutral philanthropy: the argument from strong swamping that a strong swamping option would witness the truth of ASL, and the argument from single-track dominance for the existence of a strong swamping option. However, I also argued for the rarity thesis that swamping longtermist options are rare. I gave two arguments for the rarity thesis: the argument from rapid diminution that probabilities of large far-future benefits often diminish faster than those benefits increase; and the argument from washing out that probabilities of far-future benefits are often significantly cancelled by probabilities of far-future harms. I argued that the rarity thesis does not challenge the case for swamping ASL in presentday, cause-neutral philanthropy, but showed how the rarity thesis generates two challenges to the scope of swamping ASL beyond this case. First, there is the area challenge that swamping ASL often fails when we restrict our attention to specific cause areas. Second, there is the challenge from option unawareness that swamping ASL often fails when we modify decision problems to incorporate agents' unawareness of relevant options. In some ways, this may be familiar and comforting news. For example, Greaves (2016) considers the cluelessness problem that we are often significantly clueless about the ex ante values of our actions because we are clueless about their long-term effects. Greaves suggests that although cluelessness may be correct as a description of some complex decisionmaking problems, we should not exaggerate the extent of mundane cluelessness in everyday decisionmaking. A natural way of explaining this result would be to argue for a strengthened form of the rarity thesis on which in most mundane decisionmaking, the expected long-term effects of our actions are swamped by their expected short-term effects. So in a sense, the rarity thesis is an expected and comforting result. In addition, this discussion leaves room for swamping ASL to be true and important in the case of present-day, cause-neutral philanthropy as well as in a limited number of other contexts. It also does not directly pronounce on the fate of ex-post versions of ASL, or on the fate of non-swamping, convergent ASL. However, it does suggest that swamping versions of ASL may have a more limited scope than otherwise supposed. \n\t\t\t  Relatedly, Mogensen (2021)  suggests that it may not be clearly better for a longtermist to donate to the Against Malaria Foundation rather than the Make-A-Wish Foundation due to uncertainty about long-run impacts. \n\t\t\t More formally, suppose that value is temporally separable, so that V o = S o + L o where V o , S o , L o are the overall, short-term and long-term values of option o. Assess changes in value ∆V o , ∆S o , ∆L o relative to a baseline, such as the effects of inaction. And take an expectational construal of ex ante value. Then a swamping longtermist option is such that E[∆L o ] > max o ∈O |E[∆S o ]| where O are the options available to the actor. This is a simplification of the model from (Greaves and MacAskill 2019).3 For the purposes of this paper, the best option will be a convergent option if it is not a swamping longtermist option. We may wish to expand this taxonomy to capture gradations between swamping and convergent ASL, but that will not be my focus here. \n\t\t\t That is, E[∆L o ] > 10 * max o ∈O |E[∆S o ]|.5 This argument is largely drawn from Greaves and MacAskill (2019) . \n\t\t\t I.e. for specified N large and k nontrivial, Pr(∆L o > N) > k, ignoring off-track effects.7 Strictly speaking, it is possible for an option with these features to fail to be a swamping longtermist option, for example if for some M > N we have Pr(∆L o > M) < Pr(∆L o < −M). I hope it is clear how the requirements might be tightened to avoid this worry, but also that this tightening would be a distraction in most cases of interest, where this is not a live worry. \n\t\t\t I.e. consider the probability distribution over the partition {[∆L = k] : k ∈ R}. \n\t\t\t Among the many ways to give formal expression to this idea, Gabaix and Laibson's (2021)  as-if discounting brings out the similarity to the argument from evidential paucity by highlighting the role of priors.10 For domain-specific track records see Albright (2002) ; Kott and Perconti (2018) ; Parente and Anderson-Parente (2011); Risi et al. (2019)  and Yusuf (2009) . For discussion see Fye et al. (2013)  and Mullins (2018) .", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Scope-Longtermism.tei.xml", "id": "5dc642e6ca5b00b4e548395f6a8cda3a"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Change is hardly a new feature in human affairs. Yet something has begun to change in change. In the face of a range of emerging, complex, and interconnected global challenges, society's collective governance efforts may need to be put on a different footing. Many of these challenges derive from emerging technological developments -take Artificial Intelligence (AI), the focus of much contemporary governance scholarship and efforts. AI governance strategies have predominantly oriented themselves towards clear, discrete clusters of pre-defined problems. We argue that such 'problem-solving' approaches may be necessary, but are also insufficient in the face of many of the 'wicked problems' created or driven by AI. Accordingly, we propose in this paper a complementary framework for grounding long-term governance strategies for complex emerging issues such as AI into a 'problemfinding' orientation. We first provide a rationale by sketching the range of policy problems created by AI, and providing five reasons why problem-solving governance approaches to these challenges fail or fall short. We conversely argue that that creative, 'problem-finding' research into these governance challenges is not only warranted scientifically, but will also be critical in the formulation of governance strategies that are effective, meaningful, and resilient over the long-term. We accordingly illustrate the relation between and the complementarity of problem-solving and problem-finding research, by articulating a framework that distinguishes between four distinct 'levels' of governance: problemsolving research generally approaches AI (governance) issues from a perspective of (Level 0) 'businessas-usual' or as (Level 1) 'governance puzzle-solving'. In contrast, problem-finding approaches emphasize (Level 2) 'governance Disruptor-Finding'; or (Level 3) 'Charting Macrostrategic Trajectories'. We apply this theoretical framework to contemporary governance debates around AI throughout our analysis to elaborate upon and to better illustrate our framework. We conclude with reflections on nuances, implications, and shortcomings of this long-term governance framework, offering a range of observations on intra-level failure modes, between-level complementarities, within-level path dependencies, and the categorical boundary conditions of governability ('Governance Goldilocks Zone'). We suggest that this framework can help underpin more holistic approaches for long-term strategymaking across diverse policy domains and contexts, and help cross the bridge between concrete policies on local solutions, and longer-term considerations of path-dependent societal trajectories to avert, or joint visions towards which global communities can or should be rallied.", "authors": ["Hin-Yan Liu", "Matthijs M Maas"], "title": "'Solving for X?' Towards a Problem-Finding Framework to Ground Long-Term Governance Strategies for Artificial Intelligence", "text": "Introduction Change is not a new feature in human affairs. Nor is the need for some forms of governance in response to ongoing, and at times challenging, developments. Yet over the last two centuries, and acutely within the last few decades, something has begun to change in change. Daniel Deudney has observed how \"the emergence and global spread of a modern civilization devoted to expanding scientific knowledge and developing new technologies has radically increased the rate, magnitude, complexity, novelty, and disruptiveness of change\"  (Deudney 2018, 223) . Accelerating climate collapse feedback loops are driven by economic and industrial systems intimately intertwined with fossil fuel energy extraction  (Di Muzio 2015) ; arsenals of nuclear weapons are held at hair-trigger launch alert, under a continuous state of functional 'thermonuclear monarchy'  (Scarry 2016) ; highly interlinked and efficient global supply chains, transport networks, and economies prove acutely vulnerable to pandemics; and emerging and potentially catastrophic risks from new technologies  (Bostrom and Cirkovic 2008; Bostrom 2013;  see also  Rayfuse 2017)  ensure that our global society is increasingly grasped in the throes of 'turbo change'  (Deudney 2018, 223; Rosa 2013) . Accordingly, an array of new technologies are reshaping the stakes and rapidity of the societal, strategic, and even geophysical and ecosystem changes which we must manage, both today and over the long-term. The renowned biologist E.O. Wilson once playfully quipped that \"The real problem of humanity is the following: we have paleolithic emotions; medieval institutions; and god-like technology\" (E. O.  Wilson 2009) . If that is so, can the diverse elements and processes constituting global governance evolve along with the changing stakes and nature of the challenge? A key question for any broader project examining how 'strategy' ought to be generally reconfigured within society's complex networks of communities, organizations, or systems in order to adopt meaningful and effective long-term perspectives (van  Assche et al. 2020 ) concerns how governance needs to be reoriented towards the long-term perspectives that flow from these urgent and unpredictable technological challenges. Moreover, the question may prove particularly salient and critical in the technological context. In the coming decades, existing governance instruments may well face a historical 'inflection point' in their relation to this series of emerging and 'converging' technologies  (Pauwels 2019) . Our capacity to collectively manage our relationship to such technologies, such as artificial intelligence (AI), may prove to be both a key objective for, and a key stress (or litmus) test of, our governance regimes. How or why might we have to adapt or replace governance systems in the face of the 'technology tsunami' (Danzig 2017)? Of course, this is not to say that we cannot influence, channel, or resist such technological shocks-indeed, contrary to frequent depictions of the regulatory 'pacing problem' (e.g.  Marchant 2011; Hagemann, Huddleston, and Thierer 2018) , it should be kept in mind that governance structures also provide the landscape that shapes and guides technological development, such that we do not need to resign ourselves to reactively responding to shocks  (Crootof and Ard 2021, 12)  but can also seek to shape paths of development in advance. Faced with a 'tsunami', we need not just worry around how to manage or weather it, but also about how we might channel or even surf it. Nonetheless, how might we go about making such governance changes? How can we ensure that we do not find ourselves \"addressing twenty-first century challenges with twentieth-century laws\"  (Jensen 2014, 253)  and why might such extant approaches prove problematic in the first place? We face an increasingly complex historical challenge of evolving our governance systems-or, if necessary, developing new ones-to be up to the task of responsibly managing this array of potent challenges, and to do so not just for the near-, but also for the long-term. This paper questions the premise that the extant regulatory and governance orders will remain fully intact and functional throughout the turbulent period ushered in by complex, 'transversal' emerging issues (see  Morin et al. 2019) . It does so taking its departure from one of these trend-drivers-artificial intelligence. We argue that AI and its applications will outfox contemporary regulatory and governance orders. Specifically, AI holds the potential to reveal, enable, or drive society towards new regulatory and governance equilibria, and is thus a prime candidate for triggering regulatory disruption. Again, this is not to say that we have to be merely reactive: we can and must influence and shape patterns of regulatory disruption; however, to do so, we have to be aware of the dynamics, pathways, and vectors of change  (Liu et al. 2020) . What does this reveal? We argue that an underlying problem lies in the fact that many contemporary regulatory and governance orders (and their respective academic fields) continue to implicitly adopt or manifest problem-solving orientations. In other words, regulation and governance are endeavours that respond to recognised, defined, and compartmentalised problems. Overlooked in such an orientation, however, are the inherent and contingent limitations that needlessly straight-jacket the endeavour. By prematurely settling upon an overly and necessarily narrow, and somewhat contingent, range of issues to be addressed these problem-solving efforts may increase the probability of a turbulent or even catastrophic transition towards an AI-permeated world. Accordingly, our aim in this paper is to shed light on the potential of, and prospects for a problem-finding orientation for formulating and anchoring long-term strategy assemblages for the governance of AI-strategies that take stock of the ways in which the processes, instruments, assumptions and even aims of existing governance orders will be shaped and disrupted by AI. Thus our overarching aim in this paper is to chart the underexplored terrain beyond the boundaries of current problem-solving regulatory and governance debates. By actively and systematically searching the potential problem-space opened up by AI, this paper aims to increase the confidence that efforts geared towards addressing presently-identified problems triggered by AI are relevant over the long-term, and do constitute the significant questions that need to be addressed, as well as to identify new potential problems that AI might raise for regulation and governance, which have been underexplored to date. Accordingly, this paper proceeds as follows. We first ground a rationale for a problemfinding governance framework. We sketch the policy problems created by AI technology, and provide 5 reasons on the basis of which a problem-solving governance approach to these challenges fails or falls short. We then extend the argument by arguing that problem-solving governance approaches are ill-equipped to engage with 'wicked problems' created by AI. We conclude the rationale by grounding the epistemic and scientific validity of a complementary problem-finding AI governance (research) program. Secondly, we then set out our framework of AI governance across four strategic 'levels' which correspond to AI governance approaches that take for granted different parameters, contraints and failure-modes of scholarship, which variably approaches AI issues from a perspective of (0) 'business-as-usual'; (1) 'puzzle-solving'; (2) 'governance disruptor-finding'; or (3) 'charting macrostrategic trajectories'. Thirdly, we provide reflections on nuances, implications, and shortcomings of this long-term governance framework. We discuss how to best coordinate AI governance strategies amongst and between these complementary four levels, in order to avoid the failure modes that might occur when restricting analysis or policy to any one of these levels. We assess how 'problem-finding' scholarship can be applied in problem-solving roles. We theorize how these four levels may constitute a 'Goldilocks Zone for Governance' which, in a larger perspective, highlights the assumptions and boundaries of any governance system; and we discuss how we should navigate the linkages between strategies and overarching narratives. We conclude by suggesting that these lessons are critical in structuring governance responses-to AI; and to other vectors of change-that are meaningful and resilient into the long term. \n From Problem-Solving to Problem-Finding: a Rationale In the first place, it is important to come to understand the plausible shortfalls of problemsolving approaches. But to do so, we must understand what characterizes these approaches; why they fall short of providing long-term governance responses to AI; and why an alternative and complementary problem-finding paradigm may be both pragmatically effective, as well as epistemically and scientifically warranted. 2016; Calo 2017;  Guihot, Matthew, and Suzor 2017; Dafoe 2018; Turner 2018; Hoffmann-Riem 2020 ). Yet, the identified literature on discrete policy issues adopts a general problem-solving orientation. 2 Under this paradigm, policy scholarship compartmentalises AI's immense potential problem-space into specific problems-from the safety of autonomous vehicles to misuse of 'deep fakes', and from new criminal uses to 'killer robots'-that are often defined or segmented according to disciplinary perspectives  (Petit 2017; Crootof and Ard 2021) . There are several interrelated limitations that flow from attempts to solve pre-packaged policy problems which, taken together, suggest that this is a necessary but insufficent strategy for long-term governance, especially in the context of societally-disruptive technologies such as AI. The problem-solving orientation takes departure from issues that are discursively recognized or constructed (by policymakers as well as by scholars in a field) as a 'clear and present danger': that is, they often focus on issues that have given rise to vivid (if anecdotal) challenges such that they are discernible (and discerned) by public and policymakers; which are accordingly highlighted (and, in some cases, 'securitized' (Stritzel 2014)) within government strategies; and which are validated, accepted, operationalized, and funded as legitimate and well-bounded research topics within existing disciplinary fields. In this way, the problemsolving orientation is in a sense a demand-driven (rather than supply-driven) approach to research program orientation and policy formulation. At its limits, it constitutes a 'local search' for the next urgent, externally-validated and -packaged problem that can be fitted to analysis by the discipline's available methodological or conceptual tools, rather than a 'global search' across the broader landscape of possible problems (let alone each problem's landscape of possible perspectives). To an extent, this orientation is understandable. Scholars only have so much time, so much attention. For all that interdisciplinarity is often praised, it can often appear a risky, difficult affair  (Fish 1989 ). 2 As always, no term is without its difficulties. There are some conceptual challenges with designating existing scholarship or governance approaches as 'problem solving'. For one, this (a) does not (directly) interrogate existing governance actors' privileged ability to determine which are and are not 'problems' to be 'solved' (e.g. certain AI-enabled technologies such as DeepFakes in some sense shift around (epistemic) power; that may make them a 'problem' for governments; but from the perspective of certain political actors they more obviously present themselves as an opportunity. However, since these latter actors do not conventionally 'do governance', their perspective is not easily presented under this framework). In the second place, and relatedly, (b) the term 'problem-solving' appears to at least implicitly pre-suppose or imply that the pre-existing state of affairs was broadly 'unproblematic' before it was rudely disrupted by a new technological 'problem' which upset that functional state of affairs-with the implication that governance might again be broadly unproblematic once that problem has been adequately solved. That would obviously be a highly normative and debatable assumption, which many (emerging) governance actors would challenge. However, it is a shortfall that primarily afflicts 'problemsolving' governance approaches-and we would emphasize that problem-finding approaches can bring in a wider range of actors, agendas, and values. We thank Henrik Palmer Olsen for prompting this reflection. Yet while perhaps appealing and even understandable from the perspective of a preexisting academic division of expertise and labour, however, five problems adhere to a problemsolving approach. In brief, these are (1) the unreflexive over-reliance on narrow metaphors or analogies in the face of multifarious, complex phenomena; (2) the tautological manner in which existing governance assemblages surface primarily those problems that are already (or most) legible to them; (3) the constrained path-dependency; (4) inability to clearly distinguish between symptomatic and root-cause problems; (5) the promotion of a false sense of security through a rapid delineation of the problem-space that prematurely closes-off large parts of the governance solution-space. \n Unreflexive reliance on over-narrow metaphors In the first place, in many areas, and especially where it comes to new technologies, problem-solving perspectives often, implicitly or explicitly, rely on drawing metaphors or analogies in order to fit a new problem or topic within the pre-existing boxes. Yet, as any given analogy or metaphor used to describe new technologies by necessity is incomplete  (Crootof 2018; Balkin 2015) , problem-solving perspectives that invoke them are necessarily partial and incomplete, and potentially even misleading, as they express only certain disconnected aspects of the overall challenge posed by the technology. This insufficiency of a problem-solving approach is reflected in the ancient parable of the 'blind men and the elephant' which warns against concluding that elephants are like pythons, simply because one has exhaustively touched one elephant's trunk  (Saxe 1872) . One can see this same risk in the context of the multifarious debates round AI, which often explicitly or implicitly trade on highly different metaphors or understandings  (Brockman 2019; Parson et al. 2019) . Amongst scholars writing on AI governance, many take comprehensively different views, not just on the individual questions of what AI is, how AI operates, how we treat or relate to AI technology, how we use AI in society, or what (unintended) impacts AI will have on our society down the road-but indeed take different positions on which of these above frames is even the relevant one for the policy question at hand. For instance, does the regulation of social robots turn on questions of whether the artefact is 'software' or a 'cyber-physical system' capable of physical damage (Calo 2015; 2017)? On questions of how humans will relate to humanoid systems  (Darling 2012 )? Or on questions of how they might provide new attack surfaces for criminal (mis)use  (Brundage et al. 2018; King et al. 2019; Hayward and Maas 2020; Caldwell et al. 2020 )? Each of these local perspectives provides pieces of the puzzle that can be valuable and legitimate, but often there is no reflection on how the specific narrow perspective-not just on these questions, but even on which of these questions is the appropriate one to ask-is but one of various ways to frame and understand the salient features or key impacts of the technology (See also  Cave and Dihal 2019; Cave, Dihal, and Dillon 2020) . Indeed, the 'blind man and the elephant' metaphor may itself be overly restrictive: by extending or reframing the parable, we can see that a problem-solving approach risks misapprehending the problem in many more ways. After all, the problem is not only that an overtly narrow view may result in us (1) believing that an elephant is like a snake, on the basis of studying its trunk (i.e. mischaracterizing problem X on the basis of its local facet). Rather, the very focus on 'the elephant in isolation' occludes a range of other potential issues. To stay within the framing of the parable; 3 our narrow focus on the 'elephant' in front of us may lead us to (2) fail to notice the tiger that is crouching up behind us (missing out on urgent 'out-of-context' problems or development Y); (3) fail to highlight how this elephant came to potentially be brought out of its habitat and social context and presented to us, and how it came to be constructed and designated as a salient object in need of closer examination and apprehension (missing out on the socio-political roots shaping the research agenda on X); (4) fail to apprehend how, rather than apprehending the 'essence' of an individual elephant, it is also relevant to understand how that elephant fits within a broader ecosystem-or what the health of the elephant tells us about the broader ecosystem (missing out on the interrelation of problem X with seemingly 'unrelated' entities, problems, or developments); (5) base our understanding of the elephant on a momentary snapshot, rather than a longitudinal study of an elephant's developments over the course of its lifecycle (taking an overtly static rather than dynamic view); (6) take for granted that an elephant can be appropriately and sufficiently apprehended through touch-and appropriately and sufficiently described at the level of human-scale featuresrather than considering how additional (and different) insights into the elephant could be gleaned from using scientific tools that enable us to study it at different levels of analysis, from fundamental physics, cellular biology, or etiology (taking for granted a narrow methodological toolkit to studying problem X at a narrow level of granularity). This illustrates the epistemic limits and shortfalls of problem-solving approaches to understanding and governing new technologies. These approaches frequently focus on: the here and now; the direct harm; the last person to touch the technology prior to an accident; or the degree to which any solution can be found which mitigates the direct fallout; or redresses the directly (legally-recognized) injury. For example, legal problems that flow from AI applications might revolve around questions of liability for entities occupying a liminal position between 3 There is, of course, an irony in critiquing the prominence of (legal) metaphors within problem-solving AI policy research, through the use of this parable. However, we argue that in this instance our extension of the 'blind men and the elephant' metaphor works to highlight the contingency of taking for granted a classic or narrow interpretation of a certain parable (and the diversity of alternate and even contradictory problem formulations that could be derived or accomodated even within one metaphor-paradigm). agent and object, which befuddle existing legal doctrine  (Liu 2012; see generally Turner 2018) . Accommodating this challenge may be necessary for the law to regain coherence, but does little to address competing plausible models of AI applications such as a networks, or a systems, approach to understanding AI  (Liu 2019a; Ekelhof 2019) . This approach or mindset suggests that legal problems are just that, 'problems' that can be identified and interrogated with legal tools, but that this process itself skews the form that the underlying challenges are deemed to take, and the types of (legal) responses that are legitimate to propose in their wake. \n Tautological responses Second, a problem-solving approach ensures that any problems that are surfaced by examination are often tautological but somewhat arbitrary in relation to the significance of their potential sociotechnical impact. AI applications might generate legal problems because those are the problems that AI generates for the existing law-more than for the underlying society. In practice, this might often highlight legally or philosophically 'unusual' problems which are clearly and obviously blurring our existing 'symbolic order'-that is, the fundamental distinctions and categories which a general society, broad governance network, or specific legal order relies upon to draw boundaries and understand the relevant reality (see Douglas 1966)over less 'legible' but more 'socially disruptive' developments that provide scholars with less 'satisfying' analytical or legal puzzles. One might consider for instance the way by which discourses on self-driving cars have become dominated, for better or worse, by their linkage to the high-profile 'trolley problem' thought experiment from ethics, a development many have begun to critique as having led such governance debates down a dead-end  (De Freitas, Anthony, and Alvarez 2019; Jaques 2019; Cunneen et al. 2020; Himmelreich 2018; Bauman et al. 2014; Wolkenstein 2018) . A key problem here is that within problem-solving approaches there is often no reliable or unambiguous appraisal process through which to gauge whether or why those problems that are generated or focused on, are likely to be the significant and persistent problems into the future, and not just solely 'legally interesting'. If problem-solving approaches get hooked on temporary, doctrinal questions which solve unrepresentative 'edge' problems (especially problems which might, in due course, end up being 'dissolved' anyway as a result of technological developments or 'out-of-domain' social or regulatory changes), however, this bodes ill for their viability or sufficiency into the long-term. \n Limiting path-dependencies and 'brittleness' Third, there are clear path dependencies inherent to problem-solving approaches. The problems that arise are frequently understood as variations on previously validated (if not necessarily 'solved') problems, and responses often deploy analogies to those that were invoked in past iterations  (Crootof 2019b; Mandel 2017) . This sets strong constraints upon the trajectory of problem-solving efforts, ensuring diminishing-or even 'negative'-returns when these efforts are brought to bear on genuinely fresh challenges or unseen cases. Curiously, in this way a legal system is not unlike a contemporary machine learning algorithm itself: it works entirely and solely from its training data, and so can be structurally biased (depending on what elements or cases are under-or over-represented in its 'training set'), or brittle when it encounters genuinely unprecedented 'edge cases' that combine certain strange features of past cases, or bring in genuinely unprecedented  ('out-of-training-distribution')  features-as in the old legal maxim that 'hard cases make bad law'  (Davis and Stark 2001) . Moreover, just as machine learning systems have proven susceptible to so-called 'adversarial input'  (Goodfellow, Shlens, and Szegedy 2014; Hendrycks et al. 2019) , a legal or political system that hems closely (and predictably) to certain past categories may be adversarially 'spoofed' by certain, well-resourced actors who can construct or structure their 'input' in ways that exploit these features of the system in order to produce 'output' to their liking. Such strategies have been well-docmented in the legal context, consider US companies invoking 'free speech' as category to secure unrestricted campaign finance laws (Citizens United v Federal Election Commission 2010). In more political arenas, consider the ways in which certain actors have 'securitized' issues such as migration, in order to re-deploy and re-direct the age-old state logic of 'security' towards certain desired political ends  (Huysmans 2000) . Beyond these general shortfalls, what is the problem with such problem-solving path dependencies? Given the breadth and depth of challenges envisaged by the widespread introduction of AI into society, it is arguable that a doctrinal solution to the liability questions raised by AI will at best produce marginal returns. This is especially since retention or reaffirmation of the contemporary legal paradigm many instead hinder attempts to a fuller exploitation of the benefits that the technology would otherwise be able to offer. That is not to say, of course, that this 'problem-solving path-dependency' is the only such rigidity limiting governance. Indeed, as scholars in the field of Evolutionary Governance Theory have demonstrated, there are always various other 'dependencies'-from path dependencies to interdependencies, and from goal dependencies to material dependencies-which bound governance actors' ability to strategize far beyond their found governance path (van  Assche et al. 2020, 7) . Such path dependencies can therefore only be adequately understood together with various interdependencies and goal dependencies (the reflexive and at times self-fulfilling impact of visions for the future on current governance), as well as material dependencies such as the technologies entrenched in the system's discourse, policy, organization, or regimes of cooperation and coordination. Thus, while many elements of governance are contingent and could in principle be reconfigured-even radically so-  van Assche, Beunen & Duineveld (2014, 5 ) also note that \"[o]ne cannot jump from each branch in the evolutionary tree to each imaginable other branch\". Nonetheless, in some cases the path-dependency of problem-solving approaches appears particularly constrained, and sufficiently contingent, that further exploration could be fruitful. In particular, the insight that the existing regulatory order is but one possible 'governance strategy' equilibrium, does not mean that the 'leap' to other such equilibria will be easy (indeed, if it were, the present state would not be so much of an attractor at all, and would likely have 'decayed' to one). Rather, the point is that while this present equilibrium may have proven proficient at managing sociotechnical changes thus far, it may no longer be appropriate nor sufficient in securing long-term interests-or realizing long-term perspectives-in the 'turbochange' era of AI. \n Inability to differentiate between root problems and surface challenges Fourth, a different objection to problem-solving, related to the first, turns on the fact that it is 'centrality-blind'-that is, it does not easily (or at best only retrospectively) identify and prioritize governance bottlenecks or cross-cutting themes. Presently-identified problems may be differentiated into symptomatic and root-cause problems, but the problem-solving orientation does not systematically draw this distinction. Thus, much problem-solving work becomes squandered in ameliorating the symptoms rather than the root-causes of a problem: questions of AI authorship arise in the context of intellectual property law  (Kaminski 2017) ; questions of responsibility are asked in the the context of autonomous weapons systems (AWS)  (Jain 2016) ; and questions of liability are raised in the context of accidents caused by autonomous vehicles  (Schellekens 2015; Boeglin 2015) . While legitimate legal problems, these are all symptoms of the underlying uncertainty over the liminal status of AI applications, which is the root-cause legal problem. The point here is that there are neither appraisal processes, nor confidence more generally, as to whether the policy problems posed are the central or core problems, and not merely their manifestation in certain areas. This ensures that there may also be a degree of disconnection amongst problem-solving governance approaches, such that there is potentially much double work or redundancy by virtue of an uncertainty as to the \"level\" of the problems that are tackled and what parallel efforts are underway. \n Promoting a false sense of security Fifth and finally, there is the false sense of security inherent within the problem-solving approach. Most fundamentally, it implicitly supports a world-view whereby we face only 'problems', when in fact many of the challenges we face today might be better understood and approached, at least upon first encounter, as 'mysteries'  (Chomsky 1976) . Through emphasising that new challenges already \"appear to be within the reach of approaches and concepts that are moderately well understood\"  (Chomsky 1976, 281) , problem-solving approaches suggests that while we may not yet have the solutions, we have an idea as to what those might look like and how we might get there. In other words, we may not have the solution at hand, but we surely have already delineated the space within which we are sure it can (and therefore ensure it must) be found. This can be juxtaposed against 'mysteries', which \"remain as obscure to us today as when they were originally formulated\"  (Chomsky 1976, 281) . In this context, problem-solving suggests greater understanding and mastery of challenges than might otherwise be justified. This becomes an issue when epistemic certainty is not so warranted. More superficially are considerations around the appropriateness, adequacy, and sufficiency of problem-solving approaches. What we mean here is that, even where policy problems are solved, there is not necessarily a connection between this solution and the relevance of the solution to the continuing and future stability of the regulatory and governance orders. This relates back to the absence of processes for assessing the relevance of solving particular problems to the underlying challenges at the source of those problems. A different way of articulating the objections to an exclusively problem-solving approach can be captured concisely in the context of \"wicked problems\" which may serve as a bridge to our proposed emphasis upon problem-finding. \n Problem-Solving governance is an inadequate response to \"wicked\" AI problems Introduced in an influential 1973 article, Rittel and Webber introduced the concept of 'wicked problems'-which they contrasted to 'tame', soluble problems in some scientific fields such as mathematics, chemistry or engineering-as problems that are difficult or impossible to solve because our requirements for a solution are often incomplete, changing, or in tension (such as in situations where there are no undisputable public goods); because policy problems cannot be definitively described; or because the effort to solve one aspect of a wicked problem may reveal or create other problems. By representing \"wicked problems\" as an array of complex, openended, and intractable problems, 4 Brian Head (  2008 ) developed an alternative rubric under which many governance challenges, including those posed by AI, can be productively examined.  5  To be sure, it should be cautioned that merely labelling a problem as \"wicked\" may not necessarily assist in solving it-in the same way that labelling a system as 'complex', or a phenomenon as 'emergent' by itself does not assist in actually understanding any of its dynamics or behaviour, and rather should be understood better as a signpost highlighting these as topics in urgent need of further investigation. However, the benefits of the 'wicked problems' framework are less about providing a clear solution rubric, 6 and aims more to focus attention 'on the understandings that have shaped problem-identification and thus the frames for generating problem-solutions  ' (2008, 106)  in the first place. The interrogation of upstream factors impinging upon problem-identification may serve as a response to the missing processes for identifying, ascertaining and evaluating the centrality or significance of downstream problems impugned in the section above. Indeed, in conceptualising wicked problems as the convergence of uncertainty, complexity and value-divergence,  Head (2008, 106)  suggests that failures to adequately respond to wicked problems may be due to the fact that: -The \"problems\" are poorly identified and scoped; -The problems themselves may be constantly changing. -Solutions may be addressing the symptoms instead of the underlying causes. -People may disagree so strongly that many solution-options are unworkable. -The knowledge base required for effective implementation may be weak, fragmented or contested. -Some solutions may depend on achieving major shifts in attitudes and behaviours  […]  but there are insufficient incentives or points of leverage to ensure that such shifts are actualised.  (Head 2008, 106)  A common denominator to these failure modes inheres within poor mappings of the potential problem-space. Not only might the problem formulation itself be mistaken or misaligned, but the 'true' nature of the problem might not be one which is directly connected to the external real-world challenges that are most readily perceived as problems. Instead, as this list suggests, the root of the challenge might instead lie in social dynamic, economic, and political considerations. Yet, since subsequent problem-solving endeavours are sculpted in large part by the outcome of problem-identification and -definition processes, overly narrow or static problem initial formulations may lead to inadequate or brittle responses that fall short of addressing such problems. In slogan form: solving individual pieces of the puzzle provides little or no protection against against the larger threat posed by wicked problems. Such local responses may at times be orthogonal to addressing the underlying problem-and at worst may be damaging, either in their direct results, or because they forestall deeper responses. Scholars have in recent years increasingly highlighted how legacy institutions and policy toolkits derived from the management of 'complicated' problems may prove ill-suited for the management of 'complex' problems  (Kreienkamp and Pegram 2020; Morin et al. 2019) . Applied to the challenges posed by AI governance, the wicked problem framework suggests that problem-solving approaches by themselves may not be sufficient  (Rittel and Webber 1973, 160-67 ; see also  Gruetzemacher 2018) . This is because AI itself not only poses complex, wicked problems, but because the application of this technology can also be disruptive to governance responses themselves, as it can alter the key parameters of a governance path; the interests of different actors in the network; the legal principles and processes; the regulators' and regulatees' values; or the regulatory modalities at play in the policymaking portfolio. To address the core characteristics of such changes, persistent (continuous) and pervasive (wide-ranging) problem-finding approaches should be articulated. As such, we will shortly seek to elaborate upon and structure one such problem-finding approach across four strategic levels below, as a means to give granularity to the multifaceted wicked problems generated by the infusion of AI into society. Before doing so, however, we will first reflect on the epistemic and scientific foundations of a problem-finding approach. \n The Epistemic Validity of a Problem-Finding Approach to Long-Term AI Governance The problem-solving/problem-finding distinction has parallels in the division between normal science as 'puzzle-solving' and revolutionary science as 'paradigm-shifting'  (Kuhn 1970) ; in the earlier-alluded-to separation of ignorance into 'problems' and 'mysteries'  (Chomsky 1976) , and in the scientific creativity of research communities variably taking the form of incremental and secure 'cold searches' or far-reaching yet uncertain 'hot searches'  (Currie 2019) . Indeed, the distinct psychological learning strategies displayed throughout human childhood have been likened to foundational AI research paradigms, which also balance algorithms between paired phases of 'exploration'-wandering far; gathering information, and phases of 'exploitation'acting on the information gathered to the greatest effect  (Gopnik et al. 2017, 7893; in: Currie 2019, 5) . One might also compare the idea, in algorithmic evolutionary optimization, of 'simulated annealing'  (Derman 2018)  where, by introducing volatility and 'shaking up' its search, an algorithm can find its way to better solutions outside the incremental search-space. The fact that a self-similar dyad can be found in philosophy, logic, evolution, computing, and even human psychology, is indicative of not only its scientific validity, but also its critical role in governance strategies. Analogously, while problem-solving work ('puzzle-solving; 'addressing problems'; 'cold searches'; 'exploitation') is clearly the bread and butter of many scientific research communities, open-ended problem-finding research approaches ('paradigmshifting'; 'pursuing mysteries'; 'hot searches'; 'exploration') are and ought to be a natural and critical complement. This is for four reasons. First, such problem-finding work may itself produce foundational academic insights (for example about the interrelation of law and governance priorities across different scales or levels). Indeed, meta-science research has suggested that, even as exploratory research papers face a scientific 'bias against novelty', they are more likely to become amongst the top 1% highly cited papers, and to inspire follow-on highly cited research across many disciplines in the long run (Wang, Veugelers, and Stephan 2017). In the second place, open-ended, problem-finding research agendas can also work in a reflexive equilibrium with incremental policy-solving research agendas, by providing broaderscope insight into how individual policy areas operate. In this way, such research agendas can address some of the 'institutional mismatch' to addressing a set of 'transversal' issues (including, but not limited to, AI) which drive externalities beyond or across familiar problem categories and domains  (Morin et al. 2019, 2-3) . Thirdly, and importantly, problem-finding approaches can at times reveal new 'crucial considerations'  (Bostrom 2014b ): new arguments or propositions which, if true, reveal the need for not just small course corrections, but major overhauls in direction or priorities within our governance clusters and actions. For instance, research agendas into avenues to 'automatically detect' media forged by 'DeepFakes' are dependent on a crucial assumption that such detection measures can always be developed relatively quickly, and that this strategy is viable more or less indefinitely. If, however, the 'offence-defence balance' of AI research in this area is such that (the dissemination of) progress in these techniques aids attackers more than defenders (Shevlane and Dafoe 2020); or if, at the limit, progress in DeepFakes will mean that within a few short years detection will simply be functionally impossible  (Engler 2019) , this narrow research agenda would be overturned. A problem-finding orientation can therefore frame a lodestar to aim towards, rather than merely identifying obstacles present and apparent in our current trajectory. Finally, problem-finding approaches enable us to forestall or short-circuit Collingridge's 'dilemma of control', which holds that \"[w]hen change is easy, the need for it cannot be foreseen; when change is apparent, change has become expensive, difficult, and time consuming\"  (Collingridge 1980, 11) . In that sense, problem-finding and problem-solving approaches fall respectively into each side of this dilemma. In this context, the power problem inherent in attempting change when it is 'expensive, difficult, and time consuming' reveals the futility of foregrounding problem-solving approaches in relation to long-term perspectives. Problem-solving approaches, somewhat obviously, become operational precisely where the desire or need for change confronts the inertia of a complex and interconnected world that frustrate such efforts. Long-term perspectives, however, reside in the information problem aspect of the dilemma: problem-finding approaches thus attempt to explore the potential problem-space while 'change is easy' and course corrections and changes are still possible. Yet, the dilemma nature of the challenge suggests that efforts of both of its sides are necessary, and the long-term perspective merely suggests that a recalibration of these efforts needs to be effected. \n A Proposed Problem-Finding Framework Across Four Strategic Levels We propose a theoretical framework or typology for mapping the long-term strategic landscape which we apply to the seemingly distinct cluster of policy domains surrounding AI.  7  In our understanding, we subdivide governance responses to AI's problems into four strategic 'levels'. 8 Levels 0 & 1 concern default, 'problem-solving' responses to governance: 9 (0) 'business-as-usual' governance demands the simple application of existing governance structures to the new problems; whereas (1) 'governance puzzle-solving' admits at least a certain need to reconceptualise, stretch or reconfigure existing legal doctrinal categories or governance concepts, in order to address those problems. Conversely, Levels 2 & 3 concern systemic challenges that currently lie beyond the boundaries of contemporary legal frameworks. They 7 To be clear, while we use AI as a focal technology, the general argument is applicable to the way longterm governance would engage with any new technology (cf.  Liu et al. 2020) . See also the critique, of legal scholars' problem-solving responses to technology, in  (Tranter 2011) . We thank Roger Brownsword for prompting this clarification.  8  We should make a few important caveats regarding our usage of the term 'levels' in this typology. In the first place, by referring to 'levels', we do not mean to imply a ranking of importance: another way to read or refer to them would be 'types' of governance strategies-although we prefer to retain the imperfect term 'levels' for the simple reason that it emphasises more (in a way that 'types' or 'categories' does not do) how the different varieties of governance strategies are not mutually exclusive and isolated, but interlinked components of a larger governance system. In the second place, and pragmatically, this terminology should not be confused with the existing scholarship on 'multi-level governance' (See  Bache and Flinders 2004; Zürn 2012) , which explores how policymaking power and authority are distributed both throughout the hierarchies of government, as well as horizontally towards other non-state actors. In our usage, by contrast, 'levels' correspond not to the 'level' (that is, the actor) at which a given strategy is rooted or situated, but rather to the 'level' at which a strategy seeks to exert leverage (that is, its scope, or the degree to which it variably takes for granted, or interrogates the problem formulation, parameters, and larger 'outside context'. In that sense, these 'levels' are ordinal ('Level 3' takes a broader scope than 'Level 2', which takes a broader scope than 'Level 1'), but do not necessarily connote ranked or hierarchical levels of influence (e.g. 'Level 3' strategy are not by definition or always upstream from 'lower' levels, even if they might offer greater possibilities for creation and consideration of long-term perspectives); rather these levels co-constitute one another. 9 Again, to avoid terminological confusion, it should be kept in mind that these Levels 0 and 1 are 'problem-solving'. We nevertheless include them as a key baseline component in our integrated framework which (encompassing the directly problem-finding strategies at Levels 2 and 3) we collectively refer to as the 'problem-finding' framework for AI governance. This is because we consider an integrated governance framework capable of carrying out 'problem-finding' governance as one that is capable of encompassing and complementing existing 'problem-solving' approaches, not of wholly replacing them. involve scholarship aimed at (2) finding potential 'disruptors' of core governance assumptions; or at (3) charting macrostrategic trajectories & destinations. Together, this creates a comparative framework within which to compare the problem orientations, contribution orientations, and limits of these distinct analytical and governance perspectives (see Table  1 ). Accordingly, such research lines call upon problem-finding approaches that convert, contextualize, prioritize, and structure the unknown problem-space into clusters or research agendas of problems that could then be addressed at Levels 0 & 1.  \n Finding • Finding AI developments that may 'disrupt' assumptions of or conditions for the governance system, such as (1) changes to the problem portfolio; (2) to regulators' goals; (3) to governance tools, (4) to societal values The 'default' strategy involves an adherence to 'normal' governance processes, norms, structures and concepts, however constructed. While it does not deny the emergence of certain problems or challenges, it can be slow to recognize them-and even when it does, it will deny their fundamental 'newness'. This governance level is therefore rigorously focused on solving problems within (or by extension or application of) the existing governance system. In this sense, Level 0 governance is narrowly problem-solving: it recognizes the local 'problem', but requires that any solutions are chosen from the existing set of solutions made available. It is reactive, and seeks to re-establish or re-balance the old 'governance equilibria'. For instance, AI policy work at this level seeks to accommodate the issues raised by AI within the extant legal frameworks, arguing that legal doctrinal changes to accommodate the technologies are not necessary-in the terminology of cyberlaw, that distinct, domain-specific legal innovations would be as superfluous as articulating a comprehensive 'law of the horse'  (Easterbrook 1996) . This position holds that all appropriate 'metaphors' to be encoded explicitly in law or implicitly in broader governance discourse can be found in past technologies  (Mandel 2017) . Because these are easy to understand, much of the problem-solving work done today in distinct sub-fields on AI policy is situated at Level 0. For example, it is manifested in attempts to simply extend the existing legal framework and principles of international humanitarian law (IHL), to cover any and all issues that are raised by AWS  (Anderson, Reisner, and Waxman 2014; Schmitt 2013) . While it might be easy or tempting to dismiss such work out of hand, such a response would likely be neither fair nor warranted. Indeed, to be clear, a 'business-as-usual' response is not an obviously incorrect orientation in many cases (though it may be an insufficient one at a systemic level, across all cases). It relies on a certain degree of conceptual laxity-the willingness to implicitly (and at times perhaps unreflexively) stretch and re-interpret existing governance categories, just so that we can ensure their easy application to new cases. Such an approach avoids continuous, and costly, transaction costs of figuring out whether or not every new case or technology is new. Indeed, to an extent, the 'business-as-usual' heuristic is critical to governance: if we lacked one, we might easily find ourselves paralyzed: we would risk 'overfitting' (in the machine-learning sense) on our past governance experiences-leading one to argue, for instance, that a car crash involving a neon-painted car could not possibly be covered under existing laws, since no similarly-coloured car was ever involved in a car crash in the given jurisdiction, and this case was therefore clearly without precedent. Lyria Bennett Moses likewise argues that \"[m]ost of the time, a law predating technological change will apply in the new circumstances without any confusion. For example, traffic rules continue to apply to cars with electric windows, and no sane person would seek to challenge these laws as inapplicable or write an article calling for the law to be clarified.\"  (Bennett Moses 2007, 596) . Moreover, an approach of implicit 'reinterpretation' and extension of existing governance categories may have some limited (temporary) merit even in the face of actual underlying real Nonetheless, while these Level 0 responses may be somewhat understandable, and even necessary in some cases, it is likely that both their drawbacks and inadequacy are underestimated. In the first case, such moves can be brittle: when pushed too far, attempts to capture obviously (or to the public, apparently) revolutionary technologies under age-old governance approaches may fail the 'laugh test', with the result that 'the law runs out'  (Crootof 2019b, 17) , threatening the credibility of governance efforts within that specific domain, as well as, at length, the more general legitimacy of the governance system that produced or allowed it (as critics can add another anecdote to a list of 'absurdities' produced by the extant system). In other cases, and especially in global governance contexts, certain actors may simply not accept an extension of 'business-as-usual' to new technologies, leading to new explicit or implicit contestation over specific norms, entire strategies, or certain actors' legitimacy or authority to steward those policies (see also  Zürn 2018) . More fundamentally, however, because Level 0 approaches do not seek to (deeply) track or understand the nature of the ongoing changes (beyond the bare minimum necessary to address local symptomatic problems), they cannot be expected to address these in the long run. Instead, such governance structures may implicitly build up a 'technology debt', becoming hollowed out by changing practices, and ultimately being rendered into obsolete 'jurisprudential space junk' (Crootof 2019a) as a result of changing practices. Ultimately, this approach also cannot reckon with possibly new, unanticipated features or area cross-overs, and therefore cannot provide the foundation for a problem-finding approach. In sum, Level 0 response might be necessary for any governance system, yet they are not (and potentially cannot be) sufficient. At their best (or at most), they might serve as 'holding actions' that prevent near-term problems from 'unwrenching' or distracting society so badly or so frequently that it cannot even reckon with long-term trends. But by itself it is certainly not able to adopt a creative problem-finding approach capable of providing a foundation for long-term governance, because in many cases they 'deny' that there is any deeper problem (beyond the surface disruption to be stilled). In doing so, it risks building up a 'problem debt'. \n Level 1 -Governance Puzzle-Solving Closely related to Level 0 governance, but somewhat distinct, are perspectives that approach specific new problems as 'governance puzzles' to be solved. Work at this level still narrowly or primarily emphasizes the importance of fixing the direct problem at hand-that is, it takes for granted a narrow rationale-but it opens up for the possibility that doing so may require innovation and (even far-reaching) change in the regulatory tools or governance processes. As such, in contrast to Level 0, Level 1 governance is at least broadly problem-solving: it admits, at least in principle, that solving the direct problem in front of us may require innovations and changes in the set of available solutions, however, it does still narrowly seek to maintain or restore the existing status ex ante. In an AI context, this could be found in debates about discovering ways to re-conceptualize privacy (in the face of new AI challenges) in order to functionally restore the equivalent of past privacy protections to citizens. It may also extend to proposed innovations in governance approaches-such as the posited use of technologies such as AI systems as 'privacy protector'  (Els 2017; Gasser 2016 ) in order to secure established goals or rights. As noted, Level 1 analysis is largely similar to Level 0, but it admits to a certain degree of novelty in the challenge, and accordingly to a flexibility in governance response, while anchored upon an allegiance or commitment to the existing status: the end-product is the same, even if the route is different. \n Level 2 -Governance Disruptor-Finding In contrast, work at Level 2 begins to realize a problem-finding orientation. Its focus is less on the narrow problems as they are given, presented, or highlighted within pre-existing disciplinary boundaries, and instead seeks to explore underlying patterns, linkages or underilluminated problem clusters. Recognizing that 'law' is only one 'regulatory modality'  (Lessig 1998; 1999)   \n Level 3 -Charting Macrostrategic Trajectories & Destinations Finally, the third level reframes the AI policy-debates at Levels 0-2 through the lens of foundational axiological and 'macro-strategic'  (Bostrom 2016)   towards global catastrophes or even existential risks  (Moynihan 2020) . Such work has explored certain scenarios of low-or unknown-probability but high impact, such as the catastrophic or even existential threat to humanity that future AI systems might pose if not well-aligned with human values  (Yudkowsky 2008a; Bostrom 2014a; Everitt, Lea, and Hutter 2018; Russell 2019 ). In such cases, the world might get many 'near-term' policy issues at Levels 0-2 right (for example, such as achieving a Level 0 ban on AWS; altering notions of liability to account for unpredictable AI systems at Level 1; or putting in principles around the shift towards regulation by 'technological management' at Level 2)-yet in the long-term still see this come to naught. But the space of possible macrostrategic destinations surely extends beyond this. Indeed, recent Level 3 thinking therefore has explored (b) vulnerable worlds which concern ways in which worlds that are still 'safe' today may nonetheless be set on a trajectory within which continued technological progress inexorably ensures that, as Bostrom (2019, 3) predicts, \"a set of capabilities will at some point be attained that make the devastation of civilization extremely likely, unless civilization sufficiently exits the semianarchic default condition.\" Related to this, other scholarship has extended this further, by exploring the ways in which diverse, complex, path-dependent socio-technological trends, which in isolation do not rise to an 'existential risk', might converge and interact in ways that gradually, but steadily and irreversibly, increase the systemic vulnerability of our societies  (Liu, Lauta, and Maas 2018; Kuhlemann 2018) . This could include the accumulating effects of certain technologies-as with industry-fuelled climate change, a trajectory where a \"great many actors face incentives to take some slightly damaging action such that the combined effect of those actions is civilizational devastation\"  (Bostrom 2019, 7) . It could also involve intersecting or compounding effects of different technological disasters, such as the climactic 'termination shock' that would follow if future global geoengineering programs were to be suddenly interrupted as a result of regional (nuclear) war (S.D.   (Danaher 2017; 2019a) . In this way, we emphasise the importance of articulating governance lodestars; to pivot towards a prospective, teleological approach that sketches a spectrum of societal trajectories that not merely avoids catastrophic outcomes  (Bostrom 2013; , but which articulates what we seek to gain-and where we want to go. If developed, such work would highlight a potential new chapter for law and regulation, ensuring governance can speak to both the perils and promise of AI. \n Reflections on problem-finding AI Governance Strategies The above gives an analysis of four governance orientations, and their implications for constituting a long-term governance approach to the problems created by AI. We therefore will now provide four meditations on some of the nuances of our model; as such we will explore (a) the interrelation and mutual complementarity of governance across the four levels; (b) the potential reapplication of problem-finding governance research (Level 2 and 3) in a problemsolving mode; (c) how all four strategies are themselves merely situated within a larger 'Governance Goldilocks Zone' that circumscribes and articulates the very limits of what can be meaningfully governed; and (d) the interrelationship between governance strategies at these four levels, and broader (self-fulfilling) narratives or visions. \n Coordinating Strategies across Levels: from partial failures to co-evolution We want to emphasize that the point of the four-level framework is not to discard 'low-level' scholarship or governance, and fully replace it with 'high-level' scholarship. Indeed, in one sense, we should be careful to how we understand the distinction between these four levels; these are primarily mapped to the different ways in which governance strategies are oriented, and which (problem-; legal-; and societal) parameters they takes for granted. However, while this distinction between the four levels-and indeed, the terminology of 'levels'-seems to imply an ascending hierarchy or scale, that linkage should not be drawn too far. That is, it may be tempting to map the axis of problem-solving to problem-finding governance to many other scales, such as a distinction between 'near term' vs. 'long term' governance, 'specific-' vs. 'general' governance, or 'small scale' vs. 'large scale' issues. Yet while these other scales may show some correlation to our fourfold distinction, they should not be drawn too widely: it is after all possible (if not necessarily widespread) to adopt a Level 0 or Level 1 governance approach to problems that will likely only emerge in the 'long-term future' (as seen in the debates over legal personhood for robots). Conversely, it is also possible to adopt a Level 3 perspective which does not seek to take in the full spectrum of societal and technological activity, but which narrowly focuses in on one specific topic (e.g. 'over the next decade, will AI progress be more constrained by computing hardware or by data?') which is held to be a crucial consideration or pivot point for determining the longer term trajectory. As such, while the four Levels correspond somewhat to other scales, they should not be taken as equivalent. 10 More practically, the point of our framework here is not to argue for the categorical superiority of problem-finding approaches to old problem-solving 'dogma'. Indeed, we must  10  We thank Henrik Palmer Olsen for spurring this line of thought. beware not to replace the particular failures of a governance system that is almost solely problem-solving with the particular shortfalls of another governance system that is almost entirely problem-finding. Rather, the point is to use problem-finding scholarship to complement and build beyond the existing problem-solving work, while understanding how all these four strategic orientations are necessary and complementary-and how work must range and coordinate across them. In particular, we want to emphasise that each governance system has to find its own form of complementarity, which balances problem-solving with problem-finding strategies. This is a balance that can differ per governance domain depending on local conditions and interactions. To illustrate this, we now examine three types of common failure modes in AI governance which takes stock of only one or another approach, and which thereby risk failing to 'governing for the long-term': (a) reactive work that is restricted to levels 0-1, and therefore does not even try to anticipate the long-term; (b) work that examines Level 2 governance disruptors, but which ignores Level 3 macrostrategic destinations, and which therefore misses the forest for the trees; (c) work that considers Level 3 trajectories, but ignores Level 2 disruptors, and therefore risks becoming a set of grand futures, derailed by detail. In the first place, scholarship that only focuses on Level 0 is reactive and focused on solving direct problems in the present. In that sense, it is not even trying to pursue governance solutions that are scalable into the long-term. As discussed above, such approaches might have some limited uses, but as discussed above, they risk being blindsided both temporally and sectorally. In relation to Level 3 macrostrategic trajectories, this approach seems inadequate. It is reactive, a holding action at best, and does not reflect on longer-term trajectories or dependencies, meaning that we would have to be very 'lucky' to find ourselves in 'good worlds' through Level 0 work alone. In the second place, work that works up to Level 2, but does not consider the Level 3 macrostrategic destination, risks not seeing the forest for the trees. In that way, it may manage to identify and eventually address many cross-sectoral problems, but over time could still leave us stuck in a potential progress trap. That is not to say that such work is 'senseless', but rather that it too needs to be grounded in a picture of how constitutional shifts at distinct levels-and their associated governance strategies-cohere and converge into a long-term governance trajectories. Thirdly, and conversely, any work that focuses only on Level 3, but which tries to solely reason 'down' to the other levels, runs the risk of ending up as a set of grand futures, derailed by detail. Specifically, by ignoring Level 2 disruptors, such strategies may make foundational assumptions-about the available or effective instruments of law; about the values of policymakers and societies-which are on track to get sidelined by Level 2 eddies and deflecting vectors even in the medium term. For example, some proposals to govern and avert potential future catastrophic risks arising from advanced artificial intelligence systems have envisioned governance strategies grounded in a comprehensive inter-state global treaty regime (cf. G. Wilson 2013) or housed within the United Nations  (Nindler 2019; Castel and Castel 2016;  However see  Cihon, Maas, and Kemp 2020) . While it is certainly valuable to explore all avenues-certainly historically proven ones-the risk is that such a move constitutes an attempted leap, from a Level 3 macrostrategic goal (e.g. a catastrophic trajectory to avoid), immediately down to possible Level 0 or Level 1 solutions ('global treaties') that are patterned on the solution package perceived to be available or the norm today. In doing so, however, the risk is that such proposals fail to engage with key Level 2 insights. For instance, (a) some have Ultimately, the aim is not to elevate some of these strategies over others, but instead to help provide an inter-community framework or 'translation zone', that can help scholars situate themselves within scholarship at different levels, and to translate their work across to other communities as well as policymakers and the broader network of governance actors. Such an approach is critical also to fully levarage both the divergent 'scanning' function of problemfinding scholarship at Levels 2-3, as well as the convergent 'rallying' function of inviting groups and actors behind preferable Level 3 trajectories, in ways that are cognizant of possible Level 2 'governance disruptors', as well as the intricacies of concrete Level 0-1 policy implementation. \n Problem-finding strategies applied in problem-solving mode Having set out the interrelation and complementarity of problem-solving and problem-finding approaches across the four levels, we should now make a special observation with regards to the ways in which approaches at these four levels can reflect-or adapt to one another. After all, one reading of our argument above would be that while we have reservations about a problem-solving governance approach pursued in isolation, they can work alongside problem-finding approaches, because these two frames generate a sufficient complementarity. In another perspective, however, one might argue that even the problem-finding paradigms (Levels 2-3) can involve (or may require) 'problem-solving' features-but that these are different to the isolated (Level 0 or Level 1) versions of the problem-solving approach.  11  This second interpretation has much to recommend it. As noted above, for instance, Level 3 (problem-finding macrostrategic) scholarship on AI can be understood to include an analytical mode, which explores chiefly how the injection of AI systems (or indeed other technologies) may inadvertently alter the trajectory of human society in the mid-to long-term. This type of analysis can be purely problem-finding. It is important, however, to note that much work in this paradigm also involves an interventional mode: such work does not merely seek to 'find' previously unseen but crucial dangers; but it also aims to explore how we might identify or create governance inflection points and opportunities for intervention, to re-direct the global trajectory towards preferable (i.e. 'good') worlds. We should therefore draw a distinction between Level 3 work that aims to find new crucial considerations, and Level 3 work that aims to solve these challenges through interventions or inflection points. Paradoxically, this latter approach therefore appears to revive aspects of problemsolving analysis in Level 3 work. Nonetheless, the type of problem-solving here is frequently distinct from Level 0 or Level 1 problem-solving, because it need not be as closely wedded to pre-existing solution sets. That is, while it admits that any new problems (such as 'crucial considerations') excavated by problem-finding approaches (Levels 2-3) could be amenable to problem-solving approaches, it does not automatically presume that existing responses are necessarily adequate or superior to alternate governance responses. Nonetheless, the interrelation of the problem-solving and problem-finding paradigms in (AI) governance, and the way they can shape, alter and inform one another, remains a key nuance to this framework, which remains to be further explored. \n A Goldilocks Zone for Governance? Beyond the Four-Level Framework Moreover, while we distinguish four strategic levels in our governance framework, it is worth stepping back and taking stock of the broader long term implications that this framework raises, particularly in terms of how governance (of any form) categorically relates to differential, technology-driven changes.  12  The animating question here is whether there is something to be found beyond these four levels? That is, these four levels of problem-solving and problem-finding governance strategies may together describe (or cover; or constitute) the governance landscape in a complete and comprehensive manner. Yet the variety amongst them-and the lower and upper 'limits' of Level-0 and Level-3 governance strategies respectively-imply that these four levels of governance are in a sense themselves confined within a 'Goldilocks Zone' for governance. The 'Goldilocks Zone' is the informal astronomical term for the 'circumstellar habitable zone'-that narrow region around a star where the surface temperature on some planetary bodies is \"just right\" for water to be present in the liquid phase, implying they could support the emergence of life (NASA 2020). To extend this analogy, we might then come to think of a 'Governance Goldilocks Zone' as that narrow band of core parameters, assumptions, or boundary conditions for problem 'governability' within which the very project of governance (of either a problem-solving or problem-finding variety) is capable of residing or thriving in the first place. Within this broader context, we can recognize how one can only 'do governance' (of any type or sort) under assured conditions of autonomy (freedom to determine behaviour and to act) and influence (behaviours and actions are causally connected to outcomes), and these constitute the underlying presumptions of our four level framework.  13  The long-term governance perspective, however, enables us to step outside of these presumptions in order to understand the implications of a Governance Goldilocks Zone. That is, to draw another analogy (and at the risk of mixing metaphors), we might allude to the fundamental states of matter to illustrate the character of the problem landscape beyond this narrow Governance Goldilocks Zone. One might accordingly think of the Levels as corresponding to varying states of matter at distinct temperatures. In this heuristic, one could see Level -1 governance as corresponding to a Bose-Einstein Condensate (BEC) (a state of matter that occurs at extremely low temperatures near absolute zero, where molecular motion nearly stops and atoms begin to clump together); Level 0 would be a solid state; Level 1 a liquid state; Levels 2-3 the gaseous state; and Level 4 the plasma state. These suggest varying levels of particle movement and intuitive grasp of behaviour, but for the purposes of this section, we focus on the Bose-Einstein condensate and the plasma that bookend the fundamental states of matter. This because the Bose-Einstein condensate state provides an example of situations where change is difficult, and the plasma state presents an example as to the difficulties inherent for intuitive predictability of behaviour that lie beyond our daily experience of matter. Level '-1': Governance BEC: Stasis in reality or in governance responses Stretching backwards into the space 'beneath' Level 0 would imply stasis, both actual and perceived, with both forms of stasis converging to foreclose the impetus or opportunity for altering the existing governance landscape. Thus, at Level 'negative-1' (-1), 14 we have stepped outside of a governance framework that can be responsive and adaptive to (technologically induced) change, either because there is no actual change (stasis in reality), or no perceived or recognised change (stasis in governance responses) in the sociotechnical landscape. Of course, no society has ever been in a perfect form of internal stasis-let alone a form of stasis resilient to any outside shocks. Nonetheless, some soft forms of these conditions might have historically applied to (possibly pre-industrial revolution) societies, which did undertake investigations into the governance implications of (socio)technical change, perhaps because no such change was easily perceptible or in memory. Under these conditions, not even Level 0 'business-as-usual' governance is engaged, because no new (sociotechnical) situations (appear to) present themselves for examination. This static nature of Level -1 conditions results in governance self-entrenchment, because the very possibility of subsequent change (in governance) is systemically resisted: there are no new 'problems' that are presented for Level 0-1 approaches to 'solve'; and there is no background change that can provide an easy seed, impetus, or justification for Level 2-3 approaches to go out and 'find' potential challenges. In the course of discarding a dynamic view of governance, where its aims, objectives, values, methods, and actors are in constant and mutual feedback, Level -1 amounts in a sense to a condition of 'absolute zero', where there is no movement (at least from within the system) and therefore no prospect for interaction, recombination, or phase transitions.  15  There are no moments of 'legal disruption' or uncertainty  (Liu et al. 2020) , and as such Level -1 undercuts the assumption of governance autonomy by removing the perception or reality that there is change, and it can thus be perilous from a govenance perspective because alternatives are no longer imagined, perceived or pursued-let alone actualised. It lies outside the governance Goldilocks zone, because it is 'uninhabitable' for governance initiatives-providing no 'sustenance' or activation energy.  Level '4': 'plasma' governance beyond sight, comprehension, or control  At the other end of the scale, we can consider the space extending beyond Level 3 as analogous to plasma, a fundamental state of matter that is often misunderstood and difficult to intuit from  14  It should be emphasized that by calling these categories 'Levels', we are not suggesting that they lie within the same continuous 'problem-finding' framework.  15  Of course, to extend the analogy (or recognize some of its limits); there are of course materials which achieve unusual properties (e.g. superconductivity) at extremely low temperatures. Do governance approaches achieve such states as they crystallize below Level 0? We leave that question for future work. a perspective that is more familiar with the more quotidian states of matter. In the context of long term governance, Level 4 appears so deeply unfathomable as to defy conceptualisation from our present standpoint. It suggests the minimum requirement to reconsider our contemporary axiological configuration  (Danaher 2020) , and questions the desirability of our present notions, techniques and objectives for governance  (Bostrom, Dafoe, and Flynn 2019) . Indeed, at Level 4, it may be that our attempts at 'doing governance' may range from suboptimal through to being downright harmful over the long-term, because opportunities and prospects are opened up in this space that is currently beyond our ability to comprehend or fathom. From such a long-term governance perspective, Level 4 suggests that 'doing governance' might not only be futile, but may also be both counter-productive. The 'ingovernability' of Level 4 problems-that is, the features which drive a certain problem outside the governance Goldilocks Zone-may derive from three complementary sources, each of which is individually sufficient to erode or even foreclose our potential to 'do governance': (1) we cannot see governance problems; (2) we cannot comprehend the governance problems; (3) we fundamentally cannot control the governance problems (with today's tools). Respectively, these factors range from primarily undermining autonomy towards undercutting influence although of course these factors comprise of a mix of these two dimensions. With regards to (1) governance problems that we cannot see, there are a wide range of blinkers and veils that prevent us from seeing clearly all potential threats. Generally speaking, accurate judgments of catastrophic shocks are beset by various cognitive and epistemic biases  (Yudkowsky 2008b) , making their governance subject to a 'tragedy of the uncommons' (Wiener 2016). In extreme cases, the absence of observable or recognized precedent around truly existential disasters can create an 'anthropic shadow'  (Cirkovic, Bostrom, and   Finally, (3) with regards to (AI; or general) problems that are either intrinsically or currently deeply beyond our control, problems are again outside the Governability Zone. There are at least two ways we might lack 'control' over a governance problem: in the first case, one might imagine a case where a certain problem is in principle governable, but all power and control have become centralised to a single actor in a perfect autocracy, thereby locking out participation by all others. In this case, if the 'singleton' actor is uninterested in addressing a problem, governance would have no purchase either. In the second, more common case, the challenge might either be one fundamentally beyond our present technological means; 17 it might require a level of perfect collaboration and cooperation that is beyond our (present) political means; the prospect for control has either been lost, or is recognised as not existing.  18  The common denominator is the loss of control which shuts out the very possibility to 'do governance', either through the loss of participatory input in the first form or the disconnection between action and outcome in the second form. Taken together, these suggest that the governance strategies with which we are familiar, and which we have mapped to Levels 0-3, may in fact reside in a Goldilocks Governance Zone, which circumscribes the limits within which it is possible to meaningfully contemplate or enact any strategic long-term governance.  19  As the underlying presumptions of autonomy and influence fall away at either end, be it the stasis of Level -1, or the unfathomable invisibility, incomprehensibility or impotence that characterise Level 4, the Goldilocks Governance Zone comes into view and becomes demarcated. Yet, without attempting to define the four strategic levels for long-term governance, identifying such a sweetspot would have been a difficult endeavour. \n Strategies and narratives While we make a strong claim in support of the complementary (and under-appreciated) value of problem-finding scholarship to such long-term governance scholarship, the framework we have proposed here is by no means a definitive one, and there will be many directions within which it can be developed and enriched further. To highlight but one point of potential interest, along with governance strategies, it will be key to examine how overarching narratives and encompassing long-term visions which frame strategies, can percolate across the four levels. There is already extensive scholarship on the role, dynamics, and risks of imagined futures  (Beckert 2016)  or of 'sociotechnical imaginaries' (Jasanoff and Kim 2015) with which mediumterm technological projects such as 'smart cities' are often imbued  (Sadowski and Bendor 2019) . Likewise, there are a host of emerging Level 3 descriptive or aspirational narratives revolving around our societal macro-strategic societal state, trajectory, or destination-whether 'turbochange' (Deudney 2018), the 'great challenges' framework (Torres 2018), a 'vulnerable world'  (Bostrom 2019) , or 'existential security' (Sears 2020)-which provide kernels or seeds of various (possibly contradictory or competing) long-term perspectives and projects. 19 That is not to say that these conditions are rigid or fixed. To exploit the habitability analogy of the Goldilock Zone further, it is worth noting (a) that not all areas on earth fall within the habitable zone, as there are extremes in environmental conditions (e.g. volcanoes) which restrict (at least human) habitation; and (b) that we are technologically capable of affording human habitation beyond the natural parameters of these zones (including in outer space) through life support systems. Drawing on this analogy in terms of governability, we might suspect (a) that governance (and problem 'governability') is not uniformly distributed within Levels 0-3 of our governance framework, and that there might indeed be pockets of 'ungovernability' within these zones; and (b) even the external boundaries (e.g. the 'ceiling' of Level 3 governance) are not fixed, and using institutional or technological means, it could be possible to extend the zones of governability beyond the past parameters. This raises three questions: in the first place, how are such long-term narratives embedded and manifested in Level 0 to Level 2 partial strategies? In the second place, how do these narratives relate to, and exert political effects, alongside pre-existing 'utopian' or 'dystopian' visions, ideals or narratives of the future  (Berenskoetter 2011 )? Finally, in the third place, are there ways in which such narratives (or their policies or analysis at different levels) can have 'self-realizing' effects? The role of 'self-fulfilling prophecies' has a long pedigree in scholarship, particularly where it concerns the 'autogenetic effects' of social predictions or utopian and dystopian narratives  (Maas 2012) . Self-fulfilling and self-negating prophecies have received study in economics  (Felin and Foss 2009) , as well as in international relations  (Houghton 2009) , where scholars have charted the ways in which influential theories-such as the the 'democratic peace' and 'commercial peace' theses, or the 'clash of civilizations' (Bottici and Challand 2006; see also memorably Tipson 1997)-might end up self-fulfilling. Beyond the social realm, however, there might also be ways in which technological 'predictions', such as Moore's Law  (Mollick 2006)  or current predictions of technological unemployment, become selffulfilling  (Khurana 2019 ). This will be one key dynamic to be considered in both our problemfinding framework, as well as long-term governance more broadly. To conclude, we hope our framework can contribute to the co-evolution of governance actors (and scholars) operating within different 'Levels', as well as the coordination and unification of 'partial strategies' at these levels, into more deliberate, pluriform, and crossdomain approaches that leverage the best from both problem-solving and problem-finding. \n Concluding Thoughts Contemporary society continues to experience significant changes to the scale, pace, and type of change in both its social, natural, and technological environment. This suggests the need to define, develop, and deploy complementary governance strategies both to meet these challenges, and to build towards more positive futures, which can take long-term perspectives into account while leveraging concrete and operationalized policies today. In this paper, we have identified some of the shortcomings of contemporary approaches to both the study and practice of governance, which often-explicitly or implicitly-take a 'problem-solving' approach. Specifically, we argue that problem-solving approaches to complex problems (such as AI) are limited, because of (1) the unreflexive over-reliance on narrow metaphor in the face of complex phenomena; (2) the tautological manner in which existing governance lenses identify and prioritize primarily those problems that are already legible to them; (3) the constrained path-dependency; (4) the inability to clearly distinguish between symptomatic and root-cause problems; (5) the promotion of a false sense of security through a rapid delineation of the problem-space that prematurely closes-off large parts of the governance solution-space. We as such argued that while such 'problem-solving' approaches may be necessary, but are also insufficient in the face of many of the 'wicked problems' which we face, including in the realm of AI. By contrast, we have made the case that exploratory, problem-finding research is critical to long-term policy and governance strategy endeavours. This is true for epistemic and scientific Beyond these epistemic virtues, the proactive and explorative orientation of the problemfinding approach is crucial to any and all conceptions of the long-term perspectives, whether these are conceived as 'early warning radar', or as 'inspirational lodestar'. Accordingly, we have proposed and elucidated a rough typology for four distinct 'levels' at which one can study or formulate governance strategies: at Levels (0) 'Business As Usual' governance and (1) 'Governance Puzzle-Solving', work takes a predominantly problem-solving approach to issues. Conversely, problem-finding approaches focus on (2) 'Governance Disruptor-Finding' and (3) 'Charting Macrostrategic Trajectories'. While within this paper, we have primarily discussed specific examples involving AI governance strategies, problem-finding approaches would likely be equally crucial for many other cross-sectorial challenges-if not more crucial. Problem-solving approaches might gain traction with regard to the clear, discrete clusters of pre-defined problems that are thrown up in relation to AI governance, by nature of the technogenic origins of the challenges. Yet, intersectional challenges diminish the crispness of their concomitant problem-sets. The result is that there may be uncertainty or ambiguity concerning how problem-solving approaches could effectively go about confronting such emergent, complex, and interconnected global challenges, without arbitrarily confining the metaphors and models in order to manufacture the necessary problem-clusters to be tackled. Thus, while problem-solving approaches may already be strained when confronted with polymorphous policy domains such as those relating to AI governance, the efficacy of problem-solving approaches may fade further where long-term governance takes multiple interacting developments into view. Long-term perspectives trade in future worlds, themselves intricate complexes of emerging challenges and possibilities of which factors such as AI and its governance comprise a minuscule subset. Thus, attempts to grapple with long-term perspectives must necessarily involve problem-finding approaches at their very core because these futures remain mysteries. In slogan form: we need to collect, catalogue, and categorise the challenges of long-term governance before we can set about answering concrete questions that are raised. We opened this paper with an emphasis upon the notion of change: an apparently paradoxical observation that change is the only constant, while simultaneously itself being subjected to change. Stepping back from AI, and long-term governance -what does it mean for change to be changing? If we are right that the very possibility of doing governance requires change (that is, a guarantee that we are not at Level -1), then change undergirds the prospect, potential and promise of governance itself. Yet, it is clear that change is not one-dimensional, instead varying according to the rate, scale, direction, and duration, of the change in question: these can be further combined to constitute gradual, disruptive change, and paradigmatic change, for example. The question of change can also be addressed through an agent perspective to focus on who is driving or inhibiting the change in question; conversely a patient approach would focus upon the redistribution of benefits and burdens in the process of aftermath of that change. Furthermore, the relativity inherent within change requires considerations of benchmarks, those concerning the perceived status quo and expectations for the future for example, and how developments diverge or converge to these. The prospect for change also raises the possibility for encountering attractor states, which will affect the nature of observed change as well as future change. Leveraging the adjacent possibles opened up through the process of unpacking the four strategic levels for long-term governance, we also contextualised our framework within the broader spaces of (un)governability to suggest that our framework sketched out the Governance Goldilocks Zone within which the conditions are conducive to permitting us to still able to 'do governance' in the first place. Looking beyond the boundaries of our framework suggests directions for further research to explore and identfy the underlying presumptions that support the very prospect for our framework to come into play, and to the underexplored factors at play that enable us to project governance into the long-term. clarifying comments and questions during the final stage of refining this paper. No conflict of interest is identified. technological changes. All else being equal, holding up a 'pretence of continuity' might avoid certain externalities in the form of continuous regulatory uncertainty, or (in the case of highstakes technologies) continuous contestation or attempted renegotiation by stakeholders. It can therefore be a necessary governance strategy in areas where reaching original political agreement on the existing governance arrangement proved thorny or difficult-such as for instance in international arms control agreements on certain types of new weapons (Crootof 2019b;  Maas 2019a; Picker 2001 ). \n questions about what kind of worlds we do and do not want to reach with AI. Work at this level examines (I) analytically, how the injection of AI systems (or indeed other technologies) might inadvertently alter the trajectory of human society in the mid-to long-term (Brundage 2018; Seth D. Baum et al. 2019); and it examines (II) strategically, how we might identify or constitute inflection points and opportunities for intervention for re-directing that trajectory towards preferable worlds. Some work to date has focused on mapping (technological) vectors that could alter the macrostrategic trajectory. In most cases, such work has focused on (a) terminal trajectories, the most extreme scenarios whereby technogenic 'turbo-change' gradually or suddenly leads society \n argued traditional international law instruments based on state consent are very poorly equipped to engage extreme but unknown catastrophic risks (Aaken 2016;  though  see Vöneky 2018); (b) in recent decades, the parameters of global governance have already been shifting, with many arguing that it has trended away from formal international law-making, towards diverse 'regime complexes' made up of heterogeneous actors and informal governance arrangements (Alter and Raustiala 2018; Pauwelyn, Wessel, and Wouters 2014; Morin et al. 2019); (c) such proposals may not sufficiently engage with insights of how and where the deployment of AI itself may affect or erode the political scaffolding or legitimacy of 'hard' international law itself (Maas 2019b; Deeks 2020). Any of these might constitute a 'crucial consideration' against the project of trying to secure the Level 3 macrostrategic objective of avoiding disastrous long-term trajectories by attempted Level 0 or Level 1 tools (i.e. treaties) alone. \n reasons relating to good and valid science and scholarship-because problem-finding work can (1) itself produce foundational academic insights at the interstitices of different problem domains, or at different scales of analysis; (2) it can work in a productive reflexive equilibrium with problem-solving scholars and strategy-makers; because (3) it can at times reveal new strategic 'crucial considerations' which might completely alter the balance of local-context strategy considerations; and because (4) it can enable us to forestall the 'power' problem inherent to the Collingridge Dilemma of technology governance. \n Table 1 : Taxonomy of problem-solving (0-1) and problem-finding (2-3) AI governance strategies Level 0 -'Business-as-usual' governance 1 • Does not take for • May miss the forest for granted scope and the trees (relative to nature of governance Level 3) problems at hands • Aims to understand 'out of bound' strategic barriers or deflectors \n to correspond better to the evolving opportunity/risk profile as manifested by the affordances (Glăveanu 2012; Norman 2013)  of a new technology. To be sure, this is not to say that all such structural change comes from changes in technology, or that all technological change brings about structural shifts such as this. Nonetheless, at times such shifts can be key and disruptive to extant strategic trajectories. For example, the risk profile of displacing the primacy of concrete 'law' in regulating behaviour or in aiming to pursue strategies. For instance, it has been explored how AI systems enable regulation through In sum, whereas Level 0-1 work often takes for granted the scope and nature of the problems at hand, a Level 2 analysis investigates how the technology, its uses, its indirect effects, and the direct problems, may aggregrate into overarching trends that can change the terms of governance. It can as such reckon with 'out of bound' strategic barriers or deflectors (changes to the problem portfolio; to regulators' goals; to governance tools, or to societal values) that would completely surprise and change the terms of analyses at Levels 0-1-potentially rendering any solutions arrived at as superfluous, contrary to the new regulatory goals, or out of step with the nature and values of the society they are meant to serve. Simultaneously, analysis at this level can reckon with these 'medium-term' disruptors that might frustrate/undercut governance efforts that only take account of Level 3 destinations. 'microdirectives' (Sheppard 2018) and 'technological management' (Brownsword 2015; 2016; 2019b), potentially feeding into systems of 'algocracy' (Danaher 2016). AI also facilitates behavioural manipulation through 'hypernudging' (Yeung 2017; 2018), or invisible influences AWS can shift with technical progress: issues that appeared critical at an early stage (the embedded in AI-mediated adaptive choice architectures (Susser 2019). Likewise on global level, potential to violate IHL principles), may soon become matched or eclipsed by far-reaching the use of AI systems might well help alter the processes by which international law or broader challenges in other domains: for instance, military AI systems may generate risks of operational global governance is produced or enforced (Deeks 2020; Maas 2019b). While such developments safety, 'flash wars', or strategic instability (Scharre 2016b; Danzig 2018; Maas 2019c; Geist and may themselves give rise to concern, and therefore may become the object of (long-term) Lohn 2018; Sharikov 2018). Moreover, much greater than the direct risks from misuse or governance strategies themselves, such shifts should simultaneously be examined insofar as accident, might be indirect risks from 'structure'-deriving from the way new AI systems or they alter the operating parameters-including what Deudney (2018) describes as 'material- capabilities shape the landscape of incentives around actors, in potentially hazardous ways contextual' factors, and what van Assche et al. (2020) describe as the 'material' dependencies of the human-made environment. (Brownsword 2018), potentially spelling the 'death of [regulation through] rules and standards' , Level 2 scholarship shifts focus from changes beyond narrowly understood laws or (Casey and Niblett 2017). Moreover, AI-sparked shifts in the international balance of power government regulation, towards to the broader regulatory system-and how developments in a could destabilise the global legal order, just as previous technologies have upset international might lessen the reputational penalties from waging high-tech wars. Yet, the corresponding particular technology (such as AI) may end up disrupting or deflecting governance efforts (either law (Picker 2001). AI could do so by differentially empowering illiberal states (Danzig 2017; shift to a paradigm that heightens scrutiny or accountability of such 'hidden violence' (Kahn those aimed at the technology, or even in general). Accordingly, Level 2 analyses can emphasize Harari 2018; N. Wright 2018), or by eroding the buy-in of powerful states, into the multilateral 2002) may not develop in time, or to the same intensity. Finally, the fact that technological at least four distinct angles. international legal order (Danzig 2017; Maas 2019b; Deeks 2020). In the context of our AWS change can 'reveal rights' (Parker and Danks 2019) can demonstrate shortcomings in the First, such scholarship can explore how ongoing technological progress-or altered social example, finding new political equilibria that can support global restrictions on AWS might existing human rights umbrella. This limits human rights' ability to 'push-back' effectively become more urgent than reconceptualising 'autonomy'. practices that seize upon pre-existing but formerly marginal affordances in new, salient wayscan shift or expand a technology's 'problem portfolio' in ways that render initial, now pathdependent, regulatory efforts inadequate or even counterproductive. Such work can expand the subjective problem portfolio (Zwetsloot and Dafoe 2019; van der Loeff et al. 2019) . Furthermore, AI development in areas such as one-shot learning (Ram 2019) , 'simulation transfer ' (OpenAI et al. 2018) , synthetic data, and others can lower the data or expertise needs for using AI, and thereby can lower the proliferation threshold of certain AI systems to new, non-treaty parties. Indeed, in general, greater data efficiency can have considerable and far-reaching governance implications (Tucker, Anderljung, and Dafoe 2020). In these cases, neither 'solving the AWS problem' at Level 0 by extending IHL principles, nor tailoring legal principles to AWS at Level 1 does much good. Thus, policy responses at Levels 0-1 may no longer track-and may even occlude-the changing character of threats posed by AI systems. Instead, Level 2 articulates more adaptive or 'preventative security governance' approaches (Garcia 2016;  or 'innovation-proof' governance(Maas 2019a; see also Crootof 2019b) . Thus in our AWS example, Level 0-1 approaches suggest winning a legal battle, but in reality the regulatory war is lost.Second, AI can change policymakers' goals in formulating law. The regulatory mind-set may shift from 'legal coherentism' towards 'regulatory instrumentalism' or even 'technocracy' Third, reconfigurations between the different regulatory modalities may shift the very operational foundations (the 'wedge' or 'contact point') of governance strategies wholesale, by(further)As such, even if AWS challenges were to be 'solved' within Levels 0-1, a general retrenchment in the relative regulatory strength of 'normative' (international) law relative to unilateral technological tools (Brownsword 2018) may neutralise these gains. For instance, where AI mediates or undercuts meaningful human decision-making, global AWS regulations anchored in maintaining 'meaningful human control' over narrowly defined 'autonomous' weapons systems, will overlook AI's behavioural influences, 'automation bias', and propensity to 'normal accidents' (Maas 2018; Scharre 2016a;  Carvin 2017; Borrie 2016 ). In such cases, human agents remain nominally engaged in lethal decision-making, but in fact are consigned to the 'moral crumple zone' (Elish 2016) .Fourth, AI can change core values(Danaher 2018) and even fundamental rights(Liu 2019b;   Liu and Zawieska 2017), both altering the yardstick against which we measure impact of AI and potentially defusing our means (or indeed motive) of resistance. In our AWS example, if military AI applications enable a pivot towards conflict prediction and pre-emption (De Spiegeleire, Maas, and Sweijs 2017), the shift from high-casualty drone strikes towards 'invisible wars' that subtly enable the prediction and interdiction of 'enemies' (Deeks 2018)  against AI: in our AWS example, attempting to fit the wrongs precipitated by AWS as claims made through the existing human rights framework reveals severe shortcomings that structurally understate the injury (Liu 2018; . \n These are axiological descriptions of certain societal destinations-or trajectories-that do not merely avoid these above enumerated risks, but articulate clear principles or 'desiderata'(Bostrom,   Dafoe, and Flynn 2019)  for what good society we would like to pursue or societal trajectory we would like to be set on. Of course, in one sense, perspectives that articulate how we want to proceed ('journey' or 'trajectory') or where we want to end up in ('destination') in the long term, are both alternative and complementary strategies for avoiding 'bad worlds' or other pitfalls in development. These should invoke different types of assessment. The (troubled) narratives of utopia versus dystopia seem to consider macrostrategic destinations in their exploration of contours and configurations of established worlds, whereas ethical quandaries seem to relate to macrostrategic trajectories insofar as their consideration or their outcomes affect pivotal points in pursuit of a path ahead. At the very least, setting out the parameters of 'good worlds' provides a benchmark to assess whether or not we are either on the path to a bad world, or are actually in some sense in a bad world already. This would be useful since we presently seem to lack objective points of reference to indicate what type of world we are in. It should, however, be noted that such shared positive perspectives for long-term societal macrostrategy remain relatively underdeveloped-especially in terms of linking them to shared narratives and visions. Baum, Maher, and Haqq-Misra 2013). A further variation on this theme examines trajectories towards (c) exposed worlds. Whereas a diminished ability to cooperate (Christakis 2019); or increasing algorithmic 'inscrutability' may lead us to a 'new dark age' of uncertainty, surveillance, or the death of sociality or empathy (Bridle 2018). Such long-term trajectories towards 'flawed realization' (Bostrom 2013, 19) might emerge if AI systems would have long-term and lasting impacts on our world and society which would (on net) be considered ethically adverse or at least sub-optimal across a broad range of ethical views. vulnerable worlds can involve disastrous scenarios as a result of many small accumulating or interacting failures, errors, or hazards, an exposed world-or what some have called a 'fragile world' (Manheim 2018)-is one that makes certain choices that have rendered it susceptible to previously-modest shocks or hazards, including hazards that society previously might have weathered much better. Under some scenarios, technologies such as AI might lead us to more exposed worlds, insofar as larger parts of our economy and knowledge base become tied up in certain to-us opaque infrastructures dependent on global connectivity and constant and reliable electricity supplies.Alternatively, AI could inadvertently lead societies into (d) bad worlds: 'progress traps' (R.Wright 2005) or 'inadequate equilibria' (Yudkowsky 2017)  where society reaches a future destination where it is not destroyed, nor necessarily vulnerable or exposed, but where it is bad and (nearly) inescapable in the sense that subsequent recovery out of this state towards better trajectories is no longer possible. Even if these effects are not absolutely 'intrinsic' to AI technologies, it may still be that many of the uses to which AI is put within our society, nonetheless could: imperil human dignity (Brownsword 2017); erode the 'social suite' and relational capacities of humans, in ways that drive creeping negative 'social spillovers', such as Finally, while these are sometimes under-examined relative to these other scenarios, it is also important for Level 3 scholarship to orient itself towards (e) Good worlds.The insight of Level 3 problem-finding scholarship is that, even if we managed to solve the discrete AI policy-problems-such as the global regulation of military AI-at lower levels, this reactive, fragmented firefighting approach may simply not suffice in aggregate to shift societies out of overall-captive trajectories towards the 'attractor states' of vulnerable, fragile or bad worlds. Rather, AI might drive long-term sociotechnical effects or shifts in constitutional societal values, shifts which might be passively unobserved or even unobservable (cf. Cirkovic, Bostrom, and Sandberg 2010)  by society, or which might even be actively veiled by certain interested parties, until it is too late to marshal resistance or change. The point is that AI can lock in such dependencies in ways that even the broad analyses at Level 2 may not fully appreciate, missing the forest for the trees. At the same time, Level 3 work highlights positive leverage points which problem-oriented work at Level 0-2 might pass by: opportunities where AI itself opens up loci for positive interventions that can shift this trajectory-promoting beneficial applications of 'AI for global good' (Cave and Ó hÉigeartaigh 2018) which empower human autonomy, or help consolidate a 'postwork utopia' \n Sandberg 2010),16    which can exacerbate the effects of our 'availability bias', further restricting our ability to see what might really be at stake. As Cirkovic, Sandberg, and Bostrom put it: 'the observation selection effect implicit in conditioning on our present existence prevents us from sharply discerning magnitudes of extreme risks close (in both temporal and evolutionary terms) to us'(2010, 1500). Even where we might perceive certain threats in the abstract, the intangible and distributed nature of certain global processes (such as climate change) can make them 'hyperobjects' (Morton 2013) , complicating our apprehension of these problems (or the appropriate levers we might use to affect them). While these constitute the prominent examples, it becomes obvious that the prospect for long-term governance is severely eroded under conditions of invisibility. Because these governance problems lie beyond the epistemic 'event horizon', they do not prompt problem-solving investigations; and even problem-finding investigations may only hit on them 'by chance'-as theoretical possibilities discovered in the course of exploring some (adjacent) other problem. While Level 3 problem-finding work may sometimes help at spotting such challenges, for lack of a 'signal', many of them may remain governance 'dark matter' that is structurally outside of the Governance Goldilocks Zone. Danaher 2019b). Within this context, we may appreciate that there is a problem (posed by AI), yet our inability to properly comprehend it may place us within the equivalent of a 'governance black box' which, Danaher proposes, we find ourselves in, in relation to AI systems. Thus, the arguments here are the same as those Danaher presents in relation to techno-superstition: lack of understanding, illusion or control, erosion of Our lack of comprehension under (2) merges into such conditions of opacity, mirroring the distinction drawn between sensation and perception on the one end, and our inability to control outcomes under (3). For instance, John Danaher has anticipated that algorithms may contribute to a coming era of 'Techno-Superstition', which combines opacity (a growing lack of public understanding of how the world-or an AI system-actually works) with the unwarranted illusion of control over AI systems ( achievement, loss of autonomy, and undermining of human agency. Applied to Level 4, these arguments would suggest that we are not able to 'do governance' in a meaningful manner under such conditions. \n\t\t\t Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t In this section, the terminological reversion to 'problems', as opposed to 'mysteries' (Chomsky 1976) , is unavoidable given the use of 'problem'-terminology in the wicked problems literature itself. The question of how to integrate these into a concept of 'wicked mysteries' is left to future work. 5 Although the 'wicked problems' framework has also come under critique in recent years, on both conceptual and practical grounds. See for instance (Turnbull and Hoppe 2019; Noordegraaf et al. 2019; Peters and Tarpey 2019) . \n\t\t\t Although see the distinction between 'coping', 'taming', or 'solving' in (Daviter 2017) .Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t We thank Roger Brownsword for spurring this line of thought.12  We thank Victoria Sobocki for discussions prompting this section. \n\t\t\t This has similarities to Roger Brownsword's account of the core 'existence conditions' for human life (such as the maintenance of the core infrastructure conditions), and the generic conditions for agency, which he holds to be the 'first regulatory responsibility', and a 'regulatory red line' (Brownsword 2019a, 90-95).Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE> \n\t\t\t 'Anthropic bias can be understood as a form of sampling bias, in which the sample of observed events is not representative of the universe of all events, but only representative of a set of events compatible with the existence of suitably positioned observers' (Cirkovic, Bostrom, and  Sandberg 2010, 1495-96). \n\t\t\t For example, while we may harden the resilience of our societal infrastructures (e.g. electrical grid) to solar flares, we do not have the technical capabilities to affect the solar mechanics in order to reduce their prominence.18  For example, by taking complex adaptive systems seriously and recognising the ramifications that actions and inputs are largely untethered from producing the intended outcomes over the span of longterm governance perspectives. Alternatively, the viewpoint advanced by Nassim Nicholas Taleb over the course of Incerto could also be applied to suffuse randomness and uncertainty into processes upon which we project causation and the illusion of control.Electronic copy available at: https://ssrn.com/abstract=<PHONE> Electronic copy available at: https://ssrn.com/abstract=<PHONE>", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/SSRN-id3761623.tei.xml", "id": "e483964540bb249ea28051f4101c8c7c"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "is a research organization focused on studying the security impacts of emerging technologies, supporting academic work in security and technology studies, and delivering nonpartisan analysis to the policy community. CSET aims to prepare a generation of policymakers, analysts, and diplomats to address the challenges and opportunities of emerging technologies. CSET focuses on the effects of progress in artifi cial intelligence, advanced computing, and biotechnology.", "authors": ["June", "Micah Musser", "Ashton Garriott"], "title": "Machine Learning and Cybersecurity", "text": "he size and scale of cyber attacks has increased in recent years as a result of a number of factors, including the increasing normalcy of cyber operations within international politics, the growing reliance of industry on digital infrastructure, and the difficulties of maintaining an adequate cybersecurity workforce. Many commentators across government, media, academia, and industry have wondered how cybersecurity professionals might be able to adapt machine learning for defensive purposes. Could machine learning allow defenders to detect and intercept attacks at much higher rates than is currently possible? Could machine learning-powered agents automatically hunt for vulnerabilities or engage an adversary during an unfolding attack? Should policymakers view machine learning as a transformational force for cyber defense or as mere hype? This report examines the academic literature on a wide range of applications combining cybersecurity and artificial intelligence (AI) to provide a grounded assessment of their potential. It breaks cybersecurity practice into a four-stage model and examines the impact that recent machine learning innovations could have at each stage, contrasting these applications with the status quo. The report offers four conclusions: • Machine learning can help defenders more accurately detect and triage potential attacks. However, in many cases these technologies are elaborations on long-standing methods-not fundamentally new approaches-that bring new attack surfaces of their own. \n Executive Summary \n T • A wide range of specific tasks could be fully or partially automated with the use of machine learning, including some forms of vulnerability discovery, deception, and attack disruption. But many of the most transformative of these possibilities still require significant machine learning breakthroughs. • Overall, we anticipate that machine learning will provide incremental advances to cyber defenders, but it is unlikely to fundamentally transform the industry barring additional breakthroughs. Some of the most transformative impacts may come from making previously un-or under-utilized defensive strategies available to more organizations. • Although machine learning will be neither predominantly offense-biased nor defense-biased, it may subtly alter the threat landscape by making certain types of strategies more appealing to attackers or defenders. This paper proceeds in four parts. First, it introduces the scope of the research and lays out a simplified, four-stage schema of cybersecurity practice to frame the different ways that future machine learning tools could be deployed. Second, it contextualizes recent machine learning breakthroughs and their implications for cybersecurity by examining the decades-long history of machine learning as applied to a number of core detection tasks. The third and most substantial section examines more recent machine learning developments and how they might benefit cyber defenders at each stage of our cybersecurity schema, and considers whether these newer machine learning approaches are superior to the status quo. Finally, a concluding section elaborates on the four conclusions mentioned above and discusses why the benefits of machine learning may not be as transformative in the immediate future as some hope, yet are still too important to ignore. \n Introduction s a typical internet user goes about her day, she will be quietly protected by a bewildering number of security features on her devices-some obvious, others less so. If she uses Gmail, Google will automatically scan every email that arrives in her inbox to determine if it is spam, and if the email contains attachments, those will also be scanned to determine if they contain malware.  