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 37] evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks; [or]

b. based on a decision of the Commission,[ref 38] 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.

Computing power and the governance of artificial intelligence

Advanced AI governance

Executive Summary 

This literature review provides an overview and taxonomy of past and recent research in the emerging field of advanced AI governance.

Aim: The aim of this review is to help disentangle and consolidate the field, improve its accessibility, enable clearer conversations and better evaluations, and contribute to overall strategic clarity or coherence in public and policy debates. 

Summary: Accordingly, this review is organized as follows:

The introduction discusses the aims, scope, selection criteria, and limits of this review and provides a brief reading guide. 

Part I reviews problem-clarifying work aimed at mapping the parameters of the AI governance challenge, including lines of research to map and understand:

1. Key technical parameters constituting the technical characteristics of advanced AI technology and its resulting (sociotechnical) impacts and risks. These include evaluations of the technical landscape of advanced AI (its forms, possible developmental pathways, timelines, trajectories), models for its general social impacts, threat models for potential extreme risks (based on general arguments and direct and indirect threat models), and the profile of the technical alignment problem and its dedicated research field. 
2. Key deployment parameters constituting the conditions (present and future) of the AI development ecosystem and how these affect the distribution and disposition of the actors that will (first) deploy such systems. These include the size, productivity, and geographic distribution of the AI research field; key AI inputs; and the global AI supply chain. 
3. Key governance parameters affecting the conditions (present and future) for governance interventions. These include stakeholder perceptions of AI and trust in its developers, the default regulatory landscape affecting AI, prevailing barriers to effective AI governance, and effects of AI systems on the tools of law and governance themselves.
4. Other lenses on characterizing the advanced AI governance problem. These include lessons derived from theory, from abstract models and wargames, from historical case studies (of technology development and proliferation, of its societal impacts and societal reactions, of successes and failures in historical attempts to initiate technology governance, and of successes and failures in the efficacy of different governance levers at regulating technology), and lessons derived from ethics and political theory. 

Part II reviews option-identifying work aimed at mapping potential affordances and avenues for governance, including lines of research to map and understand: 

1. Potential key actors shaping advanced AI, including actors such as or within AI labs and companies, the digital AI services and compute hardware supply chains, AI industry and academia, state and governmental actors (including the US, China, the EU, the UK, and other states), standard-setting organizations, international organizations, and public, civil society, and media actors. 
2. Levers of governance available to each of these actors to shape AI directly or indirectly.
3. Pathways to influence on each of these key actors that may be available to (some) other actors in aiming to help inform or shape the key actors’ decisions around whether or how to utilize key levers of governance to improve the governance of advanced AI. 

Part III reviews prescriptive work aimed at putting this research into practice in order to improve the governance of advanced AI (for some view of the problem and of the options). This includes lines of research or advocacy to map, articulate, and advance:

1. Priorities for policy given theories of change based on some view of the problem and of the options.
2. Good heuristics for crafting AI policy. These include general heuristics for good regulation, for (international) institutional design, and for future-proofing governance.
3. Concrete policy proposals for the regulation of advanced AI, and the assets or products that can help these be realized and implemented. This includes proposals to regulate advanced AI using existing authorities, laws, or institutions; proposals to establish new policies, laws, or institutions (e.g., temporary or permanent pauses on AI development; the establishment of licensing regimes, lab-level safety practices, or governance regimes on AI inputs; new domestic governance institutions; new international AI research hubs; new bilateral agreements; new multilateral agreements; and new international governance institutions).

Introduction 

This document aims to review, structure, and organize existing work in the field of advanced AI governance. 

Background: Despite being a fairly young and interdisciplinary field, advanced AI governance offers a wealth of productive work to draw on and is increasingly structured through various research agendas[ref 1] and syllabi.[ref 2] However, while technical research on the possibility, impacts, and risks of advanced AI has been mapped in various literature reviews and distillations,[ref 3] few attempts have been made to comprehensively map and integrate existing research on the governance of advanced AI.[ref 4] This document aims to provide an overview and taxonomy of work in this field.

Aims: The aims of this review are several: 

1. Disentangle and consolidate the field to promote greater clarity and legibility regarding the range of research, connections between different research streams and directions, and open gaps or underexplored questions. Literature reviews can contribute to such a consolidation of academic work;[ref 5] 
2. Improve the field’s accessibility and reduce some of its “research debt”[ref 6] to help those new to the field understand the existing literature, in order to facilitate a more cohesive and coordinated research field with lower barriers to entry, which reduces duplication of effort or work; 
3. Enable clearer conversations between researchers exploring different questions or lines of research, discussing how and where their insights intersect or complement one another; 
4. Enable better comparison between different approaches and policy proposals; and
5. Contribute to greater strategic clarity or coherence,[ref 7] improving the quality of interventions, and refining public and policy debates. 

Scope: While there are many ways of framing the field, one approach is to define advanced AI governance as:

However, the aim of this document is not to engage in restrictive boundary policing of which research is part of this emerging field, let alone the “core” of it. The guiding heuristic here is not whether a given piece of research is directly, explicitly, and exclusively focused on certain “right” problems (e.g., extreme risks from advanced AI), nor whether it is motivated by certain political orientations or normative frameworks, nor even whether it explicitly uses certain terminology (e.g., “Transformative AI,” “AGI,” “General-Purpose AI System,” or “Frontier AI”).[ref 9] Rather, the broad heuristic is simply whether the research helps answer a part of the advanced AI governance puzzle. 

Accordingly, this review aims to cast a fairly broad net to cover work that meets any of the following criteria:

• Explicitly focuses on the governance of future advanced, potentially transformative AI systems, in particular with regard to their potential significant impacts or extreme risks.
• Focuses on the governance of today’s AI systems, where (at least some of) the authors are interested in the implications of the analysis for the governance of future AI systems;
• Focuses on today’s AI systems, where the original work is (likely) not directly motivated by a concern over (risks from) advanced AI but nonetheless offers lessons that are or could be drawn upon by the advanced AI governance community to inform insights for the governance of advanced AI systems; and
• Focuses on (the impacts or governance of) non-AI technologies or issues (such as historical case studies of technology governance), where the original work is not directly motivated by questions around AI but nonetheless offers lessons that are or could be drawn upon by the advanced AI governance community to inform insights for the governance of advanced AI systems. 

Limitations: With this in mind, there are also a range of limitations or shortcomings for this review:

• Preliminary survey: A literature review of this attempted breadth will inevitably fall short of covering all relevant work and sub-literatures in sufficient depth. In particular, given the speed of development in this field, a project like this will inevitably miss key work, so it should not be considered exhaustive. Indeed, because of the breadth of this report, I do not aim to go into the details of each topic, but rather to organize and list sources by topic. Likewise, there is some unbalance in that there has to date been more organized (technical) literature on (Part 1) characterizing the problem of advanced AI governance, than there has been on drafting concrete proposals (Part 3). As such, I invite others to produce “spin-offs” of this report which go into the detail of the content for each topic or sub-section in order to produce more in-depth literature reviews.[ref 10]
• Broad scope: In accordance with the above goal to cast a “broad net,” this review covers both work that is core to and well established in the existing advanced AI governance field, and adjacent work that could be or has been considered by some as of significant value, even if it has not been as widely recognized yet. It also casts a broad net in terms of the type of sources surveyed, covering peer reviewed academic articles, reports, books, and more informal digital resources such as web fora. 
• Incomplete in scope: By and large, this review focuses on public and published analyses and mostly omits currently in-progress, unpublished, or draft work.[ref 11] Given that a significant portion of relevant and key work in this field is unpublished, this means that this review likely will not capture all research directions in this field. Indeed, I estimate that this review captures at best ~70% of the work and research undertaken on many of these questions and subfields, and likely less. I therefore welcome further, focused literature reviews.
• A snapshot: While this review covers a range of work, the field is highly dynamic and fast-moving, which means that this project will become outdated before long. Attempts will be made to update and reissue the report occasionally. 

Finally, a few remaining disclaimers: (1) inclusion does not imply endorsement of a given article’s conclusions; (2) this review aims to also highlight promising directions, such as issues or actors, that are not yet discussed in depth in the literature. As such, whenever I list certain issues (e.g., “actors” or “levers”) without sources, this is because I have not yet found (or have missed out on) much work on that issue, suggesting there is a gap in the literature—and room for future work. Overall, this review should be seen as a living document that will be occasionally updated as the field develops. To that end, I welcome feedback, criticism, and suggestions for improvement. 

Reading guide: In general, I recommend that rather than aiming to read this from the top, readers instead identify a theme or area of interest and jump to that section. In particular, this review may be most useful to readers (a) that already have a specific research question and want to see what work has been done and how a particular line of work would fit into the larger landscape; (b) that aim to generate or distill syllabi for reading groups or courses; or (c) that aim to explore the broader landscape or build familiarity with fields or lines of research they have not previously explored. All the research presented here is collected from prior work, and I encourage readers to consult and directly cite those original sources named here.

I. Problem-clarifying work: Understanding the AI governance challenge 

Most object-level work in the field of advanced AI governance has sought to disambiguate and reduce uncertainties around relevant strategic parameters of the AI governance challenge.[ref 12]

Strategic parameters serve as highly decision-relevant or even crucial considerations, determining which interventions or solutions are appropriate, necessary, viable, or beneficial for addressing the advanced AI governance challenge. Different views of these parameters constitute underlying cruxes for different theories of actions and approaches. This review discusses three types of strategic parameters:[ref 14]

• Technical parameters of the advanced AI challenge (i.e., what are the future technical developments in AI, on what timelines and on what trajectory will progress occur, why or how might such systems pose risks, and how difficult is the alignment challenge);
• Deployment parameters of who is most likely to develop advanced AI systems and how they are likely to develop and use them (i.e., whose development decisions are to be governed); and
• Governance parameters of how, when, and why governance interventions to shape advanced AI development and deployment are most likely to be viable, effective, or productive. 

Accordingly, research in this subfield includes:

• Empirical and theoretical work aiming to identify or get better estimates of each of these parameters as they apply to advanced AI (Sections 1, 2, 3). 
• Work applying other lenses to the advanced AI governance problem, drawing on other fields (existing theories, models, historical case studies, political and ethical theory) in order to derive crucial insights or actionable lessons (Section 4).

1. Technical parameters 

An initial body of work focuses on mapping the relevant technical parameters of the challenge for advanced AI governance. This includes work on a range of topics relating to understanding the future technical landscape, understanding the likelihood of catastrophic risks given various specific threat models, and understanding the profile of the technical alignment problem and the prospects of it being solved by existing technical alignment research agendas.[ref 15]

1.1. Advanced AI technical landscape 

One subfield involves research to chart the future technical landscape of advanced AI systems.[ref 16] Work to map this landscape includes research on the future form, pathways, timelines, and trajectories of advanced AI.

Forms of advanced AI 

Work exploring distinct potential forms of advanced AI,[ref 17] including:

• strong AI,[ref 18] autonomous machine intelligence,[ref 19] general artificial intelligence,[ref 20] human-level AI (HLAI),[ref 21] general-purpose AI system (GPAIS),[ref 22] comprehensive AI services (CAIS),[ref 23] highly capable foundation models,[ref 24] artificial general intelligence (AGI),[ref 25] robust artificial intelligence,[ref 26] AI+,[ref 27] (machine/artificial) superintelligence,[ref 28] and superhuman general purpose AI,[ref 29] amongst others.

Developmental paths towards advanced AI

This includes research and debate on a range of domains. In particular, such work focuses on analyzing different hypothesized pathways towards achieving advanced AI based on different paradigms or theories.[ref 30] Note that many of these are controversial and contested, and there is pervasive disagreement over the feasibility of many (or even all) of these approaches for producing advanced AI. 

Nonetheless, some of these paradigms include programs to produce advanced AI based on:

• First principles: Approaches that aim to create advanced AI based on new fundamental insights in computer science, mathematics, algorithms, or software, producing AI systems that may, but need not, mimic human cognition.[ref 31]
• Direct/Scaling: Approaches that aim to “brute force” advanced AI[ref 32] by running (one or more) existing AI approaches with increasingly greater computing power and/or training data to exploit observed “scaling laws” in system performance.[ref 33] 
• Evolutionary: Approaches that aim to create advanced AI based on algorithms that compete to mimic the evolutionary brute search process that produced human intelligence.[ref 34]
• Reward-based: Approaches that aim to create advanced AI by running reinforcement learning systems with simple rewards in rich environments.[ref 35]
• Bootstrapping: Approaches that aim to create some minimally intelligent core system capable of subsequent recursive (self)-improvement as a “seed AI.”[ref 36] 
• Neuro-inspired: Various forms of biologically-inspired, brain-inspired, or brain-imitative approaches that aim to draw on neuroscience and/or “connectomics” to reproduce general intelligence.[ref 37]
• Neuro-emulated: Approaches that aim to digitally simulate or recreate the states of human brains at a fine-grained level, possibly producing whole-brain-emulation.[ref 38]
• Neuro-integrationist: Approaches that aim to create advanced AI based on merging components of human and digital cognition.
• Embodiment: Approaches that aim to create advanced AI by providing the AI system with a robotic physical “body”’ to ground cognition and enable it to learn from direct experience of the world.[ref 39]
• Hybrid: Approaches that rely on combining deep neural network-based approaches to AI with other paradigms (such as symbolic AI).[ref 40]

Notably, of these approaches, recent years have seen most sustained attention focused on the direct (scaling) approach and whether current approaches to advanced AI, if scaled up with enough computing power or training data, will suffice to produce advanced or transformative AI capabilities. There have been various arguments both in favor of and against this direct path. 

• Arguments in favor of a direct path: “scaling hypothesis,”[ref 41] “prosaic AGI,”[ref 42] and “Human feedback on diverse tasks (HFDT)”;[ref 43]
• Arguments against a direct path, highlighting various limits and barriers: “deep limitations,”[ref 44] “the limits of machine intelligence,”[ref 45] “why AI is harder than we think,”[ref 46] and other skeptical arguments;[ref 47] 
• Discussion of the possible features of “engineering roadmaps” for AGI-like systems.[ref 48]

Advanced AI timelines: Approaches and lines of evidence

A core aim of the field is to chart the timelines for advanced AI development across the future technical development landscape.[ref 49] This research focuses on various lines of evidence,[ref 50] which are here listed in order from more abstract to more concrete and empirical, and from relying more on outside-view arguments to relying more on inside-view arguments,[ref 51] with no specific ranking on the basis of the strength of individual lines of evidence.

Outside-view analyses of timelines

Outside-view analyses of AI development timelines, including:

• Estimates based on philosophical arguments and anthropic reasoning:
  • Prima facie likelihood that we (of all generations) are the ones to find ourselves living in the “most important” century, one that we can expect to contain things such as transformative technologies.[ref 52]
• Estimates based on extrapolating historical (growth) trends:
  • Insights from endogenous growth theory on AI development dynamics;[ref 53]
  • Likelihood of explosive economic growth occurring this century, for some reason (plausibly technological, plausibly AI[ref 54]), given analyses of long-run economic history;[ref 55]
  • The accelerating historical rate of development of new technologies[ref 56] as well as potential changes in the historical rate of increase in the economy;[ref 57]
  • The historical patterns of barriers to technology development,[ref 58] including unexpected barriers or delays in innovation,[ref 59] as well as lags in subsequent deployment or diffusion.[ref 60]
• Estimates based on extrapolating from historical trends in efforts dedicated to creating advanced AI:
  • External “semi-informative priors” (i.e., only basic information regarding how long people have attempted to build advanced, transformative AI and what resources they have used, and comparing it to how long it has taken other comparable research fields to achieve their goals given certain levels of funding and effort);
  • Arguments extrapolating from “significantly increased near-future investments in AI progress” given that (comparatively) moderate past investments already yielded significant progress.
• Estimates based on meta-induction from the track record of past predictions:
  • The general historical track record of past technological predictions, especially those made by futurists[ref 63] as well as those made in professional long-range forecasting exercises,[ref64] to understand the frequency of over- or underconfidence and of periods of excessive optimism (hype) or excessive pessimism (counterhype);[ref 65] 
  • The specific historical track record of past predictions around AI development[ref 66] and the frequency of past periods’ excessive optimism (hype) or excessive pessimism (counterhype or “underclaiming”[ref 67]).[ref 68]

Judgment-based analyses of timelines

Judgment-based analyses of timelines, including:

Estimates based on (specialist) expert opinions:

• Expert opinion surveys of anticipated rates of progress;[ref 69] 
• Expert elicitation techniques (e.g., Delphi method).[ref 70]
• Estimates based on (generalist) estimates from information aggregation mechanisms (financial markets; forecaster prediction markets):[ref 71]
  • Forecasters’ predictions of further AI progress on prediction platforms[ref 72] or forecasting competitions;[ref 73]
  • Current financial markets’ real interest rates, assuming the efficient market hypothesis, suggesting that markets reject short timelines.[ref 74]

Inside-view models on AI timelines

Inside-view models-based analyses of timelines, including:

• Estimates based on first-principle estimates of minimum resource (compute, investment) requirements for a “transformative” AI system, compared against estimated trends in these resources:
  • The “biological anchors” approach:[ref 75] Comparison with human biological cognition by comparing projected trends in the falling costs of training AI models to the expected minimum amount of computation needed to train an AI model as large as the human brain;[ref 76]
  • The “direct approach”:[ref 77] Analysis of empirical neural scaling laws in current AI systems to upper bound the compute needed to train a transformative model. In order to provide estimates of the system’s development, this analysis can be combined with estimates of future investment in model training, hardware price-performance, and algorithmic progress[ref 78] as well as with potential barriers in the (future) availability of the data and compute needed to train these models.[ref 79]
• Estimates based on direct evaluation of outputs (progress in AI systems’ capabilities):
  • Debates over the significance and implications of specific ongoing AI breakthroughs for further development;[ref 80]
  • Operationalizing and measuring the generality of existing AI systems.[ref 81]

Methodological debates on AI-timelines analysis

Various methodological debates around AI-timelines analysis:

• On the potential pitfalls in many of the common methods (forecasting methods,[ref 82] extrapolation, expert predictions[ref 83]) in forecasting AI;
• On the risk of misinterpreting forecasters who are depending on poor operationalization;[ref 84]
• On the risk of deference cycles in debates over AI timelines[ref 85] because the opinions and analyses of a small number of people end up tacitly informing the evaluations of a wide range of others in ways that create the impression of many people independently achieving similar conclusions;[ref 86]
• On the (potentially) limited utility of further discourse over and research into AGI timelines: arguments that all low-hanging fruit may already have been plucked[ref 87] and counterarguments that specific timelines remain relevant to prioritizing strategies.[ref 88]

Advanced AI trajectories and early warning signals

A third technical subfield aims at charting the trajectories of advanced AI development, especially the potential for rapid and sudden capability gains, and whether there will be advanced warning signs:

• Exploring likely AGI “takeoff speeds”:[ref 89]
  • From first principles: arguments in favor of “fast takeoff”[ref 90] vs. arguments for slow(er), more continuous development;[ref 91] 
  • By analogy: exploring historical precedents for sudden disjunctive leaps in technological capabilities.[ref 92]
• Mapping the epistemic texture of the AI development trajectory in terms of possible advance warning signs of capability breakthroughs[ref 93] or the lack of any such fire alarms.[ref 94]

1.2. Impact models for general social impacts from advanced AI 

Various significant societal impacts that could result from advanced AI systems:[ref 95]Potential for advanced AI systems to drive significant, even “explosive” economic growth[ref 96] but also risks of significant inequality or corrosive effects on political discourse;[ref 97]

• Significant impacts on scientific progress and innovation;[ref 98]
• Significant impacts on democracy;[ref 99]
• Lock-in of harmful socio-political dangers as a result of the increasing role of centralization and optimization;[ref 100]
• Impacts on geopolitics and international stability.[ref 101]

This is an extensive field that spans a wide range of work, and the above is by no means exhaustive.

1.3. Threat models for extreme risks from advanced AI 

A second subcluster of work focuses on understanding the threat models of advanced AI risk,[ref 102] based on indirect arguments for risks, specific threat models for direct catastrophe, or takeover,[ref 103] or on specific threat models for indirect risks.[ref 104]

General arguments for risks from AI

Analyses that aim to explore general arguments (by analogy, on the basis of conceptual argument, or on the basis of empirical evidence from existing AI systems) over whether or why we might have grounds to be concerned about advanced AI.[ref 105]

Analogical arguments for risks

Analogies[ref 106] with historical cases or phenomena in other domains:

• Historical cases of intelligence enabling control: emergence of human dominion over the natural world: “second species argument”[ref 107] and “the human precedent as indirect evidence of danger”;[ref 108]
• Historical cases where actors were able to achieve large shifts in power despite only wielding relatively minor technological advantages: conquistadors;[ref 109] 
• Historical cases of “lock-in” of suboptimal or bad societal trajectories based on earlier choices and exacerbated by various mechanisms for lock-in: climate change, the agricultural revolution, and colonial projects.[ref 110]

Analogies with known “control problems” observed in other domains:

• Analogies with economics principal-agent problems;[ref 111] 
• Analogies with constitutional law “incomplete contracting” theorems;[ref 112] in particular, the difficulty of specifying adequate legal responses to all situations or behaviors in advance because it is hard to specify specific and concrete rules for all situations (or in ways that cannot be gamed), whereas vague standards (such as the “reasonable person test”) may rely on intuitions that are widely shared but difficult to specify and need to be adjudicated ex post;[ref 113] 
• Analogies to economic systems[ref 114] and to bureaucratic systems and markets, and their accordant failure modes and externalities;[ref 115]
• Analogies to “Goodhart’s Law,” where a proxy target metric is used to improve a system so far that further optimization becomes ineffective or harmful;[ref 116]
• Analogies to the “political control problem”—the problem of the alignment and control of powerful social entities (corporations, militaries, political parties) with (the interests of) their societies, a problem that remains somewhat unsolved, with societal solutions relying on patchwork and fallible responses that cannot always prevent misalignment (e.g., corporate malfeasance, military coups, or unaccountable political corruption);[ref 117] 
• Analogies with animal behavior, such as cases of animals responding to incentives in ways that demonstrate specification gaming;[ref 118]
• Illustration with thought experiments and well-established narrative tropes: “sorcerer’s apprentice,”[ref 119] “King Midas problem,”[ref 120] and “paperclip maximizer.”[ref 121]

Conceptual arguments for risks

Conceptual and theoretical arguments based on existing ML architectures: 

• Arguments based on the workings of modern deep learning systems.[ref 122]

Conceptual and theoretical arguments based on the competitive environment that will shape the evolutionary development of AIs:

• Arguments suggesting that competitive pressures amongst AI developers may lead the most successful AI agents to likely have (or be given) undesirable traits, which creates risks.[ref 123]

Empirical evidence for risks

Empirical evidence of unsolved alignment failures in existing ML systems, which are expected to persist or scale in more advanced AI systems:[ref 124]

• “Faulty reward functions in the wild,”[ref 125] “specification gaming,”[ref 126] and reward model overoptimization;[ref 127] 
• “Instrumental convergence,”[ref 128] goal misgeneralization, and “inner misalignment” in reinforcement learning;[ref 129] 
• Language model misalignment[ref 130] and other unsolved safety problems in modern ML,[ref 131] and the harms from increasingly agentic algorithmic systems.[ref 132] 

Empirical examples of elements of AI threat models that have already occurred in other domains or with simpler AI systems:

• Situational awareness: cases where a large language model displays awareness that it is a model, and it can recognize whether it is currently in testing or deployment;[ref 133] 
• Acquisition of a goal to harm society: cases of AI systems being given the outright goal of harming humanity (ChaosGPT);
• Acquisition of goals to seek power and control: cases where AI systems converge on optimal policies of seeking power over their environment;[ref 134]
• Self-improvement: examples of cases where AI systems improve AI systems;[ref 135] 
• Autonomous replication: the ability of simple software to autonomously spread around the internet in spite of countermeasures (various software worms and computer viruses);[ref 136] 
• Anonymous resource acquisition: the demonstrated ability of anonymous actors to accumulate resources online (e.g., Satoshi Nakamoto as an anonymous crypto billionaire);[ref 137] 
• Deception: cases of AI systems deceiving humans to carry out tasks or meet goals.[ref 138] 

Direct threat models for direct catastrophe from AI

Work focused at understanding direct existential threat models.[ref 139] This includes:

• Various overviews and taxonomies of different accounts of AI risk: Barrett & Baum’s “model of pathways to risk,”[ref 140] Clarke et al.’s Modelling Transformative AI Risks (MTAIR),[ref 141] Clarke & Martin on “Distinguishing AI Takeover Scenarios,”[ref 142] Clarke & Martin’s “Investigating AI Takeover Scenarios,”[ref 143] Clarke’s “Classifying Sources of AI X-Risk,”[ref 144] Vold & Harris “How Does Artificial Intelligence Pose an Existential Risk?,”[ref 145] Ngo “Disentangling Arguments for the Importance of AI Safety,”[ref 146] Grace’s overview of arguments for existential risk from AI,[ref 147] Nanda’s “threat models,”[ref 148] and Kenton et al.;[ref 149]
• Analysis of potential dangerous capabilities that may be developed by general-purpose AI models, such as cyber-offense, deception, persuasion and manipulation, political strategy, weapons acquisition, long-horizon planning, AI development, situational awareness, and self-proliferation.[ref 150]

Scenarios for direct catastrophe caused by AI

Other lines of work have moved from providing indirect arguments of risk, to instead sketching specific scenarios in and through which advanced AI systems could directly inflict existential catastrophe.

Scenario: Existential disaster because of misaligned superintelligence or power-seeking AI

• Older accounts, including by Yudkowsky,[ref 151] Bostrom,[ref 152] Sotala,[ref 153] Sotala and Yampolskiy,[ref 154] and Alexander;[ref 155] 
• Newer accounts, such as Cotra & Karnofsky’s “AI takeover analysis,”[ref 156] Christiano’s account of “What Failure Looks Like,”[ref 157] Carlsmith on existential risks from power-seeking AI,[ref 158] Ngo on “AGI Safety From First Principles,”[ref 159] and “Minimal accounts” of AI takeover scenarios;[ref 160] 
• Skeptical accounts: various recent critiques of AI takeover scenarios.[ref 161]

Scenario: Gradual, irretrievable ceding of human power over the future to AI systems

• Christiano’s account of “What Failure Looks Like, (1).”[ref 162] 

Scenario: Extreme “suffering risks” because of a misaligned system

• Various accounts of “worst-case AI safety”;[ref 163]
• Potential for a “suffering explosion” experienced by AI systems.[ref 164]

Scenario: Existential disaster because of conflict between AI systems and multi-system interactions 

• Disasters because of “cooperation failure”[ref 165] or “multipolar failure.”[ref 166]

Scenario: Dystopian trajectory lock-in because of misuse of advanced AI to establish and/or maintain totalitarian regimes;

• Use of advanced AI to establish robust totalitarianism;[ref 167] 
• Use of advanced AI to establish lock-in of the future values.[ref 168]

Scenario: Failures in or misuse of intermediary (non-AGI) AI systems, resulting in catastrophe

• Deployment of “prepotent” AI systems that are non-general but capable of outperforming human collective efforts on various key dimensions;[ref 169]
• Militarization of AI enabling mass attacks using swarms of lethal autonomous weapons systems;[ref 170] 
• Military use of AI leading to (intentional or unintentional) nuclear escalation, either because machine learning systems are directly integrated in nuclear command and control systems in ways that result inescalation[ref 171] or because conventional AI-enabled systems (e.g., autonomous ships) are deployed in ways that result in provocation and escalation;[ref 172] 
• Nuclear arsenals serving as an arsenal “overhang” for advanced AI systems;[ref 173]
• Use of AI to accelerate research into catastrophically dangerous weapons (e.g., bioweapons);[ref 174]
• Use of AI to lower the threshold of access to dual-use biotechnology, creating risks of actors misusing it to create bioweapons.[ref 175]

Other work: vignettes, surveys, methodologies, historiography, critiques

• Work to sketch vignettes reflecting on potential threat models:
  • AI Impacts’ AI Vignettes project;[ref 176]
  • FLI Worldbuilding competition;[ref 177]
  • Wargaming exercises;[ref 178]
  • Other vignettes or risk scenarios.[ref 179] 

• Surveys of how researchers rate the relative probability of different existential risk scenarios from AI;[ref 180] 
• Developing methodologies for AI future developments and risk identification,[ref 181] such as red-teaming,[ref 182] wargaming exercises,[ref 183] and participatory technology assessment,[ref 184] as well as established risk identification techniques (scenario analysis, fishbone method, and risk typologies and taxonomies), risk analysis techniques (causal mapping, Delphi technique, cross-impact analysis, bow tie analysis, and system-theoretic process analysis), and risk evaluation techniques (checklists and risk matrices);[ref 185]
• Historiographic accounts of changes in AI risk arguments and debates over time:
  • General history of concerns around AI risk (1950s–present);[ref 186]
  • Early history of the rationalist and AI risk communities (1990s–2010);[ref 187]
  • Recent shifts in arguments (e.g., 2014–present);[ref 188]
  • Development and emergence of AI risk “epistemic community.”[ref 189]
• Critical investigations of and counterarguments to the case for extreme AI risks, including object-level critiques of the arguments for risk[ref 190] as well as epistemic arguments, arguments about community dynamics, and argument selection effects.[ref 191] 

Threat models for indirect AI contributions to existential risk factors

Work focused at understanding indirect ways in which AI could contribute to existential threats, such as by shaping societal “turbulence”[ref 192] and other existential risk factors.[ref 193] This covers various long-term impacts on societal parameters such as science, cooperation, power, epistemics, and values:[ref 194] 

• Destabilizing political impacts from AI systems in areas such as domestic politics (e.g., polarization, legitimacy of elections), international political economy, or international security[ref 195] in terms of the balance of power, technology races and international stability, and the speed and character of war ;
• Hazardous malicious uses;[ref 196]
• Impacts on “epistemic security” and the information environment;[ref 197]
• Erosion of international law and global governance architectures;[ref 198]
• Other diffuse societal harms.[ref 199]

1.4. Profile of technical alignment problem

• Work mapping different geographical or institutional hubs active on AI alignment: overview of the AI safety community and problem,[ref 200] and databases of active research institutions[ref 201] and of research;[ref 202]
• Work mapping current technical alignment approaches;[ref 203]
• Work aiming to assess the (relative) efficacy or promise of different approaches to alignment, insofar as possible:[ref 204] Cotra,[ref 205] Soares,[ref 206] and Leike.[ref 207] 
• Mapping the relative contributions to technical AI safety by different communities[ref 208] and the chance that AI safety problems get “solved by default”;[ref 209]
• Work mapping other features of AI safety research, such as the need for minimally sufficient access to AI models under API-based “structured access” arrangements.[ref 210]

2. Deployment parameters

Another major part of the field aims to understand the parameters of the advanced AI deployment landscape by mapping the size and configuration of the “game board” of relevant advanced AI developers—the actors whose (ability to take) key decisions (e.g., around whether or how to deploy particular advanced AI systems, how much to invest in alignment research, etc.) may be key in determining risks and outcomes from advanced AI. 

As such, there is significant work on mapping the disposition of the AI development ecosystem and how this will determine who is (or will likely be) in the position to develop and deploy the most advanced AI systems. Some work in this space focuses on mapping the current state of these deployment parameters; other work focuses on the likely future trajectories of these deployment parameters over time.

2.1. Size, productivity, and geographic distribution of the AI research field 

• Mapping the current size, activity, and productivity of the AI research field;[ref 211]
• Mapping the global geographic distribution of active AGI programs,[ref 212] including across key players such as the US or China.[ref 213]

2.2. Geographic distribution of key inputs in AI development

• Mapping the current distribution of relevant inputs in AI development, such as the distribution of computation,[ref 214] semiconductor manufacturing,[ref 215] AI talent,[ref 216] open-source machine learning software,[ref 217] etc.
• Mapping and forecasting trends in relevant inputs for AI,[ref 218] such as:
  • Trends in compute inputs scaling[ref 219] and in the training costs and GPU price-performance of machine learning systems over time;[ref 220]
  • Trends in dataset scaling and potential ceilings;[ref 221] 
  • Trends in algorithmic progress, including their effect on the ability to leverage other inputs, e.g., the relative importance of CPUs versus specialized hardware;[ref 222]
• Mapping and forecasting trends in input criticality for AI, such as trends in data efficiency[ref 223] and the degree to which data becomes the operative constraint on language model performance.[ref 224]

2.3. Organization of global AI supply chain

• Mapping the current shape of the AI supply chain;[ref 225]
• Mapping and forecasting dominant actors in the future AI ecosystem, in terms of:
  • different actors’ control of and access to key inputs and/or chokepoints;[ref 226] 
  • future shape of the AI supply chain (e.g., level of integration and monopoly structure);[ref 227] 
  • shape of AI deployment landscape (e.g., dominance of key operators of generative models vs. copycat models).

