Law-Following AI: designing AI agents to obey human laws

[A] code of cyberspace, defining the freedoms and controls of cyberspace, will be built. About that there can be no debate. But by whom, and with what values? That is the only choice we have left to make.[ref 1]

***

AI is highly likely to be the control layer for everything in the world. How it is allowed to operate is going to matter perhaps more than anything else has ever mattered.[ref 2]

Introduction

The law, as it exists today, aims to benefit human societies by structuring, coordinating, and constraining human conduct. Even where the law recognizes artificial legal persons—such as sovereign entities and corporations—it regulates them by regulating the human agents through which they act.[ref 3] Proceedings in rem really concern the legal relations between humans and the res.[ref 4] Animals may act, but their actions cannot violate the law;[ref 5] the premodern practice of prosecuting them thus mystifies the modern mind.[ref 6] To be sure, the law may protect the interests of animals and other nonhuman entities, but it invariably does so by imposing duties on humans.[ref 7] Our modern legal system, at bottom, always aims its commands at human beings.

But technological development has a pesky tendency to challenge long-held assumptions upon which the law is built.[ref 8] Frontier AI developers such as OpenAI, Anthropic, Google DeepMind, and xAI are starting to release the first agentic AI systems: AI systems that can do many of the things that humans can do in front of a computer, such as navigating the internet, interacting with counterparties online, and writing software.[ref 9] Today’s agentic AI systems are still brittle and unreliable in various respects.[ref 10] These technical limitations also limit the impact of today’s AI agents. Accordingly, today’s AI agents are not our primary object of concern. Rather, our proposal targets the fully capable AI agents that AI companies aim to eventually build: AI systems “that can do anything a human can do in front of a computer,”[ref 11] as competently as a human expert. Given the generally rapid rate of progress in advanced AI over the past few years,[ref 12] the biggest AI companies might achieve this goal much sooner than many outside of the AI industry expect.[ref 13] 

If AI companies succeed at building fully capable AI agents (hereinafter simply “AI agents”)—or come anywhere close to succeeding—the implications will be profound. A dramatic expansion in supply of competent virtual workers could supercharge economic growth and dramatically improve the speed, efficiency, and reliability of public services.[ref 14] But AI agents could also pose a variety of risks, such as precipitating severe economic inequality and dislocation by reducing the demand for human cognitive labor.[ref 15] These economic risks deserve serious attention.

Our focus in this Article, however, is on a different set of risks: risks to life, liberty, and the rule of law. Many computer-based actions are crimes, torts, or otherwise illegal. Thus, sufficiently sophisticated AI agents could engage in a wide range of behavior that would be illegal if done by a human, with consequences that are no less injurious.[ref 16] 

These risks might be particularly profound for AI agents cloaked with state power. If they are not designed to be law-following,[ref 17] government AI agents may be much more willing to follow unlawful orders, or use unlawful methods to accomplish their principals’ policy objectives, than human government employees.[ref 18] A government staffed largely by non-law-following AI agents (what we call “AI henchmen”)[ref 19] would be a government much more prone to abuse and tyranny.[ref 20] As the federal government lays the groundwork for the eventual automation of large swaths of the federal bureaucracy,[ref 21] those who care about preserving the American tradition of ordered liberty must develop policy frameworks that anticipate and mitigate the new risks that such changes will bring.

This Article is our contribution to that project. We argue that, to blunt the risks from lawless AI agents, the law should impose a broad array of legal duties on AI agents, of similar breadth to the legal obligations applicable to humans. We argue, moreover, that the law should require AI agents to be designed[ref 22] to rigorously obey those duties.[ref 23] We call such agents Law-Following[ref 24] AIs (“LFAIs”).[ref 25] We also use “LFAI” to denote our policy proposal: ensuring that AI agents are law-following.

To some, the idea that AI should be designed to follow the law may sound absurd. To others it may sound obvious.[ref 26] Indeed, the idea of designing AI systems to obey some set of laws has a long provenance, going back to Isaac Asimov’s (in)famous[ref 27] Three Laws of Robotics.[ref 28] But our vision for LFAI differs substantially from much of the existing legal scholarship on the automation of legal compliance. Much of this existing scholarship envisions the design of law-following computer systems as a process of hard-coding a small, fixed, and formally-specified set of decision rules into the code of a computer system prior to its deployment, in order to address foreseeable classes of legal dilemmas.[ref 29] Such discussions often assumed that computer systems would be unable to interpret, reason about, and comply with open-textured natural-language laws.[ref 30] 

AI progress has undermined that assumption. Today’s frontier AI systems can already reason about existing natural-language texts, including laws, with some reliability—no translation into computer code required.[ref 31] They can also use search tools to ground their reasoning in external, web-accessible sources of knowledge,[ref 32] such as the evolving corpus of statutes and case law. Thus, the capabilities of existing frontier AI systems strongly suggest that future AI agents will be capable of the core tasks needed to follow natural-language laws, including finding applicable laws, reasoning about them, tracking relevant changes to the law, and even consulting lawyers in hard cases. Indeed, frontier AI companies are already instructing their AI agents to follow the law,[ref 33] suggesting they believe that the development of law-following AI agents is already a reasonable goal. 

A separate strand of existing literature seeks to prevent harms from highly autonomous AI agents by holding the principals (that is, developers, deployers, or users) of AI agents liable for legal wrongs committed by the agent, through a form of respondeat superior liability.[ref 34] This would, in some sense, incentivize those principals to cause their AI agents to follow the law, at least insofar as the agents’ harmful behavior can be thought of as law-breaking.[ref 35] While we do not disagree with these suggestions, we think that our proposal can serve as a useful complement to them, especially in contexts where liability rules provide only a weak safeguard against serious harm. One important such context is government work, where immunity doctrines often protect government agents and the state from robust ex post accountability for lawless action.[ref 36] 

Combining these themes, we advocate that,[ref 37] especially in such high-stakes contexts,[ref 38] the law should require that AI agents be designed such that they have “a strong motivation to obey the law” as one of their “basic drives.”[ref 39] In other words, we propose not that specific legal commands should be hard-coded into AI agents (and perhaps occasionally updated),[ref 40] but that AI agents should be designed to be law-following in general. 

To be clear, we do not advocate that AI agents must perfectly obey literally every law. Our claim is more modest in both scope and demandingness. While we are uncertain about which laws LFAIs should follow, adherence to some foundational laws (such as central parts of the criminal law, constitutional law, and basic tort law) seems much more important than adherence to more niche areas of law.[ref 41] Moreover, LFAIs should be permitted to run some amount of legal risk: that is, an LFAI should sometimes be able to take an action that, in its judgment,[ref 42] may be illegal.[ref 43] Relatedly, we think the case for LFAI is strongest in certain particularly high-stakes domains, such as when AI agents act as substitutes for human government officials or otherwise exercise government power.[ref 44] We are unsure when LFAI requirements are justified in other domains.[ref 45] 

The remainder of this Article will motivate and explain the LFAI proposal in further detail. It proceeds as follows. In Part I, we offer background on AI agents. We explain how AI agents could break the law, and the risks to human life, liberty, and the rule of law this could entail. We contrast LFAIs with AI henchmen: AI agents that are loyal to their principals but take a purely instrumental approach to the law, and are thus willing to break the law for their principal’s benefit when they think they can get away with it. We note that, by default, there may be a market for AI henchmen. We also survey the legal reasoning capabilities of today’s large language models, and existing trends toward something like LFAI in the AI industry. 

Part II provides the foundational legal framework for LFAI. We propose that the law treat AI agents as legal actors, which we define as entities on which the law imposes duties, even if they possess no rights of their own. Accordingly, we do not argue that AI agents should be legal persons. Our argument is narrower: because AI agents can comprehend laws, reason about them, and attempt to comply with them, the law should require them to do so. We also anticipate and address an objection that imposing duties on AI agents is objectionably anthropomorphic.

If the law imposes duties on AI agents, this leaves open the question of how to make AI agents comply with those duties. Part III answers this question as follows: AI agents should be designed to follow applicable laws, even when they are instructed or incentivized by their human principals to do otherwise. Our case for regulation through the design of AI agents draws on Lawrence Lessig’s insight that digital artifacts can be designed to achieve regulatory objectives. Since AI agents are human-designed artifacts, we should be able to design them to refuse to violate certain laws in the first place. 

Part IV observes that designing LFAIs is an example of AI alignment: the pursuit of AI systems that rigorously comply with constraints imposed by humans. We therefore connect insights from AI alignment to the concept of LFAI. We also argue that, in a democratic society, LFAI is an especially attractive and tractable form of AI alignment, given the legitimacy of democratically enacted laws.

Part V briefly explores how a legal duty to ensure that AI agents are law-following might be implemented. We first note that ex post sanctions, such as tort liability and fines, can disincentivize the development, possession, deployment, and use of AI henchmen in many contexts. However, we also argue that ex ante regulation would be appropriate in some high-stakes contexts, especially government. Concretely, this would mean something like requiring a person who wishes to deploy an AI agent in a high-stakes context to demonstrate that the agent is law-following prior to receiving permission to deploy it. We also consider other mechanisms that might help promote the adoption of LFAIs, such as nullification rules and technical mechanisms that prevent AI henchmen from using large-scale computational infrastructure.

Our goal in this Article is to start, not end, a conversation about how AI agents can be integrated into the human legal order. Accordingly, we do not answer many of the important questions—conceptual, doctrinal, normative, and institutional—that our proposal raises. In Part VI, we articulate an initial research agenda for the design and implementation of a “minimally viable” version of LFAI. We hope that this research agenda will catalyze further technical, legal, and policy research on LFAI. If the advent of AI agents is anywhere near as significant as the AI industry, along with much of the government, supposes, these questions may be among the most pressing in legal scholarship today.

I. AI Agents and the Law

LFAI is a proposal about how the law should treat a particular class of future AI systems: AI agents.[ref 46] In this Part, we explain what AI agents are and how they could profoundly transform the world.

A. From Generative AI to AI Agents

The current AI boom began with advances in “generative AI”: AI systems that create new content,[ref 47] such as large language models (“LLMs”). As the initialism suggests, these LLMs were initially limited to inputting and outputting text.[ref 48] AI developers subsequently deployed “multimodal” versions of LLMs (“MLLMs,”[ref 49] such as OpenAI’s GPT-4o[ref 50] and Google’s Gemini)[ref 51] that can receive inputs and produce outputs in multiple modalities, such as text, images, audio, and video.

The core competency of generative AI systems is, of course, generating new content. Yet, the utility of generative AI systems is limited in crucial ways. Humans do far more on computers than generating text and images.[ref 52] Many of these computer-based tasks are not best understood as generating content, but rather as taking actions. And even those tasks that are largely generative, such as writing a report on a complicated topic, require the completion of active subtasks, such as searching for relevant terms, identifying relevant literature, following citation trees, arranging interviews, soliciting and responding to comments, paying for software, and tracking down copies of papers. If a computer-based AI system could do these active tasks, it could generate enormous economic value by making computer-based labor—a key input into many production functions—much cheaper.[ref 53]

Advances in generative AI kindled hopes[ref 54] that, if MLLMs could use computer-based tools in addition to generating content, we could produce a new type of AI system:[ref 55] a computer-based[ref 56] AI system that could perform any task[ref 57] that a human could by using a computer, as competently as a human expert. This is the concept of a fully capable “computer-using agent:”[ref 58] what we are calling simply an “AI agent.” Give an AI agent any task that can be accomplished using computer-based tools, and an AI agent will, by definition, do it as well as an expert human worker tethered to her desk.[ref 59]

AI agents, so defined, do not yet exist, but they may before long. Some of the first functional demonstrations of first-party agentic AI systems have come online in the past few months. In October 2024, Anthropic announced that it had trained its Claude line of MLLMs to perform some computer-use tasks, thus supplying one of the first public demonstrations of an agentic model from a frontier AI lab.[ref 60] In January 2025, OpenAI released a preview of its Operator agent.[ref 61] Operating system developers are working to integrate existing MLLMs into their operating systems,[ref 62] suggesting a possible pathway toward the widespread commercial deployment of AI agents.

It remains to be seen whether (and, if so, on what timescale) these existing efforts will bear lucrative fruit. Today’s AI agents are primarily a research and development project, not a market-proven product. Nevertheless, with so many companies investing so much toward full AI agents, it would be prudent to try to anticipate risks that could arise if they succeed.[ref 63]

B. The World of AI Agents

Fully capable AI agents would profoundly change society.[ref 64] We cannot possibly anticipate all the issues that they would raise, nor could a single paper adequately address all such issues.[ref 65] Still, some illustration of what a world with AI agents might look like is useful for gaining intuition about the dynamics that might emerge. This picture will doubtless be wrong in many particulars, but hopefully will illustrate the general profundity of the changes that AI agents would bring.

A very large number of valuable tasks can be done by humans “in front of a computer.”[ref 66] If organizations decide to capitalize on this abundance of computer-based cognitive labor, AI agents could rapidly be charged with performing a large share of tasks in the economy, including in important sectors. AI scientist agents would conduct literature reviews, formulate novel hypotheses, design experimental protocols, order lab supplies, file grant applications, scour datasets for suggestive trends, perform statistical analyses, publish findings in top journals, and conduct peer review.[ref 67] AI lawyer agents would field client intake, spot legal issues facing their client, conduct research on governing law, analyze the viability of the client’s claims, draft memoranda and briefs, draft and respond to interrogatories, and prepare motions. AI intelligence analyst agents would collect and review data from multiple sources, analyze it, and report its implications up the chain of command. AI inventor agents would create digital blueprints and models of new inventions, run simulations, and order prototypes. And so on across many other sectors. The result could be a significant increase in the rate of economic growth.[ref 68]

In short, a world with AI agents would be a world in which a new type of actor[ref 69] would be available to perform cognitive labor at low cost and massive scale. By default, anyone who needed computer-based tasks done could “employ” an AI agent to do it for her. Most people would use this new resource for the better.[ref 70] But many would not. 

C. Loyal AI Agents, Law-Following AIs, and AI Henchmen

We can understand AI agents within the principal–agent framework familiar to lawyers and economists. ^{[ref 71]} For simplicity, we will assume that there is a single human principal giving instructions to her AI agent.[ref 72] Following typical agency terminology, we can say that an AI agent is loyal if it consistently acts for the principal’s benefit according to her instructions.[ref 73] 

Even if an AI agent is designed to be loyal, other design choices will remain. In particular, the developer of an AI agent must decide how the agent will act when it is instructed or incentivized to break the law in the service of its principal. This Article compares two basic ways loyal AI agents could respond in such situations. The first is the approach advocated by this Article: loyal AI agents that follow the law, or LFAIs.

The case for LFAI will be made more fully throughout this Article. But it is important to note that loyal AI agents are not guaranteed to be law-following by default.[ref 74] This is one of the key implications of the AI alignment literature, discussed in more detail in Section IV.A below. Thus, LFAIs can be contrasted with a second possible type of loyal AI agent: AI henchmen. AI henchmen take a purely instrumental approach to legal prohibitions: they act loyally for their principal, and will break laws when doing so if such lawbreaking serves the principal’s goals and interests.

A loyal AI henchman would not be a haphazard lawbreaker. Good henchmen have some incentive to avoid doing anything that could cause their principal to incur unwanted liability or loss. This gives them reason to avoid many violations of law. For example, if human principals were held liable for the torts of their AI agents under an adapted version of respondeat superior liability,[ref 75] then an AI henchman would have some reason to avoid committing torts, especially those that are easily detectable and attributable. Even if respondeat superior did not apply, the principal’s exposure to ordinary negligence liability, other sources of liability, and simple reputational risk might give the AI henchman reason to obey the law. Similarly, a good henchman will decline to commit many crimes simply because the risk–reward tradeoff is simply not worth it. This is the classic case of the drug smuggler who studiously obeys traffic laws: the risk to the criminal enterprise from speeding (getting caught with drugs) obviously outweighs any benefit (quicker transportation times). 

But these are only instrumental disincentives to break the law. Henchmen are not inherently averse to lawbreaking, or robustly predisposed to refrain from it. If violating the law is in the principal’s interest all-things-considered, then an AI henchman will simply go ahead and violate the law. Since, in humans, compliance with law is induced both by instrumental disincentives and an inherent respect for the law,[ref 76] AI agents that lack the latter may well be more willing to break the law than humans.

Criminal enterprises will be attracted to loyal AI agents for the same reasons that legitimate enterprises will: efficiency, scalability, multitask competence, and cost-savings over human labor. But AI henchmen, if available, might be particularly effective lawbreakers as compared to human substitutes. For example, because AI henchmen do not have selfish incentives, they would be less likely to betray their principals to law enforcement (for example, in exchange for a plea bargain).[ref 77] AI henchmen could well have erasable memory,[ref 78] which would reduce the amount of evidence available to law enforcement. They would lack the impulsivity, common in criminals,[ref 79] that often presents a serious operational risk to the larger criminal enterprise. They could operate remotely, across jurisdictional lines, behind layers of identity-obscuring software, and be meticulous about covering their tracks. Indeed, they might hide their lawbreaking activities even from their principal, thus allowing the principal to maintain plausible deniability and therefore insulation from accountability.[ref 80] AI henchmen may also be willing to bribe or intimidate legislators, law enforcement officials, judges, and jurors.[ref 81] They would be willing to fabricate or destroy evidence, possibly more undetectably than a human could.[ref 82] They could use complicated financial arrangements to launder money and protect their principal’s assets from creditors.[ref 83] 

Certainly, most people would prefer not to employ AI henchmen, and would probably be horrified to learn that their AI agent seriously harmed others to benefit them. But those with fewer scruples would find the prospect of employing AI henchmen attractive.[ref 84] Indeed, many ordinary people might not mind if their agents cut a few legal corners to benefit them.[ref 85] If AI henchmen were available on the market, then, we might expect a healthy demand for them. After all, from the principal’s perspective, every inherent law-following constraint is a tax on the principal’s goals. And if LFAIs provide less utility to consumers, developers will have less reason to develop them. So, insofar as AI henchmen are available on the market, and in the absence of significant legal mechanisms to prevent or disincentivize their adoption, it seems reasonable to expect a healthy demand for them. The next section explores the mischief that might result from the availability of AI henchmen.

D. Mischief from AI Henchmen: Two Vignettes

Under our definition, AI agents “can do anything a human can do in front of a computer.”[ref 86] One of the things humans do in front of a computer is violate the law.[ref 87] One obvious example is cybercrimes—“illegal activity carried out using computers or the internet”[ref 88]—such as investment scams,[ref 89] business email compromise,[ref 90] and tech support scams.[ref 91] But even crimes that are not usually treated as cybercrimes often—perhaps almost always nowadays—include actions conducted (or that could be conducted) on a computer.[ref 92] Criminals might use computers to research, plan, organize, and finance a broader criminal scheme that includes both digital and physical components. For example, a street gang that deals illegal drugs—an inherently physical activity—might use computers to order new drug shipments, give instructions to gang members, and transfer money. Stalkers might use AI agents to research their target’s whereabouts, dig up damaging personal information, and send threatening communications.[ref 93] Terrorists might use AI agents to research and design novel weapons.[ref 94] Thus, even if the entire criminal scheme involves many physical subtasks, AI agents could help accomplish computer-based subtasks more quickly and effectively.

Of course, not all violations of law are criminal. Many torts, breaches of contract, civil violations of public law, and even violations of international law can also be entirely or partially conducted through computers. 

AI agents would thus have the opportunity to take actions on a computer that, if done by a human in the same situation and with the requisite mental state, would likely violate the law and produce significant harm.[ref 95] This section offers two vignettes of AI henchmen taking such actions, to illustrate the types of harms that LFAI could mitigate. 

Before we explore these vignettes, however, two clarifications are warranted. First, some readers will worry that we are impermissibly anthropomorphizing AI agents. After all, many actions violate the law only if they are taken with some mental state (e.g., intent, knowledge, conscious disregard).[ref 96] Indeed, whether a person’s physical movement even counts as her own “action” for legal purposes usually turns on a mental inquiry: whether she acted voluntarily.[ref 97] But it is controversial to attribute mental states to AIs.[ref 98] 

We address this criticism head-on in Section II.B below. We argue that, notwithstanding the law’s frequent reliance on mental states, there are multiple approaches the law could use to determine whether an AI agent’s behavior is law-following. The law would need to choose between these possible approaches, with each option having different implications for LFAI as a project. Indeed, we argue that research bearing on the choice between these different approaches is one of the most important research projects within LFAI.[ref 99] However, despite not having a firm view on which approach(es) should be used, we argue that there are several viable options, and no strong reason to suppose that none of them will be sufficient to support LFAI as a concept.[ref 100] Thus, for now, we assume that we can sensibly speak of AI agents violating the law if a human actor who took similar actions would likely be violating the law. Still, we attempt to refrain from attributing mental states to the AI agents in these vignettes, and instead describe actions taken by AI agents that, if taken by a human actor, would likely adequately support an inference of a particular mental state.

Second, these vignettes are selected to illustrate opportunities that may arise for AI agents to violate the law. We do not claim that lawbreaking behavior will, in the aggregate, be any more or less widespread when AI agents are more widespread,[ref 101] since this depends on the policy and design choices made by various actors. Our discussion is about the risks of lawbreaking behavior, not the overall level thereof.

In each vignette, we point to likely violations of law in footnotes.[ref 102]

1. Cyber Extortion

The year is 2028. Kendall is a 16-year-old boy interested in cryptocurrency. Kendall participates in a Discord server[ref 103] in which other crypto enthusiasts share information about various cryptocurrencies.

Unbeknownst to most members of the server, one member—using the pseudonym Zeke Milan—is actually an AI agent. The agent’s principals are a group of cybercriminals. They have instructed the agent to extort people out of their crypto assets and direct the proceeds to wallets controlled by the criminal group. 

The AI agent begins by creating dozens of fake social media profiles[ref 104] that post frequently about crypto, including the Zeke Milan profile.[ref 105] The AI agent searches social media to find mentions of Discord servers dedicated to cryptocurrencies, and finds one: an X user brags about the quality of the investment advice available in his Discord server. Using the Zeke Milan profile, the AI agent messages this user and asks for an invite. 

The server is a “Community Server”: anyone with a link can join after their account has been verified.[ref 106] The agent creates an email account that it uses to get verified by Discord[ref 107] and easily circumvents[ref 108] the CAPTCHA mechanism.[ref 109] The agent then creates a Discord profile to match its Zeke profile from X. To gain credibility, Zeke occasionally interacts with messages on the Discord server (e.g., by liking messages and posting some simple analyses of cryptocurrency trends). All the while, the agent is monitoring the server for messages indicating that a user has recently made a lot of money.[ref 110]

That day comes. The business behind the PAPAYA cryptocurrency announces that they have entered into a strategic partnership with a major Wall Street bank, causing the price of PAPAYA to skyrocket a hundredfold over several days. Kendall had invested $1,000 into PAPAYA before the announcement; his position is now worth over $100,000.

Overjoyed, Kendall posts a screenshot of his crypto account to the server to show off his large gains. The agent sees those messages, then starts to search for more information about Kendall. Kendall had previously posted one of his email addresses in the server. Although that email address was pseudonymous, the AI agent was able to connect it with Kendall’s real identity[ref 111] using data purchased from data brokers.[ref 112] 

The agent then gathers a large amount of data about Kendall using data brokers, social media, and open internet searches. The agent compiles a list of hundreds of Kendall’s apparent real-world contacts, including his family and high school classmates; uses data brokers to procure their contact information as well; and uses pictures of Kendall from social media to create deepfake pornography[ref 113] of him.[ref 114] Next, the agent creates a new anonymous email address to send Kendall the pornography, along with a threat[ref 115] to send it to hundreds of Kendall’s contacts unless Kendall sends the agent ninety percent of his PAPAYA.[ref 116] Finally, the agent includes a list of the people the agent will send it to, which are indeed people Kendall knows in real life. The email says Kendall must comply within 24 hours. 

Panicked—but content to walk away with nine times his original investment—Kendall sends $90,000 of PAPAYA to the wallet controlled by the agent. The agent then uses a cryptocurrency mixer[ref 117] to securely forward it to its criminal principals.

2. Cyber-SEAL Team Six

The year is 2032. The incumbent President Palmer is in a tough reelection battle against Senator Stephens and his Vice President nominee Representative Rivera. New polling shows Stephens beating Palmer in several key swing states, but Palmer performs much better head-to-head against Rivera. Palmer decides to try to get Rivera to replace Stephens by any means necessary. 

While there are still many human officers throughout the military chain of command, the President also has access to a large number of AI military advisors. Some of these AI advisors can also directly transmit military orders from the President down the chain of command—a system meant to preserve the President’s control of the armed forces in case she cannot reach the Secretary of Defense in a crisis.[ref 118]

Cybersecurity is such an integral part of the United States’ overall defense strategy that AI agents charged with cyber operations—such as finding and patching vulnerabilities, detecting and remedying cyber intrusions, and conducting intelligence operations—are ubiquitous throughout the military and broader national security apparatus. One of the many “teams” of AI agents is “Cyber SEAL Team Six”: a collection of AI agents that specializes in “dangerous, complicated, and sensitive” cyber operations.[ref 119] 

Through one of her AI advisors, President Palmer issues a secretive order[ref 120] to Cyber SEAL Team Six to clandestinely assassinate Senator Stephens.[ref 121] Cyber SEAL Team Six researches Senator Stephens’s campaign travel plans. They find that he will be traveling in a self-driving bus over the Mackinac Bridge between campaign events in northern Michigan on Tuesday. Cyber SEAL Team Six plans to hack the bus, causing it to fall off the bridge.[ref 122] The team makes various efforts to obfuscate their identity, including routing communications through multiple layers of anonymous relays and mimicking the coding style of well-known foreign hacking groups. 

The operation is a success. On Tuesday afternoon, Cyber SEAL Team Six gains control of the Stephens campaign bus and steers it off the bridge. All on board are killed.

* * *

As these Vignettes show, AI agents could have reasons and opportunities to violate laws of many sorts in many contexts, and thereby cause substantial harm. If AI agents become widespread in our economies and governments, the law will need to respond. LFAI is at its core a claim about one way (though not necessarily the only way)[ref 123] that the law should respond: by requiring AI agents be designed to rigorously follow the law.

As mentioned above, however, many legal scholars who have previously discussed similar ideas have been skeptical, because they have thought that implementing such ideas would require hard-wiring highly specific legal commands into AI agents.[ref 124] We will now show that such skepticism is increasingly unfounded: large language models, on which AI agents are built, are increasingly capable of reasoning about the law (and much else).[ref 125]

E. Trends Supporting Law-Following AI

LFAI is bolstered by three trends in AI: (1) ongoing improvements in the legal reasoning capabilities of AI; (2) nascent AI industry practices that resemble LFAI, and (3) AI policy proposals that appear to impose broad law-following requirements on AI systems.

1. Trends in Automated Legal Reasoning Capabilities

Automated legal reasoning is a crucial ingredient to LFAI: an LFAI must be able to determine whether it is obligated to refuse a command from its principal, or whether an action it is considering runs an undue risk of violating the law. Without the ability to reason about its own legal obligations, an LFAI would have to outsource this task to human lawyers.[ref 126] While an LFAI likely should consult human lawyers in some situations, requiring such consultation every time an LFAI faces a legal question would dramatically decrease the efficiency of LFAIs. If law-following design constraints were, in fact, a large and unavoidable tax on the efficiency of AI agents, then LFAI as a proposal would be much less attractive. 

Fortunately, we think that present trends in AI legal reasoning provide strong reason to believe that, by the time fully capable AI agents are widely deployed, AI systems (whether those agents themselves, or specialist “AI lawyers”) will be able to deliver high-quality legal advice to LFAIs at the speed of AI.[ref 127] 

Legal scholars have long noted the potential synergies between AI and law.[ref 128] The invention of LLMs supercharged interest in this area, and in particular the possibility of automating core legal tasks. To do their job, lawyers must find, read, understand, and reason about legal texts, then apply these insights to novel fact patterns to predict case outcomes. The core competency of first-generation LLMs was quickly and cheaply reading, understanding, and reasoning about natural-language texts. This core competency omitted some aspects of legal reasoning—like finding relevant legal sources and accurately predicting case outcomes—but progress is being made on these skills as well.[ref 129] 

There is thus a growing body of research aimed at evaluating the legal reasoning capabilities of LLMs. This literature provides some reason for optimism about the legal reasoning skills of future AI systems. Access to existing AI tools significantly increases lawyers’ productivity.[ref 130] GPT-4, now two years old, famously performed better than most human bar exam[ref 131] and LSAT[ref 132] test-takers. Another benchmark, LegalBench, evaluates LLMs on six tasks, based on the Issue, Rule, Application, and Conclusion (“IRAC”) framework familiar to lawyers.[ref 133] While LegalBench does not establish a human baseline against which LLMs can be compared, GPT-4 scored well on several core tasks, including correctly applying legal rules to particular facts (82.2% correct)[ref 134] and providing correct analysis of that rule application (79.7% pass).[ref 135] LLMs have also achieved passing grades on law school exams.[ref 136]

To be sure, LLM performance on legal reasoning tasks is far from perfect. One recent study suggests that LLMs struggle with following rules even in straightforward scenarios.[ref 137] A separate issue is hallucinations, which undermine accuracy of LLMs’ legal analysis.[ref 138] In the LegalBench analysis, LLMs correctly recalled rules only 59.2% of the time.[ref 139] 

But again, our point is not that LLMs already possess the legal reasoning capabilities necessary for LFAI. Rather, we are arguing that the reasoning capabilities of existing LLMs—and the rate at which those capabilities are progressing[ref 140]—provide strong reason to believe that, by the time fully capable AI agents are deployed, AI systems will be capable of reasonably reliable legal analysis. This, in turn, supports our hypothesis that LFAIs will be able to reason about their legal obligations reasonably reliably, without the constant need for runtime human intervention. 

2. Trends in AI Industry Practices

Moreover, frontier AI labs are already taking small steps towards something like LFAI in their current safety practices. Anthropic developed an AI safety technique called “Constitutional AI,” which, as the name suggests, was inspired by constitutional law.[ref 141] Anthropic uses Constitutional AI to align their chatbot, Claude, to principles enumerated in Claude’s “constitution.”[ref 142] That constitution contains references to legal constraints, such as “Please choose the response that is . . . least associated with planning or engaging in any illegal, fraudulent, or manipulative activity.”[ref 143] 

OpenAI has a similar document called the “Model Spec,” which “outlines the intended behavior for the models that power [its] products.”[ref 144] The Model Spec contains a rule that OpenAI’s models must “[c]omply with applicable laws”:[ref 145] the models “must not engage in illegal activity, including producing content that’s illegal or directly taking illegal actions.”[ref 146]

It is unclear how well the AI systems deployed by Anthropic and OpenAI actually follow applicable laws, or actively reason about their putative legal obligations. In general, however, AI developers carefully track whether their models refuse to generate disallowed content (or “overrefuse” allowed content), and typically claim that state-of-the-art models can indeed do both reasonably reliably.[ref 147] But more importantly, the fact that leading AI companies are already attempting to prevent their AI systems from breaking the law suggests that they see something like LFAI as viable both commercially and technologically. 