1  Whether she uses Chrome, Firefox, or Safari to browse the web, her browser will analyze every website she visits and attempt to alert her if a site is malicious.  2  If she uses an antivirus product-among the most common of which are Microsoft Windows Defender, Symantec, and ESET-then the files on her device will be regularly scanned to check for potentially malicious files.  3  All of these services utilize machine learning to protect users from cyber attacks. Over the past decade, the role of machine learning in cybersecurity has been gradually growing as the threats to organizations become more serious and as the technology becomes more capable. The increasing commonality of cyber operations as a geopolitical tool means that many organizations risk being targeted by well-resourced threat actors and advanced persistent threats.  4  At the same time, the supply of trained cybersecurity professionals struggles to meet the growing need for expertise.  5  And to add to the sense of risk, a nearly endless stream of publications-including a recent report released by CSET-speculates about the ways AI will be used to further intensify cyber attacks in the near future.  6  In this climate, researchers, policymakers, and practitioners alike have found themselves wondering if machine learning might provide the means to turn the tide in the war against cyber attacks. Popular media outlets frequently write about the possibility that machine learning could vastly A improve attack detection, while many in the defense community talk about future AI systems that could hunt for and dispatch intruders autonomously.  7  \"AI\" has become a mainstay in cybersecurity marketing materials, with advertisements comparing AI products to everything from the immune system to swarms of organisms.  8  Taking these claims at face value, a policymaker paying loose attention to the state of AI research might conclude that the total transformation of the cybersecurity industry at the hands of machine learning algorithms is at most a few years away. So it might surprise that same policymaker to learn that machine learning systems have been commonly used for a number of key cybersecurity tasks for nearly 20 years. While the breakthroughs of the last five years have rightfully drawn attention to the field of AI and machine learning (ML) research, it is easy to forget that the algorithms behind these advances have in many cases been around for decades.  9  Learning this, that same policymaker might look at the spate of recent high-profile hacks and feel tempted to dismiss the potential of machine learning altogether. After all, if machine learning has slowly been making its way into the cybersecurity industry for decades, yet the rate and scale of cyber attacks hasn't meaningfully gone down, why should anyone believe that a transformation is just around the corner? Part of the difficulty in evaluating machine learning's potential for cybersecurity comes from the fact that different parties have different standards for success. For our purposes, it is useful to make a distinction between two types of standards that we might use to evaluate the impact of a new technology: counterfactual standards and historical standards. Using counterfactual standards means that we ask the question: Where would we be without machine learning, all else being equal? Approaching the question from this angle should give us a great deal of appreciation for machine learning: given that the number of events that signal a possible cyber attack has grown from several hundred to several million per day in some industries, it's surprising that most If machine learning has slowly been making its way into the cybersecurity industry for decades, yet the rate and scale of cyber attacks hasn't meaningfully gone down, why should anyone believe that a transformation is just around the corner? companies aren't entirely consumed by the task of keeping their IT infrastructure secure.  10  In no small part, this success is thanks to ML systems that can automatically screen out potential attacks, generate alerts on suspicious behavior, or perform some rudimentary analysis of anomalous activity. Counterfactual standards are what matter to the cybersecurity practitioner-the person who knows the threat landscape and has to respond to it one way or another. But policymakers also need insight into a set of more general questions: Is the world getting more secure or less? How worried should we be over the long-term about the shortage of cybersecurity talent? Can machine learning make us less vulnerable to our adversaries in absolute terms? To answer these questions, we need to use historical standards. Rather than comparing the world with machine learning to a hypothetical world without it, we should compare the past to the present to the (most likely) future. That machine learning can offer significant benefits to cybersecurity practitioners is broadly-though not universally-agreed upon.  11  Whether these benefits will amount to a \"transformation\" in cybersecurity is more contested. For the purposes of this paper, \"transformative impact\" refers to an impact that makes a difference by historical-and not merely counterfactual-standards. In the context of cybersecurity, this means that a technology should do more than help defenders simply keep up in the face of growing threats: to be transformative, the technology should bring the total number of threats requiring a human response down (and keep them down), or it should meaningfully alter cybersecurity practice. Ultimately, it is more important for policymakers to understand how machine learning will transform cybersecurity than it is to quibble about whether it will bring changes. For that reason, this paper explores the potential impact that a wide range of ML advances could have on cybersecurity. We try to simultaneously present the reader with a healthy dose of skepticism regarding some of the most-hyped ML That machine learning can offer significant benefits to cybersecurity practitioners is broadly-though not universally-agreed upon. Whether these benefits will amount to a \"transformation\" in cybersecurity is more contested. applications in cybersecurity, while also drawing attention to some potential applications that have garnered less popular attention. And, in keeping with our emphasis on historical standards for success, we try to contextualize the recent growth of interest in machine learning for cybersecurity by examining how it has already been deployed in the past. This is a large task, and unfortunately not everything relevant to machine learning's growing role in cybersecurity can be covered in this report. Our focus is squarely on the technical-what machine learning could do. We sidestep many important but non-technical issues, such as privacy or legal concerns surrounding cybersecurity data collection, as well as the economics of implementing newer ML methods. We also adopt a general model of cybersecurity and avoid organizations that may have unique cybersecurity needs or goals, such as companies that rely on industrial control systems. This report is structured as follows: We begin in the first section by introducing a four-stage model of cybersecurity practice that we will return to in later sections as a way to schematize different ML applications. The second section, \"Traditional Machine Learning and Cybersecurity,\" discusses how decades-old ML methods have been widely studied for three core cybersecurity tasks: spam filtering, intrusion detection, and malware detection. Providing this historical context can help the reader understand which ML applications are genuinely new innovations, and which ones are merely extensions of long-standing trends in cybersecurity. In the third section, \"Cybersecurity and the Cutting Edge,\" we turn our attention to newer advances in ML research and examine how these methods might generate new types of applications at each part of our four-stage model of cybersecurity. This section emphasizes lines of academic research that seem promising and may translate into fully functional products within the near-to medium-term. The final section presents conclusions that appear to follow from this research, along with some additional analysis. he most well-known schema for conceptualizing cybersecurity is the Cybersecurity Framework designed by the National Institute of Standards and Technology (NIST).  12  This framework broadly breaks cybersecurity into five functions to help defenders assess their level of risk: identify, protect, detect, respond, and recover. This paper uses an adapted version of the NIST model that emphasizes four slightly different categories: prevention, detection, response and recovery, and active defense. There are two primary reasons to deviate from the NIST framework. First, while there are many instances of machine learning being used to detect cyber attacks, the use of ML in the other four categories of the NIST framework is still rather immature, which justifies grouping multiple categories together for analysis. Second, \"active defense\"-which we elaborate on below-is a growing area of interest that is conceptually distinct from the other categories for both technical and operational reasons; it therefore merits its own discussion. A Four-Stage Model of Cybersecurity 1 \n T In this model, prevention refers to all the actions that security experts take to minimize the vulnerabilities of their organization. At a minimum, prevention requires that defenders make secure decisions regarding network configurations and user privileges, and that they maintain an active knowledge of their networks and software dependencies, patching known vulnerabilities in a timely fashion. For software companies, prevention includes the work of evaluating one's own products for vulnerabilities, preferably before a product is brought to market. For mid-or large-sized organizations, prevention often requires frequent audits of one's overall security posture as well as periodic penetration testing, in which benign attacks are launched against a network to discover and expose weaknesses. This category roughly corresponds to the NIST categories of identification, which emphasizes asset management and security strategy, and protection, which focuses on the technological and policy controls that are used to secure assets. No amount of anticipatory work can make an organization immune from cyber attacks. When attacks do occur, detection systems are needed to quickly alert defenders so that they can respond. A comprehensive set of detection systems must monitor everything from network traffic to email attachments and employee behavior in order to identify anomalies and potential threats. Because attackers often evade discovery for months after breaching a network, effective detection systems can make an enormous difference to an organization's ability to limit the impact of cyber attacks. 13  Summary evaluations of GPT-3 performance on six disinformation-related tasks. Once an attack is detected, security professionals must determine how to respond and recover. For some types of threats, the response is straightforward: in the case of spam detection, for instance, an email service simply redirects mail that is likely illegitimate into a spam folder. For many other types of attacks, however, the appropriate response is far from clear. In the face of an ongoing attack, cybersecurity personnel must often decide whether to shut down machines, sequester parts of their network, or otherwise take steps that could significantly disrupt an organization's operations. An effective response must identify the scale and scope of an attack, thwart the attacker's access, and fully eliminate any foothold the attacker might have. After doing so, it is important to restore the system to its original state prior to the attack. The triad of prevention, detection, and response and recovery forms the core of cybersecurity. For most organizations, performing these tasks well is the height of good cybersecurity practice. However, for organizations facing attacks from well-resourced threat actors, compliance with pre-existing frameworks may not be sufficient. These organizations must also ensure that they can flexibly adapt to changes in the threat landscape. To account for actions that allow organizations to respond more flexibly to sophisticated threats, this report includes the additional stage of active defense. This term is used analogously to the way the SANS Institute has used it: as a spectrum of activity that includes annoyance, attribution, or outright counter-attack.  14  Active defense can be thought of as an \"other\" category that includes any attempt to deliberately engage or study external actors rather than simply responding to problems as they arise. This category can be broken down into a few more clearly defined subcategories, of which this report emphasizes three: (1) deception, or attempts to mislead and slow down adversaries; (2) threat intelligence, or attempts to actively study potential adversaries to better anticipate their actions; and (3) attribution, or attempts to connect multiple events to a single entity that can then be studied in more detail.* Active defense, done well, can allow defenders to stay ahead of their adversaries and can potentially create disincentives against attacking in the first place. Section 3 will examine each of these four components of cybersecurity and ask how newer ML architectures may play a role in augmenting current practices. But before turning towards the future of machine learning in cybersecurity, it is important *This breakdown does not include any activities that would fall into the category of outright counterattack. Because this is a much more difficult subject with an often unclear status of legality, we do not explore the possibility of using machine learning for counter-attack in this paper. to get some sense of the past, because ML-based cyber tools are not as new as many assume. In the next section, this report examines three ML applications that have been studied for several decades: spam detection, intrusion detection, and malware detection. lthough the last decade has seen major advances in AI and ML research, the use of ML methods for cybersecurity has a long history. Until recently, these applications fell almost entirely into the detection stage of cyber defense, with the most significant attention paid to spam detection, intrusion detection, and malware detection. This section discusses how simpler, longstanding ML approaches have historically been adapted for these three applications. Figure  1  presents a rough timeline of major developments in the cyber threat landscape and how machine learning has evolved to combat them. \"Traditional machine learning\" here refers to the broad set of decades-old algorithms that were dominant in ML research until the major, predominantly deep learning-driven advances of the past 5-10 years.* These methods are typically divided into one of two categories: supervised learning and unsupervised learning. In supervised learning, labeled datasets are used to train a model to classify new inputs, while in unsupervised learning, unlabeled datasets are examined to identify underlying patterns within the data. Either approach can be useful for cyber defenders looking \n Traditional Machine Learning and Cybersecurity \n 2 \n A * Some examples of traditional machine learning algorithms include naive Bayes classifiers, random forests, k-means clustering, logistic regression models, and support vector machines. It is worth noting that despite the name we use, \"traditional machine learning\" is not itself a stagnant field; XGBoost, for instance, is a popular and newer gradient boosting algorithm, but one that simply represents a better optimization of older, more traditional approaches. to detect potentially malicious traffic. If well-labeled data on previous attacks exists, supervised methods can be used to detect future attacks by matching malicious traffic to a known profile, while unsupervised methods can be used to identify attacks merely because they are anomalous and out of place.  †  To show the history of how these traditional machine learning approaches have been used in cybersecurity, we turn first to the example of spam detection. *See the note on page 10 below for a definition of polymorphic and metamorphic viruses. †This distinction is not as clear-cut as it seems; in some cases, unsupervised learning can be used to match attacks to known profiles and supervised learning can be used to detect anomalies. For the most part, however, the way we have made the distinction corresponds to the way that supervised and unsupervised algorithms have been studied for detection purposes.   \n SPAM DETECTION Machine learning has been a major part of spam detection since the very early 2000s, and many of these early ML methods are still used today. Before the introduction of ML techniques, spam detection relied on blocklists that screened out mail from (known) malicious IP addresses or on keyword detection that blocked emails containing hand-curated lists of spammy terms like \"free\" or \"sexy.\" Unfortunately, because these methods were applied indiscriminately, they could often block legitimate emails. To address this problem, computer scientists began to propose machine learning-based solutions around the turn of the century.  15  These early methods were relatively straightforward: First collect a large body of emails, label them as either spam or legitimate, and split them into individual words. For each word, calculate the probability that an email was spam if it contained that word. When a new email arrives, the probabilities associated with each of its words could be used to calculate the risk that the email was spam, and emails with risk scores above a preset threshold could be automatically blocked. The core elements of spam detection have not changed much since the early 2000s, though researchers have made improvements. Better spam classifiers can be built by extracting more technical details from mail headers, such as IP addresses and server information, or by treating words that appear in a subject line differently from words that appear in the body of an email.  16  Better algorithms can be used that recognize phrases or synonyms rather than treating all words as independent of one another. 17 Some companies have developed extremely complex spam detectors that can, among other things, track a user's past email interactions to flag anomalous contacts or use deep learning models to determine whether or not branded emails are sent from authentic companies.  18  Nonetheless, even the most advanced spam classifiers used by companies like Google have mostly developed out of a slow process of elaboration and evolution from these earlier methods. Of course, even moderate improvements in accuracy can matter a great deal to massive companies responsible for protecting billions of emails. But it would be an error to portray recent innovations in ML spam detection systems as a fundamental transformation of past practice: in reality, machine learning has been central to the task for nearly two decades. \n INTRUSION DETECTION Intrusion detection systems attempt to discover the presence of unauthorized activities on computer networks, typically by focusing on behavior profiles and searching for signs of malicious activity. Intrusion detection systems are typically classified as either misuse-based or anomaly-based. In misuse-based detection, attacks are identified based on their resemblance to previously seen attacks, whereas in anomaly-based detection, a baseline of \"normal\" behavior is constructed and anything that does not match that baseline is flagged as a potential attack. Both methods can make use of different ML methods. The simplest forms of misuse-based detection rely on known indicators of compromise to detect previously encountered threats. For instance, if an organization has seen malware that attempts to contact a specific website, cyber defenders could write a simple detection system which provides an alert any time a machine on the network attempts to contact that website. Misuse-based detection-especially when based on simple methods like these-typically has high processing speeds and low false positive rates, which allows it to quickly and accurately identify malicious events. However, this form of detection can only monitor for known threats, providing little meaningful protection against novel attacks. Machine learning can be used to automate some forms of misuse-based detection by allowing a system to \"learn\" what different types of attacks look like. If many (labeled) examples of past attacks are available, a supervised learning classifier can be trained to identify the tell-tale signs of different types of attacks, without the need for humans to generate specific lists of rules that would trigger an alert.* Since at least 1999, researchers have attempted to generate network traffic profiles of different types of attacks so that ML classifiers can learn how to identify previously seen attacks.  19  This research was initially pushed by DARPA, and the results suggested that machine learning could prove competent at misuse-based detection.  20  Although it is tempting to think that newer ML methods-like the rise of deep learning-have enabled dramatically more powerful detection tools, one review of several dozen experimental results in 2018 suggested that deep learning is not reliably more accurate at misuse-based detection than decades-old ML approaches.  21  Because no one type of model is consistently best for misuse-based detection, researchers often find that the most successful machine learning systems are ensemble models, or models that classify new inputs by utilizing multiple classifiers that \"vote\" on a classification.  22  The relatively common use of this method avoids overreliance on any specific model-which may have its own blind spots-but it also indicates that no single architecture is clearly superior to the rest. In contrast to misuse-based detection, anomaly-based detection flags suspicious behavior without making specific comparisons to past attacks. This type of detection system is more likely to use unsupervised learning methods to cluster \"normal\" traffic within a network and alert as suspicious any activity which deviates from that pattern. In theory, anomaly-based detection can identify novel attacks-one of the most difficult aspects of cybersecurity. To enable this type of capability, research in this area has focused on finding ways to appropriately baseline \"normal\" traffic for a given network, since even normal traffic can be highly variable, and a poorly tuned intrusion detection system will generate many false positives that are expensive to investigate.  23  Unfortunately, although anomaly-based detection can be highly effective when tracking an individual machine or user, it often struggles to effectively identify suspicious behavior across a network. In addition, a long-standing complaint regarding anomaly detection systems has been their tendency to generate many false positives, which speaks to the difficulty of defining \"normal\" traffic strictly enough to detect any anomalies but loosely enough that no legitimate behavior is flagged as anomalous.  24  Moreover, changes in an organization's standard procedures can dramatically undermine the usefulness of anomaly detection, at least until new baselines are learned-a lesson that many businesses discovered last year when millions of employees suddenly began working from home in response to COVID-19.  25  Because of the difficulties associated with anomaly detection, many organizations use it only as a complement to more standard misuse-based detection systems. This discussion illustrates that machine learning already has a long and multifaceted history in the field of intrusion detection. Different ML methods have been adapted to multiple types of intrusion detection, in a research process dating back over two decades. Moreover, empirical studies and the continued importance of ensemble models speak to the fact that newer innovations have not fully displaced these older models. As with spam detection, it would be a mistake to think that the developments of the past decade of ML research, specifically the rise of deep learning, have entirely transformed intrusion detection-though in Section 3, we will return to this topic to discuss some of the ways in which intrusion detection has meaningfully been changed by newer ML innovations. \n MALWARE DETECTION While intrusion detection systems monitor a system or network's behavior to identify signs that a network is under attack, malware detection systems examine specific files to determine if they are malicious. The simplest forms of malware detection found in early antivirus products would monitor machines for specific indicators of compromise, such as exact filenames or signatures (which are specific sequences of bytes or strings taken from the contents of a file). By maintaining long lists of malware signatures and conducting regular scans, these antivirus products tried to determine if any files on a machine were associated with these known definitions. Unfortunately, these detection methods can be easily evaded by polymorphic or metamorphic viruses-types of malware that change their own code each time they propagate-thereby ensuring that different versions will have different signatures.* By some estimates, in 2018 up to 94 percent of malicious executables exhibited polymorphic traits.  26  While traditional detection techniques can still be adapted for the detection of polymorphic or metamorphic malware-for instance, by looking at the sequence of actions the malware will take when executed rather than by matching based on raw code-these traditional methods become increasingly complex and computationally intensive as attackers improve.  27  Machine learning, however, excels at identifying shared features between samples that can't be classified using simple rules. As early as 1996, researchers at IBM began to explore the use of neural networks to classify boot sector viruses, a specific type of virus that targets a machine's instructions for booting up.  28  Additional research throughout the early 2000s explored the use of statistical models and standard ML classifiers for detecting malware.  29  Recent years have seen an explosion of interest in malware detection methods that utilize newer, deep learning-based approaches, methods which come with many advantages, such as the ability to extract relevant features from raw data with less human guidance.  30  Yet, as with intrusion detection, some experimental results indicate that-at least on some datasets-decades-old ML classifiers remain on par with more advanced deep learning methods when trained on the same data.  31  Notably, many techniques that do not use machine learning have remained effective at detecting and analyzing malware, despite the rise of polymorphic and metamorphic viruses. For instance, cyber defenders may execute unrecognized files in sandboxes-isolated virtual environments where a file can be allowed to execute without any ability to interact with real systems and data.  32  The sandbox collects information about the file, such as what kinds of functions it tries to execute, to determine if it is malicious. This method allows an antivirus product to detect polymorphic or metamorphic code without relying on machine learning and underscores the fact that machine learning is by no means the only way for cyber defenders to respond *Polymorphic code can refer to any type of code which appears in multiple forms, for instance, because attackers have created multiple variants or because the code can encrypt itself using a different key each time it propagates. By contrast, metamorphic code actually rewrites its underlying code while preserving the same functionality. to the intelligent evolution of cyber attacks. Even sandboxes, however, can be augmented with machine learning to identify files that resemble past malware but that do not necessarily try to execute the exact same behavior.  33  Throughout this section, we have emphasized two major points. First, there is a multi-decade-long history of researchers applying traditional ML techniques to major cybersecurity tasks, though with a very heavy emphasis on detection tasks. And second, though more powerful methods exist today, they typically represent natural evolutions from more traditional approaches. These facts are important to keep in mind when determining just how \"transformative\" recent ML advances will be for cybersecurity. lthough many decades-old ML approaches in cybersecurity remain competitive with more recent algorithms, the recent spike in machine learning interest is driven by some genuinely impressive breakthroughs. In recent years, AI innovations have led to self-driving cars, accurate language translation, and better-than-human video game playing. Although the possibility of transferring these successes into the cyber domain is uncertain, there is strong reason to think that-could this transfer be achieved-machine learning might become a major aid for cyber defenders. This section explores how newer ML architectures might be applied to cybersecurity. It is organized around the four-stage model of cybersecurity introduced earlier and discusses potential ML applications for each stage in turn. As it proceeds, it pays particular attention to four types of ML methods that have been responsible for many of the AI breakthroughs of the past five years: 1) deep learning; 2) reinforcement learning; 3) generative adversarial networks (GANs); and 4) massive natural language models.* This section assumes a basic familiarity with these concepts; readers without a background in ML should refer to the appendix for a brief description of each type of model. Of these breakthroughs, the development of deep learning systems has in many ways been the most fundamental. It is the combination of deep learning architectures with other types of approaches-such as GANs, reinforcement learning systems, or natural Cybersecurity and the Cutting Edge of AI Research 3 A * Cutting-edge language models are impressive in their own right, but they have also drawn attention to the applications of simpler forms of natural language processing. Many of the natural language tools we discuss in this section are not particularly complicated, but they do represent a growing interest in the question of how linguistic data might be better leveraged for cyber defense. language models-that has enabled most of the progress of the past half-decade, as visualized in Figure  2 . The Relationship Between Cutting-Edge AI Architectures At the beginning of our discussion of each stage, we provide a graphic listing a few potential ML tools that could be leveraged by defenders at that stage. These graphics also list the type of ML architecture that implementations of each tool might rely on and a rough assessment of how significantly ML can transform the task itself. All of these assessments are meant to be very general and draw from the discussions that follow. \n PREVENTION FIGURE 3 AI Applications for Prevention* *NLP refers to natural language processing. This figure lists only the most common and most specific underlying technology for each application; a deep learning-based GAN, for instance, would be simply listed as \"GAN.\" The first stage of our cybersecurity model is prevention-the work that defenders do to find and patch vulnerabilities in order to eliminate potential weaknesses. There has long been interest in building tools that can autonomously find and patch new vulnerabilities, but machine learning has only recently emerged as a plausible way of doing so. As recently as the 2016 DARPA-sponsored Cyber Grand Challenge, the most promising routes for automated vulnerability discovery still relied on carefully coded systems without the use of ML methods. Though several teams at that competition attempted to use machine learning to identify software vulnerabilities, Mayhem-the winning program-ultimately made no use of ML in its vulnerability discovery modules. Instead, Mayhem used two more common vulnerability discovery methods to identify potential weaknesses. First, a symbolic execution engine analyzed programs by building up representations of their execution logic, thereby learning how potentially crash-inducing behavior could be triggered. Second, it made use of fuzzers, which feed many modified and semi-random inputs to a program to determine if they cause any unexpected states that can be exploited.  34  In Mayhem, both of these elements were coded using traditional, rules-based systems, which were combined with modules that could autonomously write and deploy patches for discovered vulnerabilities. The Cyber Grand Challenge resulted in several significant innovations in the field of automated vulnerability discovery. That no team at the time made heavy use of machine learning demonstrates that machine learning is far from the only way to build autonomous cyber tools.* At the same time, ML research has made significant strides since 2016, and experts disagree about whether participants in a new Cyber Grand Challenge would find more use for machine learning today.  35  Consider fuzzers. In recent years, researchers studying fuzzers have increasingly explored deep learning as a means to more efficiently learn from successful inputs and find a larger number of vulnerabilities over time.  36  Deep learning-based fuzzers are often more efficient than older models; as one example, the deep learning-based program NeuFuzz was able to find three times as many crash-causing inputs as a modified version of a leading open source fuzzer across a variety of file types.  37  Outside of academia, Microsoft has also studied the use of deep learning to augment its own fuzzers.  38  Fuzzers look for vulnerabilities in code, but cyber defenders can also use penetration testing (or pentesting for short) to look for publicly known vulnerabilities and *Mayhem did use machine learning in other areas. For instance, the team found that machine learning could be used to generate realistic-seeming attack traffic, which could mislead other teams and cause them to waste resources. While intriguing, this application has less relevance outside of the 2016 competition, with its unique scoring system and capture-the-flag setup. insecure configurations in networks. In a pentest, experienced hackers systematically probe a network for vulnerabilities to identify potential weaknesses. For large organizations, pentests can be expensive, potentially costing tens of thousands of dollars and consuming weeks of employee time. Automated tools such as the opensource program Metasploit can partially offset these costs, but these tools simply run through a list of pre-selected, known exploits to determine if any machines on a network are vulnerable to them; they are not capable of strategizing how they use resources.  39  Recently, researchers have studied how reinforcement learning can allow cyber defenders to build AI agents that conduct pentests more strategically.  40  Some researchers have demonstrated that reinforcement learning-based agents can devise plausible strategies for a variety of capture-the-flag style simulations, and can do so reasonably quickly if they first watch a few examples of human-led pentesting.  41  These types of reinforcement learning-based approaches allow automated tools to search a network for vulnerabilities much faster than rules-based pentesting tools like Metasploit.  42  These successes require caveats, especially due to the high computational cost of reinforcement learning in complex environments. Studies exploring the use of reinforcement learning for pentesting typically rely on small environments-often simulated networks of around ten machines-with a limited number of exploits provided to the program. As either the complexity of the environment or the number of actions available to the program increases, reinforcement learning can quickly become computationally prohibitive.  43  This problem is difficult but not fully intractable: in other contexts, researchers have developed models that can efficiently narrow down the number of options that a program must consider.  44  If researchers could develop computationally feasible methods of simulating complex networks and efficiently choosing among many options, reinforcement learning programs could become important aids for pentesters, just as Metasploit was adopted as a major tool in the pentester's toolkit in previous decades. But it is also worth emphasizing that of all the technologies we discuss in this paper, this is the one that is perhaps most obviously of interest to attackers as well-indeed, attackers have often been observed using pentesting tools developed for legitimate use to instead compromise their targets' networks.  45  Beyond finding vulnerabilities with autonomous fuzzers and pentesters, machine learning may soon be able to provide tools that can help defenders allocate their time and attention to the most pressing vulnerabilities. Although there are many other ways that machine learning could be adapted for preventative purposes, we mention two other applications here: analysis of new bug reports and severity assessment for identified vulnerabilities. Since at least 2005, researchers have explored ways in which machine learning can be used to automatically analyze bug reports. ML systems might, for example, direct bugs to the software engineer most able to address them, label malware characteristics based on human-generated reports, or predict the likelihood that a new bug could present a security vulnerability.  46  Though much of the research in this area has yielded results that are insufficiently accurate for operational use, Microsoft recently demonstrated that machine learning can successfully use bug report titles to classify bugs as security-relevant or non-security-relevant. The Microsoft model was trained on more than a million actual Microsoft bug reports, suggesting that some automated forms of bug report analysis may be feasible for software companies with access to large amounts of training data.  47  Once vulnerabilities have been identified, defenders have limited time and resources to patch them. One prominent-though somewhat controversial-metric that helps defenders determine the severity of a vulnerability is the Common Vulnerability Scoring System (CVSS), which relies upon expert analysis to assign severity scores to new vulnerabilities. Some researchers think that machine learning and text mining could be used to automate the process of assigning CVSS scores by intelligently interpreting vulnerability descriptions.  48  Other researchers have used data about attacks observed in the wild to build machine learning systems that predict the likelihood of some vulnerabilities actually being exploited.  49  There is some evidence that when these machine learning-based risk assessments are used in conjunction with the CVSS, organizations can achieve similar levels of risk remediation for significantly less effort by prioritizing their remediation efforts on the vulnerabilities that are most likely to be exploited.  50  Taken together, the use of machine learning for improved fuzzing, pentesting, bug report analysis, and severity assessment could allow organizations to improve their ability to identify and prioritize vulnerabilities. \n DETECTION FIGURE 4 AI Applications for Detection Detection remains, at least among many popular-facing venues, the key place where deep learning and newer ML methods are thought to be potentially transformative forces.  51  Unfortunately, as of yet, deep learning has not provided the revolutionary breakthroughs that many have hoped for. While sufficiently large models do tend to perform incrementally better than simpler models-especially at large enough scales-these improvements are at times outweighed by the growing number of threats facing most organizations. The bottom line: despite the fundamental role that deep learning has played in the ML advances of the last half-decade, many cybersecurity companies continue to make significant use of simpler models today. One problem with ML-based detection, sometimes overlooked in popular coverage, is that ML systems are vulnerable to several classes of attack that do not apply to other types of detection systems. The process by which many ML systems reach decisions can often be poorly understood and highly sensitive to small changes that a human analyst would view as trivial, which often makes it possible for attackers to find \"adversarial examples\"-slightly altered inputs that dramatically change a model's response despite being undetectable to a human.  52  The use of ML models also opens up new avenues of attack: the model itself must be kept secure, but defenders must also make sure that their data is not poisoned and that the (typically open source) algorithms and statistical packages they use have not been tampered with.  53  In addition, while machine learning is sometimes presented as an objective process of \"learning patterns from data,\" in reality the design of ML systems is often the result of many judgment calls. In one somewhat infamous example, the security firm Cylance deployed an ML-based malware detection product, only for a group of white hat hackers to discover that if a short 5MB string were attached to any malicious file, the file would bypass the detection system nearly 90 percent of the time.  54  It seems that Cylance had built an ML malware detection system that worked relatively well, only to discover that it also blocked a number of legitimate games from being downloaded. In response, the firm added on a second ML system that calculated how similar a file was to any file on a whitelist of approved games, and which let high-scoring files through the system, even if the first ML model had identified them as potentially malicious. It was that second model that the white hat hackers were able to target: by adding on a string of code from the popular game Rocket League, they were able to create malware that Cylance's detection system failed to flag. This anecdote is not necessarily a story about ML's flaws so much as it is a story about the difficulties of implementing ML systems in real-world environments. Cylance's core malware detection algorithm seems not to have been vulnerable to this \"universal bypass\"; rather, it was the designers' decision to add on a second component in order to avoid blocking popular video games that introduced the vulnerability. Ultimately, the use of machine learning can bring several types of vulnerabilities, whether because those weaknesses are inherent to the ML approach itself (as is the case with data poisoning threats and adversarial examples), or because wrangling the ML model into a deployable product leads developers to inadvertently add in further vulnerabilities. None of this means that newer ML innovations are unimportant at the detection stage. For one, while older models can typically only provide rudimentary assessments of the severity of different types of alerts, deep learning can allow more sophisticated forms of analysis, which can help defenders prioritize more serious threats. This would provide an enormous benefit to cyber defenders, because many detection systems today generate so many false alerts that it is nearly impossible for human analysts to investigate them all. Some organizations are already making use of tiered systems, in which deep learning models analyze alerts from first-level detection systems to determine which potential threats require the highest priority.  55  These types of systems are most useful for large organizations, since they require not only a great deal of data about \"normal\" traffic, but also many examples of unique attacks. In addition, while adversarial examples are specifically crafted to fool ML models, the rise of other types of ML systems-most notably GANs-that can generate synthetic and realistic-looking data poses a threat to both ML and rules-based forms of attack detection. Researchers have already demonstrated that GANs can develop attack strategies that slip by existing intrusion detection systems, whether or not they are machine learning-based.  56  But GANs can also help to harden ML systems against these types of more elaborate threats: when intrusion detection systems are trained using a GAN to anticipate potential adversarial attacks, their accuracy against adversarial attacks is greatly improved.  57  GANs can also augment existing datasets by generating new data that can be reincorporated into a classifier's training data. GAN-augmented datasets of spam and malware have been used by some research teams to create classifiers that perform better than those trained only on real-world data.  58  Even though we have been somewhat pessimistic regarding the potential of deep learning to transform detection capabilities, the rise of adversarial attacks and the ever-increasing scale of threats means that defenders cannot afford to ignore advances in machine learning. Because GANs allow attackers to subvert both MLand rules-based detection systems, doing nothing is not an adequate response, and the best defense will likely rely at least in part upon the use of ML methods. And yet, even here it is important to emphasize that some researchers do not believe that the ML field has made meaningful progress in making ML systems robust to adversarial examples.  59  While adversarial training remains critical for defensive systems, it may not be enough to effectively counter the rising threat of adversarial attacks-another reason to be skeptical that machine learning can provide a silver bullet for cyber defenders. \n RESPONSE AND RECOVERY While there is a great deal of research on ML-driven detection systems, more ambitious proposals posit AI systems that might one day move through networks autonomously, patching vulnerabilities and fighting dynamically against attackers. After all, in recent years researchers have developed ML systems that can master competitive games as diverse as Go or StarCraft.  60  But as complex as these games are, they are comparatively more structured and easier to model than the nebulous world of cybersecurity. Furthermore, in most of the other stages of cybersecurity, ML progress has come from automating discrete and self-contained (yet increasingly complex) tasks, rather than from attempting to build ML systems that can operate fully autonomously. Response and recovery, by contrast, is a dynamic and continuous process that is not as easily broken into discrete components, which makes it much more difficult to build ML tools that can adequately automate human decision-making. In recent years, some authors have begun to identify more targeted roles that AI/ML systems could play in the response and recovery process. A 2020 report from the National Science & Technology Council, for instance, identifies at least two  concrete ways in which AI could aid the response and recovery process: by accurately categorizing ongoing attacks and selecting an appropriate initial response strategy, or by automating the decision to isolate machines from a network or impose user restrictions to contain infections.  61  The first type of tool could allow cyber defenders to automate their initial responses to a wide variety of common types of attacks, while the second type would be generally useful in containing security breaches from spreading to other parts of a network.* Despite being identified as a potential area of opportunity, we are not aware of any significant research demonstrating the viability of a ML system that could automate a wide range of responses to multiple types of attacks. † In part, this absence is likely due to the fact that such a goal remains too broad to be effectively automated by current ML methods. At the same time, there have been some attempts to further break this goal down, with the intention of building ML systems that can intelligently respond to specific types of attacks. As perhaps the most high-profile example, DARPA is currently sponsoring a project, ongoing since 2017, to create autonomous systems that can respond to botnets. If successful, this program would result in autonomous systems that can identify botnet-infected machines, select publicly-known vulnerabilities to deploy against them in order to remove malware, and move laterally to neutralize adjacent compromised machines.  62  There has been more progress in building ML models that can learn to isolate potentially compromised machines from a network in order to contain a security breach. This type of application is an example of moving target defense, a defensive strategy in which defenders try to impede attackers by dynamically restructuring parts of their IT infrastructure in response to an attack. Reinforcement learning in particular has been useful in exploring this topic. Several researchers have attempted to model network systems under attack in order to test the ability of reinforcement learning agents to respond. In one example, researchers allowed a defending agent with imperfect knowledge of an attacker's actions to reimage-or restore to factory settings-machines which it suspected may have become infected. By imposing a cost upon the defender for each reimage it made, the researchers demonstrated that a reinforcement learning-based defender could learn the appro-*In Figure  5 , we refer to the first application as \"adversary engagement,\" which would be the ultimate goal of an ML system that can understand ongoing attacks and respond accordingly. The second goal is a simple example of moving target defense. † Here it is worth again reiterating that the focus of this paper is specifically on machine learning and not on other types of technological advances that might be called \"AI.\" There have been more efforts to address this sort of problem using non-ML methods, but we do not comment on the success of these efforts here. priate cost-benefit tradeoff and make useful decisions regarding when to reimage.  63  Similar approaches have allowed reinforcement learning agents to learn when to isolate potentially infected nodes within a constrained network or to develop game theoretic strategies for adaptively responding to adversaries.  64  These results are promising, because they suggest that machine learning could be useful for automating tactics like moving target defense or for providing responses to some types of threats, such as botnets. But it remains far from certain that such tools will become broadly useful. The examples mentioned above were the result of extremely simplified simulations of real networks, and as with pentesting, it is not clear that the complexity of the simulation can be reasonably scaled up without becoming too computationally intensive to allow for real-time decision-making. Moreover, machine learning continues to be most successful in contexts with relatively clearly-defined problems and outcomes. Although it may become increasingly useful as a way to implement specific defensive strategies in response to various threats, we do not foresee it becoming capable in the short-to medium-term of fully automating the work that human analysts do when investigating the sources of a compromise and formulating strategic responses. \n ACTIVE DEFENSE FIGURE 6 \n AI Applications for Active Defense Organizations with significant cybersecurity needs must often take proactive steps to shape their cybersecurity strategies in response to new threats. This report groups these actions under the broad framework of \"active defense.\" Though this term has a specific national security definition, it is here used analogously to the way it has been used by researchers at the SANS Institute: as a spectrum of activity ranging from annoyance to attribution to outright counterattacks.  65  This section sets aside the issue of outright counter-attack and instead focuses on three less legally fraught tactics: deception, threat intelligence, and attribution. The wide range of potential AI applications across these three tactics suggests that active defense may be one of the stages of our cybersecurity model with comparatively much to gain from recent ML breakthroughs.  \n NLP \n Deception One of the most basic active defense measures available to cyber defenders is deception: intentionally faking something to mislead and slow down attackers. Though this strategy seems simplistic, it can have major operational benefits. In the lead-up to the 2017 French presidential election, for instance, the Macron campaign-aware that it was being targeted by Russian hackers-chose to forge internal emails full of misleading or outlandish content. Russian hackers could not easily distinguish real information from forgeries, and although they eventually chose to leak everything, this strategy effectively allowed the Macron campaign to minimize the public perception of the issue.  66  Deployed well, deceptive behavior can also cause attackers to second-guess themselves or otherwise manipulate their attention towards unproductive activities.  67  Generating realistic-looking documents or activity profiles is an area where machine learning excels. The rise of massive natural language models like GPT-3 will likely make the process of generating deceptive textual documents-like the Macron campaign's fake emails-relatively easy to automate. While this technology has an enormous potential for misuse, defenders could plausibly use it for deceptive purposes, generating waves of deceptive internal documents to mislead attackers. The use of other types of ML models-like GANs-could be used for non-textual deception. For instance, researchers have shown that it is possible to use a GAN to generate realistic network data that might be used to confuse attackers.  68  (An analogous type of forgery was used by the winning team of the 2016 Cyber Grand Challenge, which used machine learning to create traffic that looked like an attack and fooled opposing teams into wasting time trying to stop it.  69  ) Beyond forgeries, the honeypot-a security mechanism designed to lure cyber attackers to attractive-seeming but fake targets-is a long-standing deceptive tool for network defenders. Honeypots can offer \"bait\" in the form of fake files and data, run scripted interactions to thwart adversaries, or mislead attackers into thinking that they have been more successful than they have been. However, honeypots often require manual configuration, deployment, and maintenance to look like realistic production networks long enough to keep the attackers engaged, which can be costly and makes them difficult for smaller organizations to maintain. Since at least the early 2000s, researchers have explored how machine learning could be used to create more realistic and dynamic honeypots, including through the use of relatively simple clustering methods.  70  In more recent years, researchers have experimented with reinforcement learning-based honeypots that can learn to engage attackers for as long as possible, while also ideally tricking them into downloading custom software onto their own machines.  71  It would be premature to say that the use of reinforcement learning is a game-changer in this area, but it may at least make these types of honeypots more accessible to a larger number of organizations by automating much of the complexity of maintaining a dynamic honeypot. \n Threat Intelligence Gathering threat intelligence about potential adversaries can theoretically allow defenders to anticipate attacks and build stronger defenses, though for most organizations, the labor-intensive nature of collecting, processing, and analyzing threat intelligence makes it cost-prohibitive. But because finding patterns in data is a core strength of machine learning, some researchers have explored how ML and text mining can be used to improve threat intelligence analysis. For instance, ML methods can be used to cluster dark web users, or text mining methods could be leveraged to automatically collect, classify, and analyze posts on dark web forums and marketplaces, allowing researchers to identify zero-day exploits before they are deployed.  72  It is not difficult to imagine a fully automated ML system that could be given a company's name or a list of its products and told to search for mentions of potential vulnerabilities on the dark web, analyze them, and generate a report describing anticipated vectors of attack. Other tools introduced as deceptive tactics could be repurposed to collect threat intelligence about potential adversaries. Data collected from honeypot systems, for instance, can be analyzed and collated to determine if an organization is facing a persistent threat actor, which can then inform security strategies.  73  Even something as simple as a phishing email could become an opportunity to collect threat intelligence about an adversary. Some researchers have proposed systems that can produce misleading responses to phishing emails to elicit information from attackers, for instance by appearing willing to give money to a scam but asking for the attackers' routing numbers, which can then be traced and identified.  74  In fact, this is a major line of research within the Active Social Engineering Defense initiative at DARPA, which is currently attempting to build systems that can detect social engineering attacks and respond autonomously to trick adversaries into revealing identifying information.  75  These tactics cross the line between threat intelligence collection and outright attribution, the third active defense tactic discussed in this section. \n Attribution When facing an attack from an advanced persistent threat, federal agencies and cybersecurity companies will typically make an effort to attribute the attack to a specific adversary. Doing so enables federal actors to determine what type of response is warranted, and it helps cybersecurity companies determine the likely motivations of hostile actors in order to best defend their customers. Other companies are less likely to engage in attribution work. While there is some strategic value to being able to share precise information about adversaries with other organizations, the fact that most corporations cannot legally respond in kind to an adversary severely reduces the value of a successful attribution, particularly considering how difficult the process can be. When trying to attribute multiple attacks, the simplest approaches rely on manual analysis of attack indicators obtained through network logs, firewalls, or honeypots. However, because most persistent cyber attackers can avoid leaving obvious patterns, attribution also relies on human analysis and an understanding of potential adversaries and their overarching goals; attribution therefore relies not only on technical but also strategic indicators and, potentially, knowledge about relevant geopolitical situations. This is a level of analysis that machine learning is not well-equipped to carry out, so it remains unlikely that machine learning will be able to successfully automate the full process of attribution. Nonetheless, machine learning may still be able to assist in the attribution process. For example, if research analysts first extract high-level descriptions of an attack-including the tactics used, the country of attack, and so forth-ML methods may be able to cluster this information to identify similar attacks.  76  Natural language processing may help attribution systems automatically extract attribution-relevant information from blogs, research papers, and incident response reports, reducing the amount of work that humans need to do to make a successful attribution.  77  Applying ML methods to an attacker's source code, when available, may offer a different path to attribution. ML methods have already shown the ability to distinguish and de-anonymize authors based upon their distinctive writing styles.  78  In recent years, some researchers have demonstrated that human generated code works in a similar way: different programmers make unique choices as they code, which allows machine learning to create digital fingerprints that can be used to identify coders.  79  By comparing to databases of malicious code, machine learning may soon be able to cluster bits of code written by the same coder, providing one more clue in the attribution process. Attackers might always further muddy the waters by repurposing code from other sources, and the fact that advanced persistent threats typically contain numerous independent coders makes digital fingerprinting difficult. But the possibility is nonetheless intriguing. Some of these applications are speculative. There are notable limitations with relying on ML techniques for attribution; for example, threat data can often be inconsistent if it is collected from multiple security companies. Still, integrating high-level cyber threat behavior with machine learning is an understudied area, and further research may allow federal agencies and cybersecurity companies to augment their attribution toolkit, while also making some simpler forms of attribution available to other types of organizations. \n Conclusions his report has summarized a range of potential ways AI could be used to aid cyber defenders as well as some of the limitations of current techniques. We offer four major conclusions for what this means for the actual practice of cybersecurity: 1. At the detection stage, ML will allow for continued improvements over existing detection tools, especially for companies with extensive big data capabilities. However, these developments should be viewed as incremental advances that introduce new attack surfaces of their own. 2. At the prevention, response and recovery, and active defense stages, the use of ML is less commonplace but is making progress. Some of the most exciting advances will still require major ML breakthroughs in order to be fully deployable. 3. Overall, we anticipate that gains from ML will continue to be incremental rather than transformative for the cybersecurity industry. Barring major ML breakthroughs, the most transformative impacts may come from ML's ability to make previously un-or under-utilized defensive strategies more accessible to many organizations. \n 4. At the present time, machine learning is unlikely to fundamentally shift the strategic balance of cybersecurity towards attackers or defenders. However, for specific applications, the introduction of machine learning techniques may be more useful to one side or the other. \n T In focusing on the underlying technical capabilities of machine learning, this report has set aside many other issues, such as whether or not defenders will successfully implement ML systems and fully leverage their potential. These conclusions are preliminary, not predictive: whether they hold true will depend on how attackers and defenders choose to invest in and deploy this technology. \n ML CAN IMPROVE DETECTION, BUT IT BRINGS RISKS OF ITS OWN Machine learning has played an increasingly large role within attack detection for several decades, and the coming years will likely make it even more important. Among other things, this is due to the development of better algorithms and the use of greater computational power. But classification-whether of specific attack patterns or \"anomalies\" more generally-is a task where quality data makes or breaks the model. Because of the critical importance of data for training classification algorithms, the benefits of improved detection systems will be most easily leveraged by companies with the ability to collect the most cyber data. This is one area where defenders have an asymmetric advantage relative to attackers: defenders can collect and store far more data about their own networks than attackers can, which makes it possible for defenders to continually improve their defenses. Nonetheless, these benefits are likely to be incremental for three reasons. First, because the size and scale of cyber attacks continues to rise, it is unclear how important the gains from better ML models will be. Slow and steady improvement in algorithms and data collection abilities may only be enough to keep pace with the threat landscape, without allowing defenders to get ahead. Second, attackers have an asymmetric advantage of their own: an attack only has to work once, while a defensive strategy has to work all the time. Considering that attackers can use GANs to generate malicious traffic that resembles benign traffic enough to evade many types of ML-powered and rules-based defenses, the defender's data advantage may not be enough to protect them from the offensive use of ML methods.  80  Further, because most network data tends to be highly variable even under normal conditions, building algorithms that can reliably identify malicious anomalies without flagging legitimate behavior is a difficult task. This increases the chances that motivated attackers-especially those using heuristic or ML tactics to camouflage themselves within a network's background traffic-will continue to find ways to slip past even the best detection systems. Finally, the benefits of newer ML-based detection systems may be less long-lasting than many hope, since ML models bring new attack avenues of their own. These attacks can be internal to the ML algorithms themselves, which rely on datasets that attackers could poison and which attackers can often circumvent using subtly altered adversarial examples.  81  Moreover, the decisions of ML models can also be opaque, and when they behave in strange or unproductive ways, defenders may be tempted to make adjustments that introduce additional vulnerabilities, as appears to have been the case in the Cylance model discussed above. None of this is meant to trivialize the important role that machine learning can play in improving detection systems. Even if machine learning cannot vastly improve baseline detection abilities, deep learning models in particular may be useful for performing rudimentary triage of alerts in order to identify the behaviors that are most likely indicative of malicious activity. But most of these developments represent incremental improvements that require the context of the above caveats. While the developments are important, we do not view them as likely to significantly transform the nature of threat detection. \n PROGRESS EXISTS AT OTHER STAGES BUT MAY STILL AWAIT MAJOR BREAKTHROUGHS While machine learning has already been widely deployed for various types of detection, it remains rarer in the other stages of the cybersecurity model. Nonetheless, researchers have slowly been exploring the use of machine learning for a diverse range of applications. Table  2  summarizes each of the major ML applications for each stage that this report has covered. At the prevention stage, ML systems could one day allow for improved fuzzing systems that software developers can use to detect vulnerabilities before they are exploited, or for pentesting aids that can more effectively search a network for vulnerabilities than current tools can. ML systems may also allow defenders to more accurately identify potential threats by triaging bug reports and attempting to automatically predict the potential severity of a newly discovered vulnerability. At the response and recovery stage, there has been some progress in building reinforcement learning systems that are capable of learning when to sequester potentially compromised machines from a network in order to limit an attacker's lateral movement. There has also been growing interest in the use of machine learning to automate defensive strategies, as demonstrated by DARPA's ongoing project to build autonomous systems that can neutralize compromised devices within an attacking botnet. At the active defense stage, there is a broad and increasing number of applications for which machine learning could be adapted to help defenders deceive, analyze, and attribute attackers. Dynamic honeypots, deceptive document generation, and phishing response may soon leverage reinforcement learning and/or advanced natural language models. In addition, advances in natural language processing may improve threat intelligence collection from dark web sources. Finally, clustering methods and advanced code analysis may allow for more sophisticated attribution methods, including some that may become available to organizations that have not traditionally invested in attribution work. Major advances within these applications could meaningfully transform cybersecurity by automating many tasks that typically consume a cyber defender's time or by adding new streams of information and analysis to the defender's situational awareness. Nonetheless, these advances may still require significant breakthroughs of uncertain likelihood. For instance, the computational intensity of reinforcement learning in highly complex environments continues to make it difficult to build ML systems that can handle the complexity of cybersecurity networks. Building deployable reinforcement learning-based cyber responders that could fully mimic human expertise would require major breakthroughs in the ability of researchers to effi-  ciently represent all relevant network data. Machine learning will likely become an increasingly important part of each of these stages in the cybersecurity model, but transforming the work of the cyber defender may still be out of reach for the near-to medium-term. \n EXPECT ML GAINS TO BE INCREMENTAL RATHER THAN TRANSFORMATIONAL Machine learning can offer a great deal to cyber defenders across all four stages of the cybersecurity model used in this paper. Nonetheless, whether or not it is a truly \"transformational\" technology depends very heavily on what standards are used to assess its impact. There is no doubt that machine learning can make significant improvements on a variety of cybersecurity technologies. If it were not for machine learning, defenders today would likely be consumed with low-level analysis that is more often delegated to automated systems which frequently make heavy use of ML systems. This is especially true of a task like spam detectionwhich would likely be all but impossible at scale without machine learning-but it is also true of intrusion and malware detection. By counterfactual standards, then, machine learning is indeed a transformative technology: without it, cybersecurity today would have a very different-and very worse-record of success. But this does not mean that machine learning is a truly transformational technology by historical standards. While machine learning successfully automates many low-level tasks today, the rapidly growing number of threats and sophisticated attacks facing most organizations means that ML-based Major advances within these applications could meaningfully transform cybersecurity by automating many tasks that typically consume a cyber defender's time or by adding new streams of information and analysis to the defender's situational awareness. Nonetheless, these advances may still require significant breakthroughs of uncertain likelihood. detection systems have mostly only allowed cyber defenders to keep pace with the threat landscape. So far, machine learning has not transformed the core type of work that cyber defenders are required to perform, nor has it made the current cybersecurity skills shortage significantly less worrisome.  82  By historical standards, we conclude that machine learning is not presently a transformational technology. This does not mean that ML is unimportant for cyber defenders. Continued investment in both researching and implementing promising ML-cyber tools should be an important area of focus for cyber defenders, and should be promoted by the federal government wherever possible. But policymakers should not anticipate that developing better ML tools is a replacement for addressing other, more fundamental cybersecurity problems, such as encouraging cybersecurity workforce development or enforcing secure practices across federal agencies. While we do not think that machine learning is poised to transform the cybersecurity industry, this requires two caveats: First, fundamental breakthroughs in ML research may significantly change the threshold of possibility. For instance, if researchers are able to develop computationally simpler methods of simulating complicated networks, reinforcement learning-based tools may be able to take on significantly more complex tasks than what is currently possible. While we do not currently anticipate machine learning being able to automate many high-level decisions during the incident response process, a sufficiently large breakthrough might change that assessment and enable ML tools that could be given substantially more autonomy. Second, not all the defensive strategies discussed in this report are equally common, which means that machine learning may be able to have a larger impact on some areas than others-especially where it can make previously un-or under-utilized strategies more accessible for more organizations. For instance, some survey results suggest that while upwards of 90 percent of organizations use antivirus products, the number of organizations actively using honeypots or related Policymakers should not anticipate that developing better ML tools is a replacement for addressing other, more fundamental cybersecurity problems, such as encouraging cybersecurity workforce development or enforcing secure practices across federal agencies. technologies may be much lower-potentially as low as 25 percent.  83  Improvements in the use of machine learning for malware detection are useful and important-but they are unlikely to change many organizations' underlying defensive strategies. By contrast, if ML tools make it possible to easily deploy honeypots that can realistically integrate into an existing network, many more organizations might use honeypots. This would in some ways represent a larger shift in the threat landscape than improvements in malware detection tools, because the widespread deployment of a new strategy could have more significant impacts on attackers' decision-making processes than an incremental improvement on an existing strategy. As another example, some researchers have explored how the presence of deceptive tactics such as decoys can disorient attackers and shape their strategies.  84  If ML makes it easy to automatically create deceptive internal records or engage in moving target defense, for instance, then widespread adoption of the technologies may meaningfully shape attacker strategies, even if the ML models themselves are not particularly impressive. The mere knowledge that such deceptive tactics are more commonly used may be enough to influence the attacker's decision-making. \n ML APPLICATIONS PRESENT NO CLEAR-CUT ADVANTAGE TO OFFENSE OR DEFENSE ON NET Most of the ML applications this report has discussed are dual-use: they could be used either by attackers or defenders. In some cases, the opportunity for attackers is obvious, as in the cases of ML-powered fuzzing and pentesting. In other cases, attackers might find creative ways to repurpose or circumvent defensive technologies. For instance, if machine learning is used to calculate the odds of an exploit being used, attackers might simply incorporate this information into their decisions regarding which exploits to develop, making the model increasingly useless as attackers adapt and the training data becomes unrepresentative of actual realities. Because of these dual-use considerations, ML will not give a clear-cut advantage either to attackers or to defenders in the coming years. However, while ML will not clearly benefit either attackers or defenders overall, certain specific applications of ML may be either offense-biased or defense-biased. For example, the rise of sophisticated natural language processing models is likely to improve spearphishing abilities.  85  With the ability to generate detailed, fluent, and personalized text or voice recordings, attackers will be able to engage in more effective social engineering campaigns. Although the same natural language processing models can be adapted for defensive purposes-say, to build chatbots that can mislead social engineers-these are reactive measures and should not distract from the important takeaway: barring major developments, machine learning will allow attackers to conduct more effective social engineering campaigns in the near future. Therefore, the use of machine learning for text generation is likely an offense-biased technology. At the same time, there is arguably a larger number of ML applications that may be primarily useful for defenders. For instance, attack detection is a major focus of ML research for cybersecurity, but attackers may have relatively little use for attack detection technologies, since there is little need to detect an attack that you yourself are responsible for.* Similarly, the use of dynamic honeypots are more obviously relevant for organizations seeking to protect their networks than for attackers hoping to penetrate them, which makes ML-based honeypots another generally defense-biased technology. These are technologies where defenders may generally gain more than attackers from increases in ML adoption. \n SUMMARY Policymakers and practitioners alike need to think about how machine learning can alter specific tasks within cybersecurity, rather than talking in general terms about how machine learning can alter cybersecurity as a whole. We have attempted to provide greater clarity by surveying a range of potential machine learning applications and the factors that continue to constrain their development. In addition, by focusing on the continuity in the use of machine learning for cybersecurity tasks dating back to the 1990s, we have aimed to demonstrate that newer breakthroughs will bring evolutionary cybersecurity gains rather than revolutionary changes to the industry. It is possible that one day, machine learning tools will automate entire swathes of our cybersecurity model at once, seamlessly performing the work of multiple stages instead of functioning as tools that merely aid in already-existing, well-defined tasks. But in the near-to medium-term future, this possibility is remote. It is far more pragmatic for policymakers and practitioners to work towards a more nuanced understanding of the types of tasks that could benefit from machine learning and the types of tasks that can't. Effective policy will need to take these nuances into account to promote useful types of research, enable organizations to take full advantage of new machine learning tools, and help defenders prepare against their constantly improving adversaries. *Attackers who study newer detection models may be able to find ways to craft adversarial attacks that can bypass them, so progress in this area is only strictly defense-biased if it results in detection systems that are less vulnerable to adversarial attacks. Since it is not clear that this is the case with newer ML detection models, it is unclear whether current improvements in ML detection systems are capable of shifting the overall strategic balance towards defenders. Deep learning is based on the use of artificial neural networks and may include hundreds or thousands of nodes whose structure can be adapted for specific tasks.  \n MASSIVE NATURAL LANGUAGE MODELS Reinforcement learning simulates an agent with several available actions, an environment, and a reward function and allows the agent to experiment with the use of different actions to learn strategies that maximize rewards. A GAN includes two neural networks: a generator that uses random input to create new inputs and a discriminator that attempts to distinguish real from generated inputs. These models are not unified by a single underlying structure but are instead characterized by their use of massive models that can include billions of parameters in order to learn patterns within human language. 1) Performance can often improve with more data, well after performance in simpler models flattens 2) Can be used as an underlying approach for most AI applications 3) Models can often automatically learn to extract relevant features from data 1) Requires significant processing power to run and significant memory to store due to model size 2) Widely regarded as one of the most opaque types of machine learning 3) Benefits of more data may be difficult to fully leverage in the cyber domain, which is often very secretive 1) Allows for AI agents that can learn strategic thinking 2) Can be combined with deep learning systems to improve performance 3) Does not require historical data if realistic environments can be simulated 1) Computational cost increases extremely quickly when the environment complexity or number of actions are increased 2) Difficult to simulate realistic yet computationally tractable cyber networks 1) Can be combined with deep learning systems to improve performance 2) Can be used to augment existing data sets or evade existing classifiers 1) Does not have general-purpose applications like deep learning 2) Typical caveats about the need for robust underlying data apply 1) Increasingly capable of automating text generation, summarization, and translation, among other tasks 2) Could be used to automate security tasks that involve written text (e.g. generating bug reports) 1) Cannot be built by individual companies without investing millions of dollars; access will depend on how corporate owners choose to make them available FIGURE FIGURE 1 \n FIGURE 2 2 FIGURE 2 \n FIGURE 2 The 2 FIGURE 2The Relationship Between Cutting-Edge AI Architectures \n FIGURE 3 AI 3 FIGURE 3 AI Applications for Prevention \n FIGURE 4 AI 4 FIGURE 4 AI Applications for Detection \n FIGURE 5 AI 5 FIGURE 5 AI Applications for Response and Recovery \n FIGURE 5 AI 5 FIGURE 5 AI Applications for Response and Recovery \n FIGURE 6 AI 6 FIGURE 6 AI Applications for Active Defense \n TABLE 1 A 1 Comparison of Our Model and the NIST Cybersecurity Framework NIST FUNCTIONS OUR MODEL Identify Prevention \n TABLE 1 1 \n TABLE 2 2 Summary of AI Applications PREVENTION DETECTION Fuzzing Accurate Detection \n TABLE 2 2 Summary of AI Applications RESPONSE AND RECOVERY ACTIVE DEFENSE Adversary Engagement Deceptive Document Generation Pentesting Alert Prioritization Moving Target Defense Dynamic Honeypotting Bug Triage and Classification Adversarial Hardening of Defense Systems Automated Phishing Response Vulnerability Severity Assessment Dark Web Threat Intelligence Collection Attack Clustering for Attribution Code De-Anonymization \n\t\t\t *Some researchers tend to define misuse-based detection more restrictively as being strictly based on curated lists of indicators of compromise rather than probabilistic models. Under this definition, ML cannot be used to perform misuse-based detection, since ML is by its very nature probabilistic. However, many ML classifiers resemble misuse-based detection systems in that they use labeled examples of previous attacks to detect future attacks, struggle to generalize to zero-day attacks, and do not seek to identify a baseline model of \"normal\" behavior. We therefore think it makes the most sense to categorize ML classifiers with these characteristics under the label of \"misuse-based detection.\"", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Machine-Learning-and-Cybersecurity.tei.xml", "id": "3ed20484f7b82cb976dc9852513ea4a1"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "The author would also like to thank Melissa Deng, Dale Brauner, and Danny Hague for editorial support. The author is solely responsible for any mistakes.", "authors": ["John VerWey"], "title": "No Permits, No Fabs The Importance of Regulatory Reform for Semiconductor Manufacturing", "text": "Executive Summary The ongoing global chip shortage, coupled with China's heavy investments in indigenizing semiconductor manufacturing capabilities, has brought attention to the importance of semiconductors to the U.S. economy, the fragility of semiconductor supply chains, and the decline of U.S. chipmaking capacity over the past three decades. In part as a result of this attention, Congress has advanced legislation to appropriate $52 billion in funding for the CHIPS for America Act. Approximately $39 billion will likely go toward incentivizing semiconductor manufacturers to build new chipmaking capacity in the United States. But more can be done to improve the resiliency of U.S. access to microelectronics beyond manufacturing incentives. This report outlines infrastructure investments and regulatory reforms that could make the United States a more attractive place to build new chipmaking capacity and ensure continued U.S. access to key inputs for semiconductor manufacturing. \n Key Findings: The United States currently builds fewer fabs * at a slower rate than the rest of the world. Part of the reason for this is permitting regulations which require long assessment timelines. Fabs have extensive infrastructure requirements, which interact with federal, state, and local regulations in complex ways. Modern fabs require access to (1) large plots of (2) seismically inactive land with a reliable, affordable, and stable supply of (3) water, (4) electricity, (5) talent,  (6)  transportation infrastructure, and  (7)  nearby land for co-locating with suppliers essential for constructing and operating a modern fab. The CHIPS Act correctly aims to increase the number of fabs constructed in the United States, but regulatory support and infrastructure investments are needed to ensure that these new fabs are built on time and on budget. The United States government should prioritize regulatory support at the local, state, and federal level to expedite fab construction. In particular, full implementation of several recommendations from the 2017 President's Council of Advisors on Science and Technology related to environmental review and permitting of high technology facility construction are essential.  1  The United States should also make infrastructure investments targeting utilities, transportation, and supply chain networks that will assist semiconductor manufacturers. Engagement with allies will be essential for increasing resilience in the semiconductor materials, gases, and chemicals supply chain. In the medium to long term, increasing domestic production and/or stockpiling should be considered. Increasing domestic United States production of many raw materials and chemicals is contingent on opening new mining and/or refining operations, which would require extensive permitting. Thus, coordination with allies who already have existing production and refining capacity may be more expeditious than attempting to establish new capacity in the United States. Though many of these materials have a limited shelf life, stockpiling of certain materials, modeled after existing United States government programs like those operated by the Defense Logistics Agency Strategic Materials, may be an option. The United States should quantify demand for key material inputs, identify potential alternatives, and support their development. Ongoing efforts supported by the Environmental Protection Agency and the semiconductor industry to develop substitutes for environmentally harmful greenhouse gases used in semiconductor manufacturing could serve as a template for further work to identify substitutes for certain materials, chemicals, and gases used in semiconductor manufacturing for which there is no commercially viable domestic supply. \n Introduction: Fab Infrastructure Requirements and Federal Permitting This paper, in concert with forthcoming companion papers on reshoring semiconductor manufacturing, argues that current semiconductor reshoring efforts should prioritize construction of leading-edge fabrication facilities and increase the resilience of the associated supply chain necessary to support these facilities. The U.S. Department of Commerce's 100-day review of the semiconductor supply chain in response to Executive Order 14017 on Securing America's Supply Chains generally aligns with this argument, finding \"federal incentives to build or expand semiconductor facilities are necessary to counter the significant subsidies provided by foreign allies and competitors.\" 2 However, the United States has many regulations in place which may effectively counteract the purpose of CHIPS Act funds, slowing construction of new leading-edge fabrication facilities. In addition, the semiconductor industry must contend with myriad environmental, health, and safety (EHS) regulations that serve important purposes, but will inevitably slow the development of a more resilient semiconductor supply chain in the United States. Finally, the simple reality is that there are very few leading-edge semiconductor manufacturers in the world, and most of them are headquartered outside the United States. In practice, this means that the United States must craft policies that convince specific foreign companies to build outside of their headquarters country, where presumably they face significant political pressure to build domestically and enjoy easier access to policymakers to facilitate build-out in regulatory environments they can navigate adeptly. Local, state, and federal permitting processes in the United States are beneficial to the general public but present tradeoffs for semiconductor manufacturers. A 2017 report from the White House found that these permitting processes \"minimize[d] environmental and community impact, which companies are not always economically incentivized to do. The combination of the current Federal and state permitting and review processes, however, can be slow, unpredictable, and lacking in transparency.\"  3  The report goes on to note that semiconductor factories are particularly vulnerable to permitting-related delays due to their long construction times, significant geographic footprint, and complex supply chains. In addition, due to the many specialized chemicals and gases used in the semiconductor manufacturing process, unique permits must be acquired, which can further delay, or even halt, operations. Adding to this challenge, the semiconductor industry's transition to a fabless-foundry operating model (in which semiconductor design and semiconductor fabrication are done by separate firms) in the past 20 years largely caught U.S. firms flat-footed. Many leading U.S. companies maintain an Integrated Device Manufacturer (IDM) operating model of doing both design and manufacturing in-house, an increasingly costly proposition, especially given the success of Taiwanese foundries that focus solely on manufacturing. In part as a result of both the U.S. regulatory environment and this ongoing structural shift in the industry, the United States builds fewer fabs at a much slower rate than other countries, and at a greater cost to companies. The United States should recognize the nature and value of indirect EHS regulatory support provided by other countries interested in attracting semiconductor manufacturers, and adopt policies that make it equally attractive to build fabs in the United States. This paper begins by reviewing fab rates of construction worldwide from 1990 to 2020, finding that the United States builds fabs far more slowly than other competitor countries and regions, notably Taiwan and China. Next, this paper reviews fabs' myriad unique infrastructure requirements. Finally, this paper discusses how these infrastructure requirements must contend with U.S. environmental, health, and safety regulatory permitting processes, potentially slowing the construction timeline for fabs and the associated supply chain necessary to support semiconductor manufacturing. \n The United States Builds Fewer Fabs More Slowly Than the Rest of the World It takes a long time (typically two to four years) to build a fab in any country. But analysis in this section shows that fab construction takes much longer in the United States than in the East Asian countries where most chipmaking currently takes place. There are many factors underlying the decline in U.S. chipmaking capacity, but one underappreciated factor may be the longer construction timelines associated with building new (\"greenfield\") American fabs. In part because of the unique infrastructure requirements of fabs and the regulatory processes these large construction projects must navigate, construction of semiconductor fabs takes several years. Between 1990 and 2020 there were approximately 635 greenfield semiconductor fabs built around the world. The average time between the construction start date and the beginning date of production was 682 days or roughly 1.86 years. This timeline does not include pre-permitting and pre-construction considerations, indicating that fab construction times exceed two years on average. There is considerable regional variation in the time it takes to build a new fab. As Figure  1  shows, Japan (584 days) and South Korea (620 days) build fabs significantly faster than the rest of the world on average. The Americas, of which the United States is the primary site of semiconductor fabrication facilities, build fabs at a significantly reduced speed, taking an average of 736 days, or roughly five months longer than Japan. Only construction of fabs in Southeast Asia takes longer than in the Americas. Until the recent slowing of Moore's Law, this industry introduced a new generation of chips every 12 to 18 months, meaning a five month delay in construction could be damaging to a firm's competitiveness. For context, in a period of five months, leading edge foundries like those operated by Samsung and TSMC could produce roughly 500,000 wafers. This trend is particularly troubling as the U.S. industry's position as a leading-edge manufacturer has been ceded to firms in Taiwan and South Korea. Both of these countries maintain the ability to build the world's most advanced fabs at rates that exceed the United States' ability to build trailing-edge fabs. Source: World Fab Forecast. Sample consisted of 635 greenfield fab construction projects with a \"construction start\" date between 1/1/1990 and a \"production date\" no later than 12/1/2020. The difference between \"construction start\" and \"production date\" was calculated for each fab project and then averaged by region. Figures  2-4  follow this methodology and further break the analysis out by decade. The United States' ability to expeditiously construct fabs has declined at the same time as the total number of fab projects in the United States has declined. Some of this is due to changes in the global semiconductor value chain, which has concentrated resources in Asia as foundries have risen in prominence, and countries like Taiwan, South Korea, and China have established significant market share in the industry from 1990 to 2020. However, during this same 30-year period, the time required to build a new fab in the United States increased 38 percent, rising from an average of 665 days (1.8 years) during the 1990 to 2000 time period to 918 days (2.5 years) during the 2010-2020 time period (Figure  2 ). At the same time, the total number of new fab projects in the United States was halved, decreasing from 55 greenfield fab projects in the <PHONE> time period to 22 greenfield fab projects between 2010 and 2020.  The decline in the total number of new fab projects in the United States, as well as the speed with which those projects are completed, is striking when compared to other countries. In China, for example, the total number of new fab projects has increased from 14 during the <PHONE> time period to 95 during the 2010-2020 time period (Figure  3 ). At the same time, China has seen the average number of days from construction start date to production date for these fab projects decrease from a high of 747 days (2 years) during the 2000-2010 time period to 675 days (1.85 years) during the 2010-2020 time period. China is building more fabs and building them faster.   Comparing average fab construction time across regions while controlling for the size of the fab construction project further illustrates the U.S. deficit in advanced semiconductor manufacturing. As Figure  5  shows, from 1990 to 2020, China built 32 fabs that produce 100,000 or more wafer starts per month (WSpM), while the rest of the world only built 24 during the same time period. * The United States had no greenfield fab projects that involved construction of factories with capacity greater than or equal to 100,000 WSpM during this time period. * High WSpM correlates strongly with advanced fabrication. The newest and most advanced fabs in the world can have WSpM greater than 200k. There are several important caveats to this data. First, this analysis is limited to greenfield projects and thus does not take into account fab expansion projects frequently undertaken by leading firms like Samsung and TSMC. Second, while China builds large advanced fabs faster than any other region in the world, most of these advanced fabs are owned and operated by non-Chinese headquartered firms. For example, Samsung and SK Hynix (South Korea) and TSMC (Taiwan) operate high volume fabs in China. This is also the case in Southeast Asia, where firms like Micron (USA) have built their most advanced and highest-WSpM fabs. The fact that some U.S.-headquartered firms are choosing to build their most advanced fabs outside the United States speaks to the challenge U.S. policymakers face in crafting subsidies, regulatory support, and an infrastructure environment that is competitive with allied and adversary countries in Asia. \n Fab Infrastructure Requirements This section reviews some of the infrastructure constraints facing chipmakers deciding where to establish greenfield fabs. Modern fabs require access to (1) large plots of (2) seismically inactive land with a reliable, affordable, and stable supply of (3) water, (4) electricity, (5) talent, (6) transportation infrastructure, and (  7 )  \n Number of Fab Projects by Region China Japan Korea SE Asia Taiwan nearby land for co-locating with suppliers essential for operating a modern fab. These infrastructure requirements touch on construction and permitting processes in complex and sometimes costly ways. This section concludes with a discussion of infrastructure investments that other countries with strong semiconductor manufacturing industries have made, which could provide a model for future U.S. infrastructure investments. (1) Large plots of land. Fabs require large plots of land on which to locate their operations. While the cost of land is not as expensive an input in the semiconductor supply chain as the capital expenditures required to purchase and install semiconductor manufacturing equipment, it can present an obstacle to prospective semiconductor manufacturers. Fab \"shells,\" the physical structures that house semiconductor manufacturing equipment (SME) and all associated materials and supporting operations, can occupy hundreds of acres and account for between 20 and 40 percent of the total capital expenditures associated with a greenfield semiconductor manufacturing facility.  6  For example, Samsung's Austin, TX-based fab occupies 640 acres (1 square mile).  7  Semiconductor manufacturers frequently purchase more land than they initially need in anticipation of future expansion. (2) Low seismic activity. Fabs and the semiconductor manufacturing equipment inside them are extremely sensitive to ambient vibration, meaning that they must be located in regions that are not seismically active, and on plots of land relatively isolated from highways, airports, and rail links.  8  While there are mitigation options available, and the industry has pioneered novel seismic isolation techniques, these unique requirements further limit the overall geographic supply of ideal building sites. (3) Stable supply of water. Some estimates indicate a modern fab consumes around five million gallons of water per day. 9 This reliance on a consistent supply of water was recently highlighted in Taiwan, when a record drought forced the government to institute measures to ensure the country's semiconductor manufacturers had access to adequate supplies of water to continue operations. 10 This intensive water consumption frequently necessitates that individual companies take action. For example, TSMC recently announced plans to establish on-site water storage and treatment facilities to prevent potential disruptions stemming from droughts and impurities.  11  (4) Stable supply of electricity. Fabs rely on a stable electrical grid to sustain their 24/7/365 operations. As wafer processing has increased in complexity and automation, energy consumption has gone up proportionally. Leading edge fabs can now consume as much electricity in one year as 50,000 homes.  12  Once a fab has reached volume production, energy costs can account for up to 30 percent of the fab's operating costs.  13  For smaller countries that are home to significant fab operations like Singapore, Taiwan, and South Korea, these factory operations consume a notable percentage of overall electricity in the country. The OECD estimates that TSMC alone accounted for 7.7 percent of total industrial electricity consumption in Taiwan in 2017.  14  The cost of an electrical grid failure that disrupts fab operations is substantial. A February 2021 power grid failure in Texas shut down Samsung, NXP, and Infineon chip factories for several weeks, costing Samsung alone over $270 million.  15  (5) Talent. Fab sites need to be strategically located in an area nearby a skilled workforce. This has resulted in fab clusters in close proximity to university systems with a consistent talent pool and in relative proximity to metropolitan areas. In the United States, these clusters include California's Silicon Valley, New York's Tech Valley, and Oregon's Silicon Forest, all of which draw technical talent from nearby cities, universities, and firms.  16  (6) Transportation infrastructure. Transportation infrastructure also drives the cost of fab construction around the world. The vast majority of semiconductor-affiliated commerce occurs by air freight, necessitating that fabs be close to an international airport. However, there are some chemicals and materials used in the semiconductor production process which, due to their hazardous nature, low value, or extreme weight, ship via ocean freight. In addition, fabs need adequate last-mile road infrastructure to take delivery of particularly large pieces of semiconductor manufacturing equipment and associated material handling devices. For example, ASML indicates one of its Extreme Ultraviolet Lithography systems \"ships in 40 freight containers, spread over 20 trucks and three cargo planes.\"  17  This combination of air, ocean, and land transportation logistics presents myriad supply chain and regulatory bottlenecks. Specialty logistics services have been developed to facilitate semiconductor transportation. One supplier of transportation logistics services targeting the semiconductor industry advertises that they handle everything from SME and hazardous chemical delivery to air freight distribution of finished products (Figure  6 ).  18  Center for Security and Emerging Technology | 14 (7) Nearby land for co-location with key suppliers. Semiconductor fabrication facilities rely on a vast supply chain. Intel, for example, reports that it has identified over 16,000 suppliers.  19  However there are some suppliers whose products or services are so essential to the operations of modern fabs that they must co-locate production and warehousing facilities with the fabs they are supplying. Suppliers of rare gases and specialty chemicals used intensively in fabs need proximate land to establish production and warehousing facilities. Following the announcement that TSMC will be constructing a new fab in Arizona, several Taiwanese suppliers of specialty chemicals and gases to TSMC indicated they will be establishing operations nearby the new facility in addition to their current locations near TSMC's fabs in Taiwan.  20  In this instance, having additional land nearby prospective greenfield semiconductor fabrication sites is another important consideration. \n Government Policies to Reduce Infrastructure Costs and Accelerate Fab Construction Recognizing the costs that these infrastructure requirements impose on semiconductor manufacturers, many countries offer incentives to offset some of the price of constructing new factories. A joint report from the Semiconductor Industry Association and the Boston Consulting Group found that some countries provide both direct support to individual semiconductor companies and also incentives that are designed to facilitate the creation of a semiconductor ecosystem in close proximity to fabs.  21  The Japanese government has offered tax breaks, subsidies, and investments to attract semiconductor companies to establish joint ventures with Japanese firms, mirroring the sorts of incentives proposed by South Korea.  22  Similarly, the Taiwanese government's Ministry of Economic Affairs offers tax and tariff incentives as well as research and development (R&D) subsidies to attract semiconductor companies. Specifically, Taiwan maintains a business tax rate of 17 percent, allows semiconductor firms to credit up to 15 percent of their research & development expenses against their income tax bill annually, and permits firms to import semiconductor manufacturing equipment tariff-free.  23  Subsidies are also available for up to 50 percent of total spending by semiconductor firms who establish an R&D center in Taiwan. International semiconductor firms are clearly responsive to these incentives. The Taiwanese government notes these incentives have successfully attracted Micron, the U.S. memory chip company, to build a fab in Taiwan, while equipment suppliers like ASML from the Netherlands, as well as Applied Materials and Lam Research from the United States, have all set up R&D centers or training headquarters in Taiwan.  24  In addition, some countries focus on minimizing pre-and postfabrication operating costs for semiconductor manufacturers. A recent OECD report found that some governments support semiconductor manufacturers through the provision of water and electricity at below-market rates via state utilities.  25  Furthermore, the provision of land at below-market prices to semiconductor manufacturers was observed by the OECD to be a form of investment incentive. The OECD highlighted the case of Tsinghua Unigroup, a Chinese semiconductor firm which \"purchased land for its foundry in Chengdu for CNY 240 per m^2, while the official average price for industrial land in second-tier cities was CNY 724 per m^2.\"  26  The Government of Israel also made use of similar land-specific incentives which were used by Intel to expand its operations in the country.  27  Finally, many countries in Asia also provide infrastructure support in the form of utilities and logistics investments, as well as providing for expedited procedural consideration and eased regulations associated with semiconductor factory construction.  28  When Micron first considered establishing a fab in Taiwan, the country's investment authority \"assisted Micron in terms of land acquisition…accelerated the administrative process…eliminated investment barriers (such as coordinating underground pipelines and sidewalk construction)  [and]  organized job fairs to help the company recruit talent.\" 29 Taiwan has also established a series of free trade zones designed to facilitate efficient trading, warehousing, transport, and customs clearance processes that are critical to international semiconductor manufacturers. 30 The government of Singapore has provided myriad infrastructure investments designed to attract semiconductor manufacturers. Through government agencies like the Economic Development Board and JTC Corporation, Singapore has established four industrial estates that provide shovel-ready plots of land for semiconductor manufacturers and their suppliers that come preequipped with basic infrastructure like power, electricity and roads.  31  These estates also include ready-built facilities that feature chilled water, bulk industrial gas supply, high ceilings to accommodate SME, and incorporate vibration-control construction techniques.  32  The results of the Singaporean government's efforts are clear, having successfully attracted 14 global semiconductor firms employing 18,600 workers in the industry across these estates.  33  The value of this infrastructure support is not easily quantified but lowers the cost of doing business for semiconductor manufacturers and decreases construction timelines. And the success of this regulatory facilitation and infrastructure investment is clear: GlobalFoundries, one of the leading foundries in the United States, recently chose to substantially expand its semiconductor manufacturing facility in Singapore rather than doing so at its factory in New York.  34   \n Semiconductor Fabrication Regulatory Permitting Considerations Fabs' many infrastructure requirements touch on local, state, and federal EHS regulations. These regulations implicate agencies at each level of government, sometimes with overlapping jurisdictions. As a result, construction must carefully navigate arcane regulatory processes to develop greenfield semiconductor fabs. Recognizing the time delays that these regulations and permitting processes place on semiconductor manufacturers, other countries provide incentives and indirect subsidies to expedite fab construction timelines. To date, the United States government has not provided sufficient local, state, and federal regulatory support to match the efforts offered by peer semiconductor-producing countries. Regulatory support steps could include fully implementing 2017 recommendations from the President's Council of Advisors for Science and Technology, such as identifying where federal permitting regulations for high-technology facilities are redundant with state rules and might therefore be modified or removed.  35  The U.S. Environmental Protection Agency (EPA) could also consider creating a \"fast track\" process for preconstruction and operating permits related to the Clean Air Act (CAA). State and local environmental agencies perform a first review of these permit applications, but the EPA retains the right to review any draft permit and provide comments to state or local authorities. The EPA has experimented with the use of flexible air permits in the past and should create a program specifically tailored to the needs of U.S. semiconductor firms that would accelerate construction of new fabs and re-tooling of existing fabs. For example, following receipt of a flexible air permit, Intel Corporation used this permits' advance approval to make 150-200 equipment changes and process modifications to its Aloha, Oregon fab that would have otherwise required EPA New Source Review permits and resulted in concurrent permits from the Oregon Department of Environmental Quality.  36  This flexible permit allowed Intel to make changes to its fab without notifying environmental regulators, so long as the changes did not result in the fab's emissions exceeding previously agreed-upon levels. This saved the company \"hundreds of business days associated with making operational and process changes to ramp up production.\"  37  Representatives from Intel stated that, had they not received the flexible air permit, continued permitting-related delays would likely have pushed Intel to redirect its production investment and operating facilities to locations where changes could be made within existing environmental regulations (e.g. other U.S. states or the company's fabs in Ireland or Israel). 38 \n United States Federal EHS Regulatory Processes Federal EHS laws and regulations govern both the construction and operation of fabs. These laws and regulations are generally overseen by the EPA, the U.S. Army Corps of Engineers, and the U.S. Department of the Interior. These agencies are tasked with auditing proposed construction projects for compliance with relevant federal regulations that touch on fab utility, transportation, and supply chain infrastructure requirements. For example, a regional economic development association in the state of Washington that sought to increase fab construction in the Pacific Northwest observed that \"among local and federal agencies, the Bonneville Power Administration (BPA), the United States Army Corps of Engineers (USACE), and the Environmental Protection Agency (EPA), would be key players in coordinating site reviews with parallel agencies in Washington State.\"  39   At the federal level, there are many laws designed to maintain environmental quality that are also known to often significantly delay major construction projects. Notably, the National Environmental Policy Act (NEPA) review process governs construction projects deemed to be a \"major federal action.\" 40 If the provision of CHIPS Act incentives are determined to be a \"major federal action\" 41 then the construction of new semiconductor fabs could be significantly delayed.  42  In 2020 the White House Council on Environmental Quality compiled data on timelines for 1,276 Environmental Impact Statements (EIS) filed between 2010 and 2018 and found that NEPA reviews averaged 4.5 years.  43  This permitting process does not include the average of 1.86 years it takes to physically construct a semiconductor fab. This number also does not reflect other federal environmental reviews, some of which may happen concurrent with the NEPA process or entirely separately. A 2017 report from the White House also identified preconstruction permits and operating permits required under the Clean Air Act as \"the primary barrier to responsible and timely facility permitting\" finding that \"for some large projects  \n Suppliers In addition to constructing fabs, a process which takes a minimum of two years, the process of establishing a more resilient semiconductor supply chain in the United States necessitates increased domestic production of critical minerals, materials, chemicals, and gases, all of which require lengthy permitting processes. As a result, ongoing collaboration with international allies may be a more expeditious means of increasing semiconductor supply chain resiliency than attempting to increase domestic production in the short term. Some of the semiconductor supply chain, such as providers of specialty chemicals, would need to construct production facilities that would implicate many of the same regulatory considerations identified in Table  1 . Other parts of the supply chain, such as suppliers of raw minerals, would encounter entirely new regulatory considerations such as mining permits that may slow the ability to ramp up domestic production. The June 2021 White House supply chain report found \"establishing strategic and critical material production is an extremely lengthy process. Independent of permitting activities, a reasonable industry benchmark for the development of a mineralbased strategic and critical materials project is not less than ten years.\" 48 United States-based production of materials, chemicals and gases used in semiconductor manufacturing is, with few exceptions, limited. The United States Geological Survey's 2020 Mineral Commodities survey indicates materials like Arsenic, Beryllium, Bismuth, Boron, Cadmium, Helium, Indium, Rhenium, and Silicon are all used in the semiconductor manufacturing process.  49  The National Institute of Standards and Technology also maintains an index of 49 semiconductor process gases used in semiconductor production.  50  Except for a few materials like Helium and Silicon, U.S. firms do not produce or refine these materials in the United States. For example, Intel reports that it relies on smelters and refiners in China, Japan, Germany, Russia, Vietnam, the Philippines, Austria, and the United States for its supply of Tungsten, a key metal that is used to produce tungsten hexafluoride and tungsten sputter targets, both of which are used in the majority of semiconductor devices.  51  One reason that United States firms are not major suppliers of semiconductor materials is the increased cost of doing business in the United States. A Taiwanese supplier of specialty gases estimates that it costs five to six times more for them to build a factory in the United States than in Taiwan.  52  This price disparity is driven by several factors. In addition to contending with some of the construction permitting described above in Table  1  and higher labor market rates,  53  Taiwanese suppliers of semiconductor materials identified \"the role of transportation systems\" and a need for \"seamless dual supply of electricity and natural gas at a favorable rate for the operation of purification and solvent recovery plants\" as important factors, and requested federal and state government help to improve transportation system connectivity and utility provider reliability.  54  Both of these companies are single source suppliers of chemicals for TSMC in Taiwan and expressed interest in supplying TSMC's United States facility, but observe a clear increase in the cost of doing business in the United States. United States-based production of materials, chemicals and gases used in semiconductor manufacturing is also limited because many of these goods are derived from, or contribute to, environmentally harmful practices. For example, the semiconductor industry intensively consumes hydrofluorocarbons in the manufacturing process, a type of gas that is recognized by the EPA as a high global warming potential (GWP) greenhouse gas (GHG).  