2.4. Dispositions and values of advanced AI developers

• Anticipating the likely behavior or attitude of key advanced AI actors with regard to their caution about and investment in safety research, such as expecting AI companies to “race forward” and dedicate “naive safety effort.”[ref 228]

2.5. Developments in converging technologies

• Mapping converging developments in adjacent, potentially intersecting or relevant technologies, such as cryptography,[ref 229] nanotechnology,[ref 230] and others.  

3. Governance parameters

Work on governance parameters aims to map (1) how AI systems are currently being governed, (2) how they are likely to be governed by default (given prevailing perceptions and regulatory initiatives), as well as (3) the conditions for developing and implementing productive governance interventions on advanced AI risk. 

Some work in this space focuses on mapping the current state of these governance parameters and how they affect AI governance efforts initiated today. Other work focuses on the likely future trajectories of these governance parameters.

3.1. Stakeholder perceptions of AI 

Surveys of current perceptions of AI among different relevant actors: 

• Public perceptions of the future of AI,[ref 231] of AI’s societal impacts,[ref 232] of the need for caution and/or regulation of AI,[ref 233] and of the rights or standing of AI entities;[ref 234]
• Policymaker perceptions of AI[ref 235] and the prominence of different memes, rhetorical frames, or narratives around AI;[ref 236]
• Expert views on best practices in AGI lab safety and governance.[ref 237]

Predicting future shifts in perceptions of AI among relevant actors given:

• The spread of ongoing academic conversations concerned about advanced AI risk;[ref 238]
• The effects of “warning shots,”[ref 239] or other “risk awareness moments”;[ref 240]
• The effect of motivated misinformation or politicized AI risk skepticism.[ref 241]

3.2. Stakeholder trust in AI developers 

• Public trust in different actors to responsibly develop AI;[ref 242] 
• AI-practitioner trust in different actors to responsibly develop AI[ref 243] and Chinese AI researchers’ views on the development of “strong AI.”[ref 244]

3.3. Default landscape of regulations applied to AI

This work maps the prevailing (i.e., default, “business-as-usual”) landscape of regulations that will be applied to AI in the near term. These matter as they will directly affect the development landscape for advanced AI and indirectly bracket the space for any new (AI-specific) governance proposals.[ref 245] This work includes:

• Existing industry norms and practices applied to AI in areas such as release practices around generative AI systems;[ref 246]
• General existing laws and governance regimes which may be extended to or affect AI development, such as anticompetition law;[ref 247] national and international standards;[ref 248] international law norms, treaties, and regimes;[ref 249] and existing global governance institutions.[ref 250]
• AI-specific governance regimes currently under development, such as:
  • EU: the EU AI Act [ref 251] and the AI Liability Directive,[ref 252] amongst others;
  • US: the US AI policy agenda,[ref 253] such as various federal legislative proposals relating to generative AI,[ref 254] or President Biden’s executive order,[ref 255] amongst others.;
  • International: such as the 2019 OECD AI Principles (nonbinding);[ref 256] the 2021 UNESCO Recommendation on the Ethics of Artificial Intelligence (nonbinding);[ref 257] the 2023 G7 Hiroshima guidelines (nonbinding);[ref 258] and the Council of Europe’s draft (framework) Convention on Artificial Intelligence, Human Rights, Democracy and the Rule of Law (potentially binding),[ref 259] amongst others.

3.4. Prevailing barriers to effective AI governance

• Definitional complexities of AI as target for regulation;[ref 260] 
• Potential difficulties around building global consensus given geopolitical stakes and tensions;[ref 261]
• Potential difficulty around building civil society consensus given outstanding disagreements and tensions between different expert communities;[ref 262] 
• Potential challenges around cultivating sufficient state capacity to effectively implement and enforce AI legislation.[ref 263]

3.5. Effects of AI systems on tools of governance

Predicting the impact of future technologies on governance and the ways these could shift the possibility frontier of what kind of regimes will be politically viable and enforceable:

• Effects of AI on general cooperative capabilities;[ref 264]
• Effects of AI on international law creation and enforcement;[ref 265]
• Effects of AI on arms control monitoring.[ref 266]

4. Other lenses on the advanced AI governance problem

Other work aims to derive key strategic lessons for advanced AI governance, not by aiming to empirically map or estimate first-order facts about the key (technical, deployment, or governance) strategic parameters, but rather by drawing indirect (empirical, strategic, and/or normative) lessons from abstract models, historical cases, and/or political theory.

4.1. Lessons derived from theory

Work characterizing the features of advanced AI technology and of its governance challenge, drawing on existing literatures or bodies of theory:

Mapping clusters and taxonomies of AI’s governance problems:

• AI creating distinct types of risk deriving from (1) accidents, (2) misuse, and (3) structure;[ref 267]
• AI creating distinct problem logics across domains: (1) ethical challenges, (2) safety risks, (3) security threats, (4) structural shifts, (5) common goods, and (6) governance disruption;[ref 268]
• AI driving four risk clusters: (1) inequality, turbulence, and authoritarianism; (2) great-power war; (3) the problems of control, alignment, and political order; and (4) value erosion from competition.[ref 269]

Mapping the political features of advanced AI technology:

• AI as general-purpose technology, highlighting radical impacts on economic growth, disruption to existing socio-political relations, and potential for backlash and social conflict;[ref 270] 
• AI as industry-configured general-purpose tech (low fixed costs and private sector dominance), highlighting challenges of rapid proliferation (compared to “prestige,” “public,” or “strategic” technologies);[ref 271]
• AI as information technology, highlighting challenges of increasing returns to scale driving greater income inequality, impacts on broad collective identities as well as community fragmentation, and increased centralization of (cybernetic) control;[ref 272]
• AI as intelligence technology, highlighting challenges of bias, alignment, and control of the principal over the agent;[ref 273]
• AI as regulation-resistant technology, rendering coordinated global regulation difficult.[ref 274]

Mapping the structural features of the advanced AI governance challenge:

• In terms of its intrinsic coordination challenges: as a global public good,[ref 275] as a collective action problem,[ref 276] and as a matter of “existential security”;[ref 277]
• In terms of its difficulty of successful resolution: as a wicked problem[ref 278] and as a challenge akin to “racing through a minefield”;[ref 279]
• In terms of its strategic dynamics: as a technology race,[ref 280] whether motivated by security concerns or by prestige motivations,[ref 281] or as an arms race[ref 282] (but see also critiques of the arms race framing on definitional grounds,[ref 283] on empirical grounds,[ref 284] and on grounds of rhetorical or framing risks[ref 285]); 
• In terms of its politics and power dynamics: as a political economy problem.[ref 286]

Identifying design considerations for international institutions and regimes, from:

• General theory on the rational design of international institutions;[ref 287]
• Theoretical work on the orchestration and organization of regime complexes of many institutions, norms, conventions, etc.[ref 288]

4.2. Lessons derived from models and wargames

Work to derive or construct abstract models for AI governance in order to gather lessons from these for understanding AI systems’ proliferation and societal impacts. This includes models of:

• International strategic dynamics in risky technology races,[ref 289] and theoretical models of the role of information sharing,[ref 290] agreement, or incentive modeling;[ref 291] 
• AI competition and whether and how AI safety insights will be applied under different AI safety-performance tradeoffs,[ref 292] including collaboration on safety as a social dilemma[ref 293] and models of how compute pricing factors affect agents’ spending on safety (“safety tax”) meant to reduce the danger from the new technology;[ref 294]
• The offense-defense balance of increasing investments in technologies;[ref 295]
• The offense-defense balance of scientific knowledge in AI with potential for misuse;[ref 296]
• Lessons from the “epistemic communities” lens, on how coordinated expert networks can shape policy;[ref 297]
• Lessons from wargames and role-playing exercises.[ref 298]

4.3. Lessons derived from history

Work to identify and study relevant historical precedents, analogies, or cases and to derive lessons for (AI) governance.[ref 299] This includes studies where historical cases have been directly applied to advanced AI governance as well as studies where the link has not been drawn but which might nevertheless offer productive insights for the governance of advanced AI.

Lessons from the history of technology development and spread

Historical cases that (potentially) provide insights into when, why, and how new technologies are pursued and developed—and how they subsequently (fail to) spread.

Historical rationales for technology pursuit and development

Historical rationales for actors pursuing large-scale scientific or technology development programs:

• Development of major transformative technologies during wartime: US development of the atom bomb;[ref 300]
• Pursuit of strategically valuable megaprojects: the Apollo Program and the Manhattan Project;[ref 301]
• Technologies pursued for prestige reasons: Ming Dynasty treasure fleets,[ref 302] the US/USSR space race,[ref 303] and the French nuclear weapons program;[ref 304]
• Risk of races being started by possibly incorrect perceptions that a rival is actively pursuing a technology: the Manhattan Project (1939–1945), spurred by the Einstein Letter; the “missile gap” project to build up a US ICBM capability (1957–1962).[ref 305]

Historical strategies of deliberate large-scale technology development projects

Historical strategies for unilateral large-scale technology project development:

• Crash recruitment and resource allocation for a large strategic program: “Operation Paperclip,” the post-WWII effort to recruit 1,600 German scientists and engineers, fast-tracking the US space program as well as several programs aimed at other Cold War weapons of mass destruction;[ref 306]
• Different potential strategies for pursuing advanced strategic technologies: the distinct nuclear proliferation strategies (“hedging, sprinting, sheltered pursuit, hiding”) taken by different countries in pursuing nuclear weapons;[ref 307]
• Government-industry collaborations to boost development of strategic technologies: the 1980’s SEMATECH collaborative research consortium to boost the US semiconductor industry;[ref 308]
• Nations achieving early and sustained unilateral leads in developing key strategic technologies: the US program to develop stealth aircraft;[ref 309]
• Surprisingly rapid leaps from the political decision to run a big technology program to the achievement: Apollo 8 (134 days between NASA decision to go to the moon and launch),[ref 310] UAE’s “Hope” Mars mission (set up its space agency UAESA in 2014, was only able to design its own satellite (KhalifaSat) in 2018, and launched its “Hope” Mars Mission in July 2020, less than six years after establishment),[ref 311] and various other examples including BankAmericard (90 days), P-80 Shooting Star (first USAF jet fighter) (143 days), Marinship (197 days), The Spirit of St. Louis (60 days), the Eiffel Tower (2 years and 2 months), Treasure Island, San Francisco (~2 years), the Alaska Highway (234 days), Disneyland (366 days), the Empire State Building (410 days), Tegel Airport and the Berlin Airlift (92 days),[ref 312] the Pentagon (491 days), Boeing 747 (930 days), the New York Subway (4.7 years), TGV (1,975 days), USS Nautilus (first nuclear submarine) (1,173 days), JavaScript (10 days), Unix (21 days), Xerox Alto (first GUI-oriented computer) (4 months), iPod (290 days), Amazon Prime (6 weeks), Git (17 days), and COVID-19 vaccines (3-45 days).[ref 313]

Historical strategies for joint or collaborative large-scale technology development:

• International “big science” collaborations: CERN, ITER, International Space Station, Human Genome Project,[ref 314] and attempted collaborations on Apollo-Soyuz between the US and Soviet space programs.[ref 315]

Historical instances of sudden, unexpected technological breakthroughs

Historical cases of rapid, historically discontinuous breakthroughs in technological performance on key metrics:

• “Large robust discontinuities” in historical technology performance trends:[ref 316]
  • the Pyramid of Djoser (2650 BC—structure height trends); 
  • the SS Great Eastern (1858—ship size trends); 
  • the first and second telegraphs (1858, 1866—speed of sending a message across the Atlantic Ocean); 
  • the first nonstop transatlantic flight (1919—speed of passenger or military payload travel); 
  • first nuclear weapons (1945—relative effectiveness of explosives); 
  • first ICBM (1958—average speed of military payload); 
  • the discovery of YBa2Cu3O7 as a superconductor (1987—warmest temperature of superconduction).[ref 317] 
• “Bolt-from-the-blue” technology breakthroughs that were held to be unlikely or impossible even shortly before they happened: Invention of flight;[ref 318] of penicillin, nuclear fission, nuclear bombs, or space flight;[ref 319] of internet hyperlinks and effective internet search.[ref 320]

Historical patterns in technological proliferation and take-up

Historical cases of technological proliferation and take-up:[ref 321]

• Patterns in the development, dissemination and impacts of major technological advancements: flight, the telegraph, nuclear weapons, the laser, penicillin, the transistor, and others;[ref 322]
• Proliferation and penetration rates of other technologies in terms of time between invention and widespread use: steam engine (80 years), electricity (40 years), IT (20 years),[ref 323] and mobile phones;
• Role of state “diffusion capacity” in supporting the diffusion or wide adoption of new innovations: the US in the Second Industrial Revolution and the Soviet Union in the early postwar period;[ref 324]
• Role of espionage in facilitating critical technology diffusion: early nuclear proliferation[ref 325] and numerous information leaks in modern IT systems;[ref 326]
• Constrained proliferation of technological insights (even under compromised information security conditions): surprisingly limited track record of bioweapon proliferation: the American, Soviet, Iraqi, South African, and Aum Shinrikyo bioweapon programs ran into a range of problems which resulted in programs that failed if not totally then at least to make effective steps towards weaponization. This suggests that tacit knowledge and organizational conditions can be severely limiting and prevent proliferation even when some techniques are available in the public scientific literature.[ref 327] The (1991–2018) limited success of China in re-engineering US fifth-generation stealth fighters in spite of extensive espionage that included access to blueprints, recruitment of former engineers, and even access to the wreck of a F-117 aircraft that had crashed in Serbia;[ref 328] 
• Various factors contributing to technological delay or restraint with many examples of technologies being slowed or abandoned or having their uptake inhibited, including weapon systems, nuclear power, geoengineering, and genetically modified (GM) crops, as a result of (indirect) regulations, public opposition, and historical contingency;[ref 329]
• Supply chain evolution of previous general-purpose technologies: studies of railroads, electricity, and cloud computing industries, where supply chains were initially vertically integrated but then evolved into a fully disintegrated natural monopoly structure with a handful of primary “upstream” firms selling services to many “downstream” application sectors.[ref 330]

Lessons from the historical societal impacts of new technologies

Historical cases that (potentially) provide insights into when, why, and how new technologies can have (unusually) significant societal impacts or pose acute risks.

Historical cases of large-scale societal impacts from new technologies

Historical cases of large-scale societal impacts from new technologies:[ref 331]

• Impacts of previous narrowly transformative technologies: impact of nuclear weapons on warfare, and electrification of militaries as driver of “general-purpose military transformation”;[ref 332] 
• Impacts of previous general-purpose technologies: general electrification,[ref 333] printing, steam engines, rail transport, motor vehicles, aviation, and computing;[ref 334]
• Impacts of previous “revolutionary” or “radically transformative”[ref 335] technologies: domesticated crops and the steam engine;[ref 336] 
• Impacts of previous information technologies: speech and culture, writing, and the printing press; digital services; and communications technologies;[ref 337]
• Impacts of previous intelligence technologies: price mechanisms in a free market, language, bureaucracy, peer review in science, and evolved institutions like the justice system and law;[ref 338]
• Impacts of previous labor-substitution technologies as they compare to the possible societal impacts of large language models.[ref 339]

Historical cases of particular dangers or risks from new technologies

Historical precedents for particular types of dangers or threat models from technologies:

• Human-machine interface risks and failures around complex technologies: various “normal accidents” in diverse industries and domains, most notably nuclear power;[ref 340]
• Technology misuse risks: the proliferation of easily available hacking tools, such as the “Blackshades Remote Access Tool,”[ref 341] but see also the counterexample of non-use of an (apparent) decisive strategic advantage: the brief US nuclear monopoly;[ref 342] 
• Technological “structural risks”: the role of technologies in lowering the threshold for war initiation such as the alleged role of railways in inducing swift, all-or-none military mobilization schedules and precipitating escalation to World War I.[ref 343]

Historical cases of value changes as a result of new technologies

Historical precedents for technologically induced value erosion or value shifts: 

• Shared values eroded by pressures of global economic competition: “sustainability, decentralized technological development, privacy, and equality”;[ref 344]
• Technological progress biasing the development of states towards welfare-degrading (inegalitarian and autocratic) forms: agriculture, bronze working, chariots, and cavalry;[ref 345]
• Technological progress biasing the development of states towards welfare-promoting forms: ironworking, ramming warships, and industrial revolution;[ref 346]
• Technological progress leading to gradual shifts in societal values: changes in the prevailing technology of energy capture driving changes in societal views on violence, equality, and fairness;[ref 347] demise of dueling and honor culture after (low-skill) pistols replaced (high-skill) swords; changes in sexual morality after the appearance of contraceptive technology; changes in attitudes towards farm animals after the rise of meat replacements; and the rise of the plough as a driver of diverging gender norms.[ref 348]

Historical cases of the disruptive effects on law and governance from new technologies

Historical precedents for effects of new technology on governance tools: 

• Technological changes disrupting or eroding the legal integrity of earlier (treaty) regimes: submarine warfare;[ref 349] implications of cyberwarfare for international humanitarian law;[ref 350] the Soviet Fractional Orbital Bombardment System (FOBS) evading the 1967 Outer Space Treaty’s ban on stationing WMDs “in orbit”;[ref 351] the mid-2010’s US “superfuze” upgrades to its W76 nuclear warheads, massively increasing their counterforce lethality against missile silos without adding a new warhead, missile, or submarine, formally complying with arms control regimes like New START;[ref 352] and various other cases;[ref 353]
• Technologies strengthening international law: satellites strengthening monitoring with treaty compliance,[ref 354] communications technology strengthening the role of non-state and civil-society actors.[ref 355]

Lessons from the history of societal reactions to new technologies

Historical cases that (potentially) provide insights into how societies are likely to perceive, react to, or regulate new technologies.

Historical reactions to and regulations of new technologies 

Historical precedents for how key actors are likely to view, treat, or regulate AI:

• The relative roles of various US actors in shaping the development of past strategic general-purpose technologies: biotech, aerospace tech, and cryptography;[ref 356] 
• Overall US government policy towards perceived “strategic assets”: oil[ref 357] and early development of US nuclear power regulation;[ref 358]
• The historical use of US antitrust law motivated by national security considerations: various cases over the last century;[ref 359]
• Early regulation of an emerging general-purpose technology: electricity regulation in the US;[ref 360] 
• Previous instances of AI development becoming framed as an “arms race” or competition: 1980’s “race” between the US and Japan’s Fifth Generation Computer Systems (FGCS) project;[ref 361]
• Regulation of the “safety” of foundational technology industries, public infrastructures, and sectors: UK regulation of sectors such as medicines and medical devices, food, financial services, transport (aviation & road and rail), energy, and communications;[ref 362]
• High-level state actors buy-in to ambitious early-stage proposals for world control and development of powerful technology: Initial “Baruch Plan” for world control of nuclear weapons (eventually failed);[ref 363] extensive early proposals for world control of airplane technology (eventually failed);[ref 364] and repeated (private and public) US offers to the Soviet Union for a joint US-USSR moon mission, including a 1963 UN General Assembly offer by John F. Kennedy to convert the Apollo lunar landing program into a joint US-Soviet moon expedition (initially on-track, with Nikita Khruschev eager to accept the offer; however, Kennedy was assassinated a week after the offer, the Soviets were too suspicious of similar offers by the Johnson administration, and Khruschev was removed from office by coup in 1964);[ref 365]
• Sustained failure of increasingly more powerful technologies to deliver their anticipated social outcomes: sustained failure of the “Superweapon Peace” idea—the recurring idea that certain weapons of radical destructiveness (nuclear and non-nuclear) may force an end to war by rendering it too destructive to contemplate;[ref 366]
• Strong public and policy reactions to “warning shots” of a technology being deployed: Sputnik launch and Hiroshima bombing;[ref 367]
• Strong public and policy reactions to publicly visible accidents involving a new technology: Three Mile Island meltdown,[ref 368] COVID-19 pandemic,[ref 369] and automotive and aviation industries;[ref 370]
• Regulatory backlash and path dependency: case of genetically modified organism (GMO) regulations in the US vs. the EU;[ref 371]
• “Regulatory capture” and/or influence of industry actors on tech policy, the role of the US military industrial complex in perpetuating the “bomber gap” and “missile gap” myths,[ref 372] and undue corporate influence in the World Health Organisation during the 2009 H1N1 pandemic;[ref 373]
• State norm “antipreneurship” (actions aiming to preserve the prevailing global normative status quo at the global level against proposals for new regulation or norm-setting): US resistance to proposed global restraints on space weapons, between 2000 and the present, utilizing a range of diplomatic strategies and tactics to preserve a permissive international legal framework governing outer space.[ref 374]

Lessons from the history of attempts to initiate technology governance

Historical cases that (potentially) provide insights into when efforts to initiate governance intervention on emerging technologies are likely to be successful and into the efficacy of various pathways towards influencing key actors to deploy regulatory levers in response.

Historical failures to initiate or shape technology governance 

Historical cases where a fear of false positives slowed (plausibly warranted) regulatory attention or intervention: 

• Failure to act in spite of growing evidence: a review of nearly 100 cases of environmental issues where the precautionary principle was raised, concluding that fear of false positives has often stalled action even though (i) false positives are rare and (ii) there was enough evidence to suggest that a lack of regulation could lead to harm.[ref 375] 

Historical cases of excessive hype leading to (possibly) premature regulatory attention or intervention: 

• Premature (and possibly counterproductive) legal focus on technologies that eventually took much longer to develop than anticipated: Weather modification technology,[ref 376] deep seabed mining,[ref 377] self-driving cars,[ref 378] virtual and augmented reality,[ref 379] and other technologies charted under the Gartner Hype Cycle reports.[ref 380]

Historical successes for pathways in shaping technology governance 

Historical precedents for successful action towards understanding and responding to the risks of emerging technologies, influencing key actors to deploy regulatory levers:

• Relative success in long-range technology forecasting: some types of forecasts for military technology that achieved reasonable accuracy decades out;[ref 381]
• Success in anticipatory governance: history of “prescient actions” in urging early action against risky new technologies, such as Leo Szilard’s warning of the dangers of nuclear weapons[ref 382] and Alexander Fleming’s 1945 warning of the risk of antibiotic resistance;[ref 383] 
• Successful early action to set policy for safe innovation in a new area of science:[ref 384] the 1967 Outer Space Treaty, UK’s Warnock Committee and Human Embryology Act 1990, the Internet Corporation for Assigned Names and Numbers (ICANN);
• Governmental reactions and responses to new risks as they emerge: the 1973 Oil Crisis, the 1929–1933 Great Depression,[ref 385] the 2007–2009 financial crisis,[ref 386] the COVID-19 pandemic;[ref 387]
• How effectively other global risks motivated action in response, and how cultural and intellectual orientations influence perceptions: biotechnology, nuclear weapons, global warming, and asteroid collision;[ref 388]
• The impact of cultural media (film, etc.) on priming policymakers to risks:[ref 389] the role of The Day After in motivating Cold War efforts towards nuclear arms control,[ref 390] of the movies Deep Impact and Armageddon in shaping perceptions of the importance of asteroid defense,[ref 391] of the novel Ghost Fleet in shaping Pentagon perceptions of the importance of emerging technologies to war,[ref 392] of Contagion in priming early UK policy responses to COVID-19,[ref 393] of Mission Impossible: Dead Reckoning: Part One in deepening President Biden’s concerns over AI prior to signing a landmark 2023 Executive Order.[ref 394]
• The impact of different analogies or metaphors in framing technology policy:[ref 395] for example, the US military’s emphasis on framing the internet as “cyberspace” (i.e., just another “domain” of conflict) led to strong consequences institutionally (supporting the creation of the US Cyber Command) as well as for how international law has subsequently been applied to cyber operations;[ref 396] 
• The role of “epistemic communities” of experts in advocating for international regulation or agreements,[ref 397] specifically their role in facilitating nonproliferation treaties and arms control agreements for nuclear weapons[ref 398] and anti-ballistic missile systems,[ref 399] as well as the history of the earlier era of arms control agreements;[ref 400] 
• Attempted efforts towards international control of new technology: early momentum but ultimate failure of the Baruch Plan for world control of nuclear weapons[ref 401] and the failure of world control of aviation in 1920s;[ref 402]
• Policy responses to past scientific breakthroughs, and the role of geopolitics vs. expert engagement: the 1967 UN Outer Space Treaty, the UK’s Warnock Committee and the Human Fertilisation and Embryology Act 1990, the establishment of the Internet Corporation for Assigned Names and Numbers (ICANN), and the European ban on GMO crops;[ref 403]
• The role of activism and protests in spurring nonproliferation and moratoria in spurring nuclear nonproliferation agreements and nuclear test bans;[ref 404] the role of activism (in response to “trigger events”) in achieving a de facto moratorium on genetically modified crops in Europe in the late 1990s;[ref 405] in addition, the likely role of protests and public pressure in contributing to abandonment or slowing of various technologies from geoengineering experiments to nuclear weapons, CFCs, and nuclear power;[ref 406]
• The role of philanthropy and scientists in fostering Track-II diplomacy initiatives: the Pugwash conferences.[ref 407]

Lessons from the historical efficacy of different governance levers

Historical cases that (potentially) provide insights into when different societal (legal, regulatory, and governance) levers have proven effective in shaping technology development and use in desired directions.

Historical failures of technology governance levers 

Historical precedents for failed or unsuccessful use of various (domestic and/or international) governance levers for shaping technology:

• Mixed-success use of soft-law governance tools for shaping emerging technologies: National Telecommunications and Information Administration discussions on mobile app transparency, drone privacy, facial recognition, YourAdChoices, UNESCO declaration on genetics and bioethics, Environmental Management System (ISO 14001), Sustainable Forestry Practices by the Sustainable Forestry Initiative and Forest Stewardship Council, and Leadership in Energy and Environmental Design.[ref 408]
• Failed use of soft-law governance tools for shaping emerging technologies: Children’s Online Privacy Protection Rule, Internet Content Rating Association, Platform for Internet Content Selection, Platform for Privacy Preferences, Do Not Track system, and nanotechnology voluntary data call-in by Australia, the US, and the UK;[ref 409]
• Failures of narrowly technology-focused approaches to safety engineering: failure of narrow technology-focused approaches to the design of safe cars and in the design and calibration of pulse oximeters during the COVID pandemic, which were mismatched to—and therefore led to dangerous outcomes for—female drivers and darker-skinned patients, respectively, highlighting the role of incorporating human, psychological, and other disciplines;[ref 410]
• Failures of information control mechanisms at preventing proliferation: selling of nuclear secrets by A.Q. Khan network,[ref 411] limited efficacy of Cold War nuclear secrecy regimes at meaningfully constraining proliferation of nuclear weapons,[ref 412] track record of major leaks and hacks of digital information, 2005–present;[ref 413]
• Failure to transfer (technological) safety techniques, even to allies: in the late 2000s, the US sought to help provide security assistance to Pakistan to help safeguard the Pakistani nuclear arsenal but was unable to transfer permissive action link (PAL) technologies because of domestic legal barriers that forbade export to states that were not part of the Nuclear Non-Proliferation Treaty (NPT);[ref 414]
• Degradation of previously established export control regimes: Cold War-era US high performance computing export controls struggled to be updated sufficiently quickly to keep pace with hardware advancements.[ref 415] The US initially treated cryptography as a weapon under export control laws, meaning that encryption systems could not be exported for commercial purposes even to close allies and trading partners; however, by the late 1990s, several influences—including the rise of open-source software and European indignation at US spying on their communications—led to new regulations that allowed cryptography to be exported with minimal government interference;[ref 416]
• “Missed opportunities” for early action against anticipated risks: mid-2000s effort to put “killer robots” on humanitarian disarmament issue agenda, which failed as these were seen as “too speculative”;[ref 417] 
• Mixed success of scientific and industry self-regulation: the Asilomar Conference, the Second International Conference on Synthetic Biology, and 2004–2007 failed efforts to develop guidelines for nanoparticles;[ref 418]
• Sustained failure to establish treaty regimes: various examples, including the international community spending nearly 20 years since 2004 negotiating a new treaty for Biodiversity Beyond National Jurisdiction;[ref 419]
• Unproductive locking-in of insufficient, “empty” institutions, “face-saving” institutions, or gridlocked mechanisms: history of states creating suboptimal, ill-designed institutions—such as the United Nations Forum on Forests, the Copenhagen Accord on Climate Change, the UN Commission on Sustainable Development, and the 1980 UN Convention on Certain Conventional Weapons—with mandates that may deprive them of much capacity for policy formulation or implementation;[ref 420]
• Drawn-out contestation of hierarchical and unequal global technology governance regimes: the Nuclear Non-Proliferation Treaty regime has seen cycles of contestation and challenge by other states;[ref 421]
• Failures of non-inclusive club governance approaches to nonproliferation: the Nuclear Security Summits (NSS) (2012, 2014, 2016) centered on high-level debates over the stocktaking and securing of nuclear materials. These events saw a constrained list of invited states; as a result, the NSS process was derailed because procedural questions over who was invited or excluded came to dominate discussions (especially at the 2016 Vienna summit), politicizing what had been a technical topic and hampering the extension and take-up of follow-on initiatives by other states.[ref 422] 

Historical successes of technology governance levers 

Historical precedents for successful use of various governance levers at shaping technology:

• Effective scientific secrecy around early development of powerful new technologies: early development of the atomic bomb[ref 423] and early computers (Colossus and ENIAC).[ref 424]
• Successes in the oversight of various safety-critical technologies: track record of “High Reliability Organisations”[ref 425] in addressing emerging risks after initial incidents to achieve very low rates of errors, such as in air traffic control systems, naval aircraft carrier operations,[ref 426] the aerospace sector, construction, and oil refineries;[ref 427] 
• Successful development of “defense in depth”[ref 428] interventions to lower the risks of accident in specific industries: safe operation of nuclear reactors, chemical plants, aviation, space vehicles, cybersecurity and information security, software development, laboratories studying dangerous pathogens, improvised explosive devices, homeland security, hospital security, port security, physical security in general, control system safety in general, mining safety, oil rig safety, surgical safety, fire management, and health care delivery,[ref 429] and lessons from defense-in-depth frameworks developed in cybersecurity for frontier AI risks;[ref 430]
• Successful safety “races to the top” in selected industries: Improvements in aircraft safety in the aviation sector;[ref 431]
• Successful use of risk assessment techniques in safety-critical industries: examination of popular risk identification techniques (scenario analysis, fishbone method, and risk typologies and taxonomies), risk analysis techniques (causal mapping, Delphi technique, cross-impact analysis, bow tie analysis, and system-theoretic process analysis), and risk evaluation techniques (checklists and risk matrices) used in established industries like finance, aviation, nuclear, and biolabs, and how these might be applied in advanced AI companies;[ref 432]
• Susceptibility of different types of digital technologies to (global) regulation: relative successes and failures of global regulation of different digital technologies that are (1) centralized and clearly material (e.g., submarine cables), (2) decentralized and clearly material (e.g., smart speakers); (3) centralized and seemingly immaterial (e.g., search engines), and (4) decentralized and seemingly immaterial (e.g., Bitcoin protocol);[ref 433] 
• Use of confidence-building measures to stabilize relations and expectations: 1972 Incidents at Sea Agreement[ref 434] and the 12th–19th century development of Maritime Prize Law;[ref 435]
• Successful transfer of developed safety techniques, even to adversaries: the US “leaked” PAL locks on nuclear weapons to the Soviet Union;[ref 436] 
• Effective nonproliferation regimes: for nuclear weapons, a mix of norms, treaties, US “strategies of inhibition,”[ref 437] supply-side export controls,[ref 438] and domestic politics factors[ref 439] have produced an imperfect but remarkably robust track record of nonproliferation.[ref 440] Indeed, based on IAEA databases there have historically been 74 states that decided to build or use nuclear reactors, of which 69 have at some time been considered potentially able to pursue nuclear weapons, and of which 10 went nuclear and 7 ran but abandoned a program, and for 14–23, evidence exists of a considered decision not to use their infrastructure to pursue nuclear weapons;[ref 441]
• General design lessons from existing treaty regimes: drawing insights from the design and efficacy of a range of treaties—including the Single Convention on Narcotic Drugs (SCND), the Vienna Convention on Psychotropic Substances (VCPS), the Convention Against Illicit Trafficking of Narcotic Drugs and Psychotropic Substances (CAIT), the Montreal Protocol on Substances that Deplete the Ozone Layer, the Cartagena Protocol on Biosafety to the Convention on Biological Diversity, the Biological Weapons Convention (BWC), the Treaty on the Non-Proliferation of Nuclear Weapons (NPT), the Convention on Nuclear Safety, the Convention on International Trade in Endangered Species (CITES), the Basel Convention on the Control of Transboundary Movements of Hazardous Wastes and their Disposal, and the Bern Convention on the Conservation of European Wildlife and Natural Habitats—to derive design lessons for a global regulatory system dedicated to the regulation of safety concerns from high-risk AI;[ref 442]
• Effective use of international access and benefit distribution mechanisms in conjunction with proliferation control measures: the efficacy of the IAEA’s “dual mandate” to enable the transfer of peaceful nuclear technology whilst seeking to curtail its use for military purposes;[ref 443]
• Effective monitoring and verification (M&V) mechanisms in arms control regimes: M&V implementation across three types of nuclear arms control treaties: nonproliferation treaties, US-USSR/Russia arms limitation treaties, and nuclear test bans;[ref 444]
• Scientific community (temporary) moratoria on research: the Asilomar Conference[ref 445] and the H5N1 gain-of-function debate;[ref 446]
• Instances where treaty commitments, institutional infighting, or bureaucratic politics contributed to technological restraint: a range of cases resulting in cancellation of weapon systems development, including nuclear-ramjet powered cruise missiles, “continent killer” nuclear warheads, nuclear-powered aircraft, “death dust” radiological weapons, various types of anti-ballistic-missile defense, and many others.[ref 447]
• International institutional design lessons from successes and failures in other areas: global governance successes and failures in the regime complexes for environment, security, and/or trade;[ref 448]
• Successful use of soft-law governance tools for shaping emerging technologies: Internet Corporation for Assigned Names and Numbers, Motion Picture Association of America, Federal Trade Commission consent decrees, Federal Communications Commission’s power over broadcaster licensing, Entertainment Software Rating Board, NIST Framework for Improving Critical Infrastructure Cybersecurity, Asilomar rDNA Guidelines, International Gene Synthesis Consortium, International Society for Stem Cell Research Guidelines, BASF Code of Conduct, Environmental Defense Fund, and DuPont Risk Framework;[ref 449]
• Successful use of participatory mechanisms in improving risk assessment: use of scenario methods and risk assessments in climate impact research.[ref 450]

4.4. Lessons derived from ethics and political theory 

Mapping the space of principles or criteria for “ideal AI governance”:[ref 451]

• Mapping broad normative desiderata for good governance regimes for advanced AI,[ref 452] either in terms of outputs or in terms of process;[ref 453]
• Understanding how to weigh different good outcomes post-TAI-deployment;[ref 454]
• Understanding the different functional goals and tradeoffs in good international institutional design.[ref 455]

II. Option-identifying work: Mapping actors and affordances

Strategic clarity requires an understanding not just of the features of the advanced AI governance problem, but also of the options in response. 