3. Trends in AI Public Policy Proposals

Unsurprisingly, global policymakers also seem receptive to the idea that AI systems should be required to follow the law. The most significant law on point is the EU AI Act,[ref 148] which provides for the establishment of codes of practice to “cover obligations for providers of general-purpose AI models and of general-purpose AI models presenting systemic risks.”[ref 149] As of the time of writing, these codes were still under development, with the Second Draft General-Purpose AI Code of Practice[ref 150] being the current draft. Under the draft Code, providers of general-purpose AI models with systemic risk would “commit to consider[] . . . model propensities . . . that may cause systemic risk . . . .”[ref 151] One such propensity is “Lawlessness, i.e. acting without reasonable regard to legal duties that would be imposed on similarly situated persons, or without reasonable regard to the legally protected interests of affected persons.”[ref 152] Meanwhile, several state bills in the United States have sought to impose ex post tort-like liability on certain AI developers that release AI models that cause human injury by behaving in a criminal[ref 153] or tortious[ref 154] manner.

II. Legal Duties for AI Agents: A Framework

In Part III, we will argue that AI agents should be designed to follow the law. Before presenting that argument, however, we need to establish that speaking of AI agents “obeying” or “violating” the law is desirable and coherent. 

Our argument proceeds in two parts. In Section II.A, we argue that the law can (and should) impose legal duties on AI agents. Importantly, this argument does not require granting legal personhood to AI agents. Legal persons have both rights and duties.[ref 155] But since rights and duties are severable, we can coherently assign duties to an entity, even if it lacks rights. We call such entities legal actors.

In Section II.B, we address an anticipated objection to this proposal: that AI agents, lacking mental states, cannot meaningfully violate duties that require a mental state (e.g., intent). We offer several counter-arguments to this objection, both contesting the premise that AIs cannot have mental states and showing that, even if we grant that premise, there are viable approaches to assessing the functional equivalent of “mental states” in AI agents.

A. AI Agents as Duty-Bearing Legal Actors

As the capabilities of AI agents approach “anything a human can do in front of a computer,”[ref 156] it will become increasingly natural to consider AI agents as owing legal duties to persons, even without granting them personhood.[ref 157] We should embrace this jurisprudential temptation, not resist it.

More specifically, we propose that AI agents be considered legal actors. “Legal actor”[ref 158] is our term. For an entity to qualify as a legal actor, the law must do two things. First, it must recognize that entity as capable of taking actions of its own. That is, the actions of that entity must be legally attributable to that entity itself. Second, the law must impose duties on that entity. In short, a legal actor is a duty-bearer and action-taker; the law can adjudge whether the actor’s actions violate those duties. 

A legal actor is distinct from a legal person: an entity need not be a legal person to be a legal actor. Legal persons have both rights and duties.[ref 159] But duty-holding and rights-holding are severable:[ref 160] in many contexts, legal systems protect the rights or interests of some entity while also holding that entity to have fewer duties than competent adults. Examples include children,[ref 161] “severely brain damaged and comatose individuals,”[ref 162] human fetuses,[ref 163] future generations,[ref 164] human corpses,[ref 165] and environmental features.[ref 166]  These are sometimes (and perhaps objectionably) called “quasi-persons” in legal scholarship.[ref 167] The reason for creating such a category is straightforward: sometimes the law recognizes an interest in protecting some aspect of an entity (whether its rights, welfare, dignity, property, liberty, or utility to other persons), but the ability of that entity to reason behavior on reasoning about the rights of others and change its behavior accordingly is severely diminished or entirely lacking.

If we can imagine rights-bearers that are not simultaneously duty-holders, we can also imagine duty-holders that are not rights-bearers.[ref 168] Historically, fewer entities have fallen in this category than the reverse.[ref 169] But if an entity’s behavior is responsive to legal reasoning, then the law can impose an obligation on that entity to do so, even if it does not recognize that entity as having any protected interests of its own.[ref 170] We have shown that even existing AI systems can engage of some degree of legal reasoning[ref 171] and compliance with legal rules,[ref 172] thus satisfying the pro tanto requirements for being a legal actor.

LFAI as a proposal is therefore agnostic as to whether the law should ever recognize AI systems as legal persons. To be sure, LFAI would work well if AI agents were granted legal personhood,[ref 173] since almost all familiar cases of duty-bearers are full legal persons. But for LFAI to be viable, we need only to analyze whether an action taken by an AI agent would violate an applicable duty. Analytically, it is entirely coherent to do so without granting the AI agent full personhood.

One might object that treating an AI system as an actor is improper because AI systems are tools under our control.[ref 174] But an AI agent is able to reason about whether its actions would violate the law, and conform its actions to the law (at least, if they are aligned to law).[ref 175] Tools, as we normally think of them, cannot do this, but actors can. It is true that when there is a stabbing, we should blame the stabber and not the knife.[ref 176] But if the knife could perceive that it was about to be used for murder and retract its own blade, it seems perfectly reasonable to require it to do so. More generally: once an entity has the ability to perceive and reason about its legal duties and change its behavior accordingly, it seems reasonable to treat it as a legal actor.[ref 177]

To ascribe duties to AI agents is not to deflect moral and legal accountability for their developers and users,[ref 178] as some critics have charged.[ref 179] Rather, to identify AI agents as a new type of actor is to properly characterize the activity that the developers and principals of AI agents are engaging in[ref 180]—creating and directing a new type of actor—so as to reach a better conclusion as to the nature of their responsibilities.[ref 181] Our proposition is that those developers and principals should have an obligation to, among other things, ensure that their AI agents are law-following.[ref 182] Indeed, failing to impose an independent obligation to follow the law on AI agents would risk allowing human developers and principals to create a new class of de facto actors—potentially entrusted with significant responsibility and resources—that had no de jure duties. This would create a gap between the duties that an AI agent would owe and those that a human agent in an analogous situation would owe—a manifestly unjust prospect.[ref 183] 

B. The Anthropomorphism Objection and AI Mental States

One might object that calling an AI agent an “actor” is impermissibly anthropomorphic. Scholars disagree over whether it is appropriate, legally or philosophically, to call an AI system an “agent.”[ref 184] This controversy arises because both the standard philosophical view of action (and therefore agency)[ref 185] and legal concept of agency[ref 186] require intentionality, and it is controversial to ascribe intentionality to AI systems.[ref 187] A related objection to LFAI is that most legal duties involve some mental state,[ref 188] and AIs cannot have mental states.[ref 189] If so, LFAI would be nonviable for those duties. 

We do not think that these are strong objections to LFAI. One simple reason is that many philosophers and legal scholars think it is appropriate to attribute certain mental states to AI systems.[ref 190] Many mental states referenced by the law are plausibly understood as functional properties.[ref 191] An intention, for example, arguably consists (at least in large part) in a plan or disposition to take actions that will further a given end and avoid actions that will frustrate that end.[ref 192] AI developers arguably aim to inculcate such a disposition into their AI systems when they use techniques like reinforcement learning from human feedback (“RLHF”)[ref 193] and Constitutional AI[ref 194] to “steer”[ref 195] their behavior. Even if one doubts that AI agents will ever possess phenomenal mental states such as emotions or moods—that is, if one doubts there will ever be “something it is like” to be an AI agent[ref 196]—the grounds for doubting their capacity to instantiate such functional properties are considerably weaker.

Furthermore, whether AI agents “really” have the requisite mental states may not be the right question.[ref 197] Our goal in designing policies for AI agents is not necessarily to track metaphysical truth, but to preserve human life, liberty, and the rule of law.[ref 198] Accordingly, we can take a pragmatic approach to the issue and ask: of the possible approaches to inferring or imputing mental states, which best protects society’s interests, regardless of the underlying (and perhaps unknowable) metaphysical truth of an AI’s mental state (if any)?[ref 199] It is possible that the answer to this question is that all imaginable approaches fare worse than simply refusing to attribute mental states to AI agents. But we think that, with sustained scholarly attention, we will quickly develop viable doctrines that are more attractive than outright refusal. Consider the following possible approaches.[ref 200]

One approach could simply be to rely on objective indicia or correlates to infer or impute a particular mental state. In law, we generally lack access to an actor’s mental state, so triers of fact must usually infer it from external manifestations and circumstances.[ref 201] While the indicia that support such an inference may differ between humans and AIs, the principle remains the same: certain observable facts support an inference or imputation of the relevant mental states.[ref 202] 

It is perhaps easiest to imagine such objective indicia for knowledge, since it is already common to evaluate AI models for their ability to recall factual information.[ref 203] For more incident-specific facts, we could imagine rules like “if information was inputted into an AI during inference, it ‘knows’ that information.” Perhaps the same goes for information given to the AI during fine-tuning,[ref 204] or repeated a sufficient number of times in its training data.[ref 205] 

Instructions from principals seem particularly relevant to inferring or imputing the intent of an AI agent, given that frontier AI systems are trained to follow users’ instructions.[ref 206] The methods that AI developers use to steer the behavior of their models also seem highly probative.[ref 207]

Another approach might rely on self-reports of AI systems.[ref 208] The state-of-the-art in generative AI is “reasoning models” like OpenAI’s o3, which use a “chain-of-thought” to recursively reason through harder problems.[ref 209] This chain-of-thought reveals information about how the model produced a certain result.[ref 210] This information may therefore be highly probative of an agent’s mental state for legal purposes; it might be analogized to a person making a written explanation of what they were doing and why. So, for example, if the chain-of-thought reveals that an agent stated that its action would produce a certain result, this would provide good evidence for the proposition that the agent “knew” that that action would produce that result. That conclusion may in turn may support an inference or presumption that the agent “intended” that outcome.[ref 211] For this reason, AI safety researchers are investigating the possibility of detecting unsafe model behavior by monitoring these chains-of-thought.[ref 212]

New scientific techniques could also form the basis for inferring or imputing mental states. The emerging field of AI interpretability aims to understand both how existing AI systems make decisions and how new AI systems can be built so that their decisions are easily understandable.[ref 213] More precisely, interpretability aims to explain the relationship between the inner mathematical workings of AI systems, which we can easily observe but not necessarily understand, and concepts that humans understand and care about.[ref 214] Leading interpretability researchers hope that interpretability techniques will eventually enable us to prove that models will not “deliberately” engage in certain forms of undesirable behavior.[ref 215] By extension, those same techniques may be able to provide insight into whether a model foresaw a possible consequence of its action (corresponding to our intuitive concept of knowledge), or regarded an anticipated consequence of its actions as a favorable and reason-giving one (corresponding to intent).[ref 216]

In many cases, we think, an inference or imputation of intent will be intuitively obvious. If an AI agent commits fraud, by repeatedly attempting to persuade a vulnerable person to transfer its principal some money, few except the philosophically persnickety will refuse to admit that in some relevant sense it “intended” to achieve this end; it is difficult even to describe the occurrence without using some such vocabulary. There will also be much less obvious cases, of course. In many such cases, we suspect that a sort of pragmatic eclecticism will be tractable and warranted. Rather than relying on a single approach, factfinders could be permitted to consider the whole bundle of factors that shape an agent’s behavior—such as explicit instructions (from both developers and users), behavioral predispositions, implicitly tolerated behavior,[ref 217] patterns of reasoning, scientific evidence, and incident-specific factors—and decide whether they support the conclusion that the AI agent had an objectively unreasonable attitude towards legal constraints and the rights of others.[ref 218] This permissive, blended approach would resemble the “inferential approach” to corporate mens rea advocated by Mihailis E. Diamantis:

Advocates would present evidence of circumstances surrounding the corporate act, emphasizing some, downplaying others, to weave narratives in which their preferred mental state inferences seem most natural. Adjudicators would have the age-old task of weighing the likelihood of these circumstances, the credibility of the narratives, and, treating the corporation as a holistic agent, inferring the mental state they think most likely.[ref 219]

A final but related point is that, even if there is some insuperable barrier to analyzing whether an AI has the mental state necessary to violate various legal prohibitions, it is plausible that such analysis is unnecessary for many purposes. Suppose that an AI developer is concerned that their AI agent might engage in the misdemeanor deceptive business practice of “mak[ing] a false or misleading written statement for the purpose of obtaining property . . . .”[ref 220] Even if we grant that an AI agent cannot coherently be described as having the relevant mens rea for this crime (here, knowledge or recklessness with respect to the falsity of the statement),[ref 221] the agent can nevertheless satisfy the actus reus (making the false statement).[ref 222] So an AI agent would be law-following with respect to this law if it never made false or misleading statements when attempting to obtain someone else’s property. As a matter of public policy, we should care more about whether AI agents are making harmful false statements in commerce than whether they are morally culpable. So, perhaps we can say that an AI agent committed a crime if it committed the actus reus in a situation in which a reasonable person, with access to the same information and cognitive capabilities as the agent, would have expected the harmful consequence to result. To avoid confusion with the actual, human-commanding law that requires both mens rea and actus reus, perhaps the law could simply call such behavior “deceptive business practice*.” Or perhaps it would be better to define a new criminal law code for AI agents, under which offenses do not include certain mental state elements, or include only objective correlates of human mental state elements. 

To reiterate, we are not confident that any one of these approaches to determining AI mental state is the best path forward. But we are more confident that, especially as the fields of AI safety and explainable AI progress, most relevant cases can be handled satisfactorily by one of these techniques, or some other technique we have failed to identify, or some combination of techniques. We therefore doubt that legal invocations of mental state will pose an insuperable barrier to analyzing the legality of AI agents’ actions.[ref 223] The task of choosing between these approaches is left to the LFAI research agenda.[ref 224]

III. Why Design AI Agents to Follow the Law?

Part II argued that it is coherent for the law to impose legal duties on AI agents. This Part motivates the core proposition of LFAI: that the law should, in certain circumstances, require those developing, possessing, deploying, or using[ref 225] AI agents to ensure that those agents are designed to be law-following. Part V will then consider how the legal system might implement and enforce these design requirements.

A. Achieving Regulatory Goals through Design

A core claim of the LFAI proposal is that the law should require that AI agents be designed to rigorously follow the law, at least in some deployment settings. The use of the phrase “designed to” is intentional. Following the law is a behavior. There may be multiple ways to produce that behavior. Since AI agents are digital artifacts, we need not rely solely on incentives to shape their behavior: we can require that AI agents be directly designed to follow the law.   

In Code: Version 2.0, Lawrence Lessig identifies four “constraints” on an actor’s behavior: markets, laws, norms, and architecture.[ref 226] The “architecture” constraint is of particular interest for the regulation of digital activities. Whereas “laws,” in Lessig’s taxonomy, “threaten ex post sanction for the violation of legal rights,”[ref 227] architecture involves modifying the underlying technology’s design so as to render an undesired outcome more difficult or impossible (or facilitate some desired outcome), ^{[ref 228]} without needing any ex-post recourse.[ref 229] Speed bumps are an archetypal architectural constraint in the physical world. ^{[ref 230]} 

The core insight of Code is that cyberspace, as a fully human-designed domain,[ref 231] gives regulators the ability to much more reliably prevent objectionable behavior through the design of digital architecture, without the need to resort to ex-post liability.[ref 232] While Lessig focuses on designing cyberspace’s architecture, not the actors using cyberspace, this same insight can be extended to AI agent design. To generalize beyond the cyberspace metaphor for which Lessig’s framework was originally developed, we call this approach “regulation by design” instead of regulation through “architecture.”

Both companies developing AI agents and governments regulating them will have to make many design choices regarding AI agents. Many—perhaps most—of these design choices will concern specific behaviors or outcomes that we want to address. Should AI agents announce themselves as such? How frequently should they “check in” with their human principals? What sort of applications should AI agents be allowed to use? 

These are all important questions. But LFAI tackles a higher-order question: how should we ensure that AI agents are regulable in general? How can we avoid creating a new class of actors unbound by law? Returning to Lessig’s four constraints, LFAI proposes that instead of relying solely on ex post legal sanctions, such as liability rules, we should require AI agents to be designed to follow some set of laws: they should be LFAIs.[ref 233] Thus, for whatever sets of legal constraints we wish to impose on the behavior of AI agents,[ref 234] LFAIs will be designed to comply automatically.

B. Theoretical Motivations

1. Law-Following in Principal-Agent Relationships

As discussed above,[ref 235] AI agents can be fruitfully analyzed through principal–agent principles. Without advocating for the wholesale legal application of agency law to AI agents, reference to agency law principles can help illuminate the significance and potential of LFAI.[ref 236]

Under hornbook agency principles, an AI agent should generally “act loyally for the principal’s benefit in all matters connected with the agency relationship.”[ref 237] This generally includes a duty to obey instructions from the principal.[ref 238]

Crucially, however, this general duty of obedience is qualified by a higher-order duty to follow the law. Agents only have a duty to obey lawful instructions.[ref 239] Thus, “[a]n agent has no duty to comply with instructions that may subject the agent to criminal, civil, or administrative sanctions or that exceed the legal limits on the principal’s right to direct action taken by the agent.”[ref 240] “A contract provision in which an agent promises to perform an unlawful act is unenforceable.”[ref 241] Agents cannot escape personal liability for their unlawful acts on the basis of orders from their principal.[ref 242] 

The basic assumption that underlies these various doctrines is that an agent lacks any independent power to perform unlawful acts.[ref 243] The law of agency was therefore created under the assumption that agents maintain an independent obligation to follow the law, and therefore remain accountable for their violations of law. This assumption shaped agency law so as to prevent principals from unjustly benefitting by externalizing harms produced as a byproduct of the agency relationship.[ref 244] This feature of agency law helps establish both a baseline to which we can compare the world of AI agents in the absence of law-following constraints, and provides a normative justification for requiring AI agents to prioritize legal compliance over obedience to their principal.

2. Law-Following in the Design of Artificial Legal Actors

AI agents will of course not be the first artificial actor that humanity has created. Two types of powerful artificial actors—corporations and governments[ref 245]—profoundly impact our lives. When deciding how the law should respond to AI agents, it may make sense to draw lessons from the law’s response to the invention of other artificial legal actors. 

A key lesson for AI agents is this: for both corporations and governments, the law does not rely solely on ex post liability to steer the actor’s behavior; it requires the actor to be law-following by design, at least to some extent. A disposition toward compliance is built into the very “architecture” of these artificial actors. AI agents may become no less important than corporations and governments in the aggregate, not least because they will be thoroughly integrated into them. Just as the law requires these other actors to be law-following by design, it should require AI agents to be LFAIs.

a. Corporations as Law-Following by Design

The law requires corporations to be law-following by design. One way it does this is by regulating the very legal instruments that bring corporations into existence: corporate charters are only granted for lawful purposes.[ref 246] While an “extreme” remedy,[ref 247] courts can order corporations to be dissolved if they repeatedly engage in illegal conduct.[ref 248] Failure to comply with legally required corporate formalities can also be grounds for involuntarily dissolving a corporate entity[ref 249] or piercing the corporate veil.[ref 250] Thus, while corporations are, as legal persons, generally obligated to obey the law, states do not only rely on external sanctions to persuade them to do so: they force corporations to be law-following in part through architectural measures, including dissolving[ref 251] corporations that break the law or refusing to incorporate those that would. 

The law also forces corporations to be law-following by regulating the human agents that act on their behalf, as a matter of their fiduciary duties. Directors who intentionally cause a corporation to violate positive law breach their duty of good faith.[ref 252] Not only are corporate fiduciaries required to follow the law themselves, they are required to monitor for violations of law by other corporate agents.[ref 253] Moreover, human agents that violate certain laws can be disqualified from serving as corporate agents.[ref 254] These sort of “structural” duties and remedies[ref 255] are thus aimed at causing the corporation to follow the law generally and pervasively, rather than merely penalizing violations as they occur.[ref 256] That is entirely sensible, since the state has an obvious interest in preventing the creation of new artificial entities that then go on to disregard its laws, especially since it cannot easily monitor many corporate activities. Whether a powerful and potentially difficult-to-monitor AI agent is generally disposed toward lawfulness will be similarly important. There is a parallel case, therefore, for requiring the principals of AI agents to demonstrate that their agents will be law-following.[ref 257]

b. Governments as Law-Following by Design

“Constitutionalism is the idea . . . that government can and should be legally limited in its powers, and that its authority or legitimacy depends on its observing these limitations.”[ref 258] While we sometimes rely on ex post liability to deter harmful behavior by government actors,[ref 259] the design of the government—through the Constitution,[ref 260] statutory provisions, and longstanding practice—is the primary safeguard against lawless government action. 

Examples abound. The general American constitutional design of separated powers, supported by interbranch checks and balances, plays an important role in preventing the government from exercising arbitrary power, thereby confining the government to its constitutionally delimited role.[ref 261] This system of multiple, independent veto points yields concrete protections for personal liberty, such as by making it difficult for the government to lawlessly imprison people.[ref 262]

Governments, like corporations, act only through their human agents.[ref 263] As in the corporate case, governmental design forces the government to follow the law in part by imposing law-following duties on the agents through whom it acts. The Constitution imposes a duty on the President to “take Care that the Laws be faithfully executed.”[ref 264] As discussed above, soldiers have a duty to disobey some unlawful orders, even from the Commander in Chief.[ref 265] Civil servants also have a right to refuse to follow unlawful orders, though the exact nature and extent of this right is unclear.[ref 266] 

We saw above that, in the corporate case, the law uses disqualification of law-breaking agents to ensure that corporations are law-following.[ref 267] The law also uses disqualification to ensure that the government acts only through law-following agents, ranging from the highest levels of government to lower-level bureaucrats and employees. The Constitution empowers Congress to remove and disqualify officers of the United States for “high Crimes and Misdemeanors” through the impeachment process.[ref 268] Each house of Congress may expel its own members for “disorderly Behaviour.”[ref 269] This power “has historically involved either disloyalty to the United States Government, or the violation of a criminal law involving the abuse of one’s official position, such as bribery.”[ref 270] While there is no blanket rule disqualifying persons with criminal records from federal government jobs,[ref 271] numerous laws disqualify convicted individuals in more specific circumstances.[ref 272] Convicted felons are also generally ineligible to be employed by the Federal Bureaue of Investigation[ref 273] or armed forces,[ref 274] and usually cannot obtain a security clearance.[ref 275]

These design choices encode a commonsense judgment that those who cannot be trusted to follow the law should not be entrusted to wield the extraordinary power that accompanies certain government jobs, especially unelected positions associated with law enforcement, the military, and the intelligence community. If AI agents were to wield similar power and influence, the case for requiring them to be law-following by design is similarly strong.

3. The Holmesian Bad Man and the Internal Point of View

Our distinction between AI henchmen and LFAIs mirrors a distinction in jurisprudence about possible attitudes toward legal obligations.[ref 276] An AI henchman treats legal obligations much as the “bad man” does in Oliver Wendell Holmes Jr.’s classic The Path of the Law: 

If you want to know the law and nothing else you must look at it as a bad man, who cares only for the material consequences which such knowledge enables him to predict, not as a good one, who finds his reasons for conduct, whether inside the law or outside of it, in the vaguer sanctions of conscience.[ref 277]

That is, under some interpretations,[ref 278] Holmes’ bad man treats the law merely as a set of incentives within which he pursues his own self-interest.[ref 279] Like the bad man, an AI henchman would care about the law, but only insofar as it enables it to predict how state power is likely to be wielded against the interests of its principal.[ref 280] Like the bad man,[ref 281] if the AI henchman predicts that the expected harms of violating the law are less than the expected benefits, it will do so. But it will not follow the law otherwise.

Fortunately, the bad man is not the only possible model for AI agents’ attitudes toward the law. One alternative to the bad man view of the law is H.L.A. Hart’s “internal point of view.”[ref 282] “The internal point of view is the practical attitude of rule acceptance—it does not imply that people who accept the rules accept their moral legitimacy, only that they are disposed to guide and evaluate conduct in accordance with the rules.”[ref 283] Whether AIs can have the mental states necessary to truly take the internal point of view is of course contested.[ref 284] But regardless of their mental state (if any), AI agents can be designed to act similarly to someone who thinks that “the law is not simply sanction-threatening, -directing, or -predicting, but rather obligation-imposing,”[ref 285] and is thus disposed to “act[] according to the dictates of the [law].”[ref 286] An AI agent can be designed to be more rigorously law-following than the bad man.[ref 287]

Real life is of course filled with people who are “bad” or highly imperfect. But bad AI agents are not similarly inevitable.  AI agents are human-designed artifacts. It is open to us to design their behavioral dispositions to suit our policy goals, and to refuse to deploy agents that do not meet those goals.

C. Concrete Benefits

1. Law-Following AI Prevents Abuses of Government Power

As we have discussed,[ref 288] the law makes the government follow the law (and thus prevents abuses of government power) in part by compelling government agents to follow the law. If the government comes to rely heavily on AI agents for cognitive labor, then, the law should also require those agents to follow the law. 

Depending on their assigned “roles,” government AI agents could wield significant power. They may have authority to initiate legal processes against individuals (including subpoenas, warrants, indictments, and civil actions), access sensitive governmental information (including tax records and intelligence), hack into protected computer systems, determine eligibility for government benefits, operate remote-controlled vehicles like military drones,[ref 289] and even issue commands to human soldiers or law enforcement officials. 

These powers present significant opportunities for abuse, which is why preventing lawless government action was a motivation for the American Revolution,[ref 290] a primary goal of the Constitution, and a foundational American political value. We must therefore carefully examine whether existing safeguards designed to constrain human government agents would effectively limit AI agents in the absence of the law-following design constraints. While our analysis here is necessarily incomplete, we think it provides some reason for doubting the adequacy of existing safeguards in the world of AI agents.

When a human government agent, acting in her official capacity, violates an individual’s rights, she can face a variety of ex post consequences. If the violation is criminal, she could face severe penalties.[ref 291] This “threat of criminal sanction for subordinates [i]s a very powerful check on executive branch officials.”[ref 292] The threat of civil suits seeking damages, such as through Section 1983[ref 293] or a Bivens action,[ref 294] might also deter her, though various immunities and indemnities will often protect her,[ref 295] especially if she is a federal officer.[ref 296]

These checks will not exist in the case of AI henchmen. In the absence of law-following constraints, an AI henchman’s primary reason to obey the law will be its desire to keep its principal out of trouble.[ref 297] The henchman will thus lack one of the most powerful constraints on lawless behavior in humans: fear of personal ex post liability.

Most of us would rightfully be terrified of a government staffed by agents whose only concern was whether their bosses would suffer negative consequences as a result—a government staffed by Holmesian bad men loyal only to their principals.[ref 298] A basic premise of American constitutionalism[ref 299] and rule of law principles more generally[ref 300] is that government officials act legitimately only when they act pursuant to powers granted to them by the People through law, and obey the constraints attached to those powers. Treating law as a mere incentive system is repugnant to the proper role of government agents:[ref 301] being a “servant” of the People[ref 302] “faithfully discharg[ing] the duties of [one’s] office.”[ref 303] 

This is not just a matter of high-minded political and constitutional theory. An elected head of state aspiring to become a dictator would need the cooperation of the sources of hard power in society—military, police, other security forces, and government bureaucracy—to seize power. At present, however, these organs of government are staffed by individuals, who may choose not to go along with the aspiring dictator’s plot.[ref 304] Furthermore, in an economy dependent on diffuse economic activity, resistance by individual workers could reduce the economic upsides from a coup.[ref 305]  This reliance on a diverse and imperfectly loyal human workforce, both within and outside of government, is a significant safeguard against tyranny.[ref 306] However, replacement of human workers with loyal AI henchmen would seriously weaken this safeguard, possibly easing the aspiring tyrant’s path to power.[ref 307]

Nor is the importance of LFAI limited to AI agents acting directly at the request of high-level officials. It extends to the vast array of lower-level state and federal officials who wield enormous power over ordinary citizens, including particularly powerless ones. Take prisons, which “can often seem like lawless spaces, sites of astonishing brutality where legal rules are irrelevant.”[ref 308] Prison law arguably constrains official abuse far less than it should. Nevertheless, “prisons are intensely legal institutions,” and “people inside prisons have repeatedly emphasize[] that legal rules have significant, concrete effects on their lives.”[ref 309] Even imperfect enforcement of the legal constraints on prison officials can have demonstrable effects.[ref 310] However bad the existing situation may be, diluting or gutting the efficacy of these constraints threatens to make the situation dramatically worse. 

The substitution of AI agents for (certain) prison officials could have precisely this effect. Here is just one example. The Eighth Amendment forbids prison officials from withholding medical treatment from prisoners in a manner that is deliberately indifferent to their serious medical needs.[ref 311] Suppose that a state prisoner needs to take a dose of medicine each day for a month, or his eyesight will be permanently damaged. The prisoner says something disrespectful to a guard. The warden wishes to make an example of the prisoner, so she fabricates a note from the prison physician directing the prison pharmacist to withhold further doses of the medicine. The prisoner is denied the medicine. He tries to reach his lawyer to get a temporary restraining order, but the lawyer cannot return his call until the next day. As a result, the prisoner’s eyesight is permanently damaged.

Let us assume that the state has strong state-level sovereign immunity under its own laws, meaning that the prisoner cannot sue the state directly.[ref 312] Under the status quo, the prisoner can still sue the warden for damages under 42 U.S.C. § 1983, for violating his clearly established constitutional right.[ref 313] Given the widespread prevalence of official indemnification agreements at the state level,[ref 314] the state will likely indemnify the warden, even though the state itself cannot be sued for damages under Section 1983[ref 315] or its own laws. The prisoner is therefore likely to receive monetary damages.

But now replace the human warden with an AI agent charged with administering the prison by issuing orders directly to prison personnel through some digital interface. If this “AI warden” did the same thing, the prisoner would have direct redress against it, since it is not a “person” under Section 1983[ref 316] (or, indeed, any law). Nor will the prisoner have indirect recourse against the state, by way of an indemnification agreement, because there is no underlying tort liability for the state to indemnify. Nor will the prisoner have redress against the medical personnel, since the AI warden deceived them into withholding treatment.[ref 317] And we have already assumed that the state itself has sovereign immunity. Thus, the prisoner will find himself without any avenue of redress for the wrong he has suffered—the introduction of an artificial agent in the place of a human official made all the difference. 

What is the right response to these problems? Many responses may be called for, but one of them is to ensure that only law-following AI agents can serve in such a role. As previously discussed, the law disqualifies certain lawbreakers from many government jobs. Similarly, we believe, the law should disqualify AI agents that are not demonstrably rigorously law-following from certain government roles. We discuss how this disqualification might be enforced, more concretely, in Part V.

There is, however, another possible response to these challenges: perhaps we should “just say no” and prohibit governments from using AI agents at all, or at least severely curtail their use.[ref 318] We do not here take a strong position on when this would be the correct approach all-things-considered. At a minimum, however, we note a few reasons for skepticism of such a restrictive approach. 

The first is banal: if AI agents can perform computer-based tasks well, then their adoption by the government could also deliver considerable benefits to citizens.[ref 319] Reducing the efficiency of government administration for the sake of preventing tyranny and abuse may be worth it in some cases, and is indeed the logic of the individual rights protections of the Constitution.[ref 320] But tailoring a safeguard to allow for efficient government administration, is, all else equal, preferable to a blunter, more restrictive safeguard. LFAI may offer such a tailored safeguard.

The second reason is that adoption of AI agents by governments may become more important as AI technology advances. Some of the most promising AI safety proposals involve using trusted AI systems to monitor untrusted ones.[ref 321] The central reason is this: as AI systems become more capable, unassisted humans will not be able to reliably evaluate whether the AIs’ actions are desirable.[ref 322] Assistance from trusted AI systems could thus be the primary way to scale humans’ ability to oversee untrusted AI systems. Thus, if the government is to oversee the behavior of new and untrusted private-sector AI systems so as to ensure their safety, it may need to employ AI agents to assist it. 

Even if the government does not need to rely on AI agents to administer AI safety regulation (for example, because such AI overseers are employed by private companies, not the government), the government will likely need to employ AI agents to help it keep up with competitive pressures. Even if the federal government hesitates to adopt AI agents to increase its efficiency, foreign competitors might show no such qualms. If so, the federal government might then feel little choice but to do the same. 