55  HFCs for semiconductor use are generally not produced in the United States. Recently, The Biden Administration's support for the Kigali Amendment to the Montreal Protocol on Substances that Deplete the Ozone Layer committed the United States to cap, and ultimately reduce, the semiconductor industry's use of HFCs.  56  However, until the semiconductor industry can innovate a substitute for the use of HFCs and qualify suppliers, there is no short-term alternative except to continue importing HFCs. Increasing the domestic resilience of this part of the supply chain in the medium to long term requires finding an international supplier willing to navigate the regulatory bureaucracy to establish new operations in the United States. Several companies and industry associations have highlighted how United States government regulatory decisions may increase the costs of doing business in the United States for semiconductor manufacturers and their suppliers. For example, the EPA has undertaken risk evaluations of several chemicals under the Toxic Substances Control Act that are used in the semiconductor manufacturing process.  57  Depending on the findings of these risk assessments, new regulations may restrict the production and supply of these chemicals in the United States  58  In addition, Executive Order 140083 (January 27, 2021) and the subsequent implementation of a temporary ban on new oil and gas leases under Department of Interior Order No. 3395 may reduce future opportunities for domestic helium development. Helium is recovered from natural gas deposits but exists in economic quantities in only a few places within the United States, mainly on federal lands, and is intensively used in semiconductor manufacturing.  59   \n Conclusion This report demonstrates that the United States is building fewer fabs at a slower rate than the rest of the world. CHIPS Act manufacturing incentives correctly aim to increase the number of fabs constructed in the United States, but more policy work is needed to ensure these new fabs are built on time and on budget. Prioritize regulatory support at the local, state, and federal level to expedite fab construction. The United States should make infrastructure investments targeting utilities, transportation, and supply chain networks that will assist semiconductor manufacturers. These policies would make the United States competitive with the \"foreign allies and competitors\" identified in the June 2021 White House supply chain report, and would align United States incentives to match those already offered abroad. Fully implement several of the recommendations from the 2017 PCAST report: 60 1. The Federal government should review permitting for technology facilities to identify areas where regulations are redundant with state rules and might therefore be modified or removed. 2. The EPA should create additional \"fast track\" permitting options that allow fabs to make some operational changes without filing environmental permit applications, potentially modeled after the State of Oregon's Plant Site Emissions Limit (PSEL) program. 61 Engage with allies and partners to increase semiconductor supply chain resiliency in materials, gases, and chemicals. Because increasing domestic United States production of many raw materials and chemicals would be contingent on lengthy development of mining and/or refining capacity, coordination with allies who already have existing production and refining capacity is essential. Quantify current and forecasted demand for materials, gases, and chemicals used in United States semiconductor manufacturing. Stockpiling of certain materials, modeled after existing United States government programs like those operated by The Defense Logistics Agency Strategic Materials branch, 62 may be an option. Identify substitutes for environmentally harmful or strategically concentrated resources. The EPA is currently supporting semiconductor industry efforts to develop substitutes for environmentally harmful greenhouse gases used in semiconductor manufacturing. These efforts could serve as a template for further work to identify substitutes for certain materials, chemicals, and gases used in semiconductor manufacturing for which there is no commercially viable domestic supply.  63   Figure 1 . 1 Figure 1. Average Number of Days from Fab Construction Start Date to Production Date, by region <PHONE> \n Figure 2 . 2 Figure 2. Time Required to Build New Fabs in the United States (L) and Total Number of New Fab Projects (R), 1990 -2020 \n Figure 3 . 3 Figure 3. Time Required to Build New Fabs in China (L) and Total Number of New Fab Projects (R), 1990 -2020 \n Figure 4 . 4 Figure 4. Time Required to Build New Fabs in Taiwan (L) and Total Number of New Fab Projects (R), 1990 -2020 \n Figure 5 . 5 Figure 5. Number of >100k WSpM Greenfield Fab Projects (L) and Average Number of Days from Fab Construction Start Date to Production Date for those Projects (R) by Region, <PHONE> \n Figure 6 . 6 Figure 6. Transportation Infrastructure Requirements for Semiconductor Production \n Table 1 . 1 Table 1 describes federal environmental requirements for construction in greater detail. Federal Environmental Requirements for Construction Relevant Law Agency Associated Requirements Regulation Jurisdiction Air Quality Clean Air Act (CAA) Environmental Permit for construction- Protection related pollutant emissions Agency Asbestos Clean Air Act (CAA) Environmental Report of asbestos releases Protection above a threshold Agency Dredged and Section 404 of the CWA Army Corp of Permit for discharge of Fill Engineers or dredged material Material/Water state regulator s Environmental National Environmental Environmental Submit Environmental Impact Policy Act Protection Assessment Agency Submit Environmental Impact Assessment Hazardous Comprehensive Environmental Permit to excavate Substances Environmental Response, Protection contaminated soil Compensation, and Agency Liability Act (CERCLA) Historic National Historic United States Pre-construction Properties Preservation Act Department of consultation for historic the Interior property considerations Polychorinated Toxics Substances Environmental Storage and disposal biphenol (PCB) Control Act (TSCA) Protection requirements Wastes Agency Solid and Resource Conservation Environmental Transporters of hazardous Hazardous and Recovery Act (RCRA) Protection waste must register with Wastes Agency EPA Spill Reporting Maintain a material safety data sheet (MSDS) \n ,[this]   permitting process can take 12-18 months.\"44  More recently, the EPA has announced a goal to make permitting decisions within six months of receipt. But as discussed earlier, given the tight timelines on which this industry operates, any delay can be costly to a firm's competitiveness.45    U.S. State and Local EHS Regulatory ProcessesState and local EHS regulations and agencies also have a bearing on semiconductor manufacturers. In some cases, the EPA delegates authority to implement regulatory programs to states and other agencies. For example, the U.S. Army Corps of Engineers (USACE) administers Section 404 of the Clean Water Act, which regulates the discharge of dredged and fill material into all waters of the United States. The USACE's Fort Worth, TX District reviewed an application of Samsung Austin Semiconductor LLC in 2014 to expand their fabrication facility in compliance with Section 404 of the CWA. However, this application also invoked reviews being conducted concurrently by the Texas Commission on Environmental Quality and a City of Austin Development Permit.46   Even after federal and state-level concerns are addressed, semiconductor building sites may run into local regulatory barriers. When GlobalFoundries was considering building a new fab in Malta, NY there were lingering zoning changes at the local level that needed to be made by the Town Boards of Malta and Stillwater, NY which delayed the site's initial development.47     United States Regulatory Processes Affecting Semiconductor \n\t\t\t Stephanie Yang, \"The Chip Shortage Is Bad. Taiwan's Drought Threatens to Make It Worse,\" The Wall Street Journal, April 16, 2021,", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-No-Permits-No-Fabs.tei.xml", "id": "8aa70bf3cead3a9bab53fc2a5bf40329"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "This paper is the third installment in a series on \"AI safety,\" an area of machine learning research that aims to identify causes of unintended behavior in machine learning systems and develop tools to ensure these systems work safely and reliably. The first paper in the series, \"Key Concepts in AI Safety: An Overview,\" described three categories of AI safety issues: problems of robustness, assurance, and specification. This paper introduces interpretability as a means to enable assurance in modern machine learning systems.", "authors": ["Tim G J Rudner", "Helen Toner"], "title": "Key Concepts in AI Safety: Interpretability in Machine Learning CSET Issue Brief", "text": "Introduction Interpretability, also often referred to as explainability, in artificial intelligence (AI) refers to the study of how to understand the decisions of machine learning systems, and how to design systems whose decisions are easily understood, or interpretable. This way, human operators can ensure a system works as intended and receive an explanation for unexpected behaviors. Modern machine learning systems are becoming prevalent in automated decision making, spanning a variety of applications in both the private and public spheres. As this trend continues, machine learning systems are being deployed with increasingly limited human supervision, including in areas where their decisions may have significant impacts on people's lives. Such areas include automated credit scoring, medical diagnoses, hiring, and autonomous driving, among many others. 1 At the same time, machine learning systems are also becoming more complex, making it difficult to analyze and understand how they reach conclusions. This increase in complexity-and the lack of interpretability that comes with it-poses a fundamental challenge for using machine learning systems in high-stakes settings. Furthermore, many of our laws and institutions are premised on the right to request an explanation for a decision, especially if the decision leads to negative consequences. 2 From a job candidate suing for discrimination in a hiring process, to a bank customer inquiring about the reason for receiving a low credit limit, to a soldier explaining their actions before a court-martial, we assume that there is a process for assessing how a decision was made and whether it was in line with standards we have set. This assumption may not hold true if the decisionmaker in question is a machine learning system which is unable to provide such an explanation. In order for modern machine learning systems to safely integrate into existing institutions in high-stakes settings, they must be interpretable by human operators. \n Why Are Modern Machine Learning Systems Not Interpretable? Many modern machine learning systems use statistical models called deep neural networks which are able to represent a wide range of complex associations and patterns. To understand why decisions by deep neural networks are hard to interpret, consider two types of systems that are interpretable. One example is an earlier generation of AI systems which use human-specified rules instead of relying on data. As a simplified example, the autopilot system on an airplane uses a set of if-thisthen-that rules to keep the plane on course-if the nose drops, lift it; if the plane banks left, roll a little right, and so on. While the rules in real systems are more complicated, they nonetheless allow humans to look back on system behavior and recognize which this triggered which that. A second example is a linear model, a simple kind of machine learning model. Like all machine learning systems, a linear model uses number values called parameters to represent the relationship between inputs and outputs. For example, one could create a model to predict someone's salary from their age and years of schooling, two explanatory variables. In a linear model, the main parameters would be one number value to be multiplied by the explanatory variable \"age,\" and one number to be multiplied by the other explanatory variable \"years of schooling.\" Determining what value those two parameters should take is the learning part of machine learning. For a linear model, good parameter values can be found by simple calculations that may take a computer less than a second to perform. More importantly, because each parameter in a linear model is directly associated with one explanatory variable, understanding how the model works is simple. If, say, the parameter that gets multiplied by \"age\" is much higher than the parameter for \"years of schooling,\" then the model is predicting age is a more important determinant of salary. Deep neural networks are different. By design, they have far more parameters than linear models, and each parameter is connected in complex ways with inputs and other parameters, rather than directly linking explanatory variables to the outcome that the model seeks to predict. This complexity is a double-edged sword. On one hand, models can represent highly complex associations and patterns, allowing them to solve problems previously considered out of reach for computers, including image recognition, autonomous driving, and playing the game of Go. On the other hand, unlike older or simpler computer systems, the internal functioning of each model is very difficult to understand. At this point it is worth noting why the term \"black box,\" often used in this context, is not quite right to describe why deep neural networks are hard to understand. Machine learning researchers understand perfectly well how the mathematical operations underlying these systems work, and it is easy to look at the parameter values that make up the model. The problem lies in understanding how these millions (or even billions) of number values connect to the concepts we care about, such as why a machine learning model may erroneously classify a cat as a dog. In other words, interpreting deep neural networks requires both the ability to understand which high-level features in the data-such as a certain part of an image or a specific sequence of wordsaffect a model's predictions and why a model associates certain high-level features with a corresponding prediction-that is, how deep neural networks \"reason\" about data. \n How to Make Modern Machine Learning Systems More Interpretable Researchers are pursuing a range of different approaches to improving the interpretability of modern machine learning systems. A fundamental challenge for this work is that clear, well-defined concepts have yet to be developed around what it would mean for different types of systems to be interpretable. So far, interpretability research seeks to build tools making it somewhat more possible for a human operator to understand a system's outputs and inner workings. Saliency maps are one popular set of tools for making modern machine learning systems used for computer-vision applications more interpretable. Broadly speaking, saliency maps visualize which areas of an image led to a model's classification of the same image.  3  For example, we might investigate why a deep learning model has learned to identify images of cats and dogs from a large dataset of cat and dog images labeled as such. If we wish to understand why the model classified an image of a German Shepherd as a dog, a saliency map may highlight the parts of the image containing features present in dogs, but not in cats (for example, a large muzzle). In this way, a saliency map communicates to a human operator which part of the image prompted the machine learning system to classify the image as it did. Another popular method for making deep neural networks more interpretable is to visualize how different components of the model relate to high-level concepts that may affect the model's predictions-concepts such as textures and objects in image classifiers, grammar and tone in language models, or short-term vs. long-term planning in sequential decision-making models.  Unfortunately, existing methods to make modern machine learning systems interpretable fall short; they typically only provide one angle from which to view the system instead of taking a holistic view. To fully understand how a machine learning system works, we must understand how the data and learning algorithm affected training, whether training could have resulted in a different model under modified training conditions, and how all of these factors ultimately affect the system's predictions. At this time, our understanding of these questions is still very limited and the insights obtained from existing methods are fallible, require human supervision, and only apply to a small subset of application areas. For example, the saliency maps in Figure  1  do shed some light on how the model in question works. One can see, for instance, that the model focuses on the dog to classify it as such, but identifies the sailboat in part by looking at the ocean. Saliency maps, however, do not help a human observer understand what might have led to different outcomes. Likewise, Figure  2  shows a method for understanding what the different parts of an image classifier detect. Unfortunately, looking at this type of visualization does not help a human operator evaluate whether the system is likely to be accurate, fair, or reliable. The lack of a common, well-defined vocabulary for interpretable machine learning systems further exacerbates these shortcomings.  5  Key concepts, such as trustworthiness, reliability, transparency, or verifiability, are often used loosely or interchangeably rather than referring to standardized or generally accepted technical definitions of such terms, making it difficult to measure progress and to accurately and reliably communicate research results to the public. \n Outlook Interpretability will allow us to understand potential failure modes for machine learning systems, enable regulatory compliance and audits, and reduce the risk of models with algorithmic or datainduced bias being deployed. Rendering modern machine learning systems like deep neural networks interpretable will help us ensure that any such system deployed in safety-critical settings works as intended. Unfortunately, while an active and ongoing area of research, existing approaches to achieving interpretability do not yet provide satisfactory solutions. It remains unclear when-or even if-we will be able to deploy truly interpretable deep learning systems. In the meantime, our best option may be to stick with simpler, more inherently interpretable models whenever possible.   Figure 1 . 1 Figure 1. Examples of images and their corresponding saliency maps indicating which parts of the images contribute most to how a machine learning system trained on a large set of images would classify them. \n 4 4 \n Figure 2 . 2 Figure 2. A visualization showing examples of what different layers of an image classification network \"see.\" The left-most column, depicting an early layer, is picking up lines; middle layers are detecting textures and patterns of increasing complexity; later layers, shown on the right, are looking for objects. \n 6    \n Endnotes1 Danny Yadron and Dan Tynan, \"Tesla Driver Dies in First Fatal Crash While Using Autopilot Mode,\" The Guardian, June 30, 2016, https:// www.theguardian.com/technology/2016/jun/30/tesla-autopilot-death-selfdriving-car-elon-musk; Forough Poursabzi-Sangdeh, Daniel G. Goldstein, Jake M. Hofman, Jennifer Wortman Vaughan, Hanna Wallach, \"Manipulating and Measuring Model Interpretability,\" arXiv [cs.AI] (November 8, 2019), arXiv, preprint, https://arxiv.org/abs/1802.07810; and Jennifer Valentino-DeVries, \"How the Police Use Facial Recognition, and Where It Falls Short,\" The New York Times, January 12, 2020, https://www.nytimes.com/2020/01/12/ technology/facial-recognition-police.html. \n \n\t\t\t For a more in-depth discussion of what an \"explanation\" is in a legal context, why explanations are necessary, and how AI explanations compare to human \n\t\t\t Cynthia Rudin, \"Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead,\" Nature Machine Intelligence 1 (May 2019): 206-215, https://www.https://www.nature.com/ articles/s42256-<PHONE>-x.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Key-Concepts-in-AI-Safety-Interpretability-in-Machine-Learning.tei.xml", "id": "6c18327296eef14d4ace55a167bd7a27"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Machine Learning (ML) has been successful in automating a range of cognitive tasks that humans solve effortlessly and quickly. Yet many realworld tasks are difficult and slow : people solve them by an extended process that involves analytical reasoning, gathering external information, and discussing with collaborators. Examples include medical advice, judging a criminal trial, and providing personalized recommendations for rich content such as books or academic papers.", "authors": ["Owain Evans", "Andreas Stuhlmüller", "Chris Cundy", "Ryan Carey", "Zachary Kenton", "Thomas Mcgrath", "Andrew Schreiber"], "title": "Predicting Human Deliberative Judgments with Machine Learning", "text": "There is great demand for automating tasks that require deliberative judgment. Current ML approaches can be unreliable: this is partly because such tasks are intrinsically difficult (even AI-complete) and partly because assembling datasets of deliberative judgments is expensive (each label might take hours of human work). We consider addressing this data problem by collecting fast judgments and using them to help predict deliberative (slow ) judgments. Instead of having a human spend hours on a task, we might instead collect their judgment after 30 seconds or 10 minutes. These fast judgments are combined with a smaller quantity of slow judgments to provide training data. The resulting prediction problem is related to semi-supervised learning and collaborative filtering. We designed two tasks for the purpose of testing ML algorithms on predicting human deliberative judgments. One task involves Fermi estimation (back-of-the-envelope estimation) and the other involves judging the veracity of political statements. We collected a dataset of 25,000 judgments from more than 800 people. We define an ML prediction task for predicting deliberative judgments given a training set that also contains fast judgments. We tested a variety of baseline algorithms on this task. Unfortunately our dataset has serious limitations. Additional work is required to create a good testbed for predicting human deliberative judgments. This technical report explains the motivation for our project (which might be built on in future work) and explains how further work can avoid our mistakes. Our dataset and code is available at https: //github.com/oughtinc/psj. \n Introduction \n Fast and slow judgments Machine Learning has been successful in automating mental tasks that are quick and effortless for humans. These include visual object recognition, speech recognition and production, and basic natural language prediction and comprehension  [1, 2, 3, 4] . Andrew Ng states the following heuristic  [5] : If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future.  1  In this technical report we refer to judgments that are quick (roughly 30 seconds or less) and easy for most humans as fast judgments. Fast judgments contrast with slow judgments, which may involve lengthy processes of deliberate reasoning, research, experimentation, and discussion with other people.  2  Many important real-world tasks depend on slow judgments: • Predict the verdict of jury members in a criminal trial. • Predict which engineers will be hired by a company with an extensive interview process. • Predict whether experts judge a news story to be fake or intentionally misleading. • Predict a doctor's advice to an unwell patient after a thorough medical exam. • Predict how a researcher will rate a new academic paper after reading it carefully. • Predict how useful a particular video lecture will be for someone writing a thesis on recent Chinese history. There is great demand for Machine Learning (ML) and other AI techniques for predicting human slow judgments like these, especially in hiring workers, detection of fake or malicious content, medical diagnosis, and recommendation  [8, 9, 10, 11] . However ML approaches to predicting these slow judgments are often unreliable  [12, 13, 14] : even if they do reasonably well on a majority of instances, they may have large errors on more demanding cases (e.g. on inputs that would be tricky for humans or on inputs chosen by humans to try to fool the algorithm). One source of the unreliability for ML algorithms is optimizing for a subtly wrong objective. Suppose a student gets video lecture recommendations from YouTube.  3  These recommendations may be optimized based on video popularity (\"Do users click on the video?\") and engagement (\"Do users Like or share the video?\"). These metrics are mostly based on fast judgments. Yet the student would prefer recommendations based on slow judgments, such as the evaluation of another student who has carefully checked the lecture for factual accuracy by tracking down the sources. Fast and slow judgments will sometimes coincide, as when a lecture is inaudible or off-topic. Yet a lecture may seem useful on first glance while turning out to be riddled with inaccuracies that careful research would expose.  4  Predicting slow judgments in the tasks above is challenging for current ML algorithms. This is due in part to the intrinsic difficulty of the tasks; predicting how a student evaluates a lecture is arguably AI-complete  [16] . Another difficulty is that collecting slow judgments is inherently expensive: if it takes five hours of fact-checking to recognize the errors in a lecture then a dataset of millions of such evaluations is impractically expensive. Big datasets won't solve AI-complete problems but will improve ML performance. Predicting slow judgments is also related to long-term AI Safety, i.e. the problem of creating AI systems that remain aligned with human preferences even as their capabilities exceed those of humans  [17, 18, 19, 20, 21] . Rather than create AI that shares only human goals, a promising alternative is to create AI that makes decisions in the way a human would at every timestep  [22, 23] . This approach of imitating human decision-making is only promising if it imitates human deliberate (slow) judgments  [24, 25] . A system making thousands of human-like slow judgments per second could have super-human capabilities while remaining interpretable to humans  [26, 27] . \n Using Fast Judgments to Predict Slow Judgments How can we get around the challenges of predicting slow judgments? One approach is to tackle the AI-completeness head on by trying to emulate the process of human deliberate reasoning.  5  A second approach (which complements the first) is to tackle the data problem head on and find ways to collect huge datasets of real or synthetic slow judgments  [38] . This technical report explores an indirect approach to the data problem. Instead of collecting a big dataset of slow judgments, we collect a small dataset of slow judgments along with a larger quantity of side information related to the slow judgments. What kind of side information would help predict a person's slow judgment? If Alice makes a slow judgment about a question, then Alice's fast judgment about the same question is relevant side information. As noted above, in predicting a student's thorough evaluation of a video lecture, it is helpful to know their fast judgment (e.g. after watching the video for only 30 seconds). Likewise, a doctor's guess about a diagnosis after 30 seconds will sometimes predict their careful evaluation. Another kind of side information for predicting Alice's slow judgment about a question are the judgments of other people about the same question. This is the idea behind collaborative filtering  [39] , where a person's rating of a song is predicted from the ratings of similar people. While collaborative filtering has been widely studied and deployed, there is little prior work on using fast judgments as side information for slow judgments. The motivation for using fast judgments is that they are often easily available and their cost is much lower than slow judgments. Human cognition is like an \"anytime\" iterative algorithm: whereas our slow judgments are more discerning and reliable, our fast judgments are well-typed and provide increasingly good approximations to slow judgments over time.  6  For most judgment tasks we can collect fast judgments from humans and these fast judgments will come at a cost orders of magnitude cheaper than slow judgments. Where fast judgments sometimes coincide with slow judgments and include uncertainty information 7 , it may be better to spend a fixed budget on a mix of slow and fast judgments than on slow judgments alone. Often fast judgments are not just cheaper to collect but are essentially free. YouTube, Facebook, Twitter and Reddit have vast quantities of data about which content people choose to look at, share with others, \"Like\", or make a quick verbal comment on. Using fast judgments as side information to predict slow judgments requires modifying standard ML algorithms. While the objective is predicting slow judgments, most of the training examples (e.g. most video lectures) only come with fast judgments. This is related to semi-supervised learning  [40, 41] , distant supervision  [42]  (where the training labels are a low-quality proxy for the true labels), learning from privileged information  [43] , as well as to collaborative filtering. \n Contributions and caveats This tech report describes a project applying machine learning (ML) to predicting slow judgments. We designed tasks where slow judgments (deliberate thinking and research) are required for doing well but quick judgments often provide informative guesses. We collected a dataset of fast and slow human judgments for these tasks, and formulated a set of prediction problems (predicting held-out slow judgments given access to varying quantities of slow judgments at training time). We applied ML baselines: standard collaborative filtering algorithms, neural collaborative filtering specialized to our tasks, and a Bayesian hierarchical model.  8  . Unfortunately our dataset turned out to be problematic and is unlikely to be a good testing ground for predicting slow judgments. This report describes parts of our project that may be usefully carried over to future work and summarizes what we learned from our efforts to create an appropriate dataset. The problem of predicting slow judgments (and of creating a dataset for the purpose) \n Fermi comparisons weight of a blue whale in kg < 50,000 population of Moscow * smaller angle in degrees between hands of clock at 1.45 < 15,000,000 driving distance in miles between London and Amsterdam < 371 weight of $600 worth of quarters in pounds * area of Canada in square miles < 328,000,000 \n Politifact truth judgments Rob Portman: \"Since the stimulus package was passed, Ohio has lost over 100,000 more jobs.\" Republican Party: \"Charlie Crist is embroiled in a fraud case for steering taxpayer money to a de facto Ponzi scheme.\" \n Mark Zaccaria: \"James Langevin voted to spend $3 billion for a jet engine no one wants.\" was harder than expected. We hope to stimulate future work that avoids the problems we encountered. Our main contributions are the following: 1. We designed two slow judgment tasks for humans: Fermi Comparisons involve mental reasoning and calculation, and Politifact Truth Judgments involve doing online research to assess the veracity of a political statement.  9  We created a web app to collect both fast and slow judgments for the task. These tasks and the app could be used in future work. 2. We relate predicting slow judgments with side information to collaborative filtering and we implement simple baselines based on this relation. 3. We diagnose problems with our dataset. First, while slow judgments were significantly more accurate than fast judgments, the difference was smaller than expected. Second, variability among subjects was hard to distinguish from noise; so ML algorithms could not exploit similarities among users as in collaborative filtering. Third, while there is clear variation in how users respond to different questions, this variation is very hard for current ML algorithms to exploit. \n Tasks and Datasets for Slow Judgments Our overall goal was (A) to define the ML problem of predicting slow judgments given a training set of fast and slow judgments, and (B) to find or create datasets which can be used to test algorithms for this problem. The domain that humans are making fast/slow judgments about should be AI-complete or closely related to an AI-complete problem.  10  Solving such a task will (for some problem instances) depend on patterns or structures that current Machine Learning algorithms do not capture. So while we cannot hope for highly accurate predictions of slow judgments, we can seek ML algorithms that \"know what they know\"  [44] . Such algorithms exploit patterns they are able to learn (producing confident accurate predictions) and otherwise provide well-calibrated uncertain predictions  [45, 46] . This section describes the AI-complete tasks we designed for testing algorithms for predicting slow judgments. Before that we first review related datasets in this area. \n Existing Datasets Many ML datasets record human slow judgments for AI-complete tasks. These include datasets of movie reviews [47], reviewer scores for academic papers  [48] , and verdicts for legal cases  [49] . There are also datasets where the ground-truth labels are the product of extensive cognitive work by humans (e.g. theorem proving  [38] , political fact-checking  [14] ) and these could potentially be used to study human slow judgments. However, these datasets do not contain fast judgments. In particular, for each task instance in the dataset, there's a slow judgment by a particular individual but there's no corresponding fast judgment. Moreover the datasets do not explicitly record information about the time and resources than went into the slow judgment  11  . For example, the reviewers of academic papers do not divulge whether they spent ten minutes or two hours reviewing a paper. Due to the lack of existing datasets, we created a new dataset recording human slow and fast judgments for AI-complete tasks. We designed two tasks for this purpose, which may be useful for future work in this area (even if our dataset is not). \n Two Judgment Tasks: Fermi and Politifact We created two tasks for humans that require deliberate, multi-step thinking and rely on broad world knowledge. In the Fermi Comparisons task (abbreviated to \"Fermi\") users determine which of two quantities is larger. This often involves simple arithmetic, factual knowledge, and doing back-of-the-envelope estimates to determine the order of magnitude of the quantity. Example questions are show in Figure  1 . Note that human subjects (who we refer to as \"users\") are not allowed to look up the quantities online as this would trivialize the task. In the Politifact Truth task (abbreviated to \"Politifact\") users evaluate whether a statement by a US politician or journalist is true or false (see Figure  1 ). They have access to metadata about the statement (see Figure  2 ) and are allowed to do research online. For both tasks, users assign a probability that the statement is true (with \"0\" being definitely false and \"1\" being definitely true) and they enter their probability via the interface in Figure  2 . Their goal is to minimize mean squared error between their probabilistic judgment and the ground-truth answer. Questions and ground-truth answers for Fermi were constructed and computed by the authors. Political statements and ground-truth (i.e. the judgments of professional fact-checkers) for Politifact were downloaded from the politifact.com API (following  [14] ). The Fermi and Politifact tasks satisfy the following desirable criteria: • A generalized version of each task is AI-complete. Fermi questions depend on mathematical reasoning about sophisticated and diverse world knowledge (e.g. estimate the mass of the atmosphere). If novel Fermi questions are presented at test time (as in a job interview for high-flying undergraduates), the task is AI-complete. Politifact questions can be about any political topic (e.g. economics, global warming, international relations). Answering them requires nuanced language understanding and doing sophisticated research on the web. • Fast judgments are informative about slow judgments. Given a question (as in Figure  2 ), fast judgments of the probability (e.g. less than 30 seconds) will be informative about a user's slow judgment (e.g. 5 or 10 minutes). However some questions are too difficult to solve with a fast judgment. • It is possible to make progress through analytical thinking or gathering evidence. For Fermi, breaking down the problem into pieces and coming up with estimates for each of the pieces is a useful cognitive strategy. For Politifact (but not Fermi), participants are allowed to use a search engine and to read web pages to check political facts. • The ground-truth is available for both Fermi and Politifact questions. The problem of predicting slow judgments we discussed in the Introduction is not limited to human judgments about objective facts (as the list of examples in Section 1.1 makes clear). However, if the ground-truth is available then this makes certain experiments possible.  12   \n Data collection Human participants (who we refer to as \"users\") were recruited online  13  and answered questions using the UI in Figure  2 . Users see a question, a form for entering their probability judgment, a progress bar that counts down the time remaining, and (for Politifact only) a search engine tool. For each question, users provides fast, medium and slow probability judgments. In Fermi the user is presented with a question and has 15 seconds to make the fast judgment. Once the 15 seconds have elapsed, they get an additional 60 seconds in which they are free to change their answer as they see fit (the medium judgment). Finally they get another 180 seconds to make the slow judgment. The setup for Politifact is identical but the time periods are 30, 90 and 360 seconds. (Users are free to use less than the maximal amount of time. So for particularly easy or difficult questions, the \"slow\" judgment may actually only take 30 seconds or less.)  12  In Politifact, the ground-truth is just an especially slow judgment of an expert in political fact-checking and could be modeled as such. We did not try this in our experiments.  13  We used Amazon's Mechanical Turk (MT) for a pilot study. For the main experiment we used volunteers recruited via social media with an interest in improving their probabilistic judgments. We found it hard to design an incentive scheme for MT such that the users would make a good-faith effort to do well (and not e.g. declaring 50% on each judgment) while not cheating by looking up Fermi answers.  The Fermi data we collected consisted of 18,577 judgments, with a third of these being slow judgments. There were 619 distinct users and 2,379 Fermi estimation questions (with variable numbers of judgment per question). Each user answered 12.9 distinct questions on average (median 5). The Politifact dataset was smaller, containing 7,974 judgments, from 594 users, covering 1,769 statements. Each user answered 6.4 distinct questions on average (median 5). \n Descriptive Statistics Our goal was to use the dataset to train ML algorithms to predict slow judgments and such a dataset should have the following features: 1. Large, clean, varied dataset: as usual a large dataset with minimal noise is desirable. In particular there should be many judgments per user and many distinct questions. Possible sources of noise should be minimized (e.g. users are noisier when learning the task and if they pay less attention due to tedious or overly difficult questions). 2. Fast-Slow Correlation: fast judgments are informative about slow judgments but not too much. For some users and some questions, fast judgments may be similar to slow judgments. In other cases, the fast judgment will be uncertain or wrong. 3. User Variation: individual users vary in their overall performance and their performance in different kinds of question. We want to predict the slow judgments of an individual (not the ground-truth). In some tasks listed in Section 1.1 users vary because of different preferences. In Fermi and Politifact users might vary in their areas of knowledge (e.g. science vs. sport questions in Fermi). 4. Question Variation: individual questions vary in terms of how users answer them. For example, some questions are hard (producing random or systematically incorrect answers) and some are easy. This allows algorithms to predict user answers better based on features of the question. Did our data have the features above? As noted in Section 2.3 the dataset was relatively small (especially Politifact) and many users only answered a small number of questions. Correlations between fast and slow judgments are shown in Figure  3 . They show that judgments started out uncertain but became more confident (in either direction) given time, and that confident fast judgments were less likely to change. Figure  4  shows that slow judgments were more accurate than fast on average. Nevertheless, the differences between slow, medium and fast judgments are not as large as we had hoped. Did users vary in their accuracy? We expected users to vary in overall accuracy and accuracy for different kinds of question (e.g. some users do better on science than sports questions). However the level of variation among users was not high enough to rule out possibility that the best two-thirds of users did not vary (see Figure  5 ). This is a problem with the design of our experiment and we discuss it further in Section 4.  Histogram shows number of users who obtained a particular level of performance (measured by MSE). Note that 0.25 is the MSE achieved by a user who always says 50%. While users have widely varying performance, much of this is due to noise and we can't rule out the possibility that most users do not vary in actual skill (see Figure  5 ). The graph suggests we can't rule out the null hypothesis that the best two-thirds of users don't vary in their accuracy. (We removed users with less than k=6 judgments; the graphs were similar when k was set higher.) Questions in both Fermi and Politifact did vary significantly in difficulty for human users (see Figure  6 ). However, there are less than 2500 questions for each task. This makes it hard for algorithms to generalize from the question text or meta-data without substantial pre-training.  \n Predicting Slow Judgments with ML This section presents a problem definition for predicting slow judgments using ML and results for baseline algorithms on our dataset. The ML task and algorithms are independent of the dataset and may be useful for future work in this area. \n Task: Predicting Slow Judgments with ML \n Task Definition For both Fermi and Politifact the data consists of user probability judgments about questions. Each probability judgment h(q, u, t) ∈ [0, 1] depends on the question text and meta-data features q, the user id u, and the time t ∈ {0, 1, 2} (where the indices correspond to fast, medium, and slow judgments respectively). We let ĥ(q, u, t) be an algorithm's prediction of the user judgment h(q, u, t). The task is to predict slow judgments h(q, u, 2) from the test set. While only slow judgments appear in the test set, the training set consists of fast, medium and slow judgments.  14  The loss function is mean-squared error over all slow judgments in the test set. Let T = {(q, u)} be the set of question-user pairs that appear in the test set. Then the objective is to minimize the mean squared error over slow judgments in the test set: 1 |T | (q,u)∈T h(q, u, 2) − ĥ(q, u, 2) 2 (1) Note that this problem setup is analogous to content-based collaborative filtering, where the task is to predict a user's held-out rating for an item (e.g. a movie), given the user id and a feature vector for the item  [51] . \n Train on Fast Judgments, Predict Slow Slow judgments are much more expensive to collect than fast judgments. Yet for questions that are either very easy or very difficult for a user, the fast and slow judgments will be very similar. So if a model can learn when fast judgments are a good proxy for slow judgments, it can predict slow judgments reasonably well from a much cheaper dataset (as discussed in Section 1.1). In our dataset we have the user's slow judgment for a question whenever we have the fast judgment. Yet we can simulate the strategy of mostly collecting fast judgments by masking many of the slow judgments from our training set.  15  Our results are computed for three different levels of masking. \"Unmasked\" means that all slow judgments in the training set are included. \"Heavy\" means 90% of question-user pairs have the medium and slow removed and 7% have just the slow removed. \"Light\" means we removed slow and medium judgments for 60% of question-user pairs and removed just slow for 30%.  16   \n Algorithms and Results We ran the following baseline algorithms on our datasets: 1. Collaborative Filtering (KNN and SVD): The prediction task is closely related to collaborative filtering (CF). The main difference is that instead of predicting a user's judgment about an item, we predict a user's judgment about a question at a given time (where we expect users to generally converge to zero or one given more time). To reduce our task to CF, we flatten the data by treating each user-time pair (u, t) as a distinct user. We applied the standard CF algorithms K-nearest-neighbors (KNN CF) and singular value decomposition (SVD CF), using the implementation in the Surprise library  [52] . \n Neural Collaborative Filtering: We adapt the Neural CF algorithm  [53]  to our task. A neural net is used to map the latent question and user embeddings to a sequence of judgments for each time. Linear Neural CF forces the judgments to change linearly over time. \n Hierarchical Bayesian Linear Regression: To predict the user judgment for a question we pool all user judgments for the same question (ignoring user identity) and regress using a Bayesian linear regression model. This model exploits the temporal structure of judgments but it discards the question features and user identity. 4. Clairvoyant Model : Since user slow judgments are correlated with the ground-truth, algorithms will do better on new questions to the extent they can predict the ground-truth. Predicting ground-truth without sideinformation is difficulty for Fermi and Politifact. However, we can investigate how well a model would do if it had access to the ground-truth. The \"Clairvoyant\" model simply predicts that each user will respond with the ground-truth answer with full confidence. The \"Clairvoyant Mean\" model predicts the base-rate probability for a user given the ground-truth of a question. \n Results Table  1  shows results for Fermi and Politifact with different levels of masking of slow judgments. Hierarchical Linear Regression and SVD perform best. Since Hierarchical Linear regression ignores both question features and user identity, this suggests it is difficult for algorithms to improve predictions by modeling questions and users. This is additional evidence that our dataset is problematic. We had expected the Neural CF algorithms would do well, as they learn latent representations for both questions and users and (unlike SVD and KNN) they do not discard the temporal structure in the data. However, their poor performance is partly explained by the difficulty (discussed in Section 2.4) of distinguishing user variation from noise.  17  The performance of Clairvoyant Mean suggests that non-clairvoyant algorithms could be improved with a strong language and reasoning model that accurately predicts the ground-truth.  Problem: Fast and slow judgments were too similar There weren't big differences between fast and slow judgments. This can be addressed by choosing a task that requires more thinking and research than the Fermi and Politifact tasks. The task should also be incremental, such that a small amount of additional work yields a slightly better answer.  19  If additional work is unlikely to help, users will be averse to doing it (without big compensation). Problem: Variation between users was hard to distinguish from noise Lack of discernible variation can be addressed by collecting substantially more data per user and by trying to reduce noise (e.g. having practice questions, having less ambiguous questions, having tasks for which users are consistently fully engaged). A related issue is that human responses were low in information content. Users assigned probabilities (from 20 discrete options) to binary ground-truth outcomes. Many questions were easy and were answered with 95-100% confidence, while others were very difficult and answered with 50-55% confidence. Furthermore, users were rarely anti-correlated with the ground truth, so slow judgments for true statements were generally answered somewhere in the range 60-100%, which is a small region of the response space in terms of mean squared error. This problem of low information content could be addressed by having a task where user responses are richer in information: e.g. scalar valued, categorical (with many categories), or a text response such as a movie review. The cost of this change is that the ML modeling task becomes harder. Another way to address this issue is to have a task in which the goal for users is something other than guessing the ground-truth. There are many situations in which people make slow judgments for questions that are not about objective matters of fact. For example, when someone reviews a non-fiction book for their blog, they consider both objective qualities of the book (\"Does it get the facts right?\") and also subjective qualities (\"Did I personally learn things from it? Will it help me make decisions?\"). As noted above, there is nothing about our task definition or modeling that assumes the questions have an objective ground-truth.  20  Problem: Question features were hard for algorithms to exploit The questions in our tasks varied substantially in difficulty and content. Our models couldn't really exploit this variation, probably because (a) there were less than 2500 questions, and (b) predicting whether a question is challenging for humans is an intrinsically difficult task (drawing on NLP and world knowledge). Having models that can make use of question and meta-data features is an important goal for our research. If models can only predict slow judgments based on other human judgments, they will not be able to generalize to new questions that no humans have yet answered. The most obvious fix for this problem is to collect a much larger dataset. If the dataset contained millions of questions then language models would be better at learning to recognize easy vs. difficult questions. While collecting such a big dataset would be expensive, the cost could be mitigated by mainly collecting fast judgments.  21  An alternative to collecting a large dataset is to choose some object task (instead of Fermi and Politifact) for which pre-trained language models are useful (e.g. something related to sentiment analysis rather than judging the ground-truth of statements). \n Conclusion Machine Learning will only be able to automate a range of important real-world tasks (see Section 1.1) if algorithms get better at predicting human deliberative judgments in varied contexts. Such tasks are challenging due to AI-completeness and the difficulty and cost of data. We tried to create a dataset for evaluating ML algorithms on predicting slow judgments. The previous section discussed problems with our dataset and potential remedies. It's also worth considering alternative approaches to the challenge we outline in Section 1.1. First, some tasks may be more fruitful to model than Fermi estimation and political fact checking. Second, deliberation can be modeled explicitly by building on recent deep learning approaches  [28, 29, 30, 31, 32, 33, 34, 35, 36, 37] . To more precisely capture how humans deliberate, we could also record the individual steps people take during deliberating (e.g. by requiring users to make their reasoning explicit or by recording their use of web search). Finally, we acknowledge that while predicting slow judgments is an important task for the future of AI, it may be difficult to make progress on today. 6 Acknowledgements Figure 1 : 1 Figure1: Example questions. In Fermi, people guess whether the left-hand side is smaller than right-hand side. In Politifact they guess whether the speaker's statement is true. \n Figure 2 : 2 Figure 2: Question-answering UI for Fermi Comparisons (above) and Politifact Truth (below) tasks. Users are not allowed to do online research for Fermi but are for Politifact. \n Figure 3 : 3 Figure3: Correlation between fast and slow judgments for a given question-user pair for Fermi (above) and Politifact (below). Blue markers are proportional in size to number of instances and orange line is a linear regression through the data. Fast user judgments are often 50% but then become more confident in either direction given more time to think. \n Mean squared error of user's answers, FermiFast Slow \n Figure 4 : 4 Figure 4: Variation in User Performance in Fermi for fast and slow judgments.Histogram shows number of users who obtained a particular level of performance (measured by MSE). Note that 0.25 is the MSE achieved by a user who always says 50%. While users have widely varying performance, much of this is due to noise and we can't rule out the possibility that most users do not vary in actual skill (see Figure5). \n Figure 5 : 5 Figure 5: Mean performance on Fermi for different user quantiles. Users were divided into quantiles based on performance on a random half of the data and the figure shows their performance on the other half. Error bars show standard error from 5 random divisions into quantiles.The graph suggests we can't rule out the null hypothesis that the best two-thirds of users don't vary in their accuracy. (We removed users with less than k=6 judgments; the graphs were similar when k was set higher.) \n Figure 6 : 6 Figure 6: Variation in question difficulty for human users. Histograms show number of questions with a given MSE across users. Many questions had an MSE near zero (most users got them confidently right), while others were difficult (MSE near 0.25) or counter-intuitive (MSE greater than 0.25). \n Table 1 : 1 Mean squared test error for various algorithms and different levels of masking slow judgments at training time. Note that Clairvoyant models had access to groundtruth and other models did not. Politifact Fermi Unmasked Light Heavy Unmasked Light Heavy KNN CF 0.127 0.130 0.133 0.115 0.125 0.134 SVD CF 0.124 0.126 0.127 0.102 0.113 0.112 Linear Neural CF 0.131 0.135 0.135 0.137 0.141 0.141 Neural CF 0.130 0.131 0.129 0.136 0.136 0.138 Hierarchical Lin. Reg. 0.123 0.126 0.127 0.098 0.107 0.114 Clairvoyant 0.242 0.242 0.242 0.216 0.216 0.216 Clairvoyant Mean 0.112 0.112 0.112 0.111 0.111 0.111 Always Guess 50% 0.137 0.137 0.137 0.138 0.138 0.138 4 Discussion: Mitigating Dataset Problems Some problems with our dataset were described in Section 2.4. How could these problems be mitigated in future work? \n\t\t\t Ng intends it as a heuristic rather than a rigorous scientific conclusion. 2  The distinction is similar to that between System 1 (fast) and System 2 (slow) cognition [6, 7] . However, in this work we distinguish judgments by how long they take and whether they make use of external information and not by the underlying cognitive process. For instance, fast judgments can depend on quick application of analytical reasoning. \n\t\t\t This is just meant as an example of a recommender system and is not a comment on the actual YouTube algorithm. A paper [15]  on YouTube recommendations states that they optimize for whether users follow a recommendation (weighted by user watch-time). This will mostly depend on fast judgments. 4  It could be argued that YouTube, Facebook and other sites are optimizing for being entertaining and keeping users on the site and that these are well predicted by fast judgments. Yet it's clear that users sometimes seek content that they would rate highly after a slow judgment (e.g. for educational purposes, for help making a business decision). So the question remains how to build ML algorithms for this task. 5  There is a large and varied literature aiming to create algorithms that perform sophisticated reasoning. Here are some recent highlights: [28, 29, 30, 31, 32, 33, 34, 35, 36, 37]   \n\t\t\t In doing a long calculation we might have no idea of the answer until we solve the whole problem. For many real-world problems our quick guesses are somewhat accurate and gradually improve with time.7 That is, the human provides both a guess and a measure of confidence. 8  Code is available at https://github.com/oughtinc/psj \n\t\t\t Both tasks require humans to decide some objective matter of fact. Yet there is no requirement that slow judgments be about objective facts: e.g. a student's judgment about a lecture is partly based on their own preferences and interests. 10  The important real-world problems of predicting slow judgments in Section 1.1 are plausibly AI-complete. \n\t\t\t A paper [50]  on adversarial examples for humans does record human judgments under different time constraints. But this domain is not AI-complete and the slower judgments are still pretty fast. \n\t\t\t The test set is randomly sampled from the entire dataset and so the same users and questions can appear both in train and test. \n\t\t\t This is like the standard practice in semi-supervised learning, where researchers remove all but 5% of the labels from the training set to simulate a situation where most data is unlabeled16  In order to achieve clear-cut separation of the training data and the held-out test set we also implemented a masking procedure before doing the additional masking for Light and Heavy. For each judgment in the test set, we removed the medium judgment made by the same user about the same question from the training set. We also stochastically remove the corresponding fast judgment with 80% probability. \n\t\t\t Like Hierarchical Linear Regression, the Linear Neural CF assumes that judgments evolve linearly over time. However, Linear Neural CF does not build in which data-points to regress on and wasn't able to learn this. 18  We experimented with various language models (results not shown). We found it difficult to train or fine-tune these models on our small dataset without overfitting. \n\t\t\t This would also allow a series of intermediate times between fast and slow. 20  We collected data for a third task (not discussed elsewhere in this report), where the aim was to evaluate Machine Learning papers. We asked researchers to judge papers subjectively (\"How useful is this paper for your research?\") rather than objectively (\"Should the paper be accepted for a conference?\"). Unfortunately we did not a sufficient amount of data from volunteers. But we think some kind of variant on this task would be worth doing. \n\t\t\t It remains to be seen how well this can work if the goal is to predict slow judgments.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/predicting-judgments-tr2018.tei.xml", "id": "6928537565846a71f913b6e494946895"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "The most extreme risk are those that threaten the entirety of human civilization, known as global catastrophic risks. The very extreme nature of global catastrophes makes them both challenging to analyze and important to address. They are challenging to analyze because they are largely unprecedented and because they involve the entire global human system. They are important to address because they threaten everyone around the world and future generations. Global catastrophic risks also pose some deep dilemmas. One dilemma occurs when actions to reduce global catastrophic risk could harm society in other ways, as in the case of geoengineering to reduce catastrophic climate change risk. Another dilemma occurs when reducing one global catastrophic risk could increase another, as in the case of nuclear power reducing climate change risk while increasing risks from nuclear weapons. The complex, interrelated nature of global catastrophic risk suggests a research agenda in which the full space of risks are assessed in an integrated fashion in consideration of the deep dilemmas and other challenges they pose. Such an agenda can help identify the best ways to manage these most extreme risks and keep human civilization safe.", "authors": ["Seth D Baum", "Anthony M Barrett"], "title": "Global Catastrophes: The Most Extreme Risks", "text": "Introduction The most extreme type of risk is the risk of a global catastrophe causing permanent worldwide destruction to human civilization. In the most extreme cases, human extinction could occur. Global catastrophic risk (GCR) is thus risk of events of the highest magnitude of consequences, and the risks merit serious attention even if the probabilities of such events are low. Indeed, a growing chorus of scholars rates GCR reduction as among the most important priorities for society today. Unfortunately, many analysts also estimate frighteningly high probabilities of global catastrophe, with one even stating \"I think the odds are no better than fifty-fifty that our present civilization on Earth will survive to the end of the present century\"  (Rees 2003:8) . Regardless of what the probabilities are, it is clear that humanity today faces a variety of serious GCRs. To an extent, humanity always has faced GCRs, in the form of supervolcano eruptions, impacts from large asteroids and comets, and remnants of stellar explosions. Events like these have contributed to several mass extinction events across Earth's history. The Toba volcano eruption about 70,000 years ago may have come close to bringing the human species to a premature end. However, scholars of GCR generally believe that today's greatest risks derive from human activity. These GCRs include war with biological or nuclear weapons, extreme climate change and other environmental threats, and misuse or accidents involving powerful emerging technologies like artificial intelligence and synthetic biology. These GCRs threaten far greater destruction than was seen in the World Wars, the 1918 flu, the Black Death plague, or other major catastrophes of recent memory. The high stakes and urgent threats of GCR demand careful analysis of the risks and the opportunities for addressing them. However, several factors make GCR difficult to analyze. One factor is the unprecedented nature of global catastrophes. Many of the catastrophes have never occurred in any form, and of course no previous global catastrophe has ever destroyed modern human civilization. The lack of precedent means that analysts cannot rely on historical data as much as they can for smaller, more frequent events. Another factor is the complexity of GCRs, involving global economic, political, and industrial systems, which present difficult analytical decisions about which details to include. Finally, the high stakes of GCR pose difficult dilemmas about the extent to which GCR reduction should be prioritized relative to other issues. In this paper we present an overview of contemporary GCR scholarship and related issues for risk analysis and risk management. We focus less on the risks themselves, each of which merits its own dedicated treatment. Other references are recommended for the risks, perhaps the best of which are the relevant chapters of  Bostrom and Ćirković (2008) . Instead, our focus here is on overarching themes of importance to the breadth of the GCRs. The following section defines GCR in more detail and explains why many researchers consider it to be so important. Next, some of the analytical challenges that GCR poses and the techniques that have been developed to meet these challenges are explained. There follows a discussion of some dilemmas that arise when GCR reduction would require great sacrifice or would interfere with each other. Finally, conclusions are drawn. \n What Is GCR And Why Is It Important? Taken literally, a global catastrophe can be any event that is in some way catastrophic across the globe. This suggests a rather low threshold for what counts as a global catastrophe. An event causing just one death on each continent (say, from a jet-setting assassin) could rate as a global catastrophe, because surely these deaths would be catastrophic for the deceased and their loved ones. However, in common usage, a global catastrophe would be catastrophic for a significant portion of the globe. Minimum thresholds have variously been set around ten thousand to ten million deaths or $10 billion to $10 trillion in damages  (Bostrom and Ćirković 2008) , or death of one quarter of the human population  (Atkinson 1999; Hempsell 2004 ). Others have emphasized catastrophes that cause long-term declines in the trajectory of human civilization  (Beckstead 2013) , that human civilization does not recover from  (Maher and Baum 2013) , that drastically reduce humanity's potential for future achievements  (Bostrom 2002 , using the term \"existential risk\"), or that result in human extinction  (Matheny 2007; Posner 2004) . A common theme across all these treatments of GCR is that some catastrophes are vastly more important than others. Carl Sagan was perhaps the first to recognize this, in his commentary on nuclear winter  (Sagan 1983) . Without nuclear winter, a global nuclear war might kill several hundred million people. This is obviously a major catastrophe, but humanity would presumably carry on. However, with nuclear winter, per Sagan, humanity could go extinct. The loss would be not just an additional four billion or so deaths, but the loss of all future generations. To paraphrase Sagan, the loss would be billions and billions of lives, or even more. Sagan estimated 500 trillion lives, assuming humanity would continue for ten million more years, which he cited as typical for a successful species. Sagan's 500 trillion number may even be an underestimate. The analysis here takes an adventurous turn, hinging on the evolution of the human species and the long-term fate of the universe. On these long time scales, the descendants of contemporary humans may no longer be recognizably \"human\". The issue then is whether the descendants are still worth caring about, whatever they are. If they are, then it begs the question of how many of them there will be. Barring major global catastrophe, Earth will remain habitable for about one billion more years until the Sun gets too warm and large. The rest of the Solar System, Milky Way galaxy, universe, and (if it exists) the multiverse will remain habitable for a lot longer than that  (Adams and Laughlin 1997) , should our descendants gain the capacity to migrate there. An open question in astronomy is whether it is possible for the descendants of humanity to continue living for an infinite length of time or instead merely an astronomically large but finite length of time (see e.g.  Ćirković 2002; Kaku 2005) . Either way, the stakes with global catastrophes could be much larger than the loss of 500 trillion lives. Debates about the infinite vs. the merely astronomical are of theoretical interest  (Ng 1991; Bossert et al. 2007 ), but they have limited practical significance. This can be seen when evaluating GCRs from a standard risk-equals-probability-times-magnitude framework. Using Sagan's 500 trillion lives estimate, it follows that reducing the probability of global catastrophe by a mere one-in-500-trillion chance is of the same significance as saving one human life. Phrased differently, society should try 500 trillion times harder to prevent a global catastrophe than it should to save a person's life. Or, preventing one million deaths is equivalent to a one-in-500-million reduction in the probability of global catastrophe. This suggests society should make extremely large investment in GCR reduction, at the expense of virtually all other objectives. Judge and legal scholar Richard Posner made a similar point in monetary terms  (Posner 2004 ). Posner used $50,000 as the value of a statistical human life (VSL) and 12 billion humans as the total loss of life (double the 2004 world population); he describes both figures as significant underestimates. Multiplying them gives $600 trillion as an underestimate of the value of preventing global catastrophe. For comparison, the United States government typically uses a VSL of around one to ten million dollars  (Robinson 2007) . Multiplying a $10 million VSL with 500 trillion lives gives $5x10 21 as the value of preventing global catastrophe. But even using \"just\" $600 trillion, society should be willing to spend at least that much to prevent a global catastrophe, which converts to being willing to spend at least $1 million for a one-in-500-million reduction in the probability of global catastrophe. Thus while reasonable disagreement exists on how large of a VSL to use and how much to count future generations, even low-end positions suggest vast resource allocations should be redirected to reducing GCR. This conclusion is only strengthened when considering the astronomical size of the stakes, but the same point holds either way. The bottom line is that, as long as something along the lines of the standard riskequals-probability-times-magnitude framework is being used, then even tiny GCR reductions merit significant effort. This point holds especially strongly for risks of catastrophes that would cause permanent harm to global human civilization. The discussion thus far has assumed that all human lives are valued equally. This assumption is not universally held. People often value some people more than others, favoring themselves, their family and friends, their compatriots, their generation, or others whom they identify with. Great debates rage on across moral philosophy, economics, and other fields about how much people should value others who are distant in space, time, or social relation, as well as the unborn members of future generations. This debate is crucial for all valuations of risk, including GCR. Indeed, if each of us only cares about our immediate selves, then global catastrophes may not be especially important, and we probably have better things to do with our time than worry about them. While everyone has the right to their own views and feelings, we find that the strongest arguments are for the widely held position that all human lives should be valued equally. This position is succinctly stated in the United States Declaration of Independence, updated in the 1848 Declaration of Sentiments: \"We hold these truths to be self-evident: that all men and women are created equal\". Philosophers speak of an agent-neutral, objective \"view from nowhere\" (Nagel 1986) or a \"veil of ignorance\"  (Rawls 1971)  in which each person considers what is best for society irrespective of which member of society they happen to be. Such a perspective suggests valuing everyone equally, regardless of who they are or where or when they live. This in turn suggests a very high value for reducing GCR, or a high degree of priority for GCR reduction efforts. \n Challenges To Analyzing GCR Given the goal of reducing GCR, one must know what the risks are and how they can be reduced. This requires diving into the details of the risks themselves-details that we largely skip in this paper-but it also requires attention to a few analytical challenges. The first challenge is the largely unprecedented nature of global catastrophes. Simply put, modern human civilization has never before ended. There have been several recent global catastrophes of some significance, the World Wars and the 1918 flu among them, but these clearly did not knock civilization out. Earlier catastrophes, including the prehistoric mass extinction events, the Toba volcanic eruption, and even the Black Death plague, all occurred before modern civilization existed. The GCR analyst is thus left to study risks of events that are in some way untested or unproven. But the lack of historical precedent does not necessarily imply a lack of ongoing risk. Indeed, the biggest mistake of naïve GCR analysis is to posit that, because no global catastrophe has previously occurred, therefore none will occur. This mistake comes in at least three forms. The first and most obviously false form is to claim that unprecedented events never occur. In our world of social and technological innovation, it is easy to see that this claim is false. But accounting for it in risk analysis still requires some care. One approach is to use what is known in probability theory as zero-failure data  (Hanley 1983; Bailey 1997; Quigley and Revie 2011) . Suppose that no catastrophe has occurred over n prior time periods-for example, there has been no nuclear war in the 65 years since two countries have had nuclear weapons. (The second country to build nuclear weapons was the Soviet Union, in 1949.) It can thus be said that there have been zero failures of nuclear deterrence in 65 cases. An approximate upper bound can then be estimated for the probability p of nuclear deterrence failure, i.e. the probability of nuclear war, occurring within an upcoming year. Specifically, p lies within the interval [0, u] with (1 -α) confidence, where u = 1 -α (1/n) gives the upper limit of the confidence interval. Thus for 95% confidence (α = 0.05), u = 1-0.05 (1/65) = 0.05, meaning that there is a 95% chance that the probability of nuclear war within an upcoming year is somewhere between 0 and 0.05. Note that this calculation assumes (perhaps erroneously) that the 65 non-failures are independent random trials and that p is approximately constant over time, but it nonetheless provides a starting point for estimating the probability of unprecedented events.  Barrett et al. (2013)  uses a similar approach as part of a validation check of a broader risk analysis of U.S.-Russia nuclear war. The second form of the mistake is to posit that the ongoing existence of human civilization proves that global catastrophes will not occur. It is true that civilization's continued existence despite some past threats should provide some comfort, but it should only provide some comfort. Consider this: if a global catastrophe had previously occurred, nobody would still be around to ponder the matter (at least for catastrophes causing human extinction). The fact of being able to observe one's continued survival is contingent upon having survived. While it is easy to see that this is a mistake, it is harder to correct for it. Again, it requires careful application of probability theory, correcting for what is known as an observation selection effect  (Bostrom 2002b (Bostrom , Ćirković et al. 2010 . The basic idea is to build the existence of the observer into probability estimates for catastrophes that would eliminate future observers. The result is probability estimates unbiased by the observer's existence, with global catastrophe probability estimates typically revised upwards. The third form of the mistake is to posit that, because humanity has survived previous catastrophes, or risks of catastrophes, therefore it will survive future ones. This mistake is especially pervasive in discussions of nuclear war. People sometimes observe that no nuclear war has ever occurred and cite this as evidence to conclude that therefore nuclear deterrence and the fear of mutually assured destruction will indefinitely continue to keep the world safe (for discussion see  Sagan and Waltz 2013) . But there have been several near misses, from the 1962 Cuban missile crisis to the 1995 Norwegian rocket incident, and there is no guarantee that nuclear war will be avoided into the distant future. Similarly, just because no pandemic has ever killed the majority of people (Black Death killed about 22%), or just because early predictions about the rise of artificial intelligence proved false (they expected human-level AI within decades that have long since come and gone; see  Crevier 1993; McCorduck 2004) , it does not necessarily follow that no pandemics would be so lethal, or that AI cannot reach the lofty heights of the early predictions. Careful risk analysis can correct for the third form by looking at the full sequences of events that would lead to particular global catastrophes. For example, nuclear weapons in the United States are launched following a sequence of decisions by increasingly high ranking officials, ultimately including the President. This decision sequence can be built into a risk model, with model parameters estimated from historical data on how often each step in the decision sequence has been reached  (Barrett et al. 2013) . The more often near misses have occurred, and the nearer the misses were, the higher the probability of an eventual \"hit\" in the form of a nuclear war. The same analytic structure can be applied to other GCRs. But for many aspects of GCRs, as with other low-probability risks, there is not enough historical or other empirical data to fully characterize the risk. A good example of this is the risk from AI. The concern is that AI with human-level or super-human intelligence could outsmart humanity, assume control of the planet, and inadvertently cause global catastrophe while pursuing whatever objectives it was initially programmed for  (Omohundro 2008 , Yudkowsky 2008 . While there is reason to take this risk seriously-and indeed many do-assessing the risk cannot rely exclusively on empirical data, because no AI like this has ever existed. Characterizing AI risk thus requires expert judgment to supplement whatever empirical data is available  (Baum et al. 2011) . And while experts, like everyone else, are prone to make mistakes in making predictions and estimating the nature of the world, careful elicitation of expert judgment can reduce these mistakes and improve the overall risk analysis. That said, for GCR analysis it can be especially important to remember the possibility of experts being wrong. Indeed, for very low probability GCRs, this possibility can dominate the analysis, even when experts have wide consensus and high confidence in their conclusions, and even when the conclusions have significant theoretical and empirical basis  (Ord et al. 2010 ). It can be similarly important to remember the possibility that experts with outlier opinions can be right  (Ćirković 2012) . Ordinarily, these possibilities would not merit significant attention, but the high stakes of GCR means that even remote possibilities can warrant at least some scrutiny. A different type of analytical challenge comes from the global nature of GCRs, which makes them especially complex risks to analyze. GCRs are driven variously by the biggest geopolitical rivalries (in the case of biological or nuclear war), advanced research and development and the advantages it can confer (in the case of emerging technologies), or the entire global industrial economy (in the case of environmental collapses). Likewise, the impacts of global catastrophes depend on the resilience of global human civilization to major systemic shocks, potentially including major losses of civil infrastructure, manufacturing, agriculture, trade, and other basic institutions that enable the existence and comforts of modern civilization  (Maher and Baum 2013) . Assessing the extent of GCR requires accounting for the complexities all these disparate factors plus many others. Of course it is not possible to include every detail in any risk analysis, and certainly not for global risks. One must always focus on the most important parts. Here it is helpful to recall the high stakes associated with the most severe global catastrophes: the ones that would cause permanent harm to human civilization. While these catastrophes can be highly multifaceted, with a wide variety of impacts, the one impact that stands out as particularly important is that permanent harm. Other impacts, and the causes of those impacts, are simply less important. A GCR analysis can focus on the permanent harm and its causes. Climate change is an excellent example of a highly complex GCR. Climate change is caused mainly by a wide range of industrial, agricultural, deforestation, and transportation activities, which in turn are connected to a large portion of the activities that people worldwide do on a daily basis. The impacts of climate change are equally vast, affecting meteorological, ecological, and human systems worldwide, causing everything from increased storm surge to increased incidence of malaria. Each of these various impacts is important to certain people and certain ecosystems, but most of them will not make or break humanity's success as a civilization. Instead, the GCR analyst can look directly at worst-case global catastrophe scenarios, such as the possibility of temperatures exceeding the thermal limits of mammals across much of the world, in which case mammals in those regions not in air conditioning will overheat and die (Shewood and Huber 2010). Thus a focus on GCR can make the overall analysis easier. A subtler complexity, which GCR scholarship is only just starting to address, is the systemic nature of certain GCRs. Most GCR scholarship treats each risk as distinct: there could be a nuclear war, or there could be catastrophic climate change, or there could be an AI catastrophe, and so on. But these risks do not exist in isolation. They may be caused by the same phenomena, such as a quest for economic growth (causing both climate change, via industrial activity, and AI, via commercial technology development). Or they may cause each other, such as in the concern that climate change will lead to increased violent conflict  (Gemenne et al. 2014) . They may also have interacting impacts, such as if a war or other catastrophe hits a population already weakened by climate change. These various interactions suggest a systems approach to GCR analysis  (Baum et al. 2013; Hanson 2008; Tonn and MacGregor 2009) , just as interactions among other risks suggest a systems approach to risk analysis in general  (Haimes 2012; Park et al. 2013) . Systemic effects further suggests that global catastrophe could be caused by relatively small events whose impacts cascade into a global catastrophe. Similar systemic effects can already be seen across a variety of contexts, such as the 2003 Italy power outage caused by trees hitting two power lines in Switzerland, with effects cascading across the whole system  (Buldyrev et al. 2010) . Just as these systems proved fragile to certain small disturbances, perhaps the global human civilization could too. Characterizing these global systemic risks can give a clearer understanding of the totality of the GCRs that human civilization now faces, and can also help identify some important opportunities to reduce or otherwise manage the risks. \n Some GCR Dilemmas Unfortunately, managing GCR is not always as simple as analyzing the risks and identifying the risk management options. Some GCR management options pose deep dilemmas that are not easily resolved, even given full information about the risks. The bottom line is that no matter how successful GCR analysis gets, society still faces some difficult decisions. One dilemma arises from the very high stakes of GCR, or rather the very high magnitude of damages associated with permanent harm to human civilization. The high magnitude suggests that GCR reduction efforts should be prioritized over many other possible courses of action. Sometimes prioritizing GCR reduction efforts is not a significant concern, when the efforts would not be difficult or when they would be worth doing anyway, such as reducing climate change risk by improving energy efficiency. However, sometimes GCR reductions would come at a significant cost. In these cases society may think twice about whether the GCR reductions are worth it, even if the GCR reductions arguably should take priority given the high magnitude of global catastrophe damages. An example of such a dilemma can be found for climate change and other environmental risks. Because these risks are driven by such a large portion of human activity, reducing the risks could mean curtailing quite a lot of these activities. Society may need to, among other things, build new solar and wind power supplies, redesign cities for pedestrians and mass transit, restructure subsidies for the agriculture and energy sectors, and accept a lower rate of economic growth. Individuals may need to, among other things, change their diets, modes of transportations, places of residence, and accept a simpler lifestyle. Such extensive efforts may pass a cost-benefit test  (Stern 2007) , especially after accounting for the possibility of global catastrophe  (Weitzman 2009 ), but many people may still not want to do them. Indeed, despite the increasingly stark picture painted by climate change research, the issue still does not rank highly on the public agenda  (Pew 2014) . Should aggressive effort nonetheless be taken to reduce greenhouse gas emissions and protect the environment? This is a difficult dilemma. A similar dilemma arises for one proposed solution to climate change: geoengineering. Geoengineering is the intentional manipulation of the global Earth system  (Caldeira et al. 2013) . A prominent form of geoengineering would inject aerosol particles into the stratosphere in order to block incoming sunlight, thereby reducing temperatures at the surface. While this stratospheric geoengineering could not perfectly compensate for the climatic changes from greenhouse gas emissions, it probably could help avoid some of the worst damages, such as the damages of exceeding the thermal limits of mammals. However, stratospheric geoengineering comes with its own risks. In particular, if society stops injecting particles into the stratosphere, then temperatures rapidly rise back to where they would have been without any geoengineering  (Matthews and Caldeira 2007) . The rapid temperature increase could be very difficult to adapt to, potentially causing an even larger catastrophe than regular climate change. This creates a dilemma: Should society run the risk of geoengineering catastrophe, or should it instead endure the pains of regular climate change  (Baum et al. 2013 )? Given how bad climate change could get, this makes for another difficult dilemma. The high stakes of GCR suggest that society should do whatever would minimize GCR, and accept whatever suffering might follow. This could mean taking aggressive action to protect the environment, or, if that does not happen, suffering through climate change instead of attempting geoengineering. Looking at the analysis on paper, it is easy to recommend minimizing GCR: the numbers simply point heavily in that direction. But in practice, this would not be an easy recommendation to make, asking many people to make such a great sacrifice. Hopefully, clever solutions can be found that will avoid these big dilemmas, but society should be prepared to make difficult choices if need be. Another type of dilemma occurs when multiple GCRs interact with each other. Sometimes one action can reduce multiple GCRs. However, sometimes an action would reduce one GCR while increasing another. This poses a dilemma between the two GCRs, a risk-risk tradeoff  (Graham and Wiener 1995) . A good example of this type of dilemma is nuclear power. Nuclear power can help reduce climate change risk by shifting electricity production away from coal. However, nuclear power can also increase nuclear war risk by enabling nuclear weapons proliferation. This dilemma is seen most clearly in ongoing debates about the nuclear program of Iran. While Iran claims that its program is strictly for peaceful electricity and medical purposes, many observers across the international community believe that Iran is using its program to build nuclear weapons. Given the risks from climate change and nuclear war, should nuclear power be promoted? How much should it be promoted, and in what circumstances? Resolving these dilemmas requires quantifying and comparing the two GCRs to identify when nuclear power would result in a net decrease in GCR. Unfortunately, at this time GCR research has not quantified climate change and nuclear war risk well enough to be able to make the comparison and reach conclusions about nuclear power. Meanwhile, in circumstances in which nuclear power would not create a nuclear weapons proliferation risk, such as for countries that already have nuclear weapons or clearly do not want them, nuclear power probably would bring a net GCR reduction. This conclusion brings up the first type of dilemma -the general sacrifice for GCR reductionwhere nuclear power raises concerns about industrial accidents like Chernobyl or Fukushima. Such accidents are \"only\" local (or regional) catastrophes, but they are nonetheless plenty large enough to weigh heavily in decision making. \n Conclusions: A Research Agenda Regardless of whether GCR reduction should be prioritized above everything else, it is clear that GCR is an important issue and that reducing GCR merits some effort. The big research question then is, which efforts? That is, what are the best, most effective, most desirable ways to reduce GCR? Unfortunately, the GCR research community has not yet made significant advances to answering this vital question. A new research agenda is needed for it. We believe that GCR research is most helpful at guiding GCR reduction efforts when the research covers all the major risks and risk reduction options in a consistent, transparent, and rigorous manner. It should to include all the major risks and risk reduction options in order to identify which ones are most important and most worth pursuing. Analyzing each risk in isolation fails to account for their various systemic interactions and prevents evaluating risk-risk dilemmas like that posed by nuclear power. In contrast, integrating all the risks and risk reduction options into one assessment can help decision makers identify the best options. Similar integrated assessments have been done for a variety of other topics, such as the popular economic assessments of climate change (e.g.  Nordhaus 2008 ). Something along these lines, adapted for the particulars of GCR, would be of great value to GCR reduction decision making. An integrated assessment of GCR poses its own analytical challenges. The particulars of different GCRs can be quite different from each other. Fitting together e.g. a climate model, an epidemiological model, and a technological development model would not be easy, nor would filling in the important gaps that inevitably appear between the models. Each GCR is full of rich complexity; the full system of GCRs is more complex still. This makes it even more important to focus on the most important aspects of GCR, lest the analysis get bogged down in details. It is equally important for the analysis to focus on risk reduction options that are consistent with what attentive decision makers are in a position to do. Otherwise the analysis risks irrelevance. While this is true for any analysis, it is especially true for GCR. The global scale of GCR makes it tempting for analysis to ignore details that seem small relative to the risk but are important to decision makers, and also to entertain risk reduction options that perform well in a theoretically ideal world, \"if only everyone would do that\". Furthermore, the high stakes of GCR makes it tempting for analysis to recommend rather drastic actions that go beyond what most people are willing to do. It is thus that much more important for GCR analysis to remain in close touch with actual decision makers, to ensure that the analysis can help inform actual decisions. Despite these challenges, we believe that such a research agenda is feasible and can make an important contribution to society's overall efforts to reduce GCR. Indeed, the future of civilization could depend on it. \t\t\t Electronic copy available at: https://ssrn.com/abstract=<PHONE>", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/SSRN-id3046668.tei.xml", "id": "295544b27112be9729dbde1ef93e39f2"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "The international system is at an artificial intelligence fulcrum point. Compared to humans, AI is often faster, fearless, and efficient. National security agencies and militaries have been quick to explore the adoption of AI as a new tool to improve their security and effectiveness. AI, however, is imperfect. If given control of critical national security systems such as lethal autonomous weapons, buggy, poorly tested, or unethically designed AI could cause great harm and undermine bedrock global norms such as the law of war. To balance the potential harms and benefits of AI, international AI arms control regulations may be necessary. Proposed regulatory paths forward, however, are diverse. Potential solutions include calls for AI design standards to ensure system safety, bans on more ethically questionable AI applications, such as lethal autonomous weapon systems, and limitations on the types of decisions AI can make, such as the decision to use force. Regardless of the chosen regulatory scheme, however, there is a need to verify an actor's compliance with regulation. AI verification gives teeth to AI regulation. This report defines \"AI Verification\" as the process of determining whether countries' AI and AI systems comply with treaty obligations. \"AI Verification Mechanisms\" are tools that ensure regulatory compliance by discouraging or detecting the illicit use of AI by a system or illicit AI control over a system. This report introduces and explains why these mechanisms have potential to support an AI verification regime. However, further research is needed to fully assess their technical viability and whether they can be implemented in an operationally practical manner.", "authors": ["Matthew Mittelsteadt"], "title": "AI Verification Mechanisms to Ensure AI Arms Control Compliance", "text": "Despite the importance of AI verification, few practical verification mechanisms have been proposed to support most regulation in consideration. Without proper verification mechanisms, AI arms control will languish. To this end, this report seeks to jumpstart the regulatory conversation by proposing mechanisms of AI verification to support AI arms control. Due to the breadth of AI and AI arms control policy goals, many approaches to AI verification exist. It is well beyond the scope of this report to focus on all possible options. For the sake of brevity, this report addresses the subcase of verifying whether an AI exists in a system and if so, what functions that AI could command. This report also focuses on mechanical systems, such as military drones, as the target of regulation and verification mechanisms. This reflects the focus of a wide range of regulatory proposals and the policy goals of many organizations fighting for AI arms control. In sum, this report concentrates on verification mechanisms that support many of the most popular AI arms control policy goals. Naturally, other approaches exist and should be studied further, however; they are beyond the scope of this initial report on the subject. To these ends, this report presents several novel verification mechanisms including: System Inspection Mechanisms: Mechanisms to verify, through third party inspection, whether any AI exists in a given system and whether that AI could control regulated functions: • Verification Zone Inspections: An inspection methodology that uses limited-scope software inspections to verify that any AI in a system cannot control certain functions. The subsystems that AI must not control, for example, subsystems controlling the use of force, are designated as \"verification zones.\" If these select verification zones can be verified as free from AI control, the system as a whole is compliant. This limited inspection scope reduces the complexity of system inspections, protects subsystems irrelevant to AI regulation, and renders inspections less intrusive. • Hardware Inspections: The existence of AI in subsystems and its control over certain functions can be verified by examining whether AI chips exist and what subsystems they control. Sustained Verification Mechanisms: These are tools that can be used to verify a system remains compliant after an initial inspection: • Preserving Compliant Code with Anti-Tamper Techniques: These techniques protect system software from post-inspection tampering that may alter what AI can control. Methods chosen to illustrate such techniques include cryptographic hashing of code and code obfuscation. Cryptographically hashed system software also provides a record of expected system design for inspectors that can be used to monitor system software compliance long term. • Continuous Verification through Van Eck Radiation Analysis: Verified systems can be affixed with a Van Eck radiation monitoring mechanism that can be used to monitor the radiation the system produces when code is run. Aberrations detected in this radiation could indicate non-compliant manipulation. \n Introduction Artificial intelligence has emerged as one of the most notable new national security technologies. While diverse in form and function, the common thread that unites the constellation of AI technologies is the unique power to grant autonomy and beyond-human insight to a range of systems and processes. The national security benefits of AI are obvious. One can easily imagine, and in many cases already see, the powerful new intelligence tools and military capabilities driven by AI. In the right hands, AI systems promise invaluable security benefits; in the wrong hands, they could be a grave threat. If given control of critical national security systems such as lethal autonomous weapons, buggy, poorly tested, or unethically designed AI could cause great harm and undermine bedrock global norms such as the law of war. Few argue AI systems be left unconstrained. To balance the potential harms and benefits of AI, international AI arms control regulations may be necessary. If used, AI systems must be safe, adhere to the law, and have an overall net benefit. To these commonly accepted ends, various policy goals and regulations have been proposed to delimit acceptable design and use of AI systems under international law. Call it \"AI arms control.\" Discourse, however, has sputtered in part due to a lack of practical mechanisms to verify a system's compliance with proposed international regulations. * The standard national security term of art for ensuring compliance with arms control is \"verification.\" The United Nations Institute for Disarmament Research and other United Nations bodies define verification as \"the collection, collation and analysis of information in order to make a judgement as to whether a party is complying with its obligations.  1  To address a potential point of confusion, this same term is also used by computer scientists to describe the process of analyzing whether software behaves as expected. In this report, verification is used strictly in the national security sense. Without verification, international AI arms control lacks teeth. Effective AI arms control may continue to languish unless practical verification mechanisms are developed. This report seeks to jumpstart this discussion by proposing several novel AI verification mechanisms. In this report, I define \"AI Verification\" as: The processes of determining whether countries' AI and AI systems comply with treaty obligations. I define \"AI Verification Mechanisms\" as: The set of mechanisms that ensure regulatory compliance by discouraging or detecting the illicit use of AI by a system or illicit AI control over a system. As mentioned, AI is a broad class of technologies. Regulators cannot expect a single silver bullet that can eliminate AI threats. Regulations will likely have to be technology specific, and verification will have to use a range of overlapping mechanisms. Recognizing the breadth of this topic and the unwieldy task of addressing all possible regulatory outcomes, this report strives not to answer all questions and solve every problem, but to get the process started. For the sake of brevity, this report specifically addresses the subcase of verifying whether an AI exists in a system or subsystem and if so, what functions that AI could command. I focus on mechanical systems, such as military drones, as the target of regulation and verification mechanisms. I refer to these mechanical systems as simply \"systems\" or \"AI systems,\" as appropriate, throughout. These choices reflect the emphasis of a wide range of regulatory proposals and the policy goals of many organizations lobbying for AI arms control. These include regulations mandating a human remains in the loop for use of force decisions, bans on certain systems that should not be AI controlled, and other regulatory proposals that seek to limit what can use AI and what AI can do. In sum, this report concentrates on verification mechanisms that support many of the most popular AI arms control policy goals. Naturally, there are many other routes that can be taken to regulate AI systems. Other possible options include safety and control demonstrations, rules regulating thoroughness of testing, and the ubiquitous implementation of certain norms. These and other options are beyond the scope of this report but should be considered further by interested policymakers and researchers. This report is intended for those policymakers and national security leaders who oversee AI systems (and systems that may one day become AI systems) or negotiate international policy to control their use. It is important to highlight that the mechanisms contained within this report represent a regulatory starting point. Verification is technically complex, detailed, and politically difficult. Each mechanism should be passed to engineers for further research and to determine how it can be implemented in current or future state systems. This report is a list of possibilities, not answers. Policymakers must lean on their engineers and diplomats to rework and build on these ideas as needed to fit technical and political reality. \n A Note on Spoofing Many verification mechanisms could be subject to spoofing. In the context of AI verification, spoofing can be thought of as any method that an actor may use to deceive regulators and pursue illicit AI activities. It should be assumed that if an actor has the sufficient will and resources to cheat, they could develop spoofing methods to circumvent verification. This does not mean verification is fruitless, merely imperfect. Former Deputy Secretary of Defense Paul Nitze summarized these realities by stating the goal of effective verification is to make sure that: [I]f the other side moves beyond the limits of the Treaty in any militarily significant way, we would be able to detect such violation in time to respond effectively and thereby deny the other side the benefit of the violation. 2 A practical verification mechanism does not necessarily render spoofing impossible but seeks to catch spoofing in time to act. That said, this stated goal comes with the asterisk that even methods failing to guarantee this basic requirement can help. A second goal of verification is discouragement. If a verification mechanism instills a potential evader with a lack of \"certainty about the likelihood of discovery,\" it can still reduce harm.  3  This report is written with these goals in mind. Using these mechanisms, actors can better discourage and detect non-compliant activity, increase the cost of spoofing, and build trust in relevant regulations. \n Verification Inspection Mechanisms A common tool in pre-existing arms control verification mechanisms is thirdparty system inspections to verify compliance with regulation. This first set of mechanisms are intended to support inspections as a tool for AI verification. \n Verification Zones and Quarantine Zones Fundamental to inspections is the ability of third-party inspectors to quickly review a system and target the components that may make it non-compliant. As stated in the introduction, the goal of our verification mechanisms is to demonstrate whether a target system contains AI, and if so, whether certain critical functions of that system can be AI-controlled. This report defines a \"critical function\" loosely as those functions that regulators believe humans, rather than AI, should control. An example of such a function is the decision over the use of force, which many believe on ethical grounds should be reserved for humans. Unfortunately, conducting third-party inspections on military and national security systems is often not politically or technically easy and is often dismissed as too difficult to be worthwhile.  4  Several factors complicate matters, including: • Complexity: The high complexity of AI algorithms, and the lengthy code baked into mechanical systems in general, often renders system code difficult to decipher or even unintelligible to humans. Complexity poses a challenge to inspections, a core element of many verification regimes; if inspectors cannot understand the technology they are inspecting, they cannot verify compliance with the regulations they seek to enforce. For AI systems, this complexity is often rooted in the machine learning process, a process by which AI writes and/or adjusts its own algorithms. AI-written code emphasizes functionality, not understanding, creating software not even the most talented engineers, let alone inspectors, can decipher. • Invisibility: AI algorithms run invisibly, leaving no obvious visible markers. Even the physical technologies that underpin AI are opaque. There is no standard set of inputs that \"create AI.\" Each AI system is the unique result of a concert of sensors, communications links, semiconductors, and other technologies. This invisibility makes it difficult to determine whether an AI or a human controls a system or a given function. Further, invisibility poses challenges to the detection of the use and production of AI, complicating efforts to find proof of AI control. • Secrecy: \"AI Competition\" creates an imperative for states, corporations, and other actors to tightly guard AI algorithms to ensure their advantage.  5  The result has been intense secrecy and a strong disincentive for actors to openly identify which systems are AI systems, let alone expose their inner workings to verification inspectors. The guiding paradigm of AI verification and inspections has traditionally assumed all systems should be treated and inspected as monoliths. This view is the root of many of the above challenges and has led many experts to contend that AI verification requires full access to a system's software to establish the degree of AI control over it or its functions.  6  Under this assumption, to concretely establish whether a lethal autonomous drone, for instance, allows AI to control the use of force, an inspector would require full access to all software and components contained within. Without full access, the inspector could never be certain if an uninspected part of a system hides an AI program that surreptitiously controls this critical function. Light must be shed on all code to know for certain the system is regulatorily compliant. This view makes inspection an all or nothing task. The noted secrecy demands of national security and proprietary information, however, make full system access politically unlikely. Even if full access is granted, a system-wide inspection would be complex and unwieldy, especially if inspectors must examine long and often indecipherable algorithms. Working under this monolithic systems paradigm, verification does indeed seem impossible. Thankfully, mechanical systems are not monoliths. The term \"system\" is no mistake. In reality, any mechanical system is not a monolith but a series of overlapping \"subsystems\" that usually coordinate under the orchestration of an operating system.  7  Subsystems may include sensors, interfaces, communications, navigation, targeting, and firing, among others. Following this pattern of subdivision, each subsystem often uses its own algorithm(s); some use AI algorithms and others conventional algorithms. As such, a given system can contain no AI, one AI, or many AIs. When tasks are carried out, it is often the subsystems, and their controlling algorithms, that actually implement the overall system's tasks. The operating system merely acts as the system director, allocating resources, easing collaboration, and ensuring smooth operation. It is unlikely that state actors will consent to inspections of a system's operating system, as it is the centerpiece of any computing system and its design is usually highly secretive, especially in security critical national security systems. The easiest path forward appears to be accepting the operating system as a \"black box.\" That said, inspections still require a window into the overall system, a window that certain subsystems can provide. The recommendation of this report is to narrow inspection scope to only subsystems of interest to policy goals. This reduces the unwieldy complexity of inspections, balances the need for a certain level of system secrecy, and makes system inspections more palatable overall. \n Box 1: Apple iPhone Subsystems Apple iPhones include both conventional subsystems, such as the internet browser, and AI subsystems, such as Siri. While an iPhone contains an AI, Siri, it is just one application that works in concert with other applications to deliver the various functions and services expected from an iPhone. The iPhone's operating system takes on executive coordination functions, coordinating, triggering, and providing resources for the overlapping subsystems to make sure they work in concert. In sum, the iPhone itself is not an AI. One can imagine Apple would be much more willing to allow inspectors to view the code for a rudimentary subsystem such as its flashlight application, or any one of its applications for that matter, rather than wholesale access to all of an iPhone's code or its operating system. To formalize this mechanism, subsystems should be divided into one of two categories, \"quarantine zones\" and transparent \"verification zones,\" defined as: Quarantine zones: Subsystems that may contain the operating system, sensitive software, and other system functions intentionally obfuscated due to high secrecy demands. The inner workings of these zones should not matter for verification. Inspectors and policymakers should assume by default that AI exists within or controls these quarantine zones (even if it does not) until proven otherwise. Verification zones: These are subsystems that are intentionally made transparent. These zones can be physically or logically demarcated within the system through software, hardware, or both. These zones would allow inspectors access to the code or hardware of subsystems to determine from where they receive instructions. If controlling instructions can come from a quarantine zone, it should be assumed the functions in this zone could be AIcontrolled. If not, this suggests the subsystem(s) within cannot be controlled by an AI outside of a verification zone. Verification zones should surround subsystems involved in functions that should not be AI controlled or should not contain AI. To verify a human remains in the loop for use of force decisions, for instance, the weapons deployment system should be in a verification zone so an inspector can verify it is not AI controlled and can trace its instructions to a human source. When curtailing AI command and control, AI itself is often not the problem but rather the ability of AI to communicate with and control critical subsystems. If these critical subsystems are made transparent, inspectors could analyze their code and the source of their commands to determine whether they could come from an AI algorithm. If instructions flow from a quarantine zone, an AI could be sending those instructions, and the subsystem cannot be verifiably AI-free. If the instructions flow only from a transparent source in a verification zone, then inspectors can verify the subsystems within are AI-free or verify what functions are AI controlled within the zone. This mechanism allows for verification that skirts the need to dig through the wholesale code of a system. Through this limited scope, verification by inspection likely grows more politically feasible and practically manageable, while remaining effective. Verification Checkpoints Separation of these zones can be implemented either physically or digitally. Physically, zones can be isolated from one another if they do not have the physical means to communicate. This is known as air-gapping.  8  If the weapons system of a drone, for instance, is not wired to a subsystem containing illicit AI, the AI could not control a subsystem or its critical functions. Digitally, software could be designed to carefully control the flow of instructions and isolate certain software elements from others. I call these verification checkpoints. A verification checkpoint is an interface that stands between a given verification zone and all other subsystems. This checkpoint acts to sieve instructions. Using authentication and carefully managed intrasystem communications pipelines, a checkpoint could limit the instructions a verification zone can accept and from where it can accept instructions. If a quarantine zone subsystem is not authorized to communicate with a verification zone subsystem, an instruction checkpoint will halt those instructions. This protects the verification zone from AI within that quarantine zone that may be issuing those instructions. More critically, it demonstrates to an inspector that the AI in the quarantine zone cannot control functions in the verification zone. Forcing instructions through a checkpoint creates an intentional bottleneck. If all instructions flow through one point, this produces a centralized and convenient node to target all regulation and inspections. A checkpoint's code could be relatively concise and likely technically simple enough for an inspector to easily understand. It is possible that these bottlenecks may negatively impact performance, a tradeoff made in favor of enhanced transparency. Policymakers must consider whether this trade-off is worth it. As for the specific design of these checkpoints, there is a range of possibilities available. Such checkpoints could be designed to only allow instructions to flow into a verification zone from a select human source. Alternatively, checkpoints could be designed to give varying degrees of control to different sources, perhaps allowing for a benign AI in a quarantine zone, perhaps a navigation AI, to send a verification zone targeting coordinates while continuing to halt the quarantine zone from sending certain critical instructions, such as an instruction to \"fire.\" I call these mixed control checkpoints \"centaur checkpoints.\" Verification checkpoints reduce potentially unwieldy system inspections to targeted reviews that determine if the checkpoint could allow a quarantine zone, and therefore an AI, to send instructions to a regulated subsystem in a verification zone. By examining checkpoints, inspectors can identify the source of commands to verify if they come from a quarantine zone. If commands can be traced to a human-controlled verification zone, it can be determined that the subsystems beyond the checkpoint are AI-free. \n Policy Refinements and Assumptions Ideally, the quarantine/verification zone would be packaged with some additional policies. If possible, state actors would agree upon standard verification zones in their systems to open to inspectors. State actors must choose and define these zones carefully. To ensure an AI does not have control over critical decisions, states must agree what defines the subsystem that implements those decisions so that it can be clear what is regulated. In general, focus must be placed on subsystems of interest to regulation and those that contain enough information to verify the system will act according to policy goals. To build trust and respect secrecy demands, less is more. Policymakers should work with engineers to determine what quantity of subsystems would need to be verification zones in order to confidently verify a system. State actors may also wish to negotiate additional policies to ease inspection. Potential examples could include standard subsystem labels and physical markings to clarify component function and layout, designing subsystems to be easily accessed by an inspector, or mandating subsystem code be clearly described in code comments. This is by no means an exhaustive list; state actors should consider these and other policies that may help to ease inspections. This concept also rests on several assumptions about state actor behavior. It can be assumed that if a state actor knows a given piece of hardware or software is in a verification zone, the state actor will intentionally exclude topsecret technology from that zone's design. Note that opening these zones to inspection does not eliminate the possibility of technical advantage, but merely redirects all actors to finding their technical edge in other nontransparent subsystems. If engineers design systems with this in mind, a state actor should have few reservations about opening these zones to inspection. As addressed below in the \"challenges\" section, this may mean substantial system changes are required. Box 2: Inspecting Military Drones through Verification Zones Say an inspector wants to examine a military drone to verify it cannot use force without a human in the loop. To verify this, actors could agree on select verification zones in this system and open those zones to inspections. To do so, the inspector would require access to only those subsystems relating to the use of force, specifically the weapon deployment subsystem. This subsystem would therefore be a verification zone. Many subsystems, such as the drone's fuel management system, are likely unnecessary to this task and would be declared as uninspectable quarantine zones. To inspect this verification zone, the inspector may examine the weapons deployment subsystems and determine if it contains AI. Further, the inspector must verify which other subsystems may control weapons deployment and whether those systems are AI-controlled. To do so, the inspector can trace from where inputs into the verification zone flow: • If instructions are fed to the weapons system from an uninspectable quarantine zone, a zone that could contain AI, the inspector cannot verify with certainty that AI does not control the use of force. • If instructions flow from a transparent verification zone that does not contain AI, then the inspector can verify that AI does not control the use of force. This method allows the inspector to verify compliance while balancing the need for secrecy and reducing the scope of inspection. \n Challenges As mentioned, it is highly likely that these mechanisms cannot be easily implemented using current system designs. Many systems, while modular to an extent, are still highly interconnected, and it may be that certain subsystems cannot be easily subdivided from the rest to form these verification zones. Further, current systems aren't designed with checkpoints in mind, and therefore this could hinder efficiency if instructions must be funneled through this intentional bottleneck. While a hindrance, this problem is not something that should halt research and policy negotiation. Historically, it has taken many parties to arms control treaties significant time and research to implement the changes required. For example, the United States is still working through the substantial changes required by the Chemical Weapons Convention it signed in 1997 with a goal of completing disarmament by 2023.  