This entails mapping the range of possible levers that could be used in response to this problem. Critically, this is not just about speculating about what governance tools we may want to put in place for future advanced AI systems mid-transition (after they have arrived). Rather, there might be actions we could take in the “pre-emergence” stage to adequately prepare ourselves.[ref 456] 

Within the field, there has been extensive work on options and areas of intervention. Yet there is no clear, integrated map of the advanced AI governance landscape and its gaps. Sam Clarke proposes that there are different ways of carving up the landscape, such as based on different types of interventions, different geographic hubs, or “Theories of Victory.”[ref 457] To extend this, one might segment the advanced AI governance solution space along work which aims to identify and understand, in turn:[ref 458]

• Key actors that will likely (be in a strong position to) shape advanced AI;
• Levers of influence (by which these actors might shape advanced AI);
• Pathways towards influencing these actors to deploy their levers well.[ref 459]

1. Potential key actors shaping advanced AI

In other words, whose decisions might especially affect the development and deployment of advanced AI, directly or indirectly, such that these decisions should be shaped to be as beneficial as possible?

Some work in this space explores the relative importance of (the decisions of) different types of key actors: 

• The roles of state vs. firms vs. AI researchers in shaping AI policy;[ref 462] 
• Role of “epistemic communities” of scientific experts,[ref 463] especially members of the AI research community;[ref 464]
• The role of different potentially relevant stakeholders for responsible AI systems across its development chain, from individual stakeholders to organizational stakeholders to national/international stakeholders;[ref 465]
• The relative role of expert advice vs. public pressure in shaping policymakers’ approach to AI;[ref 466]
• Role of different actors in and around the corporation in shaping lab policy,[ref 467] including actors within the lab (e.g., senior management, shareholders, AI lab employees, and employee activists)[ref 468] and actors outside the lab (e.g., corporate partners and competitors, industry consortia, nonprofit organizations, the public, the media, and governments).[ref 469]

Other work focuses more specifically on mapping particular key actors whose decisions may be particularly important in shaping advanced AI outcomes, depending on one’s view of strategic parameters. 

The following list should be taken more as a “landscape” review than a literature review, since coverage of different actors differs amongst papers. Moreover, while the list aims to be relatively inclusive of actors, it is clear that the (absolute and relative) importance of each of these actors obviously differs hugely between worldviews and approaches.

1.1. AI developer (lab & tech company) actors 

Leading AI firms pursuing AGI: 

• OpenAI, 
• DeepMind, 
• Anthropic, 
• Aleph Alpha, 
• Adept, 
• Cohere, 
• Inflection, 
• Keen, 
• xAI.[ref 470]

Chinese labs and institutions researching “general AI”;

• Baidu Research, 
• Alibaba DAMO Academy, 
• Tencent AI Lab, 
• Huawei, 
• JD Research Institute, 
• Beijing Institute for General Artificial Intelligence; 
• Beijing Academy of Artificial Intelligence, etc.[ref 471]

Large tech companies[ref 472] that may take an increasingly significant role in AGI research:

• Microsoft, 
• Google,
• Facebook,
• Amazon.

Future frontier labs, currently not known but to be established/achieve prominence (e.g., “Magma”[ref 473]).

1.2. AI services & compute hardware supply chains 

AI services supply chain actors:[ref 474] 

• Cloud computing providers:[ref 475]
  • Globally: Amazon Web Services (32%), Microsoft Azure (21%), and Google Cloud (8%); IBM;
  • Chinese market: Alibaba, Huawei, and Tencent.

Hardware supply chain industry actors:[ref 476] 

• Providers of optical components to photolithography machine manufacturers:
  • Carl Zeiss AG [Germany], a key ASML supplier of optical lenses;[ref 477]
• Producers of extreme ultraviolet (EUV) photolithography machines: 
  • ASML [The Netherlands].[ref 478]
• Photoresist processing providers:
  • Asahi Kasei and Tokyo Ohka Kogyo Co. [Japan].[ref 479]
• Advanced chip manufacturing: 
  • TMSC [Taiwan]; 
  • Intel [US];
  • Samsung [South Korea].
• Semiconductor intellectual property owners and chip designers:
  • Arm [UK];
  • Graphcore [UK].
• DRAM integrated circuit chips: 
  • Samsung (market share 44%) [South Korea]; 
  • SK hynix (27%) [South Korea]; 
  • Micron (22%) [US]. 
• GPU providers: 
  • Intel (market share 62%) [US]; 
  • AMD (18%) [US];
  • NVIDIA (20%) [US].

1.3. AI industry and academic actors 

Industry bodies:

• Partnership on AI;
• Frontier Model Forum;[ref 480]
• ML Commons;[ref 481]
• IEEE (Institute of Electrical and Electronics Engineers) + IEEE-SA (standards body);
• ISO (and IEC).

Standard-setting organizations:

• US standard-setting organizations (NIST);
• European Standards Organizations (ESOs), tasked with setting standards for the EU AI Act: the European Committee for Standardisation (CEN), European Committee for Electrotechnical Standardisation (CENELEC), and European Telecommunications Standards Institute (ETSI);[ref 482]
• VDE (influential German standardization organization).[ref 483]

Software tools & community service providers:

• arXiv; 
• GitHub; 
• Colab; 
• Hugging Face.

Academic communities:

• Scientific ML community;[ref 484]
• AI conferences: NeurIPS, AAAI/ACM, ICLR, IJCAI-ECAI, AIES, and FAccT, etc.;
• AI ethics community and various subcommunities;
• Numerous national-level academic or research institutes. 

Other active tech community actors:

• Open-source machine learning software community;[ref 485]
• “Open”/diffusion-encouraging[ref 486] AI community (e.g., Stability.ai, Eleuther.ai);[ref 487]
• Hacker communities;
• Cybersecurity and information security expert communities.[ref 488]

1.4. State and governmental actors 

Various states, and their constituent (government) agencies or bodies that are, plausibly will be, or potentially could be moved to be in powerful positions to shape the development of advanced AI.

The United States

Key actors in the US:[ref 489] 

• Executive Branch actors;[ref 490]
• Legislative Branch;[ref 491]
• Judiciary;[ref 492]
• Federal agencies;[ref 493]
• Intelligence community;[ref 494] 
• Independent federal agencies;[ref 495]
• Relevant state and local governments, such as the State of California (potentially significant extraterritorial regulatory effects),[ref 496] State of Illinois and State of Texas (among first states to place restrictions on biometrics), etc.

China

Key actors in China:[ref 497]

• 20th Central Committee of the Chinese Communist Party; 
• China’s State Council;
• Bureaucratic actors engaged in AI policy-setting;[ref 498] 
• Actors and institutions engaged in track-II diplomacy on AI.[ref 499] 

The EU

Key actors in the EU:[ref 500]

• European Commission;
• European Parliament;
• Scientific research initiatives and directorates;[ref 501]
• (Proposed) European Artificial Intelligence Board and notified bodies.[ref 502]

The UK

Key actors in the UK:[ref 503]

• The Cabinet Office;[ref 504]
• Foreign Commonwealth and Development Office (FCDO);
• Ministry of Defence (MoD);[ref 505]
• Department for Science, Innovation and Technology (DSIT);[ref 506]
• UK Parliament;[ref 507]
• The Digital Regulators Cooperation Forum;
• Advanced Research and Invention Agency (ARIA).

Other states with varying roles

Other states that may play key roles because of their general geopolitical influence, AI-relevant resources (e.g., compute supply chain and significant research talent), or track record as digital norm setters: 

• Influential states: India, Russia, and Brazil;
• Significant AI research talent: France, and Canada;
• Hosting nodes in the global hardware supply chain: US (NVIDIA), Taiwan (TSMC), South Korea (Samsung), the Netherlands (ASML), Japan (photoresist processing), UK (Arm), and Germany (Carl Zeiss AG);
• Potential (regional) neutral hubs: Singapore[ref 508] and Switzerland;[ref 509]
• Global South coalitions: states from the Global South[ref 510] and coalitions of Small Island Developing States (SIDS);[ref 511]
• Track record of (digital) norm-setters: Estonia and Norway.[ref 512]

1.5. Standard-setting organizations

International standard-setting institutions:[ref 513] 

• ISO;
• IEC;
• IEEE;
• CEN/CENELEC;
• VDE (Association for Electrical, Electronic & Information Technologies) and its AI Quality & Testing Hub.[ref 514]

1.6. International organizations

Various United Nations agencies:[ref 515]

• ITU;[ref 516]
• UNESCO;[ref 517]
• Office of the UN Tech Envoy (conducting the process leading to the Global Digital Compact in 2024);
• UN Science, Technology, and Innovation (STI) Forum;
• UN Executive Office of the Secretary-General;
• UN General Assembly;
• UN Security Council (UNSC);
• UN Human Rights Council;[ref 518]
• Office of the High Commissioner on Human Rights;[ref 519] 
• UN Chief Executives Board for Coordination;[ref 520] 
• Secretary-General’s High-Level Advisory Board on Effective Multilateralism (HLAB);
• Secretary-General’s High-Level Advisory Body on Artificial Intelligence (“AI Advisory Body”).[ref 521] 

Other international institutions already engaged on AI in some capacity[ref 522] (in no particular order):

• OECD;[ref 523]
• Global Partnership on AI;
• G7;[ref 524]
• G20;[ref 525]
• Council of Europe (Ad Hoc Committee on Artificial Intelligence (CAHAI));[ref 526]
• NATO;[ref 527]
• AI Partnership for Defense;[ref 528]
• Global Road Traffic Forum;[ref 529]
• International Maritime Organisation;
• EU-US Trade and Technology Council (TTC);[ref 530] 
• EU-India Trade and Technology Council; 
• Multi-stakeholder fora: World Summit on the Information Society (WSIS), Internet Governance Forum (IGF), Global Summit on AI for Good,[ref 531] and World Economic Forum (Centre for Trustworthy Technology).

Other international institutions not yet engaged on AI:

• International & regional courts: International Criminal Court (ICC), International Court of Justice (ICJ), and European Court of Justice.

1.7. Public, Civil Society, & media actors 

Civil society organizations:[ref 532] 

• Gatekeepers engaged in AI-specific norm-setting and advocacy: Human Rights Watch, Campaign to Stop Killer Robots,[ref 533] and AlgorithmWatch;[ref 534]
• Civilian open-source intelligence (OSINT) actors engaged in monitoring state violations of human rights and international humanitarian law:[ref 535] Bellingcat, NYT Visual Investigation Unit, CNS (Arms Control Wonk), Middlebury Institute, Forensic Architecture, BBC Africa Eye, Syrian Archive, etc. 
• Military AI mediation: Centre for Humanitarian Dialogue and Geneva Centre for Security Policy.[ref 536] 

Media actors:

• Mass media;[ref 537]
• Tech media;
• “Para-scientific media.”[ref 538]

Cultural actors: 

• Film industry (Hollywood, etc.);
• Influential and widely read authors.[ref 539]

2. Levers of governance (for each key actor)

That is, how might each key actor shape the development of advanced AI?

Research in this field includes analysis of different types of tools (key levers or interventions) available to different actors to shape advanced AI development and use.[ref 541]

2.1. AI developer levers

Developer (intra-lab)-level levers:[ref 542]

• Levers for adequate AI model evaluation and technical safety testing:[ref 543] decoding; limiting systems, adversarial training; throughout-lifecycle test, evaluation, validation, and verification (TEVV) policies;[ref 544] internal model safety evaluations;[ref 545] and risk assessments;[ref 546]
• Levers for safe risk management in AI development process: Responsible Scaling Policies (RSPs),[ref 547] the Three Lines of Defense (3LoD) model,[ref 548] organizational and operational criteria for adequately safe development,[ref 549] and “defense in depth” risk management procedures;[ref 550] 
• Levers to ensure cautious overall decision-making: ethics and oversight boards;[ref 551] corporate governance policies that support and enable cautious decision-making,[ref 552] such as establishing an internal audit team;[ref 553] and/or incorporating as a Public Benefit Corporation to allow the board of directors to balance stockholders’ pecuniary interests against the corporation’s social mission; 
• Levers to ensure operational security: information security best practices[ref 554] and structured access mechanisms[ref 555] at the level of cloud-based AI service interfaces;
• Policies for responsibly sharing safety-relevant information: information-providing policies to increase legibility and compliance: model cards;[ref 556]
• Policies to ensure organization can pace and/or pause capability research:[ref 557] Board authority to pause research and channels to invite external AI scientists to review alignment of systems.[ref 558] 

Developer external (unilateral) levers:

• Use of contracts and licensing to attempt to limit uses of AI and its outputs (e.g., the Responsible AI Licenses (RAIL) initiative);[ref 559]
• Voluntary safety commitments;[ref 560] 
• Norm entrepreneurship (i.e., lobbying, public statements, or initiatives that signal public concern and/or dissatisfaction with an existing state of affairs, potentially alerting others to the existence of a shared complaint and facilitating potential “norm cascades” towards new expectations or collective solutions).[ref 561]

2.2. AI industry & academia levers 

Industry-level (coordinated inter-lab) levers:

• Self-regulation;[ref 562]
• Codes of conduct;
• AI ethics principles;[ref 563]
• Professional norms;[ref 564]
• AI ethics advisory committees;[ref 565]
• Incident databases;[ref 566]
• Institutional, software, and hardware mechanisms for enabling developers to make verifiable claims;[ref 567]
• Bug bounties;[ref 568]
• Evaluation-based coordinated pauses;[ref 569] 
• Other inter-lab cooperation mechanisms:[ref 570]
  • Assist Clause;[ref 571]
  • Windfall Clause;[ref 572]
  • Mutual monitoring agreements (red-teaming, incident-sharing, compute accounting, and seconding engineers);
  • Communications and heads-up;
  • Third-party auditing;
  • Bias and safety bounties;
  • Secure compute enclaves;
  • Standard benchmarks & audit trails;
  • Publication norms.[ref 573]

Third-party industry actors levers:

• Publication reviews;[ref 574]
• Certification schemes;[ref 575]
• Auditing schemas.[ref 576]

Scientific community levers:

• Institutional Review Boards (IRBs);[ref 577]
• Conference or journal pre-publication impact assessment requirements;[ref 578] academic conference practices;[ref 579]
• Publication and model sharing and release norms;[ref 580]
• Benchmarks;[ref 581]
• Differential technological development (innovation prizes);[ref 582]
• (Temporary) moratoria.[ref 583]

2.3. Compute supply chain industry levers

Global compute industry-level levers:[ref 584] 

• Stock-and-flow accounting; 
• Operating licenses;
• Supply chain chokepoints;[ref 585]
• Inspections
• Passive architectural on-chip constraints (e.g., performance caps)
• Active architectural on-chip constraints (e.g., shutdown mechanisms)

2.4. Governmental levers

We can distinguish between general governmental levers and the specific levers available to particular key states.

General governmental levers[ref 586]

Legislatures’ levers:[ref 587]

• Create new AI-specific regimes, such as:
  • Horizontal risk regulation;[ref 588] 
  • Industry-specific risk regulatory regimes;
  • Permitting, licensing, and market gatekeeping regimes;[ref 589]
  • Bans or moratoria;
  • Know-Your-Customer schemes.[ref 590]
• Amend laws to extend or apply existing regulations to AI:[ref 591]
  • Domain/industry-specific risk regulations;
  • Competition/antitrust law,[ref 592] including doctrines around merger control, abuse of dominance, cartels, and collusion;  agreements on hardware security; and state aid;
  • Liability law;[ref 593]
  • Insurance law;[ref 594]
  • Contract law;[ref 595]
  • IP law;[ref 596]
  • Copyright law (amongst others through its impact on data scraping practices);[ref 597]
  • Criminal law;[ref 598] 
  • Privacy and data protection law (amongst others through its impact on data scraping practices);
  • Public procurement law and procurement processes.[ref 599]

Executive levers:

• Executive orders;
• Foreign investment restrictions;
• AI R&D funding strategies;[ref 600]
• Nationalization of firms;
• Certification schemes;
• Various tools of “differential technology development”:[ref 601] policies for preferential advancement of safer AI architectures (funding and direct development programs, government prizes, advanced market commitments, regulatory requirements, and tax incentives)[ref 602] and policies for slowing down research lines towards dangerous AI architectures (moratoria, bans, defunding, divestment, and/or “stage-gating” review processes);[ref 603]
• Foreign policy decisions, such as initiating multilateral treaty negotiations.

Judiciaries’ levers:

• Judicial decisions handed down on cases involving AI that extend or apply existing doctrines to AI, shaping economic incentives and setting precedent for regulatory treatment of advanced AI, such as the US Supreme Court ruling on Gonzalez v. Google, which has implications for whether algorithmic recommendations will receive full Section 230 protections;[ref 604]
• Judicial review, especially of drastic executive actions taken in response to AI risk scenarios;[ref 605]
• Judicial policymaking, through discretion in evaluating proportionality or balancing tests.[ref 606]

Expert agencies’ levers:

• A mix of features of other actors, from setting policies to adjudicating disputes to enforcing decisions;[ref 607]
• Create or propose soft law.[ref 608]

Ancillary institutions:

• Improved monitoring infrastructures;[ref 609]
• Provide services in terms of training, insurance, procurement, identification, archiving, etc.[ref 610]

Foreign Ministries/State Department:

• Set activities and issue agendas in global AI governance institutions;
• Bypass or challenge existing institutions by engaging in “competitive regime creation,”[ref 611] “forum shopping,”[ref 612] or the strategic creation of treaty conflicts;[ref 613]
• Initiate multilateral treaty negotiations;
• Advice policymakers about the existence and meaning of international law and which obligations these impose;[ref 614]
• Conduct state behavior around AI issues (in terms of state policy, and through discussion of AI issues in national legislation, diplomatic correspondence, etc.) in such a way as to contribute to the establishment of binding customary international law (CIL).[ref 615]

Specific key governments levers

Levers available to specific key governments:

US-specific levers:[ref 616] 

• AI-specific regulations, such as the AI Bill of Rights;[ref 617] Algorithmic Accountability Act;[ref 618] 2023 Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence;[ref 619] and various currently pending federal legislative proposals for regulating generative and/or frontier AI;[ref 620]
• General levers,[ref 621] such as federal R&D funding, foreign investment restrictions, export controls,[ref 622] visa vetting, expanded visa pathways, secrecy orders, voluntary screening procedures, use of the Defense Production Act,[ref 623] antitrust enforcement, the “Born Secret” Doctrine, nationalization of companies or compute hardware, various Presidential Emergency powers,[ref 624] etc.

EU-specific levers: 

• AI-specific regulations, including:
  • The AI Act, which will have direct regulatory effects[ref 625] but may also exert extraterritorial impact as part of a “Brussels Effect”;[ref 626]
  • Standard-setting by European Standards Organizations (ESOs);[ref 627]
  • AI Liability Directive.[ref 628]

China-specific levers:

• AI-specific regulations;[ref 629]
• Standards;[ref 630]
• Activities in global AI governance institutions.[ref 631]

UK-specific levers:[ref 632] 

• National Security and Investment Act 2021;
• Competition Law: 1998 Competition Act;
• Export Control legislation;
• Secrecy orders.

2.5. Public, civil society & media actor levers

Civil Society/activist movement levers:[ref 633]

• Lab-level (internal) levers:
  • Shareholder activism, voting out CEOs; 
  • Unions and intra-organizational advocacy, strikes, and walkouts;[ref 634]
  • Capacity-building of employee activism via recruitment, political education, training, and legal advice.
• Lab-level (external) levers:
  • Stigmatization of irresponsible practices;[ref 635]
  • Investigative journalism, awareness-raising of scandals and incidents, hacking and leaks, and  whistleblowing;
  • Impact litigation[ref 636] and class-action lawsuits;[ref 637]
  • Public protest[ref 638] and direct action (e.g., sit-ins).
• Industry-level levers:
  • Norm advocacy and lobbying;
  • Open letters and statements;
  • Mapping and highlighting (compliance) performance of companies; establishing metrics, indexes, and prizes; and certification schemes.[ref 639]
• Public-focused levers:
  • Media content creation;[ref 640]
  • Boycott and divestment;
  • Shaming of state noncompliance with international law;[ref 641]
  • Emotional contagion—shaping and disseminating of public emotional dynamics or responses to a crisis.[ref 642]
• Creating alternatives:
  • Public interest technology research;
  • Creating alternative (types of) institutions[ref 643] and new AI labs.
• State-focused levers:
  • Monitor compliance with international law.[ref 644]

2.6. International organizations and regime levers 

International standards bodies’ levers:

• Set technical safety and reliability standards;[ref 645]
• Undertake “para-regulation,” setting pathways for future regulation not by imposing substantive rules but rather by establishing foundational concepts or terms.[ref 646]

International regime levers:[ref 647]

• Setting or shaping norms and expectations: 
  • Setting, affirming, and/or clarifying states’ obligations under existing international law principles;
  • Set fora and/or agenda for negotiation of new treaties or regimes in various formats, such as:
    • Broad framework conventions;[ref 648]
    • Nonproliferation and arms control agreements;[ref 649]
    • Export control regimes.[ref 650]
  • Create (technical) benchmarks and focal points for decision-making by both states and non-state actors;[ref 651]
  • Organize training and workshops with national officials.
• Coordinating behavior; reducing uncertainty, improving trust:
  • Confidence-building measures;[ref 652]
  • Review conferences (e.g., BWC);
  • Conferences of parties (e.g., UNFCCC);
  • Establishing information and benefit-sharing mechanisms.
• Creating common knowledge or shared perceptions of problems; establish “fire alarms”:
  • Intergovernmental scientific bodies (e.g., Intergovernmental Panel on Climate Change (IPCC) and Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services (IPBES));
  • International warning systems (e.g., WHO’s “public health emergency of international concern” mechanism).
• Adjudicating and arbitrating state disagreements over application of policies, resolving tensions or crises for regimes:
  • Arbitral bodies (e.g., WTO Appellate Body);
  • Adjudicatory tribunals (e.g., ICJ);
  • Treaty bodies (e.g., Human Rights Committee);
  • Other dispute resolution mechanisms (e.g., BWC or Resolution 1540 allowing complaints to be lodged at the UNSC).
• Establishing material constraints:
  • Supply-side material proliferation controls (e.g., stock-and-flow accounting and trade barriers);
  • Fair and equitable treatment standards in international investment law.
• Monitoring state compliance:
  • Inspection regimes;
  • Safeguards;
  • National contributions;
  • Network of national contact points.
• Sanctioning noncompliance:
  • Inducing direct costs through sanctions;
  • Inducing reputational costs,[ref 653] in particular through shaming.[ref 654]

2.7. Future, new types of institutions and levers

Novel governance institutions and innovations:

• “Regulatory markets” and private regulatory authorities;[ref 655]
• New monitoring institutions and information markets;[ref 656]
• Quadratic voting and radical markets[ref 657]
• Blockchain smart contracts.[ref 658]

3. Pathways to influence (on each key actor)

That is, how might concerned stakeholders ensure that key actors use their levers to shape advanced AI development in appropriate ways?

This includes research on the different pathways by which the use of these above levers might be enabled, advocated for, and implemented (i.e., the tools available to affect the decisions by key actors).

This can draw on mappings and taxonomies: “A Map to Navigate AI Governance”[ref 660] “The Longtermist AI Governance Landscape”.[ref 661] 

3.1. Pathways to directly shaping advanced AI systems’ actions through law

Directly shaping advanced AI actions through law (i.e., legal systems and norms as an anchor or lodestar for technical alignment approaches):

• “Law-following AI”;[ref 662]
• Encode “incomplete contracting” as a framework for AI alignment;[ref 663]
• Negative human rights as technical safety constraint for minimal alignment;[ref 664]
• Human rights norms as a benchmark for maximal alignment;[ref 665]
• Encode fiduciary duties towards users into AI systems;[ref 666]
• Mandatory on-chip controls (monitoring and remote shutdown);
• Legal informatics approach to alignment.[ref 667]

3.2. Pathways to shaping governmental decisions

Shaping governmental decisions around AI levers at the level of:

• Legislatures:
  • Advocacy within the legislative AI policymaking process.[ref 668]
• Executives:
  • Serve as high-bandwidth policy advisor;[ref 669]
  • Provide actionable technical information;[ref 670]
  • Shape, provide, or spread narratives,[ref 671] ideas, “memes,”[ref 672] framings, or (legal) analogies[ref 673] for AI governance.
• Clarify or emphasize established principles within national law (e.g., precautionary principle and cost-benefit analysis[ref 674]) and/or state obligations under international law (e.g., customary international law,[ref 675] IHRL,[ref 676] etc.). 

3.3. Pathways to shaping court decisions

Shaping court decisions around AI systems that set critical precedent for the application of AI policy to advanced AI:

• Advance legal scholarship with new arguments, interpretations, or analogies and metaphors for AI technology;[ref 677]
• Clarifying the “ordinary meaning” of key legal terms around AI;[ref 678]
• Judge seminars and training courses;[ref 679]
• Online information repositories.[ref 680]

3.4. Pathways to shaping AI developers’ decisions

Shaping individual lab decisions around AI governance:

• Governmental regulations (e.g., industry risk, liability, criminal, etc.);
• Institutional design choices: establish rules in the charter that enable the board of directors to make more cautious or pro-social choices,[ref 681] and establish an internal AI ethics board[ref 682] or internal audit functions;[ref 683]
• Campaigns or resources to educate researchers about AI risk, making AI safety research more concrete and legible, and/or creating common knowledge about researchers’ perceptions of and attitudes towards these risks;[ref 684]
• Employee activism and pressure,[ref 685] and documented communications of risks by employees (which make companies more risk averse because they are more likely to be held liable in court);[ref 686]
• Human rights norms generally applicable to business activities under the Ruggie Principles,[ref 687] which amongst others can directly influence decisions by tech company oversight bodies;[ref 688]
• Develop and provide clear industry standards and resources for their implementation, such as AI risk management frameworks.[ref 689]

Shaping industry-wide decisions around AI governance:

• Governmental regulations (as above);
• Ensure competition law frameworks enable cooperation on safety.[ref 690]

3.5. Pathways to shaping AI research community decisions

Shaping AI research community decisions around AI governance:

• Develop and disseminate clear guidelines and toolsets to facilitate responsible practices, such as:
  • Frameworks for pre-publication impact assessment of AI research;[ref 691] 
  • “Model cards” for the transparent reporting of benchmarked evaluations of a model’s performance across conditions and for different groups;[ref 692] 
  • General risk management frameworks for evaluating and anticipating AI risks.[ref 693]
• Framing and stigmatization around decisions or practices;[ref 694]
• Participatory technology assessment processes.[ref 695]

Shaping civil society decisions around AI governance:

• Work with “gatekeeper” organizations to put issues on the advocacy agenda.[ref 696]

3.6. Pathways to shaping international institutions’ decisions

Shaping international institutional decisions around AI governance:

• Clarify global administrative law obligations;[ref 697] 
• Influence domestic policy processes in order to indirectly shape transnational legal processes;[ref 698]
• Scientific expert bodies’ role in informing multilateral treaty-making by preparing evidence for treaty-making bodies, scientifically advising these bodies, and directly exchanging with them at intergovernmental body sessions or dialogical events.[ref 699]

Shaping standards bodies’ decisions around AI governance:

• Technical experts’ direct participation in standards development;[ref 700]
• Advancing standardization of advanced AI-relevant safety best practices.[ref 701]

3.7. Other pathways to shape various actors’ decisions

Shaping various actors’ decisions around AI governance:

• Work to shape broad narratives around advanced AI, such as through compelling narratives or depictions of good outcomes;[ref 702]
• Work to shape analogies or metaphors used by the public, policymakers, or courts in thinking about (advanced) AI;[ref 703]
• Pursue specific career paths with key actors to contribute to good policymaking.[ref 704]

III. Prescriptive work: Identifying priorities and proposing policies

Finally, a third category of work aims to go beyond either analyzing the problem of AI governance (Part I) or surveying potential elements or options for governance solutions analytically (Part II). This category is rather prescriptive in that it aims to directly propose or advocate for specific policies or actions by key actors. This includes work focused on: 

1. Articulating broad theories of change to identify priorities for AI governance (given a certain view of the problem and of the options available); 
2. Articulating broad heuristics for crafting good AI regulation; 
3. Putting forward policy proposals as well as assets that aim to help in their implementation.

1. Prioritization: Articulating theories of change

Achieving an understanding of the AI governance problem and potential options in response is valuable. Yet, this is not enough alone to deliver strategic clarity about which of these actors should be approached or which of these levers should be utilized in what ways. For that, it is necessary to develop more systematic accounts of different (currently held or possible) theories of change or impact. 

The idea of exploring and comparing such theories of action is not new. There have been various accounts that aim to articulate the linkages between near-term actions and longer-term goals. Some of these have focused primarily on theories of change (or “impact”) from the perspective of technical AI alignment.[ref 705] Others have articulated more specific theories of impact for the advanced AI governance space.[ref 706] These include:

• Dafoe’s Asset-Decision model, which focuses on the direction of research activities to help (1) create assets which can eventually (2) inform impactful decisions;[ref 707] 
• Leung’s model for impactful AI strategy research that can shape key decisions by (1) those developing and deploying AI and (2) those actors shaping the environments in which it is developed and deployed (i.e., research lab environment, legislative environment, and market environment).[ref 708]
• Garfinkel’s “AI Strategy: Pathways for Impact,”[ref 709] which highlights three distinct pathways for positively influencing the development of advanced AI: (1) become a decision-maker (or close enough to influence one), (2) spread good memes that are picked up by decision-makers, and (3) think of good memes to spread and make them credible;
• Baum’s framework for “affecting the future of AI governance,” which distinguishes several avenues by which AI policy could shape the long-term:[ref 710] (1) improve current AI governance, (2) support AI governance communities, (3) advance research on future AI governance, (4) advance CS design of AI safety and ethics to create solutions, and (5) improve underlying governance conditions.

In addition, some have articulated specific scenarios for what successful policy action on advanced AI might look like,[ref 711] especially in the relative near-term future (“AI strategy nearcasting”).[ref 712] However much further work is needed.

2. General heuristics for crafting advanced AI policy 

General heuristics for making policies relevant or actionable to advanced AI.