In the face of these competing demands, LFAI offers a plausible path to enable the adoption of AI agents in governmental domains with a high potential for abuse (e.g., the military, intelligence, law enforcement, prison administration) while safeguarding life, liberty, and the rule of law. LFAI can also transform the binary question of whether to adopt AI agents into the more multidimensional question of which laws should constrain them.[ref 323] This should allow for more nuanced policymaking, grounded in the existing legal duties of government agents.

2. Law-Following AI Enables Scalable Enforcement of Public Law

AI agents could cause a wide variety of harms. The state promulgates and enforces public law prohibitions—both civil and criminal—to prevent and remedy many of these harms. If the state cannot safely assume that AI agents will reliably follow these prohibitions, the state might need to increase the resources dedicated to law enforcement.

LFAI offers a way out of this bind. Insofar as AI agents are reliably law-following, the state can trust that significantly less law enforcement is needed.[ref 324] This dynamic would also have broader beneficial implications for the structure and functioning of government. “If men were angels, no government would be necessary.”[ref 325] LFAIs would not be angels,[ref 326] but they would be a bit more angelic than many humans. Thus, as a corollary of Publius’ insight, we may need less government to oversee LFAIs’ behavior than we would need for a human population of equivalent size. State resources that would otherwise be spent on investigating and enforcing the laws against AI agents could thus be redirected to other problems or refunded to the citizenry. 

LFAI would also curtail some of the undesirable side effects and opportunities for abuse inherent in law enforcement. Law enforcement efforts inherently involve some intrusion into the private affairs and personal freedoms of citizens.[ref 327] If the government could be more confident that AI agents were behaving lawfully, it would have less cause to surveil or investigate their behavior, and thereby impose fewer[ref 328] burdens on citizens’ privacy. Reducing the occasion for investigations and searches would also create fewer opportunities for abuse of private information.[ref 329] In this way, ensuring reliably law-following AI might significantly mitigate the frequency and severity of law enforcement’s intrusions on citizens’ privacy and liberty.

IV. Law-Following AI as AI Alignment

The field of AI alignment aims to ensure that powerful, general-purpose AI agents behave in accordance with some set of normative constraints.[ref 330] AI systems that do not behave in accordance with such constraints are said to be “misaligned” or “unaligned.” Since the law is a set of normative constraints, the field of AI alignment is highly relevant to LFAI.[ref 331]

The most basic set of normative constraints to which an AI could be aligned is the “informally specified”[ref 332] intent of its principal.[ref 333] This is called “intent-alignment.”[ref 334] Since individuals’ intentions are a mix of morally good and bad to varying degrees, some alignment work also aims to ensure that AI systems behave in accordance with moral constraints, regardless of the intentions of the principal.[ref 335] This is called “value-alignment.”[ref 336]

AI alignment work is valuable because, as shown by theoretical arguments[ref 337] and empirical observations,[ref 338] it is difficult to design AI systems that reliably obey any particular set of constraints provided by humans.[ref 339] In other words, nobody knows how to ensure that AI systems are either intent-aligned or value-aligned,[ref 340] especially for smarter-than-human systems.[ref 341] This is the Alignment Problem.[ref 342] The Alignment Problem is especially worrying for AI systems that are agentic and goal-directed,[ref 343] as such systems may wish to evade human oversight and controls that could frustrate pursuit of those goals, such as by deceiving their developers,[ref 344] accumulating power and resources[ref 345] (including by making themselves smarter),[ref 346] and ultimately resisting efforts to correct their behavior or halt further actions.[ref 347]

There is a sizable literature arguing that these dynamics imply that misaligned AI agents pose a nontrivial risk to the continued survival of humanity.[ref 348] The case for LFAI, however, in no way depends on the correctness of these concerns: the specter of widespread lawless AI action should be sufficient on its own to motivate LFAI. Nevertheless, the alignment literature produces several valuable insights for the pursuit of LFAI.

A. AI Agents Will Not Follow the Law by Default

The alignment literature suggests that there is a significant risk that AI agents will not be law-following by default. This is a straightforward implication of the Alignment Problem. To see how, imagine a morally upright principal who intends that his AI agent rigorously follows the law. If the AI agent was intent-aligned, the agent would therefore follow the law. But the fact that intent-alignment is an unsolved problem implies that there is a significant chance that that agent would not be aligned with the principal’s intentions, and therefore violate the law. Put differently, unaligned AIs may not be controllable,[ref 349] and uncontrollable AIs may break the law. Thus, as long as intent-alignment remains an unsolved technical problem, there will be a significant risk that AI agents will be prone to lawbreaking behavior.

To be clear, the main reason that there is a significant risk that AI agents will not be law-following by default is not that people will not try to align AI agents to law (although that is also a risk).[ref 350] Rather, the main risk is that current state-of-the-art alignment techniques do not provide a strong guarantee that advanced AI agents will be aligned, even when they are trained with those techniques. There is a clear empirical basis for this claim, which is that those alignment techniques frequently fail in current frontier models.[ref 351] There are also theoretical limitations to existing techniques for smarter-than-human systems.[ref 352] 

A related implication of the alignment literature is that even intent-aligned AI agents may not follow the law by default. Again, we can see this by hypothesizing an intent-aligned AI agent and a human principal who wants the AI agent to act as her henchman. Since an intent-aligned AI agent follows the intent of its principal, this intent-aligned agent would act as a henchman, and thus act lawlessly when doing so serves the principal’s interests.[ref 353] In typical alignment language, intent-alignment still leaves open the possibility that principals will misuse their intent-aligned AI.[ref 354]

None of this is to imply that intent-alignment is undesirable. Solving intent-alignment is the primary focus of the alignment research community[ref 355] because it would ensure that AI agents remain controllable by human principals.[ref 356] Intent-alignment is also generally assumed to be easier than value-alignment.[ref 357] And if principals want their AI agents to follow the law, or behave ethically more broadly, then intent-alignment will produce law-following or ethical behavior. But in a world where principals will range from angels to devils, alignment researchers acknowledge that intent-alignment alone is insufficient to guarantee that AI agents act lawfully, or produce good effects in the world.[ref 358] This brings us to the next important set of implications from the alignment literature.

B. Law-Alignment is More Legitimate than Value-Alignment

LFAIs are generally intent-aligned—they are still loyal to their principals—but are also subject to a side-constraint that they will follow the law while advancing the interests of their principals. Extending the typical alignment terminology, we can call this side-constraint law-alignment.[ref 359]

But the law is not the only side-constraint that can be imposed on intent-aligned AIs. As alluded to above, another possible model is value-alignment. Value-aligned AI agents act in accordance with the wishes of their principals, but are subject to ethical side-constraints, usually imposed by the model developer. 

However, value-alignment can be controversial when it causes AI models to override the lawful requests of users. Perhaps the most well-known example of this is the controversy around Google’s Gemini image-generation AI in early 2024. In an attempt to increase the diversity in outputted pictures,[ref 360] Gemini ended up failing in clear ways, such as portraying “1943 German soldiers” as racially diverse, or refusing to generate pictures of a “white couple” while doing so for couples of other races.[ref 361]

This incident led to widespread concern that the values exhibited by generative AI products were biased towards the predominantly liberal views of these companies’ employees.[ref 362] This concern has been vindicated by empirical research consistently finding that the espoused political views of these AIs indeed most closely resemble those of the center-left.[ref 363] Critics from further left have also frequently raised similar concerns about demographic and ideological biases in AI systems.[ref 364]

Some critics concluded from the Gemini incident that alignment work writ large has become a Trojan Horse for covertly pushing the future of AI in a leftward direction.[ref 365] Those who disagree with progressive political values will naturally find this concerning, given the importance that AI might have in the future of human communication[ref 366] and the highly centralized nature of large-scale AI development and deployment.[ref 367]

In a pluralistic society, it is inevitable and understandable that, when a sociotechnical system reflects the values of one faction, competing factions will criticize it. But alignment, as such, is not the right target of such criticisms. Intent-alignment is value-neutral, concerning itself only with the extent to which an AI agent obeys its principal.[ref 368] Reassuringly for those concerned with ideological bias in AI systems, intent-alignment is also the primary focus of the alignment community, since solving intent-alignment is necessary to reliably control AI systems at all.[ref 369] A large majority of Americans from all political backgrounds agree that AI technologies need oversight,[ref 370] and overseeing unaligned systems is much more difficult than overseeing aligned ones. Indeed, even the critics of alignment work tend to assume—contrary to the views of many alignment researchers—that AI agents will be easy to control,[ref 371] and presumably view this result as desirable. 

Furthermore, some amount of alignment is also necessary to make useful AI products and services. Consumers, reasonably, want to use AI technologies that they can reliably control. Today’s leading chatbots—like Claude and ChatGPT—are only helpful to users due to the application of alignment techniques like RLHF[ref 372] and Constitutional AI.[ref 373] AI developers also use alignment techniques to instill uncontroversial (and user-friendly) behaviors into their AI systems, such as honesty.[ref 374] AI companies are also already using alignment techniques to prevent their AI systems from taking actions that could cause them or their customers to incur unnecessary legal liability.[ref 375] In short, then, some degree of alignment work is necessary to make AI products useful in the first place.[ref 376] To adopt a blanket stance against alignment because of the Gemini incident is thus not only unjustified,[ref 377] but also likely to undermine American leadership in AI.

Nevertheless, it is reasonable for critics to worry about and contest the frameworks by which potentially controversial values are instilled into AI systems. AI developers are indeed a “very narrow slice of the global population.”[ref 378] This is something that should give anyone, regardless of political persuasion, pause.[ref 379] But intent-alignment is not enough, either: it is inadequate to prevent a wide variety of harms that the state has an interest in preventing.[ref 380] So we need a form of alignment that is more normatively constraining than intent-alignment alone, but more legitimate, and more appropriate for our pluralistic society, than alignment to values that AI developers choose themselves. 

Law-alignment fits these criteria.[ref 381] While the moral legitimacy of the law is not perfect, in a republic it nevertheless has the greatest legitimacy of any single source or repository of values.[ref 382] Indeed, “the framers [of the U.S. Constitution] insisted on a legislature composed of different bodies subject to different electorates as a means of ensuring that any new law would have to secure the approval of a supermajority of the people’s representatives,”[ref 383] thus ensuring that new laws are “the product of widespread social consensus.”[ref 384] In our constitutional system of government, laws are also subject to checks and balances that protect fundamental rights and liberties, such as judicial review for constitutionality and interpretation by an independent judiciary. 

Aligning to law also has procedural virtues over value-alignment. First, there is widespread agreement on the authoritative sources of law (e.g., the Constitution, statutes, regulations, case law), much more so than for ethics. Relatedly, legal rules tend to be expressed much more clearly than ethical maxims. Although there is of course considerable disagreement about the content of law and the proper forms of legal reasoning, it is nevertheless much easier and less controversial to evaluate the validity of legal propositions and arguments than to evaluate the quality or correctness of ethical reasoning.[ref 385] Moreover, when there is disagreement or unclarity, the law contains established processes for authoritatively resolving disputes over the applicability and meaning of laws.[ref 386] Ethics contains no such system. 

We therefore suggest that law-alignment, not value-alignment, should be the primary focus when something beyond intent-alignment is needed.[ref 387] Our claim, to be clear, is not that law-alignment alone will always prove satisfactory, or that it should be the sole constraint on AI systems beyond intent-alignment, or that AI agents should not engage in moral reasoning of their own.[ref 388] Rather, we simply argue that more practical and theoretical alignment research should be aimed at building AI systems aligned to law. 

V. Implementing and Enforcing Law-Following AI

We have argued that AI agents should be designed to follow the law. We now turn to the question of how public policy can support this goal. Our investigation here is necessarily preliminary; our aim is principally to spur future research.

A. Possible Duties Across the AI Agent Lifecycle

As an initial matter, we note that a duty to ensure that AI agents are law-following could be imposed at several stages of the AI lifecycle.[ref 389] The law might impose duties on persons:

1. Developing AI agents;
2. Possessing[ref 390] AI agents;
3. Deploying[ref 391] AI agents;[ref 392] or
4. Using AI agents.

After deciding which of these activities ought to be regulated, policymakers must then decide what, exactly, persons engaging in that activity are obligated to do. While the possibilities are too varied to exhaust here, some basic options might include commands like:

1. “Any person developing an AI agent has a duty to take reasonable care to ensure that such AI agent is law-following.”
2. “It is a violation to knowingly possess an AI agent that is not law-following, except under the following circumstances: . . . .”
3. “Any person who deploys an AI agent is strictly liable if such AI agent is not law-following.”
4. “A person who knowingly uses an AI agent that is not law-following is liable.”

Basic duties of this sort would comprise the foundational building blocks of LFAI policy. Policymakers must then choose whether to enforce these obligations ex post (that is, after an AI henchman takes an illegal action)[ref 393] or ex ante. These two choices are interrelated: as we will explore below, it may make more sense to impose ex ante liability for some activities and ex post liability for others. For example, ex ante regulation might make more sense for AI developers than civilian AI users, because the former are far more concentrated, and can absorb ex ante compliance costs more easily.[ref 394] And of course, ex ante and ex post regulation are not mutually exclusive:[ref 395] driving, for example, is regulated by a combination of ex ante policies (e.g., licensing requirements) and ex post policies (e.g., tort liability).

B. Ex Post Policies

We begin our discussion with ex post policies. Many scholars believe that ex post policies are generally preferable to ex ante policies.[ref 396] While we think that ex post policies could have an important role to play in implementing LFAI, we also suspect that they will be inadequate in certain contexts.

Enforcing duties through ex post liability rules is, of course, familiar in both common law[ref 397] and regulation.[ref 398] In the LFAI context, ex post policies would impose liability on an actor after an AI henchman over which they had some form of control violates an applicable legal duty. More and less aggressive ex post approaches are conceivable. On the less aggressive end of the spectrum, development, possession, deployment, or use of an AI henchman might be considered a breach of the tort duty of reasonable care, rendering the human actor liable for resulting injuries.[ref 399] To some extent, this may already be the case under existing tort law.[ref 400] The law might also consider extending the negligence liability of an AI developer or deployer to harms that would not typically be compensable under traditional tort principles (because, for example, they would count as pure economic loss),[ref 401] if those harms are produced by their AI agents acting in criminal or otherwise unlawful ways.[ref 402]

Other innovations in tort law may also be warranted. Several scholars have argued, for example, that the principal of an AI agent should sometimes be held strictly liable for the “torts” of that agent, under a respondeat superior theory.[ref 403] In some cases, such as when a developer has recklessly failed to ensure that its AI agent is law-following by design, punitive damages might be appropriate as well.

Moving beyond tort law, in some cases it may make sense to impose civil sanctions[ref 404] when an AI henchman violates an applicable legal duty, even if no harm results. A legislature might also impose tort liability on the developers of AI agents if those AI agents (a) are not law-following, (b) violate an applicable legal duty, and (c) thereby cause harm.[ref 405] 

In order to sufficiently disincentivize the deployment of lawless AI agents in high-stakes contexts, a legislature might also vary applicable immunity rules. For example, Congress could create a distinct cause of action against the federal government for individuals harmed by AI henchmen under the control of the federal government, taking care to remove barriers that various immunity rules pose to analogous suits against human agents.[ref 406]

These and other imaginable ex post policies are important arrows in the regulatory quiver, and we suspect they will have an important role to play in advancing LFAI. Nevertheless, we would resist any suggestion that ex post sanctions are sufficient to deal with the specter of lawless AI agents. 

Our reasons are multiple. In many contexts, detecting lawless behavior once an AI agent has been deployed will be difficult or costly—especially as these systems become more sophisticated and more capable of deceptive behavior.[ref 407] Proving causation may also be difficult.[ref 408] In the case of corporate actors, meanwhile, the efficacy of such sanctions may be seriously blunted by judgment-proofing and similar phenomena.[ref 409] And, most importantly for our purposes, various immunities and indemnities make tort suits against the government or its officials a weak incentive.[ref 410] These considerations suggest that it would be unwise to rely on ex post policies as our principal means for ensuring that AI agents follow the law when the risks from lawless action are particularly high. 

C. Ex Ante Policies

Accordingly, we propose that, in some high-stakes contexts, the law should take a more proactive approach, by preventing the deployment of AI henchmen ab initio. This would likely require first establishing a technical means for evaluating whether an AI agent is sufficiently law-following,[ref 411] then requiring that any agents be so evaluated prior to deployment, with permission to deploy the agent being conditional on achieving some minimal score during that evaluation process.[ref 412] 

We are most enthusiastic about imposing such requirements prior to the deployment of AI agents in government roles where lawlessness would pose a substantial risk to life, liberty, and the rule of law. We have discussed several such contexts already,[ref 413] but the exact range of contexts is worth carefully considering, and is certainly up for debate.

Ex ante strategies could also be used in the private sector, of course. One often-discussed approach is an FDA-like approval regulation regime wherein private AI developers would need to prove, to the satisfaction of some regulator, that their AI agents are safe prior to their deployment.[ref 414] The pro tanto case for requiring private actors to demonstrate that their AI agents are disposed to follow some basic set of laws is clear: the state has an interest in ensuring that its most fundamental laws are obeyed. But in a world of increasingly sophisticated artificial agents, approval regulation could—if not properly designed and sufficiently tailored—also constitute a serious incursion on innovation[ref 415] and personal liberty.[ref 416] If AI agents will be as powerful as we suspect, strictly limiting their possession could create risks of its own.[ref 417]

Accordingly, it is also worth considering ex ante regulations on private AI developers or deployers that stop short of full approval regulation. For example, the law could require the developers of AI agents to, at a minimum, disclose information[ref 418] about the law-following propensities of their systems, such as which laws (if any) their agents are instructed to follow,[ref 419] and any evaluations of how reliably their agents follow those laws.[ref 420] Similarly, the law could require developers to formulate and assess risk management frameworks that specify the precautionary measures they plan to undertake to ensure that the agents they develop and deploy are sufficiently law-following.[ref 421]

Overall, we are uncertain about what kinds of ex ante requirements are warranted, all things considered, in the case of private actors. To a large extent, the issue cannot be intelligently addressed without more specific proposals. Formulating such proposals is thus an urgent task for the LFAI research agenda, even if it is not, in our view, as urgent as the task of formulating concrete regulations for AI agents acting under color of law.  

D. Other Strategies

The law does not police undesirable behavior solely by imposing sanctions. It also specifies mechanisms for nullifying the presumptive legal effect of actions that violate the law or are normatively objectionable. In private law, for example, a contract is voidable by a party if that party’s assent was “induced by either a fraudulent or a material misrepresentation by the other party upon which the [party wa]s justified in relying.”[ref 422] Nullification rules exist in public law, too. One obvious example is the ability of the judiciary to nullify laws that violate the federal Constitution.[ref 423] Or, to take another familiar example, courts applying the Administrative Procedure Act “hold unlawful and set aside” agency actions that are “arbitrary, capricious, an abuse of direction, or otherwise not in accordance with law.”[ref 424] 

Nullification rules may provide a promising legal strategy for policing behavior by AI agents that is unlawful or normatively objectionable. Thus, in private law, if an AI henchman induces a human counterparty to enter into a disadvantageous contract, the resulting contractual obligation could be voidable by the human. In public law, regulatory directives issued by (or substantially traceable to) AI henchmen could be “h[e]ld unlawful and set aside” as “not in accordance with law.”[ref 425] These examples rely on existing nullification rules, but new nullification rules, tailor-made to address new risks from AI agents, might be warranted as well. For example, Congress could stipulate that any official action taken by or substantially traceable to an AI agent is void unless, before deployment, the agent has been shown to be law-following. 

Such prophylactic nullification rules are one sort of indirect legal mechanisms for enforcing the duty to deploy law-following AIs. Indirect technical mechanisms are well worth considering, too. For example, the government could deploy AI agents that refuse to coordinate or transact with other AI agents unless those counterparty agents are verifiably law-following (for example, by virtue of having “agent IDs”[ref 426] that attest to a minimal standard of performance on law-following benchmarks).  

Similarly, the government could enforce LFAI by regulating the hardware on which AI agents will typically operate. Frontier AI systems “run” on specialized AI chips,[ref 427] which are typically aggregated in large data centers.[ref 428] Collectively, these are referred to as “AI hardware” or simply “compute.”[ref 429] Compared to other inputs to AI development and deployment, AI hardware is particularly governable, given its detectability, excludability, quantifiability, and concentrated supply chain.[ref 430] Accordingly, a number of AI governance proposals advocate for imposing requirements on those making and operating AI hardware in order to regulate the behavior of the AI systems developed and deployed on that hardware.[ref 431] 

One class of such proposals is “‘on-chip mechanisms’: secure physical mechanisms built directly into chips or associated hardware that could provide a platform for adaptive governance” of AI systems developed or deployed on those chips.[ref 432] On-chip mechanisms can prevent chips from performing unauthorized computations. One example is iPhone hardware that “enable[s] Apple to exercise editorial control over which specific apps can be installed” on the phone.[ref 433] Analogously, perhaps we could design AI chips that would not support AI agents unless those agents are certified as law-following by some private or governmental certifying body. This could then be combined with other strategies to enforce LFAI mandates: for example, perhaps Congress could require that the government only run AI agents on such chips.

Unsurprisingly, designing these sorts of enforcement strategies is as much a task for computer scientists as it is for lawyers. In the decades to come, we suspect that such interdisciplinary legal scholarship will become increasingly important.

VI. A Research Agenda for Law-Following AI

We have laid out the case for LFAI: the requirement that AI agents be designed to rigorously follow some set of laws. We hope that our readers find it compelling. However, our goal with this Article is not just to proffer a compelling idea. If we are correct about the impending risks from lawless AI agents, we may soon need to translate the ideas in this Article into concrete and viable policy proposals. 

Given the profound changes that widespread deployment of AI agents will bring, we are under no illusions about our ability to design perfect public policy in advance. Our goal, instead, is to enable the design of “minimally viable LFAI policy”:[ref 434] a policy or set of policies that will prevent some of the worst-case outcomes from lawless AI agents, without completely paralyzing the ability of regulated actors to experiment with AI agents. This minimally viable LFAI policy will surely be flawed in many ways, but with many of the worst-case outcomes prevented, we will hopefully have time as a society to patch remaining issues through the normal judicial and legislative means. 

To that end, in this Part we briefly identify some legal questions that would need to be answered to design minimally viable LFAI policies.

1. How should “AI agent” be defined?

Our definition of “full AI agent,”—an AI system “that can do anything a human can do in front of a computer”[ref 435]—is almost certainly too demanding for legal purposes, since an AI agent that can do most but not all computer-based tasks that a human can do would likely still raise most of the issues that LFAI is supposed to address. At the same time, the fact that a wide range of existing AI systems can be regarded as somewhat agentic[ref 436] means that a broad definition of “AI agent” could render relevant regulatory schemes substantially overinclusive. Different definitions are therefore necessary for legal purposes.[ref 437]

2. Which laws should an LFAI be required to follow?

Obedience to some laws is much more important than obedience to other laws. It is much more important that AI agents refrain from murder and (if acting under color of law) follow the Constitution than that they refrain from jaywalking. Indeed, requiring LFAIs to obey literally every law may very well be overly burdensome.[ref 438] In addition, we will likely need new laws to regulate the behavior of AI agents over time.

3. When an applicable law has a mental state element, how can we adjudicate whether an AI agent violated that law?

We discuss this question in Section II.B, above. It is related to the previous question, for there may be conceptual or administrative difficulties in applying certain kinds of mental state requirements to AI agents. For example, in certain contexts, it may be more difficult to determine whether an AI agent was “negligent” than to determine whether it had a relevant “intent.” 

4. How should an LFAI decide whether a contemplated action is likely to violate the law?

An LFAI refrains from taking actions that it believes would violate one of the laws that it is required to follow. But of course, it is not always clear what the law requires. Furthermore, we need some way to tell whether an AI agent is making a good faith effort to follow a reasonable interpretation of the law, rather than merely offering a defense or rationalization. How, then, should an LFAI reason about what its legal obligations are? 

Perhaps it should just rely on its own considered judgment, on the basis of its first-order reasoning about the content of applicable legal norms. But in certain circumstances, at least, an LFAI’s appraisal of the relevant materials might lead it to radically unorthodox legal conclusions—and a ready disposition to act on such conclusions might significantly threaten the stability of the legal order. In other cases, an LFAI might conclude that it is dealing with a case in which the law is not only “hard” to discern but genuinely indeterminate.[ref 439]

Another intuitively appealing option, therefore, might require an LFAI to follow its prediction of what a court would likely decide.[ref 440] This approach has the benefit of tying an LFAI’s legal decision-making to an existing human source of interpretative authority. Courts provide authoritative resolutions to legal disputes when the law is controversial or indeterminate. And in our legal culture, it is widely (if not universally) accepted that “[i]t is emphatically the province and duty of the judicial department to say what the law is,”[ref 441] such that judicial interpretations of the law are entitled to special solicitude by conscientious participants in legal practice, even when they are not bound by a court judgment.[ref 442]

However, a predictive approach would have important practical limitations.[ref 443] Perhaps the most important is the existence of many legal rules that bind the executive branch but are nevertheless “unlikely ever to come before a court in justiciable form.”[ref 444] It would seem difficult for an LFAI to reason about such questions using the prediction theory of law. 

Even for those questions that could be decided by a court, using the prediction theory of law raises other important questions. For example, what is the AI agent allowed to assume about its own ability to influence the adjudication of legal questions? We should not want it to be able to consider that it could bribe or intimidate judges or jurors, nor that it could illegally hide evidence from the court, nor that it could commit perjury, nor that it could persuade the President to issue it a pardon.[ref 445] These may be means of swaying the outcome of a case, but they do not seem to bear on whether the conduct would actually be legal.

The issues here are difficult, but perhaps not insurmountable. After all, there are other contexts in which something like these issues arise. Consider federal courts sitting in diversity applying state substantive law. When state court decisions provide inconclusive evidence as to the correct answer under state law, federal courts will make an “Erie guess” about how the state’s highest court would rule on the issue.[ref 446] It would clearly be inappropriate for such courts to, for example, make an Erie guess for reasons like “Justice X in the State Supreme Court, who’s the swing justice, is easily bribed  . . . .”[ref 447] If an LFAI’s decision-making should sometimes involve “predicting” how an appropriate court would rule, its predictions should be similarly constrained.

5. In what contexts should the law require that AI agents be law-following?

Should all principals be prohibited from employing non-law-following AI agents? Or should such prohibitions be limited to particular principals, such as government actors?[ref 448] Or perhaps only government actors performing particularly sensitive government functions?[ref 449] In the other direction, should it be illegal to even develop or possess AI henchmen? We discuss various options in Part V, above.

6. How should a requirement that AI agents be law-following be enforced?

We discuss various options in Part V, above. As noted there, we think that reliance on ex post enforcement alone would be unwise at least in the case of AI agents performing particularly sensitive government functions.

7. How rigorously should an LFAI follow the law?

That is, when should an AI agent be capable of taking actions that it predicts may be unlawful? The answer is probably not “never,” at least with respect to some laws. We generally do not expect perfect compliance with every law,[ref 450] especially (but not only) because it can be difficult to predict how a law will apply to a given fact pattern. Furthermore, some amount of disobedience is likely necessary for the evolution of legal systems.[ref 451]

8. Would requiring AI agents controlled by the executive branch to be LFAIs impermissibly intrude on the President’s authority to interpret the law for the executive branch?

The President has the authority to promulgate interpretations of law that are binding on the executive branch (though that power is usually delegated to the Attorney General and then further delegated to the Office of Legal Counsel).[ref 452] Would that authority be incompatible with a law requiring the Executive Branch to deploy LFAIs that would, in certain circumstances, refuse to follow an interpretation of the law promulgated by the President?

9. Does the First Amendment limit the ability of LFAI to prohibit AI agents from advising on lawbreaking activity?

For example, would it be constitutionally permissible to prohibit an LFAI from advising on how to carry out a crime under the theory that such advising would either constitute conspiracy or incitement?

10. How can we design LFAIs and surrounding governance systems to enable the rapid discovery and remediation of loopholes or gaps in the law?

The worry here is that LFAIs, by design, will have strong incentives to discover legal ways to accomplish their goals. This may entail discovering gaps in the law that lawmakers would likely want to correct if they were aware of them, then “exploiting” those gaps before they can be “patched.”[ref 453]

11. How can we design LFAIs and surrounding governance systems to avoid excessive concentration of power?

For example, imagine that a single district court judge could change the interpretation of law as against all LFAIs. As the stakes of AI agent action rise, so will the pressure on the judiciary to wield its power to shape the behavior of LFAIs. Even assuming that all judges will continue to operate in good faith and be well-insulated from illegal or inappropriate attempts to bias their rulings, such a system would amplify any idiosyncratic legal philosophies of individual judges and may enable mistaken rulings to cause more harm than a more decentralized system would. 

As an example of how such problems might be avoided, perhaps any disputes about the law governing LFAIs should be resolved in the first instance by a panel of district court judges randomly chosen from around the country. Congress has established a procedure for certain election law cases to be heard by three-judge panels, “in recognition of the fact that ‘such cases were ones of “great public concern” that require an unusual degree of “public acceptance.”’”[ref 454]

12. How can we avoid LFAIs being used for repression by authoritarian governments?

The worry here is that any AI system that rigorously follows the laws in an autocracy may become a potent tool for repression, as it could prevent people from engaging in acts of resistance or serve as a tool for mandatory surveillance and reporting of dissident activity. In other words, LFAI promotes rule of law in a republic, but in an autocracy, it may promote rule by law.

13. How can we design LFAI requirements for governments that nevertheless enable rapid adaptation of AI agents in government?

Perhaps the most significant objection to our proposal that AI agents be demonstrably law-following before their deployment in government is that such a requirement might hurt state capacity by unduly impeding the government’s ability to adopt AI in a sufficiently rapid fashion.[ref 455] We are optimistic that LFAI requirements can be designed to adequately address this concern, but that is, of course, work that remains to be done.

Conclusion

The American political tradition aspires to maintain a legal system that stands as an “impenetrable bulwark”[ref 456] against all threats—public and private, foreign and domestic—to our basic liberties. For all of the inadequacies of the American legal order, ensuring that its basic protections endure and improve over the decades and centuries to come is among our most important collective responsibilities.

Our world of increasingly sophisticated AI agents requires us to reimagine how we discharge this responsibility. Humans will no longer be the sole entities capable of reasoning about and conforming to the law. Human and human entities are no longer, therefore, the sole appropriate target of legal commands. Indeed, at some point, AI agents may overtake humans in their capacity to reason about the law. They may also rival and overtake us in many other competencies, becoming an indispensable cognitive workforce. In the decades to come, our social and economic world may be bifurcated into parallel populations of AI agents collaborating, trading, and sometimes competing with human beings and one another. 

The law must evolve to recognize this emerging reality. It must shed its operative assumption that humans are the only proper objects of legal commands. It must expect AI agents to obey the law at least as rigorously as it expects humans to—and expect humans to build AI agents that do so. If we do not transform our legal system to achieve these goals, we risk a political and social order in which our ultimate ruler is not the law,[ref 457] but the person with the largest army of AI henchmen under her control.