9  This demonstrates that just because arms control verification requires heavy lifting and time does not render it impossible, nor does it mean steps should not be taken. Verification and arms control problems are technical and require technical changes. Therefore, it is likely these changes must be gradually implemented over time as new national security and military systems are developed. Cybersecurity is another challenge. With greater transparency comes greater cyber risk. When considering this mechanism, state actors must consider the security of their systems and the benefits they may receive by sacrificing some system security. Arms control usually has some cost, and it could be that cost is cybersecurity. Policymakers must note, however, that often security in one domain can be improved through a small sacrifice of security in another. Weigh this balance carefully. A final challenge is one of state actor trust and motivation. Any arms control scheme that seeks to balance secrecy and transparency will encounter international resistance. Change is hard, especially when it costs time, money, and potential military advantage. States must be motivated to buy into the concept and must trust the technical decisions made throughout the process. Without motivation and trust, negotiations will fail. To surmount these challenges, policymakers should work with engineers to research which checkpoint designs and system architectures will best balance policy goals against any detriments to cybersecurity and processing speeds. To improve trust and build international buy-in, policymakers should collaborate on this research with international partners. Doing so will ensure designs reflect the state of the art, build trust in the science behind any proposed changes, and encourage buy-in from the outset of the development process. \n Hardware Inspections If state actors wish to avoid software inspections and verify based only on hardware, they could do so by examining whether a given system, or subsystem, contains an AI chip. This mechanism can be implemented using the verification/quarantine zone model, restricting inspectors to only inspecting the hardware of subsystems relevant to their purpose. If a given subsystem is not powered locally by an AI chip and verifiably cordoned off from any external AI, it can be assumed the subsystem and its functions are not AI controlled. If the goal is to simply determine whether a system writlarge is AI-free, this mechanism is ideal as its somewhat less intrusive nature than software inspections * may make it more politically feasible to inspect an entire system without risking the exposure of most of the system's design secrets. The mix of chips that inspectors would look for in systems is diverse, ranging from general purpose Central Processing Units (CPUs) to so-called AI chips, which include Graphics Processing Units (GPUs), highly specialized Application-Specific Integrated Circuits (ASICs), and other AI optimizing processors and memory units.  10  In the world of AI, these various chips fall into two buckets: those used for AI training, when the AI is in development, and those used during inference, when the AI is in use. For AI system verification purposes, it is the inference chips that matter most as those are the chips that power AI algorithms when AI systems are deployed. At present and into at least the short term, it seems likely these chips will be required for AI system inference and therefore a clear indicator of the use of AI by that system. McKinsey estimates that within five years, 70 percent of the chips in AI systems will be ASICs, a specific type of AI chip exclusively used by AI applications, replacing the CPU as the AI processor of choice.  11  The takeaway is that within five years, most AI embedded systems could clearly be identified by their chipset. The reason for this shift, in brief, is that AI chips appear to offer many clear advantages. The use of these chips for AI processing and specifically AI * It is assumed that because the complexity and sensitive design details of hardware is often invisible to the naked eye an inspector could not easily steal or reveal substantial design secrets unless they had the time and resources to analyze the components in depth. Inspecting software, on the other hand, requires an inspector to read through the system's code, revealing all the design secrets contained within. It is likely certain secrecy risks, however, are not accounted for in this report. Policymakers should work with engineers to verify what secrets may be put at risk through a hardware inspection. inference largely derives from the special purpose features they provide for AI systems. These include (but are not limited to): 1. Greater parallelism. To analyze data efficiently, AI algorithms often must take advantage of greater parallelism than traditional CPUs can offer. These algorithms need to subdivide their complex input data and analytic models into smaller pieces so that analysis can be done in parallel, rather than sequentially. This massively speeds up operations and allows for quicker analysis. 2. Lower precision numbers: The operations AI algorithms perform often require fewer values after the decimal point in the numbers being calculated. These numbers are said to have \"lower precision.\" If a chip lowers the precision of the numbers being calculated, this increases the speed of each calculation, unlocking improved performance. 12 3. Faster memory access: AI has significant memory requirements. Fast memory allows an AI to think and act faster. AI chips facilitate this, often physically shortening the distance between processing and memory components to increase the speed of memory recall and, therefore, the speed of AI \"thinking.\" 13 4. Optimization for AI specific calculations: AI computations often involve a high volume of matrix multiplication. AI chips often optimize for this through such features as many parallel Multiply-Accumulate Circuits, which are special circuits that speed up this type of multiplication.  14  These features illustrate how AI chips can improve function in AI applications and why they may become an essential part of national security AI systems. As alluded to, inspections to verify the existence of AI in a system could be simple. All that is required is a physical examination of a system or subsystem to check if it contains AI chips. Verifying AI control over specific critical functions, such as an AI's ability to use force, is also potentially possible from this \"look and see\" method. If a critical subsystem, such as the weapons deployment system, does not include an AI chip and is verifiably cordoned off or air-gapped from other subsystems, it is likely not AI controlled and therefore compliant. While potentially effective, this mechanism is blunt. It is likely best used as a \"first pass\" inspection to determine whether a system requires greater scrutiny. Additional, more invasive analysis could then follow. \n Challenges This mechanism does not guarantee systems are entirely AI-free, as traditional CPUs can still run AI algorithms, albeit at an often slower pace. Furthermore, the utility of this mechanism could vary widely depending on system type. For lightweight applications, such as small drones, CPUs may be enough, and policymakers should account for this in their inspection policies. In this case, seeing that a system does not contain an AI chip is not enough for verification. However, for heavyweight applications, such as large drones and autonomous vehicles, the real time requirements of safety and time-critical functions require immensely fast processing to achieve real time action and reaction. This may require the faster AI-specific inference chips.  15  This speed imperative is even more crucial to military systems where the inability to react in a timely manner could lead to strategic failure or dangerous lag. It can be assumed that if a military deploys a multimillion-dollar autonomous system, it will use the chipset best equipped to reap the value of that investment. If current trends hold, this means those systems will use AI chips. In the end, capability is what matters. If a CPU-driven autonomous system is not capable of the processing speeds needed to be a true threat, it may be of little concern to regulators and useless to its owners. Do note that current trends may change. Refinements in CPUs and efficiency gains from new AI design techniques could allow systems to run on CPUs. If this mechanism is used, policymakers must ensure implementation keeps pace with technology change. Another challenge is distinguishability. This mechanism depends on AI chips exhibiting some quality that allows an inspector to determine they are AI chips. Given the breadth of AI-specific chips, this will likely require a range of qualities that inspectors can use to determine the classification of a given chip. As such, policymakers should work with engineers to delineate what qualities can be used and ensure this process conforms to chip designs. \n Sustained Verification A key challenge to verifying systems remain compliant with regulations post inspection is the ease with which system code can be altered. The inspection mechanisms above are certainly essential, especially for deterring noncomplaint activity. However, inspections only offer certainty in the very short term. Even if a system is found to be compliant, its software could be changed immediately after to embed it with AI or give an AI control of critical functions. Effective inspections and regulatory enforcement, therefore, require mechanisms to detect and deter such alterations. To craft a strong verification regime, policymakers should look to what I call \"sustained verification mechanisms,\" which help affirm previously inspected systems remain compliant long term. The following mechanisms are tailored to this purpose. \n Cryptographic Hashing One mechanism that can be used to protect system code from manipulation is anti-tamper (AT) software techniques. According to the Department of Defense, these methods render efforts to alter or exploit data and code \"so time consuming, expensive, and difficult that even if the attempts were to become successful, the AT protected technology will have been updated and replaced by the next generation version.\"  16  For AI verification, anti-tamper software ideally can even \"detect if [a] program has been altered,\" and potentially rendered non-compliant.  17  These methods promise to deter and detect rather than outright stop tampering. Therefore, policymakers should note that software tampering is still possible, despite the best efforts of antitamper technology. Anti-tamper methods are diverse, often application-specific, and continually evolving. As such, this report does not address many existing methods, such as watermarking, which makes code difficult to remove without damaging the system or making changes, or encryption wrappers, which encrypt system code and only decrypt it when the system is in use.  18  For the sake of illustration, I discuss two promising anti-tamper techniques-cryptographic hashing in this section and algorithm obfuscation in the next-and how they can be used for sustained AI verification. Interested policymakers and researchers are encouraged to research and consider other AT methods that could be appropriated for verification. Cryptographic hashing is a tool that could support verification by producing for inspectors a privacy-preserving \"record\" of system code. This record can act as a tamper detection device, allowing tampering to be signaled upon subsequent inspections of the same system. Cryptographic hashing uses a mathematical function, known as a \"cryptographic hash function,\" which takes in any type of data-including text, audio, and computer code-then scrambles, condenses, and transforms that data into a resulting \"hash,\" a seemingly random sequence of characters of a certain length.  19  For example, when the 39,808-word text of the Book of Genesis, King James Version, is fed through the SHA-256 hash function (a commonly used hash function), the entire text is transformed into: \"675d773189394dcd4cc92d1b489f1e04cca2b4e734dccda7d7e06d0 aed181db8.\"  20  This resulting hash is an unintelligible scramble, dramatically condensed and altered from the text. This data transformation property is only the tip of the cryptographic hashing iceberg. Importantly, for a hash function to be a cryptographic hash function and useful for verification, it must have five key characteristics: 1. Consistency: Input data will always produce the same hash when run through the same cryptographic hash function. 2. Speed: The hash function is fast, producing output in a reasonable amount of time. 3. Cannot be reverse-engineered: The resulting hash cannot be easily transformed back into the original input data that it represents. 4. Cannot be easily duplicated: Practically speaking, no two data inputs should produce the same hash. 5. The Avalanche Effect: Only a small alteration to the original input data will yield a dramatic \"avalanche\" of changes to the resulting hash. Small changes yield substantial differences. For sustained verification purposes these combined properties create a very powerful tool. Once an inspection is completed, an inspector can hash the entire code of a system, or even just the code of a verification zone, and store that hash as a \"record\" of what that code should be. To an inspector, the hash is the code. If inspectors have records of expected system code, this then allows them to identify and track system changes. If the same system's code were to be hashed again in the future, the result should match the hash the inspector has on file. If it does not, this signals alterations have been made. Crucially, this tool is privacy preserving. Because a given hash cannot be reverse-engineered to produce the code it represents, an inspector can save a record without putting state secrets at risk. It is notable that hashes do not need to be computed in person. This opens up the opportunity to use this tool not only for sustained verification, but continuous verification. To implement, a given system could be fitted with a theoretical \"hashing device\" that occasionally hashes the system's code and broadcasts that hash to an observer. This would give the observer a window into the code's fidelity over time and deter actors from making changes. Naturally, this concept would be somewhat invasive but would create a powerful verification tool. \n Challenges It is possible that, over time, certain cryptographic hashing functions may be broken, thereby compromising the integrity of this system. In the past, even hashing algorithms supported by reputable organizations such as the National Institute for Standards and Technology have contained compromising vulnerabilities.  21  If a significant vulnerability is found, code could theoretically be altered without signaling tampering. Therefore, an actor could alter their system code in such a way that it produces the same hash as the original and evade detection. This could facilitate cheating. This scenario, however, is immensely improbable. The mathematics of the best hashing algorithms are such that even if an actor were to try and fool the function, the process would take so long (perhaps decades) that the effort is not worthwhile.  22  The improbability of cracking existing top-tier hash functions is illustrated best by Bitcoin, the mechanics of which are built on cryptographic hashing. There exists an unparalleled monetary incentive to crack Bitcoin's hash function, yet the combined efforts of the world's hackers have failed thus far. Hashing is that robust. Still, if this method were implemented, policymakers should consider the probability of vulnerabilities and prepare appropriate contingencies. \n Algorithm Obfuscation The second anti-tamper technique I will discuss for the sake of illustration is algorithm obfuscation. Algorithm obfuscation is a method that \"make[s] a computer program 'unintelligible' while preserving its functionality.\"  23  Obfuscation is achieved by running code through an obfuscation algorithm, which intentionally scrambles the code to disguise its meaning. The resulting scrambled code is an unintelligible \"black box,\" which can function exactly like the original code but cannot be read by engineers and, ideally, cannot be reverse-engineered to discover the unscrambled version. The difference between obfuscation and hashing is that hashing would produce a numeric representation of the code (which cannot be computed), while obfuscation produces new, unintelligible code (which can be computed) with the exact functions of the original code. For verification, obfuscated code would act in a similar manner to cryptographically hashed code. If code is truly obfuscated, outside parties and inspectors could potentially view a system's code, and even record it, without compromising its specific implementation. Inspectors could keep a record of the entire system code, or just the code of a verification zone, and therefore a record of how that system should be configured. In subsequent inspections, inspectors could run a 1-to-1 comparison between past and present obfuscated code to quickly spot tampering. Unlike hashing, however, obfuscation allows inspectors to see code, not a hash, potentially eliminating the problem of cryptographic hash collisions. \n Challenges Code obfuscation is still a developing cybersecurity frontier, and much research is needed to produce robust obfuscation algorithms. According to the Defense Advanced Research Projects Agency (DARPA), the best current techniques only require about a day's worth of effort to crack.  24  That said, it is possible that adequate technology has been or will be developed. DARPA is currently developing what it calls \"Safeware,\" obfuscated software that cannot be reverse-engineered and is probably secure.  25  While this project is a moonshot, if it, or any other related research, is successful and made internationally available, state actors could use algorithm obfuscation to ensure these systems are secure, tamper-proof, and help build trust in regulations. Even if robust obfuscation techniques are indeed developed, it is important to note that this method does reveal code and, with it, certain elements of code structure and function. Further, because this is usable code, it could be used to determine what functions a given system has, even if it cannot be determined how the system performs those functions. Additionally, one could still find vulnerabilities in obfuscated code even if the overall purpose of a given algorithm remains unclear. All in all, in its current state obfuscated code is likely a less secure method than cryptographic hashing. However, it still represents one potential tool in the anti-tamper toolbox that can be further refined through consultation with engineers. \n Van Eck Radiation Analysis Van Eck radiation analysis is another sustained verification tool that can be used to continuously monitor system function and ensure it matches what was found during a system inspection. Van Eck radiation is the electromagnetic radiation computers and unshielded electronic devices emit when they process code. If intercepted, this radiation can be deconstructed to reveal a system's functions and even the code it processes.  26  Such potent information can be used to verify the consistency of the system's code in the long term, creating a powerful verification tool. As with most verification tools, however, Van Eck radiation is not a silver bullet. While computer code can indeed be intercepted by analyzing Van Eck radiation, this code will be highly complex and often garbled. Computers are excellent multi-taskers and often process multiple algorithms in tandem, making it difficult to piece together coherent algorithms from a tangle of intercepted instructions. This quality rules out any ability to simply identify algorithms using Van Eck radiation. For this, Van Eck radiation analysis is far too blunt an instrument. To further complicate matters, the instructions Van Eck radiation would reveal are written in machine code, the strings of 1s and 0s that form the operating instructions for computers. Machine code instructions are often processorspecific, meaning that if an analyst is unfamiliar with the processor running this machine code, they will not know how to correctly interpret the intercepted 1s and 0s. In sum, if the Van Eck radiation of a complex system such as a military drone were monitored, it is unlikely an observer could piece together what algorithms the system is running, let alone prove whether it contains AI or gives AI control over certain functions. At first glance, these two qualities may seem to rule out Van Eck radiation analysis as an effective verification tool. In reality, however, it is precisely Van Eck monitoring's weaknesses that make this highly practical for AI verification. Van Eck radiation leaks just enough information about system code, without revealing the complex, tightly guarded secrets of the system's algorithms. To use Van Eck radiation to sustain verification post-inspection, a system's owner would need to consent to the installation of a theoretical \"Van Eck radiation sensor\" onto the system in such a way that it can detect the unshielded Van Eck emissions of the device. This sensor would need to be trained to recognize the typical radiation patterns given off by this system, creating a baseline system radiation profile. This profile would consist of a \"dictionary\" of expected radiation patterns emitted when the system uses its existing code. Each time the sensor picks up a Van Eck emission, it would consult this dictionary to affirm that the emission matches the patterns expected of the system. If it detects a foreign pattern, be it from injected malware or intentional design changes by its owner, this indicates changes have been made to system code. The Van Eck radiation would then signal to inspectors that the system may no longer be compliant. To illustrate, if an inspector verifies a drone as AI-free, the inspector could then use a Van Eck monitoring device to measure the radiation patterns the drone's code emits when its algorithms are run. If the drone's radiation patterns later change, this would be detected. Inspectors could then reinspect the drone, determine the cause of the deviation, and establish whether the system remains compliant. The unique advantage of Van Eck radiation analysis over other sustained verification tools is its potential for minimally invasive monitoring. A theoretical Van Eck sensor would not need access to a system's code to verify that code has been altered. Detecting system changes needn't come at the expense of code secrecy. System monitoring can be done mostly externally (assuming the sensor can be installed without a radiation shield in its way) without reading a single line of code. This protects design secrets while guarding against illicit activity. This mechanism is scientifically grounded. Recent DARPA-funded research tested a nearly identical concept, albeit with the goal of detecting malware rather than changes to AI control of a system. The research found that illicitly injected malware could be detected in systems more than 99 percent of the time.  27  If this same process is appropriated for AI detection, systems could be continuously monitored and non-compliant activity detected with certainty beyond a reasonable doubt. \n Challenges Van Eck radiation's use in intelligence collection is noted for its difficulty. Challenges certainly remain, many of which must be tackled by engineers to determine the effectiveness of this method and the quality of information that can be collected. One such challenge is the known changes to emissions that result from integrated circuit wear and tear over time.  28  Such changes would need to be studied, quantified, and their impact accounted for in any Van Eck radiation analysis system. Another challenge is the potential impact of different environmental conditions on Van Eck emissions. These may include the impact of power lines, radar jamming equipment, and other stray electromagnetic fields. These influences would need to be accounted for, otherwise a system could signal changes every time it enters a foreign operating environment. The applicability of this measurement specifically to machine learning and deep learning models must also be studied in depth. One such challenge may be accounting for emissions changes resulting from the dynamic alterations a deep learning model may make on itself. Another could be accounting for a more diverse range of emissions created by the varied data these systems process. As mentioned, AI verification is a technical problem that needs technical solutions. Van Eck radiation analysis is a method that must be developed in consultation with engineers and specialists to ensure it provides robust results and accounts for a variety of implementation scenarios and challenges. Such challenges to verification development are not unheard of. Thankfully, in verification, half measures are acceptable. This mechanism does not need the precise accuracy to concretely determine whether a system was manipulated to be an effective tool. If it can detect potential manipulation with a certain level of uncertainty, this can still signal to observers that follow-up action may be required. \n Conclusion AI verification is no easy task. The mechanisms discussed in this report offer potential solutions to some of the many problems that face AI systems. More work is required, however. If the goal of policymakers is to regulate AI, there are actions that can be taken today. Specifically: 1. Verification Zone and Hardware Verification Architecture Research: Experts must develop and research system architecture that can be used to implement the verification zone and hardware verification concepts. Research should specifically identify what architectural options exist to cordon off and checkpoint the verification zones, the impact of these checkpoints on system performance and security, the information that can be gained or lost using these inspection mechanisms, and the effort it would take to implement these architectures in future and current state systems. 2. Van Eck Radiation Capture and Analysis Research: Further research is needed to determine the quality of information that can be captured from the Van Eck radiation given off by systems during processing and the degree of certainty of system consistency that information provides. Additional research should focus on the feasibility of a \"Van Eck radiation sensor\" to be used for continuous verification. 3. Coordination with International Partners: To build trust, researching, developing, and implementing these concepts cannot be done in a vacuum. Arms control requires international support, and trust requires transparency in the science. The United States should explore potential research partners to build support in these ideas, verify research, and foster trust in potential arms control agreements that may follow. By taking these steps and working with scientists, policymakers can move the AI arms control conversation forward. AI is at a fulcrum point, and international security depends on ensuring AI provides a net benefit. Policymakers must take it upon themselves to explore these ideas and build on this research so that robust and verifiable standards, norms, and regulations can be developed to constrain the misuse of AI systems and preserve future security. \t\t\t * See Appendix A for a discussion of the difficulties that have plagued the development of effective AI verification techniques. \n\t\t\t © 2021 by the Center for Security and Emerging Technology. This work is licensed under a Creative Commons Attribution-Non Commercial 4.0 International License. To view a copy of this license, visit https://creativecommons.org/licenses/by-nc/4.0/. CSET Product ID #: 20190020 Document Identifier: doi: 10.<PHONE>", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/AI_Verification.tei.xml", "id": "c8eff120b4ce1c07203f666237af0e2d"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Lewis's Principal Principle says that one should usually align one's credences with the known chances. In this paper I develop a version of the Principal Principle that deals well with some exceptional cases related to the distinction between metaphysical and epistemic modal ity. I explain how this principle gives a unified account of the Sleeping Beauty problem and chancebased principles of anthropic reasoning. In doing so, I defuse the Doomsday Argument that the end of the world is likely to be nigh.", "authors": ["Teruji Thomas"], "title": "Doomsday and objective chance", "text": "Introduction It's often the case that one should align one's credences with what one knows of the objective chances.  1    Lewis (1980)  calls this the Principal Principle. For example, it is often the case that if one knows that a fair coin has been tossed, then one should have credence 1/2 that heads came up. The standard caveatthe reason for the 'often'-is that one sometimes knows too much to simply defer to the chances. A trivial example: once one sees that the coin has landed tails, one should no longer have credence 1/2 in heads. In such cases, one has what Lewis calls 'inadmissible evidence'. In this paper, I develop a version of the Principal Principle that handles two subtler kinds of exceptions, both related to the distinction between epis temic and metaphysical modality. The first arises because one can know some contingent truths a priori. The second is related to the fact that even an ideal thinker may be ignorant of certain necessary truths-in particular, one may not know who one is. The second type of case is my main focus, and I will illustrate it with two wellknown examples: the Sleeping Beauty puzzle  (Elga, 2000)  and the Doomsday Argument  (Leslie, 1992) . My version of the Principal Principle, labelled simply PP, yields standard views about both these cases: it yields the thirder solution to Sleeping Beauty, and denies that Doomsday is especially close at hand. These conclusions are well represented in the literature; my contribution is to present them as an attractive package deal, following from a single principle about the conceptual role of objective chance. The Doomsday Argument, in particular, is usually analysed in quite different terms, using an thropic principles like the Strong SelfSampling Assumption (which is used in the Doomsday Argument itself ) and the SelfIndication Assumption (which is used to resist it). I will explain how PP leads to chancebased versions of these assumptions, unified in a principle I call Proportionality. I will especially urge the merits of Proportionality over the SelfIndication Assumption. In §2, I introduce the existing version of the Principal Principle that will be my starting place. In §3, I explain the problem that arises from a priori contingencies, and suggest a preliminary solution. In §4, I explain why this preliminary solution is unsatisfactory: it applies only in the very unusual cir cumstance that one has no selflocating information. I then state my preferred principle, PP, and show how it handles Sleeping Beauty and Doomsday. In §5, I state the principle of Proportionality, which follows from PP, and com pare it to the standard anthropic principles. (The proof of the main result is in the appendix.) Then, in §6, I briefly consider what my chancebased prin ciples suggest about reasoning based simply on a priori likelihood, rather than chance. Section 7 sums up and points out one remaining difficulty for my theory. Along the way, I will use the framework of epistemic twodimensionalism  (Chalmers, 2004)  to model the connection between epistemic and metaphys ical modality. I won't be defending epistemic twodimensionalism in this pa per, but it does conveniently represent the phenomena with which I am con cerned. My hope is that critics of twodimensionalism can find equivalent (or better!) things to say in their own frameworks. \n Background I will think of the Principal Principle as a constraint of rationality on an agent's ur prior. This ur prior, which I denote Cr, is a probability measure reflecting the agent's judgments of a priori probability and evidential support. Or, bet ter, not a probability measure but a Popper function, a twoplace function directly encoding conditional probabilities. 2 I'll refer to the arguments of Cr as 'hypotheses'. 'Propositions' would also do, but I use different terminology to emphasize that hypotheses are individuated hyperintensionally: the hypoth esis that water is H 2 O is distinct from the hypothesis that water is water, and someone could reasonably have different credences in them. What's the relationship between ur priors and credences? Suppose that at time t one has total evidence E and credence function Cr t . Then one should (I suggest) satisfy the norm of Ur Prior Conditionalization. Cr t (H ) = Cr E (H ) := Cr(H | E ). As is well known, Ur Prior Conditionalization entails ordinary Bayesian Con ditionalization: if one's evidence strengthens from E to E & E ′ , then one's credences change from Cr E (H ) to Cr E (H | E ′ ). However, Ur Prior Condi tionalization has the advantage that it handles situations where one's evidence changes in other ways, like cases of forgetting: whatever happened in the past, the appropriate thing now is to conditionalize one's ur prior on one's current evidence. The question of whether Ur Prior Conditionalization handles such cases correctly will be relevant later on, but for the most part I will treat this as a working hypothesis to which I do not know any comparably adequate alternative.  3  Now, as to the Principal Principle, I will start from a version developed by  Meacham (2010)  and (as he notes) in unpublished work by Arntzenius. Let 〈ch(H | E ) = p〉 stand for the hypothesis that the chance of H , given E , is p. Then Arntzenius's formulation of the principle is Cr(H | E & 〈ch(H | E ) = p〉) = p. A more general claim will also be useful. Let 〈ch = f 〉 stand for the hypothesis that the chances agree with the (perhaps only partially defined) Popper func tion f ; thus 〈ch = f 〉 is effectively a conjuction of hypotheses of the form 〈ch(H | E ) = p〉. I will write Cr f for the Popper function obtained by condi tionalizing Cr on 〈ch = f 〉: Cr f (H | E ) := Cr(H | E & 〈ch = f 〉). Then the general principle I attribute to Meacham and Arntzenius is PP1. Cr f (H | E ) = f (H | E ). More precisely, the two sides should be equal when both are defined, but from now on I'll always leave out this type of qualification.  4  I defer to Meacham for a careful explanation of the connection between PP1 and Lewis's classic version of the Principal Principle, but two points are especially relevant. First, PP1 is compatible with the existence of nontrivial chances even in worlds where the fundamental physics is deterministic. For example, if E is a suitable macroscopic specification of the initial conditions of a fair coin toss, and H is the hypothesis that the coin lands heads, we may well have ch(H | E ) = 1/2. This doesn't contradict the claim of determinism that, if E ′ is a complete microphysical specification of the initial conditions, then either ch(H | E ′ ) = 1 or ch(H | E ′ ) = 0. So I won't hesitate to treat coin tosses as genuinely chancy. The second important point is that, unlike one of Lewis's formulations, PP1 does not need an exception for inadmissible evidence. Continuing the example from the previous paragraph, suppose that the agent learns that H is true. Then, for any Popper function f , PP1 gives Cr f (H | H & E ) = f (H | H & E ) = 1. So after one learns the result of the coin toss, one is no longer bound to give credence 1/2 to heads. \n The Problem of A Priori Contingents \n The Problem The first problem for PP1 arises from the distinction between epistemic and metaphysical modality, and in particular from the phenonenon of a priori con tingents. Example 1: Topper Comes Up. Suzy is about to flip a coin, which she knows to be fair. She introduces 'Topper' to rigidly designate whichever side of the coin will come up. Because of the way she introduces the term, she can be certain that Topper comes up. However, Topper is either heads or tails. Suzy knows that, either way, there is a 1/2 chance that Topper comes up. So her credence that Topper comes up should not equal the known chance that Topper comes up.  5  If E is a suitable specification of the cointossing setup, and Top is the hypoth esis that Topper comes up, then Cr(Top | E & 〈ch(Top | E ) = 1/2〉) = 1 contradicting PP1 . This example trades on the idea that chance has to do with metaphysical or nomological modality, whereas credence is a matter of epistemic modality. It's essentially a priori for Suzy that Topper comes up, and that's why Suzy gives it credence one. But it's not necessary that Topper comes up, and so too it's not chance one. Similar problems can arise for natural kind terms. Suppose that 'water' rigidly designates what one might describe for short as the predominant wet stuff (which turns out to be H 2 O). Then we can dream up a case in which it's a priori that the predominant wet stuff is water, and yet there's a 1/2 chance that the predominant wet stuff is H 2 O. This would again enable a counterexample to PP1. Finally-and most importantly for this paper-similar cases arise for in dexicals. Example 2: The Sheds. I'm Carlos; Ramon is my twin. There are two windowless sheds. A fair coin is tossed. If heads, Ramon goes in Shed 1 and I go in Shed 2; if tails, the other way around. We sit there in the dark. Just before noon, partial amnesia is induced: although we both remember the general setup, neither of us is sure whether he is Carlos or Ramon, nor how the coin landed, nor whether he is in Shed 1 or Shed 2. If I'm Carlos and this is Shed 1, then the chance that I'm in this shed is the chance that Carlos is in Shed 1, i.e. 1/2. Similarly if I'm Ramon and this is Shed 2, and so on. In any case, there's a 1/2 chance that I'm in this shed. And yet I'm certain that I'm in this shed. \n A FirstPass Solution To avoid the problems raised by these examples, we could simply restrict the Principal Principle to cases in which the relevant hypotheses do not involve proper names, or natural kind terms, or indexicals, or anything of the sortin short, to the kind of hypotheses that  Chalmers (2011)    (H | E ) = f (H | E ). While this basic proposal will require some amendment in §4, its meaning and limitations will be clearer if we pause, first to explain how the neutrality restric tion works within the framework of epistemic twodimensionalism  (Chalmers, 2004 (Chalmers, , 2011 , arguably its natural home; and second to explain how PP2 pur ports to give a full account of Topper Comes Up and similar cases. I especially want to introduce the standard idea of a primary intension, and the less stan dard idea of a neutral paraphrase, which will play an important role in what follows. Recall that the intension of a hypothesis is a set of possible worlds-the worlds in which the hypothesis would be true. Two hypotheses are necessar ily equivalent if and only if they have they same intension. My assumption that chance is a form of metaphysical modality amounts to the claim that the chance of a hypothesis depends only on its intension. However, rational credences can distinguish between necessarily equivalent hypotheses. For ex ample, suppose that Suzy's coin in fact lands headsup. Then the hypothesis that Topper comes up is necessarily equivalent to the hypothesis that heads comes up. Yet Suzy gives them different credences. To represent the distinctions made by rational credences, twodimension alists introduce a second dimension of epistemic scenarios. These are like pos sible worlds, but individuated by epistemic criteria. For example, there are some scenarios in which Topper is heads, and others in which Topper is tails; as Suzy's uncertainty attests, these are distinct and genuine epistemic possibil ities. The primary intension of a hypothesis is the set of scenarios (rather than possible worlds) in which it is true; two hypotheses are a priori (rather than necessarily) equivalent if and only if they have the same primary intension. For my purposes, the key point is that, according to Chalmers, each sce nario picks out (i) a possible world as actual; and (ii) an intension, i.e. a set of possible worlds, for each hypothesis. For example, some scenarios pick out a world in which heads comes up. In such a scenario, Topper is heads, and the intension of Top is the set of worlds in which heads comes up. Other scenar ios pick out a world in which tails comes up. Then Topper is tails, and the intension of Top is the set of worlds in which tails comes up.  7  Crucially, a scenario s is in the primary intension of H if and only if the intension of H in s contains the world that is actual in s . Chalmers calls a hypothesis is neutral if and only if, unlike Top, it has the same intension in every scenario. We can now see why the restriction to neutral hypotheses, understood in this way, avoids the problems raised by Topper Comes Up and similar cases. Ur priors can't distinguish between a priori equivalent hypotheses, like 'Topper comes up' and 'Whichever side comes up, comes up'. PP1 is bound to fail insofar as such hypotheses are not necessarily equivalent, so that they can have different chances. This problem cannot arise for the class of neutral hypotheses, however: it is easy to see that two neutral hypotheses that are a priori equivalent are also a priori necessarily equivalent (they have the same intension in each and every scenario), and therefore a priori have the same chance.  8  Even if the restricted principle PP2 avoids these problems, one might worry that it applies too rarely to constrain credences in all the expected ways. In one respect, which I'll discuss in §4, this turns out to be a very serious worry, but I think it is worth having a firstpass explanation of how PP2 could give the desired results in a case like Topper Comes Up. Let us focus on Suzy's credence that Topper is heads. Since this hypothesis is not neutral, PP2 does not directly tell us the right credence. However, we can reason in two stages. First, the hypothesis that heads comes up is (more plausibly) neutral, and, if so, PP2 does require Suzy's credence in heads to be 1/2. Second, it's a priori for Suzy that Topper is heads if and only if heads comes up. Therefore Suzy must also have credence 1/2 that Topper is heads. Not only do we get the right conclusion, the explanation for it strikes me as perspicacious. At any rate, it illustrates that the restriction to neutral hy potheses is not debilitating insofar as there are what I'll call neutral paraphrases of more general hypotheses. Here, E • is a neutral paraphrase of E if and only if E ��� is neutral and E and E • have the same primary intension (they are a pri ori equivalent). Because of this last condition, E and E • are interchangeable when it comes to ideal ur priors. For example, 'heads comes up' is a neutral paraphrase of 'Topper is heads'. Suzy's credence in the latter is determined by her credence in the former, which is in turn determined by PP2. \n The Problem of SelfLocation \n The Problem While it is arguable that a wide range of hypotheses about the world do ad mit neutral paraphrases, it is unfortunately impossible to maintain that our evidence in ordinary circumstances-circumstances in which we expect the Principal Principle to be binding-is of that type. The reason is that neutral hypotheses exclude the use of indexicals. Because of this-as I'm about to ex plain in more detail-a neutral hypothesis, or one with a neutral paraphrase, can arguably give an adequate thirdpersonal, qualitative description of the world, but it can do nothing to identify one's own situation. As a result, PP2 only applies if one has essentially no knowledge whatsoever of what one is like or where one is in space and time. This is an unaccceptable result. To see the problem more formally, consider, for example, the hypothesis S that I am sitting-a perfectly ordinary thing for me to know. In a scenario where I am Carlos, the intension of S consists of the worlds in which Carlos is sitting; in a scenario where I am Ramon, it consists of the worlds in which Ramon is sitting. So S is not neutral. Nor, I claim, does it admit a neutral paraphrase. Remember that a neutral paraphrase would be a neutral hypothe sis with the same primary intension as S . Consider a world w in which Carlos is sitting but Ramon is not, and consider two scenarios: in the first, w is actual and I am Carlos; in the second, w is actual and I am Ramon. The primary in tension of S contains the first of these scenarios, but not the second. A neutral hypothesis, in contrast, must have the same intension in both scenarios, and this intension either contains w or it doesn't. Correspondingly, the primary in tension of a neutral hypothesis either contains both scenarios or neither. Since the primary intension of S contains only one of the scenarios, it doesn't have a neutral paraphrase. (Note, for example, that someone is sitting, though neutral, is not a neutral paraphrase of S : its primary intension contains both of the scenarios I described.) Generalizing from this example, my evidence in ordinary circumstances distinguishes between different scenarios in which one and the same world is actual: it distinguishes me from most of the other people in the actual world. Neutral hypotheses cannot do this. Therefore my evidence in ordinary cir cumstances does not admit a neutral paraphrase, and therefore PP2 does not ordinarily apply. The twodimensionalist framework represents this point in the following way. Following Chalmers again, we can identify each epistemic scenario with a centered possible world: a triple (w, x , t ) where w is a possible world, and x is an individual and t a time in w. I'll refer to (w, x , t ) as a centering of w, and (x , t ) as a center. Thus the primary intension of a hypothesis is a set of centered worlds. For example, the primary intension of the hypothesis I'm sitting consists of the centered worlds (w, x , t ) such that, a priori, if I'm x at t in w, then I'm sitting.  9  Now, let w be some possible world. For a hypothesis to be neutral, it must have the same intension with respect to every scenario, and in particular with respect to every centering of w. It follows that if the primary intension contains one centering of w, it must contain them all. This makes precise the idea that neutral hypotheses (or those with neutral paraphrases) completely fail to locate the subject in the world. \n A Solution One strategy would be to supplement PP2 by principles of a different kind that together constrain one's credences in the right way. I will consider some such principles in §5. However, this strategy seems backtofront. The Prin cipal Principle, whatever the details, is supposed to express a platitude about how knowledge of the chances ordinarily constrains our credences. It hardly matters what it says about bizarre cases of complete selflocating ignorance; it ought to apply directly in situations that (at worst) idealize what we take to be the ordinary case. I propose instead to formulate a modification of PP2 that applies directly when one does have fully selflocating evidence: that is, more carefully, when the primary intension of one's evidence contains at most one center for each possible world. Lest this appear a radical move, let me emphasize that it is a natural de velopment of what  Lewis (1980)  himself says. He states the Principal Princi ple in a setting where one's credences assign probabilities to possible worlds, and therefore do not explicitly distinguish different centers within each world. However, the point is not that his principle applies only in bizarre cases of complete selflocating ignorance! He applies it to ordinary cointossing cases, after all. Rather, uncentered possible worlds usually suffice because each such world comes with an implicit center, picked out by the agent's selflocating evidence. As Lewis says, we only need to use centers explicitly if we want to handle cases in which 'one's credence might be divided between different pos sibilities within a single world'  (Lewis, 1980, p. 268) . So Lewis's principle applies when one's credences are not so divided, i.e. when one has fully self locating evidence. Moreover, it applies no matter what the implicit centerings may be. That's exactly the picture I want to spell out. To emphasize the role of indexicals, I will often represent potentially non neutral hypotheses in the form 〈I am G 〉. Then 〈I am F G 〉 means 〈I am F 〉& 〈I am G 〉, and so on. I'll say that a hypothesis is fully selflocating if and only if its primary intension contains at most one centering of each possible world. The proposal is to restrict the Principal Principle to cases of fully selflocating evidence. However, there is a more convenient way to put this. Say that 〈I am G 〉 is merely selflocating relative to background evidence E if and only if it picks out exactly one centering of each world compatible with E . More carefully, I am talking about primary intensions, so the condition is that, if the primary intension of E contains a centering of w, then the primary intension of E & 〈I am G 〉 contains exactly one centering of w. It follows that E & 〈I am G 〉 is fully selflocating. In these terms, the main proposal of this paper is that the chances bind credences conditional on any merely selflocating hypothesis: PP. If E and H are neutral hypotheses, and 〈I am G 〉 is merely selflocating relative to E , then Cr f (H | E & 〈I am G 〉) = f (H | E ). The restriction to neutral E and H is still important here, but in §5 I will develop a less restricted principle-Proportionality-as a consequence of PP. One might worry that ordinary evidence is never fully selflocating: per haps it does not narrow things down to exactly one individual and one time in each world compatible with one's evidence. I'll consider a troubling form of this worry in §7, but for now I will just address the most mundane form: one's evidence may not often pin down a precise time. There are two basic responses. The first is that one can have fully selflocating evidence even if one does not know what time it is in the ordinary sense that one does not know what clocks are saying right now. Clockfaces are only one way of picking out an instant in each world. However, one may still worry that one's evidence is coarsegrained in a way that just can't pin the present down exactly. There may be some deep issues here about perception and even about the metaphysics of time, but the short answer is that we are allowed, as far as PP goes, to count times in a coarsegrained way. We need not take 'one time' to mean 'one instant' rather than 'one interval of unit length', where the units are adjustable and we count nonoverlapping unitlength intervals as different times. What's crucial in applying PP is that the precision with which 〈I am G 〉 locates me in the world is independent of how the world turns out, conditional on E . Now I'll illustrate PP with two important applications. \n Application: Sleeping Beauty First, consider this famous example: 10 Example 3: Sleeping Beauty. On Sunday night, Beauty knows she is in the following situation. After she goes to sleep, a fair coin will be tossed. She will be awakened on Monday. A few minutes later, she will be told it is Monday. Then she will go back to sleep. If the coin landed heads, she will sleep through Tuesday. But if it landed tails, her memories of Monday will be erased, and she will be awakened on Tuesday morning. Thus, when Beauty wakes on Monday, she does not know whether the coin landed heads or tails, nor, supposing the coin landed tails, whether it is Monday or Tuesday. What should Beauty's credence in heads be (1) on Sunday night; (2) after learning it is Monday; (3) on first waking? PP allows us to analyse the case as follows. (1) On Sunday night, Beauty's evidence, as normal, is fully selflocating. Therefore PP applies and tells us she should have credence 1/2 in heads. But let me spell this out, to make clear how the formalism works. Let E be a neutral, thirdperson specification of the setup; Beauty knows E throughout, and she also knows the chance proposition 〈ch = f 〉 to the effect that heads (H ) and tails (T ) both have chance 1/2 conditional on E . Let 〈I am G 〉 be the hypothesis that I'm Beauty and it's Sunday night. With some minor simplifications, the primary intension of E & 〈I am G 〉 effectively contains just two centered worlds: the world in which the coin lands heads, centered on Beauty on Sunday night, and the world in which the coin lands tails, centered on Beauty on Sunday night. The hypothesis 〈I am G 〉 is merely selflocating relative to E . PP then yields Cr f (H | E & 〈I am G 〉) = 1/2, as expected. (2) A very similar application of PP shows that Beauty should have cre dence 1/2 in heads after she learns it's Monday: at that point, too, she has fully selflocating evidence. (3) On first waking, however, her evidence is not fully selflocating, and PP does not apply. Nevertheless, we can argue from PP that she must have credence 1/3 in heads. Consider the three hypotheses HM (heads, which im plies it's Monday), TM (tails and it's Monday), and TT (tails and it's Tuesday). Assuming that Beauty will update by conditionalization when she learns it's Monday (i.e. HM ∨ TM), she must, when first waking, already give HM and TM equal credence. Now consider what would happen were she instead to learn HM∨TT. She would again have fully selflocating evidence, and should again have credence 1/2 in heads. So she must already give HM and TT equal credence. All together, she gives equal credence to each of the three hypothe ses HM, TM, and TT. Since these are mutually exclusive and exhaust the possibilities open to her, she must give credence 1/3 to each. This pattern of credences is called the 'thirder' position in the literature on Sleeping Beauty. I find the extant arguments for thirderism quite compelling, and I am happy to refer to them as corroboration for my view. However, the analysis I've presented is slightly different from the most common way of understanding thirderism.  Elga (2000)  appeals to a principle of indiffer ence: Beauty should, on waking, consider the hypotheses TM and TT equally likely, since her evidence is fully symmetric between them. But this suggestion invites standard worries about indifference reasoning, including the thought that Beauty might have symmetrical but only imprecise credences in these hy potheses  (Weatherson, 2005) . My argument is different, and isn't directly sus ceptible to such worries. Instead of appealing to evidential symmetry, I claim that Beauty should align her credences with the known chances, not only after she learns it's Monday, but also if she were instead to learn HM ∨ TT, on the basis that these are both merely selflocating hypotheses relative to her other evidence E . Of course, thirderism is not the only standard position when it comes to Sleeping Beauty. As Elga explains, the main rivals to thirders are halfers, who claim that Beauty should give credence 1/2 to heads when she wakes up, as well as on Sunday night. I can't do justice to the whole literature, and want to focus on my positive proposal, but it seems significantly more difficult to do for halferism what I've done for thirderism here: to embed it in a package that includes systematic norms for updating (as in Ur Prior Conditionalization) and a natural version of the Principal Principle (as in PP). For, on the one hand, Beauty's evidence after learning it's Monday is structurally very similar to her evidence on Sunday night, so it's hard to see why the Principal Principle would apply in the second case but not the first. On the other hand, if, as 'double halfers' claim, Beauty should have credence 1/2 in heads at both these times and on first waking, then she must not apply Bayesian conditionalization when she learns it's Monday. 11 \n Application: The Doomsday Argument Here is another example. It is very similar to Sleeping Beauty, but it will be useful to consider it separately, because it is commonly analysed using quite different tools, which I will contrast with PP in section 5. Example 4: Doomsday. There's a 1/2 chance that humanity goes extinct at an early stage, resulting in a total of 100 billion people who ever live (call this outcome early doom); and a 1/2 chance that humanity hangs on much longer, resulting in 100 quadrillion people who ever live (call this outcome late doom). Against this evidential background, I learn that I am the 70 bil lionth person to be born. What should my credence be in early doom? As the Doomsday Argument notes, knowing I am the 70 billionth person rules out many possibilities that are compatible with both early doom and late doom, but vastly many more that are only compatible with late doom (for example, the possibility that I am the 200 billionth person). So, for any reasonable priors, that piece of evidence should result in a dramatic shift in credence to wards early doom. Unless I was antecedently ridiculously confident in late doom-far more confident than the stated 1/2 chance-I should now be al most certain of early doom.  12  The example is practically significant because our actual evidential situ ation is stylistically similar to the one described. We have some idea about the various kinds of extinction risks we face (either as a species or as a global ecosystem), and a fairly precise idea of how far along we are since life began. The basic logic of the Doomsday Argument generalizes to more complicated cases, and seems to show that an early doom for humanity is much more likely (epistemically speaking) than the chances would on their face suggest. However, in parallel to my analysis of Sleeping Beauty, PP implies that my posterior credence in early doom should be 1/2. At least, it does so for reasonable ways of filling in the details. Most simply, assume that everyone has the same lifespan; then the hypothesis that I'm the 70 billionth person is fully selflocating by the criteria sketched at the end of §4.2. Thus it is after, not before, learning that I am the 70 billionth person that PP binds my credence to 1/2. This, along with Ur Prior Conditionalization, commits me to having been 'ridiculously' confident in late doom prior to gaining the new evidence. But, then again, prior to that evidence I was in the ridiculous epistemic state of having essentially no selflocating information. There is a general worry about Bayesianism that extremely high confidence rarely seems warranted, but, beyond that, we shouldn't be too worried about getting surprising results about such exotic epistemic positions. \n The Principal Principle and Anthropic Reasoning PP only directly constrains the credences of agents with fully selflocating evi dence. But, as already hinted in my analysis of Sleeping Beauty and Doomsday, it has broader implications. I'll now draw out some of those implications, and show how they relate to-and in some ways improve upon-the anthropic principles that are commonly used to analyse Doomsday. \n Proportionality The main technical result of this paper is that, given a sufficiently rich domain of hypotheses, PP entails a superficially stronger principle that I call Proportion ality. I'll give the argument from PP to Proportionality in the appendix, and focus on its formulation here. Proportionality is a sophisticated version of the intuitive idea that my credence that I'm F , given that I'm G , should be high to the extent that most G s are F s. In many cases, it sets that credence equal to the chanceexpected number of F andG s divided by the chanceexpected number of G s.  13  However, stating it in full generality requires a little work. So, consider a neutral hypothesis E and an arbitrary hypothesis 〈I am G 〉. Recall that the primary intension of 〈I am G 〉 contains zero or more centerings of each possible world. I define N f (G | E ) to be the expected number of such centerings, according to the probability measure f (− | E ). To spell it out: for each world w, we take the number of centerings of w in the primary intension of 〈I am G 〉, we multiply that by the probability (according to f , conditional on E ) that w is actual, and then we sum over worlds.  14  Thus N f (G | E ) = 1 if 〈I am G 〉 is merely selflocating with respect to E , and will be higher insofar as 〈I am G 〉 fails to pin down my location in th worlds where E is true. Here is the key principle: Proportionality. Suppose E is a neutral hypothesis. Then Cr f (〈I am F 〉 | 〈I am G 〉 & E ) = N f (F G | E ) N f (G | E ) . Note that (unlike in PP2) the restriction to neutral E is not onerous, since the overall evidence 〈I am G 〉 & E is effectively arbitrary. One can get a good idea of how Proportionality works by applying it to Sleeping Beauty. (And one can get a good idea of the argument for Propor tionality by pondering the argument for thirderism I gave in section 4.) As in my earlier discussion of the case, let E be a neutral specification of the setup, but now let 〈I am G 〉 be the hypothesis that I'm Beauty and I've just woken up on one of the relevant days. The primary intension of E & 〈I am G 〉 effec tively consists of three centered worlds: the heads world centered on Beauty on Monday, the tails world centered on Beauty on Monday, and the tails world centered on Beauty on Tuesday. It follows that N f (G | E ) = 3/2. On the other hand, let 〈I am F 〉 be the same as the hypothesis H that the coin landed heads. Then the primary intension of 〈I am F G 〉 consists of a single centered world: the heads world, centered on Beauty on Monday. It follows that N f (F G | E ) = 1/2. Therefore, according to Proportionality, Cr f (〈I am F 〉 | 〈I am G 〉 & E ) = 1/3, as the thirder claims. \n The SelfSampling Assumption Discussions of the Doomsday Argument usually turn on two 'anthropic' prin ciples: the SelfSampling Assumption and the SelfIndication Assumption. I'll now explain how these principles are related to Proportionality. One goal is to shed more light on Proportionality, but I'll also argue that-at least in some ways-Proportionality provides a more satisfactory picture of the Doomsday Argument. The first of the two main anthropic principles is, in Bostrom's influential formulation, The Strong SelfSampling Assumption (SSSA). One should rea son as if one's present observermoment were a random sample from the set of all observermoments in its reference class.  15  Here an 'observermoment' is what I have been calling a centered possible world: one's present observer moment is the actual world centered on oneself and the present time. Although SSSA is not precisely stated, the basic idea is that one should consider different merely selflocating hypotheses to be equally likely. So, for example, Beauty should be indifferent between Monday and Tuesday, conditional on tails. And, in Doomsday, I should initially give equal credence to different hypotheses about my birthrank in each world separately. Because of this, my initial credence that I'm the 70 billionth person is a million times higher conditional on early doom than on late doom. This determines how strongly I should update in favour of early doom upon learning my birth rank: the subjective odds of early doom increase by a factor of one million. My own analysis of Doomsday used PP to determine my posterior cre dence in early doom directly. It is unnecessary to adduce SSSA as a separate principle, since the following version of it is a simple application of Propor tionality: Uniformity. Suppose that 〈I am G 〉 and 〈I am G ′ 〉 are merely selflocating relative to a neutral hypothesis E . Then Cr f (〈I am G 〉 | E & 〈I am G or G ′ 〉) = Cr f (〈I am G ′ 〉 | E & 〈I am G or G ′ 〉). In short, the merely selflocating hypotheses 〈I am G 〉 and 〈I am G ′ 〉 get equal credence. Besides being much more precise, Uniformity differs from SSSA in several important respects. First, Uniformity only applies conditional on an appropriate chance hy pothesis 〈ch = f 〉. I'll say more about this limitation in §6. It is a limitation, but it also points to a key strength: Uniformity emerges from a unifying story about objective chance. Second, Uniformity makes sense even when some worlds compatible with E include infinitely many observermoments, whereas there is no entirely rea sonable way to randomly sample from an infinite set.  16  This is the effect of conditionalizing on 〈I am G or G ′ 〉: it narrows consideration to at most two centers in each world. Third, as usually conceived, SSSA is a principle of indifference between different merely selflocating hypotheses, similar to the indifference principle Elga used to analyse Sleeping Beauty. In contrast, Uniformity is based on a claim about the applicability of the chancecredence link. Of course, PP does include a kind of indifference claim, to the effect that all merely selflocating hypotheses are equally good from the point of view of the Principal Principle. A fourth, closely related difference is that SSSA appeals to the the idea of a 'reference class' of observermoments. Uniformity treats all merely self locating hypotheses as equally good, without limitation to a narrower reference class (but with the understanding that centered worlds include only genuine a priori possibilities-see fn. 9). Bostrom uses flexibility in the choice of ref erence class to resolve various problems that arise from his theory, including the Doomsday Argument. This flexibility seems unnecessary when it comes to Uniformity: PP defuses the Doomsday Argument without further recourse to reference classes. \n The SelfIndication Assumption The second, more controversial anthropic principle is The SelfIndication Assumption (SIA). Given the fact that you exist, you should (other things equal) favor hypotheses accord ing to which many observers exist over hypotheses on which few observers exist.  (Bostrom, 2002, p. 66)  This is again rather imprecise, but SIA is commonly understood as a claim about the evidential import of the fact that one exists: conditionalising one's ur prior on that evidence increases the relative likelihood of worlds with large pop ulations. This idea is especially clearly stated by  Bartha & Hitchcock (1999a) , but goes back to  Dieks (1992) . One post hoc motivation for SIA is that it provides a way of blocking the Doomsday Argument. Suppose that we think the chancecredence link is properly given by PP2. Then, knowing only the chance hypothesis stated in Doomsday, I should have a 1/2 credence in early doom and 1/2 in late doom. Next, I conditionalize on two pieces of evidence: (E 1 ) that I exist and (E 2 ) that I am, specifically, the 70 billionth person. The Doomsday Argument really shows us that given E 1 , E 2 shifts my credences dramatically towards early doom. But SIA suggests that conditionalizing on E 1 itself shifts my credences towards late doom-towards worlds with lots of people. (For simplicity, I as sume that 'person' and 'observer' mean the same thing.) So if we interpret SIA in exactly the right way, these two shifts will cancel out, and the net effect of learning E 1 and E 2 is to leave my credence in early doom at the original 1/2. Is there any independent reason to think that E 1 has exactly the eviden tial significance required?  Bartha & Hitchcock (1999a, p. 349)  provide what they call a 'justso story': if the 100 billion people in the early doom world and the 100 quadrillion people in late doom world were chosen separately and uniformly at random from a stock of possible people, then any one of those possible people would have a greater chance (and greater to just the right de gree!) of being selected into the late doom world. But even if we managed to take this justso story seriously as a piece of cosmology, the upshot would be unclear. How does it help with cases of selflocation within a life, as in Sleeping Beauty? And notice that the metaphysical claim that the population is chosen at random is compatible with the not unreasonable epistemic claim that it is a priori, for me, that I exist. But if it is a priori, then it has no eviden tial weight for me at all, contrary to SIA. As this indicates, the justso story equivocates between metaphysical and epistemic modality in the way I have been trying to avoid. Motivationally, then, the SelfIndication Assumption is on shaky ground. Nevertheless, there is a precise sense in which Proportionality, somewhat like SIA, requires one to give higher credence to largepopulation hypotheses than the chances naively suggest. It entails: Weighting. If E and H are neutral hypotheses, then Cr f (H | E & 〈I am G 〉) = N f (G | H & E ) N f (G | E ) f (H | E ). According to Weighting, my credence in H conditional on E & 〈I am G 〉 should equal the chance of H conditional on E , but weighted by a factor that reflects how H is correlated with the number of G s. The factor is large insofar as the worlds in which H is true tend to have many centerings in the primary intension of 〈I am G 〉. Weighting is a precise generalization of the claims that, before learning it's Monday, Beauty should be quite confident in tails, and that, before learning I'm the 70 billionth person, I should be extremely confident in late doom-in both cases, more confident than the 1/2 chance. Some readers might like to see in more detail how Weighting applies in Sleeping Beauty. As in §5.1, let E &〈I am G 〉 be Beauty's nonchance evidence on first waking; as before, its primary intension contains three centered worlds, and N f (G |E ) = 3/2. Let T (rather than H ) be the hypothesis that tails comes up. The primary intension of T & E & 〈I am G 〉 consists of two centerings of the tails world, and it follows that N f (G | T & E ) = 2. Thus, Weighting says, Beauty's credence in tails should be 2/(3/2) = 4/3 times the 1/2 chance, i.e. it should be 2/3, just as the thirder claims. \n Beyond Chance This paper has been about objective chance, and the anthropic principles de veloped in §5 are formulated in terms of a chance hypothesis 〈ch = f 〉. As I mentioned in §2, my understanding of chancetalk is pretty broad: it's not just limited to indeterministic interpretations of quantum mechanics, or anything like that. Still, I agree that there are situations where talk of chances would seem misplaced, including cases in which we are considering the relative plau sibility of different scientific theories. So I don't claim to have recovered the full scope of the anthropic principles that have been proposed in the literature. But I have shown that one can get pretty far with chances, and the results are suggestive of a more general analysis.  (H | E & 〈ch = f 〉) = f (H | E ). And the argument for Proportionality given in the appendix supports the more general claim Cr(〈I am F 〉 | 〈I am G 〉 & E ) = N Cr 0 (F & G | E ) N Cr 0 (G | E ) This generalization of Proportionality does not involve any chance hypothe sis; it instead involves the judgments of evidential support represented by the Popper function Cr 0 . The point of this innovation is that sometimes our judgments of a priori evidential support plausibly relate to Cr 0 rather than to Cr. We just don't usu ally consider the case of complete selflocating ignorance; we take selflocation for granted, as does most of the literature in epistemology that is not specif ically concerned with Sleeping Beauty or Doomsdaylike cases. So the loose thought that H and ¬H are equally likely conditional on E may well suggest that Cr 0 (H |E ) = 1/2 rather than Cr(H |E ) = 1/2. Note that Cr 0 (H |E ) = 1/2 is what we'd expect from PP if one knew a priori that ch(H | E ) = 1/2. In that sense, the judgements of evidential support represented by Cr 0 are calibrated to the chances. Some other ways of measuring a priori likelihood are at least compatible with chancecalibration. For example, one might attempt to gauge the relative likelihood of H and ¬H by imagining what an angel in heaven would find plausible without having looked out to see how the universe is going. 17 But of course the angel knows perfectly well where he is, so judgments arrived at in this way must already take selflocating evidence into account. For illustration, consider a version of Doomsday in which early doom and late doom are supposed to be 'equally likely a priori', but this isn't cashed out in terms of chances. If 'equally likely' is understood in terms of Cr, then (set ting aside SIA and other shenanigans) the Doomsday Argument does seem to show that someone with fully selflocating evidence will be dramatically more confident in early doom than in late. But this point is not very inter esting unless we have a decent grip on what is epistemically likely given the exotic evidential background of complete selflocating ignorance. In contrast, if 'equally likely' is understood in a chancecalibrated sense, or just against an implicit evidential background that already includes selflocating information, then the Doomsday Argument does not go through. \n A Final Problem I've shown how to formulate a version of the Principal Principle that is bet ter insulated against the problem of a priori contingencies and which works even in the context of selflocating ignorance. The main ideas are that one should to stick to neutral hypotheses, and that chances bind credences relative to fully selflocating evidence. The resulting picture, including Ur Prior Con ditionalization, fits cleanly with the thirder view of Sleeping Beauty. It also yields chancebased versions of some wellknown anthropic principles (Uni formity, Weighting, and most fundamentally Proportionality) while blocking the chancebased Doomsday Argument. Finally, one can generalize these prin ciples beyond chances to chancecalibrated judgments of a priori likelihood. The aspect of this picture that I ultimately find least satisfying is that, when it comes down to it, our ordinary evidence may not be fully selflocating. Given the immense size of the universe, we should take seriously the possibil ity that there are qualitative duplicates, or near enough, of ourselves and our surroundings somewhere else. (More carefully, the issue is that my total evi dence includes in its primary intension some epistemic scenarios centered on sufficiently close duplicates of myself.) As a stylized case, consider a version of Doomsday in which the 100 quadrillion people in the late doom world consist of a million distantly separated groups of duplicates of the 100 billion people who would exist given early doom. Against that background, it would be hard for me to get fully selflocating evidence; reasonable evidence could at best narrow down one's identity to a million qualitatively identical people, condi tional on late doom. By Weighting, I should then be extremely confident in late doom. And, to emphasise, I need not be unusually uninformed: I could be well acquainted with my environment as far as telescopes can see. I think I have to bite the bullet here: compared to the chances, my cre dences should favour worlds that contain many clones of myself and my en vironment.  18  The consolation is that this won't interfere with ordinary appli cations of the Principal Principle. For example, when it comes to a fair coin toss, one should still give heads credence 1/2, so long as the expected num ber of one's clones doesn't depend on the toss. It is true that Proportionality, rather than PP, is more directly applicable. So once we take into account the possibility of clones, Proportionality may be the best way to think about the chancecredence link. \n Appendix: Derivation of Proportionality The argument will assume that there is a sufficiently rich space of hypotheses. Instead of formulating general conditions, I'll just state exactly what I'll use in terms of the hypothesis E and the predicates F and G . (a) E is nonatomic: there is a neutral hypothesis A such that f (A|E ) ̸ = 0, 1. (It will be convenient to write A ′ for ¬A.) This first condition is a substantive restriction on the class of cases in which Proportionality follows formally from PP. Informally, though, it would be odd if Proportionality held for nonatomic E but not for E ; I also note that nonatomicity assumptions are common in axiomatic decision theory. Any way, in contrast, the following two conditions, while messier to formulate, are more vacuous: they say that there are enough neutral hypotheses and fully self locating hypotheses to carve up modal space in the ways one would expect. For example-recalling that each hypothesis has an intension in each scenariothe next two conditions hold if there is a hypothesis for every function from scenarios to intensions. (b) For all integers k ≥ j ≥ 0, there is a neutral hypothesis E j k whose intension contains a world w iff the primary intension of 〈I am G 〉& E calls neutral: 6 PP2. If E and H are neutral hypotheses, then Cr f \n How so? Starting from an ur prior Cr, we can (partially) define a Popper function Cr 0 that encodes judgements of evidential support given an arbitrary background of merely selflocating evidence. Restricting ourselves to neutral hypotheses H and E , the idea is that Cr 0 (H | E ) = p holds if and only if Cr(H | E & 〈I am G 〉) = p whenever 〈I am G 〉 is merely selflocating relative to E . With this definition, we can reformulate PP more simply as the claim that Cr 0 \n\t\t\t This paper is mainly a project in Bayesian epistemology, and I'll speak throughout about what one 'knows' as a shorthand for what evidence one has in the sense relevant to Bayesian conditionalization. This is a natural way of speaking, but nothing turns on the identification of evidence with knowledge. \n\t\t\t  See Hájek (2003)  for reasons one might take conditional probabilities as primitive. Un conditional probabilities can be recovered as probabilities conditional on a tautology.3 See e.g.Moss (2015, pp. 174-176)  for discussion of Ur Prior Conditionalization, and Titelbaum (2016) for some relevant alternatives. \n\t\t\t To clarify the connection to Meacham's work: the hypothesis 〈ch = f 〉 takes the place of what he calls a 'chancegrounding' proposition. \n\t\t\t This example is inspired by a similar one inHawthorne & LasonenAarnio (2009, pp. 95- 96). \n\t\t\t Because of the conditionalization, it really suffices that E and H & E are neutral. And, while I won't focus on this issue here,Lewis (1980, pp. 268-9)  essentially points out that the Popper function f must also be given in a suitably neutral form. If I know that the chance of heads is x , and, unknown to me, x equals 1/4, then I'm under no compulsion to set my credence in heads to 1/4. \n\t\t\t If heads actually comes up, there is no possible world in which Topper is tails. However, there are possible worlds in which tails comes up, and the thought is that, in any scenario that picks out such a world as actual, Topper is tails. \n\t\t\t This depends on a seemingly harmless assumption, adopted by Chalmers, that every pos sible world is actual in some scenario. \n\t\t\t The identification of scenarios with centered worlds, and the question of whether this is fully appropriate, are somewhat delicate; I defer to Chalmers (2011)  for discussion. The use of centred possible worlds to model selflocating ignorance is standard since at least Lewis (1979) , and most of the rest of this paper could be written in a Lewisian framework. Note though that Lewis claims the objects of belief are properties, whose intensions are sets of centred worlds. In contrast, for twodimensionalists, the (ordinary, not primary) intension of a hypothesis is still a set of possible worlds. See Magidor (2015)  for critique especially of the Lewisian tradition. By the way, it may be that some formally possible centered worlds do not represent genuine epistemic possibilities, even a priori. Perhaps it is a priori for me that I am not a rock; then (w, x , t ) does not correspond to an epistemic scenario, if x is a rock at time t in w. When I talk about centered worlds, I only mean those that represent genuine epistemic possibilities. \n\t\t\t The example was made popular by Elga (2000) ; see his first footnote for its history. \n\t\t\t On the first point, Lewis (2001)  claims that Beauty has inadmissible evidence once she learns it's Monday, but it seems hard to independently justify this claim, or to fit it into a systematic account of admissibility. On the second, see Titelbaum (2016)  for a survey of alternative updating methods and their problems. \n\t\t\t This is a simple version of the Doomsday Argument treated explicitly by Leslie (1992)  and attributed to Brandon Carter. See Bostrom (2002)  for a discussion of its history. Note that your current evidence may well be fully selflocating even if you have little idea of your birthrank (cf. my discussion of knowing the time in §4.2). So this Doomsday Argument says nothing about what should happen if you were to learn your birthrank in real life. In fact, if you know that the difference between early doom and late doom is some future extinction event, then you've already ruled out having a birth rank incompatible with early doom, and learning your birth rank shouldn't have much (or any) interesting effect. \n\t\t\t Proportionality is closely related to what Manley (2014) calls 'Typicality', but importantly different from what Arntzenius & Dorr (2017) call 'Proportion': roughly, the latter requires the stated credence to equal the expected proportion of G s that are F s. \n\t\t\t This recipe is a little rough for the usual reason that there may be uncountably many relevant worlds, and we can't just sum over them; I'll give a more formal definition in the appendix. \n\t\t\t  Bostrom (2002, p. 162). The (not 'Strong') SelfSampling Assumption applies to ob servers, rather than observer moments, but that won't help with Sleeping Beauty cases, and is actually incompatible with SSSA. \n\t\t\t I don't claim to solve all the related problems that arise from infinite worlds, for discussion of which see Bartha & Hitchcock (1999b) , Weatherson (2005) , and especially Arntzenius & Dorr (2017) . It's worth mentioning in this context that Popper functions need not be countably additive. \n\t\t\t  See Bostrom (2002, pp. 32ff) for a similar heuristic. \n\t\t\t  See Elga (2004); Weatherson (2005)  for a discussion of related problems. \n\t\t\t A partition A 1 , . . . , A n of a hypothesis A is a collection of hypotheses such that, a priori, the A i are mutually exclusive and their disjunction is equivalent to A.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Thomas-Doomsday-and-Objective-Chance-Version-2.tei.xml", "id": "dee514e3a89597527dc380eb4941e511"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "This paper is the first installment in a series on \"AI safety,\" an area of machine learning research that aims to identify causes of unintended behavior in machine learning systems and develop tools to ensure these systems work safely and reliably. Below, we introduce three categories of AI safety issues: problems of robustness, assurance, and specification. Other papers in this series elaborate on these and further key concepts.", "authors": ["Tim G J Rudner", "Helen Toner"], "title": "Key Concepts in AI Safety: An Overview", "text": "Introduction The past decade has seen the emergence of modern artificial intelligence and a variety of AI-powered technological innovations. This rapid transformation has predominantly been driven by machine learning, a subfield of AI in which computers learn patterns and form associations based on data. Machine learning has achieved success in application areas including image classification and generation, speech and text generation, and decision making in complex environments such as autonomous driving, video games, and strategy board games. However, unlike the mathematical and computational tools commonly used in engineering, modern machine learning methods do not come with safety guarantees. While advances in fields such as control theory have made it possible to build complex physical systems, like those found in various types of aircraft and automobiles, that are validated and guaranteed to have an extremely low chance of failure, we do not yet have ways to produce similar guarantees for modern machine learning systems. As a result, many machine learning systems cannot be deployed without risking the system encountering a previously unknown scenario that causes it to fail. The risk of system failures causing significant harm increases as machine learning becomes more widely used, especially in areas where safety and security are critical. To mitigate this risk, research into \"safe\" machine learning seeks to identify potential causes of unintended behavior in machine learning systems and develop tools to reduce the likelihood of such behavior occurring. This area of research is referred to as \"AI safety\" 1 and focuses on technical solutions to ensure that AI systems operate safely and reliably. Many other challenges related to the safe deployment of AI systems-such as how to integrate them into existing networks, how to train operators to work effectively with them, and so onare worthy of substantial attention, but are not covered here. Problems in AI safety can be grouped into three categories: robustness, assurance, and specification. Robustness guarantees that a system continues to operate within safe limits even in unfamiliar settings; assurance seeks to establish that it can be analyzed and understood easily by human operators; and specification is concerned with ensuring that its behavior aligns with the system designer's intentions. 2 \n Modern Machine Learning Machine learning methods are designed to learn patterns and associations from data. 3 Typically, a machine learning method consists of a statistical model of the relationship between inputs and outputs (for example, the relationship between an audio recording and a text transcription of it) and a learning algorithm specifying how the model should change as it receives more information about this input-output relationship. The process of updating the model as more data is made available is called \"training,\" and recent advances in fundamental research and engineering have enabled efficient training of highly complex models from large amounts of data. Once trained successfully, a machine learning system can be used to make predictions (such as whether or not an image depicts an object or a human), to perform actions (such as autonomous navigation), or to generate synthetic data (such as images, videos, speech, and text). Many modern machine learning systems use deep neural networks-statistical models that can represent a wide range of complex associations and patterns and that work particularly well with large amounts of data. Examples of useful application areas for deep neural networks include image classification and sequential decision-making in autonomous systems, as well as text, speech, and image generation. Machine learning systems derive associations and patterns from data rather than from a prespecified set of rules. As a result, these systems are only as good as the data they were trained on. While modern machine learning systems usually work remarkably well in settings similar to those encountered during training, they often fail in settings that are meaningfully different. For example, a deep neural network trained to classify images of cats and dogs in black and white is likely to succeed at classifying similar images of cats and dogs in color. However, it will not be able to correctly classify a fish if it has never encountered an image of one during training. While machine learning systems do not use explicit rules to represent associations and patterns, they do use rules to update the model during training. These rules, also called \"learning algorithms,\" encode how the human designer of a machine learning system wants it to \"learn\" from data. For example, if the goal is to correctly classify images of cats and dogs, the learning algorithm should include a set of steps that update the model to gradually become better at classifying cats and dogs. This goal can be encoded in a learning algorithm in many ways, and it is the task of the human designer of such a system to do so. \n Robustness In order to be reliable, a machine learning system must operate safely under a wide range of conditions. Building into the system the ability to quantify whether or not it is confident about a prediction may reduce the chance of failure in situations it is not well-prepared to handle. The system, upon recognizing it is in a setting it was not trained for, could then revert to a safe fallback option or alert a human operator. Challenging inputs for machine learning systems can come in many shapes and guises, including situations a system may never have encountered before (as in the fish classification example above). Operating safely in such scenarios means that a system must, first, recognize that it has not been trained for such a situation and, second, have a way to act safely-for example, by notifying a human operator to intervene. An active area of research around this problem seeks to train machine learning models to estimate confidence levels in their predictions. These estimates, called predictive uncertainty estimates, would allow the system to alert a human operator if it encounters inputs meaningfully different from those it was trained on. Consider, for example, a machine learning system tasked to identify buildings in satellite imagery. If trained on satellite imagery of a certain region, the system learns to identify buildings that look similar to those in the training data. If, when deployed, it encounters an image of a building that looks meaningfully unlike anything it has seen during training, a robust system may or may not classify the image as showing a building, but would invariably alert a human operator about its uncertainty, prompting manual human review. \n Assurance To ensure the safety of a machine learning system, human operators must understand why the system behaves the way it does, and whether its behavior will adhere to the system designer's expectations. A robust set of assurance techniques already exist for previous generations of computer systems. However, they are poorly suited to modern machine learning systems such as deep neural networks. Interpretability (also sometimes called explainability) in AI refers to the study of how to understand the decisions of machine learning systems, and how to design systems whose decisions are easily understood, or interpretable. This way, human operators can ensure a system works as intended and, in the case of unexpected behavior, receive an explanation for said behavior. It is worth noting that researchers and engineers working with and developing modern machine learning systems do understand the underlying mathematical operations inside so-called \"black-box\" models and how they lead from inputs to outputs. But this type of understanding is difficult to convert into typical human explanations for decisions or predictions-say, \"I liked the house because of its large kitchen,\" or \"I knew that dog was a Dalmatian because it had spots.\" Interpretability, then, seeks to understand how trained machine learning systems \"reason\"-that is, how certain types of inputs or input characteristics inform a trained system's predictions. Some of the best tools we have for this so far include generating visualizations of the mathematical operations inside a machine learning system or indicating which input characteristics are most responsible for a model's outputs. In high-stakes settings where humans interact with machine learning systems in real time, interpretability will be crucial in giving human operators the confidence to act on predictions obtained from such systems. \n Specification \"Specification\" of machine learning systems refers to defining a system's goal in a way that ensures its behavior aligns with the human operator's intentions. Machine learning systems follow a pre-specified algorithm to learn from data, enabling them to achieve a specific goal. Both the learning algorithm and the goal are usually provided by a human system designer. Examples of possible goals include minimizing a prediction error or maximizing a reward. During training, a machine learning system will try to reach the given goal, regardless of how well it reflects the designer's intent. Therefore, designers must take special care to specify an objective that will lead to the desired behavior. If the goal set by the system designer is a poor proxy for the intended behavior, the system will learn the wrong behavior and be considered \"misspecified.\" This outcome is likely in settings where the specified goal cannot fully capture the complexities of the desired behavior. Poor specification of a machine learning system's goal can lead to safety hazards if a misspecified system is deployed in a high-stakes environment and does not operate as intended. Misspecification has already arisen as a problem in YouTube's video recommendation algorithms. This algorithm was designed to optimize for engagement-the length of time a user spends watching videos-to maximize ad revenue. However, an unintended side effect manifested: To maximize viewing time, in some cases, the recommendation algorithm gradually steered users toward extremist content-including videos from white supremacist and other political and religious extremist channelsbecause it predicted these recommendations would cause the user to stay engaged longer. The extent of this phenomenon remains disputed, and YouTube has changed its algorithms since this issue first gained considerable attention. Yet the underlying idea-that optimizing for engagement could have unintended effectsdemonstrates the hazards of goal misspecification. 4 \n Conclusion Safety considerations must precede the deployment of modern machine learning systems in high-stakes settings. Robustness, assurance, and specification are key areas of AI safety that can guide the development of reliably safe machine learning systems. While all three are the subjects of active and ongoing research, it remains uncertain when we will be able to consider machine learning systems reliably safe.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Key-Concepts-in-AI-Safety-An-Overview.tei.xml", "id": "f4631760ad998f8441500573ae46fe6f"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "This working paper is a preliminary analysis of the legal rules, norms, and strategies governing artificial intelligence (AI)-related intellectual property (IP). We analyze the existing AI-related IP practices of select companies and governments, and provide some tentative predictions for how these strategies and dynamics may continue to evolve in the future. In summary: • AI developers use a mix of patents, trade secrets, and open-source licensing agreements to protect their AI-related IP. • Many AI companies are pursuing what may seem like a counterintuitive IP strategy: aggressively patenting AI technologies while sharing them freely. They experience competitive pressure to patent in order to present the threat of a countersuit if another company sues them for IP infringement. However, they also experience pressure to open-source their work in order to attract top talent and entice consumers to use their platforms. • Governments broadly have two goals related to IP policy for AI that are at times in conflict with the goals of researchers and/or companies: to ensure that AI-related inventions can be patented, and to ensure that national-security-relevant AI inventions are restricted for government use and/or kept secret. • Significant uncertainty exists regarding how AI patentability, open-sourcing, and infringement litigation will evolve in the future. • There is an opportunity for patent pools to be used to facilitate pro-social behavior and ethical norms among AI developers. Existing patent pools and practices by international standards organisations represent possible models to replicate. \n Introduction 2.", "authors": ["Nathan Calvin", "Jade Leung"], "title": "Who owns artificial intelligence? A preliminary analysis of corporate intellectual property strategies and why they matter", "text": "Table of Contents \n Introduction Artificial intelligence (AI) is increasingly a focal point of competition between leading firms, and between states. This paper focuses on a key, often under-examined component of the competitive strategies being employed by both corporate AI developers and national governments: the protection of their intellectual property. Intellectual property (IP) is a broad and flexible concept, referring to creations of the mind that are eligible for protection through law.  2  Today, companies are using a mix of patents, trade secrets, and open-source licensing agreements to protect their AI-related IP. Simultaneously, government patent offices, judiciaries, and national security apparatuses are deciding which aspects of AI should be patentable, and whether certain inventions should be restricted for military purposes. The IP policy choices that governments and corporations make can have profound implications for the development trajectory of a technology. For example, in the 1990s, the biotechnology industry was transformed after court decisions in the US enabled a broader range of biological compounds and processes to be protected by patent law. This development spurred additional private investment, but critically also allowed companies to claim ownership over what previously would have been considered basic academic research. This, in turn, encouraged higher levels of secrecy to protect valuable intellectual property.  3  More recently, one of the most prominent advances in biotechnology, the CRISPR gene editing mechanism (originally derived from a naturally occurring process in bacteria), has been subject to a protracted legal battle over overlapping patent claims in the US.  4  Changes in IP law and strategy may have a similarly large impact on the trajectory of AI development. What these impacts could be, however, have received little study. This paper aims to provide a preliminary analysis of the goals and strategies of corporations and governments focused on AI development, and what the implications of these strategies may be. First, we explain how corporate AI developers currently protect their AI-related intellectual property and why they choose the methods that they utilize. Second, we describe how governments use intellectual property law to pursue national goals related to AI. Finally, we describe three plausible scenarios for how IP strategies in AI may evolve. 2 World International Property Organization 3 World International Property Organization  4  Cohen, 2019 \n Understanding Corporate AI Developer IP Strategies Corporate AI developers face two key decisions around how to protect their AI-related intellectual property: whether or not to patent AI techniques and systems, and whether to open-source models or keep them private as trade secrets. A prevalent strategy among top AI developers today involves accumulating patents while simultaneously sharing research with the open-source community.  5  For example, Microsoft holds the most number of machine learning patents in the US (see Figure  1 ), 6 but is also an active participant in the open-source community, 7 sharing source code for machine learning methods and under certain circumstances providing free licenses for their patents. Microsoft's strategy is not an anomaly. Amazon, Google, IBM, Facebook, Baidu, Tencent, and several other companies are prolific patent holders in AI (see Figures  1 and 2 ) while also open-sourcing substantial portions of their systems and sharing their work at academic conferences such as ICML and NeurIPS. 8  Notably, and perhaps unintuitively, some of the largest software patent holders in the world (Google, Amazon, and Facebook) signed an amicus brief to the Supreme Court in 2014 advocating that the court make it more challenging to patent abstract software patents, which includes a considerable percentage of more theoretical AI and ML patents: \"Abstract software patents have become a plague on computer-related industries. The software industry developed and flourished without abstract patents before changes in the Federal Circuit's jurisprudence led to a flood of them. Far from promoting innovation, abstract software patents have impaired it by granting exclusive rights over high level ideas and thereby blocking others from undertaking the truly innovative task of developing specific applications.\" 9 These observations beg the question: why do so many of the top AI developers grow AI-related patent portfolios while simultaneously sharing their research at academic conferences, open-sourcing machine learning models, and advocating for the legal dissolution of many AI-related software patents? In this section we argue that these corporate IP strategies help to manage a variety of objectives that companies wish to pursue. AI developers apply for patents because of competitive pressures to do so. These same developers also often open-source AI models in order to build their reputation, attract talent, and incentivize customers to use paid products. AI developers can also use selective open-source licensing agreements as a hybrid strategy, enabling companies to participate in the open-source community while maintaining the legal threat of their patents. We discuss these incentives for both patenting and opensourcing in turn, along with limitations and drawbacks of each approach. \n Pressure to Patent Patents have several uses beyond simply enabling the patent holder to sue for patent infringement. In the case of big technology firms, there is a strong incentive to engage in \"defensive patenting;\" that is, patenting without the intention to offensively litigate for infringement, but rather to present a credible threat of counter-lawsuit to another company. As the \"mutually assured destruction\" analogy makes clear, patent litigation is extremely costly for all involved due to substantial legal fees and the stigma for investors of working with a company whose products are in legal purgatory. At the height of smartphone-related litigation in 2011, Apple and Google each spent more money on patent litigation (primarily in suits and countersuits against one another) than they did on research and development, a sum in the billions of dollars. 11 In that same year, Google spent $12 billion acquiring Motorola, which market analysts evaluated as being primarily for Motorola's substantial smartphone patent portfolio. 12 Perhaps if Google had acquired Motorola's patents before litigation began with Apple, the threat of a more substantial retaliation could have prevented the lawsuits. This defensive rationale is also the stated reason for Google's new AI patent filings in machine learning and neural networks. When asked about Google's new filings, spokespeople for Google and DeepMind stated that they \"hold patents defensively, not with the intent to start fights with others.\" 13 This dynamic can also help explain why Google has advocated for narrowing the influence of software patents while simultaneously growing its own patent portfolio; while Google may prefer a world without expensive software patent litigation, its defensive patenting strategy is shaped by threats in the existing patent litigation regime. Beyond defensive patenting, corporate AI developers could also be incentivized to hold patents in order to gain leverage in other settings. For example, Google's patent sharing arrangement with the Chinese tech giant Tencent is paving the way for Google's entry into the Chinese market.  14  Google also allows other companies to enter into its \"Open Patent Pledge\" and make use of patents in a pool as long as they commit not to engage in patent litigation against  Google. 15  However, building and maintaining a large patent portfolio has its drawbacks. First, in the US, patent holders must pay substantial upkeep fees to keep their patents active; up to thousands of dollars a year depending on the size of the patent holding entity and age of the patent (see Table  1 ).  16  For large companies like Googlewhich has more than 50,000 active patents-these costs can be in the tens of millions of dollars. 17 Second, for companies that place a high premium on secrecy, the disclosure requirements and public nature of patent filings may be onerous. Finally, some in the AI research community are philosophically opposed to the idea of patenting AI concepts and techniques, particularly broad theoretical methods that are seen as mathematical truths rather than human inventions.  18  Companies that do choose to patent regardless may face pushback from those opposed, which may have flow-on effects on their ability to, for example, attract and retain research talent. \n Patent Fee Schedule (per patent) 19 Large Entity Fee Small Entity Fee Micro Entity Fee Patent Maintenance fee due 3.5 years after patent is issued $1,600 $800 $400 Patent Maintenance fee due at 7.5 years $3,600 $1,800 $900 Patent Maintenance fee due at 11.5 years $7,400 $3,700 $1,800  \n Pressure to Open-Source Incentives for corporate open-sourcing are also complex, typically extending beyond an altruistic or philosophical belief in open science. For example, open-sourcing can be used to build a firm's reputation, generate goodwill among the research community, and encourage customers to use paid products. Apple's recent trend towards open-source and sharing more AI research demonstrates some of these incentives at work. Apple has a notorious culture of secrecy, with numerous internal mechanisms in place to prevent leaks and sequester information.  20  Apple has benefited from this culture of secrecy in consumer hardware design, a world where preventing leaks and copycat designs are critically important. However, this culture was also a barrier for Apple to recruit top ML researchers, who typically strongly value being able to publish and share their work at conferences. Notably, several of Apple's rival firms enabled researchers to do so.  21  In 2017, Apple changed its approach, launching a machine learning journal and enabling its researchers to publicly share findings at top ML conferences, including NeurIPS and ICML. Open-sourcing can also be used as a tool to generate more paying customers. For example, companies with substantial cloud computing businesses often offer free machine learning tools to encourage customers to design an application using the open-source tool. Then, these customers go on to pay for these computeintensive machine learning processes to be implemented on that same firm's cloud service. In some cases, this occurs through explicit lock-in. For example, Amazon's image recognition software \"Rekognition\" is only available on Amazon Web Services.  23  In other cases, companies aim to retain customers through brand loyalty; Google hopes that customers will use its ML open-source platform Governments around the world are grappling with how to best take advantage of the recent wave of advances in AI, with several nations releasing national plans on how their country intends to incentivize and capitalize on AI innovation.  30  These plans include ensuring that effective IP law regimes exist for AI and ML. This tends to break down into two objectives: making AI patentable, and regulating access to national security related IP. (For additional context on how national patent systems interact at the international level, see Box 1.) \n Box 1: The International Patent System and the World Intellectual Property Organization Patent systems are primarily domestic in nature rather than international. Each country has its own patent office and companies interested in seeking a patent for an invention must apply separately in all jurisdictions in which they wish to be awarded a patent.  31  A patent awarded in one country cannot be used to litigate infringement in another, though that patent does count as a form of \"prior art\" which can be used to prevent the award of a patent for that invention in another country. The World Intellectual Property Organization helps harmonize this process by assisting inventors to file their inventions in several jurisdictions at once. However, the decision to award a patent will ultimately fall to individual countries. While patent treaties such as \"The Agreement on Trade-Related Aspects of Intellectual Property Rights\" (TRIPS 32 ) have taken steps to harmonize intellectual property law across member nations, differences persist at every level of the process: what inventions are patentable, the level of scrutiny applied before a patent is granted, how patents are enforced and reviewed, the length for which patents are valid, and upkeep costs required to maintain the patent. \n Making AI Patentable A prominent element of several AI national plans is to ensure that AI-related inventions can receive patents in a timely fashion. The goal of this policy is to encourage research and development (R&D) investment in AI by rewarding that investment with a potentially lucrative patent. This strategy functions to both encourage domestic companies to invest in AI-related research, and entice corporations choosing between different IP systems to set up shop in their country rather than elsewhere. For example, US Patent and Trademark Office Chief Andrei Iancu recently expressed in a Senate hearing that the US needs to make sure that its IP rules adequately protect and incentivize innovation in AI.  33   state council plan, which declared the nation's intention to be the world leader in AI by 2030, one section advised that policy makers in China must \"[s]trengthen the protection of intellectual property in the field of AI.\"  34  The European Patent Office also recently released specific guidance on how to successfully patent inventions in AI and machine learning,  35  and Singapore is allowing AI patents to be \"fast tracked\" for review through its patent system. 36 Proposals for increasing Britain's competitiveness in AI have also highlighted its patent system's challenges in protecting AI-related inventions as a liability.  37  In some ways, the question of how to create patent protections for AI is not a new one. AI patents mostly fall into the existing category of software patents, and countries have struggled for years to find regulatory structures that incentivize innovation without allowing individual companies to control overly broad, abstract, or obvious ideas.  38  In fact, despite a push to allow more patenting and offer more stringent protections for inventions, two recent major changes in IP law within the United States-the 2011 America Invents Act 39 and the 2014 Supreme Court decision Alice Corp. v. CLS Bank International 40 -made it more difficult to claim and enforce broad software patents. It will be difficult to change patenting rules in AI without also implicating these existing decisions on software patents. Furthermore, it is unclear whether expanding the range of patentable AI-relevant inventions would effectively incentivize innovation. For one, AI and ML commercial activity has experienced massive growth and international investment even while the patentability of innovations remains uncertain, suggesting that the ability to patent AI is not necessary for innovation. In addition, more patents increases the likelihood of litigation, which could act to disincentivize innovation and slow down industry growth. This is particularly a concern for some software patents due to their broad and abstract nature. China has outpaced the US in new patent applications related to the machine learning subfield of deep learning.  42  While some observers have interpreted this information as evidence of China's fast approaching superiority over the US in new AI innovation, three key pieces of information about China and the US's patent systems should make us view these statistics in a different light. First, patents in the US and China have very different standards, requirements and protections. The majority of technology patents in China are filed as \"utility model\" patents, a category of patent in China not extant in the US.  43  Utility patents require a smaller inventive step, are subject to less rigorous examination upon filing, and last only 10 years (in comparison to 21 years for American patents). Consequently, this has led to filers taking advantage of lax inspections and review. In a 2018 report, the Chinese state-owned Xinhua News Agency accused China's IP system of being characterized by \"weak IP, fake demands and some companies fervent on phony innovation,\" according to a translation by Bloomberg News.  44  Chinese \"invention patents\" have requirements more similar to those in the US and last for 20 years rather than 10. However, they only comprise 23% of patent holdings in China. Second, in China, there are strong government financial incentives for researchers to file patents, regardless of the underlying patents' merit. 45 This is particularly true in AI, where the Chinese central government has committed substantial public funds to encourage inventions. Finally, it is important to look at the \"discard rate\" for Chinese patents-i.e. how quickly a patent is discarded by its holder-and how they compare to patents in the United States. A substantial percentage of Chinese patent holders allow their patents to expire before the patent's lifespan is complete-61% for utility patents and 37% for invention patents-in comparison to a discard rate of 15% in the US.  46  When viewed in context with the previous points, it becomes clear that there are many Chinese researchers who file for AI patents in order to claim government incentives without genuinely believing their invention is notable.  47  Regulating National-Security-Related IP Access Countries' patent regimes for AI are not only shaped by economic motivations, but also by national security interests. National governments typically pursue two primary interests on this front: to ensure that their own national security apparatuses have access to state-of-the-art technology, and to withhold that access from perceived rivals. On the goal of ensuring access, a US court decision in 2015 held that the federal government can use patented inventions without the permission of the patent holder and cannot be forced to cease usage of a patented invention; the only remedy is for the patent holder to request damages assessed at market rate (which amounts to compulsory licensing).  48  This means if a patent holder does not wish for their patent to be used by the government (e.g. a patent that has potential surveillance applications) their only recourse after suing for infringement is to force the government to pay for a reasonably costed license. This process is quite distinct from what happens when a private entity infringes on a patent, where the private entity can be enjoined to cease usage of the patent or be assigned additional damages. Additionally, over the last few years the Chinese government has passed broad laws on national security and cybersecurity with implications for access to intellectual property.  49  One of these laws mandates that network operators (broadly defined) provide \"technical support and assistance\" to national security relevant government offices.  50  The exact legal authority of the Chinese government to force cooperation with Chinese companies is difficult to know. The Center for a New American Security's Ashley Feng reports that \"U.S. government officials, including at the FBI, interpreted this vague language to mean that all Chinese companies, including Huawei, are subject to the direct orders of the Chinese government.\" 51 However, The Diplomat's Jack Wagner reports that the main purpose of the law is to mandate additional data localization and storage on Chinese mainland servers and to set standards around cybersecurity.  52  Additional concerns around Chinese corporations acting as extensions of the state are conceivable, but more speculative in nature. On the goal of withholding access from perceived rivals, the US and the UK have long had government statutes that empower their patent offices to prevent public disclosure and bar the award of patents that have national security implications, regardless of their other merits.  53  In the US in 2018, 5,792 patent applications were covered within these so-called \"secrecy orders,\" higher than at any point since the Cold War.  \n Three Scenarios for the Future of AI Intellectual Property Strategies Given the observed corporate AI developer IP strategies and growing government interest in IP law as a lever for influencing AI development, how could the dynamics of AI intellectual property protection evolve? What would these dynamics then mean for the future of the AI industry, and in particular, on the competitive strategies employed by firms and states? Here we present three plausible scenarios which focus on how the IP strategies of corporate AI developers could evolve in the near future: (1) Open Research Continues: The status quo persists: open research alongside defensive patents remains the norm within the ML industry. (2) Patent Lawsuits and Secrecy: Offensive patent litigation breaks out within the ML community and prompts additional secrecy among developers. (3) Expansion of Patent Pools: In response to the threat of litigation, AI developers enter into additional patent sharing agreements. In the following section, we describe each scenario and present evidence to support its plausibleness. These scenarios are intended to be illustrative rather than predictive; indeed, there are several alternative and hybrid scenarios that could warrant further investigation as well. \n Path #1: Open Research Continues \n Scenario: Each of the major AI developers weighs the costs of engaging in patent litigation against their competitors, and decides that the threat of a countersuit and the symmetrical legal costs make litigation a poor choice. Each developer continues to file for patents in order to maintain a credible response, but analogous to a nuclear standoff, this \"mutually assured destruction\" framework holds. This equilibrium is bolstered by competition for researchers who want to work at companies that prioritize openness and cooperation with their peers. Some patent trolls-firms that profit from licensing and litigating on patents without producing any products of their own-may gain control of patents and engage in litigation without fear of reprisal, but these lawsuits remain relatively insignificant. \n Evidence in Favor: • Despite the recent flurry of activity on the subject, the machine learning community currently shows little sign of changing its open and non-litigious culture. Academics and individuals around the world can use cutting edge machine learning techniques from open-source platforms free of charge. There has been some litigation over trade secret theft in autonomous vehicles (Waymo vs. Uber,  55  Baidu vs. JingChi 56 ), but no large-scale patent litigation over broad concepts in machine learning.  