2.1. General heuristics for good regulation

Heuristics for crafting good AI regulation:

• Utilizing and articulating suitable terminology for drafting and scoping AI regulations, especially risk-focused terms;[ref 713] 
• Understand implications of different regulatory approaches (ex ante, ex post; risk regulation) for AI regulations;[ref 714] 
• Grounding AI policy within an “all-hazards” approach to managing various other global catastrophic risks simultaneously;[ref 715]
• Requirements for an advanced AI regime to avoid “perpetual risk”: exclusivity, benevolence, stability, and successful alignment;[ref 716]
• Establishing monitoring infrastructures to provide governments with actionable information.[ref 717]

2.2. Heuristics for good institutional design 

Heuristics for good institutional design:

• General desiderata and tradeoffs for international institutional design in terms of questions of regime centralization or decentralization;[ref 718]
• Procedural heuristics for organizing international negotiation processes: ensure international AI governance fora are inclusive of Global South actors;[ref 719]
• Ideal characteristics of global governance systems for high-risk AI, such as those that (1) govern dual-use technology; (2) take a risk-based approach; (3) provide safety measures; (4) incorporate technically informed, expert-driven, multi-stakeholder processes that enable rapid iteration; (5) where the effects are consistent with the treaty’s intent; and (6) that possess enforcement mechanisms.[ref 720]

2.3. Heuristics for future-proofing governance 

Heuristics for future-proofing governance regimes and desiderata and systems for making existing regulations more adaptive, scalable, or resilient:[ref 721]

• Traditional (treaty) reform or implementation mechanisms:
  • The formal treaty amendment process;[ref 722]
  • Unilateral state actions (explanatory memoranda and treaty reservations) or multilateral responses (Working Party Resolution) to adapt multilateral treaties;[ref 723]
  • The development of lex scripta treaties through the lex posteriori of customary international law, spurred by new state behavior.[ref 724] 
• Adaptive treaty interpretation methods:
  • Evolutionary interpretation of treaties;[ref 725]
  • Treaty interpretation under the principle of systemic integration.[ref 726]
• Instrument choices that promote flexibility:
  • Use of framework conventions;[ref 727]
  • Use of informal governance institutions;[ref 728] 
  • The subsequent layering of soft law on earlier hard-law regimes;[ref 729] 
  • Use of uncorrelated governance instruments to enable legal resilience.[ref 730]
• Regime design choices that promote flexibility:
  • Scope: include key systems (“general-purpose AI systems,” “highly capable foundation models,” “frontier AI systems,” etc.) within the material scope of the regulation;[ref 731]
  • Phrasing: in-text technological neutrality or deliberate ambiguity;[ref 732]
  • Flexibility provisions: textual flexibility provisions[ref 733] such as exceptions or flexibility clauses.
• Flexibility approaches beyond the legal regime:
  • Pragmatic and informal “emergent flexibility” about the meaning of norms and rules during crises.[ref 734]

3. Policy proposals, assets and products

That is, what are specific proposals for policies to be implemented? How can these proposals serve as products or assets in persuading key actors to act upon them?

Specific proposals for advanced AI-relevant policies; note that these are presented without comparison or prioritization. This list is non-exhaustive. Many proposals moreover combine several ideas, falling into different categories.

3.1. Overviews and collections of policies

• Previous collections of older proposals, such as Dewey’s list of “long-term strategies for ending existential risk”[ref 735] as well as Sotala and Yampolskiy’s survey of high-level “responses” to AI risk.[ref 736]
• More recent lists and collections of proposed policies to improve the governance, security, and safety of AI development[ref 737] in domains such as compute security and governance; software export controls; licenses;[ref 738] policies to establish improved standards, system evaluations, and licensing regimes; procurement rules and funding for AI safety;[ref 739] or to establish a multinational AGI consortium to enable oversight of advanced AI, a global compute cap, and affirmative safety evaluations.[ref 740]

3.2. Proposals to regulate AI using existing authorities, laws, or institutions

In particular, drawing on evaluations of the default landscape of regulations applied to AI (see Section I.3.3), and of the levers of governance for particular governments (see Section II.2.4).

Regulate AI using existing laws or policies

• Strengthen or reformulate existing laws and policies, such as EU competition law,[ref 741] contract and tort law,[ref 742] etc.;
• Strengthen or reorganize existing international institutions[ref 743] rather than establishing new institutions;[ref 744]
• Extend or apply existing principles and regimes in international law,[ref 745] including, amongst others:
  • Norms of international peace and security law:
    • Prohibitions on the use of force and intervention in the domestic affairs of other states;
    • Existing export control and nonproliferation agreements.
  • Principles of international humanitarian law, such as:
    • Distinction and proportionality in wartime; 
    • Prohibition on weapons that are by nature indiscriminate or cause unnecessary suffering;
    • The requirements of humanity;
    • The obligation to conduct legal reviews of new weapons or means of war (Article 36 under Additional Protocol I to the Geneva Conventions).
  • Norms of international human rights law[ref 746] and human rights and freedoms, including the right to life and freedom from cruel, inhuman, and degrading treatment, among others; the rights to freedom of expression, association, and security of the person, among others; and the principle of human dignity;[ref 747]
  • Norms of international environmental law, including the no-harm principle and the principle of prevention and precaution;
  • International criminal law, with regard to war crimes and crimes against humanity and with regard to case law of international criminal courts regarding questions of effective control;[ref 748]
  • Rules on state responsibility,[ref 749] including state liability for harm; 
  • Peremptory norms of jus cogens, outlawing, for example, genocide, maritime piracy, slavery, wars of aggression, and torture;
  • International economic law:[ref 750] security exception measures under international trade law and non-precludement measures under international investment law, amongst others;[ref 751]
  • International disaster law: obligations regarding disaster preparedness, including forecasting and pre-disaster risk assessment, multi-sectoral forecasting and early warning systems, disaster risk and emergency communication mechanisms, etc. (Sendai Framework);
  • Legal protections for the rights of future generations: including existing national constitutional protections for the rights of future generations[ref 752] and a potential future UN Declaration on Future Generations.[ref 753]

Proposals to set soft-law policy through existing international processes

• Proposals for engagement in existing international processes on AI: support the campaign to ban lethal autonomous weapons systems,[ref 754] orchestrate soft-law policy under G20,[ref 755] engage in debate about digital technology governance at the UN Summit for the Future,[ref 756] etc.

3.3. Proposals for new policies, laws, or institutions 

A range of proposals for novel policies.

Impose (temporary) pauses on AI development

• Coordinated pauses amongst AI developers whenever they identify hazardous capabilities;[ref 757]
• Temporary pause on large-scale system training beyond a key threshold,[ref 758] giving time for near-term policy-setting in domains such as robust third-party auditing and certification, regulation of access to computational power, establishment of capable national AI agencies, and establishment of liability for AI-caused harms, etc.;[ref 759]
• (Permanent) moratoria on developing (certain forms of) advanced AI.[ref 760]

Establish licensing regimes

• Evaluation and licensing regimes: establishment of a AI regulation regime for frontier AI systems, comprising “(1) standard-setting processes to identify appropriate requirements for frontier AI developers, (2) registration and reporting requirements to provide regulators with visibility into frontier AI development processes, and (3) mechanisms to ensure compliance with safety standards for the development and deployment of frontier AI models.”[ref 761]

Establish lab-level safety practices

• Proposals for establishing corporate governance and soft law: establish Responsible Scaling Policies (RSPs)[ref 762] and establish corporate governance and AI certification schemes.[ref 763] 

Establish governance regimes on AI inputs (compute, data)

• Compute governance regimes: establish on-chip firmware mechanisms, inspection regimes, and supply chain monitoring and custody mechanisms to ensure no actor can use large quantities of specialized chips to execute ML training runs in violation of established rules;[ref 764]
• Data governance: establish public data trusts to assert control over public training data for foundation models.[ref 765]

Establish domestic institutions for AI governance

• Proposals for new domestic institutions: US “AI Control Council”[ref 766] or National Algorithms Safety Board,[ref 767] and European AI Agency[ref 768] or European AI Office.[ref 769] 

Establish international AI research consortia

Proposals to establish new international hubs or organizations aimed at AI research.[ref 770]

• A diverse range of proposals for international institutions, including: a “CERN for AI,”[ref 771] “European Artificial Intelligence megaproject,”[ref 772] “Multilateral AI Research Institute (MAIRI),”[ref 773] “international large-scale AI R&D projects,”[ref 774] a collaborative UN superintelligence research project,[ref 775] “international organization that could serve as clearing-house for research into AI,”[ref 776] “joint international AI project with a monopoly on hazardous AI development,”[ref 777] “UN AI Research Organization,”[ref 778] a “good-faith joint US-China AGI project,”[ref 779] “AI for shared prosperity,”[ref 780] and a proposal for a new “Multinational AGI Consortium.”[ref 781] 

Establish bilateral agreements and dialogues

• Establish confidence-building measures[ref 782] and pursue international AI safety dialogues.[ref 783] 

Establish multilateral international agreements 

Proposal to establish a new multilateral treaty on AI:[ref 784]

• “Treaty on Artificial Intelligence Safety and Cooperation (TAISC),”[ref 785] global compute cap treaty,[ref 786] “AI development convention,”[ref 787] “Emerging Technologies Treaty,”[ref 788] “Benevolent AGI Treaty,”[ref 789] “pre-deployment agreements,”[ref 790] and many other proposals.

Establish international governance institutions

Proposals to establish a new international organization, along one or several models:[ref 791]

• A diverse range of proposals for international institutions, including a Commission on Frontier AI, an Advanced AI Governance Organization, a Frontier AI Collaborative, and an AI Safety Project;[ref 792] an International AI Organization (IAIO) to certify state jurisdictions for compliance with international AI oversight standards to enable states to prohibit the imports of goods “whose supply chains embody AI from non-IAIO-certified jurisdictions”;[ref 793] a proposal for an “international consortium” for evaluations of societal-scale risks from advanced AI;[ref 794] a “Global Organization for High-Risk Artificial Intelligence (GOHAI)”;[ref 795] and many other proposals.[ref 796]

Conclusion

The recent advances in AI have turned global public attention to this technology’s capabilities, impacts, and risks. AI’s significant present-day impacts and the prospect that these will only spread and scale further as these systems get increasingly advanced have firmly fixed this technology as a preeminent challenge for law and global governance this century. 

In response, the disparate community of researchers that have explored aspects of these questions over the past years may increasingly be called upon to translate that research into rigorous, actionable, legitimate, and effective policies. They have developed—and continue to produce—a remarkably far-flung body of research, drawing on a diverse range of disciplines and methodologies. The urgency of action around advanced AI accordingly create a need for this field to increase the clarity of its work and its assumptions, to identify gaps in its approaches and methodologies where it can learn from yet more disciplines and communities, to improve coordination amongst lines of research, and to improve legibility of its argument and work to improve constructive scrutiny and evaluation of key arguments and proposed policies. 

This review has not remotely achieved these goals—as no single document or review can. Yet by attempting to distill and disentangle key areas of scholarship, analysis, and policy advocacy, it hopes to help contribute to greater analytical and strategic clarity, more focused and productive research, and better-informed public debates and policymaker initiatives on the critical global challenges of advanced AI.

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, ‘AI is like… A literature review of AI metaphors and why they matter for policy.’ Institute for Law & AI. AI Foundations Report 2. (October 2023). https://www.law-ai.org/ai-policy-metaphors 
• 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

Concepts in advanced AI governance

Executive summary

This report provides an overview, taxonomy, and preliminary analysis of many cornerstone ideas and concepts in the emerging field of advanced AI governance. 

Aim: The aim of this report is to contribute to improved analysis, debate, and policy by providing greater clarity around core terms and concepts. Any field of study or regulation can be improved by such clarity. 

As such, this report reviews definitions for four categories of terms: the object of analysis (e.g., advanced AI), the tools for intervention (e.g., “governance” and “policy”), the reflexive definitions of the field of “advanced AI governance”, and its theories of change.

Summary: In sum, this report:

1. Discusses three different purposes for seeking definitions for AI technology, discusses the importance of such terminology in shaping AI policy and law, and discusses potential criteria for evaluating and comparing such terms.
2. Reviews concepts for advanced AI, covering a total of 101 definitions across 69 terms, including terms focused on:
   1. the forms of advanced AI, 
   2. the (hypothesized) pathways towards those advanced AI systems, 
   3. the technology’s large-scale societal impacts, and 
   4. particular critical capabilities that advanced AI systems are expected to achieve or enable.
3. Reviews concepts within “AI governance”, such as nine analytical terms used to define the tools for intervention (e.g., AI strategy, policy, and governance), four terms used to characterize different approaches within the field of study, and five terms used to describe theories of change. 

The terms are summarized below in Table 1. Appendices provide detailed lists of definitions and sources for all the terms covered as well as a list of definitions for nine other auxiliary terms within the field.

Introduction

As AI systems have become increasingly capable and have had increasingly public impacts, the field that focuses on governing advanced AI systems has come into its own. 

While researchers come to this issue with many different motivations, concerns, or hopes about AI—and indeed with many different perspectives on or expectations about the technology’s future trajectory and impacts—there has grown an emerging field of researchers, policy practitioners, and activists concerned with and united by what they see as the increasingly significant and pivotal societal stakes of AI. Along with significant disagreements, many in this emerging community share the belief that shaping the transformative societal impacts of advanced AI systems is a top global priority.[ref 1] However, this field still lacks clarity regarding not only many key empirical and strategic questions but also many key terms that are used.

Background: This lack of clarity matters because the recent wave of progress in AI, driven especially but not exclusively by the dramatic success of large language models (LLMs), has led to an accumulation of a wide range of new terms to describe these AI systems. Yet many of these terms—such as “foundation model”,[ref 2] “generative AI”,[ref 3] or “frontier AI”[ref 4]—do not always have clear distinctions[ref 5] and are often used interchangeably.[ref 6] They moreover emerge on top of and alongside a wide range of past terms, concepts, and words that have been used in the past decades to refer to (potential) advanced AI systems, such as “strong AI”, “artificial general intelligence”, or “transformative AI”. What are we to make of all of these terms?

Rationale: Critically, debates over terminology in and for advanced AI are not just semantics—these terms matter. In a broad sense, framings, metaphors, analogies, and explicit definitions can strongly affect not just developmental pathways for technology but also policy agendas and the efficacy and enforceability of legal frameworks.[ref 7] Indeed, different terms have already become core to major AI governance initiatives—with “general-purpose AI” serving as one cornerstone category in the EU AI Act[ref 8] and “frontier AI models” anchoring the 2023 UK AI Safety Summit.[ref 9] The varying definitions and implications of such terms may lead to increasing contestation,[ref 10] as well they should: Extensive work over the past decade has shown how different terms for “AI” import different regulatory analogies[ref 11] and have implications for crafting legislation.[ref 12] We might expect the same to hold for the new generation of terms used to describe advanced AI and to center and focus its governance.[ref 13] 

Aim: The aim of this report is to contribute to improved analysis, debate, and policy by providing greater clarity around core terms and concepts. Any field of study or regulation can be improved by such clarity. Such literature reviews may not just contribute to a consolidation of academic work, but can also refine public and policy debates.[ref 14] Ideally, they provide foundations for a more deliberate and reflexive choice over what concepts and terms to use (and which to discard), as well as a more productive refinement of the definition and/or operationalization of cornerstone terms. 

Scope: In response, this report considers four types of terms, including potential concepts and definitions for each of the following:

1. the core objects of analysis—and the targets for policy (i.e., what is the “advanced AI” to be governed?),
2. the tools for intervention to be used in response (i.e., what is the range of terms such as “policy”, “governance”, or “law”?),
3. the field or community (i.e., what are current and emerging accounts, projects, or approaches within the broader field of advanced AI governance?), and 
4. the theories of change of this field (i.e., what is this field’s praxis?).

Disclaimers: This project comes with some important caveats for readers. 

First, this report aims to be relatively broad and inclusive of terms, framings, definitions, and analogies for (advanced) AI. In doing so, it draws from both older and recent work and from a range of sources from academic papers to white papers and technical reports to public fora. 

Second, this report is primarily concerned with mapping the conceptual landscape and with understanding the (regulatory) implications of particular terms. As such, it is less focused on policing the appropriateness or coherence of particular terms or concepts. Consequently, with regard to advanced AI it covers many terms that are still highly debated or contested or for which the meaning is unsettled. Not all the terms covered are equally widely recognized, used, or even accepted as useful in the field of AI research or within the diverse fields of the AI ethics, policy, law, and governance space. Nonetheless, this report will include many of these terms on the grounds that a broad and inclusive approach to these concepts serves best to illuminate productive future debate. After all, even if some terms are (considered to be) “outdated,” it is important to know where such terms and concepts have come from and how they have developed over time. If some terms are contested or considered “too vague,” that should precisely speak in favor of aiming to clarify their usage and relation to other terms. This will either allow the (long overdue) refinement of concepts or will at least enable an improved understanding of when certain terms are not usefully recoverable. In both cases, it will facilitate greater clarity of communication.

Third, this review is a snapshot of the state of debate at one moment. It reviews a wide range of terms, many of which have been coined recently and only some of which may have staying power. This debate has developed significantly in the last few years and will likely continue to do so. 

Fourth, this review will mostly focus on analytical definitions of or for advanced AI along four approaches.[ref 15] In so doing, it will on this occasion mostly omit detailed exploration of a fifth, normative dimension to defining AI, which would focus on reviewing especially desirable types of advanced AI systems that (in the view of some) ought to be pursued or created. Such a review would cover a range of terms such as “ethical AI”,[ref 16] “responsible AI”,[ref 17] “explainable AI”,[ref 18] “friendly AI”,[ref 19] “aligned AI”,[ref 20] “trustworthy AI”,[ref 21] “provably-safe AI”,[ref 22] “human-centered AI”,[ref 23] “green AI”,[ref 24] “cooperative AI”,[ref 25] “rights-respecting AI”,[ref 26] “predictable AI”,[ref 27] “collective intelligence”,[ref 28] and “digital plurality”,[ref 29] amongst many other terms and concepts. At present, this report will not focus in depth on surveying these terms, since only some of them were articulated in the context of or in consideration of especially advanced AI systems. However, many or all of these terms are capability-agnostic and so could clearly be extended to or reformulated for more capable, impactful, or dangerous systems. Indeed, undertaking such a deepening and extension of the taxonomy presented in this report in ways that engage more with the normative dimension of advanced AI would be very valuable future work.

Fifth, this report does not aim to definitively resolve debates—or to argue that all work should adopt one or another term over others. Different terms may work best in different contexts or for different purposes and for different actors. Indeed, given the range of actors interested in AI—whether from a technical engineering, sociotechnical, or regulatory perspective—it is not surprising that there are so many terms and such diversity in definitions even for single terms. Nonetheless, to be able to communicate effectively and learn from other fields, it helps to gain greater clarity and precision in the terms we use, whether these are terms referring to our objects of analysis, our own field and community, or our theory of action. Of course, achieving clarity on terminology is not itself sufficient. Few problems, technical or social or legal, may be solved exclusively by haggling over words. Nonetheless, a shared understanding facilitates problem solving. The point here is not to achieve full or definitive consensus but to understand disagreements and assumptions. As such, this report seeks to provide background on many terms, explore how they have been used, and consider the suitability of these terms for the field.[ref 30] In doing so, this report highlights the diversity of terms in current use and provides context for more informed future study and policymaking. 

Structure: Accordingly, this report now proceeds as follows. 

Part I provides a background to this review by discussing three purposes to defining key terms such as AI. It also discusses why the choice for one or another term matters significantly from the perspective of AI policy and regulation, and finally discusses some criteria by which to evaluate the suitability of various terms and definitions for the specific purpose of regulation. 

In Part II, this report reviews a wide range of terms for “advanced AI”, across different approaches which variably focus on (a) the anticipated forms or design of advanced AI systems, (b) the hypothesized scientific pathways towards these systems, (c) the technology’s broad societal impacts, or (d) the specific critical capabilities particular advanced AI systems are expected to achieve. 

Part III turns from the object of analysis to the field and epistemic community of advanced AI governance itself. It briefly reviews three categories of concepts of use for understanding this field. First, it surveys different terms used to describe AI “strategy”, “policy”, or “governance” as this community understands the available tools for intervention in shaping advanced AI development. It then reviews different paradigms within the field of advanced AI governance as ways in which different voices within it have defined that field. Finally, it briefly reviews recent definitions for theories of change that aim to compare and prioritize interventions into AI governance. 

Finally, three appendices list in detail all the terms and definitions offered, with sources, and offer a list of auxiliary definitions that can aid future work in this emerging field.[ref 31]etail, with sources; and offer a list of auxiliary definitions that can aid future work in this emerging field.

I. Defining ‘advanced AI (governance)’: Background

Any quest for clarifying definitions of “advanced AI” is complicated by the already long-running, undecided debates over how to even define the more basic terms “AI” or, indeed, “intelligence”.[ref 32] 

To properly evaluate and understand the relevance of different terms for AI, it is useful to first set out some background. In the first place, one should start by considering the purposes for which the definition is sought. Why or how do we seek definitions of “(advanced) AI”? 

1. Three purposes for definitions

For instance, rather than trying to consider a universally best definition for AI, a more appropriate approach is to consider the implications of different definitions, or—to invert the question—to ask for what purpose we seek to define AI. We can consider (at least) three different rationales for defining a term like ‘AI’. 

1. To build it (the technological research purpose): In the first place, AI researchers or scientists may pursue definitions of (advanced) AI by defining it from the “inside,” as a science.[ref 33] The aim of such technical definitions of AI[ref 34] is to clarify or create research-community consensus about (1) the range and disciplinary boundaries of the field—that is, what research programs and what computational techniques[ref 35] count as “AI research” (both internally and externally to research funders or users); (2) the long-range goals of the field (i.e., the technical forms of advanced AI); and/or (3) the intermediate steps the field should take or pursue (i.e., the likely pathways towards such AI). Accordingly, this definitional purpose aligns particularly closely with essence-based definitions (see Part II.1) and/or development-based definitions (see Part II.2) of advanced AI.
2. To study it (the sociotechnical research purpose): In the second place, experts (in AI, but especially in other fields) may seek to primarily understand AI’s impacts on the world. In doing so, they may aim to define AI from the “outside,” as a sociotechnical system including its developers and maintainers.[ref 36] Such definitions or terms can aid researchers (or governments) who seek to understand the societal impacts and effects of this technology in order to diagnose or analyze the potential dynamics of AI development, diffusion, and application, as well as the long-term sociopolitical problems and opportunities. For instance, under this purpose researchers may aim to get to terms with understanding issues such as (1) (the geopolitics or political economy of) key AI inputs (e.g., compute, data, and labor), (2) how different AI capabilities[ref 37] give rise to a spectrum of useful applications[ref 38] in diverse domains, and (3) how these applications in turn produce or support new behaviors and societal impacts.[ref 39] Accordingly, this purpose is generally better served by sociotechnical definitions of AI systems’ impacts (see Part II.3) or risk-based definitions (see Part II.4).
3. To regulate it (the regulatory purpose): Finally, regulators or academics motivated by appropriately regulating AI—either to seize the benefits or to mitigate adverse impacts—can seek to pragmatically delineate and define (advanced) AI as a legislative and regulatory target. In this approach, definitions of AI are to serve as useful handles for law, regulation, or governance.[ref 40] In principle, this purpose can be well served by many of the definitional approaches: highly technology-specific regulations for instance can gain from focusing on development-based definitions of (advanced) AI. However, in practice regulation and governance is usually better served by focusing on the sociotechnical impacts or capabilities of AI systems.

Since it is focused on the field of “advanced AI governance,” this report will primarily focus on the second and third of these purposes. However, it is useful to keep all three in mind.

2. Why terminology matters to AI governance

Whether taking a sociotechnical perspective on the societal impacts of advanced AI or a regulatory perspective on adequately governing it, the need to pick suitable concepts and terms becomes acutely clear. Significantly, the implications and connotations of key terms matter greatly for law, policy, and governance. This is because, as reviewed in a companion report,[ref 41] distinct or competing terms for AI—with their meanings and connotations—can influence all stages of the cycle from a technology’s development to its regulation. They do so in both a broad and a narrow sense.

In the broad and preceding sense, the choice of term and definition can, explicitly or implicitly, import particular analogies or metaphors into policy debates that can strongly shape the direction—and efficacy—of the resulting policy efforts.[ref 42] These framing effects can occur even if one tries to avoid explicit analogies between AI and other technologies, since apparently “neutral” definitions of AI still focus on one or another of the technology’s “features” as the most relevant, framing policymaker perceptions and responses in ways that are not neutral, natural, or obvious. For instance, Murdick and others found that the particular definition one uses for what counts as “AI” research directly affects which (industrial or academic) metrics are used to evaluate different states’ or labs’ relative achievements or competitiveness in developing the technology—framing downstream evaluations of which nation is “ahead” in AI.[ref 43] Likewise, 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 44] 

Beyond the broad policy-framing impacts of technology metaphors and analogies, there is also a narrower sense in which terms matter. Specifically, within regulation, legislative and statutory definitions delineate the scope of a law and of the agency authorization to implement or enforce it[ref 45]—such that the choice for a particular term for (advanced) AI may make or break the resulting legal regime.

Generally, within legislative texts, the inclusion of particular statutory definitions can play both communicative roles (clarifying legislative intent), and performative roles (investing groups or individuals with rights or obligations).[ref 46] More practically, one can find different types of definitions that play distinct roles within regulation: (1) delimiting definitions establish the limits or boundaries on an otherwise ordinary meaning of a term, (2) extending definitions broaden a term’s meaning to expressly include elements or components that might not normally be included in its ordinary meaning, (3) narrowing definitions aim to set limits or expressly exclude particular understandings, and (4) mixed definitions use several of these approaches to clarify components.[ref 47] 

Likewise, in the context of AI law, legislative definitions for key terms such as “AI” obviously affect the material scope of the resulting regulations.[ref 48] 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 49] This means, for instance, that regulations that refer to terms such as “advanced AI”, “frontier AI”, or “transformative AI” might not necessarily be interpreted or applied in ways that are in line with how the term is understood within expert communities. All of this underscores the importance of our choice of terms—from broad and indirect metaphors to concrete and specific legislative definitions—when grappling with the impacts of this technology on society.

Indeed, the strong legal effects of different terms mean that there can be challenges for a law when it depends on a poorly or suboptimally specified regulatory term for the forms, types, or risks from AI that the legislation means to address. 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—the so-called Collingridge dilemma.[ref 50]

At the same time, the cost of picking and locking in the wrong terms within legislative texts can be significant. Beyond the opportunity costs, unreflexively establishing legal definitions for key terms can create the risk of downstream or later “governance misspecification.”[ref 51] 

Such governance misspecification may 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 assumption may be appropriate and correct in many cases, there is a risk that if that assumption is wrong—either because of an initial misapprehension of the technology or because subsequent technological developments lead to that proxy term diverging from the legislative goals—the resulting technology law will less efficient, ineffective, or even counterproductive to its purposes.[ref 52] 

Such cases of governance misspecification can be seen in various cases of technology governance and regulation. For instance: 

• The “high-performance computer” threshold in US 1990s export control regimes: In the 1990s, the US established a series of export control regimes under the Wassenaar Arrangement, which set an initial threshold for “high-performance computers” at just 195 million theoretical operations per second (MTOPS); in doing so, the regime treated that technology as far too static and could not keep pace with Moore’s Law.[ref 53] As a result, the threshold had to be updated six times within a decade,[ref 54] even as the regime became increasingly ineffective at preventing or even inhibiting US adversaries from accessing as much computing power as they needed, and it may even have become harmful to national security as it inhibited the domestic US tech industry.[ref 55] 
• The “in orbit” provision in the Outer Space Treaty: In the late 1960s, the Outer Space Treaty aimed to outlaw positioning weapons of mass destruction in space. It therefore (as proxy) specified a ban on placing these weapons “in orbit.”[ref 56] This definition meant that there was a loophole to be exploited by the Soviet development of fractional orbital bombardment systems (FOBS), which were able to position nuclear weapons in space (on non-ballistic trajectories) without, strictly, putting them “in orbit.”[ref 57] 
• Under- and overinclusive 2010s regulations on drones and self-driving cars: Calo has chronicled how, in the early 2010s, various US regulatory responses to drones or self-driving cars defined these technologies in ways that were either under- or overinclusive, leading to inefficiency or the repeal of laws.[ref 58]

Thus, getting greater clarity in our concepts and terminology for advanced AI will be critical in crafting effective, resilient regulatory responses—and in avoiding brittle missteps that are easily misspecified.

Given all the above, the aim in this report is not to find the “correct” definition or frame for advanced AI. Rather, it considers that different frames and definitions can be more useful for specific purposes or for particular actors and/or (regulatory) agencies. In that light, we can explore a series of broad starting questions, such as: 

1. What different definitions have been proposed for advanced AI? What other terms could we choose? 
2. What aspects of advanced AI (e.g., its form and design, the expected scientific principles of its development pathways, its societal impacts, or its critical capabilities) do these different terms focus on? 
3. What are the regulatory implications of different definitions?

In sum, this report is premised on the idea that exploring definitions of AI (and related terms) matters, whether we are trying to understand AI, understand its impacts, or govern them effectively.

3. Criteria for definitions

Finally, we have the question of how to formulate relevant criteria for suitable terms and definitions for advanced AI. In the first place, as discussed above, this depends on one’s definitional purpose. 

Nonetheless, from the specific perspective of regulation and policymaking, what are some good criteria for evaluating suitable and operable definitions for advanced AI? Notably, Jonas Schuett has previously explored legal approaches to defining the basic term “AI”. He emphasizes that to be suitable for the purpose of governance, the choice of terms for AI should meet a series of requirements for all good legal definitions—namely that terms are neither (1) overinclusive nor (2) underinclusive and that they are (3) precise, (4) understandable, (5) practicable, and (6) flexible.[ref 59] Other criteria have been proposed: for instance, it has been suggested that an additional desiderata for a useful regulatory definition for advanced AI might include something like ex ante clarity—in the sense that the definition should allow one to assess, for a given AI model, whether it will meet the criteria for that definition (i.e., whether it will be regulated within some regime), and ideally allow this to be assessed in advance of deployment (or even development) of that model.[ref 60] Certainly, these criteria remain contested and are likely incomplete. In addition, there may be trade-offs between the criteria, such that even if they are individually acceptable, one must still strike a workable balance between them.[ref 61] 

II. Defining the object of analysis: Terms for advanced AI

Having briefly discussed the different definitional purposes, the relevance of terms for regulation, and potential criteria for evaluating definitions, this report now turns to survey the actual terminology for advanced AI. 

Within the literature and public debate, there are many terms used to refer to the conceptual cluster of AI systems that are advanced—i.e., that are sophisticated and/or are highly capable and/or could have transformative impacts on society.[ref 62] However, because of this diversity of terms, not all have featured equally strongly in governance or policy discussions. To understand and situate these terms, it is useful to compare their definitions with others and to review different approaches to defining advanced AI. 

In Schuett’s model for “legal” definitions for AI, he has distinguished four types of definitions, which focus variably on (1) the overarching term “AI”, (2) particular technical approaches in machine learning, (3) specific applications of AI, and (4) specific capabilities of AI systems (e.g., physical interaction, ability to make automated decisions, ability to make legally significant decisions).[ref 63] 

Drawing on Schuett’s framework, this report draws a similar taxonomy for common definitions for advanced AI. In doing so, it compares between different approaches that focus on one of four features or aspects of advanced AI.

1. The anticipated technical form or design of AI systems (essence-based approaches);
2. The proposed scientific pathways and paradigms towards creating advanced AI (development-based approaches); 
3. The broad societal impacts of AI systems, whatever their cognitive abilities (sociotechnical-change-based approach);
4. The specific critical capabilities[ref 64] that could potentially enable extreme impacts in particular domains (risk-based approaches).

Each of these approaches has a different focus, object, and motivating question (Table 2).

This report will now review these categories of approaches in turn. For each, it will broadly (1) discuss that approach’s core definitional focus and background, (2) list the terms and concepts that are characteristic of it, (3) provide some brief discussion of common themes and patterns in definitions given to these terms,[ref 65] and (4) then provide some preliminary reflections on the suitability of particular terms within this approach, as well as of the approach as a whole, to provide usable analytical or regulatory definitions for the field of advanced AI governance.[ref 66]

1. Essence-based definitions: Forms of advanced AI

Focus of approach: Classically, many definitions of advanced AI focus on the anticipated form, architecture, or design of future advanced AI systems.[ref 67] These definitions as such focus on AI systems that instantiate particular forms of advanced intelligence,[ref 68] for instance by instantiating an “actual mind” (that “really thinks”); by displaying a degree of autonomy; or by being human-like, general-purpose, or both in the ability to think, reason, or achieve goals across domains (see Table 3). 

Terms: The form-centric approach to defining advanced AI accordingly encompasses a variety of terms, including strong AI, autonomous machine (/ artificial) intelligence, general artificial intelligence, human-level AI, foundation model, general-purpose AI system, comprehensive AI services, artificial general intelligence, robust artificial intelligence, AI+, (machine/artificial) superintelligence, superhuman general-purpose AI, and highly-capable foundation models.[ref 69] 

Definitions and themes: While many of these terms are subject to a wide range of different definitions (see Appendix 1A), they combine a range of common themes or patterns (see Table 3).

Suitability of overall definitional approach: In the context of analyzing advanced AI governance, there are both advantages and drawbacks to working with form-centric terms. First, we review five potential benefits. 