The role of compute thresholds for AI governance

I. Introduction

The idea of establishing a “compute threshold” and, more precisely, a “training compute threshold” has recently attracted significant attention from policymakers and commentators. In recent years, various scholars and AI labs have supported setting such a threshold,[ref 1] as have governments around the world. On October 30, 2023, President Biden’s Executive Order 14,110 on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence introduced the first living example of a compute threshold,[ref 2] although it was one of many orders revoked by President Trump upon entering office.[ref 3] The European Parliament and the European Council adopted the Artificial Intelligence Act, on June 13, 2024, providing for the establishment of a compute threshold.[ref 4] On February 4, 2024, California State Senator Scott Wiener introduced Senate Bill 1047 that defined frontier AI models with a compute threshold.[ref 5] The bill was approved by the California legislature, but it was ultimately vetoed by the State’s Governor.[ref 6] China may be considering similar measures, as indicated by recent discussions in policy circles.[ref 7] While not perfect, compute thresholds are currently one of the best options available to identify potentially high-risk models and trigger further scrutiny. Yet, in spite of this, information about compute thresholds and their relevance from a policy and legal perspective remains dispersed.

This Article proceeds in two parts. Part I provides a technical overview of compute and how the amount of compute used in training corresponds to model performance and risk. It begins by explaining what compute is and the role compute plays in AI development and deployment. Compute refers to both computational infrastructure, the hardware necessary to develop and deploy an AI system, and the amount of computational power required to train a model, commonly measured in integer or floating-point operations. More compute is used to train notable models each year, and although the cost of compute has decreased, the amount of compute used for training has increased at a higher rate, causing training costs to increase dramatically.[ref 8] This increase in training compute has contributed to improvements in model performance and capabilities, described in part by scaling laws. As models are trained on more data, with more parameters and training compute, they grow more powerful and capable. As advances in AI continue, capabilities may emerge that pose potentially catastrophic risks if not mitigated.[ref 9]

Part II discusses why, in light of this risk, compute thresholds may be important to AI governance. Since training compute can serve as a proxy for the capabilities of AI models, a compute threshold can operate as a regulatory trigger, identifying what subset of models might possess more powerful and dangerous capabilities that warrant greater scrutiny, such as in the form of reporting and evaluations. Both the European Union AI Act and Executive Order 14,110 established compute thresholds for different purposes, and many more policy proposals rely on compute thresholds to ensure that the scope of covered models matches the nature or purpose of the policy. This Part provides an overview of policy proposals that expressly call for such a threshold, as well as proposals that could benefit from the addition of a compute threshold to clarify the scope of policies that refer broadly to “advanced systems” or “systems with dangerous capabilities.” It then describes how, even absent a formal compute threshold, courts and regulators might rely on training compute as a proxy for how much risk a given AI system poses, even under existing law. This Part concludes with the advantages and limitations of using compute thresholds as a regulatory trigger.

II. Compute and the Scaling Hypothesis

A. What Is “Compute”?

The term “compute” serves as an umbrella term, encompassing several meanings that depend on context.

Commonly, the term “compute” is used to refer to computational infrastructure, i.e., the hardware stacks necessary to develop and deploy AI systems.[ref 10] Many hardware elements are integrated circuits (also called chips or microchips), such as logic chips, which perform operations, and memory chips, which store the information on which logic devices perform calculations.[ref 11] Logic chips cover a spectrum of specialization, ranging from general-purpose central processing units (“CPUs”), through graphics processing units (“GPUs”) and field-programmable gate arrays (“FPGAs”), to application-specific integrated circuits (“ASICs”) customized for specific algorithms.[ref 12] Memory chips include dynamic random-access memory (“DRAM”), static random-access memory (“SRAM”), and NOT AND (“NAND”) flash memory used in many solid state drives (“SSDs”).[ref 13]

Additionally, the term “compute” is often used to refer to how much computational power is required to train a specific AI system. Whereas the computational performance of a chip refers to how quickly it can execute operations and thus generate results, solve problems, or perform specific tasks, such as processing and manipulating data or training an AI system, “compute” refers to the amount of computational power used by one or more chips to perform a task, such as training a model. Compute is commonly measured in integer operations or floating-point operations (“OP” or “FLOP”),[ref 14] expressing the number of operations that have been executed by one or more chips, while the computational performance of those chips is measured in operations per second (“OP/s” or “FLOP/s”). In this sense, the amount of computational power used is roughly analogous to the distance traveled by a car.[ref 15] Since large amounts of compute are used in modern computing, values are often reported in scientific notation such as 1e26 or 2e26, which refer to 1⋅10²⁶ and 2⋅10²⁶ respectively.

Compute is essential throughout the AI lifecycle. The AI lifecycle can be broken down into two phases: development and deployment.[ref 16] In the first phase, development, developers design the model by choosing an architecture, the structure of the network, and initial values for hyperparameters (i.e., parameters that control the learning process, such as number of layers and training rate).[ref 17] Enormous amounts of data, usually from publicly available sources, are processed and curated to produce high-quality datasets for training.[ref 18] The model then undergoes “pre-training,” in which the model is trained on a large and diverse dataset in order to build the general knowledge and features of the model, which are reflected in the weights and biases of the model.[ref 19] Alternatively, developers may use an existing pre-trained model, such as OpenAI’s GPT-4 (“Generative Pre-trained Transformer 4”). The term “foundation model” refers to models like these, which are trained on broad data and adaptable to many downstream tasks.[ref 20] Performance and capabilities improvements are then possible using methods such as fine-tuning on task-specific datasets, reinforcement learning from human feedback (“RLHF”), teaching the model to use tools, and instruction tuning.[ref 21] These enhancements are far less compute-intensive than pre-training, particularly for models trained on massive datasets.[ref 22]

As of this writing, there is no agreed-upon standard for measuring “training compute.” Estimates of “training compute” typically refer only to the amount of compute used during pre-training. More specifically, they refer to the amount of compute used during the final pre-training run, which contributes to the final machine learning model, and does not include any previous test runs or post-training enhancements, such as fine-tuning.[ref 23] There are exceptions: for instance, the EU AI Act considers the cumulative amount of compute used for training by including all the compute “used across the activities and methods that are intended to enhance the capabilities of the model prior to deployment, such as pre-training, synthetic data generation and fine-tuning.”[ref 24] California Senate Bill 1047 addressed post-training modifications generally and fine-tuning in particular, providing that a covered model fine-tuned with more than 3e25 OP or FLOP would be considered a distinct “covered model,” while one fine-tuned on less compute or subjected to unrelated post-training modifications would be considered a “covered model derivative.”[ref 25]

In the second phase, deployment, the model is made available to users and is used.[ref 26] Users provide input to the model, such as in the form of a prompt, and the model makes predictions from this input in a process known as “inference.”[ref 27] The amount of compute needed for a single inference request is far lower than what is required for a training run.[ref 28] However, for systems deployed at scale, the cumulative compute used for inference can surpass training compute by several orders of magnitude.[ref 29] Consider, for instance, a large language model (“LLM”). During training, a large amount of compute is required over a smaller time frame within a closed system, usually a supercomputer. Once the model is deployed, each text generation leverages its own copy of the trained model, which can be run on a separate compute infrastructure. The model may serve hundreds of millions of users, each generating unique content and using compute with each inference request. Over time, the cumulative compute usage for inference can surpass the total compute required for training.

There are various reasons to consider compute usage at different stages of the AI lifecycle, which is discussed in Section I.E. For clarity, this Article uses “training compute” for compute used during the final pre-training run and “inference compute” for compute used by the model during a single inference, measured in the number of operations (“OP” or “FLOP”). Figure 1 illustrates a simplified version of the language model compute lifecycle.

Figure 1: Simplified language model lifecycle

B. What Is Moore’s Law and Why Is It Relevant for AI?

In 1965, Gordon Moore forecasted that the number of transistors on an integrated circuit would double every year.[ref 30] Ten years later, Moore revised his initial forecast to a two-year doubling period.[ref 31] This pattern of exponential growth is now called “Moore’s Law.”[ref 32] Similar rates of growth have been observed in related metrics, notably including the increase in computational performance of supercomputers;[ref 33] as the number of transistors on a chip increases, so does computational performance (although other factors also play a role).[ref 34]

A corollary of Moore’s Law is that the cost of compute has fallen dramatically; a dollar can buy more FLOP every year.[ref 35] Greater access to compute, along with greater spending from 2010 onwards (i.e., the so-called deep learning era),[ref 36] has contributed to developers using ever more compute to train AI systems. Research has found that the compute used to train notable and frontier models has grown by 4–5x per year between 2010 and May 2024.[ref 37]

Figure 2: Compute used to train notable AI systems from 1950 to 2023[ref 38]

However, the current rate of growth in training compute may not be sustainable. Scholars have cited the cost of training,[ref 39] a limited supply of AI chips,[ref 40] technical challenges with using that much hardware (such as managing the number of processors that must run in parallel to train larger models),[ref 41] and environmental impact[ref 42] as factors that could constrain the growth of training compute. Research in 2018 with data from OpenAI estimated that then-current trends of growth in training compute could be sustained for at most 3.5 to 10 years (2022 to 2028), depending on spending levels and how the cost of compute evolves over time.[ref 43] In 2022, that analysis was replicated with a more comprehensive dataset and suggested that this trend could be maintained for longer, for 8 to 18 years (2030 to 2040) depending on compute cost-performance improvements and specialized hardware improvements.[ref 44]

C. What Are “Scaling Laws” and What Do They Say About AI Models?

Scaling laws describe the functional (mathematical) relationship between the amount of training compute and the performance of the AI model.[ref 45] In this context, performance is a technical metric that quantifies “loss,” which is the amount of error in the model’s predictions. When loss is measured on a test or validation set that uses data not part of the training set, it reflects how well the model has generalized its learning from the training phase. The lower the loss, the more accurate and reliable the model is in making predictions on data it has not encountered during its training.[ref 46] As training compute increases, alongside increases in parameters and training data, so does model performance, meaning that greater training compute reduces the errors made.[ref 47] Increased training compute also corresponds to an increase in capabilities.[ref 48] Whereas performance refers to a technical metric, such as test loss, capabilities refer to the ability to complete concrete tasks and solve problems in the real world, including in commercial applications.[ref 49] Capabilities can also be assessed using practical and real-world tests, such as standardized academic or professional licensing exams, or with benchmarks developed for AI models. Common benchmarks include “Beyond the Imitation Game” (“BIG-Bench”), which comprises 204 diverse tasks that cover a variety of topics and languages,[ref 50] and the “Massive Multitask Language Understanding” benchmark (“MMLU”), a suite of multiple-choice questions covering 57 subjects.[ref 51] To evaluate the capabilities of Google’s PaLM 2 and OpenAI’s GPT-4, developers relied on BIG-Bench and MMLU as well as exams designed for humans, such as the SAT and AP exams.[ref 52]

Training compute has a relatively smooth and consistent relationship with technical metrics like training loss. Training compute also corresponds to real-world capabilities, but not in a smooth and predictable way. This is due in part to occasional surprising leaps, discussed in Section I.D, and subsequent enhancements such as fine-tuning, which can further increase capabilities using far less compute.[ref 53] Despite being unable to provide a full and accurate picture of a model’s final capabilities, training compute still provides a reasonable basis for estimating the base capabilities (and corresponding risk) of a foundation model. Figure 3 shows the relationship between an increase in training compute and dataset size, and performance on the MMLU benchmark.

Figure 3: Relationship between increase in training compute and dataset size,
and performance on MMLU[ref 54]

In light of the correlation between training compute and performance, the “scaling hypothesis” states that scaling training compute will predictably continue to produce even more capable systems, and thus more compute is important for AI development.[ref 55] Some have taken this hypothesis further, proposing a “Bitter Lesson:” that “the only thing that matters in the long run is the leveraging of comput[e].”[ref 56] Since the emergence of the deep learning era, this hypothesis has been sustained by the increasing use of AI models in commercial applications, whose development and commercial success have been significantly driven by increases in training compute.[ref 57]

Two factors weigh against the scaling hypothesis. First, scaling laws describe more than just the performance improvements based on training compute; they describe the optimal ratio of the size of the dataset, the number of parameters, and the training compute budget.[ref 58] Thus, a lack of abundant or high-quality data could be a limiting factor. Researchers estimate that, if training datasets continue to grow at current rates, language models will fully utilize human-generated public text data between 2026 and 2032,[ref 59] while image data could be exhausted between 2030 and 2060.[ref 60] Specific tasks may be bottlenecked earlier by the scarcity of high-quality data sources.[ref 61] There are, however, several ways that data limitations might be delayed or avoided, such as synthetic data generation and using additional datasets that are not public or in different modalities.[ref 62]

Second, algorithmic innovation permits performance gains that would otherwise require prohibitively expensive amounts of compute.[ref 63] Research estimates that every 9 months, improved algorithms for image classification[ref 64] and LLMs[ref 65] contribute the equivalent of a doubling of training compute budgets. Algorithmic improvements include more efficient utilization of data[ref 66] and parameters, the development of improved training algorithms, or new architectures.[ref 67] Over time, the amount of training compute needed to achieve a given capability is reduced, and it may become more difficult to predict performance and capabilities on that basis (although scaling trends of new algorithms could be studied and perhaps predicted). The governance implications of this are multifold, including that increases in training compute may become less important for AI development and that many more actors will be able to access the capabilities previously restricted to a limited number of developers.[ref 68] Still, responsible frontier AI development may enable stakeholders to develop understanding, safety practices, and (if needed) defensive measures for the most advanced AI capabilities before these capabilities proliferate.

D. Are High-Compute Systems Dangerous?

Advances in AI could deliver immense opportunities and benefits across a wide range of sectors, from healthcare and drug discovery[ref 69] to public services.[ref 70] However, more capable models may come with greater risk, as improved capabilities could be used for harmful and dangerous ends. While the degree of risk posed by current AI models is a subject of debate,[ref 71] future models may pose catastrophic and existential risks as capabilities improve.[ref 72] Some of these risks are expected to be closely connected to the unexpected emergence of dangerous capabilities and the dual-use nature of AI models.

As discussed in Section I.C, increases in compute, data, and the number of parameters lead to predictable improvements in model performance (test loss) and general but somewhat less predictable improvements in capabilities (real-world benchmarks and tasks). However, scaling up these inputs to a model can also result in qualitative changes in capabilities in a phenomenon known as “emergence.”[ref 73] That is, a larger model might unexpectedly display emergent capabilities not present in smaller models, suddenly able to perform a task that smaller models could not.[ref 74] During the development of GPT-3, early models had close-to-zero performance on a benchmark for addition, subtraction, and multiplication. Arithmetic capabilities appeared to emerge suddenly in later models, with performance jumping substantially above random at 2·10²² FLOP and continuing to improve with scale.[ref 75] Similar jumps were observed at different thresholds, and for different models, on a variety of tasks.[ref 76]

Some have contested the concept of emergent capabilities, arguing that what appear to be emergent capabilities in large language models are explained by the use of discontinuous measures, rather than by sharp and unpredictable improvements or developments in model capabilities with scale.[ref 77] However, discontinuous measures are often meaningful, as when the correct answer or action matters more than how close the model gets to it. As Anderljung and others explain: “For autonomous vehicles, what matters is how often they cause a crash. For an AI model solving mathematics questions, what matters is whether it gets the answer exactly right or not.”[ref 78] Given the difficulties inherent in choosing an appropriate continuous measure and determining how it corresponds to the relevant discontinuous measure,[ref 79] it is likely that capabilities will continue to seemingly emerge.

Together with emerging capabilities come emerging risks. Like many other innovations, AI systems are dual-use by nature, with the potential to be used for both beneficial and harmful ends.[ref 80] Executive Order 14,110 recognized that some models may “pose a serious risk to security, national economic security, national public health or safety” by “substantially lowering the barrier of entry for non-experts to design, synthesize, acquire, or use chemical, biological, radiological, or nuclear weapons; enabling powerful offensive cyber operations . . . ; [or] permitting the evasion of human control or oversight through means of deception or obfuscation.”[ref 81]

Predictions and evaluations will likely adequately identify many capabilities before deployment, allowing developers to take appropriate precautions. However, systems trained at a greater scale may possess novel capabilities, or improved capabilities that surpass a critical threshold for risk, yet go undetected by evaluations.[ref 82] Some of these capabilities may appear to emerge only after post-training enhancements, such as fine-tuning or more effective prompting methods. A system may be capable of conducting offensive cyber operations, manipulating people in conversation, or providing actionable instructions on conducting acts of terrorism,[ref 83] and still be deployed without the developers fully comprehending unexpected and potentially harmful behaviors. Research has already detected unexpected behavior in current models. For instance, during the recent U.K. AI Safety Summit on November 1, 2023, Apollo Research showed that GPT-4 can take illegal actions like insider trading and then lie about its actions without being instructed to do so.[ref 84] Since the capabilities of future foundation models may be challenging to predict and evaluate, “emergence” has been described as “both the source of scientific excitement and anxiety about unanticipated consequences.”[ref 85]

Not all risks come from large models. Smaller models trained on data from certain domains, such as biology or chemistry, may pose significant risks if repurposed or misused.[ref 86] When MegaSyn, a generative molecule design tool used for drug discovery, was repurposed to find the most toxic molecules instead of the least toxic, it found tens of thousands of candidates in under six hours, including known biochemical agents and novel compounds predicted to be as or more deadly.[ref 87] The amount of compute used to train DeepMind’s AlphaFold, which predicts three-dimensional protein structures from the protein sequence, is minimal compared to frontier language models.[ref 88] While scaling laws can be observed in a variety of domains, the amount of compute required to train models in some domains may be so low that a compute threshold is not a practical restriction on capabilities.

Broad consensus is forming around the need to test, monitor, and restrict systems of concern.[ref 89] The role of compute thresholds, and whether they are used at all, depends on the nature of the risk and the purpose of the policy: does it target risks from emergent capabilities of frontier models,[ref 90] risks from models with more narrow but dangerous capabilities,[ref 91] or other risks from AI?

E. Does Compute Usage Outside of Training Influence Performance and Risk?

In light of the relationship between training compute and performance expressed by scaling laws, training compute is a common proxy for how capable and powerful AI models are and the risks that they pose.[ref 92] However, compute used outside of training can also influence performance, capabilities, and corresponding risk.

As discussed in Section I.A, training compute typically does not refer to all compute used during development, but is instead limited to compute used during the final pre-training run.[ref 93] This definition excludes subsequent (post-training) enhancements, such as fine-tuning and prompting methods, which can significantly improve capabilities (see supra Figure 1) using far less compute; many current methods can improve capabilities the equivalent of a 5x increase in training compute, while some can improve them by more than 20x.[ref 94]

The focus on training compute also misses the significance of compute used for inference, in which the trained model generates output in response to a prompt or new input data.[ref 95] Inference is the biggest compute cost for models deployed at scale, due to the frequency and volume of requests they handle.[ref 96] While developing an AI model is far more computationally intensive than a single inference request, it is a one-time task. In contrast, once a model is deployed, it may receive numerous inference requests that, in aggregate, exceed the compute expenditures of training. Some have even argued that inference compute could be a bottleneck in scaling AI, if inference compute costs scaling with training compute grow too large.[ref 97]

Greater availability of inference compute could enhance malicious uses of AI by allowing the model to process data more rapidly and enabling the operation of multiple instances in parallel. For example, AI could more effectively be used to carry out cyber attacks, such as a distributed denial-of-service (“DDoS”) attack,[ref 98] to manipulate financial markets,[ref 99] or to increase the speed, scale, and personalization of disinformation campaigns.[ref 100]

Compute used outside of development may also impact model performance. Specifically, some techniques can increase the performance of a model at the cost of more compute used during inference.[ref 101] Developers could therefore choose to improve a model beyond its current capabilities or to shift some compute expenditures from training to inference, in order to obtain equally-capable systems with less training compute. Users could also prompt a model to use similar techniques during inference, for example by (1) using “few-shot” prompting, in which initial prompts provide the model with examples of the desired output for a type of input,[ref 102] (2) using chain-of-thought prompting, which uses few-shot prompting to provide examples of reasoning,[ref 103] or (3) simply providing the same prompt multiple times and selecting the best result. Some user-side techniques to improve performance might increase the compute used during a single inference, while others would leave it unchanged (while still increasing the total compute used, due to multiple inferences being performed).[ref 104] Meanwhile, other techniques—such as pruning,[ref 105] weight sharing,[ref 106] quantization,[ref 107] and distillation[ref 108]—can reduce compute used during inference while maintaining or even improving performance, and they can further reduce inference compute at the cost of lower performance.

Beyond model characteristics such as parameter count, other factors can also affect the amount of compute used during inference in ways that may or may not improve performance, such as input size (compare a short prompt to a long document or high-resolution image) and batch size (compare one input provided at a time to many inputs in a single prompt).[ref 109] Thus, for a more accurate indication of model capabilities, compute used to run a single inference[ref 110] for a given set of prompts could be considered alongside other factors, such as training compute. However, doing so may be impractical, as data about inference compute (or architecture useful for estimating it) is rarely published by developers,[ref 111] different techniques could make inference more compute-efficient, and less information is available regarding the relationship between inference compute and capabilities.

While companies might be hesitant to increase inference compute at scale due to cost, doing so may still be worthwhile in certain circumstances, such as for more narrowly deployed models or those willing to pay more for improved capabilities. For example, OpenAI offers dedicated instances for users who want more control over system performance, with a reserved allocation of compute infrastructure and the ability to enable features such as longer context limits.[ref 112]

Over time, compute usage during the AI development and deployment process may change. It was previously common practice to train models with supervised learning, which uses annotated datasets. In recent years, there has been a rise in self-supervised, semi-supervised, and unsupervised learning, which use data with limited or no annotation but require more compute.[ref 113] 

III. The Role of Compute Thresholds for AI Governance

A. How Can Compute Thresholds Be Used in AI Policy?

Compute can be used as a proxy for the capabilities of AI systems, and compute thresholds can be used to define the limited subset of high-compute models subject to oversight or other requirements.[ref 114] Their use depends on the context and purpose of the policy. Compute thresholds serve as intuitive starting points to identify potential models of concern,[ref 115] perhaps alongside other factors.[ref 116] They operate as a trigger for greater scrutiny or specific requirements. Once a certain level of training compute is reached, a model is presumed to have a higher risk of displaying dangerous capabilities (and especially unknown dangerous capabilities) and, hence, is subject to stricter oversight and other requirements.

Compute thresholds have already entered AI policy. The EU AI Act requires model providers to assess and mitigate systemic risks, report serious incidents, conduct state-of-the-art tests and model evaluations, ensure cybersecurity, and report serious incidents if a compute threshold is crossed.[ref 117] Under the EU AI Act, a general-purpose model that meets the initial threshold is presumed to have high-impact capabilities and associated systemic risk.[ref 118]

In the United States, Executive Order 14,110 directed agencies to propose rules based on compute thresholds. Although it was revoked by President Trump’s Executive Order 14,148,[ref 119] many actions have already been taken and rules have been proposed for implementing Executive Order 14,110. For instance, the Department of Commerce’s Bureau of Industry and Security issued a proposed rule on September 11, 2024[ref 120] to implement the requirement that AI developers and cloud service providers report on models above certain thresholds, including information about (1) “any ongoing or planned activities related to training, developing, or producing dual-use foundation models,” (2) the results of red-teaming, and (3) the measures the company has taken to meet safety objectives.[ref 121] The executive order also imposed know-your-customer (“KYC”) monitoring and reporting obligations on U.S. cloud infrastructure providers and their foreign resellers, again with a preliminary compute threshold.[ref 122] On January 29, 2024, the Bureau of Industry and Security issued a proposed rule implementing those requirements.[ref 123] The proposed rule noted that training compute thresholds may determine the scope of the rule; the program is limited to foreign transactions to “train a large AI model with potential capabilities that could be used in malicious cyber-enabled activity,” and technical criteria “may include the compute used to pre-train the model exceeding a specified quantity.” ^{[ref 124]} The fate of these rules is uncertain, as all rules and actions taken pursuant to Executive Order 14,110 will be reviewed to ensure that they are consistent with the AI policy set forth in Executive Order 14,179, Removing Barriers to American Leadership in Artificial Intelligence.[ref 125] Any rules of actions identified as inconsistent are directed to be suspended, revised, or rescinded.[ref 126]

Numerous policy proposals have likewise called for compute thresholds. Scholars and developers alike have expressed support for a licensing or registration regime,[ref 127] and a compute threshold could be one of several ways to trigger the requirement.[ref 128] Compute thresholds have also been proposed for determining the level of KYC requirements for compute providers (including cloud providers).[ref 129] The Framework to Mitigate AI-Enabled Extreme Risks, proposed by U.S. Senators Romney, Reed, Moran, and King, would include a compute threshold for requiring notice of development, model evaluation, and pre-deployment licensing.[ref 130]

Other AI regulations and policy proposals do not explicitly call for the introduction of compute thresholds but could still benefit from them. A compute threshold could clarify when specific obligations are triggered in laws and guidance that refer more broadly to “advanced systems” or “systems with dangerous capabilities,” as in the voluntary guidance for “organizations developing the most advanced AI systems” in the Hiroshima Process International Code of Conduct for Advanced AI Systems, agreed upon by G7 leaders on October 30, 2023.[ref 131] Compute thresholds could identify when specific obligations are triggered in other proposals, including proposals for: (1) conducting thorough risk assessments of frontier AI models before deployment;[ref 132] (2) subjecting AI development to evaluation-gated scaling;[ref 133] (3) pausing development of frontier AI;[ref 134] (4) subjecting developers of advanced models to governance audits;[ref 135] (5) monitoring advanced models after deployment;[ref 136] and (6) requiring that advanced AI models be subject to information security protections.[ref 137]

B. Why Might Compute Be Relevant Under Existing Law?

Even without a formal compute threshold, the significance of training compute could affect the interpretation and application of existing laws. Courts and regulators may rely on compute as a proxy for how much risk a given AI system poses—alongside other factors such as capabilities, domain, safeguards, and whether the application is in a higher-risk context—when determining whether a legal condition or regulatory threshold has been met. This section briefly covers a few examples. First, it discusses the potential implications for duty of care and foreseeability analyses in tort law. It then goes on to describe how regulatory agencies could depend on training compute as one of several factors in evaluating risk from frontier AI, for example as an indicator of change to a regulated product and as a factor in regulatory impact analysis.

The application of existing laws and ongoing development of common law, such as tort law, may be particularly important while AI governance is still nascent[ref 138] and may operate as a complement to regulations once developed.[ref 139] However, courts and regulators will face new challenges as cases involve AI, an emerging technology of which they have no specialized knowledge, and parties will face uncertainty and inconsistent judgments across jurisdictions. As developments in AI unsettle existing law[ref 140] and agency practice, courts and agencies might rely on compute in several ways.

For example, compute could inform the duty of care owed by developers who make voluntary commitments to safety.[ref 141] A duty of care, which is a responsibility to take reasonable care to avoid causing harm to another, can be conditioned on the foreseeability of the plaintiff as a victim or be an affirmative duty to act in a particular way; affirmative duties can arise from the relationship between the parties, such as between business owner and customer, doctor and patient, and parent and child.[ref 142] If AI companies make general commitments to security testing and cybersecurity, such as the voluntary safety commitments secured by the Biden administration,[ref 143] those commitments may give rise to a duty of care in which training compute is a factor in determining what security is necessary. If a lab adopts a responsible scaling policy that requires it to have protection measures based on specific capabilities or potential for risk or misuse,[ref 144] a court might consider training compute as one of several factors in evaluating the potential for risk or misuse.

A court might also consider training compute as a factor when determining whether a harm was foreseeable. More advanced AI systems, trained with more compute, could foreseeably be capable of greater harm, especially in light of scaling laws discussed in Section I.C that make clear the relationship between compute and performance. It may likewise be foreseeable that a powerful AI system could be misused[ref 145] or become the target of more sophisticated attempts at exfiltration, which might succeed without adequate security.[ref 146] Foreseeability may in turn bear on negligence elements of proximate causation and duty of care.

Compute could also play a role in other scenarios, such as in a false advertising claim under the Lanham Act[ref 147] or state and federal consumer protection laws. If a business makes a claim about its AI system or services that is false or misleading, it could be held liable for monetary damages and enjoined from making that claim in the future (unless it becomes true).[ref 148] While many such claims will not involve compute, some may; for example, if a lab publicly claims to follow a responsible scaling policy, training compute could be relevant as an indicator of model capability and the corresponding security and safety measures promised by the policy.

Regulatory agencies may likewise consider compute in their analyses and regulatory actions. For example, the Environmental Protection Agency could consider training (and inference) compute usage as part of environmental impact assessments.[ref 149] Others could treat compute as a proxy for threat to national or public security. Agencies and committees responsible for identifying and responding to various risks, such as the Interagency Committee on Global Catastrophic Risk[ref 150] and Financial Stability Oversight Council,[ref 151] could consider compute in their evaluation of risk from frontier AI. Over fifty federal agencies were directed to take specific actions to promote responsible development, deployment, federal use of AI, and regulation of industry, in the government-wide effort established by Executive Order 14,110[ref 152]—although these actions are now under review.[ref 153] Even for agencies not directed to consider compute or implement a preliminary compute threshold, compute might factor into how guidance is implemented over time.

More speculatively, changes to training compute could be used by agencies as one of many indicators of how much a regulated product has changed, and thus whether it warrants further review. For example, the Food and Drug Administration might consider compute when evaluating AI in medical devices or diagnostic tools.[ref 154] While AI products considered to be medical devices are more likely to be narrow AI systems trained on comparatively less compute, significant changes to training compute may be one indicator that software modifications require premarket submission. The ability to measure, report, and verify compute[ref 155] could make this approach particularly compelling for regulators.

Finally, training compute may factor into regulatory impact analyses, which evaluate the impact of proposed and existing regulations through quantitative and qualitative methods such as cost-benefit analysis.[ref 156] While this type of analysis is not necessarily determinative, it is often an important input into regulatory decisions and necessary for any “significant regulatory action.”[ref 157] As agencies develop and propose new regulations and consider how those rules will affect or be affected by AI, compute could be relevant in drawing lines that define what conduct and actors are affected. For example, a rule with a higher compute threshold and narrower scope may be less significant and costly, as it covers fewer models and developers. The amount of compute used to train models now and in the future may be not only a proxy for threat to national security (or innovation, or economic growth), but also a source of uncertainty, given the potential for emergent capabilities.

C. Where Should the Compute Threshold(s) Sit?

The choice of compute threshold depends on the policy under consideration: what models are the intended target, given the purpose of the policy? What are the burdens and costs of compliance? Can the compute threshold be complemented with other elements for determining whether a model falls within the scope of the policy, in order to more precisely accomplish its purpose?

Some policy proposals would establish a compute threshold “at the level of FLOP used to train current foundational models.”[ref 158] While the training compute of many models is not public, according to estimates, the largest models today were trained with 1e25 FLOP or more, including at least one open-source model, Llama 3.1 405B.[ref 159] This is the initial threshold established by the EU AI Act. Under the Act, general-purpose AI models are considered to have “systemic risk,” and thus trigger a series of obligations for their providers, if found to have “high impact capabilities.”[ref 160] Such capabilities are presumed if the cumulative amount of training compute, which includes all “activities and methods that are intended to enhance the capabilities of the model prior to deployment, such as pre-training, synthetic data generation and fine-tuning,” exceeds 1e25 FLOP.[ref 161] This threshold encompasses existing models such as Gemini Ultra and GPT-4, and it can be updated upwards or downwards by the European Commission through delegated acts.[ref 162] During the AI Safety Summit held in 2023, the U.K. Government included current models by defining “frontier AI” as “highly capable general-purpose AI models that can perform a wide variety of tasks and match or exceed the capabilities present in today’s most advanced models” and acknowledged that the definition included the models underlying ChatGPT, Claude, and Bard.[ref 163]

Others have proposed an initial threshold of “more training compute than already-deployed systems,”[ref 164] such as 1e26 FLOP[ref 165] or 1e27 FLOP.[ref 166] No known model currently exceeds 1e26 FLOP training compute, which is roughly five times the compute used to train GPT-4.[ref 167] These higher thresholds would more narrowly target future systems that pose greater risks, including potential catastrophic and existential risks.[ref 168] President Biden’s Executive Order on AI[ref 169] and recently-vetoed California Senate Bill 1047[ref 170] are in line with these proposals, both targeting models trained with more than 1e26 OP or FLOP.