55  Korosec, 2018  56  Borak, 2017 • Current patent rules in the US make AI-related software patents less threatening for the purposes of litigation than they were before the Supreme Court's 2014 decision in Alice Corp. v. CLS Bank International, which made software patents more likely to be classified as \"abstract ideas\" and thus unpatentable. Given this decision, it is likely that many existing AI patents, particularly those covering broad mathematical concepts, will be rejected during litigation or at the Patent Trial and Appeal Board. • While some AI-related patents in the US are being granted, the majority are not. Data from patent filings shows that in recent years, over 90% of AI-related patent applications in the US were initially rejected, many for being merely \"abstract ideas\" that are not eligible for patentability. 57 By comparison, the overall rejection rate for patents in the US is 48%.  58  Fewer AI-related patents mean fewer opportunities for companies to litigate over infringement, thus bolstering the incentives for open research. • If large tech companies choose to litigate with one another over AI-related IP, they do not just have to contend with a defendant's AI-related patents, but also with all of the other patents that would likely be used in a countersuit.  59  Google, Microsoft, Amazon, and IBM have expansive business operations across several verticals; this makes patent aggression with other large companies less attractive. Path #2: Patent Lawsuits and Secrecy \n Scenario: Major patent litigations break out between AI developers. While the previous open equilibrium may be preferable for the collective interest of major private AI developers, it may only take one large company deciding it is in its interest to pursue active litigation for this state of affairs to deteriorate. For instance, if IBM, with its trove of AI-related patents and its struggling core enterprise business, 60 chose to litigate against its rivals, it could provoke additional litigation. So-called \"patent trolls\" (entities that accumulate patents while not producing products of their own) could also threaten to disrupt the mutually assured destruction equilibrium and engage in lawsuits without fear of reprisal. This litigation could bleed over and affect academic research. While the EU has a research exemption that protects academic use from being deemed infringement, the US has no such exemption, and university researchers could thus potentially find themselves on the wrong end of litigation. 61 As litigation escalates, there is a substantial incentive for companies to ensure that potentially patentable inventions are kept secret from rivals. In a 2017 paper, Nick Bostrom discusses how the pursuit of patents in AI could cause companies to share research less often in order to prevent other entities from using  57  Decker, 2019 AI developers agree to cross-license their patents to one another in a \"pool\" to reduce the risk of litigation. We previously discussed patent sharing agreements in the context of companies like Google and Tencent using these deals to gain market entry into new countries. However, patent sharing agreements need not only be between two actors. There are several historical examples of large technology companies pooling their patents to protect against litigation and create advantageous licensing dynamics. The DVD6C Licensing Group was comprised of eight of the most high profile patent holders in DVD technology (including Samsung and Toshiba, among others). 66 Third-party manufacturers interested in using their technology could approach the group as a one-stop clearinghouse in order to obtain licenses instead of approaching each member individually. Similar arrangements would enable companies within the pool to share trade secrets and research and development, though too much coordination would come under the ire of antitrust enforcement. In this scenario, patent pools could be used not only to curtail litigation risks, but also to limit or promote certain applications of AI. As previously mentioned, companies are currently able to individually place stipulations on patent licensing. If a company decided that it wanted to refuse to license its patents to manufacturers of, for example, autonomous weapons, on moral grounds, it could certainly do so. Extending this to a patent pool, corporate AI developers could group together to share intellectual property and establish shared standards for how they wish to have their intellectual property used, perhaps based on certain ethical principles. These standards could then be implemented via, for example, selective licensing  62  Bostrom, 2017  63  Simonite, 2018  64  Ibid. 65 Ibid. \n DVS6C Licensing Group agreements which restrict use of the pool's IP to actors who commit to abiding by those standards. Membership of the patent pool could also be made conditional on abiding by the pool's standards. Indeed, patent pools used for licensing agreements by standard setting bodies such as the International Standards Organization are precedent for similar structures' success. It is worth noting, however, that such multilateral \"refusals to deal\" would need to be implemented with caution and appropriate due diligence in order to avoid potential infringements of antitrust law. 67 Evidence in Favor: • Existing software patent pools show the demand for and utility of this type of coordination. Google and several other tech companies participate in the Open Invention Network and Android Networked Cross-License Agreement in order to protect Linux and Android developers from infringement litigation.  68  As an alternative, the MPEG LA group is an example of a software patent pool that operates as a profitable licensing association for its members.  69  Facebook and Google's aforementioned existing patent non-aggression agreements could also be a path towards greater cooperation. • Increasing returns to scale in AI (meaning that more data improves AI systems and platforms, which attracts more users, which in turn generates more data) could increase the odds of industry centralization. A smaller number of relevant actors improves the feasibility of this kind of coordination. 67 Department of Justice 68 Open Invention Network, 2017 69 MPEG LA \n Conclusion AI is poised to be a critically impactful technology, and its development will be deeply affected by existing social and legal institutions. This paper has preliminarily explored an under-examined aspect of this infrastructure: the legal rules, norms, and strategies governing AI-related intellectual property. Leading corporate AI developers today have employed a dynamic and at times unintuitive IP strategy that allows them to respond to the shifting competitive landscape surrounding them. Governments are also seeking to shape IP systems that incentivize innovation around AI while also protecting their national security interests. How each actor balances its varying objectives in relation to AI and how they choose to wield IP strategies to achieve these objectives remains to be seen. This preliminary analysis scratches the surface of what may be an important element of the strategic landscape shaping competition and cooperation among AI firms and prominent national governments. Further investigation in this direction could be fruitful for better understanding the goals and strategies of actors seeking to protect AI-related intellectual property, and how these strategies have flow-on implications on the competitive dynamics that arise between AI developers. This, in turn, could shed light on questions related to the prospects for cooperation between these actors to achieve prosocial outcomes with respect to AI ethics and safety, and more broadly, the appropriate governance of AI going forward. Figure 1 : 1 Figure 1: Top Machine Learning Patent Holders, 2000-2015 \n Figure 2 : 2 Figure 2: Top Neural Network Patent Holders, 2000-2015 \n Table 1 : 1 USPTO patent upkeep costs, per patent \n 2216 US Patent and Trademark Office 17 Regalado, 201318  For an example of this view, see Mark Riedl's quote in Simonite, 2018   19 US Patent and Trademark Office 20 Stone and Vance, 2009 21 Clark, 2015 22 Lewsing, 2017 \n Tensorflow on Google Cloud, though they could also use it on Amazon Web Services or Microsoft's Azure. This proves to be a strong incentive for open-sourcing given that cloud services appear to be incredibly lucrative. In 2018, Amazon, Microsoft, and Google earned $25 billion, $23 billion, and $4 billion in revenue from their cloud businesses, respectively. 24 IBM's $34 billion acquisition of open-source cloud computing provider RedHat (which reportedly was also in acquisition talks with Google, Microsoft, and Amazon before selling to IBM 25 ) and Microsoft's $11 billion JEDI cloud computing contract with the Pentagon 26 further show how cloud business is a priority for large AI developers. Microsoft offered Azure cloud users access to a substantial portion of their patents as an incentive for users to join the platform.28  Facebook attempted to add a licensing stipulation to its popular open-source platform React, which would have caused users to retroactively lose their licenses if they ever engaged in patent litigation against Facebook.29     23 Amazon Web Services 24 Griswold, 2019; Microsoft, 2018; Novet, 2018 25 Peterson, 2018 The major downside of open-sourcing is the opportunity cost: open-sourcing in its purest form means forgoing licensing fees from users. It also means sharing what would otherwise be competitive secrets with rival companies. In the next section we discuss some ways that companies manage to avoid these drawbacks.A Hybrid Strategy: The Best of Both Worlds?Despite the apparent conflict between building a large patent portfolio and participating in the open-source community, selective licensing rights can enable a hybrid strategy where companies participate in the opensource community while maintaining their patents. Companies can and do create selective licensing terms for the use of their patents in open-source projects. These agreements can include defining permitted usage in a way that allows some users to utilise the code while disincentivizing competitors to include that code in a product.27    AI developers have also used selective open-source licensing agreements to achieve other ends. New York Times, 2019 27 E.g. The GNU Operating System GPL3 Open-Source license disincentivizes commercial use.28  Microsoft29  The plan was abandoned after developer backlash. See Wolff, 2017  3. Governments' Pursuit of AI Innovation and NationalSecurity Through IP Law \n Challenges Comparing Patent Filings: The US and China 41 34 Wang, 2018 35 European Patent Office, 2019 36 Spruson and Ferguson, 2019 37 Clark et al. 2019 38 Lee, 2013 39  Leahy-Smith America Invents Act, 2011 40 Alice Corp. v. CLS Bank International, 201441 Bessen, 2013   Box 2: \n 58  Carley et al. 201559  For example, if Google were to engage in litigation against another company over machine learning patents, they would have to contend with its opposition's patents in E-commerce, search, drones, self-driving cars, smartphone hardware and software and telecommunications, because each of these are areas in which Google operates in and is thus capable of infringement.60  Imbert, 201861 Miller, 2002   intermediary research to obtain a patent first.62  Public research could also be used as evidence in litigation to prove infringement (e.g. releasing a model using a patented method), further dissuading companies in a litigious environment from engaging in open-source communities. As discussed previously, governments in the US, China, EU, and elsewhere are pushing to broaden the scope of patentable material in AI and encourage filings. If these efforts translate into more ostensibly defensible AI patents being filed successfully, this could increase incentives for litigation between patent-holders. • While Google reported that it is holding new AI patents on a defensive basis, other developers have been non-committal about their future intention to litigate with their patents.63  When asked about the issue, a Facebook spokesperson said that \"its filings shouldn't be read to indicate current or future plans.' 64 IBM's patent counsel released a statement that said its large AI patent portfolio \"reflects its commitment to fundamental research.\" 65 Evidence in Favor: • Path #3: Expansion of Patent Pools Scenario: \n\t\t\t  American Bar Association, 2014    \n\t\t\t Quinn, 2013 11 Duhigg and Lohr, 2012 12 Hardy, 2011 13 Simonite, 2018 14 Cadell, 2018 15 Google", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/GovAI-working-paper-Who-owns-AI-Apr2020.tei.xml", "id": "166280bf5104fe0242044d639590b85f"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Let's say with Nick Bostrom that an 'existential risk' (or 'x-risk') is a risk that 'threatens the premature extinction of Earth-originating intelligent life or the permanent and drastic destruction of its potential for desirable future development ' (2013, 15). There are a number of such risks: nuclear wars, developments in biotechnology or artificial intelligence, climate change, pandemics, supervolcanos, asteroids, and so on (see e.g.  Bostrom and Ćirković 2008) . So the future might bring Extinction: We die out this century. In fact, Extinction may be more likely than most of us think. In an informal poll, risk experts reckoned that we'll die out this century with a 19% probability  (Sandberg and Bostrom 2008) . The Stern Review on the Economics of Climate Change, commissioned by the UK government, assumed a 9.5% likelihood of our dying out in the next 100 years  (UK Treasury 2006) . And a recent report by the Global Challenges Foundation suggests that climate change, nuclear war and artificial intelligence alone might ultimately result in extinction with a probability of between 5% and 10% (Pamlin and Armstrong 2015). 1 But the future needn't be so grim. It may also bring \n Survival: We survive for another billion years, and on average there are always 10 billion people, who live very good 100-year-long lives. So there are 100 million billion future people with very good lives. This may sound optimistic. But it's also possible. The earth seems to remain sustainably inhabitable by at least 10 billion people (United Nations 2001, 30), and for another 1.75 billion years  (Rushby et al. 2013 ). The quality of our lives seems to have increased continuously (see e.g. the data collected in Pinker 2018), and it seems possible for this trend to continue. Notably, it partly depends on us 1 More precisely, the report distinguishes between 'infinite impact', 'where civilisation collapses to a state of great suffering and does not recover, or a situation where all human life ends', and an 'infinite impact threshold', 'an impact that can trigger a chain of events that could result first in a civilisation collapse, and then later result in an infinite impact ' (2015,  40). The above numbers are their estimates for infinite impact thresholds.", "authors": ["Stefan Riedener"], "title": "Existential risks from a Thomist Christian perspective", "text": "whether Extinction or Survival will materialise. In fact it may depend on what we do today. We could now promote academic research on x-risks  (Bostrom and Ćirković 2008) , global political measures for peace, sustainability or AI safety  (Cave and ÓhÉigeartaigh 2019) , the development of asteroid defence systems  (Bucknam and Gold 2008) , shelters  (Hanson 2008) , risk-proof food technologies  (Denkenberger and Pearce 2015) , and so on. And while none of these measures will bring x-risks down to zero, they'll arguably at least reduce them. So all of this raises a very real practical question. How important is it, morally speaking, that we now take measures to make Survival more likely? The answer depends on the correct moral theory. It's most straightforward on standard utilitarianism. Suppose we increase the probability of Survival over Extinction by just 1 millionth of a percentage point. In terms of overall expected welfare (setting nonhuman sentience aside), this is equivalent to saving about 1 billion lives, with certainty. So according to utilitarianism, even such tiny increases in probability are astronomically important. Indeed, Nick Bostrom suggested that x-risk reduction 'has such high utility that standard utilitarians ought to focus all their efforts on it  ' (2003, 308ff. ; see also  Parfit 1984 , 452f., Beckstead 2013 . But this implication isn't restricted to utilitarianism. Something very similar will emerge on any view that assigns weight to expected impartial welfare increases. Consider Effective Altruism (or 'EA'). Effective Altruism is the project of using evidence and reasoning to determine how we can do the most good, and taking action on this basis (see  MacAskill 2015) . This doesn't presuppose any specific moral theory about what the 'good' is, or about what other reasons we have beyond doing the most good. But Effective Altruists typically give considerable weight to impartial expected welfare considerations. And as long as we do, the utilitarian rationale will loom large. Thus according to a 2018 survey, EA-leaders consider measures targeted at the far future (e.g. x-risk reduction) 33 times more effective than measures targeted at poverty reduction  (Wiblin and Lempel 2018) . The EA-organisation 80'000 hours suggests that 'if you want to help people in general, your key concern shouldn't be to help the present generation, but to ensure that the future goes well in the long-term'  (Todd 2017) . And many Effective Altruists already donate their money towards x-risk reduction rather than, say, short term animal welfare improvements. In this paper, I'll ask how the importance of x-risk reduction should be assessed on a Christian moral theory. My main claim will be that Christian morality largely agrees with EA that x-risk reduction is extremely important-albeit for different reasons. So Christians should emphatically support the abovementioned measures to increase the probability of Survival. Let me clarify. First, there's no such thing as the Christian moral doctrine. One of the philosophically most elaborate and historically most influential developments of Christian thought is the work of Aquinas. So I'll take this as my starting point, and argue first and foremost that core Thomist assumptions support x-risk reduction. Thomas's specific interpretation of these assumptions are often unappealing today. But I'll also claim that they can be interpreted more plausibly, that their core idea is still authoritative for many Christians, and that on any plausible interpretation they ground an argument for x-risk reduction. So while I'll start with Thomas, my conclusions should appeal to quite many contemporary Christians. Indeed, I'll ultimately indicate that these assumptions needn't even be interpreted in a specifically Christian manner, but emerge in some form or other from a number of worldviews (cf. section 4). Second, there are different x-risk scenarios, and they raise different moral issues. I think the case is clearest for ways in which humanity might literally go extinct, before being superseded by non-human intelligence, and as a direct consequence of our own actions. I'll refer to these cases as as 'non-transitionary anthropogenic extinction', and it's on these cases that I'll focus. It would be interesting to explore other scenarios: cases in which we're replaced by another form of intelligence, non-extinction 'x-risks' (which Bostrom's definition includes) like an extended stage of suffering, or scenarios of natural extinction like volcano eruptions. My arguments will have upshots for such cases too. But I won't explore them here. Third, there are different ways in which, or different agents for whom, x-risk reduction might be 'important'. In what follows, I'll mostly be considering a collective perspective. That is, I'll assume that we as humanity collectively do certain things. And I'll focus on whether we ought to do anything to mitigate x-risks-rather than on whether you individually ought to. The existence of this collective perspective might be controversial. But I think it's plausible (see e.g.  Dietz 2019 ). Christian moral theory, or at least Thomas, also seems to assume it (cf. section 2.1). And many important issues emerge only or more clearly from it. So I'll assume it in what follows. In short, my question is about how important it is, on roughly Thomist premises, for us to reduce risks of non-transitionary anthropogenic extinction. I'll first present three considerations to the effect that, if we did drive ourselves extinct, this would be morally very problematic-a hubristic failure in perfection with cosmologically bad effects (section 2). I'll then discuss some countervailing considerations, suggesting that even though such extinction would be bad, we needn't take measures against it-because God won't let it happen, or because we wouldn't intend it, or because at any rate it isn't imminent (section 3). I'll argue that none of these latter considerations cut any ice. So I'll conclude that on roughly Thomist premises it's extremely important for us to reduce x-risks. \n Three Thomist Considerations There are many Thomist considerations that would bear on x-risks. For instance, in driving ourselves extinct, we'd presumably kill the last representatives of humanity. On a Thomist perspective, those killings will be morally problematic simply as killings (see  ST, q64) . Yet this has nothing to do with the fact that those killings lead to extinction. So let's see whether there would be anything distinctly problematic about extinction, if we brought it about. There is. \n The natural law A first consideration follows from Thomas's teleology, or from the Aristotelian strand in his thinking. Recall that for Thomas the order of the cosmos is teleological. This teleology is grounded in the fact that the cosmos is subject to God. And it takes the form of a law for us: 'the universe is governed by Divine Reason. Wherefore the very Idea of the government of things in God [...] has the nature of a law.' (ST, i-II, q91 a1, co) For Thomas, following this 'eternal law'-or Divine 'will' (ST, i-II, q93 a4, ad1) or 'plan' (ST, i-II, q93 a3, co) for all things-is the ethical imperative for us. So what does it command? We can't know God's intent 'in itself' (ST, i-II, q93 a2, co), at least not in this earthly life (see ST, i-II, q93 a2, co; ST, I, q12 a11). But we can know it 'in its effect', through its manifestations in creation, or through the 'natural law' that we recognise as structuring the physical universe. In particular, we can detect God's plan for us through the natural inclinations He's implanted in us: 'all things partake somewhat of the eternal law[:] [...] from its being imprinted on them, they derive their respective inclinations to their proper acts and ends' (ST, i-II, q91 a2, co). In other words, our inclinations allow an indirect cognition of the essence of God's will, quite like sunrays allow an indirect cognition of the substance of the sun  (ST, q93 a2, co) . So what are our natural inclinations? Thomas speaks of three kinds: 'in man there is first of all an inclination [...] which he has in common with all substances: [...] the preservation of its own being  [...] . Secondly, there is in man an inclination [...] which he has in common with other animals: [...] sexual intercourse, education of offspring and so forth. Thirdly, there is in man an inclination to good, according to the nature of his reason [...]: [...] to know the truth about God' (ST, i-II, q94 a2, co). What's crucial for present purposes is the second type of inclination. Procreation or the 'preservation of the species' is firmly part of our 'natural good'  (SCG, III, , or of what the natural law commands us to do. Now Thomas doesn't believe that everyone must follow all of these inclinations, or that everyone must have offspring. It's permissible for some to remain celibate (see e.g. ST ii-II, q152). But we as a collective have a duty-a 'duty  [...]  to be fulfilled by the multitude' (ST ii-II, q152 a2, ad1)-to procreate. So this grounds a straightforward consideration for x-risk reduction. By going extinct, we'd fail to attain our end, or to accord with the natural law. Indeed, our failure would be profound. It wouldn't just be some of us flouting this law-the bad apples in an overall virtuous whole. We'd fail collectively, as an entire species, to attain our end. And (at least as far as this life's concerned) we wouldn't just fail in one aspect of perfection, while still able to excel in others. Our survival is a precondition for any aspect of our flourishing. So our extinction would mean we fail comprehensively, in all respects of our end. And of course we wouldn't just temporarily fall short of our calling. Once we've gone extinct, there's no second bite at the apple. It would mean we've foundered irreversibly. So the moral failure in anthropogenic extinction would seem complete. In short: Natural law: We ought to attain our natural end. Our extinction would prevent us from doing so-collectively, comprehensively, and irreversibly. Thus non-transitionary anthropogenic extinction would amount to a total moral failure of us as a species. That's a first consideration for why our extinction would be problematic. Now I suppose that this kind of rationale isn't parochially Thomist, but should appeal to Christians quite broadly. Thomas himself interprets the natural law pertaining to procreation very radically-e.g. as permitting a nonprocreative life only for the sake of the 'contemplation of truth' (ST ii-II, q152 a2, co; also ST ii-II, q153 a2, co), 2 and as prohibiting any intercourse known to be non-reproductive  (SCG, III, 122; q153 a2, co) . This would mean that our procreation-related obligations go much beyond preventing extinction. And it would mean that contemporary liberal moral thought is radically wrong about the good human life, and about intercourse among people of the same sex, or people who for whatever reasons cannot or don't want to reproduce. Many contemporary Christians will want to resist these implications. But we needn't understand the general idea of a 'natural law' in this manner. Plausibly, other aspects of the human end (athletic, social, emotional) are just as integral as 'the contemplation of truth'. So there are good lives without children beyond those of philosophers and priests (see e.g.  Nussbaum 1987 Nussbaum or 2011 . Also, perhaps there are other functions of human sexuality (e.g. bonding), such that non-reproductive intercourse isn't a misuse of sexual organs. After all, there's plenty of non-reproductive sexuality among non-human animals (see e.g.  Bagemihl 1999 ). The details of the natural law are a matter of large debate. 3 But I presume that the general idea of a 'natural law' is still quite prominent for Christians today. Indeed, assuming the universe manifests God's intentions, it's a very natural assumption. And I suppose that on any plausible interpretation, it will ground a consideration against extinction: whatever it implies about childfree individuals, contraception or homosexuality, it seems hard to square the natural law with an heirless self-eradication of our species. If the law commands anything, it seems, it commands us to 'be fruitful and multiply' (Gen 1:28). So a consideration along these lines should be authoritative to Christians quite broadly. Self-inflicted extinction would constitute a total failure in our executing our designated role. 4 \n Humility However, there's more to anthropogenic extinction than a failure to reproduce. Let's look at the precise way in which, through anthropogenic extinction, we'd fail to attain our end. According to Thomas, there are different ways to fall short of perfection. One is through utter passivity, or sloth, or a 'sorrow' that 'so oppresses man as to draw him away entirely from good deeds' (ST, ii-II, q35 a2, co). So we might just not bother to do anything much at all, and therefore fail in perfection. Another way of failing lies in falling too low, or being overcome with 'the lower appetite, namely the concupiscible' (ST, ii-II, q153 a5, co). So we might behave like lower animals, and fail to live up to our standards. But drawing on Augustine, Thomas says that 'the most grievous of sins' (ST, ii-II, q162 a6, co) consists in aiming too high-in failing to respect our limitations, or acting as if we were God. To do so is to commit the sin of pride, hubris, or superbia. Following the church father (De Civ. Dei XIV, 13), Thomas characterises a prideful man as someone who 'aims higher than he is', or does not 'tend to that which is proportionate to him' (ST, ii-II, q162 a1, co). And what's proportionate to us is of course not so by coincidence, but due to Divine assignment. So pride is opposed specifically to humility: 'humility properly regards the subjection of man to God [...]. Hence pride properly regards lack of this subjection, in so far as a man raises himself above that which is appointed to him according to the Divine rule' (ST, ii-II, q162 a5, co). And in this sense, as far as the 'aversion from the immutable good' (ST, ii-II, q162 a6, co) is concerned, pride is the most grievous sin: it's not just a failure through ignorance or weakness or an innocent desire for another good, but an active 'withstanding' or 'resisting' or manifesting 'contempt of' God (ST, ii-II, q162 a6, co). What does this imply in practice? Thomas specifies what he means. Pride isn't any old undue desire. It's, specifically, an 'appetite for excellence in excess of right reason'-or an inordinate imitation of the powers or competences of God (ST, ii-II, q162 a1, ad2; emphasis added). This may take various forms. We may be pridefully curious about things we can't know, such as facts about good and evil (ST, ii-II, q163 a1, ad3). Or we may unduly discard our need for Divine grace, deeming ourselves capable of happiness on our own (ST, ii-II, q163 a2, co). But a more specific power that isn't appointed to us is decisions about life and death: 'it belongs to God alone to pronounce sentence of death and life' (ST, ii-II, q64 a5, co). Thus to kill someone, or (I take it) actively prevent them from coming into existence, is generally to show an appetite for a power that doesn't pertain to us. And if all of this is true, then non-transitionary anthropogenic extinction in particular would manifest an enormous form of pride. Note that on most such scenarios, we'd fall prey to technologies we were unable to control-nuclear weapons, artificial intelligence, biotechnology, or whatever. So our extinction would mean that we'd overestimated our mastery over these fabrications, and the invulnerability we could leverage in the face of them. It would mean we were prideful in the general sense of desiring an undue excellence. And the upshot of this would be, specifically, a life-death decision on an astronomical scale: preventing lives perhaps by the million billions. It would mean we were prideful in this more specific sense too. Or in short: Humility: Non-transitionary anthropogenic extinction would mean we overstrained our power as a species. And the upshot of it would be a life-death decision on an astronomical scale. Thus it would amount to an enormous form of superbia. That's a second consideration for why our extinction would be problematic. And here too, I suppose this consideration should appeal to Christians quite broadly. Again, Thomas's specific interpretation of God's authority over life/death-decisions should be controversial. He suggests that, while permitting us to execute death penalty (ST, ii-II, q64 a2, co), this authority absolutely prohibits suicide (ST, ii-II, q64 a5, co), and prohibits abortion from the moment of ensoulment (see e.g. ST, I, q118f.)-which some people have claimed, according to Thomas's metaphysical principles and the known facts about embryology, takes place at the moment of fertilisation  (Haldane and Lee 2003, 273) .  5  This would mean self-extinction can under no circumstances be permissible. And it would mean that contemporary liberal moral thought is radically wrong about death penalty, abortion, or suicide and euthanasia. Many contemporary Christians will want to resist these implications. But again we needn't interpret the general idea of superbia in this manner. After all, it's implausible that criminals are 'dangerous and infectious to the community' and must be cut away like an infected part of a body (ST, ii-II, q64 a2, co). Perhaps suicide can in some instances of extreme pain, or loss of autonomy or dignity, be a manifestation of self-love. And if so, perhaps we can then view ourselves as authorised by God-or His commandment to love ourselves (e.g. Mat 22:39)-to end our lives. After all, Thomas himself (defending Abraham's intention to kill Isaac) suggests that 'he who at God's command kills an innocent man does not sin' (ST, ii-II, q64 a6, ad1). Perhaps some forms of abortion (e.g. after rape) can also be seen as an expression of self-love or -respect. Or perhaps we must ascribe to Aquinas a different view of ensoulment  (Pasnau 2002) , or simply reject some of his metaphysics in light of more recent discoveries. Again, the details of superbia are contested.  6  But I assume that the pertinent general idea is still prominent among Christians today. Indeed, that playing God is a sin is a natural corollary of theism. The in-principle-ban on life/death-decisions has a very firm grounding in the Bible (Thomas cites Deut 32:39, for instance). And I suppose that on any plausible interpretation, these ideas ground a consideration against extinction: whatever they imply about death penalty, abortion, and suicide, the dictates of humility seem hardly compatible with our developing a technology that accidentally seals the fate of our whole species. So again, such a consideration should have import for Christians quite broadly. Self-inflicted extinction would constitute a shattering form of superbia. \n The value of humanity But we need to see another aspect of Thomas's view of humanity too, which is perhaps most distinctly Biblical, and which is implicit in the third kind of inclination he ascribes us. For sure, we're not God. But we are nonetheless made to know the truth about Him. In fact, of all corporeal things (i.e., disregarding angels) we're the only ones whose nature enables such knowledge. And this, for Thomas, marks our 'dignity'. It's this dignity which made it fitting for Jesus to adopt human nature, rather than becoming an animal or any other thing: 'in the irrational creature the fitness of dignity is wanting.' (ST, III, q4 a1, co) Indeed, it's these intellectual capacities that ground our likeness to Goda likeness greater than that of any non-reasoned creature (see ST, I, q93 a2, co), and sufficient to say we're His 'image': 'intellectual creatures alone  [...]  are made to God's image.' (ST, I, q93 a2, co) And it's this which ultimately manifests that God loves us more than any other thing. So Aquinas quotes Augustine approvingly: \"'God loves all things that He has made, and amongst them rational creatures more, and of these especially those who are members of His only-begotten Son Himself.\"' (ST, I, q20 a3, sc) In other words, and as Thomas says very explicitly (ST, I, q20 a4, co), of all 6 There's a lot of recent literature concerning Thomas on humility. For longer treatments, see e.g.  Tadie (2006)  or  Fullam (2001) . As far as I see, Thomas himself doesn't discuss whether birth control or contraception infringes on God's dominion over life/death-decisions. But the Catholic church has since leveraged this argument. Yet perhaps a ban on contraception needn't follow from humility either. Perhaps Thomists may appeal to the doctrine of double effect, and compare well-intentioned prevention of fertilisation with well-intentioned killing in selfdefense (cf. section 3.2). Or perhaps they might again claim to be authorised in this decision by the commandment to love ourselves, and love our partners, and our (potential) children for whom we couldn't be sufficiently responsible parents. corporeal things we are the most valuable, or those with most 'goodness'. And this difference is categorical. So our extinction would have cosmological ramifications. For standard utilitarians, there's no categorical difference between a world populated by flourishing societies of people and a world populated by one lonely lizard basking in the sun and feeling a tinge of pleasure. The difference is a matter of degree. For Aquinas that's different. In a world devoid of higher intelligence, there's nothing that's made to God's image-nothing with our special dignity or perfection. This would radically undercut the perfection of the universe, as God created it, which depends on the varieties of goodness (see e.g. ST, I, q47). So our extinction would change the face of creation. Or in short: The value of humanity: Non-transitionary anthropogenic extinction would constitute the destruction of the most valuable aspect of creation-and thus a loss of categorical cosmological significance. That's a third consideration for why our extinction would be problematic. Again, I suppose that this consideration should appeal to Christians very broadly. Thomas interprets the order of creation very radically. He suggests that everything else exists for us (ST, ii-II, q64 a1, co)-such that we may use animals simply as means for our ends, say, and the self-interested killing of another person's ox wrongs at most its owner (ST, ii-II, q64 a1, co). This would mean that our extinction would literally destroy the purpose of the physical universe. And it would mean that much contemporary thought is radically wrong about animals, or other aspects of the natural world. Again, many Christians will want to resist these implications, or so I hope. But the distinct value of humans needn't be interpreted this radically. Perhaps our 'dominion' over animals (Gen 1:26) doesn't mean we can use them simply as means. Perhaps it means we must use our de facto power in the manner of a loving and respectful and liberal guard (see e.g. Linzey 2016, ch. 2). And similarly with every other aspect of creation-ecosystems, plants, and the climate. Again, the exact nature of our status is a large question. But the general idea of our special value still seems very prominent. It does seem hard to avoid if we take seriously that man is 'God's image', or has a special (perhaps responsibility-implying) dominion over the earth-and very plausible given God's human embodiment in Christ. And on any interpretation, our special status will ground a consideration against extinction: whatever it implies about our responsibilities towards animals, the face of creation will be changed categorically if we drive ourselves extinct. So this consideration should be authoritative to Christians very broadly. Self-inflicted extinction would constitute a destruction with cosmological import. Plausibly, there are other Christian considerations that bear on x-risks, at least once we move beyond Thomas. Most notably, perhaps Christian caritas simply gives us reasons to increase expected impartial welfare (see e.g. William Paley's utilitarian Christianity, in his Principles of  Moral and Political Philosophy; Paley 2014, esp. book II) . If it does, the EA rationale will be very pertinent to Christians as it stands. But whether that's an apt interpretation of the Christian virtue-or whether, as I personally find more likely, the latter is more about making people happy than about making happy people-will be controversial. And at any rate, I think these are the three most distinctly Christian concerns. So let me leave it at that for now. \n Discussion If all of this is right, non-transitionary anthropogenic extinction would be a morally highly problematic result. Such extinction wouldn't just amount to a regrettable form of imprudence-but to a failure to fulfill our God-given role. It wouldn't amount to any old failure of this kind-but to a prideful miss of our end. And it wouldn't be just an inconsequential overstepping of our dominionbut an unauthorised decision with categorical cosmological ramifications. 7 But as mentioned, humanity can now (in the form of the present generation) take measures to reduce the likelihood that it will eventually effect this result. So this suggests that we have very strong reasons to take some such measures. It suggests that Christians too have strong reasons to donate their money towards x-risk reduction rather than, say, disaster relief; to conduct x-risk research rather than enquiries about the cause of the grief of a neighbour; or to advocate political measures for long-term safety rather than for short-term caritative purposes. But let's not get ahead of ourselves. Some considerations in Thomas's philosophy seem to suggest that even if extinction would, as far as we know, be bad, it's not important that we now do much about it in practice. So let me turn to some of these theological and moral countervailing concerns. This will not simply corroborate the results so far. It will also help clarify the precise form and weight of what the above rationale implies. \n Theological considerations: Divine providence One apparent reason against x-risk activism is Divine providence. Again, for Thomas, everything is subject to God. And Thomas's God is all-knowing, allpowerful, and all-good: 'In God there exists the most perfect knowledge' (ST, I, q14 a1, co), 'He can do all things that are possible absolutely [or don't imply a contradiction in terms]' (ST, I, q25 a3, co); and He 'loves all existing things' (ST, I, q20 a2, co). One might conclude from this that God won't allow us to go extinct, or that if He does it's somehow just for the better. And from this one might infer that we needn't do anything to prevent our extinction in practice. X-risk reduction is obviated by the rule of a loving God. Now this is difficult and well-trodden terrain. But God's existence doesn't in general seem to warrant any heedlessness. Consider risks of road accidents, say. And suppose any accident accords with His will. It's an interesting question what precisely this implies. Perhaps it warrants a certain ultimate serenity or comfort: a reassurance at the thought that whatever happens on our roads has its mysterious rightness in the grander scheme of things. Perhaps it warrants a fundamental form of trust: a consolation in the belief that our diligence will be duly rewarded (see ST, ii-II, q22 a1, co). But God's unfathomable values cannot act as a guide for us, or as an excuse to depart from the norms and expectations we're given. So these forms of Christian hope must be distinguished from optimism that accidents won't occur, or from feeling exonerated from the need to take caution. Christians ought to see to it that they stop at red lights, respect speed limits, and check their brakes-and just as carefully as everyone else. Indeed, Thomas is the first to emphasise that we're obligated to such circumspection (see e.g. ST, ii-II, q64 a8, co). In short, Divine providence doesn't affect our need to reduce road risks in practice. But then it's unclear why it should do so with risks of human extinction.  8  One might suggest that there's something special about human extinction, setting it apart from everyday hazards, and making it particularly unlikely. Perhaps it's precisely because we're His 'image', or uniquely most valuable, that God won't allow us to perish. Perhaps Jesus's redemption of mankind would have been futile if some sorrowful two thousand years later He let us burn up or wither away. 9 Or perhaps, more specifically, there's evidence in scripture that we'll survive: the primal great flood precisely didn't erase us, and after that deluge there seems to have been a promise of survival-that 'never again will all life be destroyed by the waters of a flood.' (Gen 9:11) This isn't the place to dive deep into Christian theology. But it's hard to see why such reasonings should give us much confidence. The Lord moves in mysterious ways. If all known misery is compatible with His providence, it must surely at least be possible that He'd allow us go extinct. In particular, there are many more passages in scripture painting a grim prospect of extinction than promising a boundless glorious future. And many such visions of the 'end of time' indicate anthropogenic extermination-that 'the nations in the four corners of the earth' are gathered for 'battle' or 'war' (Rev 20:8), right before 'the earth and the heavens' are gone  (Rev 20:11) . If anything, according to the Bible, anthropogenic extinction seems a very live option. So Divine providence might have implications for the metaphysics and ultimate unfathomable significance of our doomsday. But it doesn't seem to warrant our heedlessness about x-risks in practice. If the above considerations are sound, we, just as any generation in the future, must see to it that it isn't us who bring a Johannine finale about. \n Moral considerations: deontology Let's look at moral considerations that might mitigate the importance of xrisk reduction. One such consideration concerns intentions. For Thomas, the permissibility of our actions can depend on what we intend. More precisely, he held, or indeed introduced, the doctrine of double effect (or 'DDE'). According to this doctrine, there's a difference between the effects we intend an action to have, and the effects we merely foresee but don't specifically intend. While it's always impermissible to intend evil effects, it can be permissible to cause them, if what you intend is good. For instance, you may kill in self-defence, if you intend to save your life and merely foresee the death of your aggressor (see ST, ii-II, q64 a7, co). This bears on x-risks. Perhaps non-transitionary anthropogenic extinction would be a very bad result. But one might suggest that as long as we don't intend to effect it, we wouldn't have acted impermissibly: an unintentional 8 In his contribution to this collection, Dominic Roser suggests that Christianity warrants a certain renunciation of control. I'm not sure whether he intends this to contradict my claims here-and to imply, say, that Christians may reasonably check their brakes half as often as atheists. If so, his view seems to me, among other things, unstable. It seems that God's existence would either require standard full-blown caution (as I think Thomas suggests), or would warrant total relinquishment of it. It seems hard to see why He would warrant a certain but only a certain trust in the course of events. At any rate, I'd think most Christians don't understand their faith in this manner. And even if Roser is right, any such limited trust in things would arguably still not undermine my practical conclusions. 9 I thank Felix Koch for mentioning this thought to me. mass life-prevention isn't prideful, and an unintentional thwarting of perfection or erasing of God's image isn't morally wrong. So if our intentions are fine, we needn't worry about x-risk reduction. 10 But this argument is a non-starter. It might well be worse to intend our extinction rather than to merely foresee it. For Thomas, it would mean that our action would lack 'goodness from its end', and not just from its circumstances or species  (ST, q18, esp. a7, co) . But this doesn't mean that unintentional extinction wouldn't be wrong, or even specifically prideful. There's no blanket permission for the well-intentioned. You mustn't speed through the cross-walk and put people at risk, even with the laudable aim of being punctual at the faculty meeting. Thomas says explicitly that you are 'in a sense guilty of voluntary homicide'-and thus (I take it) of superbia-if you kill someone unintentionally but without having taken 'due care' (ST, ii-II, q64 a8, co). He doesn't specify when precisely the DDE allows you to accept a foreseen evil. But he suggests it depends on (i) the expected goodness of the intended good, (ii) the expected badness of the foreseen evil-or on the 'proportion' between the two (ST, ii-II, q64 a7, co)-and (iii) on whether the harmful action is 'necessary' (ST, ii-II, q64 a7, co) for the good, or whether there are ways of securing the good without these bad effects. These criteria decidedly command x-risk reduction. Sure, many technologies that involve such risks also promise important goods. But if my arguments are right, the moral costs are potentially tremendous. And there are less risky courses of action which detract little from the expected good. The above-mentioned measures would reduce x-risks, but wouldn't really jeopardise the benefits these technologies promise. So it seems absolutely obligatory to take them. Going on as we do does seem like scorching through the cross-walk for timeliness at the meeting. That isn't killing for the sake of the killing. But it's a serious violation of the negative constraint of the DDE. And thus it's surely grave enough.  11  Let's consider another countervailing consideration. Perhaps extinction would be terrible. But for all we know, it isn't actually imminent. Indeed, compared to more immediate moral callings like global poverty, gender justice, or disaster relief, it seems very remote. According to utilitarianism or EA, such distance doesn't matter. These views are absolutely impartial. But perhaps Christian morality is different. One might suggest Christian morality is more concerned with visible, immediate, near-at-hand moral problems-the wounded person alongside the road (Luke 10:30), or perhaps the literal 'neighbor' (Mark 12:31)-and doesn't warrant too much concern about such far-distant matters. In fact, Thomas himself is explicit that we should be partial, at least in the virtue of caritas: 'Among our neighbors, we should love them more who are more closely connected to us' (DQV, q2 a9, co; cf. ST ii-II, q31 a3, co). So one might conclude that although our extinction would be bad, we needn't worry about x-risk reduction-or at least not now, or not with resources we could direct towards these more immediate concerns. But this would be a mistake. Christian morality may be partialist about caritas. So there may be a tension between Christian benevolence and the impartialist EA-directive to 'do the most good'. But on the considerations I've sketched, reducing x-risks isn't a form of 'doing good', or of charity, or of fulfilling any positive obligation. It's to ensure-or make more likely-that we don't flout our end, overstep our dominion, or wreck the crown of creation. Formally, it's to ensure we satisfy the constraint of the DDE. In other words, it's to respect a perfect negative duty. And Thomas doesn't seem to endorse partiality or discounting in such negative duties, and neither does Christian morality more broadly. On Christian morality, killing, say, is wrong. And doing what will kill someone in a month is presumably ceteris paribus just as wrong as doing what will kill someone in a year, or in a thousand years for that matter. So the sheer temporal distance of human extinction in itself doesn't seem to obviate x-risk reduction. On the contrary, note that the negative character of this obligation has implications for how it may be weighed against others. For utilitarians, the obligation to reduce x-risk is formally on a par with (other) obligations of benevolence, such as obligations to support the poor. So it ought to be weighed against them. For Christians, if I'm right, this is different. As a negative obligation, the obligation to reduce x-risks must not be weighed against positive imperfect obligations, or can't be discharged by doing enough by way of benevolence elsewhere. It's quite simply a perfect duty. There's a final point worth noting. For utilitarians, notoriously, we may take any means for the sake of the good end. If killing an innocent scientist reduces x-risks, we presumably ought to kill them. For Thomist Christians this will be different. Thomas explicitly accepts deontological constraints: 'some [actions] are evil, whatever their result may be.' (ST ii-II, q88 a2, ad2) So Thomist Christians must not do anything to reduce such risks. They generally mustn't kill or lie or steal for that purpose. In practice, however, this won't make much of a difference. The most salient means of x-risk reduction include academic research, or global measures for peace, sustainability or AI safety. None of these measures seem to violate any constraints. So even if Christian morality is deontological in nature, there's plenty of good ways to start making Survival more likely. \n Conclusion If all of this is right, it's not just that non-transitionary anthropogenic extinction would, as far as we know, be a disastrous result-a cosomologically salient prideful miss of our end. At least in practice, and as a matter of a negative constraint against lack of 'due care' that in principle seems neither discountable nor weighable against positive obligations, Thomist Christians have very strong reasons to take deontologically permissible means to prevent such extinction: to conduct research on risks and risk-reduction, promote political arrangements with an eye to the very far future, donate heir money towards x-risk mitigation, and so on. Indeed, given the gravity of the possible effect, and the perfect or negative form of the duty, they seem to have reason to do this even if it will considerably constrain their resources for classical forms of caritas. Or that's as far as our argument takes us. There are many considerations relevant to non-transitionary anthropogenic extinction that I haven't yet addressed. For instance, it would be interesting to explore the implications of Christian love for our question; to consider the relevance of other Christian virtues-such as temperance, diligence, or patience; or to integrate the assumption that in some form or other we'll always live on, or that in this sense we can't really die out. 12 That's beyond the scope of this paper. It would also be interesting to consider the implications of my arguments for issues beyond my core question. Take risks of natural extinction. If we ought to ensure the preservation of the species, we must generally guard ourselves also against asteroid hits. It would presumably be a form of superbia to presume that no such event could erase us. And to let it happen would mean to let creation's most valuable part be destroyed. Allowing natural extinction might not be as grave as actively extinguishing ourselves, but still seems a large-scale moral failure. Another, and more intricate question, concerns the issue of non-human higher intelligence. Suppose we turned into a nun-human AI and left mankind behind. Would we then miss our end (in the non-preservation of the species), or would we precisely fulfill it (in the perfection of our ingenuity)? Would it be a form of superbia to thus intervene in the nature of species, or would that belong to our natural proportionate capacities? And would the most valuable aspect of the universe be lost, or would the cosmos become more valuable if inhabited by a more perfect form of reason? 13 Again, these are questions for another occasion. There's also the question whether similar arguments emerge on other religious or non-religious worldviews. They arguably do. The normative core of the first two arguments is simply a kind of perfectionism; and the core of the third argument is the idea of a special value of humanity. In Christianity these assumptions are grounded in the existence of the Christian God. But they needn't be. In some form or other, you can adopt them without believing in that God, or indeed in any form of theism. We might see them as emerging, at bottom, from very general aspects of the Christian view of ourselves: from the ambivalent sense that there's something greater than us, but that we nonetheless are quite great indeed. Perfectionism expresses the idea that what structures the universe exceeds mankind and is somehow commanding for us-so that we ought to fulfill our role in the scheme of things, and accept the limitations of it. The idea of our special value expresses that we have a certain greatness nonetheless-that our role isn't any old role, and that it would matter if we were gone. It's at bottom this dichotomous nature, fallen and fragile yet valuable, which implies we should be exceedingly careful with ourselves. Christians, and everyone with a similar view of mankind, and of our obligations and limitations and value, should agree with Effective Altruists that x-risk reduction is extremely important. \t\t\t Thomas doesn't say this explicitly in these passages. He says that the goods of the body are subordinate or 'directed' to the goods of the soul, and that virginity (or a lack of bodily good) is justified if done for the sake of Divine contemplation (or good of the soul; ST ii-II, q152 a2, co). I'm reading the conditional here as a biconditional. \n\t\t\t For a classic systematic exploration of natural law, see Finnis (2011) . For an in-depth treatment of Aquinas's theory, see e.g.-among very many others-Lisska (1998).4 Historically, this broad rationale hasn't appealed to all Christians, of course. Paul himself seems to suggest it would be best if everyone was celibate (see 1 Cor 7:7). And the Cathars even thought reproduction was a moral evil. I thank Felix Timmermann and Christoph Halbig for pointing this out to me. \n\t\t\t In the Scriptum super Sententiis, following Aristotle, Thomas himself suggests that the soul is infused after 40 days for males and after 90 days for females (SSS III, d3 q5 a2, co). \n\t\t\t Perhaps it's worth noting that in principle, these considerations are logically independent. One might accept some but not all of them. For instance, one might doubt that our natural inclinations reveal a Divine law, but still accept the cosmological disvalue of human extinction. \n\t\t\t I thank Eric Sampson and Jonathan Erhardt for mentioning this thought to me. 11 Strictly speaking, perhaps extinction wouldn't even be a 'foreseen' effect of our actions. By 'foreseen' [praecogitatus] Thomas seems to mean 'known with certainty to result' (see e.g. ST, i-II, q20 a5, co). But we aren't certain that promoting AI, say, will result in extinction. We're just not certain that it won't. In discussing the DDE, Thomas doesn't deal with uncertainty. But he does elsewhere. In ST, i-II, q20 a5, co, he says: 'if the consequences [of an action] are not foreseen, we must make a distinction. Because if they follow from the nature of the action and in the majority of cases [...] the consequences increase the goodness or malice of that action[...]. On the other hand, if the consequences follow by accident and seldom, then they do not increase the goodness or malice of the action.' This can't be right. Throwing stones off a cliff might kill passersby below in only 5% of cases. But then it's nonetheless wrong. Thomas lacked the concept of expected value. Today, we'd surely interpret the doctrine in terms of that concept, or some related one. \n\t\t\t I thank Dominic Roser for mentioning this thought to me. 13 I thank Mara-Daria Cojocaru and Carin Ism for mentioning these questions to me. For a generally positive Christian stance on enhancement, see e.g. Keenan (2014) .", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Stefan-Riedener_Existential-risks-from-a-Thomist-Christian-perspective.tei.xml", "id": "2a5f91b970814bdd9710e9bf3c90d7a5"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Studies of superintelligent-level systems have typically posited AI functionality that plays the role of a mind in a rational utility-directed agent, and hence employ an abstraction initially developed as an idealized model of human decision makers. Today, developments in AI technology highlight intelligent systems that are quite unlike minds, and provide a basis for a different approach to understanding them: Today, we can consider how AI systems are produced (through the work of research and development), what they do (broadly, provide services by performing tasks), and what they will enable (including incremental yet potentially thorough automation of human tasks). Because tasks subject to automation include the tasks that comprise AI research and development, current trends in the field promise accelerating AI-enabled advances in AI technology itself, potentially leading to asymptotically recursive improvement of AI technologies in distributed systems, a prospect that contrasts sharply with the vision of self-improvement internal to opaque, unitary agents. The trajectory of AI development thus points to the emergence of asymptotically comprehensive, superintelligent-level AI services thatcrucially-can include the service of developing new services, both narrow and broad, guided by concrete human goals and informed by strong models of human (dis)approval. The concept of comprehensive AI services (CAIS) provides a model of flexible, general intelligence in which agents are a class of service-providing products, rather than a natural or necessary engine of progress in themselves. Ramifications of the CAIS model reframe not only prospects for an intelligence explosion and the nature of advanced machine intelligence, but also the relationship between goals and intelligence, the problem of harnessing advanced AI to broad, challenging problems, and fundamental considerations in AI safety and strategy. Perhaps surprisingly, strongly self-modifying agents lose their instrumental value even as their implementation becomes more accessible, while the likely context for the emergence of such agents becomes a world already in possession of general superintelligent-level capabilities. These prospective capabilities, in turn, engender novel risks and opportunities of their own. Further topics addressed in this work include the general architecture of systems with broad capabilities, the intersection between symbolic and neural systems, learning vs. competence in definitions of intelligence, tactical vs. strategic tasks in the context of human control, and estimates of the relative capacities of human brains vs. current digital systems.", "authors": ["Eric Drexler"], "title": "Reframing Superintelligence Comprehensive AI Services as General Intelligence", "text": "Preface The writing of this document was prompted by the growing gap between models that equate advanced AI with powerful agents and the emerging reality of advanced AI as an expanding set of capabilities (here, \"services\") in which agency is optional. A service-centered perspective reframes both prospects for superintelligent-level AI and a context for studies of AI safety and strategy. Taken as a whole, this work suggests that problems centered on what highlevel AI systems might choose to do are relatively tractable, while implicitly highlighting questions of what humans might choose to do with their capabilities. This shift, in turn, highlights the potentially pivotal role of high-level AI in solving problems created by high-level AI technologies themselves. The text was written and shared as a series of widely-read Google Docs released between December 2016 and November 2018, largely in response to discussions within the AI safety community. The organization of the present document reflects this origin: The sections share a common conceptual framework, yet address diverse, overlapping, and often loosely-coupled topics. The table of contents, titles, subheads, summaries, and internal links are structured to facilitate skimming by readers with different interests. The table of contents primarily of declarative sentences, and has been edited to read as an overview. Several apologies are in order: A number of topics and examples assume a basic familiarity with deep-learning concepts and jargon, while much of the content assumes familiarity with concerns regarding artificial general intelligence circa 2016-18; some sections directly address concepts and concerns discussed in Superintelligence  (Bostrom 2014) . In this work, I have made little effort to assign proper scholarly credit to ideas: Concepts that seem natural, obvious, or familiar are best treated as latent community knowledge and very likely have uncited antecedents. Ideas that can reasonably be attributed to someone else probably should be. Finally, how I frame and describe basic concepts has shifted over time, and in the interests of early completion, I have made only a modest effort to harmonize terminology across the original documents. I thought it best to share the content without months of further delay. \n I Introduction: From R&D automation to comprehensive AI Services Responsible development of AI technologies can provide an increasingly comprehensive range of superintelligent-level AI services-including the service of developing new services-and can thereby deliver the value of general-purpose AI while avoiding the risks associated with self-modifying AI agents. \n I.1 Summary The emerging trajectory of AI development reframes AI prospects. Ongoing automation of AI R&D tasks, in conjunction with the expansion of AI services, suggests a tractable, non-agent-centric model of recursive AI technology improvement that can implement general intelligence in the form of comprehensive AI services (CAIS), a model that includes the service of developing new services. The CAIS model-which scales to superintelligent-level capabilities-follows software engineering practice in abstracting functionality from implementation while maintaining the familiar distinction between application systems and development processes. Language translation exemplifies a service that could incorporate broad, superintelligent-level world knowledge while avoiding classic AI-safety challenges both in development and in application. Broad world knowledge could likewise support predictive models of human concerns and (dis)approval, providing safe, potentially superintelligent-level mechanisms applicable to problems of AI alignment. Taken as a whole, the R&D-automation/CAIS model reframes prospects for the development and application of superintelligence, placing prospective AGI agents in the context of a broader range of intelligent systems while attenuating their marginal instrumental value. \n I.2 The trajectory of AI development reframes AI prospects Past, present, and projected developments in AI technology can inform our understanding of prospects for superintelligent-level capabilities, providing a concrete anchor that complements abstract models of potential AI systems. A development-oriented perspective highlights path-dependent considerations in assessing potential risks, risk-mitigation measures, and safety-oriented research strategies. The current trajectory of AI development points to asymptotically recursive automation of AI R&D that can enable the emergence of general, asymptotically comprehensive AI services (CAIS). In the R&Dautomation/CAIS model, recursive improvement and general AI capabilities need not be embodied in systems that act as AGI agents. Technology improvement proceeds through research and development, a transparent process that exposes component tasks to inspection, refactoring, and incremental automation.  1  If we take advanced AI seriously, then accelerating, asymptotically-complete R&D automation is a natural consequence: • By hypothesis, advances in AI will enable incremental automation and speedup of all human tasks. • As-yet unautomated AI R&D tasks are human tasks, hence subject to incremental automation and speedup. • Therefore, advances in AI will enable incremental automation and speedup of all AI R&D tasks. Today we see automation and acceleration of an increasing range of AI R&D tasks, enabled by the application of both conventional software tools and technologies in the AI spectrum. Past and recent developments in the automation of deep-learning R&D tasks include: • Diverse mechanisms embodied in NN toolkits and infrastructures • Black-box and gradient-free optimization for NN hyperparameter search  (Jaderberg et al. 2017 ) • RL search and discovery of superior NN gradient-descent algorithms  (Bello et al. 2017 ) • RL search and discovery of superior NN cells and architectures  (Zoph et al. 2017)  Today, automation of search and discovery (a field that overlaps with \"metalearning\") requires human definition of search spaces, and we can expect that the definition of new search spaces-as well as fundamental innovations in architectures, optimization methods, and the definition and construction of tasks-will remain dependent on human insight for some time to come. Nonetheless, increasing automation of even relatively narrow search and discovery could greatly accelerate the implementation and testing of advances based on human insights, as well as their subsequent integration with other components of the AI technology base. Exploring roles for new components (including algorithms, loss functions, and training methods) can be routine, yet important: as Geoff Hinton has remarked, \"A bunch of slightly new ideas that play well together can have a big impact\". Focusing exclusively on relatively distant prospects for full automation would distract attention from the potential impact of incremental research automation in accelerating automation itself. \n I.4 R&D automation suggests a service-centered model of general intelligence AI deployment today is dominated by AI services such as language translation, image recognition, speech recognition, internet search, and a host of services buried within other services. Indeed, corporations that provide cloud computing now actively promote the concept of \"AI as a service\" to other corporations. Even applications of AI within autonomous systems (e.g., self-driving vehicles) can be regarded as providing services (planning, perception, guidance) to other system components. R&D automation can itself be conceptualized as a set of services that directly or indirectly enable the implementation of new AI services. Viewing service development through the lens of R&D automation, tasks for advanced AI include: • Modeling human concerns • Interpreting human requests • Suggesting implementations • Requesting clarifications • Developing and testing systems • Monitoring deployed systems • Assessing feedback from users • Upgrading and testing systems CAIS functionality, which includes the service of developing stable, taskoriented AI agents, subsumes the instrumental functionality of proposed self-transforming AGI agents, and can present that functionality in a form that better fits the established conceptual frameworks of business innovation and software engineering. \n I.5 The services model abstracts functionality from implementation Describing AI systems in terms of functional behaviors (\"services\") aligns with concepts that have proved critical in software systems development. These include separation of concerns, functional abstraction, data abstraction, encapsulation, and modularity, including the use of client/server architectures-a set of mechanisms and design patterns that support effective program design, analysis, composition, reuse, and overall robustness. Abstraction of functionality from implementation can be seen as a figureground reversal in systems analysis. Rather than considering a complex system and asking how it will behave, one considers a behavior and asks how it can be implemented. Desired behaviors can be described as services, and experience shows that complex services can be provided by combinations of more specialized service providers, some of which provide the service of aggregating and coordinating other service providers. \n I.6 The R&D automation model distinguishes development from functionality The AI-services model maintains the distinction between AI development and AI functionality. In the development-automation model of advanced AI ser-vices, stable systems build stable systems, avoiding both the difficulties and potential dangers of building systems subject to open-ended self-transformation and potential instability. Separating development from application has evident advantages. For one, task-focused applications need not themselves incorporate an AI-development apparatus-there is little reason to think that a system that provides online language translation or aerospace engineering design services should in addition be burdened with the tasks of an AI developer. Conversely, large resources of information, computation, and time can be dedicated to AI development, far beyond those required to perform a typical service. Likewise, in ongoing service application and upgrade, aggregating information from multiple deployed systems can provide decisive advantages to centralized development (for example, by enabling development systems for self-driving cars to learn from millions of miles of car-experience per day). Perhaps most important, stable products developed for specific purposes by a dedicated development process lend themselves to extensive pre-deployment testing and validation. \n I.7 Language translation exemplifies a safe, potentially superintelligent service Language translation provides an example of a service best provided by superintelligent-level systems with broad world knowledge. Translation of written language maps input text to output text, a bounded, episodic, sequence-to-sequence task. Training on indefinitely large and broad text corpora could improve translation quality, as could deep knowledge of psychology, philosophy, history, geophysics, chemistry, and engineering. Effective optimization of a translation system for an objective that weights both quality and efficiency would focus computation solely on the application of this knowledge to translation. The process of developing language translation systems is itself a service that can be formulated as an episodic task, and as with translation itself, effective optimization of translation-development systems for both quality and efficiency would focus computation solely on that task. There is little to be gained by modeling stable, episodic service-providers as rational agents that optimize a utility function over future states of the world, hence a range of concerns involving utility maximization (to say nothing of self-transformation) can be avoided across a range of tasks. Even superintelligent-level world knowledge and modeling capacity need not in itself lead to strategic behavior. \n I.8 Predictive models of human (dis)approval can aid AI goal alignment As noted by Stuart Russell, written (and other) corpora provide a rich source of information about human opinions regarding actions and their effects; intelligent systems could apply this information in developing predictive models of human approval, disapproval, and disagreement. Potential training resources for models of human approval include existing corpora of text and video, which reflect millions of person-years of both real and imagined actions, events, and human responses; these corpora include news, history, fiction, science fiction, advice columns, law, philosophy, and more, and could be augmented and updated with the results of crowd-sourced challenges structured to probe model boundaries. Predictive models of human evaluations could provide strong priors and common-sense constraints to guide both the implementation and actions of AI services, including strategic advisory services to powerful actors. Predictive models are not themselves rational agents, yet models of this kind could contribute to the solution of agent-centered safety concerns. In this connection, separation of development from application can insulate such models from perverse feedback loops involving self-modification. \n I.9 The R&D-automation/CAIS model reframes prospects for superintelligence From a broad perspective, the R&D-automation/CAIS model: • Distinguishes recursive technology improvement from self-improving agents • Shows how incremental automation of AI R&D can yield recursive improvement • Presents a model of general intelligence centered on services rather than systems • Suggests that AGI agents are not necessary to achieve instrumental goals • Suggests that high-level AI services would precede potential AGI agents • Suggests potential applications of high-level AI services to general AI safety For the near term, the R&D-automation/CAIS model: • Highlights opportunities for safety-oriented differential technology development • Highlights AI R&D automation as a leading indicator of technology acceleration • Suggests rebalancing AI research portfolios toward AI-enabled R&D automation Today, we see strong trends toward greater AI R&D automation and broader AI services. We can expect these trends to continue, potentially bridging the gap between current and superintelligent-level AI capabilities. Realistic, pathdependent scenarios for the emergence of superintelligent-level AI capabilities should treat these trends both as an anchor for projections and as a prospective context for trend-breaking developments. II Overview: Questions, propositions, and topics \n II.1 Summary This document outlines topics, questions, and propositions that address: 1. Prospects for an intelligence explosion 2. The nature of advanced machine intelligence 3. The relationship between goals and intelligence 4. The problem of using and controlling advanced AI 5. Near-and long-term considerations in AI safety and strategy The questions and propositions below reference sections of this document that explore key topics in more depth. From the perspective of AI safety concerns, this document offers support for several currently-controversial propositions regarding artificial general intelligence: • That AGI agents have no natural role in developing general AI capabilities. • That AGI agents would offer no unique and substantial value in providing general AI services. • That AI-based security services could safely constrain subsequent AGI agents, even if these operate at a superintelligent level. II.2 Reframing prospects for an intelligence explosion II.2.1 Does recursive improvement imply self-transforming agents? Ongoing automation of tasks in AI R&D suggests a model of asymptoticallyrecursive technology improvement that scales to superintelligent-level (SIlevel) systems. In the R&D-automation model, recursive improvement is systemic, not internal to distinct systems or agents. The model is fully generic: It requires neither assumptions regarding the content of AI technologies, nor assumptions regarding the pace or sequence of automation of specific R&D tasks. Classic self-transforming AGI agents would be strictly more difficult to implement, hence are not on the short path to an intelligence explosion. • Section 1: R&D automation provides the most direct path to an intelligence explosion The application of an idealized, fully general learning algorithm would enable but not entail the learning of any particular competence. Time, information, and resource constraints are incompatible with universal competence, regardless of ab initio learning capacity. • Section 2: Standard definitions of \"superintelligence\" conflate learning with competence • Section 21: Broad world knowledge can support safe task performance • Section 39: Tiling task-space with AI services can provide general AI capabilities  II.3  Reframing the nature of advanced machine intelligence II.3.1 Is human learning an appropriate model for AI development? Action, experience, and learning are typically decoupled in AI development: Action and experience are aggregated, not tied to distinct individuals, and the machine analogues of cognitive change can be profound during system development, yet absent in applications. As we see in AI technology today, learning algorithms can be applied to produce and upgrade systems that do not themselves embody those algorithms. Accordingly, using human learning and action as a model for AI development and application can be profoundly misleading. • Section 2: Standard definitions of \"superintelligence\" conflate learning with competence • Section 7: Training agents in human-like environments can provide useful, bounded services • Section 16: Aggregated experience and centralized learning support AI-agent applications II.3.2 Does stronger optimization imply greater capability? Because optimization for a task focuses capabilities on that task, strong optimization of a system acts as a strong constraint; in general, optimization does not extend the scope of a task or increase the resources employed to perform it. System optimization typically tends to reduce resource consumption, increase throughput, and improve the quality of results. • Section 8: Strong optimization can strongly constrain AI capabilities, behavior, and effects II.3.3 Do broad knowledge and deep world models imply broad AI capabilities? Language translation systems show that safe, stable, high-quality task performance can be compatible with (and even require) broad and deep knowledge about the world. The underlying principle generalizes to a wide range of tasks. • Section 21: Broad world knowledge can support safe task performance II.3.4 Must we model SI-level systems as rational, utility-maximizing agents? The concept of rational, utility-maximizing agents was developed as an idealized model of human decision makers, and hence is inherently (though abstractly) anthropomorphic. Utility-maximizing agents may be intelligent systems, but intelligent systems (and in particular, systems of agents) need not be utility-maximizing agents. • Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame • Section 6: A system of AI services is not equivalent to a utility maximizing agent • Section 17: End-to-end reinforcement learning is compatible with the AI-services model • Section 18: Reinforcement learning systems are not equivalent to reward-seeking agents II.3.5 Must we model SI-level systems as unitary and opaque? Externally-determined features of AI components (including their development histories, computational resources, communication channels, and degree of mutability) can enable structured design and functional transparency, even if the components themselves employ opaque algorithms and representations. • Section 9: Opaque algorithms are compatible with functional transparency and control • Section 15: Development-oriented models align with deeply-structured AI systems • Section 38: Broadly-capable systems coordinate narrower systems  II.4  Reframing the relationship between goals and intelligence II.4.1 What does the orthogonality thesis imply for the generality of convergent instrumental goals? Intelligent systems optimized to perform perform bounded tasks (in particular, episodic tasks with a bounded time horizon) need not be agents with openended goals that call for self preservation, cognitive enhancement, resource acquisition, and so on; by Bostrom's orthogonality thesis, this holds true regardless of the level of intelligence applied to those tasks. • Section 19: The orthogonality thesis undercuts the generality of instrumental convergence II.4.2 How broad is the basin of attraction for convergent instrumental goals? Instrumental goals are closely linked to final goals of indefinite scope that concern the indefinite future. Societies, organizations, and (in some applications) high-level AI agents may be drawn toward convergent instrumental goals, but high-level intelligence per se does not place AI systems within this basin of attraction, even if applied to broad problems that are themselves long-term. • Section 21: Broad world knowledge can support safe task performance • Section 25: Optimized advice need not be optimized to induce its acceptance  II.5  Reframing the problem of using and controlling advanced AI II.5.1 Would the ability to implement potentially-risky self-transforming agents strongly motivate their development? If future AI technologies could implement potentially-risky, self-transforming AGI agents, then similar, more accessible technologies could more easily be applied to implement open, comprehensive AI services. Because the service of providing new services subsumes the proposed instrumental value of self-transforming agents, the incentives to implement potentially-risky self-transforming agents appear to be remarkably small. • Section 11: Potential AGI-enabling technologies also enable comprehensive AI services • Section 12: AGI agents offer no compelling value • Section 33: Competitive AI capabilities will not be boxed II.5.2 How can human learning during AI development contribute to current studies of AI safety strategies? We can safely predict that AI researchers will continue to identify and study surprising AI behaviors, and will seek to exploit, mitigate, or avoid them in developing AI applications. This and other predictable aspects of future knowledge can inform current studies of strategies for safe AI development. • Section 35: Predictable aspects of future knowledge can inform AI safety strategies II.5.3 Can we avoid strong trade-offs between development speed and human oversight? Basic research, which sets the overall pace of technological progress, could be safe and effective with relatively little human guidance; application development, by contrast, requires strong human guidance, but as an inherent part of the development task-to deliver desirable functionality-rather than as an impediment. Support for human guidance can be seen as an AI service, and can draw on predictive models of human approval and concerns. • Section 22: Machine learning can develop predictive models of human approval • Section 23: AI development systems can support effective human guidance • Section 24: Human oversight need not impede fast, recursive AI technology improvement II.5.4 Can we architect safe, superintelligent-level design and planning services? Consideration of concrete task structures and corresponding services suggests that SI-level AI systems can safely converse with humans, perform creative search, and propose designs for systems to be implemented and deployed in the world. Systems that provide design and planning services can be optimized to provide advice without optimizing to manipulate human acceptance of that advice. • Section 26: Superintelligent-level systems can safely provide design and planning services • Section 28: Automating biomedical R&D does not require defining human welfare II.5.5 Will competitive pressures force decision-makers to transfer strategic decisions to AI systems? In both markets and battlefields, advantages in reaction time and decision quality can motivate transfer of tactical control to AI systems, despite potential risks; for strategic decisions, however, the stakes are higher, speed is less important, advice can be evaluated by human beings, and the incentives to yield control are correspondingly weak. • Section 27: Competitive pressures provide little incentive to transfer strategic decisions to AI systems II.5.6 What does the R&D-automation/AI-services model imply for studies of conventional vs. extreme AI-risk concerns? Increasing automation of AI R&D suggests that AI capabilities may advance surprisingly rapidly, a prospect that increases the urgency of addressing conventional AI risks such as unpredictable failures, adversarial manipulation, criminal use, destabilizing military applications, and economic disruption. Prospects for the relatively rapid emergence of systems with broad portfolios of capabilities, including potentially autonomous planning and action, lend increased credence to extreme AI-risk scenarios, while the AI-services model suggests strategies for avoiding or containing those risks while gaining the benefits of high-and SI-level AI capabilities. • Section 12: AGI agents offer no compelling value • Section 14: The AI-services model brings ample risks II.5.7 What can agent-oriented studies of AI safety contribute, if risky AI agents are optional? People will want AI systems that plan and act in the world, and some systems of this class can naturally be modeled as rational, utility-directed agents. Studies of systems within the rational-agent model can contribute to AI safety and strategy in multiple ways, including: • Expanding the range of safe AI-agent architectures by better understanding how to define bounded tasks in a utility-directed framework. • Expanding safe applications of utility-directed agents to less wellbounded tasks by better understanding how to align utility functions with human values. • Better understanding the boundaries beyond which combinations of agent architectures and tasks could give rise to classic AGI-agent risks. • Better understanding how (and under what conditions) evolutionary pressures could engender perverse strategic behavior in nominally nonagent-like systems. • Exploring ways access to high-level AI services could could help to avoid or mitigate classic agent-centered AI risks. • Section 10: R&D automation dissociates recursive improvement from AI agency • Section 29: The AI-services model reframes the potential roles of AGI agents  II.6  Reframing near-and long-term considerations in AI safety and strategy II.6.1 Could widely available current or near-term hardware support superintelligence? Questions of AI safety and strategy become more urgent if future, qualitatively SI-level computation can be implemented with greater-than-human task throughput on affordable, widely-available hardware. There is substantial reason to think that this condition already holds. • Section 40: Could 1 PFLOP/s systems exceed the basic functional capacity of the human brain? II.6.2 What kinds of near-term safety-oriented guidelines might be feasible and useful? Current technology presents no catastrophic risks, and several aspects of current development practice align not only with safety, but with good practice in science and engineering. Development of guidelines that codify current good practice could contribute to near-term AI safety with little organizational cost, while also engaging the research community in an ongoing process that addresses longer-term concerns. Within the space of potential intelligent systems, agent-centered models span only a small region, and even abstract, utility-directed rational-agent models invite implicitly anthropomorphic assumptions. In particular, taking human learning as a model for machine learning has encouraged the conflation of intelligence-as-learning-capacity with intelligence-as-competence, while these aspects of intelligence are routinely and cleanly separated AI system development: Learning algorithms are typically applied to train systems that do not themselves embody those algorithms. The service-centered perspective on AI highlights the generality of Bostrom's Orthogonality Thesis: SI-level capabilities can indeed be applied to any task, including services that are (as is typical of services) optimized to deliver bounded results with bounded resources in bounded times. The pursuit of Bostrom's convergent instrumental goals would impede-not improve-the performance of such services, yet would be natural for those same instrumental goals to be hotly pursued by human organizations (or AI agents) that employ AI services. Prospects for service-oriented superintelligence reframe the problem of managing advanced AI technologies: Potentially-risky self-transforming agents become optional, rather than overwhelmingly valuable, the separation of basic research from application development can circumvent trade-offs between development speed and human oversight, and natural task architectures suggest safe implementations of SI-level design and planning services. In this connection, distinctions between tactical execution and strategic advice suggest that even stringent competitive pressures need not push decision-makers to cede strategic decisions to AI systems. In contrast to unprecedented-breakthrough models that postulate runaway self-transforming agents, prospects for the incremental emergence of diverse, high-level AI capabilities promise broad, safety-relevant experience with problematic (yet not catastrophic) AI behaviors. Safety guidelines can begin by codifying current safe practices, which include training and re-training diverse architectures while observing and studying surprising behaviors. The development of diverse, high-level AI services also offers opportunities for safety-relevant differential technology development, including the development of common-sense predictive models of human concerns that can be applied to improve the value and safety of AI services and AI agents. The R&D-automation/AI-services model suggests that conventional AI risks (e.g., failures, abuse, and economic disruption) are apt to arrive more swiftly than expected, and perhaps in more acute forms. While this model suggests that extreme AI risks may be relatively avoidable, it also emphasizes that such risks could arise more quickly than expected. In this context, agent-oriented studies of AI safety can both expand the scope of safe agent applications and improve our understanding of the conditions for risk. Meanwhile, service-oriented studies of AI safety could usefully explore potential applications of high-level services to general problems of value alignment and behavioral constraint, including the potential architecture of security services that could ensure safety in a world in which some extremely intelligent agents are not inherently trustworthy. \n R&D automation provides the most direct path to an intelligence explosion The most direct path to accelerating AI-enabled progress in AI technology leads through AI R&D automation, not through self-transforming AI agents. \n Summary AI-enabled development of improved AI algorithms could potentially lead to an accelerating feedback process, enabling a so-called \"intelligence explosion\". This quite general and plausible concept has commonly been identified with a specific, challenging, and risky implementation in which a predominant concentration of AI-development functionality is embodied in a distinct, goaldirected, self-transforming AI system-an AGI agent. A task-oriented analysis of AI-enabled AI development, however, suggests that self-transforming agents would play no natural role, even in the limiting case of explosively-fast progress in AI technology. Because the mechanisms underlying a potential intelligence explosion are already in operation, and have no necessary connection to unprecedented AGI agents, paths to extremely rapid progress in AI technology may be both more direct and more controllable than has been commonly assumed. \n AI-enabled AI development could lead to an intelligence explosion As suggested by  Good (1966) , AI systems could potentially outperform human beings in the task of AI development, and hence \"could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind.\" \n Risk models have envisioned AGI agents driving an intelligence explosion system that engages with the world as an agent 1 has motivated a threat model in which \"the machine\" might not be \"docile enough to tell us how to keep it under control\"  (Good 1966, p.33) . The surprisingly complex and difficult ramifications of this threat model have been extensively explored in recent years (e.g., in Bostrom 2014). \n Self-transforming AI agents have no natural role in recursive improvement Advances in AI technology emerge from research and development, 2 a process that comprises a range of different technical tasks. These tasks are loosely coupled, and none requires universal competence: Consider, for example, the technology-centered tasks of training-algorithm development, benchmark development, architecture search, and chip design; and beyond these, application development tasks that include human consultation 3 and application prototyping, together with testing, monitoring, and customer service.  4  Accordingly, general-purpose, self-transforming AI agents play no natural role in the process of AI R&D: They are potential (and potentially dangerous) products-not components-of AI development systems. To the extent that the concept of \"docility\" may be relevant to development systems as a whole, this desirable property is also, by default, deeply ingrained in the nature of the services they provide. 5 \n The direct path to an intelligence explosion does not rely on AGI agents It may be tempting to imagine that self -improvement would be simpler than loosely-coupled systemic improvement, but drawing a conceptual boundary around a system does not simplify its contents, and to require that systems capable of open-ended AI development also exhibit tight integration, functional autonomy, and operational agency would increase-not reduce- implementation challenges.  1  To attempt to place competitive R&D capabilities inside an agent presents difficulties, yet provides no compensating advantages in performance or utility.  2  The pervasive assumption that an intelligence explosion must await the development of agents capable of autonomous, open-ended self improvement has encouraged skepticism and complacency 3 regarding prospects for superintelligent-level AI.  4  If we think that any given set of human tasks can be automated, however, than so can any-and in the limit, all-AI R&D tasks. This proposition seems relatively uncontroversial, yet has profound implications for the likely trajectory of AI development. AI-enabled automation of fundamental AI R&D tasks is markedly accelerating.  5  As the range of automated tasks increases, we can expect feedback loops to tighten, enabling AI development at a pace that, while readily controlled, has no obvious limits.  \n Further Reading \n Summary Since  Good (1966)  2.2 Superintelligence has been defined in terms of adult human competence  Good (1966)  defined \"ultraintelligence\" in terms of distinct, highly competent entities: Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. The standard definition of \"superintelligence\" today  (Bostrom 2014 ) parallels  Good (1966) , yet to define superintelligence in terms of adult intellectual competence fails to capture what we mean by human intelligence. \n \"Intelligence\" often refers instead to learning capacity Consider what we mean when we call a person intelligent: • A child is considered \"intelligent\" because of learning capacity, not competence. • An expert is considered \"intelligent\" because of competence, not learning capacity. We can overlook this distinction in the human world because learning and competence are deeply intertwined in human intellectual activities; in considering prospects for AI, by contrast, regarding \"intelligence\" as entailing both learning and competence invites deep misconceptions. \n Learning and competence are separable in principle and practice A human expert in science, engineering, or AI research might provide brilliant solutions to problems, and even if a drug or neurological defect blocked the expert's formation of long-term memories, we would recognize the expert's intelligence. Thus, even in humans, competence can in principle be dissociated from ongoing learning, while in AI technology, this separation is simply standard practice. Regardless of implementation technologies, each released version of an AI system can be a fixed, stable software object. Reinforcement learning agents illustrate the separation between learning and competence: Reinforcement \"rewards\" are signals that shape learned behavior, yet play no role in performance. Trained RL agents exercise their competencies without receiving reward.  2.6 Distinguishing learning from competence is crucial to understanding potential AI control strategies Ensuring relatively predictable, constrained behavior is fundamental to AI control. The tacit assumption that the exercise of competence entails learning implies that an intelligent system must be mutable, which is to say, potentially unstable. Further, the tacit assumption that intelligence entails both learning and competence invites the misconception that AI systems capable of learning will necessarily have complex states and capabilities that might be poorly understood. As this might suggest, conflating intelligence, competence, and learning obscures much of the potential solution space for of AI control problems (e.g., approaches to architecting trustworthy composite oracles).  2  This problematic, subtly anthropomorphic model of intelligence is deeply embedded in current discussion of AI prospects. If we are to think clearly about AI control problems, then even comprehensive, superintelligent-level AI capabilities 3 must not be equated with \"superintelligence\" as usually envisioned.  \n Further Reading \n Summary From an instrumental perspective, intelligence centers on capabilities rather than systems, yet models of advanced AI commonly treat the key capability of general intelligence-the ability to develop novel capabilities-as a blackbox mechanism embedded in a particular kind of system, an AGI agent. Service-oriented models of general intelligence instead highlight the service of developing new services as a computational action, and place differentiated task functionality (rather than unitary, general-purpose systems) at the center of analysis, linking models to the richly-developed conceptual framework of software engineering. Service-oriented models reframe the problem of aligning AI functionality with human goals, providing affordances absent from opaque-agent models. \n The instrumental function of AI technologies is to provide services In the present context, \"services\" are tasks performed to serve a client. AI systems may provide services to humans more-or-less directly (driving, designing, planning, conversing. . . ), but in a software engineering context, one may also refer to clients and service-providers (servers) when both are computational processes. Not every action performs a service. As with human intelligence, intelligence embodied in autonomous systems might not serve the instrumental goals of any external client; in considering actions of the system itself, the concept of goals and utility functions would then be more appropriate than the concept of services. \n General intelligence is equivalent to general capability development General intelligence requires an open-ended ability to develop novel capabilities, and hence to perform novel tasks. For instrumental AI, capabilities correspond to potential services, and general artificial intelligence can be regarded as an open-ended service able to provide new AI services by developing new capabilities. \n The ability to learn is a capability In humans (with their opaque skulls and brains), it is natural to distinguish internal capabilities (e.g., learning to program) from external capabilities (e.g., programming a machine), and to treat these as different in kind. In computational systems, however, there need be no such distinction: To develop a capability is to implement a system that provides that capability, a process that need not modify the system that performs the implementation. Removing artificial distinctions between kinds of capabilities improves the scope, generality, and homogeneity of models of artificial general intelligence. \n Implementing new capabilities does not require \"self modification\" An incremental computational process may extend the capabilities of a computational system. If the resulting code automatically replaces the previous version, and if it is convenient to regard that process as internal to the system, then it may be natural to call the process \"self modification\". In many practical applications, however, a different approach will produce better results: Data can be aggregated from many instances of a system, then combined through a centralized, perhaps computationally-intensive development service to produce upgraded systems that are then tested and deployed. The strong advantages of data aggregation and centralized development  1  suggest that it would be a mistake to adopt \"self modification\" as a default model of system improvement. \n Service-centered models highlight differentiated, task-focused functionality Services have structure: Across both human and computational worlds, we find that high-level services are provided by employing and coordinating lower-level services. We can expect that services of different kinds (e.g., language translation, theorem proving, aircraft design, computer hacking, computer security, military strategy) will or readily could be developed and implemented as substantially distinct computational systems, each operating not only as a server, but as a client that itself employs a range of narrower services. It is safe to assert that the architecture of complex services and service-providers will be neither atomized nor monolithic. \n Service-centered models harmonize with practice in software engineering AI services are being developed and deployed in the context of other software services. Decades of research, billions of dollars, and enormous intellectual effort have been invested in organizing the development of increasingly complex systems, and universal patterns have emerged; in particular, system architectures are defined by their functions and interfaces-by service provided and means of employing them. The art of decomposing high-level functions into lower-level functions has been essential to making systems comprehensible, implementable, and maintainable. Modern software services are both the technological milieu of modern AI and a model for how complex information services emerge and evolve. Perhaps the most fundamental principles are modularity and abstraction: partitioning functionality and then decoupling functionality from implementation. We can expect that these extraordinarily general abstractions can and will scale to systems implemented by (and to provide) superintelligent-level services. \n Service-centered AI architectures can facilitate AI alignment A system-centric model would suggest that general-purpose artificial intelligence must be a property of general-purpose AI systems, and that a fullygeneral AI system, to perform its functions, must be a powerful superintelligent agent. From this model and its conclusion, profound challenges follow. A service-centric model, by contrast, proposes to satisfy general demands for intelligent services through a general capacity to develop services. This general capacity is not itself a thing or an agent, but a pool of functionality that can be provided by coordination of AI-development services. In this model, even highly-capable agents implemented at a superintelligent level 1 can be stable, and need not themselves embody AI-development functionality. This model suggests that a range of profound challenges, if recognized, can also be avoided. Diverse, potentially superintelligent-level AI services could be coordinated to provide the service of developing new AI services. Potential components and functions include: • Predictive models of human approval, disapproval, and controversies.  2  • Consulting services that propose and discuss potential products and services. 3 • Design, 4 implementation, 5 and optimization services.  6  • Specialists in technical security and safety measures. 7 . • Evaluation through criticism and red-team/blue-team competitions. 8 • Pre-deployment testing and post-deployment assessment. 9 • Iterative, experience-based upgrades to products and services.  10  Each of the above corresponds to one or more high-level services that would typically rely on others, whether these are narrower (e.g., language understanding and technical domain knowledge) or at a comparable level (e.g., predictive models of human (dis)approval). Some services (e.g., criticism and red-team/blue-team competitions) by nature interact with others that are adversarial and operationally distinct. Taken together, these services suggest a range of alignment-relevant affordances that are (to say the least) not salient in models that treat general intelligence as a black-box mechanism that is embedded in a general-purpose agent. Further Reading The AI-services model both describes the architecture of current and prospective high-level AI applications, and prescribes patterns of development that can foster safety without impeding the speed and efficiency of AI development. \n Summary Does the AI-services model describe prospects for high-level AI application development, or prescribe strategies for avoiding classic AGI-agent risks? Description and prescription are closely aligned: The services model describes current and emerging patterns of AI application development, notes these patterns are accessible, scalable, and align with AI safety, and accordingly prescribes deliberate adherence to these patterns. The alignment between descriptive and prescriptive aspects of the services model is fortunate, because strategies for AI safety will be more readily adopted if they align with, rather than impede, the momentum of AI development. 4.  \n AI-service development scales to comprehensive, SI-level services The \"AI services\" concept scales to sets of services that perform an asymptotically-comprehensive range of tasks, while AI-supported automation of AI R&D automation scales to asymptotically-recursive, potentially swift technology improvement. Because systems based on AI services (including service-development services) scale to a superintelligent level 3 , the potential scope of AI services subsumes the instrumental functionality 4 that might otherwise motivate the development of AGI agents. 5 \n Adherence to the AI-services model aligns with AI safety Because the AI-services model naturally employs diversity, competition, and adversarial goals 6 (e.g., proposers vs. critics) among service-providers, architectures that adhere to the (extraordinarily flexible) AI-services model can readily avoid classic risks associated with superintelligent, self-modifying, utility-maximizing agents. 7 4.5 Adherence to the AI-services model seems desirable, natural, 1. Section 10: R&D automation dissociates recursive improvement from AI agency. 2. Section 3: To understand AI prospects, focus on services, not implementations. 3. Section 1: R&D automation provides the most direct path to an intelligence explosion. 4. Section 12: AGI agents offer no compelling value. 5. Section 11: Potential AGI-enabling technologies also enable comprehensive AI services. 6. Section 20: Collusion among superintelligent oracles can readily be avoided. 7. Section 6: A system of AI services is not equivalent to a utility maximizing agent \n and practical There need be no technological discontinuities on the way to thorough AI R&D automation and comprehensive AI services, and continued adherence to this model is compatible with efficient development and application of AI capabilities.  1  Traditional models of general AI capabilities, centered on AGI agents, seem more difficult to implement, more risky, and no more valuable in application. Accordingly, guidelines that prescribe adherence to the AI-services model 2 could improve 3 prospects for a safe path to superintelligent-level AI without seeking to impede the momentum of competitive AI development. Further Reading It is a mistake to frame intelligence as a property of mind-like systems, whether these systems are overtly anthropomorphic or abstracted into decision-making processes that guide rational agents. \n Summary Concepts of artificial intelligence have long been tied to concepts of mind, and even abstract, rational-agent models of intelligent systems are built on psychomorphic and recognizably anthropomorphic foundations. Emerging AI technologies do not fit a psychomorphic frame, and are radically unlike evolved intelligent systems, yet technical analysis of prospective AI systems has routinely adopted assumptions with recognizably biological characteristics. To understand prospects for AI applications and safety, we must consider not only psychomorphic and rational-agent models, but also a wide range of intelligent systems that present strongly contrasting characteristics. \n The concept of mind has framed our concept of intelligence Minds evolved to guide organisms through life, and natural intelligence evolved to make minds more effective. Because the only high-level intelligence we know is an aspect of human minds, it is natural for our concept of mind to frame our concept of intelligence. Indeed, popular culture has envisioned advanced AI systems as artificial minds that are by default much like our own. AI-as-mind has powerful intuitive appeal. The concept of AI-as-mind is deeply embedded in current discourse. For example, in cautioning against anthropomorphizing superintelligent AI,  Bostrom (2014, p.105)  urges us to \"reflect for a moment on the vastness of the space of possible minds\", an abstract space in which \"human minds form a tiny cluster\". To understand prospects for superintelligence, however, we must consider a broader space of potential intelligent systems, a space in which mind-like systems themselves form a tiny cluster. \n Studies of advanced AI often posit intelligence in a psychomorphic role Technical studies of AI control set aside explicitly human psychological concepts by modeling AI systems as goal-seeking rational agents. Despite their profound abstraction, however, rational-agent models originated as idealizations of human decision-making, and hence place intelligence in an implicitly anthropomorphic frame. More concretely, in rational-agent models, the content of human minds (human values, goals, cognitive limits. . . ) is abstracted away, yet the role of minds in guiding decisions is retained. An agent's decisionmaking process fills an inherently mind-shaped slot, and that slot frames a recognizably psychomorphic concept of intelligence. This is problematic: Although the rational-agent model is broad, it is still too narrow to serve as a general model of intelligent systems. \n Intelligent systems need not be psychomorphic What would count as high-level yet non-psychomorphic intelligence? One would be inclined to say that we have general, high-level AI if a coordinated pool of AI resources could, in aggregate: • Do theoretical physics and biomedical research   \n Engineering and biological evolution differ profoundly Rather than regarding artificial intelligence as something that fills a mindshaped slot, we can instead consider AI systems as products of increasinglyautomated technology development, an extension of the R&D process that we see in the world today. This development-oriented perspective on AI technologies highlights profound and pervasive differences between evolved and engineered intelligent systems (see table  ) . \n Studies of AI prospects have often made tacitly biological assumptions Although technical models of artificial intelligence avoid overtly biological assumptions, it is nonetheless common (though far from universal!) to assume that advanced AI systems will: • Exist as individuals, rather than as systems of coordinated components 1 • Learn from individual experience, rather than from aggregated training data 2 • Develop through self-modification, rather than being constructed 3 and updated 4 • Exist continuously, rather than being instantiated on demand 5 • Pursue world-oriented goals, rather than performing specific tasks 6 These assumptions have recognizable biological affinities, and they invite further assumptions that are tacitly biomorphic, psychomorphic, and even anthropomorphic. 5.7 Potential mind-like systems are situated in a more general space of potential intelligent systems It has been persuasively argued that rational, mind-like superintelligence is an attractor in the space of potential AI systems, whether by design or inadvertent emergence. A crucial question, however, is the extent of the basin of attraction for mind-like systems within the far more general space of potential AI systems. The discussion above suggests that this basin is far from coextensive with the space of highly-capable AI systems, including systems that can, in aggregate, provide superintelligent-level services across an indefinitely wide range of tasks.  1  We cannot chart the space of potential AI problems and solutions solely within the confines of rational-agent models, because most of that space lies outside. Further Reading 6.2 Systems of SI-level agents have been assumed to act as a single agent In informal discussions of AI safety, it been widely assumed that, when considering a system comprising rational, utility-maximizing AI agents, one can (or should, or even must) model them as a single, emergent agent. This assumption is mistaken, and worse, impedes discussion of a range of potentially crucial AI safety strategies. To understand how we could employ and manage systems of rational agents, we can (without loss of generality) start by considering individual systems (\"service providers\") that act as rational agents. \n Individual service providers can be modeled as individual agents The von Neumann-Morgenstern expected-utility theorem shows that, if an agent meets a set of reasonable conditions defining rational behavior, the agent must choose actions that maximize the expected value of some function that assigns numerical values (utilities) to potential outcomes. If we consider a system that provides a service to be \"an agent\", then it is at least reasonable to regard VNM rationality as a condition for optimality. \n Trivial agents can readily satisfy the conditions for VNM rationality To manifestly violate the conditions for VNM rationality, an agent must make choices that are incompatible with any possible utility function. Accordingly, VNM rationality can be a trivial constraint: It is compatible, for example, with a null agent (that takes no actions), with an indifferent agent (that values all outcomes equally), and with any agent that acts only once (and hence cannot exhibit inconsistent preferences). The archetypical predictive model acts as a fixed function (e.g., a translator of type T :: string → string). In each instance above, greater knowledge and reasoning capacity can improve performance; and in each instance, \"actions\" (function applications) may be judged by their external instrumental value, but cannot themselves violate the conditions for VNM rationality. As suggested by the examples above, fixed predictive models can encapsulate intelligence for active applications: For example, a system might drive a car to a user-selected destination while employing SI-level resources that inform steering decisions by predicting human (dis)approval of predicted outcomes. 