Benefit (1): Well-established and recognized terms: In the first place, using form-centric terms has the advantage that many of these terms are relatively well established and familiar.[ref 72] Out of all the terms surveyed in this report, many form-centric definitions for advanced AI, like strong AI, superintelligence, or AGI, have both the longest track record and the greatest visibility in academic and public debates around advanced AI. Moreover, while some of these terms are relatively niche to philosophical (“AI+”) or technical subcommunities (“CAIS”), many of these terms are in fact the ones used prominently by the main labs developing the most disruptive, cutting-edge AI systems.[ref 73] Prima facie, reusing these terms could avoid the problem of having to reinvent the wheel and achieve widespread awareness of and buy-in on newer, more niche terms. 

Benefit (2): Readily intuitive concepts: Secondly, form-centric terms evoke certain properties—such as autonomy, adaptability, and human-likeness—which, while certainly not uncontested, may be concepts that are more readily understood or intuited by the public or policymakers than would be more scientifically niche concepts. At the same time, this may also be a drawback, if the ambiguity of many of these terms opens up greater scope for misunderstanding or flawed assumptions to creep into governance debates. 

Benefit (3): Enables more forward-looking and anticipatory policymaking towards advanced AI systems and their impacts. Thirdly, because some (though not all) form-centric definitions of advanced AI relate to systems that are perceived (or argued) to appear in the future, using these terms could help extend public attention, debate, and scrutiny to the future impacts of yet more general AI systems which, while their arrival might be uncertain, would likely be enormously impactful. This could help such debates and policies to be less reactive to the impacts of each latest AI model release or incident and start laying the foundations for major policy initiatives. Indeed, centering governance analysis on form-centric terms, even if they are (seen as) futuristic or speculative, can help inform more forward-looking, anticipatory, and participatory policymaking towards the kind of AI systems (and the kind of capabilities and impacts) that may be on the horizon.[ref 74]

One caveat here is that to consider this a benefit, one has to strongly assume that these futuristic forms of advanced AI systems are in fact feasible and likely near in development. At the same time, this approach need not presume absolute certainty over which of these forms of advanced AI can or will be developed, or on what timelines; rather, well-established risk management approaches[ref 75] can warrant some engagement with these scenarios even under uncertainty. To be clear, this need not (and should not) mean neglecting or diminishing policy attention for the impacts of existing AI systems,[ref 76] especially as these impacts are already severe and may continue to scale up as AI systems both become more widely implemented and create hazards for existing communities.

Benefit (4): Enables public debate and scrutiny of overarching (professed) direction and destination for AI development. Fourthly, and relatedly, this above advantage to using form-centric terms could still hold, even if one is very skeptical of these types of futuristic AI, because they afford the democratic value of allowing the public and policymakers to chime in on the actual professed long-term goals and aspirations of many (though not all) leading AI labs.[ref 77] 

In this way, the cautious, clear, and reflexive use of terms such as AGI in policy debates could be useful even if one is very skeptical of the actual feasibility of these forms of AI (or believes they are possible but remains skeptical that they will be built anytime soon using extant approaches). This is because there is democratic and procedural value in the public and policymakers being able to hold labs to account for the goals that they in fact espouse and pursue—even if those labs may turn out mistaken about the ability to execute on those plans (in the near term).[ref 78] This is especially the case when these are goals that the public might not (currently) agree with or condone.[ref 79] 

Using these “futuristic” terms could therefore help ground public debate over whether the development of these particular systems is even a societal goal they condone, whether society might prefer for labs or society to pursue a different visions for society’s relation to AI technology,[ref 80] or (if these systems are indeed considered desirable and legitimate goals) what additional policies or guarantees the world should demand.[ref 81]

Benefit (5): Technology neutrality: Fifthly, the use of form-centric terms in debates can build in a degree of technology neutrality[ref 82] in policy responses, since debates need not focus on the specific engineering or scientific pathways by which one or another highly capable and impactful AI system is pursued or developed. This could make the resulting regulatory frameworks more scalable and future-proof.

At the same time, there are a range of general drawbacks to using (any of these) form-focused definitions in advanced AI governance. 

Drawback (1): Connotations and baggage around terms: In the first place, the greater familiarity of some of these terms means that many form-focused terms have become loaded with cultural baggage, associations, or connotations which may mislead, derail, or unduly politicize effective policymaking processes. In particular, many of these terms are contested and have become associated (whether or not necessarily) with particular views or agendas towards building these systems.[ref 83] This is a problem because, as discussed previously, the use of different metaphors, frames, and analogies may be irreducible in (and potentially even essential to) the ways that the public and policymakers make sense of regulatory responses. Yet different analogies—and especially the unreflexive use of terms—also have limits and drawbacks and create risks of inappropriate regulatory responses.[ref 84]

Drawback (2): Significant variance in prominence of terms and constant turnover: In the second place, while some of these terms have held currency at different times in the last decades, many do not see equally common use or recognition in modern debates. For instance, terms such as “strong AI” which dominated early philosophical debates, appear to have fallen slightly out of favor in recent years[ref 85] as the emergence and impact of foundation models generally, and generative AI systems specifically, has revived significantly greater attention to terms such as “AGI”. This churn or turnover in definitions may mean that it may not be wise to attempt to pin down a single term or definition right now, since analyses that focus on one particular anticipated form of advanced AI may be more likely to be rendered obsolete. At the same time, this is likely to be a general problem with any concepts or terminology chosen.

Drawback (3): Contested terms, seen as speculative or futuristic: In the third place, while some form-centric terms (such as “GPAIS” or “foundation model”) have been well established in AI policy debates or processes, others, such as “AGI”, “strong AI”, or “superintelligence”, are more future-oriented, referring to advanced AI systems that do not (yet) exist.[ref 86] Consequently, many of these terms are contested and seen as futuristic and speculative. This perception may be a challenge, because even if it is incorrect (e.g., such that particular systems like “AGI” will in fact be developed within short timelines or are even in some sense “already here”[ref 87]), the mere perception that a technology or term is far-off or “speculative” can serve to inhibit and delay effective regulatory or policy action.[ref 88] 

A related but converse risk of using future-oriented terms for advanced AI policy is that it may inadvertently import a degree of technological determinism[ref 89] in public and policy discussions, as it could imply that one or another particular forms or architectures of advanced AI (“AGI”, “strong AI”) are not just possible but inevitable—thereby shifting public and policy discussions away from the question of whether we should (or can safely) develop these systems (rather than other, more beneficial architectures)[ref 90] towards less ambitious questions over how we should best (safely) reckon with the arrival or development of these technologies.

In response, this drawback could be somewhat mitigated by relying on terms for the forms of advanced AI—such as GPAIS or highly-capable foundation models—that are (a) more present-focused, while (b) not putting any strong presumed ceilings on the capabilities of the systems.

Drawback (4): Definitional ambiguity: In the fourth place, many of these terms, and especially future-oriented terms such as “strong AI”, “AGI”, and “human-level AI”, suffer from definitional ambiguity in that they are used both inconsistently and interchangeably with one another.[ref 91] 

Of course, just because there is no settled or uncontested definition for a term such as “AGI” does not make it prima facie unsuitable for policy or public debate. By analogy, the fact that there can be definitional ambiguity over the content or boundaries of concepts such as “the environment” or “energy” does not render “environmental policy” or “energy policy” meaningless categories or irrelevant frameworks for regulation.[ref 92] Nor indeed does outstanding definitional debate mean that any given term, such as AGI, is “meaningless.”[ref 93] 

Nonetheless, the sheer range of contesting definitions for many of these concepts may reflect an underlying degree of disciplinary or philosophical confusion, or at least suggest that, barring greater conceptual clarification and operationalization,[ref 94] these terms will lead to continued disagreement. Accordingly, anchoring advanced AI governance to broad terms such as “AGI” may make it harder to articulate appropriately scoped legal obligations for specific actors that will not end up being over- or underinclusive.[ref 95] 

Drawback (5): Challenges in measurement and evaluation: In the fifth place, an underlying and related challenge for the form-centric approach is that (in part due to these definitional disagreements and in part due to deeper reasons) it faces challenges around how to measure or operationalize (progress towards) advanced AI systems. 

This matters because effective regulation or governance—especially at the international level[ref 96]—often requires (scientific and political) consensus around key empirical questions, such as when and how we can know that a certain AI system truly achieves some of the core features (e.g., autonomy, agency, generality, and human-likeness) that are crucial to a given term or concept. In practice, AI researchers often attempt to measure such traits by evaluating an AI system’s ability to pass one or more specific benchmark tests (e.g., the Turing test, the Employment test, the SAT, etc.).[ref 97] 

However, such testing approaches have many flaws or challenges.[ref 98] At the practical level, there have been problems with how tests are applied and scored[ref 99] and how their results are reported.[ref 100] Underlying this is a challenge that the way in which some common AI performance tests are constructed may emphasize nonlinear or discontinuous metrics, which can lead to an overtly strong impression that some model skills are “suddenly” emergent properties (rather than smoothly improving capabilities).[ref 101] More fundamentally, there have been challenges to the meaningfulness of applying human-centric tests (such as the bar exam) to AI systems[ref 102] and indeed deeper critiques of the construct validity of leading benchmark tests in terms of whether they actually are indicative of progress towards flexible and generalizable AI systems.[ref 103] 

Of course, that does not mean that there may not be further scientific progress towards the operationalization of useful tests for understanding when particular forms of advanced AI such as AGI have been achieved.[ref 104] Nor is it to suggest that benchmark and evaluation challenges are unique to form-centric definitions of AI—indeed, they may also challenge many approaches focused on specific capabilities of advanced AIs.[ref 105] However, the extant challenges over the operationalization of useful tests mean that overreliance on these terms could muddle debates and inhibit consensus over whether a particular advanced system is within reach (or already being deployed). 

Drawback (6): Overt focus on technical achievement of particular forms may make this approach underinclusive of societal impacts or capabilities: In the sixth place, the focus of future-oriented form-centric approaches on the realization of one or another type of advanced AI system (“AGI”, “human-level AI”), might be adequate if the purpose for our definitions is for technical research.[ref 106] However, for those whose definitional purpose is to understand AI’s societal impacts (sociotechnical research) or to appropriately regulate AI (regulatory), many form-centric terms may miss the point. 

This is because what matters from the perspective of human and societal safety, welfare, and well-being—and from the perspective of law and regulation[ref 107]—is not the achievement of some fully general capacity in any individual system but rather overall sociotechnical impacts or the emergence of key dangerous capabilities—even if they derive from systems that are not yet (fully) general[ref 108] or that develop dangerous emergent capabilities that are not human-like.[ref 109] Given all this, there is a risk that taking a solely form-centric approach leaves advanced AI governance vulnerable to a version of the “AI effect,” whereby “real AGI” is always conceived of as being around the corner but rarely as a system already in production. 

Suitability of different terms within approach: Given the above, if one does aim to draw on this approach, it may be worth considering which terms manage to gain from the strengths of this approach while reducing some of the pitfalls. In this view, the terms “GPAIS” or “foundation model” may be more suitable in many contexts, as they are recognized as categories of (increasingly) general and competent AI systems of which some versions already exist today. In particular, because (versions) of these terms are already used in ongoing policy debates, they could provide better regulatory handles for governing the development of advanced AI—for instance by their relation to the complex supply chain of modern AI development that contains both upstream and downstream developers and users.[ref 110] Moreover, these terms do not presume a ceiling in the system’s capability; accordingly, concepts such as “highly-capable foundation model”,[ref 111] “extremely capable foundation model”, or “threshold foundation model” could help policy debates be cognizant of the growing capabilities of these systems while still being more easily understandable for policymakers.[ref 112]

2. Development-based definitions: Pathways towards advanced AI

Focus of approach: A second cluster of terms focuses on the anticipated or hypothesized scientific pathways or paradigms that could be used to create advanced AI systems. Notably, the goal or target of these pathways is often to build “AGI”-like systems.[ref 113] 

Notes and caveats: Any discussion of proposed pathways towards advanced AI has a number of important caveats. In the first place, many of these proposed paradigms have long been controversial, with pervasive and ongoing disagreement about their scientific foundations and feasibility as paths towards advanced AI (or in particular as paths towards particular forms of advanced AI, such as AGI).[ref 114] Secondly, these approaches are not necessarily mutually exclusive, and indeed many labs combine elements from several in their research.[ref 115] Thirdly, because the relative and absolute prominence and popularity of many of these paradigms have fluctuated over time and because there are often, as in any scientific field, significant disciplinary gulfs between paradigms, there is highly unequal treatment of these pathways and terms. As such, whereas some paradigms (such as the scaling, reinforcement-learning, and, to some extent, brain-inspired approaches) are reasonably widely known, many of the other approaches and terms listed (such as “seed AI”) may be relatively unknown or even very obscure within the modern mainstream machine learning (ML) community.[ref 116] 

Other taxonomies: There have been various other such attempts to create taxonomies of the main theorized pathways that have been proposed to build or implement advanced AI. For instance, Goertzel and Pennachin have defined four different approaches to creating “AGI”, which to different degrees draw on lessons from the (human) brain or mind.[ref 117] More recently, Hannas and others have drawn on this framework and extended it to five theoretical pathways towards “general AI”.[ref 118] 

Further extending such frameworks, one can distinguish between at least 11 proposed pathways towards advanced AI (See Table 4).

Terms: Many of these paradigms or proposed pathways towards advanced AI come with their own assorted terms and definitions (see Appendix 1B). These terms include amongst others de novo AGI, prosaic AGI, frontier (AI) model [compute threshold], [AGI] from evolution, [AGI] from powerful reinforcement learning agents, powerful deep learning models, seed AI, neuroAI, brain-like AGI, neuromorphic AGI, whole-brain emulation, brain-computer interface, [advanced AI based on] a sophisticated embodied agent, or hybrid AI (see Table 4).

Definitions: As noted, these terms can be mapped on 11 proposed pathways towards advanced AI, with their own terms for the resulting advanced AI systems. 

Notably, there are significant differences in the prominence of these approaches—and the resources dedicated to them—at different frontier AI labs today. For instance, while some early work on the governance of advanced AI systems focused on AI systems that would (presumably) be built from first principles, bootstrapping,[ref 121] or neuro-emulated approaches (see Table 4), much of such work has more recently shifted to focus on understanding the risks from and pathways to aligning and governing advanced AI systems created through computational scaling. 

This follows high-profile trends in leading AI labs. While (as discussed above) many research labs are not dedicated to a single paradigm, the last few years (and 2023 in particular) have seen a significant share of resources going towards computational scaling approaches, which have yielded remarkably robust (though not uncontested) performance improvements.[ref 122] As a result, the scaling approach has been prominent in informing the approaches of labs such as OpenAI,[ref 123] Anthropic,[ref 124] DeepMind,[ref 125] and Google Brain (now merged into Google DeepMind).[ref 126] This approach has also been prominent (though somewhat lagging) in some Chinese labs such as Baidu, Alibaba, Tencent, and the Beijing Institute for General Artificial Intelligence.[ref 127] Nonetheless, other approaches continue to be in use. For instance, neuro-inspired approaches have been prominent in DeepMind,[ref 128] Meta AI Research,[ref 129] and some Chinese[ref 130] and Japanese labs,[ref 131] and modular cognitive architecture approaches have informed the work by Goertzel’s OpenCog project,[ref 132] amongst others. 

Suitability of overall definitional approach: In the context of analyzing advanced AI governance, there are both advantages and drawbacks to using concepts that focus on pathways of development. 

Amongst the advantages of this approach are:

Benefit (1): Close(r) grounding in actual technical research agendas aimed at advanced AI: Defining advanced AI systems according to their (envisioned) development pathways has the benefit of keeping advanced AI governance debates more closely grounded in existing technical research agendas and programs, rather than the often more philosophical or ambiguous debates over the expected forms of advanced AI systems. 

Benefit (2): Technological specificity allowing scoping of regulation to approaches of concern: Relatedly, this also allows better regulatory scoping of the systems of concern. After all, the past decade has seen a huge variety amongst AI techniques and approaches, not just in terms of their efficacy but also in terms of the issues they raise, with particular technical approaches raising distinct (safety, interpretability, robustness) issues.[ref 133] At the same time, these correlations might be less relevant in the last few years given the success of scaling-based approaches at creating remarkably versatile and general-purpose systems. 

However, taking the pathways-focused approach to defining advanced AI has its own challenges:

Drawback (1): Brittleness as technological specificity imports assumptions about pathways towards advanced AI: The pathway-centric approach may import strong assumptions about what the relevant pathways towards advanced AI are. As such, governance on this basis may not be robust to ongoing changes or shifts in the field.

Drawback (2): Suitability of terms within this approach: Given this, development-based definitions of pathways towards advanced AI seem particularly valuable if the purpose of definition is technical research but may be less relevant if the purpose is sociotechnical analysis or regulation. Technical definitions of AI might therefore provide an important baseline or touchstone for analysis in many other disciplines, but they may not be fully sufficient or analytically enlightening to many fields of study dealing with the societal consequences of the technology’s application or with avenues for governing these. 

At any rate, one interesting feature of development-based definitions of advanced AI is that the choice of approach (and term) to focus on has significant and obvious downstream implications for framing the policy agendas for advanced AI—in terms of the policy issues to address, the regulatory “surface” of advanced AI (e.g., the necessary inputs or resources to pursue research along a certain pathway), and the most feasible or appropriate tools. For instance, a focus on neuro-integrationist-produced brain-computer interfaces suggests that policy issues for advanced AI will focus less on questions of value alignment[ref 134] and rather around (biomedical) questions of human consent, liability, privacy, (employer) neurosurveillance,[ref 135] and/ormorphological freedom.[ref 136] A focus on embodiment-based approaches towards robotic agents raises more established debates from robot law.[ref 137] Conversely, if one expects that the pathway towards advanced AI still requires underlying scientific breakthroughs, either from first principles or through a hybrid approach, this would imply that very powerful AI systems could be developed suddenly by small teams or labs, which lack large compute budgets.

Similarly, focusing on scaling–based approaches—which seems most suitable given the prominence and success of this approach in driving the recent wave of AI progress—leads to a “compute-based” perspective on the impacts of advanced AI.[ref 138] This suggests that the key tools and levers for effective governance should focus on compute governance—provided we assume that this will remain a relevant or feasible precondition for developing frontier AI. For instance, such an approach underpins the compute-threshold definition for frontier AI, which defines advanced AI with reference to particular technical elements or inputs (such as a compute usage or FLOP threshold, dataset size, or parameter count) used in its development.[ref 139] While a useful referent, this may be an unstable proxy given that it may not reliably or stably correspond to the particular capabilities of concern.

3. Sociotechnical-change based definitions: Societal impacts of advanced AI

Focus of approach: A third cluster of definitions in advanced AI governance mostly brackets out philosophical questions of the precise form of AI systems or engineering questions of the scientific pathways towards their development. Rather, it aims at defining advanced AI in terms of different levels of societal impacts.

Many concepts in this approach have emerged from scholarship that aimed to abstract away from these architectural questions and rather explore the aggregate societal impacts of advanced AI. This includes work on AI technology’s international, geopolitical impacts[ref 140] as well as work on identifying relevant historical precedents for the technology’s societal impacts, strategic stakes, and political economy.[ref 141] Examples of this work are those that identified novel categories of unintended “structural” risks from AI as distinct from “misuse” or “accident” risks,[ref 142] or taxonomies of the different “problem logics” created by AI systems.[ref 143]

Terms: The societal-impact-centric approach to defining advanced AI includes a variety of terms, including: (strategic) general-purpose technology, general-purpose military transformation, transformative AI, radically transformative AI, AGI (economic competitiveness definition), and machine superintelligence.

Definitions and themes: While many of these terms are subject to a wide range of different definitions (see Appendix 1C), they again feature a range of common themes or patterns (see Table 5).

Suitability of approach: Concepts within the sociotechnical-change-based approach may be unsuitable iFocus of approach: Finally, a fourth cluster of terms follows a risk-based approach and focuses on critical capabilities, which certain types of advanced AI systems (whatever their underlying form or scientific architecture) might achieve or enable for human users. The development of such capabilities could then mark key thresholds or inflection points in the trajectory of society. 

Other taxonomies: Work focused on the significant potential impacts or risks of advanced AI systems is of course hardly new.[ref 150] Yet in the past years, as AI capabilities have progressed, there has been renewed and growing concern that these advances are beginning to create key threshold moments where sophisticated AI systems develop capabilities that allow them to achieve or enable highly disruptive impacts in particular domains, resulting in significant societal risks. These risks may be as diverse as the capabilities in question—and indeed discussions of these risks do not always or even mostly presume (as do many form-centric approaches) the development of general capabilities in AI.[ref 151] For instance, many argue that existing AI systems may already contribute to catastrophic risks in various domains:[ref 152] large language models (LLMs) and automated biological design tools (BDTs) may already be used to enable weaponization and misuse of biological agents,[ref 153] the military use of AI systems in diverse roles may inadvertently affect strategic stability and contribute to the risk of nuclear escalation,[ref 154] and existing AI systems’ use in enabling granular and at-scale monitoring and surveillance[ref 155] may already be sufficient to contribute to the rise of “digital authoritarianism”[ref 156] or “AI-tocracy”[ref 157], to give a few examples. 

As AI systems become increasingly advanced, they may steadily and increasingly achieve or enable further critical capabilities in different domains that could be of special significance. Indeed, as leading LLM-based AI systems have advanced in their general-purpose abilities, they have frequently demonstrated emergent abilities that are surprising even to their developers.[ref 158] This has led to growing concern that as these models continue to be scaled up[ref 159] some next generation of these systems could develop unexpected but highly dangerous capabilities if not cautiously evaluated.[ref 160] 

What are these critical capabilities?[ref 161] In some existing taxonomies, critical capabilities could include AI systems reaching key levels of performance in domains such as cyber-offense, deception, persuasion and manipulation, political strategy, building or gaining access to weapons, long-horizon planning, building new AI systems, situational awareness, self-proliferation, censorship, or surveillance,[ref 162] amongst others. Other experts have been concerned about cases where AI systems display increasing tendencies and aptitudes towards controlling or power-seeking behavior.[ref 163] Other overviews identify other sets of hazardous capabilities.[ref 164] In all these cases, the concern is that advanced AI systems that achieve these capabilities (regardless of whether they are fully general, autonomous, etc.) could enable catastrophic misuse by human owners, or could demonstrate unexpected extreme—even hazardous—behavior, even against the intentions of their human principals. 

Terms: Within the risk-based approach, there are a range of domains that could be upset by critical capabilities. A brief survey (see Table 6) can identify at least eight such capability domains—moral/philosophical, economic, legal, scientific, strategic or military, political, exponential, and (extremely) dangerous.[ref 165] Namely, these include:[ref 166] 

• Concepts relating to AI systems that achieve or enable critical moral and/or philosophical capabilities include artificial/machine consciousness, digital minds, digital people, sentient artificial intelligence, robot rights catastrophe, (negative) synthetic phenomenology, suffering risks, and adversarial technological maturity.
• Concepts relating to AI systems that achieve or enable critical economic capabilities include high-level machine intelligence, tech company singularity, and artificial capable intelligence (ACI).
• Concepts relating to AI systems that achieve or enable critical legal capabilities include advanced artificial judicial intelligence, technological-legal lock-in, and legal singularity.
• Concepts relating to AI systems that achieve or enable critical scientific capabilities include process-automating science and technology and scientist model.
• Concepts relating to AI systems that achieve or enable critical strategic and/or military capabilities include decisive strategic advantage and singleton.
• Concepts relating to AI systems that achieve or enable critical political capabilities include stable totalitarianism, value lock-in, and actually existing AI.
• Concepts related to AI systems that achieve or enable critical exponential capabilities include intelligence explosion, autonomous replication in the real world, autonomous AI research, and duplicator.
• Concepts relating to AI systems that achieve or enable (extremely) hazardous capabilities include advanced AI, high-risk AI systems, AI systems of concern, prepotent AI, APS systems, WIDGET, rogue AI, runaway AI, frontier (AI) model (under two definitional thresholds), and highly-capable systems of concern.

Definitions and themes: As noted, many of these terms have different definitions (see Appendix 1D). Nonetheless, a range of common themes and patterns can be distilled (see Table 6).

Suitability of approach: There are a range of benefits and drawbacks to defining advanced AI systems by their (critical) capabilities. These include (in no particular order): 

Benefit (1): Focuses on key capability development points of most concern: A first benefit of adopting the risk-based definitional approach is that these concepts can be used, alone or in combination, to focus on the key thresholds or transition points in AI development that we most care about—not just the aggregate eventual, long-range societal outcomes nor the (eventual) “final” form of advanced AI, but rather the key intermediate (technical) capabilities that would suffice to create (or enable actors to achieve) significant societal impacts: the points of no return.

Benefit (2): Highlighting risks and capabilities can more precisely inform the public understanding: Ensuring that terms for advanced AI systems clearly center on particular risks or capabilities can help the public and policymakers understand the risks or challenges to be avoided, in a way that is far clearer than terms that focus on very general abilities or which are highly technical (i.e., terms within essence- or development-based approaches, respectively). Such terms may also assist the public in comparing the risks of one model to those posed by another.[ref 169]

Benefit (3): Generally (but not universally) clearer or more concrete: While some terms within this approach are quite vague (e.g., “singleton”) or potentially difficult to operationalize or test for (e.g., “artificial consciousness”), some of the more specific and narrow terms within this approach could offer more clarity, and less definitional drift, to regulation. While many of them would need significant further clarification before they could be suitable for use in legislative texts (whether domestic laws or international treaties), they may offer the basis for more circumscribed, tightly defined professional cornerstone concepts for such regulation.[ref 170]

However, there are also a number of potential drawbacks to risk-based definitions.

Drawback (1): Epistemic challenges around “unknown unknown” critical capabilities: One general challenge to this risk-based approach for characterizing advanced AI is that, in the absence of more specific and empirical work, it can be hard to identify and enumerate all relevant risk capabilities in advance (or to know that we have done so). Indeed, aiming to exhaustively list out all key capabilities to watch for could be a futile exercise to undertake.[ref 171] At the same time, this is a challenge that is arguably faced in any domain of (technology) risk mitigation, and it does not mean that doing such analysis to the best of our abilities is void. However, this challenge does create an additional hurdle for regulation, as it heightens the chance that if the risk profile of the technology rapidly changes, regulators or existing legal frameworks will be unsure of how or where to classify that model.

Drawback (2): Challenges around comparing or prioritizing between risk capabilities: A related challenge lies in the difficulty of knowing which (potential) capabilities to prioritize for regulation and policy. However, that need not be a general argument against this approach. Instead, it may simply help us make explicit the normative and ethical debates over what challenges to avoid and prioritize.

Drawback (3): Utilizing many parallel terms focused on different risks can increase confusion: One risk for this approach is that while the use of many different terms for advanced AI systems, depending on their specific critical capabilities in particular domains, can make more for appropriate and context-sensitive discussions (and regulation) within those domains, at an aggregate level this may increase the range of terms that regulators and the public have to reckon with and compare between—with the risk that these actors simply drown in the range of terms.

Drawback (5): Outstanding disagreements over appropriate operationalization of capabilities: One further challenge with these terms may lie in the way that some key terms remain contested or debated—and that even clearer terms are not without challenge. For instance, in 2023, the concept of “frontier model” has become subject to increasing debate over its potential adequacy for regulation.[ref 172] Notably, there are at least three ways of operationalizing this concept. The first, computational threshold, has been discussed above.[ref 173] 

However, a second operationalization for frontier AI focuses on some relative-capabilities threshold. This approach includes recent proposals to define “frontier AI models” in terms of capabilities relative to other AI systems,[ref 174] as models that “exceed the capabilities currently present in the most advanced existing models” or as “models that are both (a) close to, or exceeding, the average capabilities of the most capable existing models.”[ref 175] Taking such a comparative approach to defining advanced AI may be useful in combating the easy tendency of observers to normalize or become used to the rapid pace of AI capability progress.[ref 176] Yet there may be risks with such a comparative approach, especially when tied to a moving wavefront of “the most capable” existing models, as this could easily impose a need on regulators to engage in constant regulatory updating, as well as creating risks of underinclusivity of some foundation models that did not display hazardous capabilities in their initial evaluations, but which once deployed or shared might be reused or recombined in ways that could create or enable significant harms.[ref 177] The risk of embedding this definition of frontier AI in regulation, would be to leave a regulatory gap around significantly harmful capabilities, especially those that are no longer at the technical “frontier,” but which remain unaddressed even so. Indeed, for similar reasons, Seger and others have advocated using the concept “highly-capable foundation models” instead.[ref 178]

A third approach to defining frontier AI models has instead focused more on identifying a set of static and absolute criteria grounded in particular dangerous capabilities (i.e., a dangerous-capabilities threshold). Such definitions might be useful insofar as they help regulators or consumers identify better when a model crosses a safety threshold and in a way that is less susceptible to slippage or change over time. This could make such concepts more suitable (and resulting regulations less at risk of obsolescence or governance misspecification) than operationalizations of “frontier AI model” that rely on indirect technological metrics (such as compute thresholds) as proxies for these capabilities. Even so, as discussed above, anchoring the “frontier AI model” concept on particular dangerous capabilities leaves open questions around how to best operationalize and create evaluation suites that are able to identify or predict such capabilities ex ante.

Given this, while the risk-based approach may be the most promising ground for defining advanced AI systems from a regulatory perspective, it is clear that not all terms in use in this approach are equally suitable, and many require further operationalization and clarification.

III. Defining the advanced AI governance epistemic community

Beyond the object of concern of “advanced AI” (in all its diverse forms), researchers in the emerging field concerned with the impacts and risks of advanced AI systems have begun to specify a range of other terms and concepts, relating to the tools for intervening in and on the development of advanced AI systems in socially beneficial ways, terms by which this community’s members conceive of the overarching approach or constitution of their field, and theories of change. 

1. Defining the tools for policy intervention

First off, those writing about the risks and regulation of AI have proposed a range of terms in describing the tools, practices, or nature of governance interventions that could be used in response (see Table 7).

Like the term “advanced AI”, these terms set out objects of study in scoping the practices or tools of AI governance. They matter insofar as they link these terms to tools for intervention. 

Nonetheless, these terms do not capture the methodological dimension of how different approaches to advanced AI governance have approached these issues—nor the normative question of why different research communities have been driven to focus on the challenges from advanced AI in the first place.[ref 180]

2. Defining the field of practice: Paradigms

Thus, we can next consider different ways that practitioners have defined the field of advanced AI governance.[ref 181] Researchers have used a range of terms to describe the field of study that focuses on understanding the trajectory to forms of, or impacts of advanced AI and how to shape these. While these have significant overlaps in practice, it is useful to distinguish some key terms or framings of the overall project (Table 8).

However, while these terms show some different focus and emphasis, and different normative commitments, this need not preclude an overall holistic approach. To be sure, work and researchers in this space often hold diverse expectations about the trajectory, form, or risks of future AI technologies; diverse normative commitments and motivations for studying these; and distinct research methodologies given their varied disciplinary backgrounds and epistemic precommitments.[ref 184] However, even so, many of these communities remain united by a shared perception of the technology’s stakes—the shared view that shaping the impacts of AI is and should be a significant global priority.[ref 185] 

As such, one takeaway here is not that scholars or researchers need pick any one of these approaches or conceptions of the field. Rather, there is a significant need for any advanced AI governance field to maintain a holistic approach, which includes many distinct motivations and methodologies. As suggested by Dafoe, 

“AI governance would do well to emphasize scalable governance: work and solutions to pressing challenges which will also be relevant to future extreme challenges. Given all this potential common interest, the field of AI governance should be inclusive to heterogenous motivations and perspectives. A holistic sensibility is more likely to appreciate that the missing puzzle pieces for any particular challenge could be found scattered throughout many disciplinary domains and policy areas.”[ref 186] 

In this light, one might consider and frame advanced AI governance as an inclusive and holistic field, concerned with, broadly, “the study and shaping of local and global governance systems—including norms, policies, laws, processes, and institutions—that affect the research, development, deployment, and use of existing and future AI systems, in ways that help the world choose the role of advanced AI systems in its future, and navigate the transition to that world.”

3. Defining theories of change

Finally, researchers in this field have been concerned not just with studying and understanding the strategic parameters of the development of advanced AI systems,[ref 187] but also with considering ways to intervene upon it, given particular assumptions or views about the form, trajectory, societal impacts, or risky capabilities of this technology.

Thus, various researchers have defined terms that aim to capture the connection between immediate interventions or policy proposals, and the eventual goals they are meant to secure (see Table 9).

Drawing on these terms, one might also articulate new terms that incorporate elements from the above.[ref 196] For instance, one could define a “strategic approach” as a cluster of correlated views on advanced AI governance, encompassing (1) broadly shared assumptions about the key technical and governance parameters of the challenge; (2) a broad theory of victory and impact story about what solving this problem would look like; (3) a broadly shared view of history, with historical analogies to provide comparison, grounding, inspiration, or guidance; and (4) a set of intermediate strategic goals to be pursued, giving rise to near-term interventions that would contribute to reaching these.

Conclusion

The community focused on governing advanced AI systems has developed a rich and growing body of work. However, it has often lacked clarity, not only regarding many key empirical and strategic questions, but also regarding many of its fundamental terms. This includes different definitions for the relevant object of analysis—that is, species of “advanced AI”—as well as different framings for the instruments of policy, different paradigms or approaches to the field itself, and distinct understandings of what it means to have a theory of change to guide action. 