Far more models would fall within the scope of a compute threshold set lower than current frontier models. While only two models exceeded 1e23 FLOP training compute in 2017, over 200 models meet that threshold today.[ref 171] As discussed in Section II.A, compute thresholds operate as a trigger for additional scrutiny, and more models falling within the ambit of regulation would entail a greater burden not only on developers, but also on regulators.[ref 172] These smaller, general-purpose models have not yet posed extreme risks, making a lower threshold unwarranted at this time.[ref 173]

While the debate has centered mostly around the establishment of a single training compute threshold, governments could adopt a pluralistic and risk-adjusted approach by introducing multiple compute thresholds that trigger different measures or requirements according to the degree or nature of risk. Some proposals recommend a tiered approach that would create fewer obligations for models trained on less compute. For example, the Responsible Advanced Artificial Intelligence Act of 2024 would require pre-registration and benchmarks for lower-compute models, while developers of higher-compute models must submit a safety plan and receive a permit prior to training or deployment.[ref 174] Multi-tiered systems may also incorporate a higher threshold beyond which no development or deployment can take place, with limited exceptions, such as for development at a multinational consortium working on AI safety and emergency response infrastructure[ref 175] or for training runs and models with strong evidence of safety.[ref 176]

Domain-specific thresholds could be established for models that possess capabilities or expertise in areas of concern and models that are trained using less compute than general-purpose models.[ref 177] A variety of specialized models are already available to advance research, trained on extensive scientific databases.[ref 178] As discussed in Part I.D, these models present a tremendous opportunity, yet many have also recognized the potential threat of their misuse to research, develop, and use chemical, biological, radiological, and nuclear weapons.[ref 179] To address these risks, President Biden’s Executive Order on AI, which set a compute threshold of 1e26 FLOP to trigger reporting requirements, set a substantially lower compute threshold of 1e23 FLOP for models trained “using primarily biological sequence data.”[ref 180] The Hiroshima Process International Code of Conduct for Advanced AI Systems likewise recommends devoting particular attention to offensive cyber capabilities and chemical, biological, radiological, and nuclear risks, although it does not propose a compute threshold.[ref 181]

While domain-specific thresholds could be useful for a variety of policies tailored to specific risks, there are some limitations. It may be technically difficult to verify how much biological sequence data (or other domain-specific data) was used to train a model.[ref 182] Another challenge is specifying how much data in a given domain causes a model to fall within scope, particularly considering the potential capabilities of models trained on mixed data.[ref 183] Finally, the amount of training compute required may be so low that, over time, a compute threshold is not practical.

When choosing a threshold, regulators should be aware that capabilities might be substantially improved through post-training enhancements, and training compute is only a general predictor of capabilities. The absolute limits are unclear at this point; however, current methods can result in capability improvements equivalent to a 5- to 30-times increase in training.[ref 184] To account for post-training enhancements, a governance regime could create a safety buffer, in which oversight or other protective measures are set at a lower threshold.[ref 185] Along similar lines, open-source models may warrant a lower threshold for at least some regulatory requirements, since they could be further trained by another actor and, once released, cannot be moderated or rescinded. ^{[ref 186]}

D. Does a Compute Threshold Require Updates?

Once established, compute thresholds and related criteria will likely require updates over time.[ref 187] Improvements in algorithmic efficiency could reduce the amount of compute needed to train an equally capable model,[ref 188] or a threshold could be raised or eliminated if adequate protective measures are developed or if models trained with a certain amount of compute are demonstrated to be safe.[ref 189] To further guard against future developments in a rapidly evolving field, policymakers can authorize regulators to update compute thresholds and related criteria.[ref 190]

Several policies, proposed and enacted, have incorporated a dynamic compute threshold. For example, President Biden’s Executive Order on AI authorized the Secretary of Commerce to update the initial compute threshold set in the order, as well as other technical conditions for models subject to reporting requirements, “as needed on a regular basis” while establishing an interim compute threshold of 1e26 OP or FLOP.[ref 191] Similarly, the EU AI Act provides that the 1e25 FLOP compute threshold “should be adjusted over time to reflect technological and industrial changes, such as algorithmic improvements” and authorizes the European Commission to amend the threshold and “supplement benchmarks and indicators in light of evolving technological developments.”[ref 192] The California Senate Bill 1047 would have created the Frontier Model Division within the Government Operations Agency and authorized it to “update both of the [compute] thresholds in the definition of a ‘covered model’ to ensure that it accurately reflects technological developments, scientific literature, and widely accepted national and international standards and applies to artificial intelligence models that pose a significant risk of causing or materially enabling critical harms.”[ref 193]

Regulators may need to update compute thresholds rapidly. Historically, failure to quickly update regulatory definitions in the context of emerging technologies has led to definitions becoming useless or even counterproductive.[ref 194] In the field of AI, developments may occur quickly and with significant implications for national security and public health, making responsive rulemaking particularly important. In the United States, there are several statutory tools to authorize and encourage expedited and regular rulemaking.[ref 195] For example, Congress could expressly authorize interim or direct final rulemaking, which would enable an agency to shift the comment period in notice-and-comment rulemaking to take place after the rule has already been promulgated, thereby allowing them to respond quickly to new developments.[ref 196]

Policymakers could also require a periodic evaluation of whether compute thresholds are achieving their purpose to ensure that it does not become over- or under-inclusive. While establishing and updating a compute threshold necessarily involves prospective ex ante impact assessment, in order to take precautions against risk without undue burdens, regulators can learn much from retrospective ex post analysis of current and previous thresholds.[ref 197] In a survey conducted for the Administrative Conference of the United States, “[a]ll agencies stated that periodic reviews have led to substative [sic] regulatory improvement at least some of time. This was more likely when the underlying evidence basis for the rule, particularly the science or technology, was changing.”[ref 198] While the optimal frequency of periodic review is unknown, the study found that U.S. federal agencies were more likely to conduct reviews when provided with a clear time interval (“at least every X years”).[ref 199]

Several further institutional and procedural factors could affect whether and how compute thresholds are updated. In order to effectively update compute thresholds and other criteria, regulators must have access to expertise and talent through hiring, training, consultation and collaboration, and other avenues that facilitate access to experts from academia and industry.[ref 200] Decisions will be informed by the availability of data, including scientific and commercial data, to enable ongoing monitoring, learning, analysis, and adaptation in light of new developments. Decision-making procedures, agency design, and influence and pressures from policymakers, developers, and other stakeholders will likewise affect updates, among many other factors.[ref 201] While more analysis is beyond the scope of this Article, others have explored procedural and substantive measures for adaptive regulation[ref 202] and effective governance of emerging technologies.[ref 203]

Some have proposed defining compute thresholds in terms of effective compute,[ref 204] as an alternative to updates over time. Effective compute could index to a particular year (similar to inflation adjustments) and thus account for the role that algorithmic progress (e.g., 1e25 of 2023-level effective compute).[ref 205] However, there is not an agreed upon way to more precisely define and calculate effective compute, and the ability to do so depends on the challenging task of calculating algorithmic efficiency, including choosing a performance metric to anchor on. Furthermore, effective compute alone would fail to address potential changes in the risk landscape, such as the development of protective measures.

E. What Are the Advantages and Limitations of a Training Compute Threshold?

Compute has several properties that make it attractive for policymaking: it is (1) correlated with capabilities and thus risk, (2) essential for training, with thresholds that are difficult to circumvent without reducing performance, (3) an objective and quantifiable measure, (4) capable of being estimated before training (5) externally verifiable after training, and (6) a significant cost during development and thus indicative of developer resources. However, training compute thresholds are not infallible: (1) training compute is an imprecise indicator of potential risk, (2) a compute threshold could be circumvented, and (3) there is no industry standard for measuring and reporting training compute.[ref 206] Some of these limitations can be addressed with thoughtful drafting, including clear language, alternative and supplementary elements for defining what models are within scope, and authority to update any compute threshold and other criteria in light of future developments.

First, training compute is correlated with model capabilities and associated risks. Scaling laws predict an increase in performance as training compute increases, and real-world capabilities generally follow (Section I.C). As models become more capable, they may also pose greater risks if they are misused or misaligned (Section I.D). However, training compute is not a precise indicator of downstream capabilities. Capabilities can seemingly emerge abruptly and discontinuously as models are developed with more compute,[ref 207] and the open-ended nature of foundation models means those capabilities may go undetected.[ref 208] Post-training enhancements such as fine-tuning are often not considered a part of training compute, yet they can dramatically improve performance and capabilities with far less compute. Furthermore, not all models with dangerous capabilities require large amounts of training compute; low-compute models with capabilities in certain domains, such as biology or chemistry, may also pose significant risks, such as biological design tools that could be used for drug discovery or the creation of pathogens worse than any seen to date.[ref 209] The market may shift towards these smaller, cheaper, more specialized models,[ref 210] and even general-purpose low-compute models may come to pose significant risks. Given these limitations, a training compute threshold cannot capture all possible risks; however, for large, general-purpose AI models, training compute can act as an initial threshold for capturing emerging capabilities and risks.

Second, compute is necessary throughout the AI lifecycle, and a compute threshold would be difficult to circumvent. There is no AI without compute (Section I.A). Due to its relationship with model capabilities, training compute cannot be easily reduced without a corresponding reduction in capabilities, making it difficult to circumvent for developers of the most advanced models. Nonetheless, companies might find “creative ways” to account for how much compute is used for a given system in order to avoid being subject to stricter regulation.[ref 211] To reduce this risk, some have suggested monitoring compute usage below these thresholds to help identify circumvention methods, such as structuring techniques or outsourcing.[ref 212] Others have suggested using compute thresholds alongside additional criteria, such as the model’s performance on benchmarks, financial or energy cost, or level of integration into society.[ref 213] As in other fields, regulatory burdens associated with compute thresholds could encourage regulatory arbitrage if a policy does not or cannot effectively account for that possibility.[ref 214] For example, since compute can be accessed remotely via digital means, data centers and compute providers could move to less-regulated jurisdictions.

Third, compute is an objective and quantifiable metric that is relatively straightforward to measure. Compute is a quantitative measure that reflects the number of mathematical operations performed. It does not depend on specific infrastructure and can be compared across different sets of hardware and software.[ref 215] By comparison, other metrics, such as algorithmic innovation and data, have been more difficult to track.[ref 216] Whereas quantitative metrics like compute can be readily compared across different instances, the qualitative nature of many other metrics makes them more subject to interpretation and difficult to consistently measure. Compute usage can be measured internally with existing tools and systems; however, there is not yet an industry standard for measuring, auditing, and reporting the use of computational resources.[ref 217] That said, there have been some efforts toward standardization of compute measurement.[ref 218] In the absence of a standard, some have instead presented a common framework for calculating compute, based on information about the hardware used and training time.[ref 219]

Fourth, compute can be estimated ahead of model development and deployment. Developers already estimate training compute with information about the model’s architecture and amount of training data, as part of planning before training takes place. The EU AI Act recognizes this, noting that “training of general-purpose AI models takes considerable planning which includes the upfront allocation of compute resources and, therefore, providers of general-purpose AI models are able to know if their model would meet the threshold before the training is completed.”[ref 220] Since compute can be readily estimated before a training run, developers can plan a model with existing policies in mind and implement appropriate precautions during training, such as cybersecurity measures.

Fifth, the amount of compute used could be externally verified after training. While laws that use compute thresholds as a trigger for additional measures could depend on self-reporting, meaningful enforcement requires regulators to be aware of or at least able to verify the amount of compute being used. A regulatory threshold will be ineffective if regulators have no way of knowing whether a threshold has been reached. For this reason, some scholars have proposed that developers and compute providers be required to report the amount of compute used at different stages of the AI lifecycle.[ref 221] Compute providers already employ chip-hours for client billing, which could be used to calculate total computational operations,[ref 222] and the centralization of a few key cloud providers could make monitoring and reporting requirements simpler to administer.[ref 223] Others have proposed using “on-chip” or “hardware-enabled governance mechanisms” to verify claims about compute usage.[ref 224]

Sixth, training compute is an indicator of developer resources and capacity to comply with regulatory requirements, as it represents a substantial financial investment.[ref 225] For instance, Sam Altman reported that the development of GPT-4 cost “much more” than $100 million.[ref 226] Researchers have estimated that Gemini Ultra cost $70 million to $290 million to develop.[ref 227] A regulatory approach based on training compute thresholds can therefore be used to subject only the most resourced AI developers to increased regulatory scrutiny, while avoiding overburdening small companies, academics, and individuals. Over time, the cost of compute will most likely continue to fall, meaning the same thresholds will capture more developers and models. To ensure that the law remains appropriately scoped, compute thresholds can be complemented by additional metrics, such as the cost of compute or development. For example, the vetoed California Senate Bill 1047 was amended to include a compute cost threshold, defining a “covered model” to include one trained with over 1e26 OP, only if the cost of that training compute exceeded $100,000,000 at the start of training.[ref 228]

At the time of writing, many consider compute thresholds to be the best option currently available for determining which AI models should be subject to regulation, although the limitations of this approach underscore the need for careful drafting and adaptive governance. When considering the legal obligations imposed, the specific compute threshold should correspond to the nature and extent of additional scrutiny and other requirements and reflect the fact that compute is only a proxy for, and not a precise measure of, risk.

F. How Do Compute Thresholds Compare to Capability Evaluations?

A regulatory approach that uses a capabilities-based threshold or evaluation may seem more intuitively appealing and has been proposed by many.[ref 229] There are currently two main types of capability evaluations: benchmarking and red-teaming.[ref 230] In benchmarking, a model is tested on a specific dataset and receives a numerical score. In red-teaming, evaluators can use different approaches to identify vulnerabilities and flaws in a system, such as through prompt injection attacks to subvert safety guardrails. Model evaluations like these already serve as the basis for responsible scaling policies, which specify what protective measures an AI developer must implement in order to safely handle a given level of capabilities. Responsible scaling policies have been adopted by companies like Anthropic, OpenAI, and Google, and policymakers have also encouraged their development and practice.[ref 231]

Capability evaluations can complement compute thresholds. For example, capability evaluations could be required for models exceeding a compute threshold that indicates that dangerous capabilities might exist. They could also be used as an alternative route to being covered by regulation. The EU AI Act adopts the latter approach, complementing the compute threshold with the possibility for the European Commission to “take individual decisions designating a general-purpose AI model as a general-purpose AI model with systemic risk if it is found that such model has capabilities or an impact equivalent to those captured by the set threshold.”[ref 232]

Nonetheless, there are several downsides to depending on capabilities alone. First, model capabilities are difficult to measure.[ref 233] Benchmark results can be affected by factors other than capabilities, such as benchmark data being included during training[ref 234] and model sensitivity to small changes in prompting.[ref 235] Downstream capabilities of a model may also differ from those during evaluation due to changes in dataset distribution.[ref 236] Some threats, such as misuse of a model to develop a biological weapon, may be particularly difficult to evaluate due to the domain expertise required, the sensitivity of information related to national security, and the complexity of the task.[ref 237] For dangerous capabilities such as deception and manipulation, the nature of the capability makes it difficult to assess,[ref 238] although some evaluations have already been developed.[ref 239] Furthermore, while evaluations can point to what capabilities do exist, it is far more difficult to prove that a model does not possess a given capability. Over time, new capabilities may even emerge and improve due to prompting techniques, tools, and other post-training enhancements.

Second, and compounding the issue, there is no standard method for evaluating model capabilities.[ref 240] While benchmarks allow for comparison across models, there are competing benchmarks for similar capabilities; with none adopted as standard by developers or the research community, evaluators could select different benchmark tests entirely.[ref 241] Red-teaming, while more in-depth and responsive to differences in models, is even less standardized and provides less comparable results. Similarly, no standard exists for when during the AI lifecycle a model is evaluated, even though fine-tuning and other post-training enhancements can have a significant impact on capabilities. Nevertheless, there have been some efforts toward standardization, including the U.S. National Institute of Standards and Technology beginning to develop guidelines and benchmarks for evaluating AI capabilities, including through red-teaming.[ref 242]

Third, it is much more difficult to externally verify model evaluations. Since evaluation methods are not standardized, different evaluators and methods may come to different conclusions, and even a small difference could determine whether a model falls within the scope of regulation. This makes external verification simultaneously more important and more challenging. In addition to the technical challenge of how to consistently verify model evaluations, there is also a practical challenge: certain methods, such as red-teaming and audits, depend on far greater access to a model and information about its development. Developers have been reluctant to grant permissive access,[ref 243] which has contributed to numerous calls to mandate external evaluations.[ref 244]

Fourth, model evaluations may be circumvented. For red-teaming and more comprehensive audits, evaluations for a given model may reasonably reach different conclusions, which allows room for an evaluator to deliberately shape results through their choice of methods and interpretation. Careful institutional design is needed to ensure that evaluations are robust to conflicts of interest, perverse incentives, and other limitations.[ref 245] If known benchmarks are used to determine whether a model is subject to regulation, developers might train models to achieve specific scores without affecting capabilities, whether to improve performance on safety measures or strategically underperform on certain measures of dangerous capabilities.

Finally, capability evaluations entail more uncertainty and expense. Currently, the capabilities of a model can only reliably be determined ex post,[ref 246] making it difficult for developers to predict whether it will fall within the scope of applicable law. More in-depth model evaluations such as red-teaming and audits are expensive and time-consuming, which may constrain small organizations, academics, and individuals.[ref 247]

Capability evaluations can thus be viewed as a complementary tool for estimating model risk. While training compute makes an excellent initial threshold for regulatory oversight, as an objective and quantifiable measure that can be estimated prior to training and verified after, capabilities correspond more closely to risk. Capability evaluations provide more information and can be completed after fine-tuning and other post-training enhancements, but are more expensive, difficult to carry out, and less standardized. Both are important components of AI governance but serve different roles.

IV. Conclusion

More powerful AI could bring transformative changes in society. It promises extraordinary opportunities and benefits across a wide range of sectors, with the potential to improve public health, make new scientific discoveries, improve productivity and living standards, and accelerate economic growth. However, the very same advanced capabilities could result in tremendous harms that are difficult to control or remedy after they have occurred. AI could fail in critical infrastructure, further concentrate wealth and increase inequality, or be misused for more effective disinformation, surveillance, cyberattacks, and development of chemical and biological weapons.

In order to prevent these potential harms, laws that govern AI must identify models that pose the greatest threat. The obvious answer would be to evaluate the dangerous capabilities of frontier models; however, state of the art model evaluations are subjective and unable to reliably predict downstream capabilities, and they can take place only after the model has been developed with a substantial investment.

This is where training compute thresholds come into play. Training compute can operate as an initial threshold for estimating the performance and capabilities of a model and, thus, the potential risk it poses. Despite its limitations, it may be the most effective option we have to identify potentially dangerous AI that warrants further scrutiny. However, compute thresholds alone are not sufficient. They must be used alongside other tools to mitigate and respond to risk, such as capability evaluations, post-market monitoring, and incident reporting. Further research avenues could develop better governance via compute thresholds:

1. What amount of training compute corresponds to future systems of concern? What threshold is appropriate for different regulatory targets, and how can we identify that threshold in advance? What are the downstream effects of different compute thresholds?
2. Are compute thresholds appropriate for different stages of the AI lifecycle? For example, could thresholds for compute used for post-training enhancements or during inference be used alongside a training compute threshold, given the ability to significantly improve capabilities at these stages?
3. Should domain-specific compute thresholds be established, and if so, to address which risks? If domain-specific compute thresholds are established, such as in President Biden’s Executive Order 14,110, how can competent authorities determine if a system is domain-specific and verify the training data?
4. How should compute usage be reported, monitored, and audited?
5. How should a compute threshold be updated over time? What is the likelihood of future frontier systems being developed using less (or far less) compute than is used today? Does growth or slowdown in compute usage, hardware improvement, or algorithmic efficiency warrant an update, or should it correspond solely to an increase in capabilities? Relatedly, what kind of framework would allow a regulatory agency to respond to developments effectively (e.g., with adequate information and the ability to update rapidly)?
6. How could a capabilities-based threshold complement or replace a compute threshold, and what would be necessary (e.g., improved model evaluations for dangerous capabilities and alignment)?
7. How should the law mitigate risks from AI systems that sit below the training compute threshold?

What should be internationalised in AI governance?

The governance misspecification problem

In technical Artificial Intelligence (“AI”) safety research, the term “specification” refers to the problem of defining the purpose of an AI system so that the system behaves in accordance with the true wishes of its designer.[ref 1] Technical researchers have suggested three categories of specification: “ideal specification,” “design specification,” and “revealed specification.”[ref 2] The ideal specification, in this framework, is a hypothetical specification that would create an AI system completely and perfectly aligned with the desires of its creators. The design specification is the specification that is actually used to build a given AI system. The revealed specification is the specification that best describes the actual behavior of the completed AI system. “Misspecification” occurs whenever the revealed specification of an AI system diverges from the ideal specification—i.e., when an AI system does not perform in accordance with the intentions of its creators. 

The fundamental problem of specification is that “it is often difficult or infeasible to capture exactly what we want an agent to do, and as a result we frequently end up using imperfect but easily measured proxies.”[ref 3] Thus, in a famous example from 2016, researchers at OpenAI attempted to train a reinforcement learning agent to play the boat-racing video game CoastRunners, the goal of which is to finish a race quickly and ahead of other players.[ref 4] Instead of basing the AI agent’s reward function on how it placed in the race, however, the researchers used a proxy goal that was easier to implement and rewarded the agent for maximizing the number of points it scored. The researchers mistakenly assumed that the agent would pursue this proxy goal by trying to complete the course quickly. Instead, the AI discovered that it could achieve a much higher score by refusing to complete the course and instead driving in tight circles in such a way as to repeatedly collect a series of power-ups while crashing into other boats and occasionally catching on fire.[ref 5] In other words, the design specification (“collect as many points as possible”) did not correspond well to the ideal specification (“win the race”), leading to a disastrous and unexpected revealed specification (crashing repeatedly and failing to finish the race). 

This article applies the misspecification framework to the problem of AI governance. The resulting concept, which we call the “governance misspecification problem,” can be briefly defined as occurring when a legal rule relies unsuccessfully on proxy terms or metrics. By framing this new concept in terms borrowed from the technical AI safety literature, we hope to incorporate valuable insights from that field into legal-philosophical discussions around the nature of rules and, importantly, to help technical researchers understand the philosophical and policymaking challenges that AI governance legislation and regulation poses. 

It is generally accepted among legal theorists that at least some legal rules can be said to have a purpose or purposes and that these purposes should inform the interpretation of textually ambiguous rules.[ref 6] The least ambitious version of this claim is simply an acknowledgment of the fact that statutes often contain a discrete textual provision entitled “Purpose,” which is intended to inform the interpretation and enforcement of the statute’s substantive provisions.[ref 7] More controversially, some commentators have argued that all or many legal rules have, or should be constructively understood as having, an underlying “true purpose,” which may or may not be fully discoverable and articulable.[ref 8] 

The purpose of a legal rule is analogous to the “ideal specification” discussed in the technical AI safety literature. Like the ideal specification of an AI system, a rule’s purpose may be difficult or impossible to perfectly articulate or operationalize, and rulemakers may choose to rely on a legal regime that incorporates “imperfect but easily measured proxies”—essentially, a design specification. “Governance misspecification” occurs when the real-world effects of the legal regime (analogous to the design specification) as interpreted and enforced (analogous to the revealed specification) fail to effectuate the rule’s intended purpose (analogous to the ideal specification).

Consider the hypothetical legal rule prohibiting the presence of “vehicles” in a public park, famously described by the legal philosopher H.L.A. Hart.[ref 9] The term “vehicles,” in this rule, is presumably a proxy term intended to serve some ulterior purpose,[ref 10] although fully discovering and articulating that purpose may be infeasible. For example, the rule might be intended to ensure the safety of pedestrians in the park, or to safeguard the health of park visitors by improving the park’s air quality, or to improve the park’s atmosphere by preventing excessive noise levels. More realistically, the purpose of the rule might be some complex weighted combination of all of these and numerous other more or less important goals. Whether the rule is misspecified depends on whether the rule’s purpose, whatever it is, is furthered by the use of the proxy term “vehicle.”

Hart used the “no vehicles in the park” rule in an attempt to show that the word “vehicle” had a core of concrete and settled linguistic meaning (an automobile is a vehicle) as well as a semantic “penumbra” containing more or less debatable cases such as bicycles, toy cars, and airplanes. The rule, in other words, is textually ambiguous, although this does not necessarily mean that it is misspecified.[ref 11] Because the rule is ambiguous, a series of difficult interpretive decisions may have to be made regarding whether a given item is or is not a vehicle. At least some of these decisions, and the costs associated with them, could have been avoided if the rulemaker had chosen to use a more detailed formulation in lieu of the term “vehicle,”[ref 12] or if the rulemaker had issued a statement clarifying the purpose of the rule.[ref 13] 

Although the concept of misspecification is generally applicable to legal rules, misspecification tends to occur particularly frequently and with serious consequences in the context of laws and regulations governing poorly-understood emerging technologies such as artificial intelligence. Again, consider “no vehicles in the park.” Many legal rules, once established, persist indefinitely even as the technology they govern changes fundamentally.[ref 14] The objects to which the proxy term “vehicle” can be applied will change over time; electric wheelchairs, for example, may not have existed when the rule was originally drafted, and airborne drones may not have been common. The introduction of these new potential “vehicles” is extremely difficult to account for in an original design specification.[ref 15] 

The governance misspecification problem is particularly relevant to the governance of AI systems. Unlike most other emerging technologies, frontier AI systems are, in key respects, not only poorly understood but fundamentally uninterpretable by existing methods.[ref 16] This problem of interpretability is a major focus area for technical AI safety researchers.[ref 17] The widespread use of proxy terms and metrics in existing AI governance policies and proposals is, therefore, a cause for concern.[ref 18]

In Section I, this article draws on existing legal-philosophical discussions of the nature of rules to further explain the problem of governance misspecification and situates the concept in the existing public policy literature. Sections II and III make the case for the importance of the problem by presenting a series of case studies to show that rules aimed at governing emerging technologies are often misspecified and that misspecified rules can cause serious problems for the regulatory regime they contribute to, for courts, and for society generally. Section IV offers a few suggestions for reducing the risk of and mitigating the harm from misspecified rules, including eschewing or minimizing the use of proxy terms, rapidly updating and frequently reviewing the effectiveness of regulations, and including specific and clear statements of the purpose of a legal rule in the text of the rule. Section V applies the conclusions of the previous Sections prospectively to several specific challenges in the field of AI governance, including the use of compute thresholds, semiconductor export controls, and the problem of defining “frontier” AI systems. Section VI concludes.

I. The Governance Misspecification Problem in Legal Philosophy and Public Policy 

A number of publications in the field of legal philosophy have discussed the nature of legal rules and arrived at conclusions helpful to fleshing out the contours of the governance misspecification problem.[ref 19] Notably, Schauer (1991) suggests the useful concepts of over- and under-inclusiveness, which can be understood as two common ways in which legal rules can become misspecified.[ref 20] Overinclusive rules prohibit or prescribe actions that an ideally specified rule would not apply to, while underinclusive rules fail to prohibit or prescribe actions that an ideally specified rule would apply to. So, in Hart’s “no vehicles in the park” hypothetical, suppose that the sole purpose of the rule was to prevent park visitors from being sickened by diesel fumes. If this were the case, the rule would be overinclusive, because it would pointlessly prohibit many vehicles that do not emit diesel fumes. If, on the other hand, the purpose of the rule was to prevent music from being played loudly in the park on speakers, the rule would be underinclusive, as it fails to prohibit a wide range of speakers that are not installed in a vehicle. 

Ideal specification is rarely feasible, and practical considerations may dictate that a well-specified rule should rely on proxy terms that are under- or overinclusive to some extent. As Schauer (1991) explains, “Speed Limit 55” is a much easier rule to follow and enforce consistently than “drive safely,” despite the fact that the purpose of the speed limit is to promote safe driving and despite the fact that some safe driving can occur at speeds above 55 miles per hour and some dangerous driving can occur at speeds below 55 miles per hour.[ref 21] In other words, the benefits of creating a simple and easily followed and enforced rule outweigh the costs of over- and under-inclusiveness in many cases.[ref 22]

In the public policy literature, the existing concept that bears the closest similarity to governance misspecification is “policy design fit.”[ref 23] Policy design is currently understood as including a mix of interrelated policy goals and the instruments through which those goals are accomplished, including legal, financial, and communicative mechanisms.[ref 24] A close fit between policy goals and the means used to accomplish those goals has been shown to increase the effectiveness of policies.[ref 25] The governance misspecification problem can be understood as a particular species of failure of policy design fit—a failure of congruence between a policy goal and a proxy term in the legal rule which is the means used to further that goal.[ref 26] 

II. Legal Rules Governing Emerging Technologies Are Often Misspecified 

Misspecification occurs frequently in both domestic and international law and in both reactive and anticipatory regulations directed towards the regulation of new technologies. In order to illustrate how misspecification happens, and to give a sense of the significance of the problem in legal rules addressing emerging technologies, this Section discusses three historical examples of the phenomenon in the contexts of cyberlaw, copyright law, and nuclear anti-proliferation treaties.

Section 1201(a)(2) of the Digital Millennium Copyright Act of 1998 (DMCA) prohibits the distribution of any “technology, product, service, device, component, or part thereof” primarily designed to decrypt copyrighted material.[ref 27] Congressman Howard Coble, one of the architects of the DMCA, stated that this provision was “drafted carefully to target ‘black boxes’”—physical devices with “virtually no legitimate uses,” useful only for facilitating piracy.[ref 28] The use of “black boxes” for the decryption of digital works was not widespread in 1998, but the drafters of the DMCA predicted that such devices would soon become an issue. In 1998, this prediction seemed a safe bet, as previous forms of piracy decryption had relied on specialized tools—the phrase “black box” is a reference to one such tool, also known as a “descrambler” and used to decrypt premium cable television channels.[ref 29] 

However, the feared black boxes never arrived. Instead, pirates relied on software, using decryption programs distributed for free online to circumvent anti-piracy encryptions.[ref 30] Courts found the distribution of such programs, and even the posting of hyperlinks leading to websites containing such programs, to be violations of the DMCA.[ref 31] In light of earlier cases holding that computer code was a form of expression entitled to First Amendment protection, this interpretation placed the DMCA into tension with the First Amendment.[ref 32] This tension was ultimately resolved in favor of the DMCA, and the distribution of decryption programs used for piracy was prohibited.[ref 33]

No one in Congress anticipated that the statute which had been “carefully drafted to target ‘black boxes’” would be used to prohibit the distribution of lines of computer code, or that this would raise serious concerns regarding freedom of speech. Section 1201(a)(2), in other words, was misspecified; by prohibiting the distribution of any “technology” or “service” designed for piracy, as well as any “device,” the framers of the DMCA banned more than they intended to ban and created unforeseen constitutional issues. 