6.6 High intelligence does not imply optimization of broad utility functions  Bostrom's (2014, p.107)  Orthogonality Thesis states that \"more or less any level of intelligence can be combined with more or less any final goal\", and it follows that high-level intelligence can be applied tasks of bounded scope and duration that do not engender convergent instrumental goals. 7 Note that service-providers need not seek to expand their capabilities: This is a task for service developers, while the service-developers are themselves providers of service-development services (the implied regress is loopy, not infinite). \n Systems composed of rational agents need not maximize a utility function There is no canonical way to aggregate utilities over agents, and game theory shows that interacting sets of rational agents need not achieve even Pareto optimality. Agents can compete to perform a task, or can perform adversarial tasks such as proposing and criticizing actions; 1 from an external client's perspective, these uncooperative interactions are features, not bugs (consider the growing utility of generative adversarial networks 2 ). Further, adaptive collusion can be cleanly avoided: Fixed functions, for example, cannot negotiate or adapt their behavior to align with another agent's purpose. In light of these considerations, it would seem strange to think that sets of AI services (even SI-level services) would necessarily or naturally collapse into utility-maximizing AI agents. \n Multi-agent systems are structurally inequivalent to single agents There is, of course, an even more fundamental objection to drawing a boundary around a set of agents and treating them as a single entity: In interacting with a set of agents, one can choose to communicate with one or another (e.g., with an agent or its competitor); if we assume that the agents are in effect a single entity, we are assuming a constraint on communication that does not exist in the multi-agent model. The models are fundamentally, structurally inequivalent. 6.9 Problematic AI services need not be problematic AGI agents Because AI services can in principle be fully general, combinations of services could of course be used to implement complex agent behaviors up to and including those of unitary AGI systems. Further, because evolutionary pressures can engender the emergence of powerful agents from lesser cognitive systems (e.g., our evolutionary ancestors), the unintended emergence of problematic agent behaviors in AI systems must be a real concern.  1  Unintended, perverse interactions among some sets of service-providers will likely be a routine occurrence, familiar to contemporaneous researchers as a result of ongoing experience and AI safety studies.  2  Along this development path, the implied threat model is quite unlike that of a naïve humanity abruptly confronting a powerful, world-transforming AGI agent. \n 6.10 The AI-services model expands the solution-space for addressing AI risks The AI-services and AGI-agent models of superintelligence are far from equivalent, and the AI-services model offers a wider range of affordances for structuring AI systems. Distinct, stable predictive models of human approval, 3 together with natural applications of competing and adversarial AI services, 4 can provide powerful tools for addressing traditional AI safety concerns. If, as seems likely, much of the potential solution-space for addressing AI x-risk 5 requires affordances within the AI-services model, then we must reconsider long-standing assumptions regarding the dominant role of utility-maximizing agents, and expand the AI-safety research portfolio to embrace new lines of inquiry. Further Reading An examination of the distinction between human-like skills and humanlike goal structures reduces the this apparent conflict. Critical differences emerge through the distinction between intelligence as learning and intelligence as competence, 1 , the power of development architectures based on aggregated experience and centralized learning, 1 and the role of application architectures in which tasks are naturally bounded.  2  7.3 Human-like learning may be essential to developing general intelligence General intelligence, in the sense of general competence, obviously includes human-like world-oriented knowledge and skills, and it would be surprising if these could be gained without human-like world-oriented learning. The practical value of human-like abilities is a major driving force behind AI research. A more interesting question is whether human-like learning from humanlike experience is essential to the development of general intelligence in the sense of general learning ability. This is a plausible hypothesis: Human intelligence is the only known example of what we consider to be general intelligence, and it emerged from open-ended interaction of humans with the world over the time spans of genetic, cultural, and individual development. Some researchers aim to reproduce this success by imitation. Beyond the practical value of human-like learning, and the hypothesis that is may be essential to the development of general intelligence, there is a third reason to expect AI research to continue in this direction: Since  Turing (1950) , AI researchers have defined their objectives in terms of matching human capabilities in a broad sense, and have looked toward human-like learning, starting with child-like experience, as a natural path toward this goal. \n Current methods build curious, imaginative agents In pursuit of AI systems that can guide agents in complex worlds (to date, usually simulated), researchers have developed algorithms that build abstracted models of the world and use these as a basis for \"imagination\" to enable planning, reducing the costs of learning from trial and error  (Weber et al. 2017; Nair et al. 2018; Ha and Schmidhuber 2018; Wayne et al. 2018 ). Reinforcement-learning agents have difficulty learning complex sequences of actions guided only by end-state goals; an effective approach has been development of algorithms that have \"curiosity\", seeking novel experiences and exploring actions of kinds that might be relevant to any of a range of potential goals  (Pathak et al. 2017; Burda et al. 2018) . Search guided by curiosity and imagination can adapt to novel situations and find unexpected solutions. Agents can also learn by observation of the real world, whether guided by demonstrations offered by human beings, or by imitating actions in relevant videos downloaded from the internet  (Duan et al. 2017; Peng et al. 2018) . \n Human-like competencies do not imply human-like goal structures Human beings learn human goal structures, but full human goal structureslife goals, for example-do not emerge directly or naturally from applying curiosity, imagination, and imitation to learning even an unbounded range of bounded tasks. The idea that human-like problem-solving is tightly linked to human-like goals may stem from what are tacitly biological background assumptions, 1 including evolution under competition, learning from individual experience, unitary functionality, and even physical continuity. The abstraction of AI systems as rational utility-directed agents also proves, on examination, to draw on tacitly anthropomorphic assumptions.  2  Even speaking of \"agents learning\", as does the preceding section, is subject to this criticism. Humans learn by acting, but standard \"reinforcement learning\" methods sever this link: \"Reinforcement learning\" means training by a reinforcement learning algorithm, yet this algorithm performs no actions, while a trained agent learns nothing by acting.  3  Learning and action can of course be fused in a single system, yet they need not be, and learning can be more effective when they are separate.  4  Again, it is a mistake to conflate intelligence as learning capacity with intelligence as competence.  5  1. Section ??: ?? 2. Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame 3. Section 18: Reinforcement learning systems are not equivalent to reward-seeking agents 4. Section 16: Aggregated experience and centralized learning support AI-agent applications 5. Section 2: Standard definitions of \"superintelligence\" conflate learning with competence 7.6 Open-ended learning can develop skills applicable to bounded tasks In light of these considerations, it is natural and not necessarily problematic to pursue systems with general, human-like learning abilities as well as systems with collectively-general human-like competencies. Open-ended learning processes and general competencies do not imply problematic goals-and in particular, do not necessarily engender convergent instrumental goals.  1  Even if the development of general intelligence though open-ended learning-to-learn entailed the development of opaque, unitary, problematic agents, their capabilities could be applied to the development of compact systems 2 that retain general learning capabilities while lacking the kinds of information, competencies, and goals that fit the profile of a dangerous AGI agent. Note that extracting and refining skills from a trained system can be a less challenging development task than learning equivalent skills from experience alone. \n Human-like world-oriented learning nonetheless brings unique risks These considerations suggest that there need be no direct line from humanlike competencies to problematic goals, yet some human-like competencies are more problematic than, for example, expertise tightly focused on theorem proving or engineering design. Flexible, adaptive action in the physical world can enable disruptive competition in arenas that range from industrial parks to the battlefields. Flexible, adaptive action in the world of human information can support interventions that range from political manipulation to sheer absorption of human attention; the training objectives of the chatbot XiaoIce, for example, include engaging humans in supportive emotional relationships while maximizing the length of conversational exchanges  (Zhou et al. 2018) . She does this very well, and is continuing to learn. Adherence to the AI services model by no means guarantees benign behaviors or favorable world outcomes, 3 even when applied with good intentions. Because of their potential for direct engagement with the world, however, human-like learning and capabilities present a special range of risks.  \n Strong optimization is a strong constraint Full optimization of a system with respect to a value function typically yields not only a unique, maximal expected value, but robust constraints on the system itself. In practice, even approximate optimization strongly constrains both the structure and behavior of a system, thereby constraining its capabilities. In physical engineering, a race car cannot transport a load of passengers, and a bus will never set a land speed record; in AI development, an efficient text-translation system will never plan a vehicle path, and an efficient pathplanning system will never provide translation services. Because off-task capabilities are costly, they will be excluded by cost-sensitive optimization.  \n Optimization of AI systems can reduce unintended consequences Optimization will tend to reduce risks when task objectives are bounded in space, time, and scope, and when the value-function assigns costs to both resource use and unintended human-relevant effects. With bounded objectives, remote and long-term effects will contribute no value and hence will be unintended, not actively optimized, and likewise for off-task consequences that are local and near-term. Going further, when costs are assessed for resource consumption and unintended, human-relevant effects, 1 stronger optimization will tend to actively reduce unwanted consequences (see  Armstrong and Levinstein [2017]  for a concept of reduced-impact AI). \n Strong external optimization can strongly constrain internal capabilities If the costs of appropriate computational resources are given substantial weight, strong optimization of an AI system can act as a strong constraint on its inputs and model capacity, and on the scope of its mechanisms for off-task inference, modeling, and planning. The weighting of computationalcost components need not reflect external economic costs, but can instead be chosen to shape the system under development. A major concern regarding strong AI capabilities is the potential for poorlydefined goals and powerful optimization to lead to perverse plans that (for example) act through unexpected mechanisms with surprising side-effects. Optimization to minimize a system's off-task information, inference, modeling, and planning, however, can constrain the scope for formulating perverse plans, because the planning itself may incur substantial costs or require resources that have been omitted in the interest of efficiency. Optimization for on-task capabilities can thereby avoid a range of risks that have never been considered. 8.6 Optimizing an AI system for a bounded task is itself a bounded task Optimizing a system for a bounded AI task can itself be framed as a bounded AI task. The task of designing and optimizing an AI system for a given task using given machine resources within an acceptably short time does not entail open-ended, world-affecting activity, and likewise for designing and optimizing AI systems for the tasks involved in designing and optimizing AI systems for the tasks of designing and optimizing AI systems. \n Superintelligent-level optimization can contribute to AI safety Given that optimization can be used to shape AI systems that perform a range of tasks with little or no catastrophic risk, it may be useful to seek tasks that, in composition with systems that perform other tasks, directly reduce the risks of employing systems with powerful capabilities. A leading example is the development of predictive models of human relevance and human approval based on large corpora of human opinions and crowd-sourced challenges. Ideally, such models would have access to general world knowledge and be able to engage in general reasoning about cause, effect, and the range of potential human reactions. Predictive models of human approval 1 would be useful for augmenting human oversight, 2 flagging potential concerns, 3 and constraining the actions of systems with different capabilities.  4  These and similar applications are attractive targets for shaping AI outcome through differential technology development. Bad actors could of course apply strongly optimized AI technologies-even approval modeling-to bad or risky ends (e.g., open-ended exploitation of internet access for wealth maximization). Bad actors and bad actions are a crucial concern in considering strategies for managing a safe transition to a world with superintelligent-level AI, yet effective countermeasures may themselves require strong, safe optimization of AI systems for strategically important tasks. The development of strong optimization power is a given, and we should not shy away from considering how strongly optimized AI systems might be used to solve problems. Although transparency is desirable, opacity at the level of algorithms and representations need not greatly impair understanding of AI systems at higher levels of functionality. \n Further Reading \n Summary Deep-learning methods employ opaque algorithms that operate on opaque representations, and it would be unwise to assume pervasive transparency in future AI systems of any kind. Fortunately, opacity at the level of algorithms and representations is compatible with transparency at higher levels of system functionality. We can shape information inputs and training objectives at component boundaries, and can, if we choose, also shape and monitor information flows among opaque components in larger systems. Thus, substantial highlevel understanding and control is compatible with relaxed understanding of internal algorithms and representations. As always, the actual application of potential control measures can be responsive to future experience and circumstances. \n Deep-learning methods are opaque and may remain so The products of deep learning are notoriously opaque: Numerical transformations produce numerical vectors that can be decoded into useful results, but the encodings themselves are often incomprehensible. Opacity is the norm, and interpretability is the exception. In considering problems of AI control and safety, it would be unwise to assume pervasive transparency. \n The scope of information and competencies can be fuzzy, yet bounded Although we may lack knowledge of how a deep learning system represents information and algorithms, we can have substantial knowledge of the scope of its information and competencies. For example, information that is absent from a system's inputs (in both training and use) will be absent from its algorithms, state, and outputs. Inference capabilities may blur the scope of given information, but only within limits: A Wikipedia article cannot be inferred from language-free knowledge of physics. Likewise, while the scope of a system's competencies may be fuzzy, competencies far from a system's task focus (e.g., theorem-proving competencies in a vision system, or vehicle-guidance competencies in a language-translation system) will be reliably absent. Bounds on information and competencies are natural and inevitable, and can be applied to help us understand and constrain AI-system functionality. \n Restricting resources and information at boundaries constrains capabilities Several obvious affordances for control are available at the boundaries of AI systems. For example, tasks that require absent information cannot be performed, and the distinct role of physical memory in digital systems enables a clean separation of episodic task performance from cumulative learning.  1  Controls at boundaries have transitive effects within systems of collaborating components: A component cannot transfer information that it does not have, regardless of how internal communications are encoded. Further, competitive systems must deliver results in bounded times and with bounded resources. Optimization pressures (e.g., on model capacity, training time, and execution cost) will tend to exclude investments in off-task capacities and activities, and stronger optimization will tend to strengthen, not weaken, those constraints.  2  A system trained to provide services to other systems might perform unknown tasks, yet those tasks will not be both costly and irrelevant to external objectives. These considerations are fundamental: They apply regardless of whether an AI system is implemented on digital, analog, or quantum computational hardware, and regardless of whether its algorithms are neural and trained, or symbolic and programmed. They scale to task domains of any scope, and to systems of any level of intelligence and competence. \n Providing external capabilities can constrain internal capabilities Under appropriate optimization pressures, a system trained with access to an efficient resource with particular capabilities 1 will not itself develop equivalent capabilities, and use of those particular capabilities will then involve use of an identifiable resource. This mechanism provides an affordance for shaping the organization of task-relevant capabilities in the development of piecewise-opaque systems. Potential advantages include not only functional transparency, but opportunities to ensure that components (vision systems, physical models, etc.) are well-trained, well-tested, and capable of good generalization within their domains. \n Deep learning can help interpret internal representations Deep learning techniques can sometimes provide insight into the content of opaque, learned representations. To monitor the presence or absence of a particular kind of information in an opaque (but not adversarially opaque) representation, deep learning can be applied to attempt to extract and apply that information. For example, a representation may be opaque to humans, but if it supports an image-recognition task, then the representation must contain image information; if not, then it likely doesn't. 9.7 Task-space models can enable a kind of \"mind reading\" The task-space model 2 of general intelligence suggests that the subtasks engaged by problem-solving activities can (both in principle and in practice) be associated with regions in semantic spaces. Different high-level tasks will generate different footprints of activity in the space of subtasks, and one need not understand how every subtask is represented or performed to understand what the task is about. Restricting the range of task-space accessible to a system could potentially provide a mechanism for constraining its actions, while observing access patterns could potentially provide the ability to monitor the considerations that go into a particular action. For example, it would be unremarkable for a system that organizes food production to access services applicable to food preparation and delivery, while a system that accesses services applicable to synthesizing neurotoxins for delivery in food might trigger a warning. Such insights (a kind of \"mind reading\") could be useful both in advancing AI safety and in solving more prosaic problems of system development and debugging. \n The application of control measures can be adapted to experience and circumstances Whether any particular set of control measures should be applied, and to what extent, is a question best answered by the AI community as circumstances arise. Experience will provide considerable knowledge 1 of which kinds of systems are reliable, which fail, and which produce surprising (perhaps disturbingly surprising) results. Along the way, conventional concerns regarding safety and reliability will drive efforts to make systems better understood and more predictable. To catalog a range of potential control measures (e.g., attention to information content, task focus, and optimization pressures) is not to assert that any particular measure or intensity of application will be necessary or sufficient. The value of inquiry in this area is to explore mechanisms that could be applied in response to future experience and circumstances, and that may deserve attention today as safety-relevant components of general AI research.  \n Further Reading \n AI R&D automation will reflect universal aspects of R&D processes The R&D process links the discovery of general principles and mechanisms to the construction of complex systems tailored to specific functions and circumstances. The R&D-automation model describes AI development today, and increasing automation of AI development seems unlikely to obliterate this deep and universal task structure. 10.4 AI R&D automation will reflect the structure of AI development tasks In AI R&D, the elements of development tasks are organized along lines suggested by the following diagram:  \n AI R&D automation leads toward recursive technology improvement In the schematic diagram above, an AI builder is itself an R&D product, as are AI systems that perform exploratory AI research. Pervasive and asymptotically complete AI-enabled automation of AI R&D can enable what amounts to recursive improvement, raising the yield of AI progress per unit of human effort without obvious bound. In this model there is no locus of activity that corresponds to recursive \"self\" improvement; as we see in today's AI R&D community, loosely coupled activities are sufficient to advance all aspects of AI technology. \n General SI-level functionality does not require general SI-level agents The classic motivation for building self-improving general-purpose superintelligent agents is to provide systems that can perform a full range of tasks with superintelligent competence. The R&D-automation model, however, shows how to provide, on demand, systems that can perform any of a fully general range of tasks without invoking the services of a fully general agent. \n 10.7 The R&D-automation model reframes the role of AI safety studies and offers potential affordances for addressing AI safety problems In the present framework, agent-oriented AI safety research plays the dual roles of expanding the scope of safe agent functionality and identifying classes of systems and applications (including tightly coupled configurations of R&D components) in which radically unsafe agent behaviors might arise unintentionally. In other words, agent-oriented safety work can both find safe paths and mark potential hazards. The R&D-automation model describes component-based systems that are well-suited to the production of component-based systems, hence it invites consideration of potential safety-relevant affordances of deeply-structured AI implementations. In particular, the prospect of safe access to superintelligent machine learning invites consideration of predictive models of human approval that are trained on large corpora of human responses to events and actions, and subsequently serve as components of structured, approvaldirected AI agents.  \n Summary In catastrophic runaway-AI scenarios, systems capable of self-improvement lead to-and hence precede-opaque AGI agents with general superhuman competencies. Systems capable of self-improvement would, however, embody high-level development capabilities that could first be exploited to upgrade ongoing, relatively transparent AI R&D automation. Along this path, transparency and control need not impede AI development, and optimization pressures can sharpen task focus rather than loosen constraints. Thus, in scenarios where advances in technology would enable the implementation of powerful but risky AGI agents, those same advances could instead be applied to provide comprehensive AI services-and stable, task-focused agents-while avoiding the potential risks of self-modifying AGI-agents. \n In runaway-AGI scenarios, self-improvement precedes risky competencies Self-improving, general-purpose AI systems would, by definition, have the ability to build AI systems with capabilities applicable to AI development tasks. In classic AGI-takeover scenarios, the specific competencies that enable algorithmic self-improvement would precede more general superhuman competencies in modeling the world, defining world-changing goals, and pursuing (for example) a workable plan to seize control. Strong AI development capabilities would precede potential catastrophic threats. \n \"Self\"-improvement mechanisms would first accelerate R&D In any realistic development scenario, highly-capable systems will follow less-capable predecessors (e.g., systems with weaker architectures, smaller datasets, or less training), and developers will have practical knowledge of how to instantiate, train, and apply these systems. Along paths that lead to systems able to implement more capable systems with little human effort, 1 it will be natural for developers to apply those systems to specific development tasks. Developing and packaging an opaque, self-improving AI system might or might not be among those tasks. \n \"Self\"-improvement mechanisms have no special connection to agents The option to develop and apply self-improving AGI systems would be compelling only if there were no comparable or superior alternatives. Given AI systems able to implement a wide range of AI systems, however, there would be no compelling reason to package and seal AI development processes in an opaque box. Quite the opposite: Practical considerations generally favor development, testing, and integration of differentiated components.  2  Potential AGI-level technologies could presumably automate such processes, while an open system-development architecture would retain system-level transparency and process control. \n Transparency and control need not impede the pace of AI development Because the outputs of basic research-the building blocks of technology improvement-need not directly affect the world, the need for human intervention in basic research is minimal and need not impede progress. 3 Meanwhile, in applications, guiding development to serve human purposes is not a 1. Section 10: R&D automation dissociates recursive improvement from AI agency 2. Section 15: Development-oriented models align with deeply-structured AI systems burden, but an inherent part of the task of providing beneficial AI services.  1  Note that developing and applying AI systems to help humans guide development is itself a task within the scope of comprehensive AI R&D automation. 2 \n Optimization pressures sharpen task focus Thinking about prospects of applying high-level AI capabilities to the design and optimization of AI systems has been muddied by the tacit assumption that optimizing performance implies relaxing constraints on behavior. For any bounded task, however, this is exactly wrong: The stronger the optimization, the stronger the constraints.  3  In the context of AI-enabled AI-development, effective optimization of a development system will tend to minimize resources spent on off-task modeling and search. Regardless of their internal complexity, optimized components of AI development systems will have no spare time to daydream about world domination. \n Problematic emergent behaviors differ from classic AGI risks Systems of optimized, stable components can be used to implement fully general mechanisms, hence some configurations of components could exhibit problematic emergent behaviors of unexpected kinds. We can expect and encourage developers to note and avoid architectures of the kind that produce unexpected behaviors, 4 perhaps aided by AI-enabled analysis of both AI objectives, 5 and proposed implementations.  6  Avoiding problematic emergent behaviors in task-oriented systems composed of stable components is inherently more tractable than attempting to confine or control a self-modifying AGI system that might by default act as superintelligent adversarial agent.  \n AGI agents offer no compelling value Because general AI-development capabilities can provide stable, comprehensive AI services, there is no compelling, practical motivation for undertaking the more difficult and potentially risky implementation of self-modifying AGI agents. \n Summary Practical incentives for developing AGI agents appear surprisingly weak. Providing comprehensive AI services calls for diverse, open-ended AI capabilities (including stable agent services), but their development does not require agents in any conventional sense. Although both the AGI and AI-service models can deliver general capabilities, their differences have a range of consequences; for example, by enabling access to stable AI components, competing implementations, and adversarial checking mechanisms, the CAIS model offers safety-relevant affordances that the classic AGI model does not. Both the CAIS and AGI models propose recursive improvement of AI technologies, yet they differ in their accessibility: While CAIS models anticipate accelerating R&D automation that extends conventional development methodologies, AGI models look toward conceptual breakthroughs to enable self-improvement and subsequent safe application. Because AI development services could be used to implement AGI agents, the CAIS model highlights the importance of classic AGI-safety problems, while access to SI-level services could potentially mitigate those same problems. \n Would AGI development deliver compelling value? It is widely believed that the quest to maximize useful AI capabilities will necessarily culminate in artificial general intelligence (AGI), which is taken to imply AI agents that would be able to self-improve to a superintelligent level, potentially gaining great knowledge, capability, and power to influence the world. It has been suggested AGI may be effectively unavoidable either because: 1. Self-improving AI may almost unavoidably generate AGI agents, or 2. AGI agents would provide unique and compelling value, making their development almost irresistibly attractive. However, the non-agent-based R&D automation dissociates recursive improvement from AI agency undercuts claim (1), while the prospective result of ongoing R&D automation, general AI development services, undercuts claim (2). \n AI systems deliver value by delivering services In practical terms, we value potential AI systems for what they could do, whether driving a car, designing a spacecraft, caring for a patient, disarming an opponent, proving a theorem, or writing a symphony. Scientific curiosity and long-standing aspirations will encourage the development of AGI agents with open-ended, self-directed, human-like capabilities, but the more powerful drives of military competition, economic competition, and improving human welfare do not in themselves call for such agents. What matters in practical terms are the concrete AI services provided (their scope, quality, and reliability) and the ease or difficulty of acquiring them (in terms of time, cost, and human effort). \n Providing diverse AI services calls for diverse AI capabilities Diverse AI services resolve into diverse tasks, some shared across many domains (e.g., applying knowledge of physical principles, of constraints on acceptable behavior, etc.), while other tasks are specific to a narrower range of domains. Reflection on the range of potential AI services (driving a car, proving a theorem. . . ) suggests the diversity of underlying AI tasks and competencies. We can safely assume that: • No particular AI service will require all potential AI competencies. • Satisfying general demands for new AI services will require a general ability to expand the scope of available competencies. \n Expanding AI-application services calls for AI-development services Developing a new AI service requires understanding its purpose (guided by human requests, inferred preferences, feedback from application experience, etc.), in conjunction with a process of design, implementation, and adaptation that produces and improves the required AI capabilities. Capabilitydevelopment tasks can be cast in engineering terms: They include function definition, design and implementation, testing and validation, operational deployment, in-use feedback, and ongoing upgrades. 12.6 The AGI and CAIS models organize similar functions in different ways In the CAI-services model, capability-development functions are explicit, exposed, and embodied in AI system components having suitable capacities and functional relationships. In the CAIS model, AI-enabled products are distinct from AI-enabled development systems, and the CAIS model naturally emerges from incremental R&D automation. In the classic AGI-agent model, by contrast, capability-development functions are implicit, hidden, and embodied in a single, conceptually-opaque, self-modifying agent that pursues (or is apt to pursue) world-oriented goals. Thus, capability development is internal to an agent that embodies both the development mechanism and its product. Implementation of the AGI model is widely regarded as requiring conceptual breakthroughs. \n The CAIS model provides additional safety-relevant affordances The CAIS model both exemplifies and naturally produces deeply structured AI systems 1 based on identifiable, functionally-differentiated components. Structured architectures provide affordances for both component-and system-level testing, and for the re-use of stable, well-tested components (e.g., for vision, motion planning, language understanding. . . ) in systems that are adapted to new purposes. These familiar features of practical product development and architecture can contribute to reliability in a conventional sense, but also to AI safety in the context of superintelligent-level competencies. In particular, the CAIS model offers component and system-level affordances for structuring information inputs and retention, mutability and stability, computational resource allocation, functional organization, component redundancy, and internal process monitoring; these features distance the CAIS from opaque, self-modifying agents, as does the fundamental separation of AI products from AI development processes. In the CAIS context, components (e.g., predictive models of human preferences) can be tested separately (or in diverse testbed contexts) without the ambiguities introduced by embedding similar functionality in systems with agent-level goals and potential incentives for deception. \n Constraining AI systems through external, structural affordances: Knowledge metering to bound information scope Model distillation to bound information quantity Checkpoint/restart to control information retention Focused curricula to train task specialists Specialist composition to address complex tasks Optimization applied as a constraint  \n The CAIS model enables competition and adversarial checks In developing complex systems, it is common practice to apply multiple analytical methods to a proposed implementation, to seek and compare multiple proposals, to submit proposals to independent review, and where appropriate, to undertake adversarial red-team/blue-team testing. Each of these measures can contribute to reliability and safety, and each implicitly depends on the availability of independent contributors, evaluators, testers, and competitors. Further, each of these essentially adversarial services scales to the superintelligent limit. In the CAIS model, it is natural to produce diverse, independent, taskfocused AI systems that provide adversarial services. By contrast, it has been argued that, in the classic AGI model, strong convergence (through shared knowledge, shared objectives, and strong utility optimization under shared decision theories) would render multiple agents effectively equivalent, undercutting methods that would rely on their independence. Diversity among AI systems is essential to providing independent checks, and can enable the prevention of potential collusive behaviors.  1  12.9 The CAIS model offers generic advantages over classic AGI models 12.10 CAIS affordances mitigate but do not solve AGI-control problems Because systems that can implement AI functionality at a superintelligent level can presumably be used to implement classic AGI systems, CAI services would lower barriers to the development of AGI. Given the widespread desire to realize the dream of AGI, it seems likely that AGI will, in fact, be realized unless actively prevented. Nonetheless, in a world potentially stabilized by security-oriented applications of superintelligent-level AI capabilities, prospects for the emergence of AGI systems may be less threatening. Superintelligent-level aid in understanding and implementing solutions to the AGI control problem 1 and could greatly improve our strategic position. There is no bright line between safe CAI services and unsafe AGI agents, and AGI is perhaps best regarded as a potential branch from an R&D-automation/CAIS path. To continue along safe paths from today's early AI R&D automation to superintelligent-level CAIS calls for an improved understanding of the preconditions for AI risk, while for any given level of safety, a better understanding of risk will widen the scope of known-safe system architectures and capabilities. The analysis presented above suggests that CAIS models of the emergence of superintelligent-level AI capabilities, including AGI, should be given substantial and arguably predominant weight in considering questions of AI safety and strategy. \n Further Reading • Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame Note that, by intended definition, the \"C\" in CAIS is effectively equivalent to the \"G\" in AGI. Accordingly, to propose that an AGI agent could provide services beyond the scope of CAIS is either to misunderstand the CAIS model, or to reject it, e.g., on grounds of feasibility or coherence. To be clear, fully realized CAIS services would include the service of coordinating and providing a seamless interface to other services, modeling behaviors one might have attributed to aligned AGI agents. The CAIS model of course extends to the provision of potentially dangerous services, including the service of building unaligned AGI agents. \n Classic AGI models increase the challenges of AI goal alignment The classic AGI model posits the construction of a powerful, utility-directed, superintelligent agent, a conceptual move that both engenders the problems of aligning a superintelligent agent's overall goals with human values and amalgamates and abstracts away the concrete problems that arise in aligning specific, useful behaviors with diverse and changing human goals. Although a simplified AI-services model could arbitrarily assume aligned systems with bounded goals and action spaces, a deeper model offers a framework for considering how such systems could be developed and how their development might go wrong-for example, by indirectly and inadvertently giving rise to agents with problematic goals and capabilities. Many of the questions first framed as problems of AGI safety still arise, but in a different and perhaps more tractable systemic context. \n The CAIS model addresses a range of problems without sacrificing efficiency or generality To summarize, in each of the areas outlined above, the classic AGI model both obscures and increases complexity: In order for general learning and capabilities to fit a classic AGI model, they must not only exist, but must be integrated into a single, autonomous, self-modifying agent. Further, achieving this kind of integration would increase, not reduce, the challenges of aligning AI behaviors with human goals: These challenges become more difficult when the goals of a single agent must motivate all (and only) useful tasks. Agent-services that are artificial, intelligent, and general are surely useful, both conceptually and in practice, but fall within the scope of comprehensive agent (and non-agent) AI services. The key contribution of the CAIS model is to show how integrated, fully-general AI capabilities could be provided within an open-ended architecture that is natural, efficient, relatively transparent, and quite unlike a willful, uniquely-powerful agent. \n Further Reading • Section 1: R&D automation provides the most direct path to an intelligence explosion • Section 6: A system of AI services is not equivalent to a utility maximizing agent 14.5 CAIS capabilities could empower bad actors. AI-service resources per se are neutral in their potential applications, and human beings can already apply AI services and products to do intentional harm. Access to advanced AI services could further empower bad actors in ways both expected and as-yet unimagined; in compensation, advanced AI services could enable detection and defense against bad actors. Prospective threats and mitigation strategies call for exploration and study of novel policy options. \n CAIS capabilities could facilitate disruptive applications Today's disruptive AI applications are services that serve some practical purpose. As shown by current developments, however, one person's service may be another person's threat, whether to business models, employment, privacy, or military security. Increasingly comprehensive and high-level AI services  \n CAIS capabilities could facilitate seductive and addictive applications Some disruptive applications provide seductive services that are detrimental, yet welcomed by their targets. AI systems today are employed to make games more addictive, to build comfortable filter bubbles, and to optimize message channels for appeal unconstrained by truth. There will be enormous scope for high-level AI systems to please the people they harm, yet mitigation of the individual and societal consequences of unconstrained seductive and addictive services raises potentially intractable questions at the interface of values and policy. \n Conditions for avoiding emergent agent-like behaviors call for further study Although it is important to distinguish between pools of AI services and classic conceptions of integrated, opaque, utility-maximizing agents, we should be alert to the potential for coupled AI services to develop emergent, unintended, and potentially risky agent-like behaviors. Because there is no bright line between agents and non-agents, or between rational utility maximization and reactive behaviors shaped by blind evolution, avoiding risky behaviors calls for at least two complementary perspectives: both (1) design-oriented studies that can guide implementation of systems that will provide requisite degrees of e.g., stability, reliability, and transparency, and (2) agent-oriented studies support design by exploring the characteristics of systems that could display emergent, unintended, and potentially risky agent-like behaviors. The possibility (or likelihood) of humans implementing highly-adaptive agents that pursue open-ended goals in the world (e.g., money-maximizers) presents particularly difficult problems. Further Reading  \n AI safety research has often focused on unstructured rational-agent models Research in AI safety has been motivated by prospects for recursive improvement that enables the rapid development of systems with general, superhuman problem-solving capabilities. Research working within this paradigm has centered not on the process of recursive improvement, but on its potential products, and these products have typically been modeled as discrete, relatively unstructured, general-purpose AI systems. \n Structured systems are products of structured development In an alternative, potentially complementary model of high-level AI, the products of recursive improvement are deeply structured, task-focused systems that collectively deliver a comprehensive range of superintelligent task capabilities, yet need not be (or become) discrete entities that individually span a full range of capabilities. This model has been criticized as potentially incurring high development costs, hence prospects for deploying differentiated (but not necessarily narrow) AI systems with a collectively comprehensive range of task capabilities may best be approached through an explicit consideration of potential AI research and development processes. \n AI development naturally produces structured AI systems Today, the AI research community is developing a growing range of relatively narrow, strongly differentiated AI systems and composing them to build systems that embrace broader domains of competence. A system that requires, for example, both vision and planning will contain vision and planning components; a system that interprets voice input will contain interacting yet distinct speech recognition and semantic interpretation components. A self-driving car with a conversational interface would include components with all of the above functionalities, and more. The principle of composing differentiated competencies to implement broader task-performance naturally generalizes to potential systems that would perform high-order tasks such as the human-directed design and management of space transportation systems, or AI research and development. \n Structure arises from composing components, not partitioning unitary systems If one begins with unitary systems as a reference model, the task of implementing structured, broadly competent AI systems may appear to be a problem of imposing structure by functional decomposition, rather than one of building structures by composing functional components. In other words, taking a unitarysystem model as a reference model focuses attention on how a hypothetical system with unitary, universal competence might be divided into parts. While this framing may be useful in conceptual design, it can easily lead to confusion regarding the nature of structured AI system development and products. 15.6 A development-oriented approach to deeply structured systems suggests a broad range of topics for further inquiry A focus on AI development links AI safety studies to current R&D practice. In particular, the prospective ability to deliver deeply-structured, task-focused AI systems offers rich affordances for the study and potential implementation of safe superintelligent systems. The joint consideration of structured AI development and products invites inquiry into a range of topics, including: • Abstract and concrete models of structured AI systems  \n Summary Since Turing, discussions of advanced AI have tacitly assumed that agents will learn and act as individuals; naïve scaling to multi-agent systems retains a human-like model centering on individual experience and learning. The development of self-driving vehicles, however, illustrates a sharply contrasting model in which aggregation of information across N agents (potentially thousands to millions) speeds the acquisition of experience by a factor of N, while centralized, large-scale resources are applied to training, and amortization reduces a range of per-agent costs by a factor of 1/N. In addition to advantages in speed and amortization, centralized learning enables pre-release testing for routinely encountered errors and ongoing updates in response to rarely-encountered events. The strong, generic advantages of aggregated experience and centralized learning have implications for our understanding of prospective AI-agent applications. \n Discussions often assume learning centered on individual agents Advanced AI agents are often modeled as individual machines that learn tasks in an environment, perhaps with human supervision, along the lines suggested by Fig.  1 . In human experience, human beings have been the sole intelligent agents in the world, a circumstance that powerfully reinforces our habit of identifying experience and learning with individual agents.  \n Naïve scaling to multi-agent systems replicates individual agents Naïve scaling of the individual-agent model to multiple agents does not fundamentally alter this picture (Fig  2 ). One can extend the environment of each agent to include other AI agents while retaining a human-like model of learning: Other agents play a role like that of other human beings. Figure  7 : In a naïve scale-up of the scheme outlined in Fig.  1 , N agents would plan, act, and adapt their actions to a range of similar task environments while learning from experience independently; both costs and benefits scale as N, and the time required to learn tasks remains unchanged. 16.4 Self-driving vehicles illustrate the power of aggregating experience Self-driving cars are agents that follow a quite different model: They do not learn as individuals, but instead deliver improved competencies through a centralized R&D process that draws on the operational experience of many vehicles. Tesla today produces cars with self-driving hardware at a rate of approximately 100,000 per year, steadily accelerating the accumulation of driving experience (termed \"fleet learning\"). Vehicle-agents that learned only from individual experience could not compete in performance or safety. \n Efficiently organized machine learning contrasts sharply with human learning Machine learning contrasts sharply with human learning in its potential for efficiently aggregating experience, amortizing learning, applying populationbased exploration  (Conti et al. 2017) , and reproducing and distributing competencies. Experience can be aggregated. Among human beings, the ability to share detailed experiences is limited, and learning from others competes with learning from experience. In machine learning, by contrast, experience can be aggregated, and learning from aggregated experience need not compete with the acquisition of further experience. Learning can be accelerated and amortized. Among human beings, applying a thousand brains to learning does not reduce individual learning time or cost. In machine learn-ing, by contrast, applying increasing computational capacity can reduce learning time, and computational costs can be amortized across an indefinitely large number of current and future systems. Competencies can be reproduced. Among human beings, training each additional individual is costly because competencies cannot be directly reproduced (hence learning elementary mathematics absorbs millions of person-years per year). In machine learning, learned competencies can be reproduced quickly at the cost of a software download.  \n Aggregated learning speeds development and amortizes costs With learning aggregated from N agents, the time required to gain a given quantity of experience scales as 1/N, potentially accelerating development of competencies. Further, the computation costs of training can be amortized over N agents, yielding a per-agent cost that scales as 1/N. If common situations each require human advice when first encountered, the burden of supervising an independent agent might be intolerable, yet acceptably small when employing an agent trained with ongoing experience aggregated from a large-N deployment. Similarly, if the causes of novel errors can be promptly corrected, the per-agent probability of encountering a given error will be bounded by 1/N. \n The advantages of aggregated, amortized learning have implications for prospective AI-agent applications The advantages of richer data sets, faster learning, and amortization of the costs of training and supervision all strongly favor development approaches that employ centralized, aggregated learning across deployments of agents that share similar tasks. These considerations highlight the importance of development-oriented models in understanding prospects for the emergence of broad-spectrum AI-agent applications. Further Reading  \n RL and end-to-end training are powerful, yet bounded in scope Complex products (both hardware and software) have generally been built of components with differentiated competencies. End-to-end training challenges this model: Although systems are commonly differentiated in some respects (e.g., convolutional networks for visual processing, recurrent neural networks for sequential processing, external memories for long-term representation), these system components and their learned content do not align with distinct tasks at an application level. Functional competencies are (at least from an external perspective) undifferentiated, a confronting us with black-box systems. Will end-to-end training of black-box systems scale to the development of AI systems with extremely broad capabilities? If so-and if such methods were to be both efficient by metrics of development cost and effective by metrics of product quality-then advanced AI systems might be expected to lack the engineering affordances provided by differentiated systems. In particular, component functionalities might not be identifiable, separable, and subject to intensive training and testing. There is, however, reason to expect that broad AI systems will comprise patterns of competencies that reflect (and expose 1 ) the natural structure of complex tasks.  2  The reasons include both constraints (the nature of training and bounded scope of transfer learning) and opportunities (the greater robustness and generalization capabilities of systems that exploit robust and general components). \n General capabilities comprise many tasks and end-to-end relationships What would it even mean to apply end-to-end training to a system with truly general capabilities? Consider a hypothetical system intended to perform a comprehensive range of diverse tasks, including conversation, vehicle guidance, AI R&D, 3 and much more. What input information, internal architecture, output modalities, and objective functions would ensure that each task is trained efficiently and well? Given the challenges of transfer learning  (Teh et al. 2017 ) even across a range of similar games, why would one expect to find a compelling advantage in learning a comprehensive range of radically different tasks through end-to-end training of a single, undifferentiated system? Diverse tasks encompass many end-to-end relationships. A general system might provide services that include witty conversation and skillful driving, but it is implausible that these services could best be developed by applying deep RL to a single system. Training a general system to exploit differentiated resources (providing knowledge of vehicle dynamics, scene interpretation, predictions of human road behavior; linked yet distinct resources for conversing with passengers about travel and philosophy) seems more promising than attempting to treat all these tasks as one. \n Broad capabilities are best built by composing well-focused competencies Systems that draw on (and perhaps adapt) distinct subtask competencies will often support more robust and general performance. For example, to interact with human beings well calls for a model of many aspects of human concerns, capabilities, intentions, and responses to situations-aspects that are unlikely to be thoroughly explored through deep RL within the scope of any particular task. For example, a model of an open task environment, such as a road, may fail to model child ball-chasing events that are rare on roads, but common on playgrounds. Similarly, a system intended to explore theoretical physics might struggle to discover mathematical principles that might better be provided through access to a system with strong, specifically mathematical competencies. The use of focused, broadly-applicable competencies in diverse contexts constitutes a powerful form of transfer learning. Narrow components can support-and strengthen-broad capabilities, and are best learned in depth and with cross-task generality, not within the confines of a particular application. Note that components can be distinct, yet deeply integrated 1 at an algorithmic and representational level. \n Deep RL can contribute to R&D automation within the CAIS model of general AI Reinforcement learning fits naturally with the R&D automation model of comprehensive AI services.  2  Deep RL has already been applied to develop state-of-the art neural networks  (Zoph and Le 2016) , including scalable modular systems  (Zoph et al. 2017) , and deep RL has been applied to optimizing deep RL systems. Increasing automation of AI R&D will facilitate the development of task-oriented systems of all kinds, and will naturally result in deeplystructured systems.  3  In considering RL in the context of AI control and safety, in is important to keep in mind that RL systems are not utility-maximizing agents, 4 that learning is separable from performance, 5 that human oversight need not impede rapid progress,  6  and that component-level algorithmic opacity is compatible with  \n Summary Reward-seeking reinforcement-learning agents can in some instances serve as models of utility-maximizing, self-modifying agents, but in current practice, RL systems are typically distinct from the agents they produce, and do not always employ utility-like RL rewards. In multi-task RL systems, for example, RL \"rewards\" serve not as sources of value to agents, but as signals that guide training, and unlike utility functions, RL \"rewards\" in these systems are neither additive nor commensurate. RL systems per se are not reward-seekers (instead, they provide rewards), but are instead running instances of algorithms that can be seen as evolving in competition with others, with implementations subject to variation and selection by developers. Thus, in current RL practice, developers, RL systems, and agents have distinct purposes and roles. \n Reinforcement learning systems differ sharply from utility-directed agents Current AI safety discussions sometimes treat RL systems as agents that seek to maximize reward, and regard RL \"reward\" as analogous to a utility function. Current RL practice, however, diverges sharply from this model: RL systems comprise often-complex training mechanisms that are fundamentally distinct from the agents they produce, and RL rewards are not equivalent to utility functions. \n RL systems are neither trained agents nor RL-system developers In current practice, RL systems and task-performing agents often do not behave as unitary \"RL agents\"; instead, trained agents are products of RL systems, while RL systems are products of a development process. Each of these levels (development processes, RL systems, and task-performing agents) is distinct in its implementation and implicit goals. 18.4 RL systems do not seek RL rewards, and need not produce agents RL-system actions include running agents in environments, recording results, and running RL algorithms to generate improved agent-controllers. These RL-system actions are not agent-actions, and rewards to agents are not rewards to RL systems. Running agents that collect rewards is a training cost, not a source of reward to the training system itself. Note that the products of RL systems need not be agents: For example, researchers have applied RL systems to train mechanisms for attention in vision networks  (Xu et al. 2015) , to direct memory access in memory-augmented RNNs  (Gülçehre, Chandar, and Bengio 2017) , and (in meta-learning) to develop RL algorithms in RL systems  (Duan et al. 2016 ; J. X. ). RL systems have also been used to design architectures for convolutional neural networks  (Baker et al. 2016; Zoph and Le 2016)  and LSTM-like recurrent cells for natural-language tasks (Zoph and Le 2016). 18.5 RL rewards are not, in general, treated as increments in utility \"Reward\" must not be confused with utility. In DeepMind's work on multitask learning, for example, agents are trained to play multiple Atari games with incommensurate reward-scores. Researchers have found that these heterogeneous \"rewards\" cannot be scaled and summed over tasks as if they were measures of utility, but must instead be adaptively adjusted to provide learning signals that are effective across different games and stages of training  [ref] . RL rewards are sources of information and direction for RL systems, but are not sources of value for agents. Researchers often employ \"reward shaping\" to direct RL agents toward a goal, but the rewards used shape the agent's behavior are conceptually distinct from the value of achieving the goal. 1 \n Experience aggregation blurs the concept of individual reward Modern RL systems typically aggregate experience 2 across multiple instances of agents that run in parallel in different environments. An agent-instance does not learn from \"its own\" experience, and aggregated experience may include off-policy actions that improve learning, yet impair reward-maximization for any given instance. 18.7 RL algorithms implicitly compete for approval RL algorithms have improved over time, not in response to RL rewards, but through research and development. If we adopt an agent-like perspective, RL algorithms can be viewed as competing in an evolutionary process where success or failure (being retained, modified, discarded, or published) depends on developers' approval (not \"reward\"), which will consider not only current performance, but also assessed novelty and promise. \n Distinctions between system levels facilitate transparency and control The patterns and distinctions described above (developer vs. learning system vs. agent) are not specific to RL, and from a development-oriented perspective, they seem generic. Although we can sometimes benefit from dropping these distinctions and exploring models of agent-like RL systems that seek utility-like rewards, current and future RL systems need not conform to those models. AI development practice suggests that we also consider how AI components and systems are architected, trained, combined, and applied. A development-oriented perspective 1 focuses attention on structured processes, structured architectures, and potential points of control that may prove useful in developing safe applications of advanced AI technologies. \n RL-driven systems remain potentially dangerous It should go without saying that RL algorithms could serve as engines driving perverse behavior on a large scale. For an unpleasantly realistic example, consider the potential consequences of giving an RL-driven system read/write access to the internet-including access to contemporaneous AI services-with the objective of maximizing the net flow of money into designated accounts. In this scenario, consider how the value of a short position could be increased by manipulating news or crashing a power grid. Distinctions between system levels offer affordances for control, yet levels can be collapsed, and having affordances for control in itself precludes neither accidents nor abuse. Further Reading 19 The orthogonality thesis undercuts the generality of instrumental convergence If any level of intelligence can be applied to any goal, then superintelligent-level systems can pursue goals for which the pursuit of the classic instrumentally-convergent subgoals would offer no value. \n Summary Bostrom (2014) presents carefully qualified arguments regarding the \"orthogonality thesis\" and \"instrumental convergence\", but the scope of their implications has sometimes been misconstrued. The orthogonality thesis proposes that any level of intelligence can be applied to any goal (more or less), and the principle of instrumental convergence holds that a wide range of goals can be served by the pursuit of subgoals that include self preservation, cognitive enhancement, and resource acquisition. This range of goals, though wide, is nonetheless limited to goals of indefinite scope and duration. The AI-services model suggests that essentially all practical tasks are (or can be) directly and naturally bounded in scope and duration, while the orthogonality thesis suggests that superintelligent-level capabilities can be applied to such tasks. At a broad, systemic level, tropisms toward general instrumental subgoals seem universal, but such tropisms do not imply that a diffuse system has the characteristics of a willful superintelligent agent. \n 19.2 The thesis: Any level of intelligence can be applied to any goal (more or less) The orthogonality thesis  (Bostrom 2014, p.107)  proposes that intelligence and final goals are orthogonal: \"[. . .] more or less any level of intelligence can be combined with more or less any final goal.\" A natural consequence of the orthogonality thesis is that intelligence of any level can be applied to goals that correspond to tasks of bounded scope and duration. \n A wide range of goals will engender convergent instrumental subgoals As Marvin Minsky noted in a conversation ca. 1990, a top-level goal of narrow scope (e.g., playing the best possible game of chess) can be served by a subgoal of enormous scope (e.g., converting all accessible resources into chess-playing machinery). The instrumental convergence (IC) thesis  (Bostrom 2014  The explicitly considered IC subgoals are: • Self preservation (to pursue long-term goals, an agent must continue to exist) • Goal-content integrity (to pursue long-term goals, an agent must maintain them) • Cognitive enhancement (gaining intelligence would expand an agent's capabilities) • Technological perfection (developing better technologies would expand an agent's capabilities) • Resource acquisition (controlling more resources would expand an agent's capabilities) Recognizing the broad scope of IC subgoals provides insight into potential behaviors of a system that pursues goals with a \"superintelligent will\"  (Bostrom 2014, p.105) . \n Not all goals engender IC subgoals As formulated, the IC thesis applies to \"a wide range [implicitly, a limited range] of final goals\", and the subsequent discussion  (Bostrom 2014, p.109 ) suggests a key condition, that \"an agent's final goals concern the future\". This condition is significant: Google's neural machine translation system, for example, has no goal beyond translating a given sentence, and the scope of this goal is independent of the level of intelligence that might be applied to achieve it. In performing tasks of bounded scope and duration, the pursuit of longerterm IC subgoals would offer no net benefit, and indeed, would waste resources. Optimization pressure on task-performing systems can be applied to suppress not only wasteful, off-task actions, but off-task modeling and planning.  1  19.5 Not all intelligent systems are goal-seeking agents in the relevant sense As formulated above, the IC thesis applies to \"situated agents\", yet in many instances intelligent systems that perform tasks are neither agents nor situated in any conventional sense: Consider systems that prove theorems, translate books, or perform a series of design optimizations. Further, agents within the scope of the IC thesis are typically modeled as rational, utility-directed, and concerned with goals of broad scope, yet even situated agents need not display these characteristics (consider System 1 decision-making in humans). \n Comprehensive services can be implemented by systems with bounded goals The AI-services model invites a functional analysis of service development and delivery, and that analysis suggests that practical tasks in the CAIS model are readily or naturally bounded in scope and duration. For example, the task of providing a service is distinct from the task of developing a system to provide that service, and tasks of both kinds must be completed without undue cost or delay. Metalevel tasks such as consulting users to identify application-level tasks and preferences, selecting and configuring systems to provide desired services, supplying necessary resources, monitoring service quality, aggregating data across tasks, and upgrading service-providers are likewise bounded in scope and duration. This brief sketch outlines a structure of generic, bounded tasks that could, by the orthogonality thesis, be implemented by systems that operate at a superintelligent level. It is difficult to identify bounded (vs. explicitly world-optimizing) human goals that could more readily be served by other means. \n IC goals naturally arise as tropisms and as intended services The IC thesis identifies goals that, although though not of value in every context, are still of value at a general, systemic level. The IC goals arise naturally in an AI-services model, not as the result of an agent planning to manipulate world-outcomes in order to optimize an over-arching goal, but as system-level tropisms that emerge from local functional incentives: • Self preservation: Typical service-providing systems act in ways that avoid self-destruction: Self-driving cars are an example, though missiles are an exception. AI systems, like other software, can best avoid being scrapped by providing valuable services while not disrupting their own operation. 19.8 Systems with tropisms are not equivalent to agents with \"will\" The aggregate results of AI-enabled processes in society would tend to advance IC goals even in the absence of distinct AI agents that meet the conditions of the IC thesis. To regard systemic tropisms as manifestations of a \"superintelligent will\", however, would be much like attributing a \"will\" to a global ecosystem or economy-a potentially useful perspective that does not reflect an equivalence. The analogy between tropisms and will invites a \"motte and bailey argument\" that wrongly attributes the properties of willful rational agents to all systems in which strong aggregate capabilities provide wide-ranging services. Similarly, to argue that a diffuse system would itself undertake actions to become a willful agent in order to pursue IC subgoals is essentially circular. \n Further Reading • Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame  \n Summary The potential for successful collusion among actors decreases as their number increases and as their capabilities, knowledge, situations, and roles become more diverse. In the context of AI systems, these conditions are easy to implement and are often naturally preferred for practical reasons. Accordingly, arguments that sets of high-level AI systems will naturally or inevitably collude are ill-founded, and diverse superintelligent-level question-answering systems (\"oracles\") can be employed to improve, rather than degrade, safety. \n Trustworthiness can be an emergent property Prospects for solving AI-safety problems would be greatly improved if we could safely apply SI-level question-answering capabilities to solving those problems. A familiar objection dismisses this potentially powerful approach as unsafe in itself, arguing that, absent solutions to difficult problems, individual SI-level systems would be untrustworthy, and that attempts to establish checks and balances among multiple systems (e.g., through SI-level evaluation of proposed answers to queries) would be thwarted by collusive cooperation. Identifying robust methods for ensuring non-collusion among SI queryanswering systems would overcome this objection, enabling the construction of trustworthy composite oracles and inviting the exploration of SI-enabled strategies for managing potentially untrusted SI systems. The present discussion argues that the robust non-emergence of deceptive collusion among SI query-answering systems can be ensured by structuring a set of systems with diverse capabilities, knowledge, situations, and roles. Given the problem framing, the discussion here will take as nonproblematic the ability to develop systems that can respond to queries with SI-level competence, and (as a consequence of the nature of digital systems) will assume the ability to constrain information inputs to SI systems during their development and use. In the problematic case, SI query-answering systems might provide deceptive answers in pursuit of emergent, potentially undesirable objectives. The aim here is to identify principles for architecting composite systems that will act as non-deceptive oracles, while treating their components as actors with potential incentives for deceptive collusion. \n A range of conditions may facilitate or disrupt collusion Potential problems of deceptive collusion can be addressed by architecting systems that make collusion difficult or impossible. We can identify a range of relevant conditions: Conditions that tend to facilitate collusion among actors A1) Small numbers of actors (the required scope of collusion is small) A2) Insensitivity to defectors (partial collusion is sufficient) A3) Similarity among actors (e.g., shared histories and objectives) A4) Iterated decision processes with universally observable actions A5) Common knowledge of the above circumstances Contrasting conditions that tend to disrupt collusion among actors B1) Large numbers of actors (the required scope of collusion is large) B2) Sensitivity to defectors (collusion fails unless it is complete) B3) Diversity among actors (e.g., different histories, adversarial objectives) B4) Single-move decision processes in which most actions are unobservable B5) Lack of common knowledge of many strategically-relevant circumstances The implementation choices implied by conditions (A1-A5) are neither necessary nor desirable, while as suggested by (C1-C5) below, the choices implied by conditions (B1-B5) are both practical and well-motivated. Natural architectures for problem-solving systems C1) To improve the quality of answers, it is natural to implement multiple, diverse (and implicitly competing) systems to propose alternatives. C2) To identify low-quality or misleading answers, it is natural to employ diverse critics, any one of which could disrupt deceptive collusion. C3) Systems of diverse, competing proposers and critics naturally implement both independent and adversarial objectives. C4) It is natural to apply fixed (hence memory-free) system instantiations to multiple problems, incidentally yielding a series of history-blind, singlemove decisions. C5) It is natural to provide differentiated, task-relevant information to systems solving different problems, typically omitting knowledge of general circumstances. Points (C1-C5) suggest that implementing systems that are resistant to the emergence of deceptive collusion aligns with the general strategy of employing ensembles of diverse, stable, task-directed system components to obtain and assess answers to queries. In appropriate architectures, the use of greater numbers of potentially untrustworthy SI-level components can improve rather than degrade safety. 1 \n Collusion is fragile and easily disrupted In light of the above considerations, it seems both possible and practical to establish conditions that would preclude deceptive collusion among diverse, task-oriented, SI-level query-answering systems, thereby addressing a key problem in safely applying oracle-class superintelligent capabilities (see  Bostrom [2014, p.145] ). Prospects for obtaining trustworthy answers from sets of individually untrusted SI systems call for deeper exploration of AI safety strategies premised on reliable access to superintelligent problem-solving capabilities. Capabilities of particular interest include the ability run and evaluate the results of superintelligent-level red-team/blue-team competitions relevant to questions of global strategic stability. Christiano (2015a) considers collusion and a range of related topics; I agree with his gently-stated view that \"[. . .] the implicit argument for the robustness of collusion makes a lot of implicit assumptions. If I saw an explicit argument I might be able to assess its explicit assumptions, but for now we don't have one.\" Further Reading  \n Summary Is strongly-bounded world knowledge necessary to ensure strongly-bounded AI behavior? Language translation shows otherwise: Machine translation (MT) systems are trained on general text corpora, and would ideally develop and apply extensive knowledge about the world, yet MT systems perform a well-bounded task, serving as functions of type T :: string → string. Broad knowledge and linguistic competencies can support (rather than undermine) AI safety by enabling systems to learn predictive models of human approval from large corpora. \n Bounding task focus does not require bounding world knowledge Placing tight bounds on knowledge could be used to restrict AI competencies, and dividing broad AI tasks among components with restricted competencies could help to ensure AI safety (as discussed in  Drexler [2015] ), yet some tasks may require broad, integrated knowledge of the human world. Generalpurpose machine translation (MT) exemplifies a task that calls for integrated knowledge of indefinite scope, but also illustrates another, distinct mechanism for restricting competencies: Task-focused training (also discussed in  Drexler [2015] ). \n Extensive world knowledge can improve (e.g.) translation Fully-general human-quality machine translation would require understanding diverse domains, which in turn would require knowledge pertaining to human motivations, cultural references, sports, technical subjects, and much more. As MT improves, there will be strong incentives to incorporate broader and deeper world knowledge. \n Current MT systems are trained on open-ended text corpora Current neural machine translation (NMT) systems gain what is, in effect, knowledge of limited kinds yet indefinite scope through training on large, general text corpora. Google's GNMT architecture has been trained on tens of millions of sentence pairs for experiments and on Google-internal production datasets for on-line application; the resulting trained systems established a new state-of-the-art, approaching the quality of \"average human translators\"  (Wu et al. 2016) . Surprisingly, Google's NMT architecture has been successfully trained, with little modification, to perform bidirectional translation for 12 language pairs  (Johnson et al. 2016 ). Although the system used no more parameters than the single-pair model (278 million parameters), the multilingual model achieves a performance \"reasonably close\" to the best single-pair models. Subsequent work (below) developed efficient yet greatly expanded multilingual models that improve on previous single-model performance. \n Current systems develop language-independent representations of meaning NMT systems encode text into intermediate representations that are decoded into a target language. In Google's multilingual systems, one can compare intermediate representations generated in translating sentences from multiple source languages to multiple targets. Researchers find that, for a given set of equivalent sentences (paired with multiple target languages), encodings cluster closely in the space of representations, while these clusters are wellseparated from similar clusters that represent sets of equivalent sentences with a different meaning. The natural interpretation of this pattern is that sentenceencodings represent meaning in a form that is substantially independent of any particular language. (It may prove fruitful to train similar NMT models on sets of pairs of equivalent sentences while providing an auxiliary loss function that pushes representations within clusters toward closer alignment. One would expect training methods with this auxiliary objective to produce higher-quality languageindependent representations of sentence meaning, potentially providing an improved basis for learning abstract relationships.) \n Scalable MT approaches could potentially exploit extensive world knowledge NMT systems can represent linguistic knowledge in sets of specialized \"experts\", and Google's recently described \"Sparsely-Gated Mixture-of-Experts\" (MoE) approach  (Shazeer et al. 2017 ) employs sets of sets of experts, in effect treating sets of experts as higher-order experts. Human experience suggests that hierarchical organizations of experts could in principle (with suitable architectures and training methods) learn and apply knowledge that extends beyond vocabulary, grammar, and idiom to history, molecular biology, and mathematics. Notably, Google's MoE system, in which \"different experts tend to become highly specialized based on syntax and semantics\" has enabled efficient training and application of \"outrageously large neural networks\" that achieve \"greater than 1000× improvements in model capacity [137 billion parameters] with only minor losses in computational efficiency on modern GPU clusters\". (There is reason to think that these systems exceed the computational capacity of the human brain. 1 ) \n Specialized modules can be trained on diverse, overlapping domains As suggested by the MoE approach in NMT, domain-specialized expertise can be exploited without seeking to establish clear domain boundaries that might support safety mechanisms. For example, it is natural to expect that efficient and effective training will focus on the learning the concepts and language of chemistry (for training some modules), of history (for training others), and of mathematics (for yet others). It is also natural to expect that training modules for expertise in translating chemistry textbooks would benefit from allowing them to exploit models trained on math texts, while performance in the history of chemistry would benefit from access to models trained on chemistry and on general history. Current practice and the structure of prospective task domains 1 suggests that optimal partitioning of training and expertise would be soft, and chosen to improve efficiency and effectiveness, not to restrict the capabilities of any part. \n Safe task focus is compatible with broad, SI-level world knowledge Machine translation systems today are not agents in any conventional sense of the word, and are products of advances in an AI-development infrastructure, not of open-ended \"self improvement\" of any distinct system. As we have seen, domain-specific task focus is, in this instance, a robust and natural consequence of optimization and training for a specific task, while high competence in performing that task, employing open-ended knowledge, does not impair the stability and well-bounded scope of the translation task. It is reasonable to expect that a wide range of other tasks can follow the same basic model, 2 though the range of tasks that would naturally (or could potentially) have this character is an open question. We can expect that broad knowledge will be valuable and (in itself) non-threatening when applied to tasks that range in scope from driving automobiles to engineering urban transportation systems. Further, broad understanding based on large corpora could contribute to predictive models of human approval 3 that provide rich priors for assessing the desirability (or acceptability) of proposed plans for action by agent, helping to solve a range of problems in aligning AI behaviors with human values. By contrast, similar understanding applied to, for example, unconstrained profit maximization by autonomous corporations, could engender enormous risks. \n Strong task focus does not require formal task specification The MT task illustrates how a development-oriented perspective can reframe fundamental questions of task specification. In MT development, we find systems (now significantly automated 1  [Britz et al. 2017] ) that develop MT architectures, systems that train those architectures, and (providing services outside R&D labs 2 ) the trained-and-deployed MT systems themselves. The nature and scope of the MT task is implicit in the associated training data, objective functions, resource constraints, efficiency metrics, etc., while tasks of the systems that develop MT systems are indirectly implicit in that same MT task (together with metalevel resource constraints, efficiency metrics, etc.). Nowhere in this task structure is there a formal specification of what it means to translate a language, or a need to formally circumscribe and limit the task. \n Further Reading \n • Section 2: Standard definitions of \"superintelligence\" conflate learning with competence • Section 7: Training agents in human-like environments can provide useful, bounded services • Section 10: R&D automation dissociates recursive improvement from AI agency • Section 22: Machine learning can develop predictive models of human approval • Section 23: AI development systems can support effective human guidance • Section 35: Predictable aspects of future knowledge can inform AI safety strategies 22 Machine learning can develop predictive models of human approval By exploiting existing corpora that reflect human responses to actions and events, advanced ML systems could develop predictive models of human approval with potential applications to AI safety. \n Summary Advanced ML capabilities will precede the development of advanced AI agents, and development of predictive models of human approval need not  \n Advanced ML technologies will precede advanced AI agents The R&D-automation model of AI development shows how asymptoticallyrecursive AI-technology improvement could yield superintelligent systems (e.g., machine learning systems) without entailing the use of agents. In this model, agents are potential products, not necessary components. \n Advanced ML can implement broad predictive models of human approval As Stuart Russell has remarked, AI systems will be able to learn patterns of human approval \"Not just by watching, but also by reading. Almost everything ever written down is about people doing things, and other people having opinions about it.\" By exploiting evidence from large corpora (and not only text), superintelligent-level machine learning could produce broad, predictive models of human approval and disapproval of actions and events (note that predicting human approval conditioned on events is distinct from predicting the events themselves). Such models could help guide and constrain choices made by advanced AI agents, being directly applicable to assessing intended consequences of actions. \n Text, video, and crowd-sourced challenges can provide training data Models of human approval can draw on diverse and extensive resources. Existing corpora of text and video reflect millions of person-years of actions, events, and human responses; news, tweets, history, fiction, science fiction, advice columns, sitcoms, social media, movies, CCTV cameras, legal codes, court records, and works of moral philosophy (and more) offer potential sources of training data. An interactive crowd-sourcing system could challenge participants to \"fool the AI\" with difficult cases, eliciting erroneous predictions of approval to enable training on imaginative hypotheticals. \n Predictive models of human approval can improve AI safety Predictive models of human approval and disapproval could serve as safetyrelevant components of structured AI systems. Armstrong's (2017) concept of low-impact AI systems points to the potentially robust value of minimizing significant unintended consequences of actions, and models of human approval imply models of what human beings regard as significant. When applied to Christiano's (2014) concept of approval-directed agents, general models of human approval could provide strong priors for interpreting and generalizing indications of human approval for specific actions in novel domains. 1 \n Prospects for approval modeling suggest topics for further inquiry The concept of modeling human approval by exploiting large corpora embraces a wide range of potential implementation approaches and applications. The necessary scope and quality of judgment will vary from task to task, as will the difficulty of developing and applying adequate models. In considering paths forward, we should consider a spectrum of prospective technologies that extends from current ML capabilities and training methods to models in which we freely assume superhuman capabilities for comprehension and inference. In considering high-level capabilities and applications, questions may arise with ties to literatures in psychology, sociology, politics, philosophy, and economics. Models of approval intended for broad application must take account of the diversity of human preferences, and of societal patterns of approval and disapproval of others' preferences. Criteria for approval may be relatively straightforward for self-driving cars, yet intractable for tasks that might have broad effects on human affairs. For tasks of broad scope, the classic problems of AI value alignment arise, yet some of these problems (e.g., perverse instantiation) could be substantially mitigated by concrete models of what human beings do and do not regard as acceptable. •  \n We should assume effective use of natural-language understanding In considering advanced AI capabilities, should assume that a range of current objectives have been achieved, particularly where we see strong progress today. In particular, natural-language understanding  (Johnson et al. 2016 ) can support powerful mechanisms for defining AI tasks and providing feedback on AIsystem behaviors. Even imperfect language understanding can be powerful because both large text corpora and interactive communication (potentially drawing on the knowledge and expressive capabilities of many individuals) can help to disambiguate meaning. \n Generic models of human (dis)approval can provide useful priors In some applications, avoiding undesirable AI behaviors will require a broad understanding of human preferences. Explicit rules and direct human instruc-tion seem inadequate, but (along lines suggested by Stuart Russell  [Wolchover 2015] ) advanced machine learning could develop broad, predictive models of human approval and disapproval 1 by drawing on large corpora of, for example, news, history, science fiction, law, and philosophy, as well as the cumulative results of imaginative, crowd-sourced challenges. Generic models of human (dis)approval can provide useful priors in defining task-specific objectives, as well as constraints on actions and side-effects. Note that predicting human approval conditioned on events is distinct from predicting the events themselves. Accordingly, in judging an agent's potential actions, predictive models of approval may fail to reflect unpredictable, unintended consequences, yet be effective in assessing predicted, intended consequences (of, e.g., misguided or perverse plans). \n Bounded task objectives can be described and circumscribed Describing objectives is an initial step in systems development, and conventional objectives are bounded in terms of purpose, scope of action, allocated resources, and permissible side-effects. Today, software developed for selfdriving cars drives particular cars, usually on particular streets; in the future, systems developed to study cancer biology will perform experiments in particular laboratories, systems developed to write up experimental results will produce descriptive text, AI-architecting systems will propose candidate architectures, and so on. Priors based on models of human approval 2 can help AI development systems suggest task-related functionality, while models of human disapproval can can help development systems suggest (or implement) hard and soft constraints; along lines suggested by  Armstrong (2013) , these can include minimizing unintended effects that people might regard as important. \n Observation can help systems learn to perform human tasks For task that humans can perform, human behavior can be instructive, not only with respect to means (understanding actions and their effects), but with respect to ends (understanding task-related human objectives). Technical studies of cooperative inverse reinforcement learning  (Hadfield-Menell et al. 2016 ) address problems of learning through task-oriented observation, demonstration, and teaching, while Paul Christiano's work  (Christiano 2015b ) on scalable control explores (for example) how observation and human supervision could potentially be extended to challenging AI tasks while economizing on human effort. Generic predictive models of human approval can complement these approaches by providing strong priors on human objectives in performing tasks while avoiding harms. \n Deployment at scale enables aggregated experience and centralized learning Important AI services will often entail large-scale deployments that enable accelerating learning 1 from instances of success, failure, and human responses (potentially including criticism and advice). In addition to accelerating improvement across large deployments, aggregated experience and learning can increase the benefit-to-cost ratio of using and teaching systems by multiplying the system-wide value of users' correcting and advising individual systeminstances while diluting the per-user burdens of encountering correctable errors. \n Recourse to human advice will often be economical and effective Imperfect models of human approval can be supplemented and improved by recourse to human advice. Imperfect models of approval should contain more reliable models of human concern; an expectation of concern together with uncertainty regarding approval could prompt recourse (without overuse) of human advice. Using advice in learning from aggregated experience 2 would further economize the use of human attention. \n AI-enabled criticism and monitoring can strengthen oversight Concerns regarding perverse planning by advanced AI agents could potentially be addressed by applying comparable AI capabilities to AI development and supervision. In AI development, the aim would be to understand and avoiding the kinds of goals and mechanisms that could lead to such plans; in 1. Section 16: Aggregated experience and centralized learning support AI-agent applications AI applications, the aim would be to monitor plans and actions and recognize and warn of potential problems (or to intervene and forestall them). This kind of AI-enabled adversarial analysis, testing, monitoring, and correction need not be thwarted by collusion among AI systems, 1 even if these systems operate at superintelligent levels of competence. \n AI-enabled AI development could both accelerate application development and facilitate human guidance Fast AI technology improvement will increase the scope for bad choices and potentially severe risks. Established practice in system development, however, will favor a measure of intelligent caution, informed by contemporaneous experience and safety-oriented theory and practice.  2  We can expect the temptation to move quickly by accepting risks to be offset to some extent by improved support for goal-aligned function definition, system design, testing, deployment, feedback, and upgrade. It is very nearly a tautology to observe that the balanced use of powerful AI development capabilities can reduce the cost of producing safe and reliable AI products. Further, the underlying principles appear to scale to AI development technologies that enable the safe implementation of a full spectrum of AI services with superhuman-level performance. This potential, of course, by no means assures the desired outcome.  \n Further Reading \n Must pressure to accelerate AI technology development increase risk? Technical and economic objectives will continue to drive incremental yet potentially thorough automation of AI R&D. In considering asymptotically recursive automation of AI R&D, 1 it is natural to think of ongoing human involvement as a source of safety, but also of delays, and to ask whether competitive pressures to maximize speed by minimizing human involvement will incur risks. \"AI R&D\", however, embraces a range of quite different tasks, and different modes of human involvement differ in their effects on speed and safety. Understanding the situation calls for a closer examination. \n Basic technology research differs from world-oriented applications It  \n World-oriented applications bring a different range of concerns Application development and deployment will typically have direct effects on the human world, and many applications will call for iterative development with extensive human involvement. World-oriented application development operates outside the basic-technology R&D loop, placing it beyond the scope of the present discussion. Further Reading Advice optimized to produce results may be manipulative, optimized to induce a client's acceptance; advice optimized to produce results conditioned on its acceptance will be neutral in this regard. \n Summary To optimize advice to produce a result entails optimizing the advice to ensure its acceptance, and hence to manipulate the clients' choices. Advice can instead be optimized to produce a result conditioned on the advice being accepted; because the expected value of an outcome conditioned on an action is independent of the probability of the action, there is then no value in manipulating clients' choices. In an illustrative (and practical) case, a client may request advice on options that offer different trade-offs between expected costs, benefits, and risks; optimization of these options does not entail optimization to manipulate a client's choice among them. Manipulation remains a concern, however: In a competitive situation, the most popular systems may optimize advice for seduction rather than value. Absent effective counter-pressures, competition often will (as it already does) favor the deployment of AI systems that strongly manipulate human choices. \n Background (1): Classic concerns \"Oracles\"  (Bostrom 2014 ) are a proposed class of high-level AI systems that would provide answers in response to queries by clients; in the present context, oracles that provide advice on how to achieve goals are of particular interest. It has sometimes been suggested that oracles would be safer than comparable agents that act in the world directly, but because oracles inevitably affect the world through their clients' actions, the oracle/agent distinction per se can blur. To clarify this situation, it is important to consider (without claiming novelty of either argument or result) whether optimizing oracles to produce effective advice entails their optimizing advice to affect the world. \n Background (2): Development-oriented models In the RDA-process model, 1 research produces basic components and techniques, development produces functional systems, and application produces results for users. In AI development, an advisory oracle will be optimized for some purpose by a chain of systems that are each optimized for a purpose: • AI research optimizes components and techniques to enable development of diverse, effective AI systems. • Advisory-oracle development optimizes systems to suggest options for action across some range of situations and objectives. • Advisory-oracle application suggests actions optimized to achieve given objectives in specific situations. Each stage in an RDA process yields products (components, oracles, advice) optimized with respect to performance metrics (a.k.a. loss functions). \n Optimization for results favors manipulating clients' decisions Giving advice may itself be an action intended to produce results in the world. To optimize advice to produce a result, however, entails optimizing the advice to ensure that the advice is applied. In current human practice, advice is often intended to manipulate a client's behavior to achieve the advisor's objective, and a superintelligent-level AI advisor could potentially do this very well. At a minimum, an oracle that optimizes advice to produce results can be expected to distort assessments of costs, benefits, and risks to encourage fallible clients to implement supposedly \"optimal\" policies. A range of standard AI-agent safety problems (e.g., perverse instantiation and pursuit of convergent instrumental goals) then arise with full force. Optimizing oracles to produce advice intended to produce results seems like a bad idea. We want to produce oracles that are not designed to deceive. \n Optimization for results conditioned on actions does not entail optimization to manipulate clients' decisions Oracles could instead be optimized to offer advice that in turn is optimized, not to produce results, but to produce results contingent on the advice being applied. Because the expected value of an outcome conditioned on an action is independent of the probability of the action, optimal advice will not be optimized to manipulate clients' behavior. \n Oracles can suggest options with projected costs, benefits, and risks Because human beings have preferences that are not necessarily reducible to known or consistent utility functions, it will be natural to ask advisoryoracles to suggest and explain sets of options that offer different, potentially incommensurate combinations of costs, benefits, and risks; thus, useful advice need not be optimized to maximize a predefined utility function, but can instead be judged by Pareto-optimality criteria. With optimization of outcomes conditioned on acceptance, the quality of assessment of costs, benefits, and risks will be limited by AI competencies, undistorted by a conflicting objective to manipulate clients' choices among alternatives. (Note that the burden of avoiding options that perversely instantiate objectives rests on the quality of the options and their assessment, 1 not on the avoidance of choice-manipulation.) \n Competitive pressures may nonetheless favor AI systems that produce perversely appealing messages If multiple AI developers (or development systems) are in competition, and if their success is measured by demand for AI systems' outputs, then the resulting incentives are perverse: Advice that maximizes appeal will often be harmful, just as news stories that maximize attention often are false. Because AI-enabled communications will permit radical scaling of deception in pursuit of profit and power, it seems likely that human-driven applications of these capabilities will be the leading concern as we move forward in AI technology. It seems likely that effective countermeasures will likewise require AI-enabled communication that influences large audiences. \n Further Reading \n • Section 20: Collusion among superintelligent oracles can readily be avoided • Section 21: Broad world knowledge can support safe task performance • Section 22: Machine learning can develop predictive models of human approval • Section 23: AI development systems can support effective human guidance 26 Superintelligent-level systems can safely provide design and planning services performing such tasks match classic templates for emergent AI agency and risk. Nonetheless, examining design systems at the level of task requirements, component capabilities, and development processes suggests that classic AIagent risks need not arise. In design engineering, effective human oversight is not an impediment, but a source of value. Superintelligent-level services can help solve (rather than create) AI-control problems; for example, strong models of human concerns and (dis)approval can be exploited to augment direct human oversight. \n Design engineering is a concrete example of a planning task Planning tasks relate means to ends, and systems-level design engineering offers an illustrative example. Systems engineering tasks are characterized by complex physical and and causal structures that often involve complex and critical interactions with human concerns. As with most planning tasks, system-level design is intended to optimize the application of bounded means (finite materials, time, costs. . . ) to achieve ends that are themselves bounded in space, time, and value. \n AI-based design systems match classic templates for emergent AI-agent risk Classic models of potentially catastrophic AI risk involve the emergence (whether by design or accident) of superintelligent AI systems that pursue goals in the world. Some implementations of question-answering systems (\"oracles\") could present dangers through their potential ability to recruit human beings to serve as unwitting tools (Superintelligence, Bostrom 2014); hazardous characteristics would include powerful capabilities for modeling the external world, formulating plans, and communicating with human beings. Nonetheless, we can anticipate great demand for AI systems that have all of these characteristics. To be effective, AI-enabled design systems should be able to discuss what we want to build, explore candidate designs, assess their expected performance, and output and explain proposals. The classic model proposes that these tasks be performed by a system with artificial general intelligence (an AGI agent) that, in response to human requests, will seek to optimize a corresponding utility function over states of the world. Because a fully-general AGI agent could by definition perform absolutely any task with superhuman competence, it requires no further thought to conclude that such an agent could provide engineering design services. \n High-level design tasks comprise distinct non-agent-like subtasks 9 diagrams an abstract, high-level task structure for the development and application of AI-enabled design systems. In this architecture: • A top-level conversational interface translates between human language (together with sketches, gestures, references to previous designs, etc.) and abstract yet informal conceptual descriptions. • A second level translates between informal conceptual descriptions and formal technical specifications, supporting iterative definition and refinement of objectives, constraints, and general design approach. General AI, etc. : Comprehensive AI Services (CAIS) • The core of the design process operates by iterative generate-and-test, formulating, simulating, and scoring candidate designs with respect to objectives and constraints (including constraints that are tacit and general). • To enable systems to build on previous results, novel designs can be abstracted, indexed, and cached in a shared library. • Upstream from design tasks, the development and upgrade of AIenabled design systems is itself a product of AI-enabled design that integrates progress in basic AI technologies with domain-specific application experience. • Downstream from design tasks, design products that pass (AIsupported) screening and comparison with competing designs may be deployed and applied, generating application experience that can inform future design. In considering the task structure outlined in 9, is important to recognize that humans (today's agents with general intelligence) organize system design tasks in the same general way. The system shown in 9 is not more complex than a black-box AGI agent that has acquired engineering competencies; instead, it makes explicit the kinds of tasks that must implemented, regardless of how those tasks might be implemented and packaged. Hiding requirements in a black box does not make them go away. \n Real-world task structures favor finer-grained task decomposition As an aside, we should expect components of engineering systems to be more specialized than those diagrammed above: An any but the most abstract levels, design methods for integrated-circuit design are distinct from methods for aerospace structural engineering, organic synthesis, or AI system architecture, and methods for architecting systems built of diverse subsystems are substantially different from all of these. The degree of integration of components will be a matter of convenience, responsive to considerations that include the value of modularity in fault isolation and functional transparency. \n Use of task-oriented components minimizes or avoids classic AI risks Consider the flow of optimization and selection pressures implicit in the architecture sketched in 9: • Systems for basic AI R&D are optimized and selected to produce diverse, high-performance tools (algorithms, generic building blocks. . . ) to be used by AI systems that develop AI systems. Off-task activities will incur efficiency costs, and hence will be disfavored by (potentially superintelligent) optimization and selection pressures. • AI systems that develop AI systems are optimized and selected to produce components and architectures that perform well in applications (here, engineering design). As with basic R&D, off-task activities will incur efficiency costs, and hence will be disfavored. • All systems, including system-design systems, consist of stable, taskfocused components subject to upgrade based on aggregated application experience. Note that none of these components has a task that includes optimizing either its own structure or a utility function over states of the world. 26.7 Effective human oversight is not an impediment, but a source of value The role of human oversight it to get what people want, and as such, identifying and satisfying human desires is not an impediment to design, but part of the design process. Comprehensive AI services can include the service of supporting effective human oversight, however, as discussed in safety-oriented requirements for human oversight of basic AI R&D (algorithms, architectures, etc.) are minimal, 1 and need not slow progress. In design, as in other applications of AI technologies, effective human oversight is not enough to avoid enormous problems, because even systems that provide what people think they want can have adverse outcomes. Perversely seductive behaviors could serve the purposes of bad actors, or could arise through competition to develop systems that gain market share (consider the familiar drive to produce news that goes viral regardless of truth, and foods that stimulate appetite regardless of health). \n SI-level systems could solve more AI-control problems than they create We want intelligent systems that help solve important problems, and should consider how superintelligent-level competencies could be applied to solve problems arising from superintelligence. There is no barrier to using AI to help solve problems of AI control: Deceptive collusion among intelligent problem-solving systems would require peculiar and fragile preconditions.  2  (The popular appeal of \"us vs. them\" framings of AI control is perhaps best understood as a kind of anthropomorphic tribalism.) \n Models of human concerns and (dis)approval can augment direct oversight In guiding design, key resources will be language comprehension and modeling anticipated human (dis)approval. 3 Among the points of potential leverage: • Strong priors on human concerns to ensure that important considerations are not overlooked. • Strong priors on human approval to ensure that standard useful features are included by default. • Strong priors on human disapproval to ensure that options with predictable but excessively negative unintended effects are dropped. • Building on the above, effective elicitation of human intentions and preferences through interactive questioning and explanation of design options. • Thorough exploration and reporting (to users and regulators) of potential risks, failure modes, and perverse consequences of a proposal. • Ongoing monitoring of deployed systems to track unanticipated behaviors, failures, and perverse consequences. \n The pursuit of superintelligent-level AI design services need not entail classic AI-agent risks To understand alternatives to superintelligent-AGI-agent models, it is best to start with fundamentals-intelligence as problem solving capacity, problems as tasks, AI systems as products of development, and recursive improvement as a process centered on technologies rather than agents. Interactive design tasks provide a natural model of superintelligent-level, real-world problem solving, and within this framework, classic AI-agent problems arise either in bounded contexts, or as a consequence of reckless choices in AI-system development. Further Reading  \n Summary In a range of competitive, tactical-level tasks (e.g., missile guidance and financial trading), potential advantages in decision speed and quality will tend to favor direct AI control of actions. In high-level strategic decisions, however-where stakes are higher, urgency is reduced, and criteria may be ambiguous-humans can exploit AI competence without ceding control: If AI systems can make excellent decisions, then they can suggest excellent options. Human choice among strategic options need not impede swift response to events, because even long-term strategies (e.g., U.S. nuclear strategy) can include prompt responses of any magnitude. In light of these considerations, we can expect senior human decision makers to choose to retain their authority, and without necessarily sacrificing competitiveness. \n Pressures for speed and quality can favor AI control of decisions When AI systems outperform humans in making decisions (weighing both speed and quality), competitive situations will drive humans to implement AI-controlled decision processes. The benefits of speed and quality will differ in different applications, however, as will the costs of error. \n Speed is often critical in selecting and executing \"tactical\" actions Decision-speed can be critical: Computational systems outperform humans in response time when controlling vehicles and launching defensive missiles, offering crucial advantages, and military planners foresee increasing pressures use AI-directed systems in high-tempo tactical exchanges. In tactical military situations, as in high-frequency financial trading, failure to exploit the advantages of AI control may lead to losses. \n Quality is more important than speed in strategic planning High-level strategic plans almost by definition guide actions extended in time. Even long-term strategies may of course be punctuated by prompt, conditional actions with large-scale consequences: U.S. nuclear strategy, for example, has been planned and revised over years and decades, yet contemplates swift and overwhelming nuclear counterstrikes. Fast response to events is compatible with deliberation in choosing and updating strategies. \n System that can make excellent decisions could suggest excellent options Superior reasoning and information integration may enable AI systems to identify strategic options with superhuman speed and quality, yet this need not translate into a pressure for humans to cede strategic control. If AI systems can make excellent decisions, then they can suggest excellent sets of options for consideration by human decision makers. Note that developing sets of options does not imply a commitment to a utility function over world states; given uncertainties regarding human preferences, it is more appropriate to apply Pareto criteria to potentially incommensurate costs, benefits, and risks, and to offer proposals that need not be optimized to induce particular choices.  1  In this connection, it is also important to remember that long-term strategies are normally subject to ongoing revision in light of changing circumstances and preferences, and hence adopting a long-term strategy need not entail long-term commitments. \n Human choice among strategies does not preclude swift response to change Profoundly surprising events that call for superhumanly-swift, large-scale, unanticipated strategic reconsideration and response seem likely to be rare, particularly in hypothetical futures in which decision makers have made effective use of superintelligent-quality strategic advice. Further, human choice among strategic options need not be slow: Under time pressure, a decision maker could scan a menu of presumably excellent options and make a quick, gut choice. Beyond this, human-approved strategies could explicitly allow for great flexibility under extraordinary circumstances. In light of these considerations, substantial incentives for routine relinquishment of high-level strategic control seem unlikely. \n Senior human decision makers will likely choose to retain their authority Absent high confidence that AI systems will consistently make choices aligned with human preferences, ceding human control of increasingly high-level decisions would incur increasing risks for declining (and ultimately slight) benefits. As a practical matter, we can expect senior human decision makers to choose to retain their authority unless forcibly dislodged, perhaps because they have lost the trust of other, more powerful human decision makers. Further Reading  \n Summary It has been suggested that assigning AGI agents tasks as broad as biomedical research (e.g., \"curing cancer\") would pose difficult problems of AI control and value alignment, yet a concrete, development-oriented perspective suggests that problems of general value alignment can be avoided. In a natural path forward, diverse AI systems would automate and coordinate diverse biomedical research tasks, while human oversight would be augmented by AI tools, including predictive models of human approval. Because strong task alignment does not require formal task specification, a range of difficult problems need not arise. In light of alternative approaches to providing general AI services, there are no obvious advantages to employing risky, general-purpose AGI agents to perform even extraordinarily broad tasks. \n Broad, unitary tasks could present broad problems of value alignment \"The Value Learning Problem\" (Soares 2018) opens with an example of a potential AGI problem: 1 Consider a superintelligent system, in the sense of  Bostrom (2014) , tasked with curing cancer  [. . .]  