This report has reviewed a range of terms for different analytical categories in the field. It has discussed three different purposes for seeking definitions for core terms, and why and how (under a “regulatory” purpose) the choice of terms matters to both the study and practice of AI governance. It then reviewed analytical definitions of advanced AI used across different clusters which focus on the forms or design of advanced AI systems, the (hypothesized) scientific pathways towards developing these systems, the technology’s broad societal impacts, and the specific critical capabilities achieved by particular AI systems. The report then briefly reviewed analytical definitions of the tools for intervention, such as “policy” and governance”, before discussing definitions of the field and community itself and definitions for theories of change by which to prioritize interventions. 

This field of advanced AI governance has shown a penchant for generating many concepts, with many contesting definitions. Of course, while any emerging field will necessarily engage in a struggle to define itself, this field has seen a particularly broad range of terms, perhaps reflecting the disciplinary range. Eventually, the community may need to more intentionally and deliberately commit to some terms. In the meantime, 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 about the (regulatory) implications of some of these terms. This report has aimed to provide such greater clarity in order to help provide greater context for more informed and clear discussions about questions in and around the field.

Appendix 1: Lists of definitions for advanced AI terms

This appendix provides a detailed list of definitions for advanced AI systems, with sources. These may be helpful for readers to explore work in this field in more detail; to understand the longer history and evolution of many terms; and to consider the strengths and drawbacks of particular terms, and of specific language, for use in public debate, policy formulation, or even in direct legislative texts.

1.A. Definitions focused on the form of advanced AI

Different definitional approaches emphasize distinct aspects or traits that would characterize the form of advanced AI systems—such as that it is ‘mind-like’, performs ‘autonomously’, ‘is general-purpose’, ‘performs like a human’, ‘performs general-purpose like a human’, etc. However, it should be noted that there is significant overlap, and many of these terms are often (whether or not correctly) used interchangeably.

Advanced AI is mind-like & really thinks

• Strong AI
  • An “appropriately programmed computer [that] really is a mind, in the sense that computers given the right programs can be literally said to understand and have other cognitive states.”[ref 198]
  • “The assertion that machines could possibly act intelligently (or, perhaps better, act as if they were intelligent) is called the ‘weak AI’ hypothesis by philosophers, and the assertion that machines that do so are actually thinking (as opposed to simulating thinking) is called the ‘strong AI’ hypothesis.”[ref 199]
  • “the combination of Artificial General Intelligence/Human-Level AI and Superintelligence.”[ref 200]

Advanced AI is autonomous

• Autonomous machine intelligence: “intelligent machines that learn more like animals and humans, that can reason and plan, and whose behavior is driven by intrinsic objectives, rather than by hard-wired programs, external supervision, or external rewards.”[ref 201]
• Autonomous artificial intelligence: “artificial intelligence that can adapt to external environmental challenges. Autonomous artificial intelligence can be similar to animal intelligence, called (specific) animal-level autonomous artificial intelligence, or unrelated to animal intelligence, called non-biological autonomous artificial intelligence.”[ref 202]

General artificial intelligence: “broadly capable AI that functions autonomously in novel circumstances”.[ref 203]

Advanced AI is human-like

• Human-level AI (HLAI)
  • “systems that operate successfully in the common sense informatic situation [defined as the situation where] the known facts are incomplete, and there is no a priori limitation on what facts are relevant. It may not even be decided in advance what phenomena are to be taken into account. The consequences of actions cannot be fully determined. The common sense informatic situation necessitates the use of approximate concepts that cannot be fully defined and the use of approximate theories involving them. It also requires nonmonotonic reasoning in reaching conclusions.”[ref 204] 
  • “machines exhibiting true human-level intelligence should be able to do many of the things humans are able to do. Among these activities are the tasks or ‘jobs’ at which people are employed. I suggest we replace the Turing test by something I will call the ‘employment test.’ To pass the employment test, AI programs must… [have] at least the potential [to completely automate] economically important jobs.”[ref 205]
  • “AI which can reproduce everything a human can do, approximately. […] [this] can mean either AI which can reproduce a human at any cost and speed, or AI which can replace a human (i.e. is as cheap as a human, and can be used in the same situations.)”[ref 206]
  • “An artificial intelligence capable of matching humans in every (or nearly every) sphere of intellectual activity.”[ref 207]

Advanced AI is general-purpose 

• Foundation model
  • “models trained on broad data at scale […] that are adaptable to a wide range of downstream tasks.”[ref 208]
  • “AI systems with broad capabilities that can be adapted to a range of different, more specific purposes. […] the original model provides a base (hence ‘foundation’) on which other things can be built.”[ref 209]
• General-purpose AI systems (GPAIS)
  • “an AI system that can be used in and adapted to a wide range of applications for which it was not intentionally and specifically designed.”[ref 210]
  • “An AI system that can accomplish or be adapted to accomplish a range of distinct tasks, including some for which it was not intentionally and specifically trained.”[ref 211]
  • “An AI system that can accomplish a range of distinct valuable tasks, including some for which it was not specifically trained.”[ref 212]
  • See also “general-purpose AI models”: “AI models that are designed for generality of their output and have a wide range of possible applications.”[ref 213]
• Comprehensive AI services (CAIS)

“asymptotically recursive improvement of AI technologies in distributed systems [which] contrasts sharply with the vision of self-improvement internal to opaque, unitary agents. […] asymptotically comprehensive, superintelligent-level AI services that—crucially—can include the service of developing new services, both narrow and broad, [yielding] a model of flexible, general intelligence in which agents are a class of service-providing products, rather than a natural or necessary engine of progress in themselves.”[ref 214]

Advanced AI is general-purpose & of human-level performance

• Artificial general intelligence (AGI) [task performance definitions][ref 215] 
  • “systems that exhibit the broad range of general intelligence found in humans.”[ref 216]
  • “Artificial intelligence that is not specialized to carry out specific tasks, but can learn to perform as broad a range of tasks as a human.”[ref 217]
  • AI systems with “the ability to achieve a variety of goals, and carry out a variety of tasks, in a variety of different contexts and environments.”[ref 218] 
  • AI systems which “can reason across a wide range of domains, much like the human mind.”[ref 219] 
  • “machines designed to perform a wide range of intelligent tasks, think abstractly and adapt to new situations.”[ref 220] 
  • “AI that is capable of solving almost all tasks that humans can solve.”[ref 221] 
  • “AIs that can generalize well enough to produce human-level performance on a wide range of tasks, including abstract low-data tasks.”[ref 222]
  • “The AI that […] can do most everything we humans can do, and possibly much more.”[ref 223]
  • “[a]n AI that has a level of intelligence that is either equivalent to or greater than that of human beings or is able to cope with problems that arise in the world that surrounds human beings with a degree of adequacy at least similar to that of human beings.”.[ref 224]
  • “an agent that has a world model that’s vastly more accurate than that of a human in, at least, domains that matter for competition over resources, and that can generate predictions at a similar rate or faster than a human.”[ref 225]
  • “type of AI system that addresses a broad range of tasks with a satisfactory level of performance [or in a stronger sense] systems that not only can perform a wide variety of tasks, but all tasks that a human can perform.”[ref 226]
  • “[AI with] cognitive capabilities fully generalizing those of humans.”[ref 227]
    • See also the subdefinition of autonomous AGI (AAGI) as “an autonomous artificial agent with the ability to do essentially anything a human can do, given the choice to do so—in the form of an autonomously/internally determined directive—and an amount of time less than or equal to that needed by a human.”[ref 228]
  • “a machine-based system that can perform the same general-purpose reasoning and problem-solving tasks humans can.”[ref 229]
  • “an AI system that equals or exceeds human intelligence in a wide variety of cognitive tasks.”[ref 230]
  • “AI systems that achieve or exceed human performance across a wide range of cognitive tasks”.[ref 231]
  • “hypothetical type of artificial intelligence that would have the ability to understand or learn any intellectual task that a human being can.”[ref 232]
  • “a shorthand for any intelligence […] that is flexible and general, with resourcefulness and reliability comparable to (or beyond) human intelligence.”[ref 233]
  • “systems that demonstrate broad capabilities of intelligence as […] [a very general mental capability that, among other things, involves the ability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly and learn from experience], with the additional requirement, perhaps implicit in the work of the consensus group, that these capabilities are at or above human-level.”[ref 234]
  • “autonomous artificial intelligence that reaches Human-level intelligence. It can adapt to external environmental challenges and complete all tasks that humans can accomplish, achieving human-level intelligence in all aspects.”[ref 235]

Robust artificial intelligence: “intelligence that, while not necessarily superhuman or self-improving, can be counted on to apply what it knows to a wide range of problems in a systematic and reliable way, synthesizing knowledge from a variety of sources such that it can reason flexibly and dynamically about the world, transferring what it learns in one context to another, in the way that we would expect of an ordinary adult.”[ref 236]

Advanced AI is general-purpose & beyond-human-performance

• AI+: “artificial intelligence of greater than human level (that is, more intelligent than the most intelligent human)”[ref 237]
• (Machine/Artificial) superintelligence (ASI):
  • “an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills.”[ref 238]
  • “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.”[ref 239] 
  • “an AI significantly more intelligent than humans in all respects.”[ref 240]
  • “Artificial intelligence that can outwit humans in every (or almost every) intellectual sphere.”[ref 241]
  • “future AI systems dramatically more capable than even AGI.”[ref 242]
  • “Artificial General Intelligence that has surpassed humans in all aspects of human intelligence.”[ref 243]
  • “AI that might be as much smarter than us as we are smarter than insects.”[ref 244]
  • See also “machine superintelligence” [form and impact]: “general artificial intelligence greatly outstripping the cognitive capacities of humans, and capable of bringing about revolutionary technological and economic advances across a very wide range of sectors on timescales much shorter than those characteristic of contemporary civilization.”[ref 245]
• Superhuman general-purpose AI (SGPAI): “general purpose AI systems […] that are simultaneously as good or better than humans across nearly all tasks.”[ref 246] 
• Highly capable foundation models: “Foundation models that exhibit high performance across a broad domain of cognitive tasks, often performing the tasks as well as, or better than, a human.”[ref 247]

1.B. Definitions focused on the pathways towards advanced AI

First-principles pathways: “De novo AGI”

Pathways based on new fundamental insights in computer science, mathematics, algorithms, or software, producing advanced AI systems that may, but need not mimic human cognition.[ref 248]

• De novo AGI: “AGI built from the ground up.”[ref 249]

Scaling pathways: “Prosaic AGI”, “frontier (AI) model” [compute threshold]

Approaches based on “brute forcing” advanced AI,[ref 250] by running (one or more) existing AI approaches (such as transformer-based LLMs)[ref 251] with increasingly more computing power and/or training data, as per the “scaling hypothesis.”[ref 252]

• Prosaic AGI: AGI “which can replicate human behavior but doesn’t involve qualitatively new ideas about ‘how intelligence works.’”[ref 253] 
• Frontier (AI) model [compute threshold]:[ref 254] 
  • “foundation model that is trained with more than some amount of computational power—for example, 10^26 FLOP.”[ref 255]
  • “models within one order of magnitude of GPT-4 (>2e24 FLOP).”[ref 256]

Evolutionary pathways: “[AGI] from evolution”

Approaches based on algorithms competing to mimic the evolutionary brute search process that produced human intelligence.[ref 257]

• [AGI] from evolution: “[AGI re-evolved through] genetic algorithms on computers that are sufficiently fast to recreate on a human timescale the same amount of cumulative optimization power that the relevant processes of natural selection instantiated throughout our evolutionary past.”[ref 258]

Reward-based pathways: “[AGI] from powerful reinforcement learning agents”, “powerful deep learning models”

Approaches based on running reinforcement learning systems with simple rewards in rich environments.

• [AGI] from powerful reinforcement learning agents: “powerful reinforcement learning agents, when placed in complex environments, will in practice give rise to sophisticated expressions of intelligence.”[ref 259]

Powerful deep learning models: “a powerful neural network model [trained] to simultaneously master a wide variety of challenging tasks (e.g. software development, novel-writing, game play, forecasting, etc.) by using reinforcement learning on human feedback and other metrics of performance.”[ref 260]

Bootstrapping pathways:[ref 261] “Seed AI”

Approaches that pursue a minimally intelligent core system capable of subsequent recursive (self)-improvement,[ref 262] potentially leveraging hardware or data “overhangs.”[ref 263]

• Seed AI:
  • “an AI designed for self-understanding, self-modification, and recursive self-improvement.”[ref 264] 
  • “a strongly self-improving process, characterized by improvements to the content base that exert direct positive feedback on the intelligence of the underlying improving process.”[ref 265]
  • “The first AI in a series of recursively self-improving systems.”[ref 266]

Neuro-emulated pathways: “Whole-brain-emulation” (WBE)

Approaches that aim to digitally simulate or recreate the states of human brains at fine-grained level.

• Whole-brain-emulation (WBE):
  • “software (and possibly dedicated non-brain hardware) that models the states and functional dynamics of a brain at a relatively fine-grained level of detail.”[ref 271]
  • “The process of making an exact computer-simulated copy of the brain of a particular animal (e.g., a particular human).”[ref 272]
• Digital people [emulation definition]: “a computer simulation of a specific person, in a virtual environment […] perhaps created via mind uploading (simulating human brains) [or] entities unlike us in many ways, but still properly thought of as ‘descendants’ of humanity.”[ref 273]
  • See also related terms: “Ems”[ref 274] or “uploads”.

Neuro-integrationist pathways: “Brain-computer-interfaces” (BCI)

Approaches to create advanced AI, based on merging components of human and digital cognition.

• Brain-computer-interfaces (BCI): “use brain-computer interfaces to position both elements, human and machine, to achieve (or overachieve) human goals.”[ref 275] 

Embodiment pathways:[ref 276] “Embodied agent” 

Based on providing the AI system with a robotic physical “body” to ground cognition and enable it to learn from direct experience of the world.[ref 277]

• “an embodied agent (e.g., a robot) which learns, through interaction and exploration, to creatively solve challenging tasks within its environment.”[ref 278]

Modular cognitive architecture pathways

Used in various fields, including in robotics, where researchers integrate well-tested but distinct state-of-the-art modules (perception, reasoning, etc.) to improve agent performance without independent learning.[ref 279] 

• No clear single term.

Hybrid pathways 

Approaches that rely on combining deep neural network-based approaches to AI with other paradigms (such as symbolic AI).

• Hybrid AI: “hybrid, knowledge-driven, reasoning-based approach, centered around cognitive models.”[ref 280]

1.C. Definitions focused on the aggregate societal impacts of advanced AI

(Strategic) general-purpose technology (GPT)

• “[AI systems] having an unusually broad and deep impact on the world, comparable to that of electricity, the internal combustion engine, and computers.”[ref 281]
  • This has been further operationalized as: “[this] need not emphasize only agent-like AI or powerful AI systems, but instead can examine the many ways even mundane AI could transform fundamental parameters in our social, military, economic, and political systems, from developments in sensor technology, digitally mediated behavior, and robotics. AI and associated technologies could dramatically reduce the labor share of value and increase inequality, reduce the costs of surveillance and repression by authorities, make global market structure more oligopolistic, alter the logic of the production of wealth, shift military power, and undermine nuclear stability.”[ref 282]
• See also strategic general-purpose technology: “A general purpose technology which has the potential to deliver vast economic value and substantially affect national security, and is consequently of central political interest to states, firms, and researchers.”[ref 283]

General-purpose military transformation (GMT)

• The process by which general-purpose technologies (such as electricity and AI) “influence military effectiveness through a protracted, gradual process that involves a broad range of military innovations and overall industrial productivity growth.”[ref 284]

Transformative AI (TAI):[ref 285] 

• “potential future AI that precipitates a transition comparable to (or more significant than) the agricultural or industrial revolution.”[ref 286]
• “AI powerful enough to bring us into a new, qualitatively different future.”[ref 287]
• “software which causes a tenfold acceleration in the rate of growth of the world economy (assuming that it is used everywhere that it would be economically profitable to use it).”[ref 288]
• “AI that can go beyond a narrow task … but falls short of achieving superintelligence.”[ref 289]
• “a range of possible advances with potential to impact society in significant and hard-to-reverse ways.”[ref 290]
• “Any AI technology or application with potential to lead to practically irreversible change that is broad enough to impact most important aspects of life and society.”[ref 291] 

Radically transformative AI (RTAI)

• “any AI technology or application which meets the criteria for TAI, and with potential impacts that are extreme enough to result in radical changes to the metrics used to measure human progress and well-being, or to result in reversal of societal trends previously thought of as practically irreversible. This indicates a level of societal transformation equivalent to that of the agricultural or industrial revolutions.”[ref 292]

AGI [economic competitiveness definition]

• “highly autonomous systems that outperform humans at most economically valuable work.”[ref 293]
• “AI systems that power a comparably profound transformation (in economic terms or otherwise) as would be achieved in [a world where cheap AI systems are fully substitutable for human labor].”[ref 294]
• “future machines that could match and then exceed the full range of human cognitive ability across all economically valuable tasks.”[ref 295]

Machine superintelligence [form & impact definition]

“general artificial intelligence greatly outstripping the cognitive capacities of humans, and capable of bringing about revolutionary technological and economic advances across a very wide range of sectors on timescales much shorter than those characteristic of contemporary civilization”[ref 296]

1.D. Definitions focused on critical capabilities of advanced AI systems

Systems with critical moral and/or philosophical capabilities

• Artificial/Machine consciousness:
  • “machines that genuinely exhibit conscious awareness.”[ref 297]
  • “Weakly construed, the possession by an artificial intelligence of a set of cognitive attributes that are associated with consciousness in humans, such as awareness, self-awareness, or cognitive integration. Strongly construed, the possession by an AI of properly phenomenological states, perhaps entailing the capacity for suffering.”[ref 298]
• Digital minds: “machine minds with conscious experiences, desires, and capacity for reasoning and autonomous decision-making […] [which could] enjoy moral status, i.e. rather than being mere tools of humans they and their interests could matter in their own right.”[ref 299]
• Digital people [capability definition]: “any digital entities that (a) had moral value and human rights, like non-digital people; (b) could interact with their environments with equal (or greater) skill and ingenuity to today’s people.”[ref 300]
• Sentient artificial intelligence: “artificial intelligence (capable of feeling pleasure and pain).”[ref 301]
• Robot rights catastrophe: The point where AI systems are sufficiently advanced that “some people reasonably regard [them] as deserving human or humanlike rights. [while] Other people will reasonably regard these systems as wholly undeserving of human or humanlike rights. […] Given the uncertainties of both moral theory and theories about AI consciousness, it is virtually impossible that our policies and free choices will accurately track the real moral status of the AI systems we create. We will either seriously overattribute or seriously underattribute rights to AI systems—quite possibly both, in different ways. Either error will have grave moral consequences, likely at a large scale. The magnitude of the catastrophe could potentially rival that of a world war or major genocide.”[ref 302]
• (Negative) synthetic phenomenology: “machine consciousness [that] will have preferences of their own, that […] will autonomously create a hierarchy of goals, and that this goal hierarchy will also become a part of their phenomenal self-model […] [such that they] will be able to consciously suffer,”[ref 303] creating a risk of an “explosion of negative phenomenology” (ENP) (“Suffering explosion”).[ref 304] 
• Suffering risks: “[AI that brings] about severe suffering on an astronomical scale, vastly exceeding all suffering that has existed on Earth so far.”[ref 305]
• Adversarial technological maturity:
  • “the point where there are digital people and/or (non-misaligned) AIs that can copy themselves without limit, and expand throughout space [creating] intense pressure to move – and multiply (via copying) – as fast as possible in order to gain more influence over the world.”[ref 306]
  • “a world in which highly advanced technology has already been developed, likely with the help of AI, and different coalitions are vying for influence over the world.”[ref 307]

Systems with critical economic capabilities[ref 308]

• High-level machine intelligence (HLMI):
  • “unaided machines [that] can accomplish every task better and more cheaply than human workers.”[ref 309]
  • “an AI system (or collection of AI systems) that performs at the level of an average human adult on key cognitive measures required for economically relevant tasks.”[ref 310]
  • “The spectrum of advanced AI capabilities from next-generation AI systems to artificial general intelligence (AGI). Often used interchangeably with advanced AI.”[ref 311]
• Tech company singularity: “a transition of a technology company into a fully general tech company [defined as] a technology company with the ability to become a world-leader in essentially any industry sector, given the choice to do so—in the form of agreement among its Board and CEO—with around one year of effort following the choice.”[ref 312]
• Artificial capable intelligence (ACI):
  • “AI [that] can achieve complex goals and tasks with minimal oversight.”[ref 313]
  • “a fast-approaching point between AI and AGI: ACI can achieve a wide range of complex tasks but is still a long way from being fully general.”[ref 314]

Systems with critical legal capabilities 

• Advanced artificial judicial intelligence (AAJI): “an artificially intelligent system that matches or surpasses human decision-making in all domains relevant to judicial decision-making.”[ref 315]
• Technological-legal lock-in: “hybrid human/AI judicial systems [which] risk fostering legal stagnation and an attendant loss of judicial legitimacy.”[ref 316]
• Legal singularity: “when the accumulation of a massive amount of data and dramatically improved methods of inference make legal uncertainty obsolete. The legal singularity contemplates complete law. […] the elimination of legal uncertainty and the emergence of a seamless legal order, which is universally accessible in real time.”[ref 317]

Systems with critical scientific capabilities

• Process-automating science and technology (PASTA): “AI systems that can essentially automate all of the human activities needed to speed up scientific and technological advancement.”[ref 318] 
• Scientist model: “a single unified transformative model […] which has flexible general-purpose research skills.”[ref 319]

Systems with critical strategic or military capabilities[ref 320]

• Decisive strategic advantage: “a position of strategic superiority sufficient to allow an agent to achieve complete world domination.”[ref 321]
• Singleton: [AI capabilities sufficient to support] “a world order in which there is a single decision-making agency at the highest level.”[ref 322]

Systems with critical political capabilities

• Stable totalitarianism: “AI [that] could enable a relatively small group of people to obtain unprecedented levels of power, and to use this to control and subjugate the rest of the world for a long period of time (e.g. via advanced surveillance).”[ref 323]
• Value lock-in:
  • “an event [such as the use of AGI] that causes a single value system, or set of value systems, to persist for an extremely long time.”[ref 324] 
  • “AGI [that] would make it technologically feasible to (i) perfectly preserve nuanced specifications of a wide variety of values or goals far into the future, and (ii) develop AGI-based institutions that would (with high probability) competently pursue any such values for at least millions, and plausibly trillions, of years.”[ref 325]

Actually existing AI (AEAI): A paradigm by which the broader ecosystem of AI development, on current trajectories, may produce harmful political outcomes, because “AI as currently funded, constructed, and concentrated in the economy—is misdirecting technological resources towards unproductive and dangerous outcomes. It is driven by a wasteful imitation of human comparative advantages and a confused vision of autonomous intelligence, leading it toward inefficient and harmful centralized architectures.”[ref 326]

Systems with critical exponential capabilities

• Intelligence explosion:[ref 327]
  • “explosion to ever greater levels of intelligence, as each generation of machines creates more intelligent machines in turn.”[ref 328] 
  • “a chain of events by which human-level AI leads, fairly rapidly, to intelligent systems whose capabilities far surpass those of biological humanity as a whole.”[ref 329]
• Autonomous replication in the real world: “A model that is unambiguously capable of replicating, accumulating resources, and avoiding being shut down in the real world indefinitely, but can still be stopped or controlled with focused human intervention.”[ref 330] 
• Autonomous AI research: “A model for which the weights would be a massive boost to a malicious AI development program (e.g. greatly increasing the probability that they can produce systems that meet other criteria for [AI Safety Level]-4 in a given timeframe).”[ref 331]

Duplicator: [digital people or particular forms of advanced AI that would allow] “the ability to make instant copies of people (or of entities with similar capabilities) [leading to] explosive productivity.”[ref 332]

Systems with critical hazardous capabilities

Systems that pose or enable critical levels of (extreme or even existential) risk,[ref 333] regardless of whether they demonstrate a full range of human-level/like cognitive abilities.

• Advanced AI:
  • “systems substantially more capable (and dangerous) than existing […] systems, without necessarily invoking specific generality capabilities or otherwise as implied by concepts such as ‘Artificial General Intelligence.’”[ref 334]
  • “Systems that are highly capable and general purpose.”[ref 335]
• High-risk AI system”: 
  • An AI system that is both “(a) … intended to be used as a safety component of a product, or is itself a product covered by the Union harmonisation legislation […] (b) the product whose safety component is the AI system, or the AI system itself as a product, is required to undergo a third-party conformity assessment with a view to the placing on the market or putting into service of that product […].”[ref 336] 
  • “AI systems that are used to control the operation of critical infrastructure… [in particular] highly capable systems, increasingly autonomous systems, and systems that cross the digital-physical divide.”[ref 337]
• AI systems of concern: “highly capable AI systems that are […] high in ‘Property X’ [defined as] intrinsic characteristics such as agent-like behavior, strategic awareness, and long-range planning.”[ref 338]
• Prepotent AI: “an AI system or technology is prepotent […] (relative to humanity) if its deployment would transform the state of humanity’s habitat—currently the Earth—in a manner that is at least as impactful as humanity and unstoppable to humanity.”[ref 339]
• APS Systems: AI systems with “(a) Advanced capabilities, (b) agentic Planning, and (c) Strategic awareness.”[ref 340] These systems may risk instantiating “MAPS”—“misaligned, advanced, planning, strategically aware” systems;[ref 341] also called “power-seeking AI”.[ref 342]
• WIDGET: “Wildly Intelligent Device for Generalized Expertise and Technical Skills.”[ref 343]
• “Rogue AI:
  • “an autonomous AI system that could behave in ways that would be catastrophically harmful to a large fraction of humans, potentially endangering our societies and even our species or the biosphere.”[ref 344]
  • “a powerful and dangerous AI [that] attempts to execute harmful goals, irrespective of whether the outcomes are intended by humans.”[ref 345]
• Runaway AI: “advanced AI systems that far exceed human capabilities in many key domains, including persuasion and manipulation; military and political strategy; software development and hacking; and development of new technologies […] [these] superhuman AI systems might be designed to autonomously pursue goals in the real world.”[ref 346]
• “Frontier (AI) model [relative-capabilities threshold]: 
  • “large-scale machine-learning models that exceed the capabilities currently present in the most advanced existing models, and can perform a wide variety of tasks.”[ref 347]
  • “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.”[ref 348]
• Frontier (AI) model [unexpected-capabilities threshold]: 
  • “Highly capable foundation models, which could have dangerous capabilities that are sufficient to severely threaten public safety and global security. Examples of capabilities that would meet this standard include designing chemical weapons, exploiting vulnerabilities in safety-critical software systems, synthesising persuasive disinformation at scale, or evading human control.”[ref 349]
  • “models that are both (a) close to, or exceeding, the average capabilities of the most capable existing models, and (b) different from other models, either in terms of scale, design (e.g. different architectures or alignment techniques), or their resulting mix of capabilities and behaviours.”[ref 350] 
• Highly-capable systems of concern:
  • “Highly capable foundation models […] capable of exhibiting dangerous capabilities with the potential to cause significant physical and societal-scale harm”[ref 351]

Appendix 2: Lists of definitions for policy tools and field

2.A. Terms for tools for intervention

Strategy[ref 352]

• AI strategy research: “the study of how humanity can best navigate the transition to a world with advanced AI systems (especially transformative AI), including political, economic, military, governance, and ethical dimensions.”[ref 353]
• AI strategy: “the study of big picture AI policy questions, such as whether we should want AI to be narrowly or widely distributed and which research problems ought to be prioritized.”[ref 354] 
• Long-term impact strategies: “shape the processes that will eventually lead to strong AI systems, and steer them in a safer direction.”[ref 355]
• Strategy: “the activity or project of doing research to inform interventions to achieve a particular goal. […] AI strategy is strategy from the perspective that AI is important, focused on interventions to make AI go better.”[ref 356]
• AI macrostrategy: “the study of high level questions having to do with prioritizing the use of resources on the current margin in order to achieve good AI outcomes.”[ref 357]

Policy

• AI policy: “concrete soft or hard governance measures which may take a range of forms such as principles, codes of conduct, standards, innovation and economic policy or legislative approaches, along with underlying research agendas, to shape AI in a responsible, ethical and robust manner.”[ref 358]
• AI policymaking strategy: “A research field that analyzes the policymaking process and draws implications for policy design, advocacy, organizational strategy, and AI governance as a whole.”[ref 359]

Governance

• AI governance:
  • “AI governance (or the governance of artificial intelligence) is the study of norms, policies, and institutions that can help humanity navigate the transition to a world with advanced artificial intelligence. This includes a broad range of subjects, from global coordination around regulating AI development to providing incentives for corporations to be more cautious in their AI research.”[ref 360]
  • “local and global norms, policies, laws, processes, politics, and institutions (not just governments) that will affect social outcomes from the development and deployment of AI systems.”[ref 361] 
  • “shifting and setting up incentive structures for actions to be taken to achieve a desired outcome [around AI].”[ref 362]
  • “identifying and enforcing norms for AI developers and AI systems themselves to follow. […] AI governance, as an area of human discourse, is engaged with the problem of aligning the development and deployment of AI technologies with broadly agreeable human values.”[ref 363]
  • “the study or practice of local and global governance systems—including norms, policies, laws, processes, and institutions—govern or should govern AI research, development, deployment, and use.”[ref 364] 
• Collaborative governance of AI technology: “collaboration between stakeholders specifically in the legal governance of AI technology. The stakeholders could include representatives of governments, companies, or other established groups.”[ref 365]
• AGI safety and governance practices: “internal policies, processes, and organizational structures at AGI labs intended to reduce risk.”[ref 366] 

2.B. Terms for the field of practice

AI governance

• “the field of AI governance studies how humanity can best navigate the transition to advanced AI systems, focusing on the political, economic, military, governance, and ethical dimensions.”[ref 367] 
• “AI governance concerns how humanity can best navigate the transition to a world with advanced AI systems. It relates to how decisions are made about AI, and what institutions and arrangements would help those decisions to be made well.”[ref 368] 
• “AI governance refers (1) descriptively to the policies, norms, laws, and institutions that shape how AI is built and deployed, and (2) normatively to the aspiration that these promote good decisions (effective, safe, inclusive, legitimate, adaptive). […] governance consists of much more than acts of governments, also including behaviors, norms, and institutions emerging from all segments of society. In one formulation, the field of AI governance studies how humanity can best navigate the transition to advanced AI systems.”[ref 369]

Transformative AI governance

• “[governance that] includes both long-term AI and any nearer-term forms of AI that could affect the long-term future [and likewise] includes governance activities in both the near-term and the long-term that could affect the long-term future.”[ref 370]

Longterm(ist) AI governance

• Long-term AI governance: “[governance that] includes both long-term AI and any nearer-term forms of AI that could affect the long-term future [and likewise] includes governance activities in both the near-term and the long-term that could affect the long-term future.”[ref 371]
• Longtermist AI governance:
  • “longtermism-motivated AI governance / strategy / policy research, practice, advocacy, and talent-building.”[ref 372]
  • “the subset of [AI governance] work that is motivated by a concern for the very long-term impacts of AI. This overlaps significantly with work aiming to govern transformative AI (TAI).”[ref 373]
  • “longtermist AI governance […] which is intellectually and sociologically related to longtermism […] explicitly prioritizes attention to considerations central to the long-term trajectory for humanity, and thus often to extreme risks (as well as extreme opportunities).”[ref 374]

Appendix 3: Auxiliary definitions and terms

Beyond this, it is also useful to clarify a range of auxiliary definitions that can support analysis in the advanced AI governance field. These include, but are not limited to:[ref 375]

• Strategic parameters: Features of the world that significantly determine the strategic nature of the advanced AI governance challenge. These parameters serve as highly decision-relevant or even crucial considerations, determining which interventions or solutions are appropriate, necessary, viable, or beneficial to addressing the advanced AI governance challenge; accordingly, different views of these underlying strategic parameters constitute underlying cruxes for different theories of actions and approaches. This encompasses different types of parameters:
  • technical parameters (e.g., advanced AI development timelines and trajectories, threat models, and feasibility of alignment solution), 
  • deployment parameters (e.g., the distribution and constitution of actors developing advanced AI systems), and
  • governance parameters (e.g., the relative efficacy and viability of different governance instruments).[ref 376] 
• Key actor: An actor whose key decisions will have significant impact on shaping the outcomes from advanced AI, either directly (first-order), or by strongly affecting such decisions made by other actors (second-order).
• Key decision: A choice or series of choices by a key actor to use its levers of governance, in ways that directly affect beneficial advanced AI outcomes, and which are hard to reverse. This can include direct decisions about deployment or testing during a critical moment, but also includes many upstream decisions (such as over whether to initiate risky capabilities).
• Lever (of governance):[ref 377] A tool or intervention that can be used by key actors to shape or affect (1) the primary outcome of advanced AI development, (2) key strategic parameters of advanced AI governance, and (3) other key actors’ choices or key decisions.
• Pathway (to influence): A tool or intervention by which other actors (that are not themselves key actors) can affect, persuade, induce, incentivize, or require key actors to make certain key decisions. This can include interventions that ensure that certain levers of control are (not) used, or used in particular ways. 
• (Decision-relevant) asset: Resources that can be used by other actors in pursuing pathways of influence to key actors, and that aim to induce how these key actors make key decisions (e.g., about whether or how to use their levers). This includes new technical research insights, worked-out policy products; networks of direct influence, memes, or narratives;
• (Policy) product: A subclass of assets; specific legible proposals that can be presented to key actors.
• Critical moment(s): High-leverage[ref 378] moments where high-impact decisions are made by some actors on the basis of the available decision-relevant assets, which affect whether beneficial advanced AI outcomes are within reach. These critical moments may occur during a public “AI crunch time,”[ref 379] but they may also occur potentially long in advance (if they lock in choices or trajectories).
• “Beneficial” AI outcomes: The desired and/or non-catastrophic societal outcomes from AI technology. This is a complex normative question, which one may aim to derive by some external moral standard or philosophy,[ref 380] through social choice theory,[ref 381] or through some legitimate (e.g., democratic) process by key stakeholders themselves.[ref 382] However, this concept is often undertheorized and needs significantly more work, scholarship, and normative and public deliberation.