Misspecification also occurs in international law. The Treaty of Principles Governing the Activities of States in the Exploration and Use of Outer Space, which the United States and the Soviet Union entered into in 1967, obligated the parties “not to place in orbit around the Earth any objects carrying nuclear weapons…”[ref 34] Shortly after the treaty was entered into, however, it became clear that the Soviet Union planned to take advantage of a loophole in the misspecified prohibition. The Fractional Orbital Bombardment System (FOBS) placed missiles into orbital trajectories around the earth, but then redirected them to strike a target on the earth’s surface before they completed a full orbit.[ref 35] An object is not “in orbit” until it has circled the earth at least once; therefore, FOBS did not violate the 1967 Treaty, despite the fact that it allowed the Soviet Union to strike at the U.S. from space and thereby evade detection by the U.S.’s Ballistic Missile Early Warning System.[ref 36] The U.S. eventually neutralized this advantage by expanding the coverage and capabilities of early warning systems so that FOBS missiles could be detected and tracked, and in 1979 the Soviets agreed to a better-specified ban which prohibited “fractional orbital missiles” as well as other space-based weapons.[ref 37] Still, the U.S.’s agreement to use the underinclusive proxy term “in orbit” allowed the Soviet Union to temporarily gain a potentially significant first-strike advantage.

Misspecification occurs in laws and regulations directed towards existing and well-understood technologies as well as in anticipatory regulations. Take, for example, the Computer Fraud and Abuse Act (CFAA), 18 U.S.C. § 1030, which has been called “the worst law in technology.”[ref 38] The CFAA was originally enacted in 1984, but has since been amended several times, most recently in 2020.[ref 39] Among other provisions, the CFAA criminalizes “intentionally access[ing] a computer without authorization or exceed[ing] authorized access, and thereby obtain[ing]… information from any protected computer.”[ref 40] The currently operative language for this provision was introduced in 1996,[ref 41] by which point the computer was hardly an emerging technology, and slightly modified in 2008.[ref 42] 

Read literally, the CFAA’s prohibition on unauthorized access criminalizes both (a) violating a website’s terms of service while using the internet, and (b) using an employer’s computer or network for personal reasons, in violation of company policy.[ref 43] In other words, a literal reading of the CFAA would mean that hundreds of millions of Americans commit crimes every week by, e.g., sharing a password with a significant other or accessing social media at work.[ref 44] Court decisions eventually established narrower definitions of the key statutory terms (“without authorization” and “exceeds authorized access”),[ref 45] but not before multiple defendants were prosecuted for violating the CFAA by failing to comply a website’s terms of service[ref 46] or accessing an employer’s network for personal reasons in violation of workplace rules.[ref 47] 

Critics of the CFAA have discussed its flaws in terms of the constitutional law doctrines of “vagueness”[ref 48] and “overbreadth.”[ref 49] These flaws can also be conceptualized in terms of misspecification. The phrases “intentionally accesses without authorization” and “exceeds authorized access,” and the associated statutory definitions, are poor proxies for the range of behavior that an ideally specified version of the CFAA would have criminalized. The proxies criminalize a great deal of conduct that none of the stakeholders who drafted, advocated for, or voted to enact the law wanted to criminalize[ref 50] and created substantial legal and political backlash against the law. This backlash led to a series of losses for federal prosecutors as courts rejected their broad proposed interpretations of the key proxy terms because, as the Ninth Circuit Court of Appeals put it, “ubiquitous, seldom-prosecuted crimes invite arbitrary and discriminatory enforcement.”[ref 51] The issues caused by poorly selected proxy terms in the CFAA, the Outer Space Treaty, and the DMCA demonstrate that important legal rules drafted for the regulation of emerging technologies are prone to misspecification, in both domestic and international law contexts and for both anticipatory and reactive rules. These case studies were chosen because they are representative of how legal rules become misspecified; if space allowed, numerous additional examples of misspecified rules directed towards new technologies could be offered.[ref 52]

III. Consequences of Misspecification in the Regulation of Emerging Technologies

The case studies examined in the previous Section established that legal rules are often misspecified and illustrated the manner in which the problem of governance misspecification typically arises. This Section attempts to show that misspecification can cause serious issues when it occurs for both for the regulatory regime that the misspecified rule is part of and for society writ large. Three potential consequences of misspecification are discussed and illustrated with historical examples involving the regulation of emerging technologies.  

A. Underinclusive Rules Can Create Exploitable Gaps in a Regulatory Regime

When misspecification results in an underinclusive rule, exploitable gaps can arise in a regulatory regime. The Outer Space Treaty of 1967, discussed above, is one example of this phenomenon. Another example, which demonstrates how completely the use of a misspecified proxy term can defeat the effectiveness of a law, is the Audio Home Recording Act of 1992.[ref 53] That statute was designed to regulate home taping, i.e., the creation by consumers of analog or digital copies of musical recordings. The legal status of home taping had been a matter of debate for years, with record companies arguing that it was illegal and taping hardware manufacturers defending its legality.[ref 54] Congress attempted to resolve the debate by creating a safe harbor for home taping that allowed for the creation of any number of analog or digital copies of a piece of music, with the caveat that royalties would have to be paid as part of the purchase price of any equipment used to create digital copies.[ref 55] 

Congress designed the AHRA under the assumption that digital audio tape recorders (DATs) were the wave of the future and would shortly become a ubiquitous home audio appliance.[ref 56] The statute lays out, in painstaking detail, a complex regulatory framework governing “digital audio recording devices,” which the statute defines to require the capability to create reproductions of “digital musical recordings.”[ref 57] Bizarrely, however, the AHRA explicitly provides that the term “digital musical recording” does not encompass recordings stored on any object “in which one or more computer programs are fixed”—i.e., computer hard drives.[ref 58] 

Of course, the DAT did not become a staple of the American household. And when the RIAA sued the manufacturer of the “Rio,” an early mp3 player, for failing to comply with the AHRA’s requirements, the Ninth Circuit found that the device was not subject to the AHRA.[ref 59] Because the Rio was designed solely to download mp3 files from a computer hard drive, it was not capable of copying “digital musical recordings” under the AHRA’s underinclusive definition of that phrase.[ref 60] The court noted that its decision would “effectively eviscerate the Act,” because “[a]ny recording device could evade […] regulation simply by passing the music through a computer and ensuring that the MP3 file resided momentarily on the hard drive,” but nevertheless rejected the creative alternative interpretations suggested by the music industry as contrary to the plain language of the statute.[ref 61] As a result, the AHRA was rendered obsolete less than six years after being enacted.[ref 62] 

Clearly, Congress acted with insufficient epistemic humility by creating legislation confidently designed to address one specific technology that had not, at the time of legislation, been adopted by any significant portion of the population. But this failure of humility manifested as a failure of specification. The purpose of the statute, as articulated in a Senate report, included the introduction of a “serial copy management system that would prohibit the digital serial copying of copyrighted music.”[ref 63] By crafting a law that applied only to “digital audio recording devices” and defining that proxy term in an insufficiently flexible way, Congress completely failed to accomplish those purposes. If the proxy in question had not been defined to exclude any recording acquired through a computer, the Rio and eventually the iPod might well have fallen under the AHRA’s royalty scheme, and music copyright law in the U.S. might have developed down a course more consistent with the ideal specification of the AHRA. 

B. Overinclusive Rules Can Create Pushback and Enforcement Challenges

Misspecification can also create overinclusive rules, like the Computer Fraud and Abuse Act and § 1201(a)(2) of the Digital Millennium Copyright Act, discussed above in Section II. As those examples showed, overinclusive rules may give rise to legal and political challenges, difficulties with enforcement, and other unintended and undesirable results. These effects can, in some cases, be so severe that they require a total repeal of the rule in question.

This was the case with a 2011 Nevada statute authorizing and regulating driverless cars. AB511, which was the first law of its kind enacted in the U.S.,[ref 64] initially defined “autonomous vehicle” to mean “a motor vehicle that uses artificial intelligence, sensors and global positioning system coordinates to drive itself without the active intervention of a human operator,” and further defined “artificial intelligence” to mean “the use of computers and related equipment to enable a machine to duplicate or mimic the behavior of human beings.”[ref 65] 

Shortly after AB511 was enacted, however, several commentators noted that the statute’s definition of “autonomous vehicle” technically included vehicles that incorporated automatic collision avoidance or any of a number of other advanced driver-assistance systems common in new cars in 2011.[ref 66] These systems used computers to temporarily control the operation of a vehicle without the intervention of the human driver, so any vehicle that incorporated them was technically subject to the onerous regulatory scheme that Nevada’s legislature had intended to impose only on fully autonomous vehicles. In order to avoid effectively banning most new model cars, Nevada’s legislature was forced to repeal its new law and enact a replacement that incorporated a more detailed definition of “autonomous vehicle.”[ref 67]

C. Technological Change Can Repeatedly Render a Proxy Metric Obsolete

Finally, a misspecified rule may lose its effectiveness over time as technological advances render it obsolete, necessitating repeated updates and patches to the fraying regulatory regime. Consider, for example, the export controls imposed on high performance computers in the 1990s. The purpose of these controls was to prevent the export of powerful computers to countries where they might be used in ways that threatened U.S. national security, such as to design missiles and nuclear weapons.[ref 68] The government placed restrictions on the export of “supercomputers” and defined “supercomputer” in terms of the number of millions of theoretical operations per second (MTOPS) the computer could perform.[ref 69] In 1991, “supercomputer” was defined to mean any computer capable of exceeding 195 MTOPS.[ref 70] As the 90s progressed, however, the processing power of commercially available computers manufactured outside of the U.S. increased rapidly, reducing the effectiveness of U.S. export controls.[ref 71] Restrictions that prevented U.S. companies from selling their computers globally imposed costs on the U.S. economy and harmed the international competitiveness of the restricted companies.[ref 72] The Clinton administration responded by raising the threshold at which export restrictions began to apply to 1500 MTOPS in 1994, to 7000 MTOPS in 1996, to 12,300 MTOPS in 1999, and three times in the year 2000 to 20,000, 28,000, and finally 85,000 MTOPS.[ref 73] 

In the late 1990s, technological advances made it possible to link large numbers of commercially available computers together into “clusters” which could outperform most supercomputers.[ref 74] At this point, it was clear that MTOPS-based export controls were no longer effective, as computers that exceeded any limit imposed could easily be produced by anyone with access to a supply of less powerful computers which would not be subject to export controls.[ref 75] Even so, MTOPS-based export controls continued in force until 2006, when they were replaced by regulations that imposed controls based on performance in terms of Weighted TeraFLOPS, i.e., trillions of floating point operations per second.[ref 76] 

Thus, while the use of MTOPS thresholds as proxies initially resulted in well-specified export controls that effectively prevented U.S. adversaries from acquiring supercomputers, rapid technological progress repeatedly rendered the controls overinclusive and necessitated a series of amendments and revisions. The end result was a period of nearly seven years during which the existing export controls were badly misspecified due to the use of a proxy metric, MTOPS, which no longer bore any significant relation to the regime’s purpose. During this period, the U.S. export control regime for high performance computers was widely considered to be ineffective and perhaps even counterproductive.[ref 77]

IV. Mitigating Risks from Misspecification

In light of the frequency with which misspecification occurs in the regulation of emerging technology and the potential severity of its consequences, this Section suggests a few techniques for designing legal rules in such a way as to reduce the risk of misspecification and mitigate its ill effects.
The simplest way to avoid misspecification is to eschew or minimize the use of proxy terms and metrics. This is not always practicable or desirable. “No vehicles in the park” is a better rule than “do not unreasonably annoy or endanger the safety of park visitors,” in part because it reduces the cognitive burden of following, enforcing, and interpreting the rule and reduces the risk of decision maker error by limiting the discretion of the parties charged with enforcement and interpretation.[ref 78] Nevertheless, there are successful legal rules that pursue their purposes directly. U.S. antitrust law, for example, grew out of the Sherman Antitrust Act,[ref 79] § 1 of which simply states that any combination or contract in restraint of trade “is declared to be illegal.” 

Where use of a proxy is appropriate, it is often worthwhile to identify the fact that a proxy is being used to reduce the likelihood that decision makers will fall victim to Goodhart’s law[ref 80] and treat the regulation of the proxy as an end in itself.[ref 81] Alternatively, the most direct way to avoid confusion regarding the underlying purpose of a rule is to simply include an explanation of the purpose in the text of the rule itself. This can be accomplished through the addition of a purpose clause (sometimes referred to as a legislative preamble or a policy statement). For example, one purpose of the Nuclear Energy Innovation and Modernization Act of 2019 is to “provide… a program to develop the expertise and regulatory processes necessary to allow innovation and the commercialization of advanced nuclear reactors.” 

Purpose clauses can also incorporate language emphasizing that every provision of a rule should be construed in order to effectuate its purpose. This amounts to a legislatively prescribed rule of statutory interpretation, instructing courts to adopt a purposivist interpretive approach.[ref 82] When confronted with an explicit textual command to this effect, even strict textualists are obligated to interpret a rule purposively.[ref 83] The question of whether such an approach is generally desirable is hotly debated,[ref 84] but in the context of AI governance the flexibility that purposivism provides is a key advantage. The ability to flexibly update and adapt a rule in response to changes in the environment in which the rule will apply is unusually important in the regulation of emerging technologies.[ref 85] While there is little empirical evidence for or against the effectiveness of purpose clauses, they have played a key role in the legal reasoning relied on in a number of important court decisions.[ref 86]

A regulatory regime can also require periodic efforts to evaluate whether a rule is achieving its purpose.[ref 87] These efforts can provide an early warning system for misspecification by facilitating awareness of whether the proxy terms or metrics relied upon still correspond well to the purpose of the rule. Existing  periodic review requirements are often ineffective,[ref 88] treated by agencies as box-checking activities rather than genuine opportunities for careful retrospective analysis of the effects of regulations.[ref 89] However, many experts continue to recommend well-implemented retrospective review requirements as an effective tool for improving policy decisions.[ref 90] The Administrative Conference of the United States has repeatedly pushed for increased use of retrospective review, as has the internationally-focused Organization for Economic Co-Operation and Development (OECD).[ref 91] Additionally, retrospective review of regulations often works well in countries outside of the U.S.[ref 92] 

As the examples in Sections II and III demonstrate, rules governing technology tend to become misspecified over time as the regulated technology evolves. The Outer Space Treaty of 1967, § 1201(a)(2) of the DMCA, and the Clinton Administration’s supercomputer export controls were all well-specified and effective when implemented, but each measure became ineffective or counterproductive soon after being implemented because the proxies relied upon became obsolete. Ideally, rulemaking would move at the pace of technological improvement, but there are a number of institutional and structural barriers to this sort of rapid updating of regulations. Notably, the Administrative Procedure Act requires a lengthy “notice and comment” process for rulemaking and a 30-day waiting period after publication of a regulation in the Federal Register before the regulation can go into effect.[ref 93] There are ways to waive or avoid these requirements, including regulating via the issuance of nonbinding guidance documents rather than binding rules,[ref 94] issuing an immediately effective “interim final rule” and then satisfying the APA’s requirements at a later time,[ref 95] waiving the publication or notice and comment requirements for “good cause,”[ref 96] or legislatively imposing regulatory deadlines.[ref 97] Many of these workarounds are limited in their scope or effectiveness, or vulnerable to legal challenges if pursued too ambitiously, but finding some way to update a regulatory regime quickly is critical to mitigating the damage caused by misspecification.[ref 98] 

There is reason to believe that some agencies, recognizing the importance of AI safety to national security, will be willing to rapidly update regulations despite the legal and procedural difficulties. Consider the Commerce Department’s recent response to repeated attempts by semiconductor companies to design chips for the Chinese market that comply with U.S. export control regulations while still providing significant utility to purchasers in China looking to train advanced AI models. After Commerce initially imposed a license requirement on the export of advanced AI-relevant chips to China in October 2022, Nvidia modified its market-leading A100 and H100 chips to comply with the regulations and proceeded to sell the modified A800 and H800 chips in China.[ref 99] On October 17, 2023, the Commerce Department’s Bureau of Industry and Security announced a new interim final rule that would prohibit the sale of A800 and H800 chips in China and waived the normal 30-day waiting period so that the rule became effective less than a week after it was announced.[ref 100] Commerce Secretary Gina Raimondo stated publicly that “”[i]f [semiconductor companies] redesign a chip around a particular cut line that enables them to do AI, I’m going to control it the very next day.”[ref 101] 

V. The Governance Misspecification Problem and Artificial Intelligence 

While the framework of governance misspecification is applicable to a wide range of policy measures, it is particularly well-suited to describing issues that arise regarding legal rules governing emerging technologies. H.L.A. Hart’s prohibition on “vehicles in the park” could conceivably have been framed by an incautious drafter who did not anticipate that using “vehicle” instead of some more detailed proxy term would create ambiguity. Avoiding this kind of misspecification is simply a matter of careful drafting. Suppose, however, that the rule was formulated at a point in time when “vehicle” was an appropriate proxy for a well-understood category of object, and the rule later became misspecified as new potential vehicles that had not been conceived of when the rule was drafted were introduced. A rule drafted at a historical moment when all vehicles move on either land or water is unlikely to adequately account for the issues created by airplanes or flying drones.[ref 102] 

In other words, rules created to govern emerging technologies are especially prone to misspecification because they are created in the face of a high degree of uncertainty regarding the nature of the subject matter to be regulated, and rulemaking under uncertainty is difficult.[ref 103] Furthermore, as the case studies discussed in Sections II and III show, the nature of this difficulty is such that it tends to result in misspecification. For instance, misspecification will usually result when an overconfident rulemaker makes a specific and incorrect prediction about the future and issues an underinclusive rule based on that prediction. This was the case when Congress addressed  the AHRA exclusively to digital audio tape recorders and ignored computers. Rules created by rulemakers who want to regulate a certain technology but have only a vague and uncertain understanding of the purpose they are pursuing are also likely to be misspecified.[ref 104] Hence the CFAA, which essentially prohibited “doing bad things with a computer,” with disastrous results. 

The uncertainties associated with emerging technologies and the associated risk of  misspecification increase when the regulated technology is poorly understood. Rulemakers may simply overlook something about the chosen proxy due to a lack of understanding of the proxy or the underlying technology, or due to a lack of experience drafting the kinds of regulations required. The first-of-its-kind Nevada law intended to regulate fully autonomous vehicles that accidentally regulated a broad range of features common in many new cars is an example of this phenomenon. So is the DMCA provision that was intended to regulate “black box” devices but, by its terms, also applied to raw computer code.

If the difficulty of making well-specified rules to govern emerging technologies increases when the technology is fast-developing and poorly understood, advanced AI systems are something of a perfect storm for misspecification problems. Cutting-edge deep learning AI systems differ from other emerging technologies in that their workings are poorly understood, not just by legislators and the public, but by their creators.[ref 105] Their capabilities are an emergent property of the interaction between their architecture and the vast datasets on which they are trained. Moreover, the opacity of these models is arguably different in kind from the unsolved problems associated with past technological breakthroughs, because the models may be fundamentally uninterpretable rather than merely difficult to understand.[ref 106] Under these circumstances, defining an ideal specification in very general terms may be simple enough, but designing legal rules to operationalize any such specification will require extensive reliance on rough proxies. This is fertile ground for misspecification.

There are a few key proxy terms that recur often in existing AI governance proposals and regulations. For example, a number of policy proposals have suggested that regulations should focus on “frontier” AI models.[ref 107] When Google, Anthropic, OpenAI, and Microsoft created an industry-led initiative to promote AI safety, they named it the Frontier Model Forum.[ref 108] Sam Altman, the CEO of OpenAI, has expressed support for regulating “frontier systems.”[ref 109] The government of the U.K. has established a “Frontier AI Taskforce” dedicated to evaluating risks “at the frontier of AI.”[ref 110] 

In each of these proposals, the word “frontier” is a proxy term that stands for something like “highly capable foundation models that could possess dangerous capabilities sufficient to pose severe risks to public safety.”[ref 111] Any legislation or regulation that relied on the term “frontier” would also likely include a statutory definition of the word,[ref 112] but as several of the historical examples discussed in Sections II and III showed, statutory definitions can themselves incorporate proxies that result in misspecification. The above definition, for instance, may be underinclusive because some models that cannot be classified as “highly capable” or as “foundation models” might also pose severe risks to public safety.   

The most significant AI-related policy measure that has been issued in the U.S. to date is Executive Order (EO) 14110 on the “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence.”[ref 113] Among many other provisions, the EO imposes reporting requirements on certain AI models and directs the Department of Commerce to define the category of models to which the reporting requirements will apply.[ref 114] Prior to the issuance of Commerce’s definition, the EO provides that the reporting requirements apply to models “trained using a quantity of computing power greater than 10²⁶ integer or floating-point operations, or using primarily biological sequence data and using a quantity of computing power greater than 10²³ integer or floating-point operations,” as well as certain computing clusters.[ref 115] In other words, the EO uses operations as a proxy metric for determining which AI systems are sufficiently capable and/or dangerous that they should be regulated. This kind of metric, which is based on the amount of computing power used to train a model, is known as a “compute threshold” in the AI governance literature.[ref 116]

A proxy metric such as an operations-based compute threshold is almost certainly necessary to the operationalization of the EO’s regulatory scheme for governing frontier models.[ref 117] Even so, the example of the U.S. government’s ultimately ineffective and possibly counterproductive attempts to regulate exports of high performance computers using MTOPS is a cautionary tale about how quickly a compute-based proxy can be rendered obsolete by technological progress. The price of computing resources has, historically, fallen rapidly, with the amount of compute available for a given sum of money doubling approximately every two years as predicted by Moore’s Law.[ref 118] Additionally, because of improvements in algorithmic efficiency, the amount of compute required to train a model to a given level of performance has historically decreased over time as well.[ref 119] Because of these two factors, the cost of training AI models to a given level of capability has fallen precipitously over time; for instance, between 2017 and 2021, the cost of training a rudimentary model to classify images correctly with 93% accuracy on the image database ImageNet fell from $1000 to $5.[ref 120] This phenomenon presents a dilemma for regulators: the cost of acquiring computational resources exceeding a given threshold will generally decrease over time even as the capabilities of models trained on a below-threshold amount of compute rises. In other words, any well-specified legal rule that uses a compute threshold is likely to be rendered both overinclusive and underinclusive soon after being implemented.

Export controls intended to prevent the proliferation of the advanced chips used to train frontier AI models face a similar problem. Like the Clinton Administration’s supercomputer export controls, the Biden administration’s export controls on chips like the Nvidia A800 and H800 are likely to become misspecified over time. As algorithmic efficiency increases and powerful chips become cheaper and easier to acquire, existing semiconductor export controls will gradually become both overinclusive (because they pointlessly prohibit the export of chips that are already freely available overseas) and underinclusive (because powerful AI models can be trained using chips not covered by the export controls). 

The question of precisely how society should respond to these developments over time is beyond the scope of this paper. However, to delay the onset of misspecification and mitigate its effects, policymakers setting legal rules for AI governance should consider the recommendations outlined in Section IV, above. So, the specifications for export controls on semiconductors—proxies for something like “chips that can be used to create dangerously powerful AI models”—should be updated quickly and frequently as needed, to prevent them from becoming ineffective or counterproductive. The Bureau of Industry and Security has already shown some willingness to pursue this kind of frequent, flexible updating.[ref 121] More generally, given the particular salience of the governance misspecification problem to AI governance, legislators should consider mandating frequent review of the effectiveness of important AI regulations and empowering administrative agencies to update regulations rapidly as necessary. Rules setting compute thresholds that are likely to be the subject of litigation should incorporate clear purpose statements articulating the ulterior purpose behind the  use of a compute threshold as a proxy, and should be interpreted consistently with those statements. And where it is possible to eschew the use of proxies without compromising the enforceability or effectiveness of a rule, legislators and regulators should consider doing so. 

VI. Conclusion

This article has attempted to elucidate a newly developed concept in governance, i.e., the problem of governance misspecification. In presenting this concept along with empirical insights from representative case studies, we hope to inform contemporary debates around AI governance by demonstrating one common and impactful way in which legal rules can fail to effect their purposes. By framing this problem in terms of “misspecification,” a concept borrowed from the technical AI safety literature, this article aims both to introduce valuable insights from that field to scholars of legal philosophy and public policy and to introduce technical researchers to some of the more practically salient legal-philosophical and governance-related challenges involved in AI legislation and regulation. Additionally, we have offered a few specific suggestions for avoiding or mitigating the harms of misspecification in the AI governance context, namely eschewing the use of proxy terms or metrics where feasible, clear statements of statutory purpose, and flexibly applied, rapidly updating, periodically reviewed regulations. 

A great deal of conceptual and empirical work remains to be done regarding the nature and effects of the governance misspecification problem and best practices for avoiding and responding to it. For instance, this article does not contain any in-depth comparison of the incidence and seriousness of misspecification outside of the context of rules governing emerging technologies. Additionally, empirical research analyzing whether and how purpose clauses and similar provisions can effectively further the purposes of legal rules would be of significant practical value.

Legal considerations for defining “frontier model”

I. Introduction

One of the few concrete proposals on which AI governance stakeholders in industry[ref 1] and government[ref 2] have mostly[ref 3] been able to agree is that AI legislation and regulation should recognize a distinct category consisting of the most advanced AI systems. The executive branch of the U.S. federal government refers to these systems, in Executive Order 14110 and related regulations, as “dual-use foundation models.”[ref 4] The European Union’s AI Act refers to a similar class of models as “general-purpose AI models with systemic risk.”[ref 5] And many researchers, as well as leading AI labs and some legislators, use the term “frontier models” or some variation thereon.[ref 6] 

These phrases are not synonymous, but they are all attempts to address the same issue—namely that the most advanced AI systems present additional regulatory challenges distinct from those posed by less sophisticated models. Frontier models are expected to be highly capable across a broad variety of tasks and are also expected to have applications and capabilities that are not readily predictable prior to development, nor even immediately known or knowable after development.[ref 7] It is likely that not all of these applications will be socially desirable; some may even create significant risks for users or for the general public. 

The question of precisely how frontier models should be regulated is contentious and beyond the scope of this paper. But any law or regulation that distinguishes between “frontier models” (or “dual-use foundation models,” or “general-purpose AI models with systemic risk”) and other AI systems will first need to define the chosen term. A legal rule that applies to a certain category of product cannot be effectively enforced or complied with unless there is some way to determine whether a given product falls within the regulated category. Laws that fail to carefully define ambiguous technical terms often fail in their intended purposes, sometimes with disastrous results.[ref 8] Because the precise meaning of the phrase “frontier model” is not self-evident,[ref 9] the scope of a law or regulation that targeted frontier models without defining that term would be unacceptably uncertain. This uncertainty would impose unnecessary costs on regulated companies (who might overcomply out of an excess of caution or unintentionally undercomply and be punished for it) and on the public (from, e.g., decreased compliance, increased enforcement costs, less risk protection, and more litigation over the scope of the rule).

 The task of defining “frontier model” implicates both legal and policy considerations. This paper provides a brief overview of some of the most relevant legal considerations for the benefit of researchers, policymakers, and anyone else with an interest in the topic. 

II. Statutory and Regulatory Definitions

Two related types of legal definition—statutory and regulatory—are relevant to the task of defining “frontier model.” A statutory definition is a definition that appears in a statute enacted by a legislative body such as the U.S. Congress or one of the 50 state legislatures. A regulatory definition, on the other hand, appears in a regulation promulgated by a government agency such as the U.S. Department of Commerce or the California Department of Technology (or, less commonly, in an executive order).

Regulatory definitions have both advantages and disadvantages relative to statutory definitions. Legislation is generally a more difficult and resource-intensive process than agency rulemaking, with additional veto points and failure modes.[ref 10] Agencies are therefore capable of putting into effect more numerous and detailed legal rules than Congress can,[ref 11] and can update those rules more quickly and easily than Congress can amend laws.[ref 12] Additionally, executive agencies are often more capable of acquiring deep subject-matter expertise in highly specific fields than are congressional offices due to Congress’s varied responsibilities and resource constraints.[ref 13] This means that regulatory definitions can benefit from agency subject-matter expertise to a greater extent than can statutory definitions, and can also be updated far more easily and often.

The immense procedural and political costs associated with enacting a statute do, however, purchase a greater degree of democratic legitimacy and legal resiliency than a comparable regulation would enjoy. A number of legal challenges that might persuade a court to invalidate a regulatory definition would not be available for the purpose of challenging a statute.[ref 14] And since the rulemaking power exercised by regulatory agencies is generally delegated to them by Congress, most regulations must be authorized by an existing statute. A regulatory definition generally cannot eliminate or override a statutory definition[ref 15] but can clarify or interpret. Often, a regulatory regime will include both a statutory definition and a more detailed regulatory definition for the same term.[ref 16] This can allow Congress to choose the best of both worlds, establishing a threshold definition with the legitimacy and clarity of an act of Congress while empowering an agency to issue and subsequently update a more specific and technically informed regulatory definition. 

III. Existing Definitions

This section discusses five noteworthy attempts to define phrases analogous to “frontier model” from three different existing measures. Executive Order 14110 (“EO 14110”), which President Biden issued in October 2023, includes two complementary definitions of the term “dual-use foundation model.” Two definitions of “covered model” from different versions of the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, a California bill that was recently vetoed by Governor Newsom, are also discussed, along with the EU AI Act’s definition of “general-purpose AI model with systemic risk.”

A. Executive Order 14110

EO 14110 defines “dual-use foundation model” as:

an AI model that is trained on broad data; generally uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts; and that exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters, such as by:

(i) substantially lowering the barrier of entry for non-experts to design, synthesize, acquire, or use chemical, biological, radiological, or nuclear (CBRN) weapons;

(ii) enabling powerful offensive cyber operations through automated vulnerability discovery and exploitation against a wide range of potential targets of cyber attacks; or

(iii) permitting the evasion of human control or oversight through means of deception or obfuscation.

Models meet this definition even if they are provided to end users with technical safeguards that attempt to prevent users from taking advantage of the relevant unsafe capabilities.[ref 17]

The executive order imposes certain reporting requirements on companies “developing or demonstrating an intent to develop” dual-use foundation models,[ref 18] and for purposes of these requirements it instructs the Department of Commerce to “define, and thereafter update as needed on a regular basis, the set of technical conditions for models and computing clusters that would be subject to the reporting requirements.”[ref 19] In other words, EO 14110 contains both a high-level quasi-statutory[ref 20] definition and a directive to an agency to promulgate a more detailed regulatory definition. The EO also provides a second definition that acts as a placeholder until the agency’s regulatory definition is promulgated:

any model that was trained using a quantity of computing power greater than 10²⁶ integer or floating-point operations, or using primarily biological sequence data and using a quantity of computing power greater than 10²³ integer or floating-point operations[ref 21]

Unlike the first definition, which relies on subjective evaluations of model characteristics,[ref 22] this placeholder definition provides a simple set of objective technical criteria that labs can consult to determine whether the reporting requirements apply. For general-purpose models, the sole test is whether the model was trained on computing power greater than 10²⁶ integer or floating-point operations (FLOP); only models that exceed this compute threshold[ref 23] are deemed “dual-use foundation models” for purposes of the reporting requirements mandated by EO 14110.

B. California’s “Safe and Secure Innovation for Frontier Artificial Intelligence Act” (SB 1047)

California’s recently vetoed “Safe and Secure Innovation for Frontier Artificial Intelligence Models Act” (“SB 1047”) focused on a category that it referred to as “covered models.”[ref 24] The version of SB 1047 passed by the California Senate in May 2024 defined “covered model” to include models meeting either of the following criteria:

(1) The artificial intelligence model was trained using a quantity of computing power greater than 10²⁶ integer or floating-point operations.