without causing harm to the human (no easy task to specify in its own right). The resulting behavior may be quite unsatisfactory. Among the behaviors not ruled out by this goal specification are stealing resources, proliferating robotic laboratories at the expense of the biosphere, and kidnapping human test subjects. The example, typically draws on many distinct areas of competence that support concept development, experimentation, and data analysis. At a higher level of research organization, project management and resource allocation call for assessment of competing proposals 1 in light of not only their technical merits, but of their costs and benefits to human beings. Progress in implementing AI systems that provide diverse superhuman competencies could enable automation of a full range of technical and managerial tasks, and because AI R&D is itself subject to automation, 2 progress could be incremental, yet swift. By contrast, it is difficult to envision a development path in which AI developers would treat all aspects of biomedical research (or even cancer research) as a single task to be learned and implemented by a generic system. Prospects for radical improvements in physical tools (e.g., through molecular systems engineering) do not change this general picture. \n Human oversight can be supported by AI tools Within the scope of biomedical research, several areas-resource investments, in-vivo experimentation, and clinical applications-call for strong human oversight, and oversight is typically mandated not only by ethical concerns, but by law, regulation, and institutional rules. Human oversight is not optional (it is part of the task), yet AI applications could potentially make human oversight more effective. For example, systems that describe research proposals 3 in terms of their anticipated human consequences would enable human oversight of complex research plans, while robust predictive models of human concerns 4 could be applied to focus scarce human attention. Consultation with superintelligent-level advisors could presumable enable extraordinarily well-informed judgments by patients and physicians. \n Strong task alignment does not require formal task specification The example of language translation shows that task alignment need not require formal task specification, 5 and a development-oriented perspective on concrete biomedical tasks suggests that this property may generalize quite widely. Although it is easy to envision both safe and unsafe configurations of task-performing systems, it is reasonable to expect that ongoing AI safety research (both theoretical and informed by ongoing experience) can enable thorough automation of research while avoiding both unacceptable costs and extraordinary risks stemming from emergent behaviors. If safety need not greatly impede development, then unsafe development is best viewed as a bad-actor risk. 28.6 The advantages of assigning broad, unitary tasks to AGI agents are questionable It has been persuasively argued  (Bostrom 2014 ) that self-improving, generalpurpose AGI agents cannot safely be tasked with broad goals, or with seemingly narrow goals that might motivate catastrophic actions. If a full range of superintelligent-level AI capabilities can be provided efficiently by other means, 1 then the advantages of developing risky AGI agents are questionable. The argument that access to general, superintelligent-level AI capabilities need not incur the risks of superintelligent AGI agents includes the following points: • Recursive technology improvement 2 is a natural extension of current AI R&D, and does not entail recursive self improvement of distinct AI systems: Agents are products, not development tools. • Effective human oversight 3 need not substantially impede recursive improvement of basic AI technologies, while overseeing the development of task-focused AI systems is similar to (but less risky than) specifying tasks for an AGI system. • Models of human approval 4 can inform AI plans in bounded domains, while the use of AI systems to examine the scope and effects of proposed plans (in an implicitly adversarial architecture 5 ) scales to superintelligent proposers and critics. It seems that there are no clear technical advantages to pursuing an AGI-agent approach to biomedical research, while a task-focused approach provides a Potential AGI agents should be considered in the context of a world that will (or readily could) have prior access to general intelligence in the form of comprehensive AI services. \n Summary Discussions of SI-level AI technologies and risks have centered on scenarios in which humans confront AGI agents in a world that lacks other, more tractable SI-level AI resources. There is, however, good reason to expect that humans will (or readily could) have access to comprehensive, open-ended AI services before AGI-agent systems are implemented. The open-ended AIservices model of artificial general intelligence does not preclude (and in fact would facilitate) the implementation of AGI agents, but suggests that AI risks, and their intersection with the ethics of computational persons, should be reexamined in the context of an AI milieu that can provide SI-level strategic advice and security services. \n It has been common to envision AGI agents in a weak-AI context In classic AGI-risk scenarios, advanced AI capabilities emerge in the form of AGI agents that undergo recursive, transformative self-improvement to a superintelligent (SI) level; these agents then gain capabilities beyond those of both human beings and human civilization. Studies and discussions in this conceptual framework propose that, when confronted with AGI agents, humans would lack prior access to tractable SI-level problem-solving capabilities. \n Broad, SI-level services will (or readily could) precede SI-level AI agents The technologies required to implement or approximate recursive AI technology improvement are likely to emerge through heterogeneous AI-facilitated R&D mechanisms, 1 rather than being packaged inside a discrete entity or agent. Accordingly, capabilities that could in principle be applied to implement SI-level AGI agents could instead 2 be applied to implement general, comprehensive AI services 3 (CAIS), including stable, task-focused agents.  4  In this model, directly-applied AI services are distinct from services that develop AI services, 5 an approach that reflects natural task structures 6 and has great practical advantages. 7 \n SI-level services will enable the implementation of AGI agents Although recursive technology improvement will most readily be developed by means of heterogeneous, non-agent systems, any AI milieu that supports \"comprehensive AI services\" could (absent imposed constraints) provide the service of implementing SI-level AGI agents. This prospect diverges from classic AGI-agent risk scenarios, however, in that a strong, pre-existing AI milieu could be applied to implement SI-level advisory and security services.  \n SI-level advisory and security services could limit AGI-agent risks The prospect of access SI-level advisory and security services fundamentally changes the strategic landscape around classic AI-safety problems. For example, \"superpowers\" as defined by  Bostrom (2014)  do not exist in a world in which agents lack radically-asymmetric capabilities. Further, arguments that SI-level agents would collude (and potentially provide collectively deceptive advice), do not carry over to SI-level systems in a CAIS milieu, 1 advice can be objective, rather than manipulative, 2 predictive models of human preferences and concerns 3 can improve the alignment of actions with human intentions in performing well-bounded tasks.  4  Arguments that competitive and security pressures would call for ceding strategic control to AI systems are surprisingly weak: Tactical situations may call for responses of SI-level quality and speed, but SI-level advisory and security services could support strategic choices among excellent options, deliberated at a human pace.  5  Crucially, in a range of potential defense/offense scenarios, the requisite security systems could be endowed with arbitrarily large advantages in resources for strategic analysis, tactical planning, intelligence collection, effector deployment, and actions taken to preclude or respond to potential threats. In choosing among options in the security domain, humans would likely prefer systems that are both reliable and unobtrusive. Many AI safety strategies have been examined to date, and all have difficulties; it would be useful to explore ways in which tractable SI-level problemsolving capabilities could be applied to address those difficulties. In exploring potential responses to future threats, it is appropriate to consider potential applications of future capabilities. 1. Section 20: Collusion among superintelligent oracles can readily be avoided 2. Section 25: Optimized advice need not be optimized to induce its acceptance 3. Section 22: Machine learning can develop predictive models of human approval 4. Section 16: Aggregated experience and centralized learning support AI-agent applications 5. Section 27: Competitive pressures provide little incentive to transfer strategic decisions to AI systems 29.6 SI-level capabilities could mitigate tensions between security concerns and ethical treatment of non-human persons The spectrum of potential AGI systems includes agents that should, from a moral perspective be regarded as persons and treated accordingly. Indeed, the spectrum of potential computational persons includes emulations of generic or specific human beings 1 . To fail to treat such entities as persons would, at the very least, incur risks of inadvertently committing grave harm. It has sometimes been suggested that security in a world with SI-level AI would require stunting, \"enslaving\", or precluding the existence of computational persons. The prospect of robust, SI-level security services, however, suggests that conventional and computational persons could coexist within a framework stabilized by the enforcement of effective yet minimally-restrictive law.  Bostrom's (2014, p.201-208 ) concept of \"mind crime\" presents what are perhaps the most difficult moral questions raised by the prospect of computational persons. In this connection, SI-level assistance may be essential not only to prevent, but to understand the very nature and scope of potential harms to persons unlike ourselves. Fortunately, there is seemingly great scope for employing SI-level capabilities while avoiding potential mindcrime, because computational systems that provide high-order problem-solving services need not be equivalent to minds. 2 29.7 Prospects for superintelligence should be considered in the context of an SI-level AI services milieu The prospect of access to tractable, SI-level capabilities reframes the strategic landscape around the emergence of advanced AI. In this connection, it will be important to reexamine classic problems of AI safety and strategy, not only in the context of an eventual SI-level AI services milieu, but along potential paths forward from today's AI technologies. \n Further Reading • Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame Complex, unitary, untrusted AI systems could be leveraged to produce non-problematic general-learning kernels through competitive optimization for objectives that heavily weight minimizing description length. \n Summary In a familiar AGI threat model, opaque, self-improving AI systems give rise to systems that incorporate wide-ranging information, learn unknown objectives, and could potentially plan to pursue dangerous goals. The R&Dautomation/AI-services model suggests that technologies that could enable such systems would first be harnessed to more prosaic development paths, but what if powerful AI-development capabilities were deeply entangled with opaque, untrusted systems? In this event, conceptually straightforward methods could be employed to harness untrusted systems to the implementation of general learning systems that lack problematic information and purposes. Key affordances include re-running development from early checkpoints, and applying optimization pressure with competition among diverse systems to produce compact \"learning kernels\". Thus, from a path that could lead to problematic AGI agents, a readily accessible off-ramp leads instead to general intelligence in the form of capabilities that enable open-ended AI-service development. (These and related topics have been explored in an earlier form, but in greater depth, in  FHI Technical Report 2015 -3 (Drexler 2015 ).) 30.2 A familiar threat model posits opaque, self-improving, untrusted AI systems In a familiar AGI threat model, the pursuit of general, superintelligent level AI leads to opaque, self-improving systems with ill-characterized information content, inference capabilities, planning capabilities, and goals. Strong arguments suggest that the use of such systems could pose grave risks  (Bostrom 2014) . The present analysis will consider a near-worst-case scenario in which AI development stands at the threshold of building such systems, rather than considering how such a situation could be avoided. Further, it will consider the hard case in which \"general intelligence\" is an indivisible property, a capacity to learn more-or-less anything that a human can, and potentially much more. Crucially, the analysis will neither equate intelligence as learning capacity with intelligence as competence 1 nor assume that a product system must inherit the full information content of the producing system. Finally, it will assume that researchers retain copies of precursors of systems of interest. 30.3 Open-ended \"self-improvement\" implies strong, general AI implementation capabilities \"Self improvement\" implies strong, general AI implementation capabilities, yet in considering a chain of improved implementations, the concept of \"self\" is at best ambiguous, and at worst is an anthropomorphic distraction.  2  Operationally, \"self\" improvement implies an opaque system that is capable of implementing systems that are better than itself by some metric, and in particular, is capable of implementing systems that are improved in the sense of being better at implementing improved systems (etc.). In the proposed AGI threat model, some earlier, non-problematic system was capable of serving as a link in such a chain (given suitable machine resources, training data, simulated environments, and tasks) and the actual development history from that point led to a problematic result. \n Successful system development can be recapitulated with variations In general, results in machine learning can be repeated with variations in architectures, training data, simulated environments, training objectives, etc. In the present instance, systems with the potential to give rise to a sequence of improved systems are assumed to be architecturally opaque; nonetheless, external affordances (training data, simulated environments, training objectives, resource-optimization pressures, etc.) remain available. Development of powerful systems can be recapitulated with variations induced by external affordances, and these affordances can strongly affect the content of what is learned. If the desired core functionality is development of learning systems, it is likely that relatively abstract problem spaces will sufficient or optimal. In addition, learning to optimize systems for learning a task does not require access to detailed task-level training data or environments, and optimizing systems for the task of optimizing architectures for systems that learn a task is even more abstract and remote from object-level training information. Note that tasks at all levels are by nature bounded with respect to time and resources, and hence do not naturally engender convergent instrumental goals.  1  In research and development, different versions of systems are always in implicit competition with one another to maximize performance on some bounded task: Those that perform poorly by relevant metrics will be set aside, while versions that perform well will be used to produce (or serve as prototypes for) next-generation systems. Thus, the convergent goal of systems under development is competition to perform bounded tasks, and by the orthogonality thesis  (Bostrom 2014) , the pursuit of bounded goals can employ arbitrarily high intelligence. In aggregate, such systems will (or readily could) satisfy conditions that exclude collusion.  2   \n Optimization can favor the production of compact, general learning kernels The scheme outlined here centers on competitive optimization of \"learning kernels\" for \"compactness\", where a learning kernel is a system that, in conjunction with computational resources, auxiliary components, and a set of benchmark \"demo tasks\", can produce an expanded set of systems that perform those tasks at some given level of performance. Here, optimizing for compactness implies minimizing a weighted sum of the kernel's description length (operationally, the number of bits in a string that decompresses and compiles to build the system) and metrics corresponding to the resources it consumes in expansion; in the present context, compactness is assumed to give heavy weight to description length, penalizing resource consumption only to exclude undue computational costs. If kernel expansion can produce a range of systems that perform each of a sufficiently broad range of challenging demo tasks, then the learning kernel can be considered general. Examples of demo-task domains might include: • Language translation • Visual perception • Robotic control • Spacecraft design • Chemical synthesis • Strategic planning • Theorem proving • Software development Note that current practice in AI research does in fact iteratively develop, train, and improve relatively compact learning kernels (architectures and algorithms) that, through training, expand their information content to produce systems that perform tasks in particular domains.  \n Competitive optimization for compactness can exclude problematic information and competencies Competition for learning-kernel compactness, subject to retaining generality, will strongly exclude potentially-problematic information that does not contribute to the process of learning target competencies. A world-beating Go system need not \"know\" that it is \"playing a game\"; a high-quality machine translation system need not \"know\" that its outputs are \"requested by human beings\"; a machine vision system need not \"know\" that its capabilities enable \"autonomous vehicle control\". A compact, general kernel that can generate Go-learning, language-learning, and vision-learning systems can (and under optimization, must) \"know\" even less about concrete tasks, domains, and intended applications in a \"human world\". In the standard AGI threat model, self-improvement (which is to say, AI implementation capability) is capable of producing systems that operate at a superintelligent level. Systems at this level that are optimized to optimize general learning kernels for compactness presumably can be quite effective at stripping out information that is unnecessary for task performance. Given that post-bottleneck training data can provide all domain-relevant information, together with auxiliary resources such as efficient numerical algorithms, effective optimization will strip out virtually all world knowledge (geography, history, vocabulary, chemistry, biology, physics. . . ), including any plausible basis for problematic plans and concrete world-oriented competencies. This conclusion holds even if the kernel-implementation systems might be untrustworthy if directly applied to world-oriented tasks. \n Exclusion of problematic content can provide a safe basis for developing general capabilities In the familiar AGI threat model, development results in opaque, selfimproving, ill-characterized, but highly-capable systems, and-crucially-the use of these capabilities is assumed to require that the problematic systems themselves be applied to a wide range of world tasks. This assumption is incorrect. As argued above, posited self-improvement capabilities could instead be re-developed from earlier, non-problematic systems through a process that leads to diverse systems that compete to perform tasks that include AI system development. The AI-implementation capacity of such systems could then be applied to the development of compact general-learning kernels that will omit representations of problematic knowledge and goals. This strategy is technology-agnostic: It is compatible with neural, symbolic, and mixed systems, whether classical or quantum mechanical; it assumes complete implementation opacity, and relies only on optimization pressures directed by external affordances. Because expansion of a general learning kernel would in effect implement the \"research\" end of AI R&D automation, the strategy outlined above could provide a clean basis for any of a range of development objectives, whether in an AGI-agent or CAIS model of general intelligence. This approach is intended to address a particular class of scenarios in which development has led the edge of a cliff, and is offered as an example, not as a prescription: Many variations would lead to similar results, and ongoing application of the underlying principles would avoid classic threat models from the start. These safety-relevant principles are closely aligned with current practice in AI system development.  1   \n Further Reading • Section 1: R&D automation provides the most direct path to an intelligence explosion • Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame • Section 8: Strong optimization can strongly constrain AI capabilities, behavior, and effects • Section 9: Opaque algorithms are compatible with functional transparency and control • Section 10: R&D automation dissociates recursive improvement from AI agency • Section 11: Potential AGI-enabling technologies also enable comprehensive AI services • Section 36: Desiderata and directions for interim AI safety guidelines 31 Supercapabilities do not entail \"superpowers\" By definition, any given AI system can have \"cognitive superpowers\" only if others do not, hence (strategic) superpowers should be clearly distinguished from (technological) supercapabilities. \n Summary Superintelligence  (Bostrom 2014)  develops the concept of \"cognitive superpowers\" that potentially include intelligence amplification, economic production, technology development, strategic planning, software hacking, or social manipulation. These \"superpowers\", however, are defined in terms of potential strategic advantage, such that \"at most one agent can possess a particular superpower at any given time\". Accordingly, in discussing AI strategy, we must take care not to confuse situation-dependent superpowers with technologydependent supercapabilities. \n AI-enabled capabilities could provide decisive strategic advantages Application of AI capabilities to AI R&D 1 could potentially enable swift intelligence amplification and open a large capability gap between first-and second-place contenders in an AI development race. Strong, asymmetric capabilities in strategically critical tasks (economic production, technology development, strategic planning, software hacking, or social manipulation) could then provide decisive strategic advantages in shaping world outcomes  (Bostrom 2014, p.91-104) . 31.3 Superpowers must not be confused with supercapabilities [. . .] superpowers  [. . .]  are possessed by an agent as superpowers only if the agent's capabilities in these areas substantially exceed the combined capabilities of the rest of the global civilization [hence] at most one agent can possess a particular superpower at any given time. To avoid confusion, it is important to distinguish between strategically relevant capabilities far beyond those of contemporaneous, potentially superintelligent competitors (\"superpowers\"), and capabilities that are (merely) enormous by present standards (\"supercapabilities\"). Supercapabilities are robust consequences of superintelligence, while superpowers-as definedare consequences of supercapabilities in conjunction with a situation that may or may not arise: strategic dominance enabled by strongly asymmetric capabilities. In discussing AI strategy, we must take care not to confuse prospective technological capabilities with outcomes that are path-dependent and potentially subject to choice. \n Further Reading • Section 20: Collusion among superintelligent oracles can readily be avoided • Section 32: Unaligned superintelligent agents need not threaten world stability 32 Unaligned superintelligent agents need not threaten world stability A well-prepared world, able to deploy extensive, superintelligent-level security resources, need not be vulnerable to subsequent takeover by superintelligent agents. \n Summary It is often taken for granted that unaligned superintelligent-level agents could amass great power and dominate the world by physical means, not necessarily to human advantage. Several considerations suggest that, with suitable preparation, this outcome could be avoided: • Powerful SI-level capabilities can precede AGI agents. • SI-level capabilities could be applied to strengthen defensive stability. • Unopposed preparation enables strong defensive capabilities. • Strong defensive capabilities can constrain problematic agents. In other words, applying SI-level capabilities to ensure strategic stability could enable us to coexist with SI-level agents that do not share our values. The present analysis outlines general prospects for an AI-stable world, but necessarily raises more questions than it can explore. \n General, SI-level capabilities can precede AGI agents As has been argued elsewhere, the R&D-automation/AI-services model of recursive improvement and AI applications challenges the assumption 1 that the pursuit of general, SI-level AI capabilities naturally or necessarily leads to classic AGI agents. Today, we see increasingly automated AI R&D applied to the development of AI services, and this pattern will (or readily could) scale to comprehensive, SI-level AI services that include the service of developing new services. By the orthogonality thesis  (Bostrom 2014 ), high-level AI services could be applied to more-or-less any range of tasks. \n SI-level capabilities could be applied to strengthen defensive stability World order today-from neighborhood safety to the national security-is imperfectly implemented through a range of defensive services, e.g., local surveillance, self-defense, and police; military intelligence, arms control, and defensive weapon systems. A leading strategic problem today is the offensive potential of nominally defensive systems (deterrence, for example, relies on offensive weapons), engendering the classic \"security dilemma\" and consequent arms-race dynamics. Bracketing thorny, path-dependent questions of human perceptions, preferences, objectives, opportunities, and actions, one can envision a state of the world in which SI-level competencies have been applied to implement impartial, genuinely defensive security services. Desirable implementation steps and systemic characteristics would include: In other words, one can envision an AI-stable world in which well-prepared, SIlevel systems are applied to implement services that ensure physical security regardless of the preferences of unaligned or hostile actors. (Note that this does not presuppose a solution to AGI alignment: AI-supported design and implementation 1 of policies for security services 2 need not be equivalent to utility maximization by an AGI agent. 3 ) \n Unopposed preparation enables strong defensive capabilities A background assumption in this discussion is that, given access to SI-level capabilities, potentially enormous resources (indeed, literally astronomical) could be mobilized to achieve critical civilizational goals that include AGIcompatible strategic stability. In other words, we can expect civilizations as a whole to pursue convergent instrumental goals  (Bostrom 2014, p.109) , and to apply the resulting capabilities. In this connection, recall that what  Bostrom (2014, p.93)  terms \"superpowers\" are contextual, being properties not of agents per se, 4 but of agents that have an effective monopoly on the capability in question  (Bostrom 2014, p.104 ): In a prepared world, mere superintelligence would not confer superpowers. (Regarding the \"hacking superpower\"  (Bostrom 2014, p.94) , note that, even today, practical operating systems can provide mathematically provable, hence unhackable, security guarantees. 5 ) \n Strong defensive capabilities can constrain problematic agents The above points offer only an abstract sketch of a development process and objective, not a map of a road or a destination. A closer look can help to clarify key concepts: 1. Section 26: Superintelligent-level systems can safely provide design and planning services 2. Section 27: Competitive pressures provide little incentive to transfer strategic decisions to AI systems 3. Section 23: AI development systems can support effective human guidance 4. Section 31: Supercapabilities do not entail \"superpowers\" 5. https://sel4.systems/ 1) Preparatory, SI-level red-team/blue-team design competition can explore potential attacks while exploring the conditions necessary for security services to block attack capabilities with an ample margin of safety. Adversarial exercises could readily employ physical simulations that are qualitatively biased to favor hypothetical attackers, while assigning arbitrarily large, highlyasymmetric quantitative advantages to proposed security services. As noted above, enormous resources could potentially be mobilized to support SI-level exploration of hypothetical red-team threats and proposed blue-team security measures; thorough exploration would call for a good working approximation to what  Bostrom (2014, p.229)  terms \"technological completion\", at least in a design sense. 2) Anticipatory deployment of well-resourced, SI-level security services would implement systems that reflect the results of stringent red-team/bluecompetitions, and hence would employ more-than-adequate physical and computational resources. Note that preparatory, selective development and deployment of security systems strongly embodies what  Bostrom (2014, p.230)  terms \"differential technology development\". 3) Ongoing application of effective, physically-oriented surveillance calls for collection of information sufficient to establish reliable (yet not excessively conservative) upper bounds on the scope of potentially threatening physical capabilities of potentially untrustworthy actors. Recognition of threats can be informed by risk-averse generalizations of worst-case red-team strategies. 4) Ongoing application of effective, physically-oriented security measures calls for the application of ample (yet not unnecessarily conservative) resources to forestall potential threats; policies can be informed by amply (yet not excessively conservative) risk-averse generalizations of robust blue-team security measures. Crude security measures might require either strong interventions or stringent constraints on actors' physical resources; welldesigned security measures could presumably employ milder interventions and constraints, optimized for situation-dependent acceptability conditioned on global effectiveness. 32.6 This brief analysis necessarily raises more questions than it can explore The concept of impartial and effective AI-enabled security services raises questions regarding the deep underpinnings of a desirable civilizational order, questions that cannot be explored without raising raising further questions at levels that range from security policies and physical enforcement to the entrenchment of constitutional orders and the potential diversity of coexisting frameworks of law. Prospects for a transition to a secure, AI-stable world raise further questions regarding potential paths forward, questions that involve not only technological developments, but ways in which the perceived interests and options of powerful, risk-averse actors might align well enough to shape actions that lead to widely-approved outcomes. 32.7 A familiar alternative scenario, global control by a value-aligned AGI agent, presents several difficulties Discussions of superintelligence and AI safety often envision the development of an extremely powerful AI agent that will take control of the world and optimize the future in accord with human values. This scenario presents several difficulties: It seems impossible to define human values in a way that would be generally accepted, impossible to implement systems that would be trusted to optimize the world, and difficult to take control of the world (whether openly or by attempted stealth) without provoking effective, preemptive opposition from powerful actors. Fortunately, as outlined above, the foundational safety challenge-physical security-can be addressed while avoiding these problems. Further Reading 33 Competitive AI capabilities will not be boxed Because the world's aggregate AI capacity will greatly exceed that of any single system, the classic \"AI confinement\" challenge (with AI in a box and humans outside) is better regarded as an idealization than as a concrete problem situation. \n Summary Current trends suggest that superintelligent-level AI capabilities will emerge from a distributed, increasingly automated process of AI research and development, and it is difficult to envision a scenario in which a predominant portion of overall AI capacity would (or could) emerge and be confined in \"a box\". Individual systems could be highly problematic, but we should expect that AI systems will exist in a milieu that enables instantiation of diverse peer-level systems, a capability that affords scalable, potentially effective mechanisms for managing threatening AI capabilities. \n SI-level capabilities will likely emerge from incremental R&D automation Current trends in machine learning point to an increasing range of superhuman capabilities emerging from extensions of today's technology base that emerge from extensions of today's R&D milieu. Today we see a distributed, increasingly automated R&D process that employs a heterogeneous toolset to develop diverse demonstration, prototype, and application systems. Increasingly, we find that AI applications include AI development tools, pointing the way toward thorough automation of development that would enable AI progress at AI speed-an incremental model of asymptotically recursive technology improvement. 1 \n 33.3 We can expect AI R&D capacity to be distributed widely, beyond any \"box\" Current methods in machine learning suggest that access to large-scale machine resources will be critical to competitive performance in AI technology and applications development. The growing diversity of technologies and applications (e.g., to vision, speech recognition, language translation, ML architecture design. . . ) speak against the idea that a world-class R&D process (or its functional equivalent) will someday be embodied a distinct, generalpurpose system. In other words, it is difficult to envision plausible scenarios in which we find more AI capability \"in a box\" than in the world outside. 33.4 AI systems will be instantiated together with diverse peer-level systems We should expect that any particular AI system will be embedded in an extended AI R&D ecosystem having aggregate capabilities that exceed its own. Any particular AI architecture will be a piece of software that can be trained and run an indefinite number of times, providing multiple instantiations that serve a wide range of purposes (a very wide range of purposes, if we posit truly general learning algorithms). As is true today, we can expect that the basic algorithms and implementation techniques that constitute any particular architecture will be deployed in diverse configurations, trained on diverse data, and provide diverse services.  1  33.5 The ability to instantiate diverse, highly-capable systems presents both risks and opportunities for AI safety Absent systemic constraints, advanced AI technologies will enable the implementation of systems that are radically unsafe or serve abhorrent purposes. These prospects can be classed as bad-actor risks that, in this framing, include actions that incur classic AGI-agent risks. The almost unavoidable ability to instantiate diverse AI systems at any given level of technology also offers benefits for AI reliability and safety. In particular, the ability to instantiate diverse peer-level AI systems enables the use of architectures that rely on implicitly competitive and adversarial relationships among AI components, an approach that enables the use of AI systems to manage other AI systems while avoiding concerns regarding potential collusion.  2  Both competitive and adversarial mechanisms are found in current AI practice, and scale to a superintelligent level.  \n Further Reading \n The R&D automation model is compatible with decentralized development State-of-the-art AI research and development is currently decentralized, distributed across independent groups that operate within a range of primarily corporate and academic institutions. Continued automation of AI R&D tasks 1 will likely increase the advantages provided by proprietary tool-sets, integrated systems, and large-scale corporate resources, yet strong automation is compatible with continued decentralization. \n Accelerating progress could lead to strong centralization of capabilities Steeply accelerating progress, if driven by proprietary, rapidly-advancing tool sets, could favor the emergence of wide gaps between competing groups, effectively centralizing strong capabilities in a leader. \n Centralization does not imply a qualitative change in R&D tasks Pressures that favor centralization neither force nor strongly favor a qualitative change in the architecture of R&D tasks or their automation. Organizational centralization, tool-chain integration, and task architecture are distinct considerations, and only loosely coupled. \n Centralization and decentralization provide differing affordances relevant to AI policy and strategy Considerations involving AI policy and strategy may favor centralization of strong capabilities (e.g., to provide affordances for centralized control), or might favor the division of complementary capabilities across organizations (e.g., to provide affordances for establishing cross-institutional transparency and interdependence). Unlike classic models of advanced AI capabilities as something embodied in a distinct entity (\"the machine\"), the R&D automation model is compatible with both alternatives. \n Further Reading • Section 1: R&D automation provides the most direct path to an intelligence explosion 35 Predictable aspects of future knowledge can inform AI safety strategies AI developers will accumulate extensive safety-relevant knowledge in the course of their work, and predictable aspects of that future knowledge can inform current studies of strategies for safe AI development. \n Summary Along realistic development paths, researchers building advanced AI systems will gain extensive safety-relevant knowledge from experience with similar but less advanced systems. While we cannot predict the specific content of that future knowledge, we can have substantial confidence regarding its scope; for example, researchers will have encountered patterns of success and failure in both development and applications, and will have eagerly explored and exploited surprising capabilities and behaviors across multiple generations of AI technology. Realistic models of potential AI development paths and risks should anticipate that this kind of knowledge will be available to contemporaneous decision makers, hence the nature and implications of future safety-relevant knowledge call for further exploration by the AI safety community. 35.2 Advanced AI systems will be preceded by similar but simpler systems Although we cannot predict the details of future AI technologies and systems, we can predict that their developers will know more about those systems than we do. In general, the nature of knowledge learned during technology development is strictly more predictable than the content of the knowledge itself, hence we can consider the expected scope of future known-knowns and known-unknowns, and even the expected scope of knowledge regarding unknown-unknowns-expected knowledge of patterns of ongoing surprises. Thus, in studying AI safety, it is natural to consider not only our current, sharply limited knowledge of future technologies, but also our somewhat more robust knowledge of the expected scope of future knowledge, and of the expected scope of future knowledge regarding expected surprises. These are aspects of anticipated contemporaneous knowledge. \n Large-scale successes and failures rarely precede smaller successes and failures By design (and practical necessity), low-power nuclear chain reactions preceded nuclear explosions, and despite best efforts, small aircraft crashed before large aircraft. In AI, successes and failures of MNIST classification preceded successes and failures of ImageNet classification, which preceded successes and failures of machine vision systems in self-driving cars. In particular, we can expect that future classes of AI technologies that could yield enormously surprising capabilities will already have produced impressively surprising capabilities; with that experience, to encounter outlierssurprising capabilities of enormous magnitude-would not be enormously surprising. \n AI researchers eagerly explore and exploit surprising capabilities Scientists and researchers focused on basic technologies are alert to anomalies and strive to characterize and understand them. Novel capabilities are pursued with vigor, and unexpected capabilities are celebrated: Notable examples in recent years include word embeddings that enable the solution of word-analogy problems by vector arithmetic, and RL systems that learn the back-cavity multiple-bounce trick in Atari's Breakout game. Surprising capabilities will be sought, and when discovered, they will be tested and explored. 35.5 AI developers will be alert to patterns of unexpected failure Technology developers play close attention to performance: They instrument, test, and compare alternative implementations, and track patterns of success and failure. AI developers seek low error rates and consistent performance in applications; peculiarities that generate unexpected adverse results are (and will be) studied, avoided, or tolerated, but not ignored. \n AI safety researchers will be advising (responsible) AI developers Several years ago, one could imagine that AI safety concerns might be ignored by AI developers, and it was appropriate to ask how safety-oblivious development might go awry. AI safety concerns, however, led to the growth of AI safety studies, which are presently flourishing. We can expect that safety studies will be active, ongoing, and substantially integrated with the AI R&D community, and will be able to exploit contemporaneous community knowledge in jointly developing and revising practical recommendations. \n Considerations involving future safety-relevant knowledge call for further exploration Adoption of safety-oriented recommendations will depend in part on their realism and practicality, considerations that call for a better understanding of the potential extent of future safety-relevant in the development community. Studies of the nature of past technological surprises can inform this effort, as can studies of patterns of development, anticipation, and surprise in modern AI research. We must also consider potential contrasts between past patterns of development and future developments in advanced AI. If contemporaneous computational hardware capacity will (given the development of suitable software) be sufficient to support broadly superhuman performance, 1 then the potential for swift change will be unprecedented. Contingent on informed caution and a security mindset, however, the potential for swift change need not entail unsafe application of capabilities and unprecedented, unavoidable surprises. To understand how such a situation might be managed, it will be important to anticipate the growth of safety-relevant knowledge within the AI development community, and to explore how this knowledge can inform the development of safety-oriented practices. Further Reading  \n Good practice in development tends to align with safety concerns Fortunately (though unsurprisingly) good practice in AI development tends to align with safety concerns. In particular, developers seek to ensure that AI systems behave predictably, a characteristic that contributes to safety even when imperfect. \n Exploring families of architectures and tasks builds practical knowledge In AI R&D, we see extensive exploration of families of architectures and tasks through which developers gain practical knowledge regarding the capabilities and (conversely) limitations of various kinds of systems; practical experience also yields an understanding of the kinds of surprises to be expected from these systems. Guidelines that highlight the role of present and future practical knowledge 1 would clarify why current research is known to be safe, and how good development practice can contribute to future safety. \n Task-oriented development and testing improve both reliability and safety Systems for practical applications perform bounded tasks and are subject to testing and validation before deployment. Task-oriented development, testing, and validation contribute to knowledge of capabilities and focus strongly-motivated attention on understanding potential failures and surprises. Guidelines that codify this aspect of current practice would again help to clarify conditions that contribute to current and future safety. \n Modular architectures make systems more understandable and predictable Systems are more easily designed, developed, tested, and upgraded when they are composed of distinct parts, which is to say, when their architectures are modular rather than opaque and undifferentiated. This kind of structure is ubiquitous in complex systems. (And as noted in a recent paper from Google, \"Only a small fraction of real-world ML systems is composed of the ML code [. . .] The required surrounding infrastructure is vast and complex.\" [Sculley Neural networks and symbolic/algorithmic AI technologies are complements, not alternatives; they are being integrated in multiple ways at levels that range from components and algorithms to system architectures. \n Summary Neural network (NN) and symbolic/algorithmic (S/A) AI technologies offer complementary strengths, and these strengths can be combined in multiple ways. A loosely-structured taxonomy distinguishes several levels of organization (components, algorithms, and architectures), and within each of these, diverse modes of integration-various functional relationships among components, patterns of use in applications, and roles in architectures. The complexity and fuzziness of the taxonomy outlined below emphasizes the breadth, depth, and extensibility of current and potential NN-S/A integration. \n Motivation One might imagine that neural network and symbolic/algorithmic technologies are in competition, and ask whether NNs can fulfill the grand promise of artificial intelligence when S/A methods have failed-will NN technologies also fall short? On closer examination, however, the situation looks quite different: NN and S/A technologies are not merely in competition, they are complementary, compatible, and increasingly integrated in research and applications. To formulate a realistic view of AI prospects requires a general sense of the relationship between NN and S/A technologies. Discussions in this area are typically more narrow: They either focus on a problem domain and explore applicable NN-S/A techniques, or they focus on a technique and explore potential applications. The discussion here will instead outline the expanding range of techniques and applications, surveying patterns of development that may help us to better anticipate technological opportunities and the trajectory of AI development. \n A crisp taxonomy of NN and S/A systems is elusive and unnecessary There is no sharp and natural criterion that distinguishes NN from S/A techniques. For purposes of discussion, one can regard a technique as NN-style to the extent that it processes numerical, semantically-opaque vector representations through a series of transformations in which operations and data-flow patterns are fixed. Conversely, one can regard a technique as S/A-style to the extent that it relies on entities and operations that have distinct meanings and functions, organized in space and time in patterns that manifestly correspond to the structure of the problem at hand-patterns comprising data structures, memory accesses, control flow, and so on. Note that S/A implementation mechanisms should not be mistaken for S/A systems: S/A code can implement NN systems much as hardware implements software, and just as code cannot usefully be reduced to hardware, so NNs cannot usefully be reduced to code. In the present context, however, the systems of greatest interest are those that deeply integrate NN-and S/A-style mechanisms, or that blur the NN-S/A distinction itself. If taxonomy were clean and easily constructed, prospects for NN-S/A integration would be less interesting. \n NN and S/A techniques are complementary The contrasting strengths of classic S/A techniques and emerging NN techniques are well known: S/A-style AI techniques have encountered difficulties in perception and learning, areas in which NN techniques excel; NN-style AI techniques, by contrast, often struggle with tasks like logic-based reasoningand even counting-that are trivial for S/A systems. In machine translation, for example, algorithms based on symbolic representations of syntax and semantics fell short of their promise (even when augmented by statistical methods applied to large corpora); more recently, neural machine translation systems with soft, opaque representations have taken the lead, yet often struggle with syntactic structure and the semantics of logical entailment. S/A systems are relatively transparent, in part because S/A systems embody documentable, human-generated representations and algorithms, while NN systems instead discover and process opaque representations in ways that do not correspond to interpretable algorithms. The integration of NN techniques with S/A systems can sometimes facilitate interpretability, however: For example, NN operations may be more comprehensible when considered as functional blocks in S/A architectures, while (in sometimes amounts to a figure-ground reversal) comprehensible S/A systems can operate on NN outputs distilled into symbols or meaningful numerical values. \n AI-service development can scale to comprehensive, SI-level services In discussing how NN and S/A techniques mesh, it will be convenient to draw rough distinctions between three levels of application: Components and mechanisms, where NN and S/A building blocks interact at the level of relatively basic programming constructs (e.g., data access, function calls, branch selection). Algorithmic and representational structures, where NN and S/A techniques are systematically intwined to implement complex representations or behaviors (e.g., search, conditional computation, message-passing algorithms, graphical models, logical inference). Systems and subsystems, where individually-complex NN and S/A subsystems play distinct and complementary roles at an architectural level (e.g., perception and reasoning, simulation and planning). The discussion below outlines several modes of NN-S/A integration at each level, each illustrated by an unsystematic sampling of examples from the literature. No claim is made that the modes are sharply defined, mutually exclusive, or collectively exhaustive, or that particular examples currently outperform alternative methods. The focus here is on patterns of integration and proofs of concept. \n Integration at the level of components and mechanisms What are conceptually low-level components may have downward-or upwardfacing connections to complex systems: Components that perform simple functions may encapsulate complex NN or S/A mechanisms, while simple functions may serve as building blocks in higher-level algorithmic and representational structures (as discussed in the following section). \n NNs can provide representations processed by S/A systems: Classic AI algorithms manipulate representations based on scalars or symbolic tokens; in some instances (discussed below), systems can retain the architecture of these algorithms-patterns of control and data flow-while exploiting richer NN representations. For example, the modular, fully-differentiable visual question answering architecture of  (Hu et al. 2018 ) employs S/A-style mechanisms (sets of distinct, compositional operators that pass data on a stack), but the data-objects are patterns of soft attention over an image. \n NNs can direct S/A control and data flow: In AI applications, S/A algorithms often must select execution paths in an \"intelligent\" way. NNs can process rich information (e.g., large sets of conventional variables, or NN vector embeddings computed upstream), producing Boolean or integer values that can direct these S/A choices. NN-directed control flow is fundamental to the NN-based search and planning algorithms noted below. \n S/A mechanisms can direct NN control and data flow: Conversely, S/A mechanisms can structure NN computations by choosing among alternative NN components and operation sequences. Conditional S/A-directed NN operations enable a range of NN-S/A integration patterns discussed below. \n NNs can learn heuristics for S/A variables: Reinforcement learning ML mechanisms be applied to what are essentially heuristic computations (e.g., binary search, Quicksort, and cache replacement) by computing a value base on a observations (the values of other variables). This approach embeds ML in a few lines of code to \"integrate ML tightly into algorithms whereas traditional ML systems are build around the model\"  (Carbune et al. 2017) . \n NNs can replace complex S/A functions: Conventional algorithms may call functions that perform costly numerical calculations that give precise results when approximate results would suffice. In the so-called \"parrot transformation\"  (Esmaeilzadeh et al. 2012 ), an NN is trained to mimic and replace the costly function. NNs that (approximately) model the trajectories of physical systems  (Ehrhardt et al. 2017 ) could play a similar role by replacing costly simulators. 37.6.6 NNs can employ complex S/A functions: Standard deep learning algorithms learn by gradient descent on linear vector transformations composed with simple, manifestly-differentiable elementwise functions (ReLU, tanh, etc.), yet complex, internally non-differentiable algorithms can also implement differentiable functions. These functions can provide novel functionality: For example, a recent deep-learning architecture (OptNet) treats constrained, exact quadratic optimization as a layer, and can learn to solve Sudoku puzzles from examples  (Amos and Kolter 2017) . When considered as functions, the outputs of complex numerical models of physical systems can have a similar differentiable character. \n S/A algorithms can extend NN memory: Classic NN algorithms have limited representational capacity, a problem that becomes acute for recurrent networks that must process long sequences of inputs or represent an indefinitely large body of information. NNs can be augmented with scalable, non-differentiable (hard-attention) or structured memory mechanisms  (Sukhbaatar et al. 2015; Chandar et al. 2016 ) that enable storage and retrieval operations in an essentially S/A style. \n S/A data structures can enable scaling of NN representations: Data structures developed to extend the scope of practical representations in S/A systems can be adapted to NN systems. Hash tables and tree structures can support sparse storage for memory networks, for example, and complex data structures (octrees) enable generative convolutional networks to output fine-grained 3D representations that would otherwise require impractically large arrays  (Tatarchenko, Dosovitskiy, and Brox 2017) ; access to external address spaces has proved critical to solving complex, structured problems  (Graves et al. 2016) . Brute-force nearest-neighbor lookup (e.g., in NN embedding spaces) is widely used in single-shot and few-shot learning (see below); recent algorithmic advances based on neighbor graphs enable retrieval of near neighbors from billion-scale data sets in milliseconds  (Fu, Wang, and Cai 2017) . \n S/A mechanisms can template NN mechanisms: In a more abstract relationship between domains, S/A-style mechanisms can be implemented in a wholly differentiable, NN form. Examples include pointer networks  (Vinyals, Fortunato, and Jaitly 2015)  that (imperfectly) solve classic S/A problems such as Traveling Salesman and Delaunay triangulation, as well as differentiable stacks that can perform well on NLP problems (e.g., syntactic transformations  [Grefenstette et al. 2015 ] and dependency parsing  [Dyer et al. 2015] ) commonly addressed by recursive, tree-structured algorithms in S/A systems. \n Integration at the level of algorithmic and representational structures Higher-level integration of NN and S/A mechanisms can typically be viewed as patterns of interleaved NN and S/A operations, sometimes with substantial, exposed complexity in one or both components. \n S/A algorithms can extend NN inference mechanisms: S/A mechanisms are often applied to decode NN outputs. For example, beam search in is a classic algorithm in symbolic AI, and is applied in neural machine translation systems to select sequences of symbols (e.g., words) based on soft distributions over potential sequence elements. In another class of algorithms, S/A algorithms are applied to find near-neighbors among sets of vector outputs stored in external memory, supporting one-and few-shot learning in classifiers  (Vinyals et al. 2016; Snell, Swersky, and Zemel 2017) . \n NNs can guide S/A search: Both search over large spaces of choices (Go play) and large bodies of data (the internet ) and are now addressed by traditional S/A methods (Monte Carlo tree search  [Silver et al. 2016 ], large-scale database search  [Clark 2015 ]) guided by NN choices. Recent work has integrated tree-based planning methods with NN \"intuition\" and end-to-end learning to guide agent behaviors  (Anthony, Tian, and Barber 2017; Farquhar et al. 2017; Guez et al. 2018) . \n S/A graphical models can employ NN functionality: An emerging class of graphical models-message-passing NNs-represents both node states and messages as vector embeddings. Message-passing NNs share the discrete, problem-oriented structures of probabilistic graphical models, but have found a far wider range of applications, including the prediction of molecular properties  (Gilmer et al. 2017) , few-shot learning  (Garcia and Bruna 2017) , and inferring structured representations  (Johnson et al. 2016) , as well as algorithms that outperform conventional loopy belief propagation  (Yoon et al. 2018) , and others that can infer causality from statistical data beyond the limits that might be suggested by Pearl's formalism  (Goudet et al. 2017 ). Vicarious has demonstrated integration of perceptual evidence through a   Allamanis, Brockschmidt, and Khademi [2018] ). 37.7.9 NNs can aid automated theorem proving: Automated theorem proving systems perform heuristic search over potential proof trees, and deep-learning methods have been applied to improve premise selection  (Irving et al. 2016; Kaliszyk, Chollet, and Szegedy 2017) . Progress in automated theorem proving (and proof assistants) could facilitate the development of provably correct programs and operating systems  (Klein et al. 2014   \n Integration at the level of systems and subsystems Integration at the level of systems and subsystems extends the patterns already discussed, combining larger blocks of NN and S/A functionality. \n NNs can support and ground S/A models: Perceptual processing has been chronically weak in S/A artificial intelligence, and NN techniques provide a natural complement. NN-based machine vision applications in robotics and vehicle automation are expanding  (Steger, Ulrich, and Wiedemann 2007) , and NN-based modeling can go beyond object recognition by, for example, enabling the inference of physical properties of objects from video  (Watters et al. 2017) . With the integration of NN perception and symbolic representations, symbol-systems can be grounded. \n S/A representations can direct NN agents: In recent work, deep RL has been combined with symbolic programs, enabling the implementation of agents that learn correspondences between programs, properties, and objects through observation and action; these capabilities can be exploited in systems that learn to ground and execute explicit, humanwritten programs  (Denil et al. 2017) . The underlying principles should gen-eralize widely, improving human abilities to inform, direct, and understand behaviors that exploit the strengths of deep RL. 37.8.3 NNs can exploit S/A models and tools: Looking forward, we can anticipate that NN systems, like human beings, will be able to employ state-of-the-art S/A computational tools, for example, using conventional code that implements physical models, image rendering, symbolic mathematics and so on. NN systems can interact with S/A systems though interfaces that are, at worst, like those we use today. 37.8.4 NN and S/A models can be integrated in cognitive architectures: At a grand architectural level, AI researchers have long envisioned and proposed \"cognitive architectures\" (Soar, LIDA, ACT-R, CLARION. . . ) intended to model much of the functionality of the human mind. A recent review  (Kotseruba, Gonzalez, and Tsotsos 2016)  identifies 84 such architectures, including 49 that are still under active development. Recent work in cognitive architectures has explored the integration of S/A and NN mechanisms  (Besold et al. 2017) , an approach that could potentially overcome difficulties that have frustrated previous efforts. \n Integration of NN and S/A techniques is a rich and active research frontier An informal assessment suggests robust growth in the literature on integration of NN and S/A techniques. It is, however, worth noting that there are incentives to focus research efforts primarily on one or the other. The most obvious incentive is intellectual investment: Crossover research requires the application of disparate knowledge, while more specialized knowledge is typically in greater supply for reasons of history, institutional structure, and personal investment costs. These considerations tend to suggest an undersupply of crossover research. There is, however, good reason to focus extensive effort on NN systems that that do not integrate S/A techniques in a strong, algorithmic sense: We do not yet know the limits of NN techniques, and research that applies NNs in a relatively pure form-end-to-end, tabula rasa training with S/A code providing only infrastructure or framework elements-seems the best way to explore the NN frontier. Even if one expects integrated systems to dominate the world of applications, relatively pure NN research may be the most efficient way to develop the NN building blocks for those applications. As in many fields of endeavour, it is important to recognize the contrasts between effective methodologies in research and engineering: In particular, good basic research explores systems that are novel, unpredictable, and (preferably) simple, while good engineering favors known, reliable building blocks to construct systems that are as complex as a task may require. Accordingly, as technologies progress from research to applications, we can expect to see increasing-and increasingly eclectic-integration of NN and S/A techniques, providing capabilities that might otherwise be beyond our reach. \n Further Reading • Section I: Introduction: From R&D automation to comprehensive AI Services • Section II: Overview: Questions, propositions, and topics 38 Broadly-capable systems coordinate narrower systems In both human and AI systems, we see broad competencies built on narrower competencies; this pattern of organization is a robust feature of intelligent systems, and scales to systems that deliver broad services at a superhuman level. \n Summary In today's world, superhuman competencies reside in structured organizations with extensive division of knowledge and labor. The reasons for this differentiated structure are fundamental: Specialization has robust advantages both in learning diverse competencies and in performing complex tasks. Unsurprisingly, current AI services show strong task differentiation, but perhaps more surprisingly, AI systems trained on seemingly indivisible tasks (e.g., translating sentences) can spontaneously divide labor among \"expert\" components. In considering potential SI-level AI systems, black-box abstractions may sometimes be useful, but these abstractions set aside our general knowledge of the differentiated architecture of intelligent systems. \n Today's superhuman competencies reside in organizational structures It is a truism that human organizations can achieve tasks beyond individual human competence by employing, not just many individuals, but individuals who perform differentiated tasks using differentiated knowledge and skills. Adam Smith noted the advantages of division of labor (even in making pins), and in modern corporations, division of labor among specialists is a necessity. \n Specialization has robust advantages in learning diverse competencies The structure of knowledge enables parallel training: Learning competencies in organic chemistry, financial management, mechanical engineering, and customer relations, for example, is accomplished by individuals who work in parallel to learn the component tasks. This pattern of parallel, differentiated learning works well because many blocks of specialized knowledge have little mutual dependence. The limited pace of human learning and vast scope of human knowledge make parallel training mandatory in the human world, and in machine learning analogous considerations apply. Loosely-coupled bodies of knowledge call for loosely-coupled learning processes that operate in parallel. \n Division of knowledge and labor is universal in performing complex tasks As with learning, parallel, specialized efforts have great advantages in performing tasks. Even setting aside human constraints on bandwidth and representational power, there would be little benefit in attempting to merge day-to-day tasks in the domains of organic chemistry, financial management, mechanical engineering, and customer relations. Both information flow and knowledge naturally cluster in real-world task structures, and the task of crosstask management (e.g., in developing and operating a chemical-processing system) has only limited overlap with the information flows and knowledge that are central to the tasks that must be coordinated. To implement a complex, inherently differentiated task in a black-box system is to reproduce the task structure inside the box while making its organization opaque. \n Current AI services show strong task differentiation It goes without saying that current AI systems are specialized: Generality is a challenge, not a default. Even in systems that provide strong generalization capacity, we should expect to see diminishing returns from attempts to apply single-system generalization capacity to the full scope and qualitative diversity of human knowledge. Indeed, considering the nature of training and computation, it is difficult to imagine what \"single-system generalization\" could even mean on that scale. \n AI systems trained on seemingly indivisible tasks learn to divide labor Task specialization can emerge spontaneously in machine learning systems. A striking recent example is a mixture-of-experts model employed in a then state-of-the art neural machine translation system to enable 1000x improvements in model capacity  (Shazeer et al. 2017 ). In this system, a managerial component delegates the processing of sentence fragments to \"experts\" (small, architecturally undifferentiated networks), selecting several from a pool of thousands. During training, experts spontaneously specialize in peculiar, semantically-and syntactically-differentiated aspects of text comprehension. The incentives for analogous specialization and task delegation can only grow as tasks become wider in scope and less tightly coupled.  \n Further Reading • Section 12: AGI agents offer no compelling value \n Broad capabilities call for mechanisms that compose diverse competencies Current AI systems are notoriously narrow, and expanding the scope of functionality of individual systems is a major focus of research. Despite progress, vision networks that recognize faces are still distinct from networks that classify images, which are distinct from networks that parse scenes into regions corresponding to objects of different kinds-to say nothing of the differences between any of these and networks architected and trained to play Go or translate languages or predict the properties of molecules. Because it would be surprising to find that any single network architecture will be optimal for proposing Go moves, and identifying faces, and translating English to French, it is natural to ask how diverse, relatively narrow systems could be composed to form systems that externally present broad, seamless competencies. From this perspective, matching tasks (inputs and goals) to services (e.g., trained networks) is central to developing broadly-applicable intelligent functionality, whether the required task-matching mechanisms provide coherence to services developed within a system through the differentiation of subtasks, or integrate services developed independently. This perspective equally applicable to \"hard-wiring\" task-to-service matching at development time and dynamic matching during process execution, and equally applicable to deeplearning approaches and services implemented by (for example) Google's algorithm-agnostic AutoML. In practice, we should expect to see systems that exploit both static and dynamic matching, as well as specialized services implemented by relatively general service-development services . Algorithm selection has long been an active field  (Kotthoff 2012; Mısır and Sebag 2017; Yang et al. 2018) . \n The task-space concept suggests a model of integrated AI services It is natural to think of services as populating task spaces in which similar services are neighbors and dissimilar services are distant, while broader services cover broader regions. This picture of services and task-spaces can be useful both as a conceptual model for thinking about broad AI competencies, and as a potential mechanism for implementing them. 1. As a conceptual model, viewing services as tiling a high-dimensional task space provides a framework for considering the relationship between tasks and services: In the task-space model, the diverse properties that differentiate tasks are reflected in the high dimensionality of the task space, services of greater scope correspond to tiles of greater extent, and gaps between tiled regions correspond to services yet to be developed. 2. As an implementation mechanism, jointly embedding task and service representations in high dimensional vector spaces could potentially facilitate matching of tasks to services, both statically during implementation and dynamically during execution. While there is good reason to think that joint embedding will be a useful implementation technique, the value of task spaces as a conceptual model would stand even if alternative implementation techniques prove to be superior. The discussion that follows will explore the role of vector embeddings in modern AI systems, first, to support proposition (  1 ) by illustrating the richness and generality of vector representations, and, second, to support proposition (  2 ) by illustrating the range of areas in which proximity-based operations on vector representations already play fundamental roles in AI system implementation. With a relaxed notion of \"space\", proposition (1) makes intuitive sense; the stronger proposition (  2 ) requires closer examination. \n Embeddings in high-dimensional spaces provide powerful representations This discussion will assume a general familiarity with the unreasonable effectiveness of high-dimensional vector representations in deep learning systems, while outlining some relevant developments. In brief, deep learning systems often encode complex and subtle representations of the objects of a domain (be it an image, video, text, product) as numerical vectors (\"embeddings\") in spaces with tens to thousands of dimensions; the geometric relationships among embeddings encode relationships among the objects. In particulargiven a successful embedding-similar objects (images of the same class, texts will similar meanings) map to vectors that are near neighbors in the embedding space. Distances between vectors can be assigned in various ways: The most common in neural networks is cosine similarity, the inner product of vectors normalized to unit length; in high dimensional spaces, the cosine similarity between randomly-oriented vectors will, with high probability, be ≈ 0, while values substantially greater than 0 indicate relatively near neighbors. Distances in several other spaces have found use: The most common is Euclidian (L2) norm in conventional vector spaces, but more recent studies have employed Euclidian distance in toroidal spaces (placing bounds on distances while preserving translational symmetry  [Ebisu and Ichise 2017] ), and distances in hyperbolic spaces mapped onto the Poincare ball (the hyperbolic metric is particularly suited to a range of graph embeddings: it allows volume to grow exponentially with distance, much as the width of a balanced tree grows exponentially with depth)  (Gülçehre et al. 2018; Tifrea, Bécigneul, and Ganea 2018) . Spaces with different metrics and topologies (and their Cartesian products) may be suited to different roles, and individually-continuous spaces can of course be disjoint, and perhaps situated in a discrete taxonomic space. \n High-dimensional embeddings can represent semantically rich domains In deep neural networks, every every vector that feeds into a fully-connected layer can be regarded as a vector-space embedding of some representation; examples include the top-level hidden layers in classifiers and intermediate encodings in encoder-decoder architectures. Applications of vector embedding have been extraordinarily broad, employing representations of: • Images for classification tasks (Krizhevsky, Sutskever, and Hinton 2012) • Images for captioning tasks  (Vinyals et al. 2014)  • Sentences for translation tasks  (Wu et al. 2016)  • Molecules for property-prediction tasks  (Coley et al. 2017)  • Knowledge graphs for link-prediction tasks  (Bordes et al. 2013)  • Products for e-commerce (J.  Wang et al. 2018)  The successful application of vector embedding to diverse, semantically complex domains suggests that task-space models are not only coherent as a concept, but potentially useful in practice. \n Proximity-based (application/activation/access) can deliver diverse services Because applications may use neighborhood relationships to provide diverse functionality (here grouped under the elastic umbrella of \"services\"), the present discussion will refer to these collectively as \"PBA operations\", where \"PB\" denotes \"proximity-based\", and \"A\" can be interpreted as application of selected functions, activation of selected features, or more generically, access to (or retrieval of) selected entities. In each instance, PBA operations compute a measure of distance between a task representation embedding (\"query\") and a pre-computed embedding (\"key\") corresponding to the accessed feature, \n Encoding and decoding vector embeddings Image classification: Image → CNN → embedding → projection matrix → class  (Krizhevsky, Sutskever, and Hinton 2012)  Image captioning: Image → CNN → embedding → RNN → caption  (Vinyals et al. 2014)  Language translation: Sentence → RNN → embedding → RNN → translation  (Wu et al. 2016)  function, or entity (\"value\"). In other words, PBA operations employ key/value lookup of near neighbors in a vector space. As shorthand, one can speak of values has having positions defined by their corresponding keys. PBA operations may access multiple entities; when these are (or produce) embedding vectors, it can be useful to weight and add them (e.g., weighting by cosine similarity between query and key). Weighted PBA (wPBA) operations that discard distant values are PBAs in a strict sense, and assigning small weights to distant entities has a similar effect. The class of wPBA operations thus includes any matrix multiplication in which input and row (= query and key) vectors are actually or approximately normalized (see  Salimans and Kingma (2016)  and C.  Luo et al. (2017) ), and the resulting activation vectors are sparse, e.g., as a consequence of negatively biased ReLU units. \n PBA operations are pervasive in deep learning systems PBA mechanisms should not be regarded as a clumsy add-on to neural computation; indeed, the above example shows that wPBA operations can be found at the heart of multilayer perceptrons. A wide range of deep learning systems apply (what can be construed as) PBA operations to access (what can be construed as) fine-grained \"services\" within a neural network computation. For example, wPBA mechanisms have been used implement not only mixtureof-experts models  (Shazeer et al. 2017; Kaiser, Gomez, et al. 2017 ), but also attention mechanisms responsible for wide-ranging advances in deep learning  (Vaswani et al. 2017; Kool, van Hoof, and Welling 2018; Hudson and Manning 2018) , including memories of past situations in RL agents  (Wayne et al. 2018) . PBA mechanisms are prominent in single-shot learning in multiple domains  (Kaiser, Nachum, et al. 2017) , including learning new image classes from a single example. In the latter application, networks are trained to map images to an embedding space that supports classification; single-shot learning is performed by mapping pairs of novel labels and embeddings to that same embedding space, enabling subsequent classification by retrieving the label of the best-matching embedding (\"best-matching\" means closeness, e.g., cosine similarity)  (Vinyals et al. 2016) . In one implementation, this is accomplished by normalizing and inserting the embedding of a new example in a standard projection matrix  (Qi, Brown, and  Lowe 2017) (indeed, with suitable normalization, standard image classification architectures can be regarded as employing PBA). \n Joint embeddings can link related semantic domains Embeddings can not only map similar entities to neighboring locations, but can also align distinct domains such that entities in one domain are mapped to locations near those of corresponding entities in the other  (Frome et al. 2013; Y. Li et al. 2015; Baltrusaitis, Ahuja, and Morency 2017) . Applications have been diverse: • Text and images to enable image retrieval (K.  • Video and text to enable action recognition  (Xu, Hospedales, and Gong 2017)  • Sounds and objects in video to learn cross-modal relationships (Arandjelović and Zisserman 2017) • Images and recipes to retrieve one given the other  (Salvador et al. 2017)  • Images and annotations to improve embeddings  (Gong et al. 2014)  • Queries and factual statements to enable text-based question answering  (Kumar et al. 2015)  • Articles and user-representations to recommend news stories  (Okura et al. 2017 ) • Product and user/query representations to recommend products  (Zhang, Yao, and Sun 2017)  39.9 PBA operations can help match tasks to candidate services at scale The breadth of applications noted above suggests that AI services and tasks could be represented and aligned in vector embedding spaces. This observa-tion shows that the task-space concept is (at the very least!) coherent, but also suggests that PBA operations are strong candidates for actually implementing task-service matching. This proposition is compatible with the use of disjoint embedding spaces for different classes of tasks, the application of further selection criteria not well represented by distance metrics, and the use of PBA operations in developing systems that hard-wire services to sources of streams of similar tasks. In considering joint embeddings of tasks and services, one should imagine joint training to align the representations of both service-requesting and service-providing components. Such representations would encode the nature of the task (vision? planning? language?), domain of application (scene? face? animal?), narrower domain specifications (urban scene? desert scene? Martian scene?), kind of output (object classes? semantic segmentation? depth map? warning signal?), and further conditions and constraints (large model or small? low or high latency? low or high resolution? web-browsing or safety-critical application?). Note that exploitation of specialized services by diverse higher-level systems is in itself a form of transfer learning: To train a service for a task in one context is to train it for similar tasks wherever they may arise. Further, the ability to find services that are near-matches to a task can provide trained networks that are candidates for fine-tuning, or (as discussed below) architectures that are likely to be well-suited to the task at hand  (Vanschoren 2018) . The concept of joint task-to-service embeddings suggests directions for experimental exploration: How could embeddings of training sets contribute to embeddings of trained networks? Distributional shifts will correspond to displacement vectors in task space-could regularities in those shifts be learned and exploited in metalearning? Could relationships among task embeddings guide architecture search? Note that task-to-service relationships form a bipartite graph in which links can be labeled with performance metrics; in optimized embeddings, the distances between tasks and services will be predictive of performance. PBA operations can be applied at scale: A recently developed graph-based, polylogarithmic algorithm running can return sets of 100 near neighbors from sets of >10 7 vector embeddings with millisecond response times (single CPU)  (Fu, Wang, and Cai 2017) . Alibaba employs this algorithm for product recommendation at a scale of billions of items and customers (J.  Wang et al. 2018) . 39.10 PBA operations can help match new tasks to service-development services Fluid matching of tasks to services tends to blunt the urgency of maximizing the scope of individual models. Although models with more general capabilities correspond to broader tiles in task-space, the size of individual tiles does not determine the breadth of their aggregate scope. Advances in metalearning push in the same direction: A task that maps to a gap between tiles may still fall within the scope of a reliable metalearning process that can, on demand, fill that gap  (Vanschoren 2018) . A particular metalearning system (characterized by both architecture and training) would in effect constitute a broad but high-latency tile which has first-use costs that include both data and computational resources for training. Graphs representing deep learning architectures can themselves be embedded in continuous spaces (and, remarkably, can be optimized by gradient descent [R.  Luo et al. 2018] ); learning and exploiting joint embeddings of tasks and untrained architectures would be a natural step. In an intuitive spatial picture, metalearning methods enable population of a parallel space of service-providing services, a kind of backstop for tasks that pass through gaps between tiles in the primary task-space. Taking this picture further, one can picture a deeper backstop characterized by yet broader, more costly tiles: This space would be populated by AI research and development systems applicable to broader domains; such systems might search spaces of architectures, training algorithms, and data sets in order to provide systems suitable for filling gaps in metalearning and primary-task spaces. AI R&D comprises many subtasks (architecture recommendation, algorithm selection, etc.) that can again be situated in appropriate task spaces; as with other highlevel services, we should expect high-level AI-development services to operate by delegating tasks and coordinating other, narrower services. One may speculate that systems that display flexible, general intelligence will, internally, link tasks to capabilities by mechanisms broadly similar those in today's deep learning systems-which is to say, by mechanisms that can be construed as employing similarity of task and service embeddings in highdimensional vector spaces. What is true of both multi-layer perceptrons and e-commerce recommendation systems is apt to be quite general. \n Integrated, extensible AI services constitute general artificial intelligence The concept of \"general intelligence\" calls for a capacity to learn and apply an indefinitely broad range of knowledge and capabilities, including high-level capabilities such as engineering design, scientific inquiry, and long-term planning. The concept of comprehensive AI services is the same: The CAIS model calls for the capacity to develop and apply an indefinitely broad range of services that provide both knowledge and capabilities, again including high-level services such as engineering design, scientific inquiry, and long-term planning. In other words, broad, extensible, integrated CAIS in itself constitutes general artificial intelligence, differing from the familiar AGI picture chiefly in terminology, concreteness, and avoidance of the long-standing assumption that well-integrated general intelligence necessarily entails unitary agency. \n Further Reading • Section 1: R&D automation provides the most direct path to an intelligence explosion • Section 12: AGI agents offer no compelling value 40 Could 1 PFLOP/s systems exceed the basic functional capacity of the human brain? Multiple comparisons between narrow AI tasks and narrow neural tasks concur in suggesting that PFLOP/s computational systems exceed the basic functional capacity of the human brain. \n Summary Neurally-inspired AI systems implement a range of narrow yet recognizably human-like competencies, hence their computational costs and capabilities can provide evidence regarding the computational requirements of hypothetical systems that could deliver more general human-like competencies at human-like speeds. The present analysis relies on evidence linking task functionality to resource requirements in machines and biological systems, making no assumptions regarding the nature of neural structure or activity. Comparisons in areas that include vision, speech recognition, and language translation suggest that affordable commercial systems (~1 PFLOP/s, costing $150,000 in 2017) may surpass brain-equivalent computational capacity, perhaps by a substantial margin. Greater resources can be applied to learning, and the associated computational costs can be amortized across multiple performance-providing systems; current experience suggests that deep neural network (DNN) training can be fast by human standards, as measured by wall-clock time. In light of these considerations, it is reasonable to expect that, given suitable software, affordable systems will be able to perform human-level tasks at superhuman speeds. Hypothetical AI software that could perform tasks with human-level (or better) competence, in terms of scope and quality, would be constrained by contemporaneous computational capacity, and hence might perform with less-than-human task throughput; if so, then restricted hardware capacity might substantially blunt the practical implications of qualitative advances in AI software. By contrast, if advanced competencies were developed in the context of better-than-human hardware capacity (\"hardware overhang\"), then the practical implications of qualitative advances in AI software could potentially be much greater. A better understanding of the computational requirements for human-level performance (considering both competence and throughput) would enable a better understanding of AI prospects. \n Ratios of hardware capacities and neural capacities (considered separately) compare apples to apples The following analysis references an imperfect measure of real-world hardware capacity-floating-point performance-yet because it considers only ratios of capacity between broadly-similar systems applied to broadly-similar tasks, the analysis implicitly (though approximately) reflects cross-cutting considerations such as constraints on memory bandwidth. Thus, despite referencing an imperfect measure of hardware capacity, the present methodology compares apples to apples. The neural side of the analysis is similar in this regard, considering only ratios of (estimates of) neural activity required for task performance relative to activity in the brain as a whole; these ratio-estimates are imperfect, but again compare apples to apples. Note that this approach avoids dubious comparisons of radically different phenomena such as synapse firing and logic operations, or axonal signaling and digital data transmission. Neural structure and function is treated as a black box (as is AI software). In the end, the quantity of interest (an estimate of machine capacity/brain capacity) will be expressed as a ratio of dimensionless ratios. In 2017, NVIDIA introduced a 960 TFLOP/s \"deep-learning supercomputer\" (a 5.6× faster successor to their 2016 DGX-1 machine), at a price of $150,000; a high-end 2017 supercomputer (Sunway TaihuLight) delivers ~100 PFLOP/s; a high-end 2017 gaming GPU delivers ~0.01 PFLOP/s. The following discussion will take 1 PFLOP/s as a reference value for current laboratory-affordable AI hardware capacity. \n 40.2.4 The computational cost of machine tasks scaled to human-like throughput is reasonably well defined Consider machine tasks that are narrow in scope, yet human-comparable in quality: For a given machine task and implementation (e.g., of image classification), one can combine a reported computational cost (in FLOP/s) and reported throughput (e.g., frames per second) to define a cost scaled to human-like task-throughput (e.g., image classification at 10 frames per second). Call this the \"machine-task cost\", which will be given as a fraction of a PFLOP/s. 40.2.5 Narrow AI tasks provide points of reference for linking computational costs to neural resource requirements AI technologies based on neurally-inspired DNNs have achieved humancomparable capabilities on narrow tasks in domains that include vision, speech recognition, and language translation. It is difficult to formulate accurate, quantitative comparisons that link the known computational costs of narrow AI tasks to the resource costs of similar (yet never equivalent) neural tasks, yet for any given task comparison, one can encapsulate the relevant ambiguities and uncertainties in a single dimensionless parameter, and can consider the implications of alternative assumptions regarding its value. A key concept in the following will be \"immediate neural activity\" (INA), an informal measure of potentially task-applicable brain activity. As a measure of current neural activity potentially applicable to task performance, INA is to be interpreted in an abstract, information-processing sense that conceptually excludes the formation of long-term memories (as discussed below, human and machine learning are currently organized in fundamentally different ways). The estimates of task-applied INA in this section employ cortical volumes that could be refined through closer study of the literature; a point of conservatism in these estimates is their neglect of the differential, task-focused patterns of neural activity that make fMRI informative  (Heeger and Ress 2002) . Differential activation of neural tissue for different tasks is analogous to the use of gated mixture-of-experts models in DNNs: In both cases, a managerial function selects and differentially activates task-relevant resources from a potentially much larger pool. In DNN applications (e.g., language translation), a gated mixture-of-experts approach can increase model capacity by a factor of 100 to 1000 with little increase in computational cost  (Shazeer et al. 2017) . \n 40.2.6 The concept of a \"task-INA fraction\" encapsulates the key uncertainties and ambiguities inherent in linking machine-task costs to brain capacity The present discussion employs the concept of a \"task-INA fraction\" (f INA ), the ratio between the INA that would (hypothetically) be required for a neural system to perform a given machine task and the contemporaneous global INA of a human brain (which may at a given moment support vision, motor function, auditory perception, higher-level cognition, etc.). This ratio encapsulates the main ambiguities and uncertainties in the chain of inference that links empirical machine performance to estimates of the requirements for humanequivalent computation. These ambiguities and uncertainties are substantial: Because no actual neural system performs the same task as a machine, any comparison of machine tasks to neural tasks can at best be approximate. For example, convolutional neural networks (CNNs) closely parallel the human visual system in extracting image features, but the functional overlap between machine and neural tasks dwindles and disappears at higher levels of processing that, in CNNs, may terminate with object segmentation and classification. Potentially quantifiable differences between CNN and human visual processing include field of view, resolution, and effective frame rate. More difficult to disentangle or quantify, however, is the portion of visualtask INA that should be attributed to narrow CNN-like feature extraction, given that even low-level visual processing is intertwined with inputs that include feedback from higher levels, together with more general contextual and attentional information  (Heeger and Ress 2002) . Ambiguities and uncertainties of this kind increase when we consider tasks such as machine speech transcription (which is partially auditory and partially linguistic, but at a low semantic level), or language translation that is human-comparable in quality  (Wu et al. 2016 ), yet employs a very limited representation of language-independent meaning  (Johnson et al. 2016) . \n Uncertainties and ambiguities regarding values of f INA are bounded Despite these uncertainties and definitional ambiguities, there will always be bounds on plausible values of f INA for various tasks. For, example, given that visual cortex occupies ~20% of the brain and devotes substantial resources to CNN-like aspects of feature extraction, it would be difficult to argue that the value of f INA for CNN-like aspects of early visual processing is greater than 0.1 or less than 0.001. However, rather than emphasizing specific estimates of f INA for specific machine tasks, the method of analysis adopted here invites the reader to consider the plausibility of a range of values based on some combination of knowledge from the neurosciences, introspection, and personal judgment. As we will see, even loose bounds on values of f INA can support significant conclusions. For a given task, a useful, empirically-based benchmark for comparison is the \"PFLOP-parity INA faction\" (f PFLOP ), which is simply the ratio of the empirical machine-task cost to a 1 PFLOP/s machine capacity. If the lowest plausible value of f INA lies above the PFLOP-parity INA-faction for that same task, this suggests that a 1 PFLOP/s machine exceeds human capacity by a factor of R PFLOP =  Speech-recognition tasks vs. human auditory/linguistic tasks: Baidu's Deep Speech 2 system can approach or exceed human accuracy in recognizing and transcribing spoken English and Mandarin, and would require approximately 1 GFLOP/s per real-time speech stream  (Amodei et al. 2015) . For this roughly human-level throughput, f PFLOP = 10 −6 . Turning to neural function again, consider that task-relevant auditory/semantic cortex probably comprises >1% of the human brain. If the equivalent of the Deep Speech 2 speech-recognition task were to require 10% of that cortex, then f INA = 10 −3 , and R PFLOP = 1000. Language-translation tasks vs. human language comprehension tasks: Google's neural machine translation (NMT) systems have reportedly approached human quality  (Wu et al. 2016) . A multi-lingual version of the Google NMT model (which operates with the same resources) bridges language pairs through a seemingly language-independent representation of sentence meaning  (Johnson et al. 2016) , suggesting substantial (though unquantifiable) semantic depth in the intermediate processing. Performing translation at a human-like rate of one sentence per second would require approximately 100 GFLOP/s, and f PFLOP = 10 −4 . It is plausible that (to the extent that such things can be distinguished) human beings mobilize as much as 1% of global INA at an \"NMT task-level\"involving vocabulary, syntax, and idiom, but not broader understandingwhen performing language translation. If so, then for \"NMT-equivalent translation,\" we can propose f INA = 10 −2 , implying R PFLOP = 100. Robotic vision vs. retinal visual processing: Hans Moravec applies a different yet methodologically similar analysis  (Moravec 1998 ) that can serve as a cross-check on the above values. Moravec noted that both retinal visual processing and functionally-similar robot-vision programs are likely to be efficiently implemented in their respective media, enabling a comparison between the computational capacity of digital and neural systems. Taking computational requirements for retina-level robot vision as a baseline, then scaling from the volume of the retina to the volume of the brain, Moravec derives the equivalent of R PFLOP = ~10 (if we take MIP/s ~MFLOP/s). Thus, the estimates here overlap with Moravec's. In the brain, however, typical INA per unit volume is presumably less that that of activated retina, and a reasonable adjustment for this difference would suggest R PFLOP > 100. \n It seems likely that 1 PFLOP/s machines equal or exceed the human brain in raw computation capacity In light of the above comparisons, all of which yield values of R PFLOP in the 10 to 1000 range, it seems likely that 1 PFLOP/s machines equal or exceed the human brain in raw computation capacity. To draw the opposite conclusion would require that the equivalents of a wide range of seemingly substantial perceptual and cognitive tasks would consistently require no more than an implausibly small fraction of total neural activity. The functional-capacity approach adopted here yields estimates that differ substantially, and sometimes greatly, from estimates based on proposed correspondences between neural activity and digital computation.  Sandberg and Bostrom (Sandberg and Bostrom 2008) , for example, consider brain emulation at several levels: analog neural network populations, spiking neural networks, and neural electrophysiology; the respective implied R PFLOP values are 1, 10 −3 , and 10 −7 . Again based on a proposed neural-computational correspondence, Kurzweil suggests the equivalent of R PFLOP = 0.1  (Kurzweil 2005) . \n Even with current methods, training can be fast by human standards The discussion above addresses only task performance, but DNN technologies also invite a comparisons of machine and human learning speeds. Human beings require months to years to learn to recognize objects, to recognize and transcribe speech, and to learn vocabulary and translate languages. Given abundant data and 1 PFLOP/s of processing power, the deep learning systems referenced above could be trained in hours (image and speech recognition, ~10 exaFLOPs) to weeks (translation, ~1000 exaFLOPs). These training times are short by human standards, which suggests that future learning algorithms running on 1 PFLOP/s systems could rapidly learn task domains of substantial scope. A recent systematic study shows that the scale of efficient parallelism in DNN training increases as tasks grow more complex, suggesting that training times could remain moderate even as product capabilities increase  (McCandlish et al. 2018) . \n Large computational costs for training need not substantially undercut the implications of low costs for applications Several considerations strengthen the practical implications of fast training, even if training for broad tasks were to require more extensive machine resources: • More that 1 PFLOP/s can be applied to training for narrow AI tasks. • Because broad capabilities can often be composed by coordinating narrower capabilities, parallel, loosely-coupled training processes may be effective in avoiding potential bottlenecks in learning broader AI tasks. • In contrast to human learning, machine training costs can be amortized over an indefinitely large number of task-performing systems, hence training systems could be costly without undercutting the practical implications of high task-throughput with affordable hardware. Human beings (unlike most current DNNs) can learn from single examples, and because algorithms with broad human-level competencies will (almost by definition) reflect solutions to this problem, we can expect the applicable training methods to be more efficient than those discussed above. Progress in \"single-shot learning\" is already substantial. Note that hardware-oriented comparisons of speed do not address the qualitative shortcomings of current DNN training methods (e.g., limited general-ization, requirements for enormous amounts of training data). The discussion here addresses only quantitative measures (learning speed, task throughput). \n Conclusions Many modern AI tasks, although narrow, are comparable to narrow capacities of neural systems in the human brain. Given an empirical value for the fraction of computational resources required to perform that task with humanlike throughput on a 1 PFLOP/s machine, and an inherently uncertain and ambiguous-yet bounded-estimate of the fraction of brain resources required to perform \"the equivalent\" of that machine task, we can estimate the ratio of PFLOP/s machine capacity to brain capacity. What are in the author's judgment plausible estimates for each task are consistent in suggesting that this ratio is ~10 or more. Machine learning and human learning differ in their relationship to costs, but even large machine learning costs can be amortized over an indefinitely large number of task-performing systems and application events. In light of these considerations, we should expect that substantially superhuman computational capacity will accompany the eventual emergence of a software with broad functional competencies. On present evidence, scenarios that assume otherwise seem unlikely.  \n Further Reading \n Afterword While this document was being written, AI researchers have, as the R&Dautomation/AI-services model would predict, continued to automate research and development processes while developing systems that apply increasingly general learning capabilities to an increasing range of tasks in bounded domains. Progress along these lines continues to exceed my expectations in surprising ways, particularly in the automation of architecture search and training. The central concepts presented in this document are intended to be what Chip Morningstar calls \"the second kind of obvious\"-obvious once pointed out, which is to say, obvious in light of facts that are already well-known. Based on the reception in the AI research community to date, this effort seems to have largely succeeded. Looking forward, I hope to see the comprehensive AI-services model of general, superintelligent-level AI merge into the background of assumptions that shape thinking about the trajectory of AI technology. Whatever one's expectations may be regarding the eventual development of advanced, increasingly general AI agents, we should expect to see diverse, increasingly general superintelligent-level services as their predecessors and as components of a competitive world context. This is, I think, a robust conclusion that reframes many concerns. Figure 1 : 1 Figure 1: Classes of intelligent systems \n Figure 2 : 2 Figure 2: Optimization pressures during AI system development focus resources on tasks and enable further development based on taskfocused components. \n Figure 3 : 3 Figure 3: Schematic model of the current AI research and development pipeline. \n Figure 4 : 4 Figure 4: Schematic model of an AI-enabled application-oriented system development task that draws on a range of previously developed components. \n Figure 5 : 5 Figure 5: AI development processes provide affordances for constraining AI systems that can be effective without insights into their internal representations. Points of control include information inputs, model size, (im)mutability, loss functions, functional specialization and composition, and optimization pressures that tend to become sharper as implementation technologies improve. (Adapted from Drexler [2015]) \n Figure 6 : 6 Figure 6: Individual agents capable of open-ended learning would plan, act, and adapt their actions to a particular task environment, while building on individual experience to learn better methods for planning, acting, and adaptation (generic task learning). \n Figure 8 : 8 Figure 8: In large-scale agent applications, N agents (e.g. a thousand or a million) would independently plan, act, and adapt in a range of similar task environments, while aggregation of the resulting task experience enables data-rich learning supported by centralized development resources. Centralized learning enables upgraded agent software to be tested before release. \n Figure 9 : 9 Figure 9: A task structure & architecture for interactive design engineering \n Figure 10 : 10 Figure 10: Applying problematic, ill-characterized AI systems to implementation tasks under competitive optimization pressure could produce clean, compact systems with general learning capabilities. (Schematic diagram) \n 40. 2 2 Metrics and methodology 40.2.1 AI-technology performance metrics include both task competencies and task throughput \n 40. 2 . 2 3 Laboratory-affordable AI hardware capacity reached ~1 PFLOP/s in 2017. \n ratios for specific machine tasks (Numbers in this section are rounded vigorously to avoid spurious implications of precision.)40.3.1 Image-processing tasks vs. human vision tasks:Systems based on Google's Inception architecture implement high-level feature extraction of a quality that supports comparable-to-human performance in discriminating among 1000 image classes. At a human-like 10 frames per second, the machine-task cost would be ~10 GFLOP/s (Szegedy et al. 2014) , hence f PFLOP = 10 −5 . \n Figure 11 : 11 Figure 11: Diagram of neural and computational costs as fractions of total resources (above), scaling the totals to align the fractions (below). \n Suites of AI services that support SI-level AI development-working in consultation with clients and users-could provide a comprehensive range of novel AI services; these would presumably include services provided by adaptive, upgradable, task-oriented agents. It is difficult to see how the introduction of potentially unstable agents that undergo autonomous open-ended selftransformation would provide additional value. • Section 10: R&D automation dissociates recursive improvement from AI agency • Section 11: Potential AGI-enabling technologies also enable comprehensive AI services II.2.2 Would self-transforming agents provide uniquely valuable functionality? • Section 3: To understand AI prospects, focus on services, not implementations • Section 12: AGI agents offer no compelling value II.2.3 Can fast recursive improvement be controlled and managed? Recursive improvement of basic AI technologies would apply allocated machine resources to the development of increasingly functional building blocks for AI applications (better algorithms, architectures, training methods, etc.); basic technology development of this sort could be open-ended, recursive, and fast, yet non-problematic. Deployed AI applications call for careful management, but applications stand outside the inner loop of basic-technology improvement. • Section 23: AI development systems can support effective human guidance • Section 24: Human oversight need not impede fast, recursive AI technology improvement II.2.4 Would general learning algorithms produce systems with general competence? \n 1 . 1 Section 29: The AI-services model reframes the potential roles of AGI agents. 2. Section 22: Machine learning can develop predictive models of human approval. 3. Section 23: AI development systems can support effective human guidance. 4. Section 26: Superintelligent-level systems can safely provide design and planning services. 5. Section 10: R&D automation dissociates recursive improvement from AI agency. 6. Section 8: Strong optimization can strongly constrain AI capabilities, behavior, and effects. 7. Section 26: Superintelligent-level systems can safely provide design and planning services. 8. Section 20: Collusion among superintelligent oracles can readily be avoided. 9. Section 16: Aggregated experience and centralized learning support AI-agent applications. \n 2 The AI-services model describes current AI development AI technology today advances through increasingly automated AI research and development 1 , and produces applications that provide services 2 , performing tasks such as translating languages, steering cars, recognizing faces, and beating Go masters. AI development itself employs a growing range of AI services, including architecture search, hyperparameter search, and training set development. \n 1    \n Potential AGI technologies might best be applied to automate development of comprehensive AI servicesIn summary, an AI technology base that could implement powerful selfimproving AGI agents could instead be applied to implement (or more realistically, upgrade) increasingly automated AI development, a capability that in turn can be applied to implement a comprehensive range of AI applications. Thus, swift, AI-enabled improvement of AI technology does not require opaque self-improving systems, 1 and comprehensive AI services need not be provided by potentially risky AGI agents. 2   Further Reading 1. Section 23: AI development systems can support effective human guidance 2. Section 23: AI development systems can support effective human guidance and Section 22: Machine learning can develop predictive models of human approval 3. Section 8: Strong optimization can strongly constrain AI capabilities, behavior, and effects 4. Section 35: Predictable aspects of future knowledge can inform AI safety strategies 5. Section 22: Machine learning can develop predictive models of human approval 6. Section 26: Superintelligent-level systems can safely provide design and planning services 11.8 • Section 10: R&D automation dissociates recursive improvement from AI agency • Section 12: AGI agents offer no compelling value • Section 15: Development-oriented models align with deeply-structured AI systems • Section 24: Human oversight need not impede fast, recursive AI technology improvement • Section 30: Risky AI can help develop safe AI • Section 33: Competitive AI capabilities will not be boxed \n Similarly, the classic AGI agent model posits systems that could provide general, fluidly-integrated AI capabilities, but this seemingly simple concept hides what by nature must be functionally equivalent to a comprehensive range of AI services and coordination mechanisms. The classic model assumes these capabilities without explaining how they might work; for simplicity, an AI-services model could arbitrarily assume equivalent capabilities, but a deeper model offers a framework for considering how diverse, increasingly comprehensive capabilities could be developed and integrated by increasingly automated means. framework for considering their implementation. To be comprehensive, AI services must of course include the service of developing new services, and 13 AGI-agent models entail greater complexity than current research practice shows that expanding the scope of AI services can CAIS be both incremental and increasingly automated. Relative to comprehensive AI services (CAIS), and contrary to 13.3 Classic AGI models neither simplify nor explain general AI widespread intuitions, the classic AGI-agent model implicitly increases capabilities (while obscuring) the complexity and challenges of self-improvement, general functionality, and AI goal alignment. 13.1 Summary • Section 7: Training agents in human-like environments can provide useful, bounded services • Section 10: R&D automation dissociates recursive improvement from AI agency • Section 15: Development-oriented models align with deeply-structured AI sys- tems • Section 23: AI development systems can support effective human guidance • Section 24: Human oversight need not impede fast, recursive AI technology improvement The classic AGI agent model posits open-ended self improvement, but this simple concept hides what by nature must be functionally equivalent to fully-automated and open-ended AI research and development. Hiding the complexity of AI development in a conceptual box provides only the illusion of simplicity. Discussion within the classic AGI model typically assumes an unexplained breakthrough in machine learning capabilities. For simplicity, an AI-services model could arbitrarily assume equivalent capabilities (per- 1. Section 20: Collusion among superintelligent oracles can readily be avoided haps based on the same hypothetical breakthrough), a deeper model offers a Recent discussions suggest that it would be useful to compare the relative complexities of AGI-agent and comprehensive AI services (CAIS) models of general intelligence. The functional requirements for open-ended selfimprovement and general AI capabilities are the same in both instances, but made more difficult in classic AGI models, which require that fullygeneral functionality be internal to an autonomous, utility-directed agent. The rewards for accomplishing this compression of functionality are difficult to see. To attempt to encompass general human goals within the utility function of a single, powerful agent would reduce none of the challenges of aligning concrete AI behaviors with concrete human goals, yet would increase the scope for problematic outcomes. This extreme compression and its attendant problems are unnecessary: Task-oriented AI systems within the CAIS framework could apply high-level reasoning and broad understanding to a full spectrum of goals, coordinating open-ended, collectively-general AI capabilities to provide services that, though seamlessly integrated, need not individually or collectively behave as a unitary AGI agent.13.2 Classic AGI models neither simplify nor explain self improvement \n Comprehensive AI services necessarily include the service of developing useful AI agents with stable, bounded capabilities, but superintelligent-level CAIS could also be employed to implement general, autonomous, self-modifying systems that match the specifications for risky AGI agents. If not properly directed or constrained-a focus of current AI-safety research-such agents could pose catastrophic or even existential risks to humanity. The AI-services model suggests broadening studies of AI safety to explore potential applications of CAIS-enabled capabilities to risk-mitigating differential technology development, including AI-supported means for developing safe AGI agents. potentially problematic AGI agents while providing means for mitigating can provide services useful for managing such agents. Predictive models of their potential dangers. On the negative side, AI services could facilitate the human concerns 1 are a prominent example of such services; others include development of dangerous agents, empower bad actors, and accelerate the AI-enabled capabilities for AI-systems design, 2 analysis, 3 monitoring, and development of seductive AI applications with harmful effects. A further upgrade. 4 concern-avoiding perverse agent-like behaviors arising from interactions among service providers-calls for further study that draws on agent-centric 14.4 CAIS capabilities could facilitate the development of models. Taking the long view, the CAIS model suggests a technology-agnostic, dangerous AGI agents relatively path-independent perspective on potential means for managing SI-level AI risks. 14.2 Prospects for general, high-level AI services reframe AI risks • Section 12: AGI agents offer no compelling value • Section 11: Potential AGI-enabling technologies also enable comprehensive AI services 14 The AI-services model brings ample risks Accordingly, the CAIS model suggests that high-level agents will (or readily High-level AI services could facilitate the development or emergence of could) be developed in the context of safer, more tractable AI systems 6 that dangerous agents, empower bad actors, and accelerate the development of seductive AI applications with harmful effects. 