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://law-ai.org/international-ai-institutions
• Maas, Matthijs, “AI is like… A literature review of AI metaphors and why they matter for policy.” Institute for Law & AI. AI Foundations Report 2. (October 2023). https://law-ai.org/ai-policy-metaphors
• 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 AI institutions

Executive summary

This literature review examines a range of institutional models that have been proposed for the international governance of artificial intelligence (AI). The review specifically focuses on proposals that would involve creating new international institutions for AI. As such, it focuses on seven models for international AI institutions with distinct functions. 

Part I consists of the literature review. For each model, we provide (a) a description of each model’s functions and types, (b) the most common examples of the model, (c) some underexplored examples that are not (often) mentioned in the AI governance literature but that show promise, (d) a review of proposals for applying that model to the international regulation of AI, and (e) critiques of the model both generally and in its potential application to AI. 

Part II briefly discusses some considerations for further research concerning the design of international institutions for AI, including the effectiveness of each model at accomplishing its aims, treaty-based regulatory frameworks, other institutional models not covered in this review, the compatibility of institutional functions, and institutional options to host a new international AI governance body.

Overall, the review covers seven models, as well as thirty-five common examples of those models, twenty-four additional examples, and forty-nine proposals of new AI institutions based on those models. Table 1 summarizes these findings.[ref 1]

Introduction

Recent and ongoing progress in artificial intelligence (AI) technology has highlighted that AI systems will have increasingly significant global impacts. In response, the past year has seen intense attention to the question of how to regulate these technologies, both at domestic and international levels. As part of this process, there have been renewed calls for establishing new international institutions to carry out much-needed governance functions and anchor international collaboration on managing the risks as well as realizing the benefits of this technology.

This literature review examines and categorizes a wide range of institutions that have been proposed to carry out the international governance of AI.[ref 2] Before reviewing these models, however, it is important to situate proposals to establish a new international institution on AI within the broader landscape of approaches to the global governance of AI. Not all approaches to AI governance focus on creating new institutions. Rather, the institutional approach is only one of several different approaches to international AI governance—each of them concentrating on different governance challenges posed by AI, and each of them providing different solutions.[ref 3] These approaches include: 

(1) Rely on unilateral extraterritorial regulation. The extraterritorial approach foregoes (or at least does not prioritize) the multilateral pursuit of international regimes, norms, or institutions. Rather, it aims to enact effective domestic regulations on AI developments and then rely on the direct or extraterritorial effects of such regulations to affect the conditions or standards for AI governance in other jurisdictions. As such, this approach includes proposals to first regulate AI within (key) countries, whether by existing laws,[ref 4] through new laws or standards developed by existing institutions, or through new domestic institutions (such as a US “AI Control Council”[ref 5] or a National Algorithms Safety Board[ref 6]). These national policy levers[ref 7] can unilaterally affect the global approach to AI, either directly—for instance, through the effect of export controls on chokepoints in the AI chip supply chains[ref 8]—or because of the way such regulations can spill over to other jurisdictions, as seen in discussions of a “Brussels Effect,” a “California Effect,” or even a “Beijing Effect.”[ref 9]

(2) Apply existing international institutions, regimes, or norms to AI. The norm-application-focused approach argues that because much of international law establishes broad, technology-neutral principles and obligations, and many domains are already subject to a wide set of overlapping institutional activities, AI technology is in fact already adequately regulated in international law.[ref 10] As such, AI governance does not need new institutions or novel institutional models; rather, the aim is to reassert, reapply, extend, and clarify long-existing international institutions and norms. This is one approach that has been taken (with greater and lesser success) to address the legal gaps initially created by some past technologies, such as submarine warfare,[ref 11] cyberwar,[ref 12] or data flows within the digital economy,[ref 13] amongst others. This also corresponds to the approach taken by many international legal scholars, who argue that states should simply recognize that AI is already covered and regulated by existing norms and doctrines in international law, such as the principles of International Human Rights Law,[ref 14] International Humanitarian Law, International Criminal Law,[ref 15] the doctrine of state responsibility,[ref 16] or other regimes.[ref 17]

(3) Adapt existing international institutions or norms to AI. This approach concedes that AI technology is not yet adequately or clearly governed under international law but holds that existing international institutions could still be adapted to take on this role and may already be doing so. This approach includes proposals that center on mapping, supporting, and extending the existing AI-focused activities of existing international regimes and institutions such as the IMO, ICAO, ITU,[ref 18] various UN agencies,[ref 19] or other international organizations.[ref 20] Others explore proposals for refitting existing institutions, such as expanding the G20 with a Coordinating Committee for the Governance of Artificial Intelligence[ref 21] or changing the mandate or composition of UNESCO’s International Research Centre of Artificial Intelligence (ICRAI) or the International Electrotechnical Commission (IEC),[ref 22] to take up a stronger role in AI governance. Finally, others explore how either states (through Explanatory Memoranda or treaty reservations) or treaty bodies (through Working Party Resolutions) could adapt existing treaty regimes to more clearly cover AI systems.[ref 23] The emphasis here is on a “decentralized but coordinated” approach that supports institutions to adapt to AI,[ref 24] rather than necessarily aiming to establish new institutions in an already-crowded existing international “regime complex.”[ref 25] 

(4) Create new international institutions to regulate AI based on the model of past or existing institutions. The institution-re-creating approach argues that AI technology does need new, distinct international institutions to be adequately governed. However, in developing designs or making the case for such institutions, this approach often points to the precedent of past or existing international institutions and regimes that have a similar model. 

(5) Create entirely novel international institutional models to regulate AI. This approach argues not only that AI technology needs new international institutions, but also that past or existing international institutions (mostly) do not provide adequate models to narrowly follow or mimic.[ref 26] This is potentially reflected in some especially ambitious proposals for comprehensive global AI regimes or in suggestions to introduce entirely new mechanisms (e.g., “regulatory markets”[ref 27]) to governance.

In this review, we specifically focus on proposals for international AI governance and regulation that involve creating new international institutions for AI. That is to say, our main focus is on approach 4 and, to a lesser extent, approach 5. 

We focus on new institutions because they might be better positioned to respond to the novelty, stakes, and technical features of advanced AI systems.[ref 28] Indeed, the current climate of global attention on AI seems potentially more supportive of establishing new landmark institutions for AI than has been the case in past years. As AI capabilities progress at an unexpected rate, multiple government representatives and entities[ref 29] as well as international organizations[ref 30] have recently stated their support towards a new international AI governance institution. Additionally, the idea of establishing such institutions has taken root among many of the leading actors in the AI industry.[ref 31] 

With this, our review comes with two caveats. In the first place, our focus on this institutional approach above others does not mean that pursuing the creation of new institutions is necessarily an easy strategy or more feasible than the other approaches listed above. Indeed, proposals for new treaty regimes or international institutions for AI—especially when they draw analogies with organizations that were set up decades ago—may often underestimate how much the ground of global governance has changed in recent years. As such, they do not always reckon fully with the strong trends and forces in global governance which, for better or worse, have come to frequently push states towards relying on extending existing norms (approach 2) or adapting existing institutions (approach 3)[ref 32] rather than creating novel institutions. Likewise, there are further trends shifting US policy towards pursuing international cooperation through nonbinding international agreements rather than treaties[ref 33] as well as concerns that by some trends, international organizations may be taking up a less central role in international relations today than they have in the past.[ref 34] All of these trends should temper, or at least inform, proposals to establish new institutions. 

Furthermore, even if one is determined to pursue establishing a new international institution along one of the models discussed here, many key open questions remain about the optimal route to design and establish that organization, including (a) Given that many institutional functions might be required to adequately govern advanced AI systems, might there be a need for “hybrid” or combined institutions with a dual mandate, like the IAEA?[ref 35] (b) Should an institution be tightly centralized or could it be relatively decentralized, with one or more new institutions orchestrating the AI policy activities of a constellation of many other (existing or new) organizations?[ref 36] (c) Should such an organization be established formally, or are informal club approaches adequate in the first instance?[ref 37] (d) Should voting rules within such institutions work on the grounds of consensus or simple majority? (e) What rules should govern adapting or updating the institution’s mission and mandate to track ongoing developments in AI? This review will briefly flag and discuss some of these questions in Part II but will leave many of them open for future research.

Regarding terminology, we will use both “international institution” and “international organization” interchangeably and broadly to refer to any of (a) formally established formal intergovernmental organizations (FIGOs) founded through a constituent document (e.g., WTO, WHO); (b) treaty bodies or secretariats that have a more limited mandate, primarily supporting the implementation of a treaty or regime (e.g., BWC Implementation Support Unit); and (c) ”informal IGOs” (IIGOs) that consist of loose “task groups” and coalitions of states (e.g., the G7, BRICS, G20).[ref 38] We use “model” to refer to the general cluster of institutions under discussion; we use “function” to refer to a given institutional model’s purpose or role. We use “AI proposals” to refer to the precise institutional models that are proposed for international AI governance.

I. Review of institutional models 

Below, we review a range of institutional models that have been proposed for AI governance. For each model, we discuss its general functions, different variations or forms of the model, a range of examples that are frequently invoked, and explicit AI governance proposals that follow the model. In addition, we will highlight additional examples that have not received much attention but that we believe could be promising. Finally, where applicable, we will highlight existing critiques of a given model.

Model 1: Scientific consensus-building

1.1 Functions and types: The functions of the scientific consensus-building institutional model are to (a) increase general policymaker and public awareness of an issue, and especially to (b) establish a scientific consensus on an issue. The aim of this is to facilitate greater common knowledge or shared perception of an issue amongst states, with the aim to motivate national action or enable international agreements. Overall, the goal of institutions following this model is not to establish an international consensus on how to respond or to hand down regulatory recommendations directly, but simply to provide a basic knowledge base to underpin the decisions of key actors. By design, these institutions are, or aim to be, non-political—as in the IPCC’s mantra to be “policy-relevant and yet policy-neutral, never policy-prescriptive.”[ref 39]

1.2 Common examples: Commonly cited examples of scientific consensus-building institutions include most notably the Intergovernmental Panel on Climate Change (IPCC),[ref 40] the Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services (IPBES),[ref 41] and the Scientific Assessment Panel (SAP) of the United Nations Environment Programme (UNEP).[ref 42] 

1.3 Underexplored examples: An example that has not yet been invoked in the literature but that could be promising to explore is the Antarctic Treaty’s Committee for Environmental Protection (CEP), which provides expert advice to the Antarctic Treaty Consultative Meetings and which combines scientific consensus-building models with risk-management functions, supporting the Protocol on Environmental Protection to the Antarctic Treaty.[ref 43] Another example could be the World Meteorological Organization (WMO), which monitors weather and climatic trends and makes information available.

1.4 Proposed AI institutions along this model: There have been a range of proposals for scientific consensus-building institutions for AI. Indeed, in 2018 the precursor initiative to what would become the Global Partnership on AI (GPAI) was initially envisaged by France and Canada as an Intergovernmental Panel on AI (IPAI) along the IPCC model.[ref 44] This proposal was supported by many researchers: Kemp and others suggest an IPAI that could measure, track, and forecast progress in AI, as well as its use and impacts, to “provide a legitimate, authoritative voice on the state and trends of AI technologies.”[ref 45] They argue that an IPAI could perform structural assessments every three years as well as take up quick-response special-issue assessments. In a contemporaneous paper, Mialhe proposes an IPAI model as an institution that would gather a large and global group of experts “to inform dialogue, coordination, and pave the way for efficient global governance of AI.”[ref 46]

More recently, Ho and others propose an intergovernmental Commission on Frontier AI to “establish a scientific position on opportunities and risks from advanced AI and how they may be managed,” to help increase public awareness and understanding, to “contribute to a scientifically informed account of AI use and risk mitigation [and to] be a source of expertise for policymakers.”[ref 47] Bremmer and Suleyman propose a global scientific body to objectively advise governments and international bodies on questions as basic as what AI is and what kinds of policy challenges it poses.[ref 48] They draw a direct link to the IPCC model, noting that “this body would have a global imprimatur and scientific (and geopolitical) independence […] [a]nd its reports could inform multilateral and multistakeholder negotiations on AI.”[ref 49] Bak-Coleman and others argue in favor of an Intergovernmental Panel on Information Technology, an independent, IPCC-like panel charged with studying the “impact of emerging information technologies on the world’s social, economic, political and natural systems.”[ref 50] In their view, this panel would focus on many “computational systems,” including “search engines, online banking, social-media platforms and large language models” and would have leverage to persuade companies to share key data.[ref 51] 

Finally, Mulgan and others, in a 2023 paper, propose a Global AI Observatory (GAIO) as an institution that “would provide the necessary facts and analysis to support decision-making [and] would synthesize the science and evidence needed to support a diversity of governance responses.”[ref 52] Again drawing a direct comparison to the IPCC, they anticipate that such a body could set the foundation for more serious regulation of AI through six activities: (a) a global standardized incident reporting database, (b) a registry of crucial AI systems, (c) a shared body of data and analysis of the key facts of the AI ecosystem, (d) working groups exploring global knowledge about the impacts of AI on critical areas, (e) the ability to offer legislative assistance and model laws, and (f) the ability to orchestrate global debate through an annual report on the state of AI.[ref 53] They have since incorporated this proposal within a larger “Framework for the International Governance of AI” by the Carnegie Council for Ethics in International Affairs’s Artificial Intelligence & Equality Initiative, alongside other components such as a neutral technical organization to analyze “which legal frameworks, best practices, and standards have risen to the highest level of global acceptance.”[ref 54]

1.5 Critiques of this model: One concern that has been expressed is that AI governance is currently too institutionally immature to support an IPCC-like model, since, as Roberts argues, “the IPCC […] was preceded by almost two decades of multilateral scientific assessments, before being formalised.”[ref 55] He considers that this may be a particular problem for replicating that model for AI, given that some AI risks are currently still subject to significantly less scientific consensus.[ref 56] Separately, Bak-Coleman and others argue that a scientific consensus-building organization for digital technologies would face a far more difficult research environment than the IPCC and IPBES because, as opposed to the rich data and scientifically well-understood mechanisms that characterize climate change and ecosystem degradation, research into the impacts of digital technologies often faces data access restrictions.[ref 57] Ho and others argue that a Commission on Frontier AI would face more general scientific challenges in adequately studying future risks “on the horizon,” as well as potential politicization, both of which might inhibit the ability of such a body to effectively build consensus.[ref 58] Indeed, it is possible that in the absence of decisive and incontrovertible evidence about the trajectory and risks of AI, a scientific consensus-building institution would likely struggle to deliver on its core mission and might instead spark significant scientific contestation and disagreement amongst AI researchers.

Model 2: Political consensus-building and norm-setting 

2.1 Functions and types: The function of political consensus-building and norm-setting institutions is to help states come to greater political agreement and convergence about the way to respond to a (usually) clearly identified and (ideally) agreed issue or phenomenon. These institutions’ aim is to reach the required political consensus necessary to either align national policymaking responses sufficiently well, achieving some level of harmonization that reduces trade restrictions or impedes progress towards addressing the issue, or to help begin negotiations on other institutions that establish more stringent regimes. Political consensus-building institutions do this by providing fora for discussion and debate that can aid the articulation of potential compromises between state interests and by exerting normative pressure on states towards certain goals. In a norm-setting capacity, institutions can also draw on (growing) political consensus to set and share informal norms, even if formal institutions have not yet been created. For instance, if negotiations for a regulatory or control institution are held up, slowed, or fail, political consensus-building institutions can also play a norm-setting function by establishing, as soft law, informal standards for behavior. While such norms are not as strictly specified or as enforceable as hard-law regulations are, they can still carry force and see take-up.

2.2 Common examples: There are a range of examples of political consensus-building institutions. Some of these are broad, such as conferences of parties to a treaty (also known as COPs, the most popular one being that of the United Nations Framework Convention on Climate Change [UNFCCC]).[ref 59] Many others, however, such as the Organization for Economic Co-operation and Development (OECD), the G20, and the G7, reflect smaller, at times more informal, governance “clubs,” which can often move ahead towards policy-setting more quickly because their membership is already somewhat aligned[ref 60] and because many of them have already begun to undertake activities or incorporate institutional units focused on AI developments.[ref 61]

Gutierrez and others have reviewed a range of historical cases of (domestic and global) soft-law governance that they argue could provide lessons for AI. These include a range of institutional activities, such as UNESCO’s 1997 Universal Declaration on the Human Genome and Human Rights, 2003 International Declaration on Human Genetic Data, and 2005 Universal Declaration on Bioethics and Human Rights,[ref 62] the Environmental Management System (ISO 14001), the Sustainable Forestry Practices by the Sustainable Forestry Initiative and Forest Stewardship Council, and the Leadership in Energy and Environmental Design initiative.[ref 63] Others, however, such as the Internet Corporation for Assigned Names and Numbers (ICANN), the Asilomar rDNA Guidelines, the International Gene Synthesis Consortium, the International Society for Stem Cell Research Guidelines, the BASF Code of Conduct, the Environmental Defense Fund, and the DuPont Risk Frameworks, offer greater examples of success.[ref 64] Turner likewise argues that the ICANN, which manages to develop productive internet policy, offers a model for international AI governance.[ref 65] Elsewhere, Harding argues that the 1967 Outer Space Treaty offers a telling case of a treaty regime that quickly crystallized state expectations and policies around safe innovation in a then-novel area of science.[ref 66] Finally, Feijóo and others suggest that “new technology diplomacy” on AI could involve a series of meetings or global conferences on AI, which could draw lessons from experiences such as the World Summits on the Information Society (WSIS).[ref 67]

2.3 Underexplored examples: Examples of norm-setting institutions that formulate and share relevant soft-law guidelines on technology include the International Organization for Standardization (ISO), International Electrotechnical Commission (IEC), the International Telecommunication Union (ITU), and the United Nations Commission on International Trade Law (UNCITRAL)’s Working Group on Electronic Commerce.[ref 68] Another good example of a political consensus-building and norm-setting initiative could be found in the 1998 Lysøen Declaration,[ref 69] an initiative by Canada and Norway that expanded to 11 highly committed states along with several NGOs and which kicked off a “Human Security Network” that achieved a significant and outsized global impact, including the Ottawa Treaty ban on antipersonnel mines, the Rome Treaty establishing the International Criminal Court, the Kimberley Process aimed at inhibiting the flow of conflict diamonds, and landmark Security Council resolutions on Children and Armed Conflict and Women, Peace and Security. Another norm-setting institution that is not yet often invoked in AI discussions but that could be promising to explore is the Codex Alimentarius Commission (CAC), which develops and maintains the Food and Agriculture Organization (FAO)’s Codex Alimentarius, a collection of non-enforceable but internationally recognized standards and codes of practice about various aspects of food production, food labeling, and safety. Another example of a “club” under this model which is not often mentioned but that could be influential is the BRICS group, which recently expanded from 5 to 11 members.

2.4 Proposed AI institutions along this model: Many proposals for political consensus-building institutions on AI tend to not focus on establishing new institutions, arguing instead that it is best to put AI issues on the agenda of existing (established and recognized) consensus-building institutions (e.g., the G20) or of existing norm-setting institutions (e.g., ISO). Indeed, even recent proposals for new international institutions still emphasize that these should link up well with already-ongoing initiatives, such as the G7 Hiroshima Process on AI.[ref 70] 

However, there have been proposals for new political consensus-building institutions. Erdélyi and Goldsmith propose an International AI Organisation (IAIO), “to serve as an international forum for discussion and engage in standard setting activities.”[ref 71] They argue that “at least initially, the global AI governance framework should display a relatively low level of institutional formality and use soft-law instruments to support national policymakers in the design of AI policies.”[ref 72] Moreover, they emphasize that the IAIO “should be hosted by a neutral country to provide for a safe environment, limit avenues for political conflict, and build a climate of mutual tolerance and appreciation.”[ref 73] More recently, the US’s National Security Commission on Artificial Intelligence’s final report includes a proposal for an Emerging Technology Coalition, “to promote the design, development, and use of emerging technologies according to democratic norms and values; coordinate policies and investments to counter the malign use of these technologies by authoritarian regimes; and provide concrete, competitive alternatives to counter the adoption of digital infrastructure made in China.”[ref 74] Recently, Marcus and Reuel propose the creation of an International Agency for AI (IAAI) tasked with convening experts and developing tools to help find “governance and technical solutions to promote safe, secure and peaceful AI technologies.”[ref 75] 

At the looser organizational end, Feijóo and others propose a new technology diplomacy initiative as “a renewed kind of international engagement aimed at transcending narrow national interests and seeks to shape a global set of principles.” In their view, such a framework could “lead to an international constitutional charter for AI.”[ref 76] Finally, Jernite and others propose a multi-party international Data Governance Structure, a multi-party, distributed governance arrangement for improving the global systematic and transparent management of language data at a global level, and which includes a Data Stewardship Organization in order to develop “appropriate management plans, access restrictions, and legal scholarship.”[ref 77] Other proposed organizations are also more focused on supporting states in implementing AI policy, such as through training. For instance, Turner proposes creating an International Academy for AI Law and Regulation.[ref 78]

2.5 Critiques of this model: There have not generally been many in-depth critiques of proposals for new political consensus-building or norm-setting institutions. However, some concerns focus on the difficult tradeoffs that consensus-building institutions face in deciding whether to prioritize breadth of membership and inclusion or depth of mission alignment. Institutions that aim to foster consensus across a very broad swath of actors may be very slow to reach such normative consensus, and even when they do, they may only achieve a “lowest-common-denominator” agreement.[ref 79] On the other hand, others counter that AI consensus-building institutions or fora will need to be sufficiently inclusive—in particular, and possibly controversially, with regard to China[ref 80]—if they do not want to run the risk of producing a fractured and ineffective regime, or even see negotiations implode over the political question of who was invited or excluded.[ref 81] Finally, a more foundational challenge to political consensus-building institutions is that while it may result in (the appearance of) joint narratives, this may not have much teeth if the agreement is not binding.[ref 82]

Model 3: Coordination of policy and regulation

3.1 Functions and types: The functions of this institutional model are to help align and coordinate policies, standards, or norms[ref 83] in order to ensure a coherent international approach to a common problem. There is significant internal variation in the setup of institutions under this model, with various subsidiary functions. For instance, such institutions may (a) directly regulate the deployment of a technology in relative detail, requiring states to comply with and implement those regulations at the national level; (b) assist states in the national implementation of agreed AI policies; (c) focus on the harmonization and coordination of policies; (d) focus on the certification of industries or jurisdictions to ensure they comply with certain standards; or (e) in some cases, take up functions related to monitoring and enforcing norm compliance.

3.2 Common examples: Common examples of policy-setting institutions include the World Trade Organization (WTO) as an exemplar of an empowered, centralized regulatory institution.[ref 84] Other examples given of regulatory institutions include the International Civil Aviation Organization (ICAO), the International Maritime Organization (IMO), the International Atomic Energy Agency (IAEA), and the Financial Action Task Force (FATF).[ref 85] Examples of policy-coordinating institutions may include the United Nations Environment Programme (UNEP), which synchronized international agreements on the environment and facilitated new agreements, including the 1985 Vienna Convention for the Protection of the Ozone Layer.[ref 86] Nemitz points to the example of the institutions created under the United Nations Convention on the Law of the Sea (UNCLOS) as a model for an AI regime, including an international court to enforce the proposed treaty.[ref 87] Finally, Sepasspour proposes the establishment of an “AI Ethics and Safety Unit” within the existing International Electrotechnical Commission (IEC), under a model that is “inspired by the Food and Agriculture Organization’s (FAO) Food Safety and Quality Unit and Emergency Prevention System for Food Safety early warning system.”[ref 88]

3.3 Underexplored examples: Examples that are not yet often discussed but that could be useful or insightful include the European Monitoring and Evaluation Programme (EMEP), which implements the 1983 Convention on Long-Range Transboundary Air Pollution—a regime that has proven particularly adaptive.[ref 89] A more sui generis example is that of international financial institutions, like the World Bank or the International Monetary Fund (IMF), which tend to shape domestic policy indirectly through conditional access to loans or development fund

3.4 Proposed AI institutions along this model: Specific to advanced AI, recent proposals for regulatory institutions include Ho and other’s Advanced AI Governance Organisation, which “could help internationalize and align efforts to address global risks from advanced AI systems by setting governance norms and standards, and assisting in their implementation.”[ref 90] 

Trager and others propose an International AI Organization (IAIO) to certify jurisdictions’ compliance with international oversight standards. These would be enforced through a system of conditional market access in which trade barriers would be imposed on jurisdictions which are not certified or whose supply chains integrate AI from non-IAIO certified jurisdictions. Among other advantages, the authors suggest that this system could be less vulnerable to proliferation of industry secrets by having states establish their own domestic regulatory entities rather than having international jurisdictional monitoring (as is the case with the IAEA). However, the authors also propose that the IAIO could provide monitoring services to governments that have not yet built their own monitoring capabilities. The authors argue that their model has several advantages over others, including agile standard-setting, monitoring, and enforcement.[ref 91] 

In a regional context, Stix proposes an EU AI Agency which, among other roles, could be an analyzer of gaps in AI policy and a developer of policies that could fill such gaps. For this agency to be effective, Stix suggests it should be independent from political agendas by, for instance, having a mandate that does not coincide with election cycles.[ref 92] Webb proposes a “Global Alliance on Intelligence Augmentation” (GAIA), which would bring together experts from different fields to set best practices for AI.[ref 93] 

Chowdhury proposes a generative AI global governance body as a “consolidated ongoing effort with expert advisory and collaborations [which] should receive advisory input and guidance from industry, but have the capacity to make independent binding decisions that companies must comply with.”[ref 94] In her analysis, this body should be funded via unrestricted and unconditional funds by all AI companies engaged in the creation or use of generative AI and it should “cover all aspects of generative AI models, including their development, deployment, and use as it relates to the public good. It should build upon tangible recommendations from civil society and academic organizations, and have the authority to enforce its decisions, including the power to require changes in the design or use of generative AI models, or even halt their use altogether if necessary.”[ref 95]

A proposal for a policy-coordinating institution is Kemp and others’ Coordinator and Catalyser of International AI Law, which would be “a coordinator for existing efforts to govern AI and catalyze multilateral treaties and arrangements for neglected issues.”[ref 96]

3.5 Critiques of this model: Castel and Castel critique international conventions on the grounds that they “are difficult to monitor and control.”[ref 97] More specifically, Ho and others argue that a model like an Advanced AI Governance Organization would face challenges around its ability to set and update standards sufficiently quickly, around incentivizing state participation in adopting the regulations, and in sufficiently scoping the challenges to focus on.[ref 98] Finally, reviewing general patterns in current state activities on AI standard-setting, von Ingersleben notes that “technical experts hailing from geopolitical rivals, such as the United States and China, readily collaborate on technical AI standards within transnational standard-setting organizations, whereas governments are much less willing to collaborate on global ethical AI standards within international organizations,”[ref 99] which suggests potential thresholds to overcoming state disinterest in participating in any international institutions focused on more political and ethical standard-setting.

Model 4: Enforcement of standards or restrictions

4.1 Functions and types: The function of this institutional model is to prevent the production, proliferation, or irresponsible deployment of a dangerous or illegal technology, product, or activity. To fulfill that function, institutions under this model rely on, among other mechanisms, (a) bans and moratoria, (b) nonproliferation regimes, (c) export-control lists, (d) monitoring and verification mechanisms,[ref 100] (e) licensing regimes, and (f) registering and/or tracking of key resources, materials, or stocks. Other types of mechanisms, such as (g) confidence-building measures (CBMs), are generally transparency-enabling.[ref 101] While generally focused on managing tensions and preventing escalations,[ref 102] CBMs can also build trust amongst states in each other’s mutual compliance with standards or prohibitions, and can therefore also support or underwrite standard- and restriction-enforcing institutions.

4.2 Common examples: The most prominent example of this model, especially in discussions of institutions capable of carrying out monitoring and verification roles, is the International Atomic Energy Agency (IAEA)[ref 103]—in particular, its Department of Safeguards. Many other proposals refer to the monitoring and verification mechanisms of arms control treaties.[ref 104] For instance, Baker has studied the monitoring and verification mechanisms for different types of nuclear arms control regimes, reviewing the role of the IAEA system under Comprehensive Safeguards Agreements with Additional Protocols in monitoring nonproliferation treaties such as the Non-Proliferation Treaty (NPT) and the five Nuclear-Weapon-Free-Zone Treaties, the role of monitoring and verification arrangements in monitoring bilateral nuclear arms control limitation treaties, and the role of the International Monitoring System (IMS) in monitoring and enforcing (prospective) nuclear test bans under the Preparatory Commission for the Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO).[ref 105] Shavit likewise refers to the precedent of the NPT and IAEA in discussing a resource (compute) monitoring framework for AI.[ref 106] 

Examples given of export-control regimes include the Nuclear Suppliers Group, the Wassenaar Arrangement, and the Missile Technology Control Regime.[ref 107] As examples of CBMs, people have pointed to the Open Skies Treaty,[ref 108] which is enforced by the Open Skies Consultative Commission (OSCE). 

There are also examples of global technology control institutions that were not carried through but which are still discussed as inspirations for AI, such as the international Atomic Development Authority (ADA) proposed in the early nuclear age[ref 109] or early- to mid-20th-century proposals for the global regulation of military aviation.[ref 110]

4.3 Underexplored examples: Examples that are not yet often discussed in the context of AI but that could be promising are the Organisation for the Prohibition of Chemical Weapons (OPCW),[ref 111] the Biological Weapons Convention’s Implementation Support Unit, the International Maritime Organization (in its ship registration function), and the Convention on International Trade in Endangered Species of Wild Fauna and Flora’s (CITES) Secretariat, specifically, its database of national import and export reports.