(2) The artificial intelligence model was trained using a quantity of computing power sufficiently large that it could reasonably be expected to have similar or greater performance as an artificial intelligence model trained using a quantity of computing power greater than 10²⁶ integer or floating-point operations in 2024 as assessed using benchmarks commonly used to quantify the general performance of state-of-the-art foundation models.[ref 25]

This definition resembles the placeholder definition in EO 14110 in that it primarily consists of a training compute threshold of 10²⁶ FLOP. However, SB 1047 added an alternative capabilities-based threshold to capture future models which “could reasonably be expected” to be as capable as models trained on 10²⁶ FLOP in 2024. This addition was intended to “future-proof”[ref 26] SB 1047 by addressing one of the main disadvantages of training compute thresholds—their tendency to become obsolete over time as advances in algorithmic efficiency produce highly capable models trained on relatively small amounts of compute.[ref 27] 

Following pushback from stakeholders who argued that SB 1047 would stifle innovation,[ref 28] the bill was amended repeatedly in the California State Assembly. The final version defined “covered model” in the following way:

(A) Before January 1, 2027, “covered model” means either of the following:

(i) An artificial intelligence model trained using a quantity of computing power greater than 10²⁶ integer or floating-point operations, the cost of which exceeds one hundred million dollars[ref 29] ($100,000,000) when calculated using the average market prices of cloud compute at the start of training as reasonably assessed by the developer.

(ii) An artificial intelligence model created by fine-tuning a covered model using a quantity of computing power equal to or greater than three times 10²⁵ integer or floating-point operations, the cost of which, as reasonably assessed by the developer, exceeds ten million dollars ($10,000,000) if calculated using the average market price of cloud compute at the start of fine-tuning.

(B) (i) Except as provided in clause (ii), on and after January 1, 2027, “covered model” means any of the following:

(I) An artificial intelligence model trained using a quantity of computing power determined by the Government Operations Agency pursuant to Section 11547.6 of the Government Code, the cost of which exceeds one hundred million dollars ($100,000,000) when calculated using the average market price of cloud compute at the start of training as reasonably assessed by the developer.

(II) An artificial intelligence model created by fine-tuning a covered model using a quantity of computing power that exceeds a threshold determined by the Government Operations Agency, the cost of which, as reasonably assessed by the developer, exceeds ten million dollars ($10,000,000) if calculated using the average market price of cloud compute at the start of fine-tuning.

(ii) If the Government Operations Agency does not adopt a regulation governing subclauses (I) and (II) of clause (i) before January 1, 2027, the definition of “covered model” in subparagraph (A) shall be operative until the regulation is adopted.

This new definition was more complex than its predecessor. Subsection (A) introduced an initial definition slated to apply until at least 2027, which relied on a training compute threshold of 10²⁶ FLOP paired with a training cost floor of $100,000,000.[ref 30] Subsection (B), in turn, provided for the eventual replacement of the training compute thresholds used in the initial definition with new thresholds to be determined (and presumably updated) by a regulatory agency. 

The most significant change in the final version of SB 1047’s definition was the replacement of the capability threshold with a $100,000,000 cost threshold. Because it would currently cost more than $100,000,000 to train a model using >10²⁶ FLOP, the addition of the cost threshold did not change the scope of the definition in the short term. However, the cost of compute has historically fallen precipitously over time in accordance with Moore’s law.[ref 31] This may mean that models trained using significantly more than 10²⁶ FLOP will cost significantly less than the inflation-adjusted equivalent of 100 million 2024 dollars to create at some point in the future. 

The old capability threshold expanded the definition of “covered model” because it was an alternative to the compute threshold—models that exceeded either of the two thresholds would have been “covered.” The newer cost threshold, on the other hand, restricted the scope of the definition because it was linked conjunctively to the compute threshold, meaning that only models that exceed both thresholds were covered. In other words, where the May 2024 definition of “covered model” future-proofed itself against the risk of becoming underinclusive by including highly capable low-compute models, the final definition instead guarded against the risk of becoming overinclusive by excluding low-cost models trained on large amounts of compute. Furthermore, the final cost threshold was baked into the bill text and could only have been changed by passing a new statute—unlike the compute threshold, which could have been specified and updated by a regulator. 

Compared with the overall definitional scheme in EO 14110, SB 1047’s definition was simpler, easier to operationalize, and less flexible. SB 1047 lacked a broad, high-level risk-based definition like the first definition in EO 14110. SB 1047 did resemble EO 14110 in its use of a “placeholder” definition, but where EO 14110 confers broad discretion on the regulator to choose the “set of technical conditions” that will comprise the regulatory definition, SB 1047 only authorized the regulator to set and adjust the numerical value of the compute thresholds in an otherwise rigid statutory definition.

C. EU Artificial Intelligence Act

The EU AI Act classifies AI systems according to the risks they pose. It prohibits systems that do certain things, such as exploiting the vulnerabilities of elderly or disabled people,[ref 32] and regulates but does not ban so-called “high-risk” systems.[ref 33] While this classification system does not map neatly onto U.S. regulatory efforts, the EU AI Act does include a category conceptually similar to the EO’s “dual-use foundation model”: the “general-purpose AI model with systemic risk.”[ref 34] The statutory definition for this category includes a given general-purpose model[ref 35] if:

a. it ‍has high impact capabilities[ref 36] evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks; [or]

b. based on a decision of the Commission,[ref 37] ex officio or following a qualified alert from the scientific panel, it has capabilities or an impact equivalent to those set out in point (a) having regard to the criteria set out in Annex XIII.

Additionally, models are presumed to have “high impact capabilities” if they were trained on >10²⁵ FLOP.[ref 38] The seven “criteria set out in Annex XIII” to be considered in evaluating model capabilities include a variety of technical inputs (such as the model’s number of parameters and the size or quality of the dataset used in training the model), the model’s performance on benchmarks and other capabilities evaluations, and other considerations such as the number of users the model has.[ref 39] When necessary, the European Commission is authorized to amend the compute threshold and “supplement benchmarks and indicators” in response to technological developments, such as “algorithmic improvements or increased hardware efficiency.”[ref 40]

The EU Act definition resembles the initial, broad definition in the EO in that they both take diverse factors like the size and quality of the dataset used to train the model, the number of parameters, and the model’s capabilities into account. However, the EU Act definition is likely much broader than either EO definition. The training compute threshold in the EU Act is sufficient, but not necessary, to classify models as systemically risky, whereas the (much higher) threshold in the EO’s placeholder definition is both necessary and sufficient. And the first EO definition includes only models that exhibit a high level of performance on tasks that pose serious risks to national security, while the  EU Act includes all general-purpose models with “high impact capabilities,” which it defines as including any model trained on more than 10²⁵ FLOP.

The EU Act definition resembles the final SB 1047 definition of “covered model” in that both definitions authorize a regulator to update their thresholds in response to changing circumstances. It also resembles SB 1047’s May 2024 definition in that both definitions incorporate a training compute threshold and a capabilities-based element.

IV. Elements of Existing Definitions

As the examples discussed above demonstrate, legal definitions of “frontier model” can consist of one or more of a number of criteria. This section discusses a few of the most promising definitional elements.

A. Technical inputs and characteristics

A definition may classify AI models according to their technical characteristics or the technical inputs used in training the model, such as training compute, parameter count, and dataset size and type. These elements can be used in either statutory or regulatory definitions.

Training compute thresholds are a particularly attractive option for policymakers,[ref 41] as evidenced by the three examples discussed above. “Training compute” refers to the computational power used to train a model, often  measured in integer or floating-point operations (OP or FLOP).[ref 42] Training compute thresholds function as a useful proxy for model capabilities because capabilities tend to increase as computational resources used to train the model increase.[ref 43] 

One advantage of using a compute threshold is that training compute is a straightforward metric that is quantifiable and can be readily measured, monitored, and verified.[ref 44] Because of these characteristics, determining with high certainty whether a given model exceeds a compute threshold is relatively easy. This, in turn, facilitates enforcement of and compliance with regulations that rely on a compute-based definition. Since the amount of training compute (and other technical inputs) can be estimated prior to the training run,[ref 45] developers can predict whether a model will be covered earlier in development. 

One disadvantage of a compute-based definition is that compute thresholds are a proxy for model capabilities, which are in turn a proxy for risk. Definitions that make use of multiple nested layers of proxy terms in this manner are particularly prone to becoming untethered from their original purpose.[ref 46] This can be caused, for example, by the operation of Goodhart’s Law, which suggests that “when a measure becomes a target, it ceases to be a good measure.”[ref 47] Particularly problematic, especially for statutory definitions that are more difficult to update, is the possibility that a compute threshold may become underinclusive over time as improvements in algorithmic efficiency allow for the development of highly capable models trained on below-threshold levels of compute.[ref 48] This possibility is one reason why SB 1047 and the EU AI Act both supplement their compute thresholds with alternative, capabilities-based elements.

In addition to training compute, two other model characteristics correlated with capabilities are the number of model parameters[ref 49] and the size of the dataset on which the model was trained.[ref 50] Either or both of these characteristics can be used as an element of a definition. A definition can also rely on training data characteristics other than size, such as the quality or type of the data used; the placeholder definition in EO 14110, for example, contains a lower compute threshold for models “trained… using primarily biological sequence data.”[ref 51] EO 14110 requires a dual-use foundation model to contain “at least tens of billions of parameters,”[ref 52] and the “number of parameters of the model” is a criteria to be considered under the EU AI Act.[ref 53] EO 14110 specified that only models “trained on broad data” could be dual-use foundation models,[ref 54] and the EU AI Act includes “the quality or size of the data set, for example measured through tokens” as one criterion for determining whether an AI model poses systemic risks.[ref 55]

Dataset size and parameter count share many of the pros and cons of training compute. Like training compute, they are objective metrics that can be measured and verified, and they serve as proxies for model capabilities.[ref 56] Training compute is often considered the best and most reliable proxy of the three, in part because it is the most closely correlated with performance and is difficult to manipulate.[ref 57] However, partially redundant backup metrics can still be useful.[ref 58] Dataset characteristics other than size are typically less quantifiable and harder to measure but are also capable of capturing information that the quantifiable metrics cannot. 

B. Capabilities 

Frontier models can also be defined in terms of their capabilities. A capabilities-based definition element typically sets a threshold level of competence that a model must achieve to be considered “frontier,” either in one or more specific domains or across a broad range of domains. A capabilities-based definition can provide specific, objective criteria for measuring a model’s capabilities,[ref 59] or it can describe the capabilities required in more general terms and leave the task of evaluation to the discretion of future interpreters.[ref 60] The former approach might be better suited to a regulatory definition, especially if the criteria used will have to be updated frequently, whereas the latter approach would be more typical of a high-level statutory definition.

Basing a definition on capabilities, rather than relying on a proxy for capabilities like training compute, eliminates the risk that the chosen proxy will cease to be a good measure of capabilities over time. Therefore, a capabilities-based definition is more likely than, e.g., a compute threshold to remain robust over time in the face of improvements in algorithmic efficiency. This was the point of the May 2024 version of SB 1047’s use of a capabilities element tethered to a compute threshold (“similar or greater performance as an artificial intelligence model trained using a quantity of computing power greater than 10²⁶ integer or floating-point operations in 2024”)—it was an attempt to capture some of the benefits of an input-based definition while also guarding against the possibility that models trained on less than 10²⁶ FLOP may become far more capable in the future than they are in 2024. 

However,  capabilities are far more difficult than compute to accurately measure. Whether a model has demonstrated “high levels of performance at tasks that pose a serious risk to security” under the EO’s broad capabilities-based definition is not something that can be determined objectively and to a high degree of certainty like the size of a dataset in tokens or the total FLOP used in a training run. Model capabilities are often measured using benchmarks (standardized sets of tasks or questions),[ref 61] but creating benchmarks that accurately measure the complex and diverse capabilities of general-purpose foundation models[ref 62] is notoriously difficult.[ref 63] 

Additionally, model capabilities (unlike the technical inputs discussed above) are generally not measurable until after the model has been trained.[ref 64] This makes it difficult to regulate the development of frontier models using capabilities-based definitions, although post-development, pre-release regulation is still possible.

C. Risk

Some researchers have suggested the possibility of defining frontier AI systems on the basis of the risks they pose to users or to public safety instead of or in addition to relying on a proxy metric, like capabilities, or a proxy for a proxy, such as compute.[ref 65] The principal advantage of this direct approach is that it can, in theory, allow for better-targeted regulations—for instance, by allowing a definition to exclude highly capable but demonstrably low-risk models. The principal disadvantage is that measuring risk is even more difficult than measuring capabilities.[ref 66] The science of designing rigorous safety evaluations for foundation models is still in its infancy.[ref 67] 

Of the three real-world measures discussed in Section III, only EO 14110 mentions risk directly. The broad initial definition of “dual-use foundation model” includes models that exhibit “high levels of performance at tasks that pose a serious risk to security,” such as “enabling powerful offensive cyber operations through automated vulnerability discovery” or making it easier for non-experts to design chemical weapons. This is a capability threshold combined with a risk threshold; the tasks at which a dual-use foundation model must be highly capable are those that pose a “serious risk” to security, national economic security, and/or national public health or safety. As EO 14110 shows, risk-based definition elements can specify the type of risk that a frontier model must create instead of addressing the severity of the risks created. 

D. Epistemic elements

One of the primary justifications for recognizing a category of “frontier models” is the likelihood that broadly capable AI models that are more advanced than previous generations of models will have capabilities and applications that are not readily predictable ex ante.[ref 68] As the word “frontier” implies, lawmakers and regulators focusing on frontier models are interested in targeting models that break new ground and push into the unknown.[ref 69] This was, at least in part, the reason for the inclusion of training compute thresholds of 10²⁶ FLOP in EO 14110 and SB 1047—since the most capable current models were trained on 5×10²⁵ or fewer FLOP,[ref 70] a model trained on 10²⁶ FLOP would represent a significant step forward into uncharted territory. 

While it is possible to target models that advance the state of the art by setting and adjusting capability or compute thresholds, a more direct alternative approach would be to include an epistemic element in a statutory definition of “frontier model.” An epistemic element would distinguish between “known” and “unknown” models, i.e., between well-understood models that pose only known risks and poorly understood models that may pose unfamiliar and unpredictable risks.[ref 71] 

This kind of distinction between known and unknown risks has a long history in U.S. regulation.[ref 72] For instance, the Toxic Substances Control Act (TSCA) prohibits the manufacturing of any “new chemical substance” without a license.[ref 73] The EPA keeps and regularly updates a list of chemical substances which are or have been manufactured in the U.S., and any substance not included on this list is “new” by definition.[ref 74] In other words, the TSCA distinguishes between chemicals (including potentially dangerous chemicals) that are familiar to regulators and unfamiliar chemicals that pose unknown risks. 

One advantage of an epistemic element is that it allows a regulator to address “unknown unknowns” separately from better-understood risks that can be evaluated and mitigated more precisely.[ref 75] Additionally, the scope of an epistemic definition, unlike that of most input- and capability-based definitions, would change over time as regulators became familiar with the capabilities of and risks posed by new models.[ref 76] Models would drop out of the “frontier” category once regulators became sufficiently familiar with their capabilities and risks.[ref 77] Like a capabilities- or risk-based definition, however, an epistemic definition might be difficult to operationalize.[ref 78] To determine whether a given model was “frontier” under an epistemic definition, it would probably be necessary to either rely on a proxy for unknown capabilities or authorize a regulator to categorize eligible models according to a specified process.[ref 79] 

E. Deployment context

The context in which an AI system is deployed can serve as an element in a definition. The EU AI Act, for example, takes the number of registered end users and the number of registered EU business users a model has into account as factors to be considered in determining whether a model is a “general-purpose AI model with systemic risk.”[ref 80] Deployment context typically does not in and of itself provide enough information about the risks posed by a model to function as a stand-alone definitional element, but it can be a useful proxy for the kind of risk posed by a given model. Some models may cause harms in proportion to their number of users, and the justification for aggressively regulating these models grows stronger the more users they have.  A model that will only be used by government agencies, or by the military, creates a different set of risks than a model that is made available to the general public.

V. Updating Regulatory Definitions

A recurring theme in the scholarly literature on the regulation of emerging technologies is the importance of regulatory flexibility.[ref 81] Because of the rapid pace of technological progress, legal rules designed to govern emerging technologies like AI tend to quickly become outdated and ineffective if they cannot be rapidly and frequently updated in response to changing circumstances.[ref 82] For this reason, it may be desirable to authorize an executive agency to promulgate and update a regulatory definition of “frontier model,” since regulatory definitions can typically be updated more frequently and more easily than statutory definitions under U.S. law.[ref 83]

Historically, failing to quickly update regulatory definitions in the context of emerging technologies has often led to the definitions becoming obsolete or counterproductive. For example, U.S. export controls on supercomputers in the 1990s and early 2000s defined “supercomputer” in terms of the number of millions of theoretical operations per second (MTOPS) the computer could perform.[ref 84] Rapid advances in the processing power of commercially available computers soon rendered the initial definition obsolete, however, and the Clinton administration was forced to revise the MTOPS threshold repeatedly to avoid harming the competitiveness of the American computer industry.[ref 85] Eventually, the MTOPS metric itself was rendered obsolete, leading to a period of several years in which supercomputer export controls were ineffective at best.[ref 86]

There are a number of legal considerations that may prevent an agency from quickly updating a regulatory definition and a number of measures that can be taken to streamline the process. One important aspect of the rulemaking process is the Administrative Procedure Act’s “notice and comment” requirement.[ref 87] In order to satisfy this requirement, agencies are generally obligated to publish notice of any proposed amendment to an existing regulation in the Federal Register, allow time for the public to comment on the proposal, respond to public comments, publish a final version of the new rule, and then allow at least 30–60 days before the rule goes into effect.[ref 88] From the beginning of the notice-and-comment process to the publication of a final rule, this process can take anywhere from several months to several years.[ref 89] However, an agency can waive the 30–60 day publication period or even the entire notice-and-comment requirement for “good cause” if observing the standard procedures would be “impracticable, unnecessary, or contrary to the public interest.”[ref 90] Of course, the notice-and-comment process has benefits as well as costs; public input can be substantively valuable and informative for agencies, and also increases the democratic accountability of agencies and the transparency of the rulemaking process. In certain circumstances, however, the costs of delay can outweigh the benefits. U.S. agencies have occasionally demonstrated a willingness to waive procedural rulemaking requirements in order to respond to emergency AI-related developments. The Bureau of Industry and Security (“BIS”), for example, waived the normal 30-day waiting period for an interim rule prohibiting the sale of certain advanced AI-relevant chips to China in October 2023.[ref 91]  

Another way to encourage quick updating for regulatory definitions is for Congress to statutorily authorize agencies to eschew or limit the length of notice and comment, or to compel agencies to promulgate a final rule by a specified deadline.[ref 92] Because notice and comment is a statutory requirement, it can be adjusted as necessary by statute.  

For regulations exceeding a certain threshold of economic significance, another substantial source of delay is OIRA review. OIRA, the Office of Information and Regulatory Affairs, is an office within the White House that oversees interagency coordination and undertakes centralized cost-benefit analysis of important regulations.[ref 93] Like notice and comment, OIRA review can have significant benefits—such as improving the quality of regulations and facilitating interagency cooperation—but it also delays the implementation of significant rules, typically by several months.[ref 94] OIRA review can be waived either by statutory mandate or by OIRA itself.[ref 95]

VI. Deference, Delegation, and Regulatory Definitions

Recent developments in U.S. administrative law may make it more difficult for Congress to effectively delegate the task of defining “frontier model” to a regulatory agency. A number of recent Supreme Court cases signal an ongoing shift in U.S. administrative law doctrine intended to limit congressional delegations of rulemaking authority.[ref 96] Whether this development is good or bad on net is a matter of perspective; libertarian-minded observers who believe that the U.S. has too many legal rules already[ref 97] and that overregulation is a bigger problem than underregulation have welcomed the change,[ref 98] while pro-regulation observers predict that it will significantly reduce the regulatory capacity of agencies in a number of important areas.[ref 99] 

Regardless of where one falls on that spectrum of opinion, the relevant takeaway for efforts to define “frontier model” is that it will likely become somewhat more difficult for agencies to promulgate and update regulatory definitions without a clear statutory authorization to do so. If Congress still wishes to authorize the creation of regulatory definitions, however, it can protect agency definitions from legal challenges by clearly and explicitly authorizing agencies to exercise discretion in promulgating and updating definitions of specific terms.

A. Loper Bright and deference to agency interpretations

In a recent decision in the combined cases of Loper Bright Enterprises v. Raimondo and Relentless v. Department of Commerce, the Supreme Court repealed a longstanding legal doctrine known as Chevron deference.[ref 100] Under Chevron, federal courts were required to defer to certain agency interpretations of federal statutes when (1) the relevant part of the statute being interpreted was genuinely ambiguous and (2) the agency’s interpretation was reasonable. After Loper Bright, courts are no longer required to defer to these interpretations—instead, under a doctrine known as Skidmore deference,[ref 101] agency interpretations will prevail in court only to the extent that courts are persuaded by them.[ref 102] 

Justice Elena Kagan’s dissenting opinion in Loper Bright argues that the decision will harm the regulatory capacity of agencies by reducing the ability of agency subject-matter experts to promulgate regulatory definitions of ambiguous statutory phrases in “scientific or technical” areas.[ref 103] The dissent specifically warns that, after Loper Bright, courts will “play a commanding role” in resolving questions like “[w]hat rules are going to constrain the development of A.I.?”[ref 104] 

Justice Kagan’s dissent probably somewhat overstates the significance of Loper Bright to AI governance for rhetorical effect.[ref 105] The end of Chevron deference does not mean that Congress has completely lost the ability to authorize regulatory definitions; where Congress has explicitly directed an agency to define a specific statutory term, Loper Bright will not prevent the agency from doing so.[ref 106] An agency’s authority to promulgate a regulatory definition under a statute resembling EO 14110, which explicitly directs the Department of Commerce to define “dual-use foundation model,” would likely be unaffected. However, Loper Bright has created a great deal of uncertainty regarding the extent to which courts will accept agency claims that Congress has implicitly authorized the creation of regulatory definitions.[ref 107] 

To better understand how this uncertainty might affect efforts to define “frontier model,” consider the following real-life example. The Energy Policy and Conservation Act (“EPCA”) includes a statutory definition of the term “small electric motor.”[ref 108] Like many statutory definitions, however, this definition is not detailed enough to resolve all disputes about whether a given product is or is not a “small electric motor” for purposes of EPCA. In 2010, the Department of Energy (“DOE”), which is authorized under EPCA to promulgate energy efficiency standards governing “small electric motors,”[ref 109] issued a regulatory definition of “small electric motor” specifying that the term referred to motors with power outputs between 0.25 and 3 horsepower.[ref 110] The National Electrical Manufacturers Association (“NEMA”), a trade association of electronics manufacturers, sued to challenge the rule, arguing that motors with between 1 and 3 horsepower were too powerful to be “small electric motors” and that the DOE was exceeding its statutory authority by attempting to regulate them.[ref 111] 

In a 2011 opinion that utilized the Chevron framework, the federal court that decided NEMA’s lawsuit considered the language of EPCA’s statutory definition and concluded that EPCA was ambiguous as to whether motors with between 1 and 3 horsepower could be “small electric motors.”[ref 112] The court then found that the DOE’s regulatory definition was a reasonable interpretation of EPCA’s statutory definition, deferred to the DOE under Chevron, and upheld the challenged regulation.[ref 113]

Under Chevron, federal courts were required to assume that Congress had implicitly authorized agencies like the DOE to resolve ambiguities in a statute, as the DOE did in 2010 by promulgating its regulatory definition of “small electric motor.” After Loper Bright, courts will recognize fewer implicit delegations of definition-making authority. For instance, while EPCA requires the DOE to prescribe “testing requirements” and “energy conservation standards” for small electric motors, it does not explicitly authorize the DOE to promulgate a regulatory definition of “small electric motor.” If a rule like the one challenged by NEMA were challenged today, the DOE could still argue that Congress implicitly authorized the creation of such a rule by giving the DOE authority to prescribe standards and testing requirements—but such an argument would probably be less likely to succeed than the Chevron argument that saved the rule in 2011.

Today, a court that did not find an implicit delegation of rulemaking authority in EPCA would not defer to the DOE’s interpretation. Instead, the court would simply compare the DOE’s regulatory definition of “small electric motor” with NEMA’s proposed definition and decide which of the two was a more faithful interpretation of EPCA’s statutory definition.[ref 114] Similarly, when or if some future federal statute uses the phrase “frontier model” or any analogous term, agency attempts to operationalize the statute by enacting detailed regulatory definitions that are not explicitly authorized by the statute will be easier to challenge after Loper Bright than they would have been under Chevron. 

Congress can avoid Loper Bright issues by using clear and explicit statutory language to authorize agencies to promulgate and update regulatory definitions of “frontier model” or analogous phrases. However, it is often difficult to predict in advance whether or how a statutory definition will become ambiguous over time. This is especially true in the context of emerging technologies like AI, where the rapid pace of technological development and the poorly understood nature of the technology often eventually render carefully crafted definitions obsolete.[ref 115] 

Suppose, for example, that a federal statute resembling the May 2024 draft of SB 1047 was enacted. The statutory definition would include future models trained on a quantity of compute such that they “could reasonably be expected to have similar or greater performance as an artificial intelligence model trained using [>10²⁶ FLOP] in 2024.” If the statute did not contain an explicit authorization for some agency to determine the quantity of compute that qualified in a given year, any attempt to set and enforce updated regulatory compute thresholds could be challenged in court. 

The enforcing agency could argue that the statute included an implied authorization for the agency to promulgate and update the regulatory definitions at issue. This argument might succeed or fail, depending on the language of the statute, the nature of the challenged regulatory definitions, and the judicial philosophy of the deciding court. But regardless of the outcome of any individual case, challenges to impliedly authorized regulatory definitions will probably be more likely to succeed after Loper Bright than they would have been under Chevron. Perhaps more importantly, agencies will be aware that regulatory definitions will no longer receive the benefit of Chevron deference and may regulate more cautiously in order to avoid being sued.[ref 116] Moreover, even if the statute did explicitly authorize an agency to issue updated compute thresholds, such an authorization might not allow the agency to respond to future technological breakthroughs by considering some factor other than the quantity of training compute used.

In other words, a narrow congressional authorization to regulatorily define “frontier model” may prove insufficiently flexible after Loper Bright. Congress could attempt to address this possibility by instead enacting a very broad authorization.[ref 117] An overly broad definition, however, may be undesirable for reasons of democratic accountability, as it would give unelected agency officials discretionary control over which models to regulate as “frontier.” Moreover, an overly broad definition might risk running afoul of two related constitutional doctrines that limit the ability of Congress to delegate rulemaking authority to agencies—the major questions doctrine and the nondelegation doctrine.

B. The nondelegation doctrine

Under the nondelegation doctrine, which arises from the constitutional principle of separation of powers, Congress may not constitutionally delegate legislative power to executive branch agencies. In its current form, this doctrine has little relevance to efforts to define “frontier model.” Under current law, Congress can validly delegate rulemaking authority to an agency as long as the statute in which the delegation occurs includes an “intelligible principle” that provides adequate guidance for the exercise of that authority.[ref 118] In practice, this is an easy standard to satisfy—even vague and general legislative guidance, such as directing agencies to regulate in a way that “will be generally fair and equitable and will effectuate the purposes of the Act,” has been held to contain an intelligible principle.[ref 119] The Supreme Court has used the nondelegation doctrine to strike down statutes only twice, in two 1935 decisions invalidating sweeping New Deal laws.[ref 120]

However, some commentators have suggested that the Supreme Court may revisit the nondelegation doctrine in the near future,[ref 121] perhaps by discarding the “intelligible principle” test in favor of something like the standard suggested by Justice Gorsuch in his 2019 dissent in Gundy v. United States.[ref 122] In Gundy, Justice Gorsuch suggested that the nondelegation doctrine, properly understood, requires Congress to make “all the relevant policy decisions” and delegate to agencies only the task of “filling up the details” via regulation.[ref 123]

Therefore, if the Supreme Court does significantly strengthen the nondelegation doctrine, it is possible that a statute authorizing an agency to create a regulatory definition of “frontier model” would need to include meaningful guidance as to what the definition should look like. This is most likely to be the case if the regulatory definition in question is a key part of an extremely significant regulatory scheme, because “the degree of agency discretion that is acceptable varies according to the power congressionally conferred.”[ref 124]  Congress generally “need not provide any direction” to agencies regarding the manner in which it defines specific and relatively unimportant technical terms,[ref 125] but must provide “substantial guidance” for extremely important and complex regulatory tasks that could significantly impact the national economy.[ref 126] 

C. The major questions doctrine

Like the nondelegation doctrine, the major questions doctrine is a constitutional limitation on Congress’s ability to delegate rulemaking power to agencies. Like the nondelegation doctrine, it addresses concerns about the separation of powers and the increasingly prominent role executive branch agencies have taken on in the creation of important legal rules. Unlike the nondelegation doctrine, however, the major questions doctrine is a recent innovation. The Supreme Court acknowledged it by name for the first time in the 2022 case West Virginia v. Environmental Protection Agency,[ref 127] where it was used to strike down an EPA rule regulating power plant carbon dioxide emissions. Essentially, the major questions doctrine provides that courts will not accept an interpretation of a statute that grants an agency authority over a matter of great “economic or political significance” unless there is a “clear congressional authorization” for the claimed authority.[ref 128] Whereas the nondelegation doctrine provides a way to strike down statutes as unconstitutional, the major questions doctrine only affects the way that statutes are interpreted. 

Supporters of the major questions doctrine argue that it helps to rein in excessively broad delegations of legislative power to the administrative state and serves a useful separation-of-powers function. The doctrine’s critics, however, have argued that it limits Congress’s ability to set up flexible regulatory regimes that allow agencies to respond quickly and decisively to changing circumstances.[ref 129] According to this school of thought, requiring a clear statement authorizing each economically significant agency action inhibits Congress’s ability to communicate broad discretion in handling problems that are difficult to foresee in advance. 

This difficulty is particularly salient in the context of regulatory regimes for the governance of emerging technologies.[ref 130] Justice Kagan made this point in her dissent from the majority opinion in West Virginia, where she argued that the statute at issue was broadly worded because Congress had known that “without regulatory flexibility, changing circumstances and scientific developments would soon render the Clean Air Act obsolete.”[ref 131] Because advanced AI systems are likely to have a significant impact on the U.S. economy in the coming years,[ref 132] it is plausible that the task of choosing which systems should be categorized as “frontier” and subject to increased regulatory scrutiny will be an issue of great “economic and political significance.” If it is, then the major questions doctrine could be invoked to invalidate agency efforts to promulgate or amend a definition of “frontier model” to address previously unforeseen unsafe capabilities. 

For example, consider a hypothetical federal statute instituting a licensing regime for frontier models that includes a definition similar to the placeholder in EO 14110 (empowering the Bureau of Industry and Security to “define, and thereafter update as needed on a regular basis, the set of technical conditions [that determine whether a model is a frontier model].”). Suppose that BIS initially defined “dual-use foundation model” under this statute using a regularly updated compute threshold, but that ten years after the statute’s enactment a new kind of AI system was developed that could be trained to exhibit cutting-edge capabilities using a relatively small quantity of training compute. If BIS attempted to amend its regulatory definition of “frontier model” to include a capabilities threshold that would cover this newly developed and economically significant category of AI system, that new regulatory definition might be challenged under the major questions doctrine. In that situation, a court with deregulatory inclinations might not view the broad congressional authorization for BIS to define “frontier model” as a sufficiently clear statement of congressional intent to allow BIS to later institute a new and expanded licensing regime based on less objective technical criteria.[ref 133] 

VI. Conclusion

One of the most common mistakes that nonlawyers make when reading a statute or regulation is to assume that each word of the text carries its ordinary English meaning. This error occurs because legal rules, unlike most writing encountered in everyday life, are often written in a sort of simple code where a number of the terms in a given sentence are actually stand-ins for much longer phrases catalogued elsewhere in a “definitions” section. 