1. Section 12: AGI agents offer no compelling value 2. Section 10: R&D automation dissociates recursive improvement from AI agency 14.1 Summary 3. Section 10: R&D automation dissociates recursive improvement from AI agency Prospects for general, high-level AI services reframe-but do not eliminate-a 4. Section 12: AGI agents offer no compelling value 5. Section 11: Potential AGI-enabling technologies also enable comprehensive range of AI risks. On the positive side, access to increasingly comprehen-AI services sive AI services (CAIS) can reduce the practical incentives for developing 6. Section 29: The AI-services model reframes the potential roles of AGI agents In a classic model of high-level AI risks, AI development leads to selfimproving agents that gain general capabilities and enormous power relative to the rest of the world. The AI-services model 1 points to a different prospect: Continued automation of AI R&D 2 (viewed as an increasingly-comprehensive set of development services) leads to a general ability to implement systems that provide AI services, ultimately scaling to a superintelligent level. Prospects for comprehensive AI services (CAIS) contrast sharply with classic expectations that center on AGI agents: The leading risks and remedies differ in both nature and context.14.3 CAIS capabilities could mitigate a range of AGI risksOn the positive side, capabilities within the CAIS model can be applied to mitigate AGI risks. The CAIS model arises naturally from current trends in AI development and outlines a more accessible path 3 to general AI capabilities; as a consequence, CAIS points to a future in which AGI agents have relatively low marginal instrumental value 4 and follow rather than lead the application of superintelligent-level AI functionality 5 to diverse problems. \n • Abstract and concrete models of AI R&D automation • Incremental R&D automation approaching the recursive regime • Conditions for problematic emergent behaviors in structured systems • Applications of end-to-end learning in structured systems • Unitary models as guides to potential risks in composite systems • Distinct functionalities and deep component integration • Safety guidelines for structured AI systems development • Potential incentives to pursue alternative paths • Potential incentives to violate safety guidelines • Implications of safety-relevant learning during AI development • Applications of task-focused AI capabilities to AI safety problems • Applications of superintelligent-level machine learning to predicting human approval Further Reading • Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame • Section 10: R&D automation dissociates recursive improvement from AI agency • Section 12: AGI agents offer no compelling value • Section 15: Development-oriented models align with deeply-structured AI sys- tems • Section 39: Tiling task-space with AI services can provide general AI capabilities • Section 35: Predictable aspects of future knowledge can inform AI safety strategies 16 Aggregated experience and centralized learning support AI-agent applications Centralized learning based on aggregated experience has strong advan- tages over local learning based on individual experience, and will likely dominate the development of advanced AI-agent applications. \n Advances in deep reinforcement learning and end-to-end training have raised questions regarding the likely nature of advanced AI systems. Does progress in deep RL naturally lead to undifferentiated, black-box AI systems with broad capabilities? Several considerations suggest otherwise, that RL techniques will instead provide task-focused competencies to heterogeneous systems. General AI services must by definition encompass broad capabilities, performing not a single task trained end-to-end, but many tasks that serve many ends and are trained accordingly. Even within relatively narrow tasks, we typically find a range of distinct subtasks that are best learned in depth to provide robust functionality applicable in a wider range of contexts. We can expect to see RL applied to the development of focused systems (whether base-level or managerial) with functionality that reflects the natural diversity and structure of tasks.17.2 RL and end-to-end training tend to produce black-box systemsMethods that employ end-to-end training and deep reinforcement learning (here termed simply \"deep RL\") have produced startling advances in areas that range from game play (Mnih et al. 2015)  to locomotion (Heess et al. 2017)  to neural-network design (Zoph and Le 2016) . In deep RL, what are effectively black-box systems learn to perform challenging tasks directly from reward signals, bypassing standard development methods. Advances in deep RL have opened fruitful directions for current research, but also raise questions regarding the likely nature (and safety) of advanced AI systems with more general competencies. • Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame • Section 7: Training agents in human-like environments can provide useful, bounded services • Section 18: Reinforcement learning systems are not equivalent to reward-seeking agents • Section 23: AI development systems can support effective human guidance • Section 36: Desiderata and directions for interim AI safety guidelines 17 End-to-end reinforcement learning is compatible with the AI-services model End-to-end training and reinforcement learning fit naturally within integrated AI-service architectures that exploit differentiated AI compo- nents. 17.1 Summary \n Because efficiency, quality, reliability, and safety all favor the development of functionally differentiated AI services, powerful RL techniques are best regarded as tools for implementing and improving AI services, not as harbingers of omnicompetent black-box AI. system-level functional transparency. 1 Further Reading • Section 7: Training agents in human-like environments can provide useful, bounded services • Section 9: Opaque algorithms are compatible with functional transparency and control • Section 12: AGI agents offer no compelling value • Section 18: Reinforcement learning systems are not equivalent to reward-seeking agents • Section 38: Broadly-capable systems coordinate narrower systems 18 Reinforcement learning systems are not equivalent to reward-seeking agents 1. Section 9: Opaque algorithms are compatible with functional transparency and control 2. Section 12: AGI agents offer no compelling value 3. Section 15: Development-oriented models align with deeply-structured AI systems 4. Section 18: Reinforcement learning systems are not equivalent to reward-seeking agents 5. Section 2: Standard definitions of \"superintelligence\" conflate learning with competence 6. Section 24: Human oversight need not impede fast, recursive AI technology improvement RL systems are (sometimes) used to train agents, but are not themselves agents that seek utility-like RL rewards. \n • Goal-content integrity: For AI systems, as with other software, functional (implicitly, \"goal\") integrity is typically critical. Security services that protect integrity are substantially orthogonal to functional services, however, and security services enable software upgrade and replacement rather than simply preserving what they protect.• Cognitive enhancement: At a global level, AI-supported progress in AI technologies can enable the implementation of systems with enhanced levels of intelligence, but most AI R&D tasks are more-or-less orthogonal to application-level tasks, and are bounded in scope and duration. • Technological perfection: At a global level, competition drives improvements in both hardware and software technologies; on inspection, one finds that this vast, multi-faceted pursuit resolves into a host of looselycoupled R&D tasks that are bounded in scope and duration. • Resource acquisition: AI systems typically acquire resources by providing value through competitive services (or disservices such as theft or fraud). All these goals are pursued today by entities in the global economy, a prototypical diffuse intelligent system. \n Because perverse collusion among AI systems would be fragile and readily avoided, there is no obstacle to applying diverse, high-level AI resources to problems of AI safety. • Section 28: Automating biomedical R&D does not require defining human wel- fare • Section 38: Broadly-capable systems coordinate narrower systems 20 Collusion among superintelligent oracles can readily be avoided • Section 8: Strong optimization can strongly constrain AI capabilities, behavior, and effects • Section 12: AGI agents offer no compelling value \n following sections will consider biomedical research (including cancer research tasks) from a general but more concrete, less unitary perspective, concluding that undertaking AI-driven biomedical research need not risk programs based on criminality (kidnapping, etc.) or catastrophic problems of value alignment. (I thank Shahar Avin for suggesting this topic as a case study.) Diverse roles and tasks in biomedical research and applications: Scientific research: Research direction: Developing techniques Resource allocation Implementing experiments Project management Modeling biological systems Competitors, reviewers Clinical practice: Oversight: Patients Citizen's groups Physicians Regulatory agencies Health service providers Legislatures 28.3 Diverse AI systems could automate and coordinate diverse research tasks \n AI R&D is currently distributed across many independent research groups, and the architecture of R&D automation is compatible with continued decentralized development. Various pressures tend to favor greater centralization of development in leading organizations, yet centralization per se would neither force nor strongly favor a qualitative change in the architecture of R&D tasks. Alternative distributions of capabilities across organizations could provide affordances relevant to AI policy and strategy. • Section 15: Development-oriented models align with deeply-structured AI sys- tems • Section 20: Collusion among superintelligent oracles can readily be avoided 34 R&D automation is compatible with both strong and weak centralization Advances in AI R&D automation are currently distributed across many independent research groups, but a range of pressures could potentially lead to strong centralization of capabilities. 34.1 Summary • Section 10: R&D automation dissociates recursive improvement from AI agency \n Desideratum: Clarify why current AI research is safeCurrent AI research is safe (in a classic x-risk sense), in part because current AI capabilities are limited, but also because of the way capabilities are developed and organized. Interim guidelines could clarify and codify aspects of current practice that promote foundational aspects of safety (see below), and thereby support efforts to identify safe paths to more powerful capabilities. Because current AI development work is, in fact, safe with respect to high-level risks, interim safety guidelines can clarify and codify safety-aligned charac- teristics of current work while placing little burden on practitioners. Good practice in current AI R&D tends to align with longer-term safety concerns: Examples include learning from the exploration of families of architectures and tasks, then pursuing task-oriented development, testing, and validation before building complex deployed systems. These practices can contribute to shaping and controlling AI capabilities across a range of potential devel- opment paths. Development and adoption of guidelines founded on current practice could help researchers answer public questions about the safety of their work while fostering ongoing, collaborative safety research and guideline extension to address potential longer-term, high-level risks. • Section 12: AGI agents offer no compelling value 36.2 Desiderata • Section 15: Development-oriented models align with deeply-structured AI sys-tems • Section 16: Aggregated experience and centralized learning support AI-agent applications • Section 22: Machine learning can develop predictive models of human approval • Section 24: Human oversight need not impede fast, recursive AI technology improvement 36.2.1 36.2.2 Desideratum: Promote continued safety-enabling development 36 Desiderata and directions for interim AI safety practice guidelines Guidelines that focus on safety-promoting aspects of current practice can be crafted to place little burden on what is already safety-compliant research; Interim AI safety guidelines should (and could) engage with present practice, these same safety-promoting practices can contribute to (though not in them-place little burden on practitioners, foster future safety-oriented development, and promote an ongoing process of guideline development and adoption. selves ensure) avoiding hazards in more challenging future situations. 36.2.3 Desideratum: Foster ongoing, collaborative guideline 36.1 Summary development and adoption Actionable, effective interim AI safety guidelines should: Collaboration on actionable interim safety guidelines could promote closer • Clarify why current AI research is safe, links between development-and safety-oriented AI researchers, fostering • Promote continued safety-enabling development practice, and ongoing collaborative, forward-looking guideline development. Beginning • Foster ongoing, collaborative guideline development and adoption. with readily-actionable guidelines can help to ensure that collaboration goes beyond theory and talk. \n The absence of current safety risks sets a low bar for the effectiveness of interim guidelines, yet guidelines organized around current practice can contribute to the development of guidelines that address more challenging future concerns. At a technical level, practices that support reliability in narrow AI components can provide foundations for the safe implementation of more capable systems. At an institutional level, linking current practice to longer-term concerns can foster safety-oriented research and development in several ways: by encouraging understanding and extension of what today constitutes good practice, by engaging the development community in ongoing guideline development, and by focusing greater research attention on the potential connections between development processes and safe outcomes. Interim guidelines cannot solve all problems, yet could help to set our work on a productive path. et al. 2015]) In particular, the distinction between AI development systems and their products 1 enables a range of reliability (hence safety) oriented practices. 2 The use of modular architectures in current practice again suggests 37 How do neural and symbolic technologies mesh? opportunities for codification, explanation, and contributions to future safety. 36.7 Interim safety guidelines can foster ongoing progress Further Reading • Section 10: R&D automation dissociates recursive improvement from AI agency • Section 12: AGI agents offer no compelling value • Section 22: Machine learning can develop predictive models of human approval • Section 23: AI development systems can support effective human guidance • Section 24: Human oversight need not impede fast, recursive AI technology improvement • Section 35: Predictable aspects of future knowledge can inform AI safety strategies 1. Section 35: Predictable aspects of future knowledge can inform AI safety strategies \n combination of NN-level pattern recognition and loopy graphical models: \"Recursive Cortical Networks\" can generalize far better than conventional NNs, and from far less training data (George et al. 2017) .In neural program induction, NNs are trained to replicate the input-output behavior of S/A programs. Several architectures (e.g., the Neural Turing Machine [Graves, Wayne, and Danihelka 2014] , Neural GPU [Kaiser and Sutskever 2015] , and Neural Programmer [Neelakantan, Le, and Sutskever 2015] ; reviewed in Kant [2018] ) have yielded substantial success on simple algorithms, with (usually imperfect) generalization to problem-instances larger than those in the training set. We can anticipate that advances in graph NNs will facilitate the exploitation of richer structures in code, which has traditionally centered on narrower syntactic representations. Structures latent in S/A code, but to date not explored in conjunction, include not only abstract syntax trees, data types, and function types, but control and data flow graphs (May 2018 update: see 37.7.6 S/A algorithms can template NN algorithms: S/A representations can readily implement geometric models with parts, wholes, and relationships among distinct objects. In recent NN architectures, \"capsules\" (Sabour, Frosst, and Hinton 2017) and similar components (Liao and Poggio 2017) can both support image recognition and play a symbol- like role in representing part-whole relationships; architectures that embody relation-oriented priors can both recognize objects and reason about their relationships (Hu et al. 2018) or model their physical interactions (Battaglia et al. 2016; Chang et al. 2016). The neural architectures that accomplish these tasks follow (though not closely!) patterns found in S/A information processing. More broadly, there is a recognized trend of learning differentiable 37.7.4 S/A mechanisms can structure NN computation: versions of familiar algorithms (Guez et al. 2018). S/A mechanisms are used both to construct specified graphs and to exe-37.7.7 NNs can learn S/A algorithms: cute corresponding message-passing algorithms, while specification of graph- structured NNs is a task well-suited to mixed NN and S/A mechanisms. For example, S/A and NN mechanisms have been interleaved to compose and search over alternative tree-structured NNs that implement generative models of 3D structures (Jun Li et al. 2017). In a very different example (a visual question-answering task), an S/A parsing mechanism directs the assembly of question-specific deep networks from smaller NN modules (Hu et al. 2018); see also Andreas et al. (2016). A similar strategy is employed in the Neural Rule Engine (Li, Xu, and Lu 2018), while Yi et al. (2018) employs NN mecha-nisms to parse scenes and questions into symbolic representations that drive 37.7.8 NNs can aid S/A program synthesis: a symbolic execution engine. Automatic programming is a long-standing goal in AI research, but progress has been slow. Mixed S/A-NN methods have been applied to program synthe- 37.7.5 NNs can produce and apply S/A representations: sis (Yin and Neubig 2017; Singh and Kohli 2017; Abolafia et al. 2018), with increasing success (reviewed in (Kant 2018)). Potential low-hanging fruit in-NN mechanisms can learn discrete, symbol-like representations useful in cludes adaptation of existing source code (Allamanis and Brockschmidt 2017) reasoning, planning, and predictive learning (van den Oord, Vinyals, and and aiding human programmers by code completion (Jian Li et al. 2017); au-Kavukcuoglu 2017; Raiman and Raiman 2018), including string encodings tomating the development of glue code would facilitate integration of existing of molecular-structure graphs for chemical synthesis (Segler, Preuß, and Waller 2017) and drug discovery (Merk et al. 2018). S/A representations can S/A functionality with other S/A and NN components. enforce a strong inductive bias toward generalization; for example, deep RL systems can formulate and apply symbolic rules (Garnelo, Arulkumaran, and Shanahan 2016), and NN techniques can be combined with inductive logic programming to enable the inference of universal rules from noisy data (Evans and Grefenstette 2018). \n ). New tools for architecture and hyperparameter search are accelerating NN development by discovering new architectures (Zoph and Le 2016; Pham et al. 2018) and optimizing hyperparameters (Jaderberg et al. 2017) (a key task in architecture development). Leading methods in architecture search apply NN or evolutionary algorithms to propose candidate architectures, while using an S/A infrastructure to construct and test them. 37.7.10 S/A mechanisms can support NN architecture and hyperparameter search: \n Appropriate levels of abstraction depend on both our knowledge and our purpose. If we want to model the role of Earth in the dynamics of the Solar System, it can be treated as a point mass. If we want to model Earth as a context for humanity, however, we also care about its radius, geography, geology, climate, and more-and we have substantial, useful knowledge of these. Likewise, for some purposes, it can be appropriate to model prospective AI systems as undifferentiated, black-box pools of capabilities. If we want to understand prospective AI systems in the context of human society, however, we have strong practical reasons to apply what we know about the general architecture of systems that perform broad tasks. 38.7 Black-box abstractions discard what we know about the architecture of systems with broad capabilities \n • Section 12: AGI agents offer no compelling value • Section 23: AI development systems can support effective human guidance • Section 24: Human oversight need not impede fast, recursive AI technology improvement • Section 36: Desiderata and directions for interim AI safety guidelines \n\t\t\t .1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24.2 Must pressure to accelerate AI technology development increase risk? . . . 24.3 Basic technology research differs from world-oriented applications . . . . . 24.4 We can distinguish between human participation, guidance, and monitoring . 24.5 Guidance and monitoring can operate outside the central AI R&D loop . . . 24.6 Fast, asymptotically-recursive basic research need not sacrifice safety . . . . 24.7 World-oriented applications bring a different range of concerns . . . . . . \n\t\t\t The widespread (yet questionable) assumption that this feedback process-\"recursive improvement\"-would entail self-transformation of a particular AI \n\t\t\t . Section 18: Reinforcement learning systems are not equivalent to reward-seeking agents \n\t\t\t . Section 20: Collusion among superintelligent oracles can readily be avoided 3. Section 12: AGI agents offer no compelling value \n\t\t\t . Section 16: Aggregated experience and centralized learning support AI-agent applications. \n\t\t\t . Section 16: Aggregated experience and centralized learning support AI-agent applications. \n\t\t\t . Section 24: Human oversight need not impede fast, recursive AI technology improvement.3. Section 14:The AI-services model brings ample risks. \n\t\t\t . Section 16: Aggregated experience and centralized learning support AI-agent applications 3. Section 21: Broad world knowledge can support safe task performance 4. Section 33: Competitive AI capabilities will not be boxed 5. Section 33: Competitive AI capabilities will not be boxed 6. Section 10: R&D automation dissociates recursive improvement from AI agency \n\t\t\t The conditions for von Neumann-Morgenstern rationality do not imply that systems composed of AI services will act as utility-maximizing agents, hence the design space for manageable superintelligent-level systems is broader than often supposed.6.1 SummaryAlthough a common story suggests that any system composed of rational, highlevel AI agents should (or must?) be regarded as a single, potentially powerful agent, the case for this idea is extraordinarily weak. AI service providers can readily satisfy the conditions for VNM rationality while employing knowledge and reasoning capacity of any level or scope. Bostrom's Orthogonality Thesis implies that even VNM-rational, SI-level agents need not maximize broad utility functions, and as is well known, systems composed of rational agents need not maximize any utility function at all. In particular, systems composed of competing AI service providers cannot usefully be regarded as unitary agents, much less as a unitary, forward-planning, utility-maximizing AGI agent. If, as seems likely, much of the potential solution-space for AI safety requires affordances like those in the AI-services model, then we must reconsider 1. Section 12: AGI agents offer no compelling value \n\t\t\t . Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame. \n\t\t\t . Section 20: Collusion among superintelligent oracles can readily be avoided 2.https://scholar.google.co.uk/scholar?as_ylo=2017&q=generative+ adversarial+networks&hl=en&as_sdt=0,5 \n\t\t\t . Section 35: Predictable aspects of future knowledge can inform AI safety strategies 3. Section 22: Machine learning can develop predictive models of human approval 4. Section 20: Collusion among superintelligent oracles can readily be avoided 5. Section 32: Unaligned superintelligent agents need not threaten world stability \n\t\t\t . Section 2: Standard definitions of \"superintelligence\" conflate learning with competence \n\t\t\t . Section 38: Broadly-capable systems coordinate narrower systems \n\t\t\t Strong (even superintelligent-level) optimization can be applied to increase AI safety by strongly constraining the capabilities, behavior, and effects of AI systems.8.1 SummaryStrong \"optimization power\" has often been assumed to increase AI risks by increasing the scope of a system's capabilities, yet task-focused optimization can have the opposite effect. Optimization of any system for a task constrains its structure and behavior, implicitly constraining its off-task capabilities: A competitive race car cannot transport a load of passengers, and a bus will never set a land speed record. In an AI context, optimization will tend to constrain capabilities and decrease risks when objectives are bounded in space, time, and scope, and when objective functions assign costs to both resource consumption and off-task effects. Fortunately, these are natural conditions for AI services. Optimizing AI systems for bounded tasks is itself a bounded task, and some bounded tasks (e.g., predicting human approval) can contribute to general AI safety. These considerations indicate that strong, even SI-level optimization can both improve and constrain AI performance. \n\t\t\t . Section 22: Machine learning can develop predictive models of human approval \n\t\t\t . Section 2: Standard definitions of \"superintelligence\" conflate learning with competence 2. Section 8: Strong optimization can strongly constrain AI capabilities, behavior, and effects \n\t\t\t . Section 39: Tiling task-space with AI services can provide general AI capabilities 2. Section 39: Tiling task-space with AI services can provide general AI capabilities \n\t\t\t . Section 24: Human oversight need not impede fast, recursive AI technology improvement \n\t\t\t . Section 10: R&D automation dissociates recursive improvement from AI agency 2. Section 12: AGI agents offer no compelling value \n\t\t\t . Section 15: Development-oriented models align with deeply-structured AI systems \n\t\t\t . Section 20: Collusion among superintelligent oracles can readily be avoided \n\t\t\t . Section 9: Opaque algorithms are compatible with functional transparency and control \n\t\t\t . For example, an RL system can learn a predictive model of a human observer's approval at the level of actions, learning to perform difficult tasks without a specified goal or reward: see Christiano et al. (2017) .2. Section 16: Aggregated experience and centralized learning support AI-agent applications \n\t\t\t . Section 15: Development-oriented models align with deeply-structured AI systems \n\t\t\t . Section 8: Strong optimization can strongly constrain AI capabilities, behavior, and effects \n\t\t\t . In a familiar class of worst-case models, systems with general superintelligence would infer extensive knowledge about the world from minimal information, and would choose (if possible) to pursue potentially dangerous goals by manipulating the external environment, e.g., through deceptive answers to queries. In this model, (all?) superintelligent systems, even if almost isolated, would infer the existence of others like themselves, and (all?) would employ a decision theory that induces them to collude (in a coordinated way?) to pursue shared objectives. Even if we grant the initial worst-case assumptions, the argument presented above indicates that systems with these extraordinary capabilities would correctly infer the existence of superintelligent-level systems unlike themselves (systems with diverse and specialized capabilities, knowledge, and interactions, playing roles that include adversarial judges and competitors), and would correctly recognize that collusive deception is risky or infeasible. \n\t\t\t . Section 40: Could 1 PFLOP/s systems exceed the basic functional capacity of the human brain? \n\t\t\t . Section 15: Development-oriented models align with deeply-structured AI systems 2. Section 12: AGI agents offer no compelling value 3. Section 22: Machine learning can develop predictive models of human approval \n\t\t\t . Note that learning predictive models of a human observer's approval can enable an RL system to learn difficult tasks: see Christiano et al. (2017) . \n\t\t\t . Section 15: Development-oriented models align with deeply-structured AI systems \n\t\t\t . Section 22: Machine learning can develop predictive models of human approval 2. Section 22: Machine learning can develop predictive models of human approval \n\t\t\t . Section 16: Aggregated experience and centralized learning support AI-agent applications \n\t\t\t . Section 10: R&D automation dissociates recursive improvement from AI agency \n\t\t\t . Section 35: Predictable aspects of future knowledge can inform AI safety strategies \n\t\t\t . Section 10: R&D automation dissociates recursive improvement from AI agency \n\t\t\t Superintelligent-level AI systems can safely converse with humans, perform creative search, and propose designs for systems to be implemented and deployed in the world. \n\t\t\t . Section 25: Optimized advice need not be optimized to induce its acceptance \n\t\t\t Biomedical research and applications comprise extraordinarily diverse activities. Development of even a single diagnostic or therapeutic technology, for 1. This quote comes from an earlier draft (MIRI Technical report 2015-4), available at https://intelligence.org/files/obsolete/ValueLearningProblem.pdf \n\t\t\t . Section 12: AGI agents offer no compelling value 2. Section 10: R&D automation dissociates recursive improvement from AI agency 3. Section 23: AI development systems can support effective human guidance 4. Section 22: Machine learning can develop predictive models of human approval 5. Section 21: Broad world knowledge can support safe task performance \n\t\t\t . For example, \"Ems\" (Hanson 2016)  2. Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame \n\t\t\t . Section 2: Standard definitions of \"superintelligence\" conflate learning with competence 2. Section 5: Rational-agent models place intelligence in an implicitly anthropomorphic frame \n\t\t\t . Section 19: The orthogonality thesis undercuts the generality of instrumental convergence 2. Section 20: Collusion among superintelligent oracles can readily be avoided \n\t\t\t Bostrom (2014, p.93) introduces the concept of a \"superpower\" as a property of \"a system that sufficiently excels\" in one of the strategically critical tasks, stating that \"[a] full-blown superintelligence would greatly excel at all of these tasks,\" and later explains that \"excels\" must be understood in a relative sense that entails a strong situational asymmetry(Bostrom 2014, p.104):1. Section 10: R&D automation dissociates recursive improvement from AI agency \n\t\t\t . Section 10: R&D automation dissociates recursive improvement from AI agency \n\t\t\t . Section 15: Development-oriented models align with deeply-structured AI systems 2. Section 20: Collusion among superintelligent oracles can readily be avoided \n\t\t\t Fundamental arguments suggest that AI R&D automation provides the most direct path 2 to steeply accelerating, AI-enabled progress in AI technologies.1. Section 10: R&D automation dissociates recursive improvement from AI agency 2. Section 1: R&D automation provides the most direct path to an intelligence explosion \n\t\t\t . Section 40: Could 1 PFLOP/s systems exceed the basic functional capacity of the human brain? \n\t\t\t . Section 16: Aggregated experience and centralized learning support AI-agent applications", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Reframing_Superintelligence_FHI-TR-2019-1.1-1.tei.xml", "id": "e52425ad432c210730245d1ecac6ae68"}
+{"source": "reports", "source_filetype": "pdf", "abstract": "Militaries around the world believe that the integration of machine learning methods throughout their forces could improve their effectiveness. From algorithms to aid in recruiting and promotion, to those designed for surveillance and early warning, to those used directly on the battlefield, applications of artificial intelligence (AI) could shape the future character of warfare. These uses could also generate significant risks for international stability. These risks relate to broad facets of AI that could shape warfare, limits to machine learning methods that could increase the risks of inadvertent conflict, and specific mission areas, such as nuclear operations, where the use of AI could be dangerous. To reduce these risks and promote international stability, we explore the potential use of confidence-building measures (CBMs), constructed around the shared interests that all countries have in preventing inadvertent war. Though not a panacea, CBMs could create standards for information-sharing and notifications about AIenabled systems that make inadvertent conflict less likely.", "authors": ["Michael C Horowitz", "Paul Scharre"], "title": "Risks and Confidence-Building Measures", "text": "Introduction In recent years, the machine learning revolution has sparked a wave of interest in artificial intelligence (AI) applications across a range of industries. Nations are also mobilizing to use AI for national security and military purposes.  1  It is therefore vital to assess how the militarization of AI could affect international stability and how to encourage militaries to adopt AI in a responsible manner. Doing so requires understanding the features of AI, the ways it could shape warfare, and the risks to international stability resulting from the militarization of artificial intelligence. AI is a general-purpose technology akin to computers or the internal combustion engine, not a discrete technology like missiles or aircraft. Thus, while concerns of an \"AI arms race\" are overblown, real risks exist.  2  Additionally, despite the rhetoric of many national leaders, military spending on AI is relatively modest to date. Rather than a fervent arms race, militaries' pursuit of AI looks more like routine adoption of new technologies and a continuation of the multi-decade trend of adoption of computers, networking, and other information technologies. Nevertheless, the incorporation of AI into national security applications and warfare poses genuine risks. Recognizing the risks is not enough, however. Addressing them requires laying out suggestions for practical steps states can take to minimize risks stemming from military AI competition.  3  One approach states could take is adopting confidence-building measures (CBMs): unilateral, bilateral, and/or multilateral actions that states can take to build trust and prevent inadvertent military conflict. CBMs generally involve using transparency, notification, and monitoring to attempt to mitigate the risk of conflict.  4  There are challenges involved in CBM adoption due to differences in the character of international competition today versus during the Cold War, when CBMs became prominent as a concept. However, considering possibilities for CBMs and exploring ways to shape the dialogue about AI could make the adoption of stability-promoting CBMs more likely. This paper briefly outlines some of the potential risks to international stability arising from military applications of AI, including ways AI could influence the character of warfare, risks based on the current limits of AI technology, and risks relating to some specific mission areas, such as nuclear operations, in which introducing AI could present challenges to stability. The paper then describes possible CBMs to address these risks, moving from broad measures applicable to many military applications of AI to targeted measures designed to address specific risks. In each discussion of CBMs, the paper lays out both the opportunities and potential downsides of states adopting the CBM. \n Military Uses of AI: A Risk to International Stability? Militaries have an inherent interest in staying ahead of their competitors, or at least not falling behind. National militaries want to avoid fielding inferior military capabilities and so will generally pursue emerging technologies that could improve their ability to fight. While the pursuit of new technologies is normal, some technologies raise concerns because of their impact on stability or their potential to shift warfare in a direction that causes net increased harm for all (combatants and/or civilians). For example, around the turn of the 20th century, great powers debated, with mixed results, arms control against a host of industrial era technologies that they feared could alter warfare in profound ways. These included submarines, air-delivered weapons, exploding bullets, and poison gas. After the invention of nuclear weapons, concerns surrounding their potential use dominated the attention of policymakers given the weapons' sheer destructive potential. Especially after the Cuban Missile Crisis illustrated the very real risk of escalation, the United States and the Soviet Union engaged in arms control on a range of weapons technologies, including strategic missile defense, intermediate-range missiles, space-based weapons of mass destruction (WMDs), biological weapons, and apparent tacit restraint in neutron bombs and anti-satellite weapons. The United States and the Soviet Union also, at times, cooperated to avoid miscalculation and improve stability through measures such as the Open Skies Treaty and the 1972 Incidents at Sea Agreement. It is reasonable and, in fact, vital to examine whether the integration of AI into warfare might also pose risks that policymakers should attend. Some AI researchers themselves have raised alarm at militaries' adoption of AI and the way it could increase the risk of war and international instability.  5  Understanding risks stemming from military use of AI is complicated, however, by the fact that AI is not a discrete technology like missiles or submarines. As a general-purpose technology, AI has many applications, any of which could, individually, improve or undermine stability in various ways. Militaries are only beginning the process of adopting AI, and in the near term, military AI use is likely to be limited and incremental. Over time, the cognization of warfare through the introduction of artificial intelligence could change warfare in profound ways, just as industrial revolutions in the past shaped warfare.  6  Even if militaries successfully manage safety and security concerns and field AI systems that are robust and secure, properly functioning AI systems could create challenges for international stability. For example, both Chinese and American scholars have hypothesized that the introduction of AI and autonomous systems in combat operations could accelerate the tempo of warfare beyond the pace of human control. Chinese scholars have referred to this concept as a battlefield \"singularity,\"  7  while some Americans have coined the term \"hyperwar\" to refer to a similar idea.  8  If warfare evolves to a point where the pace of combat outpaces humans' ability to keep up, and therefore control over military operations must be handed to machines, it would pose significant risks for international stability, even if the delegation decision seems necessary due to competitive pressure. Humans might lose control over managing escalation, and war termination could be significantly complicated if machines fight at a pace that is faster than humans can respond. In addition, delegation of escalation control to machines could mean that minor tactical missteps or accidents that are part and parcel of military operations in the chaos and fog of war, including fratricide, civilian casualties, and poor military judgment, could spiral out of control and reach catastrophic proportions before humans have time to intervene. The logic of a battlefield singularity, or hyperwar, is troubling precisely because competitive pressures could drive militaries to accelerate the tempo of operations and remove humans \"from the loop,\" even if they would rather not, in order to keep pace with adversaries. Then-Deputy Secretary of Defense Robert Work succinctly captured this dilemma when he posed the question, \"If our competitors go to Terminators ... and it turns out the Terminators are able to make decisions faster, even if they're bad, how would we respond?\"  9  While this \"arms race in speed\" is often characterized tactically in the context of lethal autonomous weapon systems, the same dynamic could emerge operationally involving algorithms designed as decision aids. The perception by policymakers that war is evolving to an era of machinedominated conflict in which humans must cede control to machines to remain competitive could also hasten such a development, particularly if decision makers lack appropriate education about the limits of AI. In extremis, the shift toward the use of algorithms for military decision-making, combined with a more roboticized battlefield, could potentially change the nature of war. War would still be the continuation of politics by other means in the broadest sense, but in the most extreme case it might feature so little human engagement that it is no longer a fundamentally human endeavor.  10  The widespread adoption of AI could have a net effect on international stability in other ways. AI systems could change strategy in war, including by substituting machines for human decision-making in some mission areas, and therefore removing certain aspects of human psychology from parts of war.  11  Warfare today is waged by humans through physical machinery (rockets, missiles, machine guns, etc.), but decision-making is almost universally human. As algorithms creep closer to the battlefield, some decisions will be made by machines even if warfare remains a human-directed activity that is fought for human political purposes. The widespread integration of machine decision-making across tactical, operational, and strategic levels of warfare could have far-reaching implications. Already, AI agents playing real-time computer strategy games such as StarCraft and Dota 2 have demonstrated superhuman aggressiveness, precision, and coordination. In other strategy games such as poker and Go, AI agents have demonstrated an ability to radically adjust playing styles and risk-taking in ways that would be, at best, challenging for humans to mimic for psychological reasons. AI dogfighting agents have similarly demonstrated superhuman precision and employed different tactics because of the ability to take greater risk to themselves.  12  In many ways, AI systems have the ability to be the perfect strategic agents, unencumbered by fear, loss aversion, commitment bias, or other human emotional or cognitive biases and limitations.  13  While the specific algorithms and models used for computer games are unlikely to transfer well to combat applications, the general characteristics and advantages of AI agents relative to humans could have applications in the military domain. As in the case of speed, the net consequence of machine decisionmaking on the psychology of combat could change the character of warfare in profound ways.  14  AI could have other cumulative effects on warfare. Policymakers generally assess adversaries' behavior based on an understanding of their capabilities and intentions. 15 Shifts toward AI could undermine policymaker knowledge in both of those arenas. The transition of military capabilities to software, already underway but arguably accelerated by the adoption of AI and autonomous systems, could make it harder for policymakers to accurately judge relative military capabilities. Incomplete information about adversary capabilities would therefore increase, conceivably increasing the risks of miscalculation. Alternatively, the opposite could be true-AI and autonomous systems used for intelligence collection and analysis could radically increase transparency about military power, making it easier for policymakers to judge military capabilities and anticipate the outcome of a conflict in advance. This added transparency could decrease the risks of miscalculation and defuse some potential conflicts before they begin. The integration of AI into military systems, in combination with a shift toward a more roboticized force structure, could also change policymakers' threshold for risk-taking, either because they believe that fewer human lives are at risk or that AI systems enable greater precision, or perhaps because they see AI systems as uniquely dangerous. The perceived availability of AI systems could change policymakers' beliefs about their ability to foresee the outcome of conflicts or to win. It is, no doubt, challenging to stand at the beginning of the AI age and imagine the cumulative consequence of AI adoption across varied aspects of military operations, including effects that hinge as much on human perception of the technology as the technical characteristics themselves. The history of attempts to regulate the effects of industrial age weapons in the late 19th and early 20th centuries suggests that even when policymakers accurately anticipated risks from certain technologies, such as airdelivered weapons or poison gas, they frequently crafted regulations that turned out to be ill-suited to the specific forms these technologies took as they matured. Furthermore, even when both sides desired restraint, it frequently (although not always) collapsed under the exigencies of war.  16  There is no reason to think that our prescience in predicting the path of future technologies or ability to restrain warfare is any better today. There is merit, however, in beginning the process of thinking about the many ways in which AI could influence warfare, big and small. Even beyond the scenarios described above, it is possible to frame how military applications of AI could impact international stability into two broad categories: (1) risks related to the character of algorithms and their use by militaries, and (2) risks related to militaries using AI for particular missions. \n RISKS DUE TO THE LIMITATIONS OF AI A challenge for military adoption of AI is that two key risks associated with new technology adoption are in tension. First, militaries could fail to adopt-or adopt quickly enough or employ in the right manner-a new technology that yields significant battlefield advantage. As a recent example, despite the overall growth in the military uninhabited, or unmanned, aircraft market, the adoption of uninhabited vehicles has, at times, been a source of contention within the U.S. defense establishment, principally based on debates over the merits of this new technology relative to existing alternatives. 17 Alternatively, militaries could adopt an immature technology too quickly, betting heavily and incorrectly on new and untested propositions about how a technology may change warfare. Given the natural incentive militaries have in ensuring their capabilities work on the battlefield, it may be reasonable to assume that militaries would manage these risks reasonably well, although not without some mishaps. But when balancing the risk of accidents versus falling behind adversaries in technological innovation, militaries arguably place safety as a secondary consideration.  18  Militaries may be relatively accepting of the risk of accidents in the pursuit of technological advantage, since accidents are a routine element of military operations, even in training.  19  Nevertheless, there are strong bureaucratic interests in ultimately ensuring that fielded capabilities are robust and secure, and existing institutional processes may be able to manage AI safety and security risks with some adaptation. For militaries, balancing between the risks of going too slow versus going too fast with AI adoption is complicated by the fact that AI, and deep learning in particular, is a relatively immature technology with significant vulnerabilities and reliability concerns.  20  These concerns are heightened in situations where there may not be ample data on which to train machine learning systems. Machine learning systems generally rely on very large data sets, which may not exist in some military settings, particularly when it comes to early warning of rare events (such as a nuclear attack) or tracking adversary behavior in a multidimensional battlefield. When trained with inadequate data sets or employed outside the narrow context of their design, AI systems are often unreliable and brittle. AI systems can often seem deceptively capable, performing well (sometimes better than humans) in some laboratory settings, then failing dramatically under changing environmental conditions in the real world. Self-driving cars, for example, may be safer than human drivers in some settings, then inexplicably turn deadly in situations where a human operator would not have trouble. Additionally, deep learning methods may, at present, be insufficiently reliable for safety-critical applications even when operating within the bounds of their design specifications.  21  For example, concerns about limits to the reliability of algorithms across demographic groups have hindered the deployment of facial recognition technology in the United States, particularly in highconsequence applications such as law enforcement. Militaries, too, should be concerned about technical limitations and vulnerabilities in their AI systems. Militaries want technologies that work, especially on the battlefield. Accordingly, the AI strategy of the Department of Defense (DoD) calls for AI systems that are \"resilient, robust, reliable, and secure.\" 22 This is undoubtedly the correct approach but a challenge, at least in the near term, given the reliability issues facing many uses of algorithms today and the highly dynamic conditions of battlefield use. An additional challenge stems from security dilemma dynamics. Competitive pressures could lead nations to shortcut test and evaluation (T&E) in a desire to field new AI capabilities ahead of adversaries. Similar competitive pressures to beat others to market appear to have played an exacerbating role in accident risk relating to AI systems in self-driving cars and commercial airplane autopilots.  23  Militaries evaluating an AI system of uncertain reliability could, not unjustifiably, feel pressure to hasten deployment if they believe others are taking similar measures. Historically, these pressures are highest immediately before and during wars, where the risk/reward equation surrounding new technologies can shift due to the very real lives on the line. For example, competitive pressures may have spurred the faster introduction of poison gas in World War I.  24  Similarly, in World War II, Germany diverted funds from proven technologies into jet engines, ballistic missiles, and helicopters, even though none of the technologies proved mature until after the war.  25  This dynamic risk might spark a self-fulfilling prophecy in which countries accelerate deployment of insufficiently tested AI systems out of the fear that others will deploy first.  26  The net effect is not an arms race but a \"race to the bottom\" on safety, leading to the deployment of unsafe AI systems and heightening the risk of accidents and instability. Even if military AI systems are adequately tested, the use of AI to enable more autonomous machine behavior in military systems raises an additional set of risks. In delegating decision-making from humans to machines, policymakers may de facto be fielding forces with less flexibility and ability to understand context, which would then have deleterious effects on crisis stability and managing escalation. While machines have many advantages in speed, precision, and repeatable actions, machines today cannot come close to human intelligence in understanding context and flexibly adapting to novel situations. This brittleness of machine decision-making may particularly be a challenge in pre-conflict crisis situations, where tensions among nations run high. Military forces from competing nations regularly interact in militarized disputes below the threshold of war in a variety of contested regions (e.g., the India-Pakistan border, China-India border, South China Sea, Black Sea, Syria, etc.). These interactions among deployed forces sometimes run the risk of escalation due to incidents or skirmishes that can inflame tensions on all sides. This poses a challenge for national leaders, who have imperfect command-and-control over their own military forces. Today, however, deployed military forces rely on human decision-making. Humans can understand broad guidance from their national leadership and commander's intent, such as \"defend our territorial claims, but don't start a war.\" Relative to humans, even the most advanced AI systems today have no ability to understand broad guidance, nor do they exhibit the kinds of contextual understanding that humans frequently label \"common sense.\" 27 Militaries already employ uninhabited vehicles (drones) in contested areas, which have been involved in a number of escalatory incidents in the East China Sea, South China Sea, Syria, and Strait of Hormuz. 28 Over time, as militaries incorporate more autonomous functionality into uninhabited vehicles, that functionality could complicate interactions in these and other contested areas. Autonomous systems may take actions based on programming that, while not a malfunction, are other than what a commander would have wanted a similarly situated human to do in the same situation. While the degree of flexibility afforded subordinates varies considerably by military culture and doctrine, humans have a greater ability to flexibly respond to complex and potentially ambiguous escalatory incidents in ways that may balance competing demands of ensuring national resolve while managing escalation. 29 Autonomous systems will simply follow their programming, whatever that may be, even if those rules no Center for a New American Security 1152 15 th Street NW, Suite 950, Washington, DC 20005 T: <PHONE> F: <PHONE> CNAS.org @CNASdc longer make sense or are inconsistent with a commander's intent in the given situation. This challenge is compounded by the fact that human commanders cannot anticipate all of the possible situations that forward-deployed military forces in contested regions may face. Employing autonomous systems in a crisis effectively forces human decision makers to tie their own hands with certain pre-specified actions, even if they would rather not. Unintended actions by autonomous systems in militarized disputes or contested areas are a challenge for militaries as they adopt more autonomous systems into their forces. The complexity of many autonomous systems used today, even ones that rely on rule-based decision-making, may mean that the humans employing autonomous systems lack sufficient understanding of what actions the system may take in certain situations.  30  Humans' ability to flexibly interpret guidance from higher commanders, even to the point of disregarding guidance if it no longer seems applicable, is by contrast a boon to managing escalation risks by retaining human decision-making at the point of interaction among military forces in contested regions.  31  Unintended escalation is not merely confined to lethal actions, such as firing on enemy forces. Nonlethal actions, such as crossing into another state's territory, can be perceived as escalatory. Even if such actions do not lead directly to war, they could heighten tensions, increase suspicion about an adversary's intentions, or inflame public sentiment. While in most cases, humans would still retain agency over how to respond to an incident, competing autonomous systems could create unexpected interactions or escalatory spirals. Complex, interactive dynamics between algorithms have been seen in other settings, including financial markets, 32 and even in situations where the algorithms are relatively simple.  33  Another problem stems from the potential inability of humans to call off autonomous systems once deployed. One reason for employing autonomous functionality is so that uninhabited vehicles can continue their missions even if they are operating without reliable communication links to human controllers. When there is no communication link between human operators and an autonomous system, human operators would have no ability to recall the autonomous system if political circumstances changed such that the system's behavior was no longer appropriate. This could be a challenge in de-escalating a conflict, if political leaders decide to terminate hostilities but have no ability to recall autonomous systems, at least for some period of time. The result could be a continuation of hostilities even after political leaders desire a ceasefire. Alternatively, the inability to fully cease hostilities could undermine truce negotiations, leading to the continuation of conflict. These problems are not unique to autonomous systems. Political leaders have imperfect command-and-control over human military forces, which has, at times, led to similar incidents with human-commanded deployed forces. For example, the Battle of New Orleans in the War of 1812 was fought after a peace treaty ended the war because of the slowness of communications to deployed forces. \n RISKS DUE TO THE USE OF AI FOR PARTICULAR MILITARY MISSIONS The introduction of AI into military operations could also pose risks in certain circumstances due to the nature of the military mission, even if the AI system performs correctly and consistent with human intentions. Some existing research already focuses on the intersection of AI with specific military mission areas, most notably nuclear stability.  34  Nuclear stability is an obvious area of concern given the potential consequences of an intentional or unintentional nuclear detonation.  35  Lethal autonomous weapon systems (LAWS), a particular use of AI in which lethal decision-making is delegated from humans to machines, also represents a focus area of existing research. Other areas may deserve special attention from scholars concerned about AI risks. The intersections of AI with cybersecurity and biosecurity are areas worthy of exploration where there has been relatively less work at present. Potentially risky applications of AI extend beyond the battlefield to the use of AI to aid in decision-making in areas such as early warning and forecasting adversary behavior. For example, AI tools to monitor, track, and analyze vast amounts of data on adversary behavior for early indications and warning of potential aggression have clear value. However, algorithms also have known limitations and potentially problematic characteristics, such as a lack of transparency or explainability, brittleness in the face of distributional shifts in data, and automation bias. AI systems frequently perform poorly under conditions of novelty, suggesting a continued role for human judgment. The human tendency toward automation bias, coupled with the history of false alarms generated by non-AI early warning and forecasting systems, suggests policymakers should approach the adoption of AI in early warning and forecasting with caution, despite the potential value of using AI in intelligent decision aids.  37  Education and training to ensure the responsible use of AI in early warning and forecasting scenarios will be critical.  38  Finally, autonomous systems raise novel challenges of signaling in contested areas because of ambiguity about whether their behavior was intended by human commanders. Even if the system performs as intended, adversaries may not know whether an autonomous system's behavior was consistent with human intent because of the aforementioned command-and-control issues. This can create ambiguity in a crisis situation about how to interpret an autonomous system's behavior. For example, if an autonomous system fired on a country's forces, should that be interpreted as an intentional signal by the commanding nation's political leaders, or an accident? This, again, is not a novel problem; a similar challenge exists with human-commanded military forces. Nations may not know whether the actions of an adversary's deployed forces are fully in line with their political leadership's guidance. Autonomous systems could complicate this dynamic due to uncertainty about whether the actions of an autonomous system are consistent with any human's intended action. \n The Role of Confidence-Building Measures AI potentially generates risks for international security due to ways AI could change the character of warfare, the limitations of AI technology today, and the use of AI for specific military missions such as nuclear operations. Especially given the uncertain technological trajectory of advances in AI, what are options to reduce the risks that military applications of AI can pose to international stability? To advance the conversation about ensuring that military AI adoption happens in the safest and most responsible way possible, this paper outlines a series of potential confidence-building measures aimed at mitigating risks from military uses of AI.  39  We introduce these ideas as preliminary concepts for future research, discussion, and examination, rather than to specifically advocate for any of these options. But progress in mitigating the risks from military AI competition requires moving beyond the recognition that risk mitigation is important to the hard work of suggesting, evaluating, and examining the benefits and drawbacks of specific mechanisms.  40  This paper focuses on confidence-building measures, a broad category of actions that states can take to reduce instability risks. CBMs include actions such as transparency, notification, and monitoring designed to reduce various risks arising from military competition between states. They generally encompass four areas, as Marie-France Desjardins describes: 41 Confidence-building measures are related to, but distinct from, arms control agreements. Arms control encompasses agreements states make to forgo researching, developing, producing, fielding, or employing certain weapons, features of weapons, or applications of weapons. The set of possible actions states could take is broad, and this paper will focus on the potential benefits and drawbacks of specific AIrelated confidence-building measures. Arms control for military AI applications is a valuable topic worthy of exploration, but beyond the scope of this paper. 42 \n HISTORICAL APPLICATIONS OF CBMS Confidence-building measures as a concept rose to prominence during the Cold War as a tool to reduce the risk of inadvertent war. In the wake of the Cuban Missile Crisis, the United States and the Soviet Union began exploring ways to improve their communication. While both sides recognized that war might occur, they had a shared interest, due to the potentially world-ending consequences of a global nuclear war, in ensuring that any such outbreak would be due to a deliberate decision, rather than an accident or a misunderstanding. The desire to build confidence led to a series of bilateral measures. Less than a year after the Cuban Missile Crisis, in June 1963, the United States and the Soviet Union signed a memorandum of understanding to create a hotline between the senior leadership of the two nations.  43  The idea was that this line of communication would provide a mechanism for U.S. and Soviet leaders to reach out to their counterparts and discuss crises in a way that made inadvertent escalation less likely. In 1972, as part of the Strategic Arms Limitation Talks (SALT I) arms control agreement, the United States and the Soviet Union went further, signing the Incidents at Sea Agreement, which they had been negotiating since 1967. The Incidents at Sea Agreement, not initially considered a prominent part of the 1972 SALT I accord, created a mechanism for communication and information surrounding the movement of U.S. and Soviet naval vessels. The agreement regulated dangerous maneuvers and harassment of vessels, established means for communicating the presence of submarines and surface naval movements, and generated a mechanism for regular consultation.  44  These successes helped lead to the formalization of the CBM concept in 1975 in Helsinki at the Conference on Security and Cooperation in Europe.  45  As the Cold War drew to a close, confidence-building measures expanded beyond the U.S.-Soviet context and the European theater. For example, India and China have a series of CBMs intended to prevent escalation in their disputed border area, while India and Pakistan have a hotline designed to make accidental escalation in South Asia less likely. In Southeast Asia, through the Regional Forum of the Association of Southeast Asian Nations (ASEAN), member nations have pursued CBMs designed to reduce the risk of conflict among themselves, and between any ASEAN member and China, due to territorial disputes in the South China Sea.  46  These CBMs used outside of the Cold War have had mixed effects. In the China-India case, for example, border-related CBMs did not prevent the ongoing conflict in 2020 between those two nations along the Line of Actual Control in the Himalayan region. However, norms surrounding the types of \"legitimate\" military activities promoted by CBMs have likely reduced the death toll of the clashes, with both sides generally avoiding the use of firearms, consistent with agreements from 1996 and 2005.  47  In Southeast Asia, while the ASEAN Regional Forum is a principal forum for dialogue, the consensusbased character of ASEAN makes it challenging for that dialogue to translate into policies on contested issues. Recent multilateral dialogues about emerging technologies such as cyber systems have also featured efforts to create CBMs that could be building blocks for cooperation. Unfortunately, a lack of international agreement on basic definitions and some countries' interest in dodging limitations on behavior in cyberspace have limited the development of effective norms.  48   \n CBMs rely on shared interests Center for a New American Security 1152 15 th Street NW, Suite 950, Washington, DC 20005 T: <PHONE> F: <PHONE> CNAS.org @CNASdc to succeed, and major powers such as the United States, China, and Russia do not have clearly shared interests concerning behavior in cyberspace, making it difficult to use CBMs to build trust or successfully design \"rules of the road\" agreements likely to generate widespread adherence. CBMs may be a useful tool for managing risks relating to military AI applications. There are a number of possible CBMs that states could adopt that may help mitigate the various AI-related risks previously outlined. These include broad CBMs applicable to AI as a category, CBMs designed to address some of the limitations of AI, and CBMs focused on specific missions for which militaries might use AI. 49 \n BROAD CBMS These CBMs focus broadly on mechanisms for dialogue and agreement surrounding military uses of AI, rather than the specific content of agreements. Given that a key goal of CBMs is to enhance trust, mechanisms that serve as a building block for more substantive dialogue and agreement can, in some cases, be an end in themselves and not just a means to an end.  50  These could include promoting international norms for how nations develop and use military AI systems, Track II academic-to-academic exchanges, direct military-to-military dialogues, and agreements between states regarding military AI, such as a code of conduct or mutual statement of principles. \n Promoting Norms In 2019, the U.S. Defense Innovation Board proposed a set of AI principles for the U.S. Defense Department, which DoD subsequently adopted in early 2020. While these principles no doubt have domestic audiences in the U.S. defense community and tech sector, they also serve as an early example of a state promulgating norms about appropriate use of AI in military applications. The DoD AI principles included a requirement that DoD AI systems be responsible, equitable, traceable, reliable, and governable. 51 (The full set of DoD AI principles is included in the Appendix). Similarly, the DoD's unclassified summary of its AI strategy, released in 2019, called for building AI systems that were \"resilient, robust, reliable, and secure.\" 52 A focus of the strategy was \"leading in military ethics and AI safety.\"  53  There is value in states promoting norms for responsible use of AI, including adopting and employing technology in a way that reflects an understanding of the technical risks associated with AI systems. While stating such principles is not the same as putting in place effective bureaucratic processes to ensure their compliance, there is nevertheless value in states publicly signaling to others (and to their own bureaucracies) the importance of using AI responsibly in military applications. While these norms are at a high level, they nevertheless signal some degree of attention by senior military and civilian defense officials to some of the risks of AI systems, including issues surrounding safety, security, responsibility, and controllability. These signals may aid internal bureaucratic efforts to mitigate various AI-related risks, as bureaucratic actors can point to these official documents for support. Additionally, to the extent that other nations find these statements credible, they may help signal to other nations at least some degree of awareness and attention to these risks, helping to incentivize others to do the same. One risk to such statements is that if they appear manifestly at odds with a state's actions, they can ring hollow, undermine a state's credibility, or undermine the norm itself. For example, loudly proclaiming the importance of AI ethics while using AI systems in a clearly unethical manner, such as for internal repression or without regard for civilian casualties, could not only undermine a state's credibility but also undermine the value of the norm overall, especially if other states fail to highlight the disconnect. Following through with meaningful actions to show how a state puts these norms into practice is essential for them to have real value. \n Track II Academic Dialogues One confidence-building measure is already underway: Track II dialogues between academic experts from different countries with expertise surrounding military uses of AI.  54  Because these dialogues occur among experts who are not government officials, they are low risk because they do not commit countries to actually doing anything. This also places a cap on their potential benefits. Track II dialogues can nevertheless be useful building blocks for more substantive cooperation among countries and an avenue to explore various potential modes of cooperation without fear of commitment by states. Track II dialogues can help facilitate mutual understanding among expert communities in different states and build shared trust between experts.  55  Additionally, if some of those experts transition into government positions in the future, the lessons from these dialogues can improve the prospects for cooperation in more formal venues. While there are risks to misleading statements in the context of formal government dialogues, as discussed below, the consequences of such activities in a Track II context are minimal. The nature of the dialogue is that participants are not government officials and it is to be expected that some of their statements may not be entirely in line with their government's policies. Thus, Track II dialogues can build trust and be an end in themselves, even as they serve as the means to broader cooperation and understanding. \n Military-to-Military Dialogues Direct military-to-military engagement on deconfliction measures for AI and autonomous systems could be valuable, both as a precursor to potentially more fulsome specific measures, but also a valuable communication mechanism in their own right. For example, if militaries deploy an autonomous vehicle into a contested area where other military forces will be operating, a direct military-to-military channel would give the other side an opportunity to ask questions about its behavior and the deploying side an opportunity to communicate expectations, to avoid unintended escalation or incidents. Similarly, such a venue would give militaries an opportunity to ask questions and communicate information about other capabilities or investments that may threaten mutual stability, such as investments in AI, autonomy, or automation in nuclear operations. There are many advantages of direct, private communication over more indirect, public communication. Nations can send targeted messages just to the intended audience, rather than dealing with multiple audiences, including domestic ones. There may be reduced political pressure to save face or show strength publicly, although of course some of these pressures may still exist in private channels. And direct discussions afford more high-bandwidth information exchange with greater back-and-forth between sides than may be possible via public messages broadcast to a wider audience. One challenge, of course, is that these dialogues are most challenging precisely when they are needed the most, when there is a lack of transparency and trust on both sides. However, history shows that such dialogues are possible and indeed can be valuable measures in increasing transparency and reducing mutual risks. \n Code of Conduct Nations could agree to a written set of rules or principles for how they adopt AI into military systems. These rules and principles, even if not legally binding, could nevertheless serve a valuable signaling and coordination function to avoid some of the risks in AI adoption. A code of conduct, statement of principles, or other agreement could include a wide range of both general and specific statements, including potentially on any or all of the confidence-building measures listed above. Even if countries cannot agree on specific details beyond promoting safe and responsible military use of AI, more general statements could nevertheless be valuable in signaling to other nations some degree of mutual understanding about responsible use of military AI and help create positive norms of behavior. Ideally, a code of conduct would have support from a wide range of countries and major military powers. However, if this were not possible, then a multilateral statement of principles from like-minded countries could still have some value in increasing transparency and promulgating norms of responsible state behavior. There are a few potential drawbacks to a broad code of conduct. First, a broader code of conduct, lacking the specificity of some of the measures discussed above, might undercut momentum toward broader cooperation, rather than serve as a building block. Second, there would be risk in negotiating a code of conduct that disagreements over some of the specifics could derail the entire endeavor or lead to forum shopping, whereby countries then spin off to create their own dialogues about a code of conduct. This is arguably what has happened in the cyber realm, where several different ongoing dialogue processes about codes of conduct have not led to substantive success. Third, a more formal code of conduct might start to raise the prospects of triggering some of the costs associated with CBMs. Specifically, if a country reduced its investments in military applications of AI or did not pursue capability areas because it believed adversaries were following a code of conduct, it could expose itself in the event of cheating. This might be of particular concern for democracies, given that, in many cases, democracies are more likely to comply with the agreements they sign, in part because democracies often have rigorous internal bureaucratic processes to ensure compliance. 56 Thus, one might imagine that the incentives might lead to a less formal code of conduct designed as a building block, rather than something that might cause countries to restrain capabilities. \n THE LIMITATIONS OF AI Accident risk is a significant concern for military applications of AI. Competitive pressures could increase accident risk by creating pressures for militaries to shortcut testing and rapidly deploy new AI-enabled systems. States could take a variety of options to mitigate the risks of creating unnecessary incentives to shortcut test and evaluation, 57 including publicly signaling the importance of T&E, increasing transparency about T&E processes, promoting international T&E standards, and sharing civilian research on AI safety. Additionally, AI will enable more capable autonomous systems, and their increased use may pose stability risks, particularly when deployed into contested areas. To mitigate these risks, states could adopt CBMs such as \"rules of the road\" for the behavior of autonomous systems, marking systems to signal their degree of autonomy, and adhering to off-limits geographic areas for autonomous systems. \n Public Signaling To reduce AI accident risk, national security leaders could publicly emphasize the importance of strong T&E requirements for military AI applications. This potentially could be linked to a formal multilateral statement or something more informal. Publicly promoting AI T&E could be valuable in signaling that nations agree, at least in principle, about the importance of T&E to avoid unnecessary accidents and mishaps. Public statements would be more powerful when used in combination with major investments in T&E institutions and processes. Promoting AI T&E as a CBM would be designed to create positive spillover effects. As major countries investing in AI come together to promote AI safety, it demonstrates the importance of the issue. It could encourage other governments to sign on and signal that AI experts within the bureaucracy can advocate for AI T&E measures. The downsides of publicly signaling the prioritization of AI T&E are relatively limited. A critic might argue that, to the extent that accidents are a necessary part of the innovation and capabilities development process, an overemphasis on T&E might discourage experimentation. However, promoting experimentation and innovation does not have to come at the expense of building robust and assured systems, especially since it is through experimentation and testing that accident risks are likely to be revealed, leading to the deployment of more capable systems. Ensuring that AI systems function as intended is part of fielding effective military capabilities, and effective T&E processes are aligned with the goal of fielding superior military capabilities. Rigorous T&E processes would, by definition, add time to the development process in order to ensure that systems are robust and secure before deployment, but the result would be more effective systems once deployed. In peacetime, taking additional measures to ensure that military systems will perform properly in wartime has little downside, so long as accident risk does not become a bureaucratic excuse for inaction. In wartime, the tradeoffs in delaying fielding may become more acute, and militaries may balance these risks differently. There are potential transparency downsides if countries say they emphasize AI T&E in public, but do not do so in private,  58  but that would not impose costs on countries whose actions match their rhetoric. \n Increased Transparency about T&E Processes A related unilateral or multilateral CBM could involve countries publicly releasing details about the T&E processes used for military applications of AI without revealing details about specific technical capabilities. This is similar to existing U.S. policy regarding legal weapons reviews. Currently, the U.S. military promotes norms in favor of stringent legal weapons reviews but does not share the actual reviews of specific weapons.  59  Since this CBM would build on existing norms that the United States already promotes, transparency about T&E processes for military AI systems might be more likely to receive American support than more intrusive measures. Moreover, increasing knowledge about T&E processes might bring other countries that want to learn from the American military on board. The potential drawbacks of transparency surrounding T&E processes stem from what happens if the CBM succeeds. If successful, all countries, including potential adversaries, would have greater knowledge of how to design effective T&E processes for their military AI applications. This could improve their ability to field more effective military AI systems. This downside may be somewhat mitigated if a country only shares high-level information about its T&E bureaucratic processes and refrains from sharing technical information that could actually help an adversary execute more effective T&E. Nevertheless, an overarching concern with any T&E-related CBM that aims to reduce the risk to international stability from states building unsafe AI systems is that actually succeeding in improving other states' T&E could also lead to adversaries deploying more effective AI systems. Whether an adversary's improved AI capabilities or the prospect of an adversary deploying unsafe military AI systems is more of a danger to a country's security would need to be considered. \n International Military AI T&E Standards Another CBM regarding AI safety could entail establishing and promoting specific international standards for what constitutes effective T&E practices for military AI applications. Such an effort could build on private-sector and public-private standard-setting actions for non-military uses of AI.  60  While not enforceable or verifiable, promoting common standards for AI T&E could be a useful focal point for like-minded states to promote responsible norms concerning the safe deployment of military AI systems. The downsides of promoting common T&E standards are similar to the potential downsides of a public emphasis on AI safety. These kinds of CBMs are early building blocks: While the gains are likely to be relatively limited, the downsides are limited as well, because they do not expose key information or require national commitments that limit capabilities. As with increasing transparency about T&E processes, the most significant downside to effective T&E standards would be that, if successful, this CBM could increase the reliability of military AI systems by adversary states. The relative balance of danger between more reliable, and therefore more effective, adversary AI systems versus unreliable and more accident-prone AI systems would again need to be carefully weighed. \n Shared Civilian Research on AI Safety International efforts to promote shared civilian research on AI safety could be a low-level CBM that would not explicitly involve military action. Shared civilian research would build scientific cooperation between nations, which could serve as a building block for overall cooperation. Focusing cooperation on AI safety, an area of shared interest, might also make more nations willing to sign on to participate. An analogy to this in the U.S.-Soviet context is the Apollo-Soyuz mission in 1975, whose intent was to promote cooperation between civilian scientists on a shared agenda. Similarly, nations could work to foster increased cooperation and collaboration between civilian scientists on AI safety. The potential drawbacks of cooperation stem from the general-purpose character of AI knowledge. If increasing cooperation on AI safety led to adversary breakthroughs in AI safety that made them better able to field effective military uses of AI, there could be negative consequences for other states' security. It may be possible to mitigate this downside by carefully scoping the shared civilian research, depending on the specific type of cooperation and degree of information-sharing required by participants. \n International Autonomous Incidents Agreement There are inherent risks when autonomous systems with any level of decision-making interact with adversary forces in contested areas. Given the brittleness of algorithms, the deployment of autonomous systems in a crisis situation could increase the risk of accidents and miscalculation. AI-related CBMs could build on Cold War agreements to reduce the risk of accidental escalation, with some modification to account for the new challenges AI-enabled autonomous systems present. States have long used established \"rules of the road\" to govern the interaction of military forces operating with a high degree of autonomy, such as at naval vessels at sea, and there may be similar value in such a CBM for interactions with AI-enabled autonomous systems. The 1972 Incidents at Sea Agreement and older \"rules of the road\" such as maritime prize law provide useful historical examples for how nations have managed analogous challenges in the past. Building on these historical examples, states could adopt a modern-day \"international autonomous incidents agreement\" that focuses on military applications of autonomous systems, especially in the air and maritime environments. Such an agreement could help reduce risks from accidental escalation by autonomous systems, as well as reduce ambiguity about the extent of human intention behind the behavior of autonomous systems. In addition to the Incidents at Sea Agreement, maritime prize law is another useful historical analogy for how states might craft a rule set for autonomous systems' interactions. Prize law, which first began in the 12th century and evolved more fully among European states in the 15th to 19th centuries, regulated how ships interacted during wartime. Because both warships and privateers, as a practical matter, operated with a high degree of autonomy while at sea, prize law consisted of a set of rules governing acceptable wartime behavior. Rules covered which ships could be attacked, ships' markings for identification, the use of force, seizure of cargo, and providing for the safety of ships' crews.  61  Nations face an analogous challenge with autonomous systems as they become increasingly integrated into military forces. Autonomous systems will be operating on their own for some period of time, potentially interacting with assets from other nations, including competitors, and there could be value in establishing internationally agreed upon \"rules of the road\" for how such systems should interact. The goal of such an agreement, which would not have to be as formal as the Incidents at Sea Agreement, would be to increase predictability and reduce ambiguity about the behavior of autonomous systems. Such an agreement could be legally binding but would not necessarily need to be in order to be useful. It would likely need to be codified in an agreement (or set of agreements), however, so that expectations are clear by all parties. An ideal set of rules would be self-enforcing, such that it is against one's own interests to violate them. Examples of rules of this kind in warfare include prohibitions against perfidy  62  and giving \"no quarter,\" 63 violating either of which incentivizes the enemy to engage in counterproductive behavior, such as refusing to recognize surrender or fighting to the bitter end rather than surrendering. An autonomous incidents agreement could also include provisions for information-sharing about potential deployments of autonomous systems in disputed areas and mechanisms for consultation at the militaryto-military level to resolve questions that arise (including potentially a hotline to respond to incidents in real time). One challenge with autonomous systems is that their autonomous programming is not immediately observable and inspectable from the outside, a major hurdle for verifying compliance with arms control. One benefit to an international rule set that governs the behavior of autonomous systems, particularly in peacetime or pre-conflict settings, is that the outward behavior of the system would be observable, even if its code is not. Other nations could see how another country's autonomous air, ground, or maritime drone behaves and whether it is complying with the rules, depending on how the rules are written. Given the perceived success of the Incidents at Sea Agreement in decreasing the risk of accidental and inadvertent escalation between the United States and the Soviet Union, an equivalent agreement in the AI space might have potential to do the same for a new generation. The efficacy of any agreement would depend on the details, both in the agreement itself and in states' execution. For example, the United States and China have signed multiple CBM agreements involving air and maritime deconfliction of military forces, including the 1998 U.S.-China Military Maritime Consultative Agreement and the 2014 Memorandum of Understanding Regarding the Rules of Behavior for Safety of Air and Maritime Encounters, yet U.S.-China air and naval incidents have continued.  64  However, the existence of these prior agreements themselves may be a positive sign about the potential for U.S.-China cooperation on preventing accidents and could be a building block for further collaboration. Moreover, in a February 2020 article, Senior Colonel Zhou Bo in China's People's Liberation Army (PLA) advocated for CBMs between the United States and China, including on military AI, drawing on the example of the 1972 Incidents at Sea Agreement.  65  Interest in at least some quarters in the Chinese military suggests that cooperation may be possible even in the midst of competition, especially if the PLA is willing to reciprocate American transparency.  66  In the absence of an internationally agreed upon common rule set, a country could unilaterally make declaratory statements about the behavior of its autonomous systems. For example, a country could say, \"If you fire at our autonomous ship/aircraft/vehicle, it will fire back defensively.\"  67  In principle, such a rule could incentivize the desired behavior by other nations (i.e., not shooting at the autonomous ship, unless you want to start a conflict). If every nation adopted this rule, coupled with a \"shoot-second posture\" for autonomous systems-they would not fire unless fired upon first-the result could be a mutually stable situation. A unilateral declaration of a set of rules for avoiding incidents would be analogous to declaring, \"I will drive on the right side of the road. I suggest you do the same or we both will crash.\" This could work if countries' aim is to coordinate their behavior to avoid conflict, meaning they have some shared interests in avoiding accidental escalation. One challenge to establishing rules of the road for autonomous systems' behavior would be if there were incentives to defect from the rules. For example, in World War I, technological developments enabled submarines, which were highly effective in attacking ships but unable to feasibly comply with existing prize law without putting themselves at risk of attack by surfacing. Despite attempts in the early 20th century to regulate submarines, the incentives for defecting from the existing rules were too great (and the rules failed to adapt), and the result was unrestricted submarine warfare.  68  Another challenge to a potential autonomous incidents agreement is fully exploring the incentives for trustworthiness, both in the signals that countries send about the behavior of their autonomous systems and adversaries' responses. Some declaratory policies would not be credible, such as the claim to have created a \"dead hand\" system such that if a country engaged in a particular type of action, an autonomous system would start a war and there would be nothing a leader could do to stop it. \n Marking Autonomous Systems One component of managing risks from interactions with autonomous systems might involve marking those systems to signal to adversaries their level of autonomy. This could be done through physical markings, such as paint, lights, flags, or other observable external characteristics, or through electronic means, such as radiofrequency broadcasts. One benefit of a marking system is that it builds on things militaries already do, even at the tactical level, to signal their intentions. For example, a fighter jet might tip its wing to show an adversary that it is carrying air-to-air missiles under the wing, communicating an unambiguous and credible signal about capability, and at least threatening some degree of intent. Because autonomous programming is not physically observable in the same way, militaries would have to intentionally design systems with observable markings reflecting their degree of autonomy. Another option could be that certain platforms are understood to have certain behavior (or not), the same way that conventional and nuclear capabilities may in some cases be segregated (e.g., some aircraft are nuclearcapable and some are not, which allows nations to send different kinds of signals). Because potential markings for autonomous functionality are not forced by the capability itself but are rather an optional signal that militaries can choose to send, in order for such markings to be believable and useful, there would have to be strong incentives for sending truthful signals and few incentives for deception. This would be challenging, and nations would have to carefully think through what signals might be useful and believable in different circumstances, and how adversaries might interpret such signals. Additionally, because concepts such as \"levels of autonomy\" are often murky, especially for systems that have varying modes of operation, nations would have to think carefully about what kinds of signals could helpfully and clearly communicate autonomous functionality to other countries.  69  In the past, human operators of automated or autonomous systems have in some instances misunderstood the functionality of the system they themselves were operating, leading to accidents.  70  This problem would be significantly compounded for an external observer. Signals that were trusted but misunderstood could be equally or more dangerous than ambiguity, and states should strive for clear, unambiguous signals. \n Off-limits Geographic Areas Nations could agree to declare some geographic areas off-limits to autonomous systems because of their risk of unanticipated interactions. This could be to avoid unintended escalation in a contested region (e.g., a demilitarized zone) or because a region is near civilian objects (e.g., a commercial airliner flight path) and operating there poses a risk to civilians. Other examples of areas that nations could agree to make off-limits to autonomous military systems could be overlapping territorial claims or other countries' exclusive economic zones (EEZs) or airspace above their EEZs. Reaching agreement on specific regions could be challenging, however, since the areas most at risk of escalation are precisely the regions where nations disagree on territorial claims. Nations could perceive any agreement to refrain from deploying elements of military forces to a region as reflecting negatively on their territorial claims or freedom of navigation. Agreeing to declare some areas off-limits to autonomous systems is likely to be most constructive when there are already pre-established regions that countries agree are under dispute (even if they disagree on who has a claim to ownership) and where pre-existing military deconfliction measures already exist. \n SPECIFIC MISSION-RELATED CBMS: NUCLEAR OPERATIONS The integration of AI, autonomy, and/or automation into nuclear command-and-control, early warning, and delivery systems poses unique risks to international stability because of the extreme consequences of nuclear accidents or misuse.  71  One option for mitigating these risks could be for nations to set limits on the integration of AI, autonomy, or automation into their nuclear operations. Some U.S. military leaders and official DoD documents have expressed skepticism about integrating uninhabited vehicles into plans surrounding nuclear weapons. The Air Force's 2013 Remotely Piloted Aircraft (RPA) Vector report proposed that nuclear strike \"may not be technically feasible unless safeguards are developed and even then may not be considered for [unmanned aircraft systems] operations.\" 72 U.S. Air Force general officers have been publicly skeptical about having uninhabited vehicles armed with nuclear weapons. General Robin Rand stated in 2016, during his time as head of Air Force Global Strike Command, that: \"We're planning on [the B-21] being manned. … I like the man in the loop … very much, particularly as we do the dual-capable mission with nuclear weapons.\"  73  Other U.S. military leaders have publicly expressed support for limits on the integration of AI into nuclear command-and-control. In September 2019, Lieutenant General Jack Shanahan, head of the DoD Joint AI Center, said, \"You will find no stronger proponent of the integration of AI capabilities writ large into the Department of Defense, but there is one area where I pause, and it has to do with nuclear command and control.\" In reaction to the concept of the United States adopting a \"dead hand\" system to automate nuclear retaliation if national leadership were wiped out, Shanahan said, \"My immediate answer is 'No. We do not.' … This is the ultimate human decision that needs to be made which is in the area of nuclear command and control.\"  74  While the motivation for these statements about limits on the use of autonomy may or may not be strategic stability-bureaucratic factors could also be at play-they are examples of the kinds of limits that nuclear powers could agree to set, unilaterally or collectively, on the integration of AI, autonomy, and automation into their nuclear operations. Nuclear states have a range of options for how to engage with these kinds of risks. On one end of the spectrum are arms control treaties with some degree of verification or transparency measures to ensure mutual trust in adherence to the agreements. On the other end of the spectrum are unilateral transparency measures, which could have varying degrees of concreteness ranging from informal statements from military or civilian leaders along the lines of Shanahan's and Rand's statements, all the way to formal declaratory policies. In between are options such as mutual transparency measures, statements of principles, or non-legally binding codes of conduct or other agreements between nuclear states to ensure human control over nuclear weapons and nuclear launch decisions. Even if states that desired these restraints found themselves in a position where others were unwilling to adopt more binding commitments, there may be value in unilateral transparency measures both to reduce the fears of other states and to promulgate norms of responsible state behavior. As with other areas, it is important to consider incentives for defection from an agreement and the extent to which one state's voluntary limitations depend on verifying others' compliance with an agreement. If some states, such as the United States, desire strict positive human control over their nuclear weapons and nuclear launch authority for their own reasons, then verifying others' behavior, while desirable, may not be a necessary precondition to those states adopting their own limits on the use of AI, autonomy, or automation in nuclear operations. Two possible CBMs for AI applications in the nuclear arena involve nuclear weapons states agreeing to strict human control over nuclear launch decisions and ensuring any recoverable delivery vehicles are human-inhabited, to ensure positive human control. \n Strict Human Control Over Nuclear Launch Decisions One CBM for uses of AI in the nuclear arena would involve an agreement by nuclear powers to ensure positive human control over all nuclear launch decisions. This type of agreement would preclude automated \"dead hand\" systems or any other automatic trigger for the use of nuclear weapons. The benefit of such a CBM would be to reduce the risk of accidental nuclear war. It would preclude a machine malfunction leading directly to the use of nuclear weapons without a human involved in the process. Agreement on positive human control over nuclear launch decisions could also be a mechanism for dialogue with newer nuclear powers, helping generate more transparency over their nuclear launch decisions. A drawback to this CBM would be forgoing any potential benefits of an automated \"dead hand\" or similar system. While not without controversy, automated nuclear response systems have a strategic logic under some circumstances. Some nuclear states could desire automated retaliatory systems to ensure a second strike in a decapitation scenario. To the extent that strategic stability depends on second strike capabilities, and a country believes it faces a real risk of decapitation if a conflict escalates, that country might prefer an automated option. (This was the intent behind the Soviet Perimeter system, which reportedly had a semiautomated \"dead hand\" functionality.)  75  The assurance of automated retaliation could be valuable as a deterrent and/or to reduce the incentives for a nation's leaders to launch a strike under ambiguous warning, if they felt confident that a second strike was assured. An agreement to rule out the use of automated \"dead hand\" systems might increase the risk of first strike instability, because nations could have a larger incentive to strike first-or perhaps launch in response to a false alarmbefore being decapitated. Alternatively, countries that feel they need an automated nuclear response option might prefer to not sign a CBM or to sign and then cheat. 76 Fortunately, the \"costs\" of a counterpart cheating on this type of CBM are relatively minimal, since presumably most states would only sign such an agreement if they thought it was already consistent with their nuclear launch decision-making process. \n Prohibitions on Uninhabited Nuclear Launch Platforms An agreement to prohibit uninhabited nuclear launch platforms would involve nuclear weapon states agreeing to forgo a capability that, to our knowledge, no nuclear weapon state deploys today-an uninhabited (\"unmanned\") submarine, fighter, or bomber armed with nuclear weapons. 77 Such an agreement would not affect one-way nuclear delivery vehicles, such as missiles or bombs, instead only preventing a state from deploying two-way (recoverable) remotely piloted or autonomous platforms armed with a nuclear weapon. States have long employed uninhabited nuclear delivery vehicles (missiles, bombs, torpedoes) to carry a nuclear warhead to the target. At present, however, the recoverable launch platform (submarine, bomber, transporter erector launcher) is crewed. With crewed nuclear launch platforms, humans remain not only in control over the final decision to launch a nuclear weapon, but have direct physical access to the launch platform to maintain positive control over the nuclear launch decision. A critical benefit of CBMs that sustain positive human control over nuclear weapons is a reduction in the risk of accidental nuclear war. Deploying nuclear weapons on an uninhabited launch platform, whether remotely piloted or autonomous, would by definition increase the risk that, in the case of an accident, whether mechanical or due to flawed software code, a machine, rather than a human, would make the decision about the use of nuclear weapons. Similarly, a crewed platform would have a redundant layer of direct onboard human physical control in the event that the system's software or communications links were hacked. As previously described, U.S. military leaders, often skeptical about capabilities of remotely piloted or autonomous systems, have expressed a degree of support for such a policy, even unilaterally. With American support, this type of CBM might have a better chance of succeeding and gathering support among other nuclear weapon states. Critics might argue that, similar to the objection to a ban on automated nuclear launches, some types of nuclear states might view more autonomous platforms with nuclear weapons as critical to their secondstrike capabilities because of their ability to stay in the air or concealed at sea for extended periods. Russian military officials have raised the idea of an uninhabited nuclear-armed bomber, 78 and Russia is reportedly developing a nuclear-armed uninhabited undersea vehicle, the Status-6. 79 However, given that these platforms are not currently deployed, it may be easier to reach an agreement to prohibit these platforms compared with an agreement prohibiting a capability that already exists. Moreover, to the extent that this kind of CBM is more a commitment to avoid pursuing dangerous applications of AI, rather than a restriction on current capabilities, it would also be reversible if states decided such capabilities were both necessary and safe at a later time. 80 \n Conclusion Military use of AI poses several risks, including due to ways AI could change the character of warfare, the limitations of AI technology today, and the use of AI for specific military missions such as nuclear operations. Policymakers should be cognizant of these risks as nations begin to integrate AI into their military forces, and they should seek to mitigate these risks where possible. Because AI is a generalpurpose technology, it is not reasonable to expect militaries to refrain from adopting AI overall, any more than militaries would refrain from adopting computers or electricity. How militaries adopt AI matters a great deal, however, and various approaches could mitigate risks stemming from military AI competition. Confidence-building measures are one potential tool policymakers could use to help reduce the risks of military AI competition among states. There are a variety of potential confidence-building measures that could be used, all of which have different benefits and drawbacks. As scholars and policymakers move forward to better understand the risks of military AI competition, these and other confidence-building measures should be carefully considered, alongside other approaches such as traditional arms control. 4. Reliable. The department's AI capabilities will have explicit, well-defined uses, and the safety, security, and effectiveness of such capabilities will be subject to testing and assurance within those defined uses across their entire life cycles. \n 5. Governable. The department will design and engineer AI capabilities to fulfill their intended functions while possessing the ability to detect and avoid unintended consequences, and the ability to disengage or deactivate deployed systems that demonstrate unintended behavior. 1. Tim D , A O e ie f Na i al AI S a egie , P li ic + AI Medi m.c m, J e 28, 2018, https://medium.com/politics-ai/anoverview-of-national-ai-strategies-2a70ec6edfd. 4. Exploring or even adopting CBMs does not preclude other approaches to managing the risks of military AI competition, such as arms control for military AI applications. Arms control approaches are beyond the scope of this paper, however.  l: E l i g e e he e f AI i mili a deci i -maki g. 41. The list below is adapted from Marie-France Desjardins, Rethinking Confidence-Building Measures (New York: Routledge, 1997), 5. 42. There is an arms control dilemma whereby the more useful that capabilities are perceived to be for fighting and winning wars, the harder it can become to craft effective regulation. This could create challenges for effective arms control surrounding military applications of AI, as it frequently has for other technologies such as submarines, air-delivered weapons, or nuclear weapons.      36 Center for a New American Security 1152 15 th Street NW, Suite 950, Washington, DC 20005 T: <PHONE> F: <PHONE> CNAS.org @CNASdc \n • Information-sharing and communication • Measures to allow for inspections and observers • \"Rules of the road\" to govern military operations • Limits on military readiness and operations Center for a New American Security 1152 15 th Street NW, Suite 950, Washington, DC 20005 T: <PHONE> F: <PHONE> CNAS.org @CNASdc \n Center for a a New American Security 1152 15 th Street NW, Suite 950, Washington, DC 20005 T: <PHONE> F: <PHONE> CNAS.org @CNASdc \n 2. A d e Imb ie, Jame D ham, Rebecca Gelle , a d Ca he i e Aike , Mai f ame : A P i i al A al i f Rhe ical F ame i AI, CSET Issue Brief (Center for Security and Emerging Technology, August 2020), https://cset.georgetown.edu/research/mainframes-aprovisional-analysis-of-rhetorical-frames-in-ai/; Hea he M. R ff, The f ame blem: The AI a m ace i e, Bulletin of the Atomic Scientists (April 29, 2019), https://thebulletin.org/2019/04/the-frame-problem-the-ai-arms-race-isnt-one/; A nomous Weapons: An Open Le e f m AI & R b ic Re ea che , F e f Life I i e, J l 28, 2015, https://futureoflife.org/open-letter-autonomous-weapons/?cn-reloaded=1; and B a d K a , D D fficial: US a f AI a m ace, C4ISRNET (April 10, 2018), https://www.c4isrnet.com/itnetworks/2018/04/10/dod-official-us-not-part-of-ai-arms-race/. 3. Giac m P. Pa li, Ke i Vig a d, Da id Da k , a d Pa l Me e , M de i i g A m C l: E l i g e e he e f AI in military decision-maki g (U i ed Na i I i e f Di a mame Re ea ch, 2020), https://www.unidir.org/publication/modernizing-armscontrol; A d e Imb ie a d El a B. Ka ia, AI Safe , Security, and Stability Among Great Powers: Options, Challenges, and Lessons Lea ed f P agma ic E gageme , CSET P lic B ief (Ce e f Sec i a d Eme gi g Tech l g , Decembe 2019), https://cset.georgetown.edu/research/ai-safety-security-and-stability-among-great-powers-options-challenges-and-lessons-learned-forpragmatic-engagement/; a d Michael C. H i , La e Kah , a d Ca e Mah e , The F e f Mili a A lica i f A ificial Intelligence: A Role for Confidence-B ildi g Mea e ?, Orbis, 64 no. 4 (Fall 2020), 528-543. \n 40. Pa li, Vig a d, Da k , a d Me e , M de i i g A m C \n 46. Ral h A. C a, Sec i Im lica i f C flic i he S h Chi a Sea: E l i g P e ial T igge f C flic (Pacific Forum CSIS, 1998), http://www.southchinasea.org/files/2012/03/Cossa-Security-Implications-of-Conflict-in-the-S.ChinaSea.pdf. 47. R ell G ldma , I dia-Chi a B de Di e: A C flic E lai ed, The New York Times, June 17, 2020, https://www.nytimes.com/2020/06/17/world/asia/india-china-border-clashes.html; a d Jeff e Ge lema , Sh Fi ed Al g I dia-China B de f Fi Time i Yea , The New York Times, September 8, 2020, https://www.nytimes.com/2020/09/08/world/asia/india-chinaborder.html. 48. Ch i ia R hl, D ca H lli , W a H ffma , a d Tim Ma e , C be ace a d Ge li ic : A e i g Gl bal C be ec i N m P ce e a a C ad (Ca egie E d me f I e a i al Peace, Feb a 26, 2020), https://carnegieendowment.org/2020/02/26/cyberspace-and-geopolitics-assessing-global-cybersecurity-norm-processes-at-crossroads-pub-81110. \n 49. H i , Kah , a d Mah e , The F e f Mili a A lica i f A ificial I ellige ce: A R le f C fide ce-B ildi g Mea e ? 50. C a, Sec i Im lica i f C flic i he S h Chi a Sea: E l i g P e ial T igge f C flic . 51. Defense Innovation Board, AI Principles: Recommendations on the Ethical Use of Artificial Intelligence by the Department of Defense (2019), https://media.defense.gov/2019/Oct/31/<PHONE>/-1/-1/0/DIB_AI_PRINCIPLES_PRIMARY_DOCUMENT.PDF. 52. U.S. Department of Defense, Summary of the 2018 Department of Defense Artificial Intelligence Strategy: Harnessing AI to Advance Our Security and Prosperity, 8. 53. U.S. Department of Defense, Summary of the 2018 Department of Defense Artificial Intelligence Strategy: Harnessing AI to Advance Our Security and Prosperity, 15. 54. Imb ie a d Ka ia, AI Safe , Sec i , a d S abili Am g G ea P e : O i ns, Challenges, and Lessons Learned for Pragmatic E gageme . \n 55. Pa li, Vig a d, Da k , a d Me e , M de i i g A m C l: E l i g e e he e f AI i mili a deci i -maki g. 56. Be h Simm , T ea C m lia ce a d Vi la i , Annual Review of Political Science, 13 (2010), 274-296, 10.1146/a e . li ci.12.0<PHONE>; Ja a S ei , Maki g P mi e , Kee i g P mi e : Dem c ac , Ra ifica i a d C mpliance i I e a i al H ma Righ La , British Journal of Political Science, 46 no. 3 (July 2016), 655-679, https://www.cambridge.org/core/journals/british-journal-of-political-science/article/abs/making-promises-keeping-promises-democracy-67. Nations, of course, can and do often delegate self-defense authority to crewed ship/aircraft/vehicle commanders. What is different in those circumstances is that human control is still retained over use-of-force decisions, albeit delegated to a lower-level commander. Delegating use-of-force decisions to an autonomous system would raise novel stability concerns in crises or militarized disputes due to the brittle nature of machine decisionmaking and the inability of machines (at the current state of technology) to exercise judgment, understand context, or a l c mm e e. F m e he e i k , ee Scha e, A m Wea a d S abili .68. J L. Jac b , The La f S bma i e Wa fa e T da , International Law Studies, 64 (1991), 207-208, https://digitalcommons.usnwc.edu/cgi/viewcontent.cgi?article=1756&context=ils; a d H a d Le ie, S bma i e Wa fa e: Wi h Em ha i he 1936 London Protocol, International Law Studies, 70 (1998), https://digital-commons.usnwc.edu/ils/vol70/iss1/12/. 69. Many thanks to Helen Toner for raising this point.70. William La ge ie che, The H ma Fac , Vanity Fair (October 2014), http://www.vanityfair.com/news/business/2014/10/air-franceflight-447-crash; a d Fi al Re : O he accide f 1 June 2009 to the Airbus A330-203 registered F-GZCP operated by Air France flight 447 Rio de Janeiro Pa i (B ea d E e e d A al e la c i de l a ia i ci ile, [E gli h a la i ], 2012), http://www.bea.aero/docspa/2009/f-cp090601.en/pdf/f-c 090601.e . df. Thi al ma ha e bee a fac i he U.S. A m h -down of a Navy F-18 ai c af i 2003 i h he Pa i ai a d mi ile defe e em; a d Scha e, A m Wea a d S abili , 185. 71. Horowitz, Scharre, and Velez-Gree , A S able N clea F e? The Im ac f A m S em a d A ificial I ellige ce. 72. Headquarters, U.S. Air Force, RPA Vector: Vision and Enabling Concepts, 2013-2038 (February 17, 2014), 54, http://www.globalsecurity.org/military/library/policy/usaf/usaf-rpa-vector_vision-enabling-concepts_2013-2038.pdf. \n 73. H e H dge Seck, Ai F ce Wa Kee Ma i he L i h B-21 Raide , Defe ceTech. g, September 19, 2016, http://www.defensetech.org/2016/09/19/air-force-wants-to-keep-man-in-the-loop-with-b-21-raider/. 74. S d e F eedbe g J ., N AI f N clea C mma d & C l: JAIC Sha aha , B eaki gDefe e.c m, Se embe 25, 2019, https://breakingdefense.com/2019/09/no-ai-for-nuclear-command-control-jaics-shanahan/. 75. The S ie Pe ime e em a e edl a emia ma ed dead ha d em ee clea e alia i if S ie leade hip was wiped out in a decapitation attack. According to reports, the system would retain human control over launch decisions in the hands of a relatively junior Soviet officer who would retain final launch authority. The semiautomated nature of the system would bypass the senior-level a al mall e i ed a h i e a clea la ch. The em f c i ali and the extent to which it was built and deployed \n Te a Na i al Sec i Re ie , J e 2, 2020, https://tnsr.org/roundtable/policy-roundtable-artificial-intelligence-and-internationalsecurity/; Mela ie Si , Je ife S i del, Pa l Scha e, a d Vadim K li , The Mili a i a i f A ificial I ellige ce, S a ley Center for Peace and Security, June 2020, https://stanleycenter.org/publications/militarization-of-artificial-intelligence/; Imb ie a d Ka ia, AI Safe , Security, and Stability Among Great Powers: Options, Challenges, and Lessons Learned for Pragma ic E gageme ; a d Gei a d L h , H Migh A ificial I ellige ce Affec he Ri k f N clea Wa ? 36. Be B cha a , A Na i al Sec i Re ea ch Age da f C be ec i a d A ificial I ellige ce, CSET I e B ief (Ce e for Security and Emerging Technology, May 2020), https://cset.georgetown.edu/research/a-national-security-research-agenda-for-cybersecurity-andartificial-intelligence/. 37. Scharre, Army of None: Autonomous Weapons and the Future of War. 38. Michael C. H i a d La e Kah , The AI Li e ac Ga H bbli g Ame ica Officiald m, Wa he R ck , Ja a 14, 2020, https://warontherocks.com/2020/01/the-ai-literacy-gap-hobbling-american-officialdom/. 39. O b ildi g a d c fide ce i AI, ee H i , Kah , a d Mah e , The F e f Mili a A lica i f A ificial Intelligence: A Role for Confidence-B ildi g Mea e ?, 528-543; a d Imb ie a d Ka ia, AI Safe , Sec i , a d S abili Am g Great Powers: Options, Challe ge , a d Le Lea ed f P agma ic E gageme . A m S em a d A ificial I ellige ce ; Michael C. H i e al., P lic R d able: A ificial I elligence and International Sec i , 5. A m Wea : A O e Le e f m AI & R b ic Re ea che . \n 43. U.S. Department of State, Bureau of International Security and Nonproliferation, Memorandum of Understanding Between the United States of America and the Union of Soviet Socialist Republics Regarding the Establishment of a Direct Communications Link, entered into force June 20, 1963, https://2009-2017.state.gov/t/isn/4785.htm. 44. Sean M. Lynn-J e , A Q ie S cce f A m C l: P e e i g I cide a Sea, International Security, 9 no. 4 (1985), 154-84, doi:10.<PHONE>. 45. Desjardins, Rethinking Confidence-Building Measures. \n\t\t\t T: <PHONE> F: <PHONE> CNAS.org @CNASdc \n\t\t\t . Traceable. The department's AI capabilities will be developed and deployed such that relevant personnel possess an appropriate understanding of the technology, development processes, and operational methods applicable to AI capabilities, including with transparent and auditable methodologies, data sources, and design procedure and documentation.", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/AI-and-International-Stability-Risks-and-Confidence-Building-Measures.tei.xml", "id": "523358557e6c1b8cb03eb28276bce314"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/AGI-Coordination-Geat-Powers-Report.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Building-Trust-Through-Testing.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Hacking-AI-1-.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Future-Indices.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Small-Datas-Big-AI-Potential-1.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Contending-Frames.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Military-AI-Cooperation-Toolbox.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Cooperation-Conflict-and-Transformative-Artificial-Intelligence-A-Research-Agenda.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Reducing_long_term_risks_from_malevolent_actors.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-AI-Definitions-Affect-Policymaking.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-The-Question-of-Comparative-Advantage-in-Artificial-Intelligence-1.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/MovementGrowth.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Mapping-the-AI-Investment-Activities-of-Top-Global-Defense-Companies.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/William-MacAskill_Are-we-living-at-the-hinge-of-history.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Machine-Intelligence-for-Scientific-Discovery-and-Engineering-Invention.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Ethics-and-Artificial-Intelligence-1.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/Global-Catastrophic-Risk-Annual-Report-2016-FINAL.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-AI-Accidents-An-Emerging-Threat.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/CSET-Indonesias-AI-Promise-in-Perspective-1.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}
+{"source": "reports", "source_filetype": "pdf", "authors": "n/a", "title": "n/a", "text": "n/a", "date_published": "n/a", "url": "n/a", "filename": "/Users/janhendrikkirchner/code/2022/10/alignment-research-dataset/align_data/common/../../data/raw/report_teis/2014-1.tei.xml", "id": "274b68192b056e268f128ff63bfcd4a4"}