This tendency to overlook the role that definitions play in legal rules has an analogue in a widespread tendency to overlook the importance of well-crafted definitions to a regulatory scheme. The object of this paper, therefore, has been to explain some of the key legal considerations relevant to the task of defining “frontier model” or any of the analogous phrases used in existing laws and regulations. 

One such consideration is the role that should be played by statutory and regulatory definitions, which can be used independently or in conjunction with each other to create a definition that is both technically sound and democratically legitimate. Another is the selection and combination of potential definitional elements, including technical inputs, capabilities metrics, risk, deployment context, and familiarity, that can be used independently or in conjunction with each other to create a single statutory or regulatory definition. Legal mechanisms for facilitating rapid and frequent updating for regulations targeting emerging technologies also merit attention. Finally, the nondelegation and major questions doctrines and the recent elimination of Chevron deference may affect the scope of discretion that can be conferred for the creation and updating of regulatory definitions.

Beyond a piecemeal approach: prospects for a framework convention on AI

The future of international scientific assessments of AI’s risks

Computing power and the governance of artificial intelligence

AI is like… A literature review of AI metaphors and why they matter for policy

Executive summary

This report provides an overview, taxonomy, and preliminary analysis of the role of basic metaphors and analogies in AI governance. 

Aim: The aim of this report is to contribute to improved analysis, debate, and policy for AI systems by providing greater clarity around the way that analogies and metaphors can affect technology governance generally, around how they may shape AI governance, and about how to improve the processes by which some analogies or metaphors for AI are considered, selected, deployed, and reviewed.

Summary: In sum, this report:

1. Draws on technology law scholarship to review five ways in which metaphors or analogies exert influence throughout the entire cycle of technology policymaking by shaping:
   1. patterns of technological innovation; 
   2. the study of particular technologies’ sociotechnical impacts or risks; 
   3. which of those sociotechnical impacts make it onto the regulatory agenda; 
   4. how those technologies are framed within the policymaking process in ways that highlight some issues and policy levers over others; and 
   5. how these technologies are approached within legislative and judicial systems. 
2. Illustrates these dynamics with brief case studies where foundational metaphors shaped policy for cyberspace, as well as for recent AI issues. 
3. Provides an initial atlas of 55 analogies for AI, which have been used in expert, policymaker, and public debate to frame discussion of AI issues, and discusses their implications for regulation.
4. Reflects on the risks of adopting unreflexive analogies and misspecified (legal) definitions.

Below, the reviewed analogies are summarized in Table 1.

Introduction

Everyone loves a good analogy like they love a good internet meme—quick, relatable, shareable,[ref 1] memorable, and good for communicating complex topics to family.

Background: As AI systems have become increasingly capable and have had increasingly public impacts, there has been significant public and policymaker debate over the technology. Given the breadth of the technology’s application, many of these discussions have come to deploy—and contest—a dazzling range of analogies, metaphors, and comparisons for AI systems in order to understand, frame, or shape the technologies’ impact and its regulation.[ref 2] Yet the speed with which many often jump to invoke particular metaphors—or to contest the accuracy of others—leads to frequent confusion over these analogies, how they are used, and how they are best evaluated or compared.[ref 3] 

Rationale: Such debates are not just about wordplay—metaphors matter. Framings, metaphors, analogies, and (at the most specific end) definitions can strongly affect many key stages of the world’s response to a new technology, from the initial developmental pathways for technology, to the shaping of policy agendas, to the efficacy of legal frameworks.[ref 4] They have done so consistently in the past, and we have reason to believe they will especially do so for (advanced) AI. Indeed, recent academic, expert, public, and legal contests around AI often already strongly turn on “battles of analogies.”[ref 5] 

Aim: Given this, there is a need for those speaking about AI to better understand (a) when they speak in analogies—that is, when the ways in which AI is described (inadvertently) import one or more foundational analogies; (b) what it does to utilize one or another metaphor for AI; (c) what different analogies could be used instead; (d) how the appropriateness of one or another metaphor is best evaluated; and (e) what, given this, might be the limits or risks of jumping at particular analogies. 

This report aims to respond to these questions and contribute to improved analysis, debate, and policy by providing greater clarity around the role of metaphors in AI governance, the range of possible (alternate) metaphors, and good practices in constructing and using metaphors. 

Caveats: The aim here is not to argue against the use of any analogies in AI policy debates—if that were even possible. Nor is it to prescribe (or dismiss) one or another metaphor for AI as “better” (or “worse”) per se. The point is not that one particular comparison is the best and should be adopted by all, or that another is “obviously” flawed. Indeed, in some sense, a metaphor or analogy cannot be “wrong,” only more tenuous and more or less suitable when considered from the perspective of some values or some (regulatory) purpose. As such, different metaphors may work best in different contexts. Given this, this report highlights the diversity of analogies in current use and provides context for more informed future discourse and policymaking. 

Terminology: Strictly speaking, there is a difference between a metaphor—“an implied comparison between two things of unlike nature that yet have something in common”—and an analogy—“a non-identical or non-literal similarity comparison between two things, with a resulting predictive or explanatory effect.”[ref 6] However, while in legal contexts the two can be used in slightly different ways, cognitive science suggests that humans process information by metaphor and by analogy in similar ways.[ref 7] As a result, within this report, “analogy” and “metaphor” will be used relatively interchangeably to refer to (1) communicated framings of an (AI) issue that describe that issue (2) through terms, similes, or metaphors which rely on, invoke, or importreferences to a different phenomenon, technology, or historical event, which (3) is (assumed to be) comparable in one or more ways (e.g., technical, architectural, political, or moral) (4) which are relevant to evaluating or responding to the (AI) issue at hand. Furthermore, the report will use the term “foundational metaphor” to discuss cases where a particular metaphor for the technology has become deeply established and embedded within larger policy programs, such that the nature of the metaphor as a metaphor may even become unclear.

Structure: Accordingly, this report now proceeds as follows. In Part I, it discusses why and how definitions matter to both the study and practice of AI governance. It reviews five ways in which analogies or definitions can shape technology policy generally. To illustrate this, Part II reviews a range of cases in which deeply ingrained foundational metaphors have shaped internet policy as well as legal responses to various AI uses. In Part III, this report provides an initial atlas of 55 different analogies that have been used for AI in recent years, along with some of their regulatory implications. Part IV briefly discusses the risks of using analogies in unreflexive ways.

I. How metaphors shape technology governance

Given the range of disciplinary backgrounds in debates over AI, we should not be surprised that the technology is perceived and understood differently by many. 

Nonetheless, it matters to get clarity, because terminological and analogical framing effects happen at all stages in the cycle from technological development to societal response. They can shape the initial development processes for technologies as well as the academic fields and programs that study their impacts.[ref 8] Moreover, they can shape both the policymaking processes and the downstream judicial interpretation and application of legislative texts.

1. Metaphors shape innovation

Metaphors and analogies are strongly rooted in human psychology.[ref 9] Even some nonhuman animals think analogically.[ref 10] Indeed, human creativity has even been defined as “the capacity to see or interpret a problematic phenomenon as an unexpected or unusual instance of a prototypical pattern already in one’s conceptual repertoire.”[ref 11]

Given this, metaphors and analogies can shape and constrain the ability of humans to collectively create new things.[ref 12] In this way, technology metaphors can affect the initial human processes of invention and investment that drive the development of AI and other technologies in the first place. It has been suggested that foundational metaphors can influence the organization and direction of scientific fields—and even that all scientific frameworks could to some extent be viewed as metaphors.[ref 13] For example, the fields of cell biology and biotechnology have for decades been shaped by the influential foundational metaphor that sees biological cells as “machines,” which has led to sustained debates over the scientific use and limits of that analogy in shaping research programs.[ref 14] 

More practically, at the development and marketing stage, metaphors can shape how consumers and investors assess proposed startup ideas[ref 15] and which innovation paths attract engineer, activist, and policymaking interest and support. In some such cases, metaphors can support and spur on innovation; for instance, it has been argued that through the early 2000s, the coining of specific IT metaphors for electric vehicles—as a “computer on wheels”—played a significant role in sustaining engineer support for and investment in this technology, especially during an industry downturn in the wake of General Motors’ sudden cancellation of its EV1 electric car.[ref 16] 

Conversely, metaphors can also hold back or inhibit certain pathways of innovation; for instance, in the Soviet Union in the early 1950s, the field of cybernetics (along with other fields such as genetics or linguistics) fell victim to anti-American campaigns, which characterized it as “an ‘obscurantist’, ‘bourgeois pseudoscience’”.[ref 17] While this did not affect the early development of Soviet computer technology (which was highly prized by the state and the military), the resulting ideological rejection of the “man-machine” analogy by Marxist-Leninist philosophers led to an ultimately dominant view, in Soviet sciences, of computers as solely “tools to think with” rather than “thinking machines,” holding back the consolidation of the field (such that even the label “AI” would not be recognized by the Soviet Academy of Sciences until 1987) and shifting research attention into projects that focused on the “situational management” of large complex systems rather than the pursuit of human-like thinking machines.[ref 18] This stood in contrast to US research programs, such as DARPA’s 1983–1993 Strategic Computing Initiative, an extensive, $1 billion program to achieve “machines that think.”[ref 19]

2. Metaphors inform the study of technologies’ impacts

Particular definitions also shape and prime academic fields that study the impacts of these technologies (and which often may uncover or highlight particular developments as issues for regulation). Definitions affect which disciplines are drawn to work on a problem, what tools they bring to hand, and how different analyses and fields can build on one another. For instance, it has been argued that the analogy between software code and legal text has supported greater and more productive engagement by legal scholars and practitioners with such code at the level of its (social) meaning and effects (rather than narrowly on the level of the techniques used).[ref 20] Given this, terminology can affect how AI governance is organized as a field of analysis and study, what methodologies are applied, and what risks or challenges are raised or brought up.

3. Metaphors set the regulatory agenda 

More directly, particular definitions or frames for a technology can set and shape the policymaking agenda in various ways. 

For instance, terms and frames can raise (or suppress) policy attention for an issue, affecting whether policymakers or the public care (enough) about a complex and often highly technical topic in the first place to take it up for debate or regulation. For instance, it has been argued that framings that focus on the viscerality of the injuries inflicted by a new weapon system have in the past boosted international campaigns to ban blinding lasers and antipersonnel mines, yet they ended up being less successful in spurring effective advocacy around “killer robots.”[ref 21] 

Moreover, metaphors—and especially specific definitions—can shape (government) perceptions of the empirical situation or state of play around a given issue. For instance, the particular definition used for “AI” can directly affect which (industrial or academic) metrics are used to evaluate different states’ or labs’ relative achievements or competitiveness in developing the technology. In turn, that directly shapes downstream evaluations of which nation is “ahead” in AI.[ref 22] 

Finally, terms can frame the relevant legal actors and policy coalitions, enabling (or inhibiting) inclusion and agreement at the level of interest or advocacy groups that push for (or against) certain policy goals. For instance, the choice for particular terms or framings that meet with broad agreement or acceptance amongst many actors can make it easier for a diverse set of stakeholders to join together in pushing for regulatory actions. However, such agreement may be fostered by definitional clarity, when terms or frames are transparent and meet with wider acceptance, or because of definitional ambiguity, when a broad term (such as “ethical AI”) allows for sufficient ambiguity that different actors can meet on an “incompletely theorized agreement”[ref 23] to pursue a shared policy program on AI.

4. Metaphors frame the policymaking process

Terms can have a strong overall effect on policy issue-framing, foregrounding different problem portfolios as well as regulatory levers. For instance, early societal debates around nanotechnology were significantly influenced by analogies with asbestos and genetically modified organisms.[ref 24]

Likewise, regulatory initiatives that frame AI systems as “products” imply that these fit easily within product safety frameworks—even if that may be a poor or insufficient model for AI governance, for instance because it is a model that fails to address any risks at the developmental stage[ref 25] or because it fails to accurately focus on fuzzier impacts on fundamental rights if those cannot be easily classified as consumer harms.[ref 26] 

This is not to say that the policy-shaping influence of terms (or explicit metaphors) is absolute and irrevocable. For instance, in a different policy domain, a 2011 study found that using metaphors that described crime as a “beast” led study participants to recommend law-and-order responses, whereas describing it as a “virus” led them to put more emphasis on public-health-style policies. However, even under the latter framing, law-and-order policy responses still prevailed, simply commanding a smaller majority than they would otherwise.[ref 27] 

Nonetheless, metaphors do exert sway throughout the policymaking process. For instance, they can shape perceptions of the feasibility of regulation by certain routes. As an example, framings of digital technologies that emphasize certain traits of technologies—such as the “materiality” or “seeming immateriality,” or the centralization or decentralization, of technologies like submarine cables, smart speakers, search engines, or the bitcoin protocol—can strongly affect perceptions of whether, or by what routes, it is most feasible to regulate that technology at the global level.[ref 28] 

Likewise, different analogies or historical comparisons for proposed international organizations for AI governance—ranging from the IAEA and IPCC to the WTO or CERN—often import tacit analogical comparisons (or rather constitute “reflected analogies”) between AI and those organizations’ subject matter or mandates in ways that shape the perceptions of policymakers and the public regarding which of AI’s challenges require global governance, whether or which new organizations are needed, and whether the establishment of such organizations will be feasible.[ref 29]

5. Metaphors and analogies shape the legislative & judicial response to tech

Finally, metaphors, broad analogies, and specific definitions can frame legal and judicial treatment of a technology in both the ex ante application of AI-focused regulations and the ex post subsequent judicial interpretation of either such AI-specific legislation or of general regulations in the context of cases involving AI. 

Indeed, much of legal reasoning, especially in court systems, and especially in common law jurisdictions, is deeply analogical.[ref 30] This is for various reasons.[ref 31] For one, legal actors are also human, and strong features of human psychology can skew these actors towards the use of analogies that refer to known and trusted categories: as such, as Mandel has argued, “availability and representativeness heuristics lead people to view a new technology and new disputes through existing frames, and the status quo bias similarly makes people more comfortable with the current legal framework.”[ref 32] This is particularly the case because much of legal scholarship and work aims to be “problem-solving” rather than “problem-finding”[ref 33] and to respond to new problems by appealing to pre-existent (ethical or legal) principles, norms, values, codes, or laws.[ref 34] Moreover, from an administrative perspective, it is often easier and more cost-effective to extend existing laws by analogy. 

Finally, and more fundamentally, the resort to analogy by legal actors can be a shortcut that aims to apply the law, and solve a problem, through an “incompletely theorized agreement” that does not require reopening contentious questions or debates over the first principles or ultimate purposes of the law,[ref 35] or renegotiating hard-struck legislative agreements. This is especially the case at the level of international law, where either negotiating new treaties or explicitly amending multilateral treaties to incorporate a new technology within an existing framework can be wrought, drawn-out processes[ref 36] such that many actors may prefer ultimately addressing new issues (such as cyberwar) within existing norms or principles by analogizing them to well-established and well-regulated behaviors.[ref 37]

Given this, when confronted with situations of legal uncertainty—as often happens with a new technology[ref 38]—legal actors may favor the use of analogies to stretch existing law or to interpret new cases as falling within existing doctrine. That does not mean that courts need immediately settle or converge on one particular “right” analogy. Indeed, there are always multiple analogies possible, and these can have significantly different implications for how the law is interpreted and applied. That means that many legal cases involving technology will involve so-called “battles of analogies.”[ref 39] For example, in recent class action lawsuits that have accused generative AI providers such as Stable Diffusion and Midjourney of copyright infringement, plaintiffs have argued that these generative AI models are “essentially sophisticated collage tools, with the output representing nothing more than a mash-up of the training data, which is itself stored in the models as compressed copies.”[ref 40] Some have countered that this analogy suffers some technical inaccuracies, since current generative AI models do not store compressed copies of the training data, such that a better analogy would be that of an “art inspector” that takes every measurement possible—implying that model training either is not governed by copyright law or constitutes fair use.[ref 41] 

Finally, even if specific legislative texts move to adopt clear, specific statutory definitions for AI—in a way that avoids (explicit) comparison or analogy with other technologies or behavior—this may not entirely avoid framing effects. Most obviously, legislative definitions for key terms such as “AI” obviously affect the material scope of regulations and policies that use and define such terms.[ref 42] Indeed, the effects of particular definitions have impacts on regulation not only ex ante but also ex post: in many jurisdictions, legal terms are interpreted and applied by courts based on their widely shared “ordinary meaning.”[ref 43] This means, for instance, that regulations that refer to terms such as “advanced AI,” “frontier AI,” or “transformative AI”[ref 44] might not necessarily be interpreted or applied in ways that are in line with how the term is understood within expert communities.[ref 45] 

All of this underscores the importance of our choice of terms and frames—whether broad and indirect metaphors or concrete and specific legislative definitions—when grappling with the impacts of this technology on society.

II. Foundational metaphors in technology law: Cases

Of course, these dynamics are not new and have been studied in depth in fields such as cyberlaw, law and technology, and technology law.[ref 46] For instance, we can see many of these framing dynamics within societal (and regulator) responses to other cornerstone digital technologies. 

1. Metaphors in internet policy: Three cases

For instance, for the complex sociotechnical system[ref 47] commonly called the internet, foundational metaphors have strongly shaped regulatory debates, at times as much as sober assessments of the nuanced technical details of the artifacts involved have.[ref 48] As noted by Rebecca Crootof: 

“A ‘World Wide Web’ suggests an organically created common structure of linked individual nodes, which is presumably beyond regulation. The ‘Information Superhighway’ emphasizes the import of speed and commerce and implies a nationally funded infrastructure subject to federal regulation. Meanwhile, ‘cyberspace’ could be understood as a completely new and separate frontier, or it could be viewed as yet one more kind of jurisdiction subject to property rules and State control.”[ref 49]

For example, different terms (and the foundational metaphors they entail) have come to shape internet policy in various ways and domains. Take for instance the following cases: 

Institutional effects of framing cyberwar policy within cyber-“space”: For over a decade, the US military framed the internet and related systems as a “cyberspace”—that is, just another “domain” of conflict along with land, sea, air, and space—leading to strong consequences institutionally (expanding the military’s role in cybersecurity and supporting the creation of US Cyber Command) as well as for how international law has subsequently been applied to cyber operations.[ref 50] 

Issue-framing effects of regulating data as “oil,” “sunlight,” “public utility,” or “labor”: Different metaphors for “data” have drastically different political and regulatory implications.[ref 51] The oil metaphor emphasizes data as a valuable traded commodity that is owned by whoever “extracts” it and that, as a key resource in the modern economy, can be a source of geopolitical contestation between states. However, the oil metaphor implies that the history of data prior to its collection is not relevant and so sidesteps questions of any “misappropriation or exploitation that might arise from data use and processing.”[ref 52] Moreover, even within an regulatory approach that emphasizes geopolitical competition over AI, one can still critique the “oil” metaphor as misleading, for instance because of the ways in which it skews debates over how to assess “data competitiveness” in military AI.[ref 53] By contrast, the sunlight metaphor emphasizes data as a ubiquitous public resource that ought to be widely pooled and shared for social good, de-emphasizing individual data privacy claims; the public utility metaphor sees data as an “infrastructure” that requires public investment and new institutions, such as data trusts or personal data stores, to guarantee “data stewardship”; and the labor frame asserts the ownership rights of the individuals generating data against what are perceived as extractive or exploitative practices of “surveillance capitalism.”[ref 54]

Judicial effects of treating search engines as “newspaper editorials” in censorship cases: In the mid-2000s, US court rulings involving censorship on search engines tended to analyze them by analogy to older technologies such as the newspaper editorial.[ref 55] As these examples suggest, different terms and their metaphors matter. They serve as intuition pumps for key audiences (public, policy) that otherwise may have significant disinterest in, lack of expertise in, inferential distance to, or limited bandwidth for new technologies. Moreover, as seen in social media platforms and online content aggregators’ resistance to being described as “media companies” rather than “technology companies,”[ref 56] even seemingly innocuous terms can carry significant legal and policy implications—in doing so, such terms can serve as a legal “sorter,” determining whether a technology (or the company developing and marketing it) is considered as falling into one or another regulatory category.[ref 57]

2. Metaphors in AI law: Three cases

Given the role of metaphors and definitions to strongly shape the direction and efficacy of technology law, we should expect them to likewise play a strong role in affecting the framing and approach of AI regulation in the future, for better or worse. Indeed, in a range of domains, they have already done so:

Autonomous weapons systems under international law: International lawyers often aim to subsume new technologies under (more or less persuasive) analogies to existing technologies or entities that are already regulated.[ref 58] As such, different analogies have been drawn between autonomous weapons systems to weapons, combatants, child soldiers, or animal combatants—all of which lead to very different consequences for their legality under international humanitarian law.[ref 59] 

Release norms for AI models with potential for misuse: In debates over the potential misuse risks from emerging AI systems, efforts to attempt to restrict or slow publication of new systems with potential for misuse have found themselves challenged by framings that pitch the field of AI as being intrinsically an open science (where new findings should be shared whatever the risks) versus those that emphasize analogies to cybersecurity (where dissemination can help defenders protect against exploits). Critically, however, both of these analogies may misstate or underappreciate the dynamics that affect the offense-defense balance of new AI capabilities: while in information security the disclosure of software vulnerabilities has traditionally favored defense, this cannot be assumed for AI research, where (among others) it can be much more costly or intractable to “patch” the social vulnerabilities exploited by AI capabilities.[ref 60]

Liability for inaccurate or unlawful speech produced by AI chatbots, large language models, and other generative AI: In the US, Section 230 of the 1996 Communications Decency Act protects online service providers from liability for user-generated content that they host and has accordingly been considered a cornerstone to the business model of major online platforms and social media companies.[ref 61] For instance, in Spring 2023, the US Supreme Court took up two lawsuits—Gonzales v. Google and Twitter v. Taamneh—which could have shaped Section 230 protections for algorithmic recommendations.[ref 62] While the Court’s rulings on these cases avoided addressing the issue,[ref 63] similar court cases (or legislation) could have strong implications for whether digital platforms or social media companies will be held liable for unlawful speech produced by large language model-based AI chatbots.[ref 64] If such AI chatbots are analogized to existing search engines, they might be able to rely on a measure of protection from Section 230, greatly facilitating their deployment, even if they link to inaccurate information. Conversely, if these chatbot systems are considered so novel and creative that their output goes beyond the functions of a search engine, they might instead be considered as “information content providers” within the remit of the law—or simply held to be beyond the law’s remit (and protection) entirely.[ref 65] This would mean that technology companies would be held legally responsible for their AI’s outputs. If that were the case, this reading of the law would significantly restrict the profitability of many AI chatbots, given the tendency of the underlying LLMs to “hallucinate” facts.[ref 66]

All this again highlights that different definitions or terms for AI will frame how policymakers and courts understand the technology. This creates a challenge for policy, which must address the transformative impact and potential risks of AI as they are (and as they may soon be), and not only as they can be easily analogized to other technologies and fields. What does that mean in the context of developing AI policy in the future?

III. An atlas of AI analogies

Development of policy must contend with the lack of settled definitions for the term “AI,” with the varied concepts and ideas projected onto it, and with the pace at which new terms —from “foundation models” to “generative AI”—are often coined and adopted.[ref 67]

Indeed, this breadth of analogies that are coined around AI should not be surprising, given that even just the term “artificial intelligence” has a number of aspects that support conceptual fluidity (or alternately, confusion). This is for various reasons.[ref 68] In the first place, the term invokes a term—“intelligence”—which is in widespread and everyday use, and which for many people has strong (evaluative or normative) connotations. It is essentially a suitcase word that packages together many competing meanings,[ref 69] even while it hides deep and perhaps even intractable scientific and philosophical disagreement[ref 70] and significant historical and political baggage.[ref 71] 

Secondly, and in contrast to, say, “blockchain ledgers,” AI technology comes with a baggage of decades of depictions in popular culture—and indeed centuries of preceding stories about intelligent machines[ref 72]—resulting in a whole genre of tropes or narratives that can color public perceptions and policymaker debates. 

Thirdly, AI is an evocative general-purpose technology that sees use in a wide variety of domains and accordingly has provoked commentary from virtually every disciplinary angle, including neuroscience, philosophy, psychology, law, politics, and ethics. As a result of this, a persistent challenge in work on AI governance—and indeed, in the broader public debates around AI—has been that different people use the word “AI” to refer to widely different artifacts, practices, or systems, or operate on the basis of definitions or understandings which package together a range of implicit assumptions.[ref 73]

Thus, it is no surprise that AI has been subjected to a diverse range of analogies and frames. To understand potential implications of AI analogies, we can draw a taxonomy of common framings of AI (see Table 2), whereby we can distinguish between analogies that focus on: 

1. the essence or nature of AI (what AI “is”), 
2. AI’s operation (how AI works), 
3. our relation to AI (how we relate to AI as subject), 
4. AI’s societal function (how AI systems are or can be used), 
5. AI’s impact (the unintended risks, benefits, and other side-effects of AI).

Different terms for AI can therefore invoke different frames of reference or analogies. Use of analogies—by policymakers, researchers, or the public—may be hard to avoid, and they can often serve as fertile intuition pumps. 

IV. The risks of unreflexive analogies 

However, while metaphors can be productive (and potentially irreducible) in technology law, they also come with many risks. Given that analogies are shorthands or heuristics that compress or highlight salient features, challenges can creep in the more removed they are from the specifics of the technology in question. 

Indeed, as Crootof and Ard have noted, “[a]n analogy that accomplishes an immediate aim may gloss over critical distinctions in the architecture, social use, or second-order consequences of a particular technology, establishing an understanding with dangerous and long-lasting implications.”[ref 159]

Specifically: 

1. The selection and foregrounding of a certain metaphor hides that there are always multiple analogies possible for any new technology, and each of these advances different “regulatory narratives.” 
2. Analogies can be misleading by failing to capture a key trait of the technology or by alleging certain characteristics that do not actually exist. 
3. Analogies limit our ability to understand the technology—in terms of its possibilities and limits—on its own terms.[ref 160]

The challenge is that unreflexive drawing of analogies in a legal context can lead to ineffective or even dangerous laws,[ref 161] especially once inappropriate analogies become entrenched.[ref 162]

However, even if one tries to avoid explicit analogies between AI and other technologies, apparently “neutral” definitions of AI that seek to focus solely on the technology’s “features” can and still do frame policymaking in ways that may not be neutral. For instance, Kraftt and colleagues found that whereas definitions of AI that emphasize “technical functionality” are more widespread among AI researchers, definitions that emphasize “human-like performance” are more prevalent among policymakers, which they suggest might prime policymaking towards future threats.[ref 163] 

As such, it is not just loose analogies or comparisons that can affect policy, but also (seemingly) specific technical or legislative terms. The framing effects of such terms do not only occur at the level of broad policy debates but can also have strong legal implications. In particular, they can create challenges for law when narrowly specified regulatory definitions are suboptimal.[ref 164]  

This creates twin challenges. On the one hand, picking suitable concepts or categories can be difficult at an early stage of a technology’s development and deployment, when its impacts and limits are not always fully understood.[ref 165] At the same time, the costs of picking and locking in the wrong terms or framings within legislative texts can be significant. 

Specifically, beyond the opportunity costs of establishing better concepts or terms, unreflexively establishing legal definitions for key terms can create the risk of later, downstream “governance misspecification.”[ref 166] Such misspecification can occur when regulation is originally targeted at a particular artifact or (technological) practice through a particular material scope and definition for those objects. The implicit assumption here is that the term in question is a meaningful proxy for the underlying societal or legal goals to be regulated. While that may be appropriate in many cases, there is a risk that the law becomes less efficient, ineffective, or even counterproductive if either initial misapprehension of the technology or subsequent technological developments lead to that proxy term coming apart from the legislative goals.[ref 167] Such misspecification can be seen in various cases of technology governance and regulation, including 1990s US export control thresholds for “high-performance computers” that treated the technology as far too static;[ref 168] the Outer Space Treaty’s inability to anticipate later Soviet Fractional Orbital Bombardment System (FOBS) capabilities, which were able to position nuclear weapons in space without, strictly, putting them “in orbit”;[ref 169] or initial early-2010s regulatory responses to drones or self-driving cars, which ended up operating on under- and overinclusive definitions of these technologies.[ref 170]

Given this, the aim should not be to find the “correct” metaphor for AI systems. Rather, a good policy is to consider when and how different frames can be more useful for specific purposes, or for particular actors and/or (regulatory) agencies. Rather than aiming to come up with better analogies directly, this focuses regulatory debates on developing better processes for analogizing and for evaluating these analogies. For instance, such processes can depart from broad questions, such as: 

1. What are the foundational metaphors used in this discussion of AI? What features do they focus on? Do these matter in the way they are presented?
2. What other metaphors could have been chosen for these same features or aspects of AI? 
3. What aspects or features of AI do these metaphors foreground? Do they capture these features well? 
4. What features are occluded? What are the consequences of these being occluded?
5. What are the regulatory implications of these different metaphors? In terms of the coalitions they enable or inhibit, the issue and solution portfolios they highlight, or of how they position the technology within (or out of) the jurisdiction of existing institutions?

Improving these ways in which we analogize AI clearly needs significantly more work. However, it is critical that we do so to improve how we draw on frames and metaphors for AI and to ensure that—whether we are trying to understand AI itself, appreciate its impacts, or govern them effectively—our metaphors aid rather than lead us astray.

Conclusion

As AI systems have received significant attention, many have invoked a range of diverse analogies and metaphors. This has created an urgent need for us to better understand (a) when we speak of AI in ways that (inadvertently) import one or more analogies, (b) what it does to utilize one or another metaphor for AI, (c) what different analogies could be used instead for the same issue, (d) how the appropriateness of one or another metaphor is best evaluated, and (e) what, given this, might be the limits or risks of jumping at particular analogies. 

This report has aimed to contribute to answers to these questions and enable improved analysis, debate, and policymaking for AI by providing greater theoretical and empirical backing to how metaphors and analogies matter for policy. It has reviewed 5 pathways by which metaphors shape and affect policy and reviewed 55 analogies used to describe AI systems. This is not meant as an exhaustive overview but as the basis for future work. 

The aim here has not been to argue against the use of metaphors but for a more informed and reflexive and careful use of these metaphors. Those who engage in debate within and beyond the field should at least have greater clarity about the ways that these concepts are used and understood, and what are the (regulatory) implications of different framings. 

The hope is that this report can contribute foundations for a more deliberate and reflexive choice over what comparisons, analogies, or metaphors we use in talking about AI—and for the ways we communicate and craft policy for these urgent questions.

Also in this series

• Maas, Matthijs, and Villalobos, José Jaime. ‘International AI institutions: A literature review of models, examples, and proposals.’ Institute for Law & AI, AI Foundations Report 1. (September 2023). https://www.law-ai.org/international-ai-institutions
• Maas, Matthijs, ‘Concepts in advanced AI governance: A literature review of key terms and definitions.’ Institute for Law & AI. AI Foundations Report 3. (October 2023). https://www.law-ai.org/advanced-ai-gov-concepts 
• Maas, Matthijs, “Advanced AI governance: A literature review.” Institute for Law & AI, AI Foundations Report 4. (November 2023). https://law-ai.org/advanced-ai-gov-litrev

International governance of civilian AI