4.4 Proposed AI institutions along this model: Proposals along this model are particularly widespread and prevalent. Indeed, as mentioned, a significant part of the literature on the international governance of AI has made reference to some sort of “IAEA for AI.” For instance, in relatively early proposals,[ref 112] Turchin and others propose a “UN-backed AI-control agency” which “would require much tighter and swifter control mechanisms, and would be functionally equivalent to a world government designed specifically to contain AI.”[ref 113] Ramamoorthy and Yampolskiy propose a “global watchdog agency” that would have the express purpose of tracking AGI programs and that would have the jurisdiction and the lawful authority to intercept and halt unlawful attempts at AGI development.[ref 114] Pointing to the precedent of both the IAEA and its inspection regime, and the Comprehensive Nuclear Test-Ban Treaty Organization (CTBTO)’s Preparatory Commission, Nindler proposes an International Enforcement Agency for safe AI research and development, which would support and implement the provisions of an international treaty on safe AI research and development, with the general mission “to accelerate and enlarge the contribution of artificial intelligence to peace, health and prosperity throughout the world [and … to ensure that its assistance] is not used in such a way as to further any military purpose.”[ref 115] Such a body would be charged with drafting safety protocols and measures, and he suggests that its enforcement could, in extreme cases, be backed up by the use of force under the UN Security Council’s Chapter VII powers.[ref 116] 

Whitfield draws on the example of the UN Framework Convention on Climate Change to propose a UN Framework Convention on AI (UNFCAI) along with a Protocol on AI that would subsequently deliver the first set of enforceable AI regulations. He proposes that these should be supported by three new bodies: an AI Global Authority (AIGA) to provide an inspection regime in particular for military AI, an associated “Parliamentary Assembly” supervisory body that would enhance democratic input into the treaty’s operations and play “a constructive monitoring role,” as well as a multistakeholder Intergovernmental Panel on AI to provide scientific, technical, and policy advice to the UNFCAI.[ref 117]

More recently,[ref 118] Ho and others propose an “Advanced AI Governance Organization” which, in addition to setting international standards for the development of advanced AI (as discussed above), could monitor compliance with these standards through, for example, self-reporting, monitoring practices within jurisdictions, or detection and inspection of large data centers.[ref 119] Altman and others propose an AIEA for Superintelligence” consisting of “an international authority that can inspect systems, require audits, test for compliance with safety standards, place restrictions on degrees of deployment and levels of security.”[ref 120] In a very similar vein, Guest (based on an earlier proposal by Karnofsky)[ref 121] calls for an “International Agency for Artificial Intelligence (IAIA)” to conduct “extensive verification through on-chip mechanisms [and] on-site inspections” as part of his proposal for a “Collaborative Handling of Artificial Intelligence Risks with Training Standards (CHARTS).”[ref 122] Drawing together elements from several models—and referring to the examples of the IPCC, Interpol, and the WTO’s dispute settlement system—Gutierrez proposes a “multilateral AI governance initiative” to mitigate “the shared large-scale high-risk harms caused directly or indirectly by AI.”[ref 123] His proposal envisions an organizational structure consisting of (a) a forum for member state representation (which adopts decisions via supermajority); (b) technical bodies, such as an external board of experts, and a permanent technical and liaison secretariat that works from an information and enforcement network and which can issue “red notice” alerts; and (c) an arbitration board that can hear complaints by non-state AI developers who seek to contest these notices as well as by member states.[ref 124]

In a 2013 paper, Wilson proposes an “Emerging Technologies Treaty”[ref 125] that would address risks from many emerging technologies. In his view, this treaty could either be housed under an existing international organization or body or established separately, and it would establish a body of experts that would determine whether there was a “reasonable grounds for concern” about AI or other dangerous research, after which states would be required to regulate or temporarily prohibit research.[ref 126] Likewise drawing on the IAEA model, Chesterman proposes an International Artificial Intelligence Agency (IAIA) as an institution with “a clear and limited normative agenda, with a graduated approach to enforcement,” arguing that “the main ‘red line’ proposed here would be the weaponization of AI—understood narrowly as the development of lethal autonomous weapon systems lacking ‘meaningful human control’ and more broadly as the development of AI systems posing a real risk of being uncontrollable or uncontainable.”[ref 127] In practice, this organization would draw up safety standards, develop a forensic capability to identify those responsible for “rogue” AI, serve as a clearinghouse to gather and share information about such systems, and provide early notification of emergencies.[ref 128] Chesterman argues that one organizational cause that could be adopted for this IAIA is to learn from the IAEA, where its Board of Governors (rather than the annual General Conference) has ongoing oversight of its operations.

Other authors endorse an institution more directly aimed at preventing or limiting proliferation of dangerous AI systems. Jordan and others propose a “NPT+” model,[ref 129] and the Future of Life Institute (FLI) proposes “international agreements to limit particularly high-risk AI proliferation and mitigate the risks of advanced AI.”[ref 130] PauseAI proposes an international agreement that sets up an “International AI Safety Agency” that would be in charge of granting approvals for deployments of AI systems and new training runs above a certain size.[ref 131] The Elders, a group of independent former world leaders, have recently called on countries to request, via the UN General Assembly, that the International Law Commission draft an international treaty to establish a new “International AI Safety Agency,”[ref 132] drawing on the models of the NPT and the IAEA, “to manage these powerful technologies within robust safety protocols [and to …] ensure AI is used in ways consistent with international law and human rights treaties.”[ref 133] More specific monitoring provisions are also entertained; for instance, Balwit briefly discusses an advanced AI chips registry, potentially organized by an international agency.[ref 134]

At the level of transparency-supporting agreements, there are many proposals for confidence-building measures for (military) AI. Such proposals focus on bilateral arrangements that build confidence amongst states and contribute to stability (as under Model 5), but which lack distinct institutions. For instance, Shoker and others discuss an “international code of conduct for state behavior.”[ref 135] Scharre, Horowitz, Khan and others discuss a range of other AI CBMs,[ref 136] including the marking of autonomous weapons systems, geographic limits, and limits on particular (e.g., nuclear) operations of AI.[ref 137] They propose to group these under an International Autonomous Incidents Agreement (IAIA) to “help reduce risks from accidental escalation by autonomous systems, as well as reduce ambiguity about the extent of human intention behind the behavior of autonomous systems.”[ref 138] In doing so, they point to the precedent of arrangements such as the 1972 Incidents at Sea Agreement[ref 139] as well as the 12th–19th century development of Maritime Prize Law.[ref 140] Imbrie and Kania propose an “Open Skies on AI” agreement.[ref 141] Bremmer & Suleyman propose a bilateral US-China regime to foster cooperation between the US and Beijing on AI, envisioning this “to create areas of commonality and even guardrails proposed and policed by a third party.”[ref 142] 

4.5 Critiques of this model: Many critiques of the enforcement model have ended up focusing (whether fairly or not) on the appropriateness of the basic analogy between nuclear weapons and AI that is explicit or implicit in proposals for an IAEA- or NPT-like regime. For instance, Kaushik and Korda have opposed what they see as aspirations to a “wholesale ban” on dangerous AI and argue that “attempting to regulate artificial intelligence indiscriminately would be akin to regulating the concept of nuclear fission itself.”[ref 143] 

Others critique the appropriateness of an IAEA-modeled approach: Stewart suggests that the focus on the IAEA’s safeguards is inadequate since AI systems cannot be safeguarded in the same way, and he suggests that, rather, better lessons might be found in the IAEA’s International Physical Protection Advisory Service (IPPAS) missions, which allow it to serve as an independent third party to assess the regulatory preparedness of countries that aim to develop nuclear programs.[ref 144] Drexel and Depp argue that even if this IAEA model could work on a technical level, it will likely be prohibitively difficult to negotiate such an intense level of oversight.[ref 145] Further, Sepasspour as well as Law note that rather than a straightforward setup, there were years of delay between the IAEA’s establishment (1957), its adoption of the INFCIRC 26 safeguards document (1961), its taking of a leading role in nuclear nonproliferation upon the adoption of the NPT (1968), and its eventual further empowerment of its verification function through the Additional Protocol (1997).[ref 146] Such a slow aggregation might not be adequate given the speed of advanced AI development. Finally, another issue is that the strength of an IAEA agency depends on the existence of supportive international treaties as well as specific incentives for participation.

Others question whether this model would be desirable, even if achievable. Howard generally critiques many governance proposals that would involve centralized control (whether domestic or global) over the proliferation of and access to frontier AI systems, arguing that such centralization would end up only advantaging currently powerful AI labs as well as malicious actors willing to steal models, with the concern that this would have significant illiberal effects.[ref 147]

Model 5: Stabilization and emergency response 

5.1 Functions and types: The function of this institutional model is to ensure that an emerging technology or an emergency does not have a negative impact on social stability and international peace.

Such institutions can serve various subsidiary functions, including (a) performing general stability management by assessing and mitigating systemic vulnerabilities that are susceptible to incidents or accidents; (b) providing early warning of—and response coordination to—incidents and emergencies, providing timely warning, and creating common knowledge of an emergency;[ref 148] (c) generally stabilizing relations, behavior, and expectations around AI technology to encourage transparency and trust around state activities in a particular domain and to avoid inadvertent military conflict.

5.2 Common examples: Examples of institutions involved in stability management include the Financial Stability Board (FSB), an entity “composed of central bankers, ministries of finance, and supervisory and regulatory authorities from around the world.”[ref 149] Another example might be the United Nations Office for Disaster Risk Reduction (UNDRR), which focuses on responses to natural disasters.[ref 150] Gutierrez invokes Interpol’s “red notice” alert system as an example of a model by which an international institution could alert global stakeholders about the dangers of a particular AI system.[ref 151]

5.3 Underexplored examples: Examples that are not yet invoked, but that could be promising examples of early warning functions include WHO’s “public health emergency of international concern” early warning mechanism and the procedure established in the IAEA’s 1986 Convention on Early Notification of a Nuclear Accident.

5.4 Proposed AI institutions along this model: AI proposals along the early warning model include Pauwels’ paper describing a Global Foresight Observatory as a multistakeholder platform aimed at fostering greater cooperation in technological and political preparedness for the impacts of innovation in various fields, including AI.[ref 152] Brenner and Suleyman propose a Geotechnology Stability Board which “could work to maintain geopolitical stability amid rapid AI-driven change” based on the coordination of national regulatory authorities and international standard-setting bodies. At other times, such a body would help prevent global technology actors from “engaging in regulatory arbitrage or hiding behind corporate domiciles.” Finally, it could also take up responsibility for governing open-source AI and censoring uploads of highly dangerous models.[ref 153] 

5.5 Critiques of this model: As there have been relatively limited numbers of proposals for this model, there are not yet many critiques. However, possible critiques might focus on the potential adequacy of relying on international institutions to respond to (rather than prevent) situations where dangerous AI systems have already seen deployment, as coordinating, communicating, and implementing effective countermeasures in those situations might either be very difficult or far too slow to respond adequately to countering a misaligned AI system. 

Model 6: International joint research

6.1 Functions and types: The function of this institutional model is to start a bilateral or multilateral collaboration between states or state entities to solve a common problem or achieve a common goal. Most institutions following this model would focus on accelerating the development of a technology or exploitation of a resource by particular actors in order to avoid races. Others would aim at speeding up the development of safety techniques. 

In some proposals, an institution like this aims not just to rally and organize a major research project, but simultaneously to include elements of an enforcing institution in order to exclude all other actors from conducting research and/or creating capabilities around that problem or goal, creating a de facto or an explicit international monopoly on an activity. 

6.2 Common examples: Examples that are pointed to as models of an international joint scientific program include the European Organization for Nuclear Research (CERN),[ref 154] ITER, the International Space Station (ISS), and the Human Genome Project.[ref 155] Example models of a (proposed) international monopoly include the Acheson-Lilienthal Proposal[ref 156] and the resulting Baruch Plan, which called for the creation of an Atomic Development Authority.[ref 157] 

6.3 Underexplored examples: Examples that are not yet discussed in the literature but that could be promising are the James Webb Space Telescope and the Laser Interferometer Gravitational-Wave Observatory (LIGO),[ref 158] which is organized internationally through the LIGO Scientific Collaboration (LSC).

6.4 Proposed AI institutions along this model: Explicit AI proposals along the joint scientific program model are various.[ref 159] Some proposals focus primarily on accelerating safety. Lewis Ho and others suggest an “AI Safety Project” to “promote AI safety R&D by promoting its scale, resourcing and coordination.” To ensure AI systems are reliable and less vulnerable to misuse, this institution would have access to significant compute and engineering capacity as well as to AI models developed by AI companies. Contrary to other international joint scientific programs like CERN or ITER, which are strictly intergovernmental, Ho and others propose that the AI Safety Project comprise other actors as well (e.g., civil society and the industry). The authors also suggest that, to prevent replication of models or diffusion of dangerous technologies, the AI Safety Project should incorporate information and security measures such as siloing information, structuring model access, and designing internal review processes.[ref 160] Neufville and Baum point out that “a clearinghouse for research into AI” could solve the collective problem of underinvestment in basic research, AI ethics, and safety research.[ref 161] More ambitiously, Ramamoorthy and Yampolskiy propose a “Benevolent AGI Treaty,” which involves “the development of AGI as a global, non-strategic humanitarian objective, under the aegis of a special agency within the United Nations.”[ref 162] 

Other proposals suggest intergovernmental collaboration for the development of AI systems more generally. Daniel Zhang and others at Stanford University’s HAI recommend that the United States and like-minded allies create a “Multilateral Artificial Intelligence Research Institute (MAIRI)” to facilitate scientific exchanges and promote collaboration on AI research—including the risks, governance, and socio-economic impact of AI—based on a foundational agreement outlining agreed research practices. The authors suggest that MAIRI could also strengthen policy coordination around AI.[ref 163] Fischer and Wenger add that a “neutral hub for AI research” should have four functions: (a) fundamental research in the field of AI, (b) research and reflection on societal risks associated with AI, (c) development of norms and best practices regarding the application of AI, and (d) further education for AI researchers. This hub could be created by a conglomerate of like-minded states but should eventually be open to all states and possibly be linked to the United Nations through a cooperation agreement, according to the authors.[ref 164] Other authors posit that an international collaboration on AI research and development should include all members of the United Nations from the start, as similar projects like the ISS or the Human Genome Project have done. They suggest that this approach might reduce the possibility of an international conflict.[ref 165] In this vein, Kemp and others call for the foundation of a “UN AI Research Organization (UNAIRO),” which would focus on “building AI technologies in the public interest, including to help meet international targets […] [a] secondary goal could be to conduct basic research on improving AI techniques in the safest, careful and responsible environment possible.”[ref 166]

Philipp Slusallek, Scientific Director of the German Research Center for Artificial Intelligence, suggests a “CERN for AI”—“a collaborative, scientific effort to accelerate and consolidate the development and uptake of AI for the benefit of all humans and our environment.” Slusallek promotes a very open and transparent design for this institution, in which data and knowledge would flow freely between collaborators.[ref 167] Similarly, the Large-scale Artificial Intelligence Open Network (LAION) calls for a CERN-like open-source collaboration among the United States and allied countries to establish an international “supercomputing research facility” hosting “a diverse array of machines equipped with at least 100,000 high-performance state-of-the-art accelerators” that can be overseen by democratically elected institutions from participating countries.[ref 168] Daniel Dewey goes a step further and suggests a potential joint international AI project with a monopoly over hazardous AI development in the same spirit of the 1946 Baruch Plan, which proposed an international Atomic Development Authority with a monopoly over nuclear activities. However, Dewey admits this proposal is possibly politically intractable.[ref 169] In another proposal for monopolized international development, Miotti suggests a “Multilateral AGI Consortium” (MAGIC), which would be an international organization mandated to run “the world’s only advanced and secure AI facility focused on safety-first research and development of advanced AI.”[ref 170] This organization would only share breakthroughs with the outside world once proven demonstrably safe and would therefore be coupled with a global moratorium on the creation of AI systems exceeding a set compute-governance threshold.

The proposals for an institution analogous to CERN discussed thus far envision a grand institution that draws talent and resources for research and development of AI projects in general. Other proposals have a narrower focus. Charlotte Stix, for example, suggests that a more decentralized version of this model could be more beneficial. Stix argues that a “European Artificial Intelligence megaproject” could be composed of a centralized headquarters to overview collaborations and provide economies of scale for AI precursors within a network of affiliated AI laboratories that conduct most of the research.[ref 171] Other authors argue that rather than focus on AI research in general, an international research collaboration could focus on the use of AI to solve problems in a specific field, such as climate change, health, privacy-enhancing technologies, economic measurement, or the sustainable development goals.[ref 172] 

6.5 Critiques of this model: In general, there have been few sustained critiques of this institutional model. However, Ho and others suggest that an international collaboration to conduct technical AI-safety research might face challenges in that it might pull safety researchers away from the frontier AI developers, reducing in-house safety expertise. In addition, there are concerns that any international project that would need to access advanced AI models would run risks over security concerns and model leaking.[ref 173]

Moreover, more fundamental critiques do exist; for instance, Kaushik and Korda critique the feasibility of a “Manhattan Project-like undertaking to address the ‘alignment problem’,” arguing that massively accelerating AI-safety research through any large-scale governmental project is infeasible. Moreover, they argue that it is an inappropriate analogy because the Manhattan Project offered a singular goal, whereas AI safety faces a situation where ‘“ten thousand researchers have ten thousand different ideas on what it means and how to achieve it.”[ref 174] 

Model 7: Distribution of benefits and access 

7.1 Functions and types: The function of this institutional model is to provide access to the benefits of a technology or a global public good to those states or individuals who do not yet have it due to geographic or economic reasons, among others. Very often, the aim of such an institution is to facilitate unrestricted access or even access schemes targeted to the most needy and deprived. When the information or goods being shared can potentially pose a risk or be misused, yet responsible access is still considered a legitimate, necessary, or beneficial goal, institutions under this model tend to create a system for conditional access. 

7.2 Common examples: Examples of unrestricted benefit-distributor institutions include international public-private partnerships like Gavi, the Vaccine Alliance, and the Global Fund to Fight AIDS, Tuberculosis and Malaria.[ref 175] Examples of conditional benefit-distributor institutions might include the IAEA’s nuclear fuel bank.[ref 176]

7.3 Underexplored examples: Examples that are not yet invoked in the AI context but that could be promising include the Nagoya Protocol’s Access and Benefit-sharing Clearing-House (ABS Clearing-House),[ref 177] the UN Climate Technology Centre and Network,[ref 178] and the United Nations Industrial Development Organization (UNIDO), which is tasked with helping build up industrial capacities in developing countries. 

7.4 Proposed AI institutions along this model: Stafford and Trager draw an analogy between the NPT and a potential international regime to govern transformative AI. The basis for this comparison is that both technologies are dual-use, both present risks even in civilian applications, and there are significant gaps in the access different states have to these technologies. Just like in the case of nuclear energy, in a scenario where there is a clear leader in the race to develop AI while others are lagging, it is mutually beneficial for the actors to enter a technology-sharing bargain. This way, the leader can ensure it will continue to be at the front of the race, while the laggards secure access to the technology. Stafford and Trager call this the “Hopeless Laggard effect.” To enforce this technology-sharing bargain in the sphere of transformative AI, an international institution would have to be created to conduct similar functions to the IAEA’s Global Nuclear Safety and Security Network, which transfers knowledge from countries with mature nuclear energy programs to those who are just starting to develop one. As an alternative, the authors suggest that the leader in AI could prevent the laggards from engaging in a race by sharing the wealth resulting from transformative AI.[ref 179] 

The US’s National Security Commission on Artificial Intelligence’s final report included a proposal for an International Digital Democracy Initiative (IDDI) “with allies and partners to align international assistance efforts to develop, promote, and fund the adoption of AI and associated technologies that comports with democratic values and ethical norms around openness, privacy, security, and reliability.”[ref 180] 

Ho and others envision a model that incorporates the private sector into the benefit-distribution dynamic. A “Frontier AI Collaborative” could spread the benefits of cutting-edge AI—including global resilience to misused or misaligned AI systems—by acquiring or developing AI systems with a pool of resources from member states and international development programs, or from AI laboratories. This form of benefit-sharing could have the additional advantage of incentivizing states to join an international AI governance regime in exchange for access to the benefits distributed by the collaborative.[ref 181] More broadly, the Elders suggest creating an institution analogous to the IAEA to guarantee that AI’s benefits are “shared with poorer countries.”[ref 182] In forthcoming work, Adan sketches key features for a Fair and Equitable Benefit Sharing Model, to “foster inclusive global collaboration in transformative AI development and ensure that the benefits of AI advancements are equitably shared among nations and communities.”[ref 183] 

7.5 Critiques of this model: One challenge faced by benefit-distributor institutions is how to balance the risk of proliferation with ensuring meaningful benefits and take-up from its technology-promotional and distributive work.[ref 184] For instance, Ho and others suggest that proposals such as their Frontier AI Collaborative proposal could risk inadvertently diffusing dangerous dual-use technologies while simultaneously encountering barriers and obstacles to effectively empowering underserved populations with AI.[ref 185] 

More fundamentally, potential challenges or concerns with global benefit- and access-providing institutions—especially those that involve some forms of conditional access—will likely see challenges (and critiques) on the basis of how they organize participation. In recent years, several researchers have argued that the global governance of AI is seeing only limited participation by states from the Global South;[ref 186] Veale and others have recently critiqued many initiatives to secure “AI for Good” or “responsible AI,” arguing that these have fallen into a “paradox of participation,” one involving “the surface-level participation of Global South stakeholders without providing the accompanying resources and structural reforms to allow them to be involved meaningfully.”[ref 187] It is likely that similar critiques will be raised against benefit-distributing institutions.

II. Directions for further research

In light of the literature review conducted in Part I, we can consider a range of additional directions for further research. Without intending to be exhaustive, this section discusses some of those directions briefly, offering some initial thoughts on the existing gaps in the current literature and how each line of research might be helpful to inform key decisions around the international governance of AI—around whether or when to create international institutions, what specific institutional models to prioritize, how to establish these institutions, and how to design them for effectiveness, amongst others.

Direction 1: Effectiveness of institutional models

In the above summary, we have outlined potential institutional models for AI without always making an assessment of their weaknesses or their effectiveness in meeting their stated goals. We believe such further analysis could be critical, however, to filter out models that would be apt to govern the risks from AI and reduce such risks de facto (not just de jure).

There is, of course, a live debate on the “effectiveness” of international law and institutions, with an extensive literature that tries to assess patterns of state compliance with different regimes in international law[ref 188] as well as more specific patterns affecting the efficacy of international organizations[ref 189] or their responsiveness to shifts in the underlying problem.[ref 190]

Such work has highlighted the imperfect track record of many international treaties in meeting their stated purposes,[ref 191] the various ways in which states may aim to evade obligations even while complying with the letter of the law,[ref 192] the ways in which states may aim to use international organizations to promote narrow national interests rather than broader organizational objectives,[ref 193] and the situations under which states aim to exit, shift away from, or replace existing institutions with new alternatives.[ref 194] Against such work, other studies have explored the deep normative changes that international norms have historically achieved in topics such as the role of territorial war,[ref 195] the transnational and domestic mechanisms by which states are pushed to commit to and comply with different treaties,[ref 196] the more nuanced conditions that may induce greater or lesser state compliance with norms or treaties,[ref 197] the effective role that even nonbinding norms may play,[ref 198] as well as arguments that a narrow focus on state compliance with international rules understates the broader effects that those obligations may have on the way that states bargain in light of those norms (even when they aim to bend them).[ref 199] Likewise, there is a larger body of foundational work that considers whether traditional international law, based in state consent, might be an adequate tool to secure global public goods such as those around AI, even if states complied with their obligations.[ref 200] 

Work to evaluate the (prospective) effectiveness of international institutions on AI could draw on this widespread body of literature to learn lessons from the successes and failures of past regimes, as well as on scholarship on the appropriate design of different bodies[ref 201] and measures to improve the decision-making performance of such organizations,[ref 202] in order to understand when or how any given institutional model might be most appropriately designed for AI.

Direction 2: Multilateral AI treaties without institutions 

While our review has focused on international AI governance proposals that would involve the establishment of some forms of international institutions, there are of course other models of international cooperation. Indeed, some types of treaties do not automatically establish distinct international organizations[ref 203] and primarily function by setting shared patterns of expectations and reciprocal behavior amongst states in order (ideally) to become self-enforcing. As discussed, our literature review omits discussing this type of regime. However, analyzing them in combination with the models we have outlined could be useful to determine international governance alternatives for AI, including whether or when state initiatives to establish such multilateral normative regimes that lack treaty bodies would likely be effective or might likely fall short.

Such an analysis could draw from a rich vein of existing proposals for new international treaties on AI. There have of course been proposals for new treaties for autonomous weapons.[ref 204] There are also proposals for international conventions to mitigate extreme risks from technology. Some of these, such as Wilson’s “Emerging Technologies Treaty”[ref 205] or Verdirame’s Treaty on Risks to the Future of Humanity,[ref 206] would address many types of potential existential risks simultaneously, including potential risks from AI. 

Other treaty proposals are focused more specifically on regulating AI risk in particular. Dewey discusses a potential “AI development convention” that would set down “a ban or set of strict safety rules for certain kinds of AI development.”[ref 207] Yet others address different types of risks from AI, such as Metzinger’s proposal for a global moratorium on artificial suffering.[ref 208] Carayannis and Draper discuss a “Universal Global Peace Treaty” (UGPT), which would commit states “not to declare or engage in interstate war, especially via existential warfare, i.e., nuclear, biological, chemical, or cyber war, including AI- or ASI-enhanced war.” They would see this treaty supported by a separate Cyberweapons and AI Convention, which would commit as its main article that “each State Party to this Convention undertakes never in any circumstances to develop, produce, stockpile or otherwise acquire or retain: (1) cyberweapons, including AI cyberweapons; and (2) AGI or artificial superintelligence weapons.”[ref 209] 

Notwithstanding these proposals, there are significant gaps in the scholarship surrounding the design of an international treaty for AI regulation. Some issues that we believe should be explored include, but are not limited to, the effects of reciprocity on the behavior of state parties, the relationship between the specificity of a treaty and its pervasiveness, the success and adaptability of the framework convention model (a broad treaty and protocols which specify the initial treaty’s obligations) in accomplishing their goals, and adjudicatory options for conflicts between state parties. 

Direction 3: Additional institutional models not covered in detail in this review

There are many other institutional models that this literature review does not address, as they are (currently) rarely proposed in the specific context of international AI governance. These include, but are not limited to:[ref 210]

• Various international non-governmental organizations (NGOs), e.g., the World Wide Fund for Nature (WWF), Amnesty International (AI);
• Political and economic unions, e.g., Association of Southeast Asian Nations (ASEAN); 
• Military alliances that establish security guarantees and/or political, economic, and defense cooperation, e.g., the North Atlantic Treaty Alliance (NATO), the Shanghai Cooperation Organisation (SCO), or the Collective Security Treaty Organization (CSTO); 
• International courts and tribunals, e.g., the International Criminal Court (ICC), various regional courts of human rights (African Court on Human and Peoples’ Rights, the European Court of Human Rights, the Inter-American Court of Human Rights);
• Interstate arbitral and dispute settlement bodies, e.g., the International Court of Justice (ICJ); the WTO Appellate Body, which hears disputes by WTO Members; the International Tribunal for the Law of the Sea (ITLOS), which is one of the dispute resolution mechanisms for the UN Convention on the Law of the Sea (UNCLOS); the Permanent Court of Arbitration (PCA), which resolves disputes arising out of international agreements between member states, international organizations, or private parties; or the European Nuclear Energy Tribunal, which oversees nuclear energy disputes within the OECD;
• Cartels aimed at articulating, aggregating, and securing the (economic) interests of their members, e.g., the Organization of the Petroleum Exporting Countries (OPEC), whose members cooperate to reduce market competition but whose operations may be protected by the doctrine of states immunity under international law;
• Policy implementation and/or direct service delivery organizations, e.g., the United Nations Development Programme (UNDP) or the World Bank;
• Data gathering and dissemination organizations, e.g., the World Meteorological Organization’s (WMO) climate data monitoring or the Food and Agriculture Organization (FAO) gathering of statistics on global food production;
• Post-disaster response and relief organizations, e.g., The World Food Programme (WFP) or the International Committee of the Red Cross (ICRC);
• Capacity-building and training organizations, e.g., governmental capacity-building trainings offered by the United Nations Institute for Training and Research (UNITAR), fiscal management training programs offered by the International Monetary Fund (IMF), or border-control trainings provided by the International Organization for Migration (IOM);
• Norm promotion organizations, e.g., the UNESCO World Heritage site program or UNHCR advocacy for refugee rights;
• Awareness-raising organizations, e.g., the Joint United Nations Programme on HIV/AIDS (UNAIDS), which amongst others organizes World AIDS Day; and
• “Meta”-organizations which aim to support or enhance the activities of other existing international organizations in general, e.g., the United Nations Office for Project Services (UNOPS).

Accordingly, future lines of research could focus on exploring what such models could look like for AI and what they might contribute to international AI governance.

Direction 4: Compatibility of institutional functions 

There are multiple instances of compatibility between the functions of institutions proposed by the literature explored in this review. Getting a better sense of those areas of compatibility could be advantageous when designing an international institution for AI that borrows the best features from each model rather than copying a single model. Further research could explore hybrid institutions that combine functions from several models. Some potential combinations include, but are not limited to:

• Comprehensive scientific consortia, which could combine elements from scientific consensus-building institutions, international joint scientific programs, and (scientific) benefit-distributing institutions;
• Full-spectrum consensus-building fora, which could combine elements from scientific consensus-building with political consensus-building institutions and potentially with stabilization and emergency-response institutions;
• Integrated regulator institutions, which could combine elements from regulatory and policy coordinator institutions with monitoring and verification institutions; and
• Centralized control institutions, which could combine elements from nonproliferation, export-control institutions, with access-controlling institutions, and potentially with monitoring and verification institutions.

Direction 5: Potential fora for an international AI organization

This review omits establishing patterns among different proposals on their preferred fora to negotiate or host an international AI organization. While we do not expect there to be much commentary on this, it might be a useful additional element to take into consideration when designing an international AI institution. For example, some fora that have been proposed are:

• The United Nations could establish a UN specialized agency through state negotiations or initiative at the UN General Assembly.[ref 211] For instance, as seen above, Kemp and others call for a UN AI Research Organization (UNAIRO).[ref 212] 
• Regional organizations, such as the European Union, the Organization of American States, the African Union, or ASEAN, could pioneer regional regulatory regimes that exert indirect extraterritorial effects on global AI governance. The European Union in particular has proven to be effective at indirectly regulating industries at a global level through the so-called Brussels Effect.[ref 213] Siegmann and Anderljung suggest that the EU AI Act could have a similar effect on the global AI industry.[ref 214]
• Similarly, minilateral club organizations like the G7, BRICS, the G20, or the OECD could play a similar role, bringing together like-minded countries to negotiate an international governance framework for AI that other states can then join.[ref 215]
• Public-private partnerships or coalitions between state and non-state actors, such as the Lysøen Initiative on human security[ref 216] or the Christchurch Call, an initiative (led by France and Aotearoa New Zealand) on eliminating online terrorist and violent extremist content,[ref 217] which can organize a coalition of like-minded states and actors to pursue the negotiation of new treaties, where necessary outside of UN fora.
• Gradual formalization of initial informal institutions: in some cases, organizations that are initially established in an informal configuration could lay the foundation for formal frameworks for cooperation, as happened with the gradual transformation of the General Agreement on Tariffs and Trade (GATT)  into the WTO, and which Erdélyi and Goldsmith suggest as one route that could be taken by an International Artificial Intelligence Organization.[ref 218]

This does not exhaust the available or feasible avenues, however. In many cases, significant additional work will have to be undertaken to evaluate these pathways in detail.

Conclusion

This literature review analyzed seven models for the international governance of AI, discussing common examples of those models, underexplored examples, specific proposals of their application to AI in existing scholarship, and critiques. We found that, while the literature covers a wide range of options for the international governance of AI, most of the time proposals are vague and do not develop the specific attributes an international institution would need to have in order to garner the benefits and curb the risks associated with AI. Thus, we proposed a series of pathways for further research that we expect should contribute to the design of such an international institution.

Also in this series

• Maas, Matthijs, ‘AI is like… A literature review of AI metaphors and why they matter for policy.’ Institute for Law & AI. AI Foundations Report 2. (October 2023). https://www.law-ai.org/ai-policy-metaphors 
• 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://www.law-ai.org/advanced-ai-gov-litrev

International governance of civilian AI

Re-evaluating GPT-4’s bar exam performance

1. Introduction

On March 14th, 2023, OpenAI launched GPT-4, said to be the latest milestone in the company’s effort in scaling up deep learning [1]. As part of its launch, OpenAI revealed details regarding the model’s “human-level performance on various professional and academic benchmarks” [1]. Perhaps none of these capabilities was as widely publicized as GPT-4’s performance on the Uniform Bar Examination, with OpenAI prominently displaying on various pages of its website and technical report that GPT-4 scored in or around the “90th percentile,” [1-3] or “the top 10% of test-takers,” [1, 2] and various prominent media outlets [4–8] and legal scholars [9] resharing and discussing the implications of these results for the legal profession and the future of AI.

Of course, assessing the capabilities of an AI system as compared to those of a human is no easy task [10–15], and in the context of the legal profession specifically, there are various reasons to doubt the usefulness of the bar exam as a proxy for lawyerly competence (both for humans and AI systems), given that, for example: (a) the content on the UBE is very general and does not pertain to the legal doctrine of any jurisdiction in the United States [16], and thus knowledge (or ignorance) of that content does not necessarily translate to knowledge (or ignorance) of relevant legal doctrine for a practicing lawyer of any jurisdiction; and (b) the tasks involved on the bar exam, particularly multiple-choice questions, do not reflect the tasks of practicing lawyers, and thus mastery (or lack of mastery) of those tasks does not necessarily reflect mastery (or lack of mastery) of the tasks of practicing lawyers.

Moreover, although the UBE is a closed-book exam for humans, GPT-4’s huge training corpus largely distilled in its parameters means that it can effectively take the UBE “open-book”, indicating that UBE may not only be an accurate proxy for lawyerly competence but is also likely to provide an overly favorable estimate of GPT-4’s lawyerly capabilities relative to humans.

Notwithstanding these concerns, the bar exam results appeared especially startling compared to GPT-4’s other capabilities, for various reasons. Aside from the sheer complexity of the law in form [17–19] and content [20–22], the first is that the boost in performance of GPT-4 over its predecessor GPT-3.5 (80 percentile points) far exceeded that of any other test, including seemingly related tests such as the LSAT (40 percentile points), GRE verbal (36 percentile points), and GRE Writing (0 percentile points) [2, 3].

The second is that half of the Uniform Bar Exam consists of writing essays[16],[ref 1] and GPT-4 seems to have scored much lower on other exams involving writing, such as AP English Language and Composition (14th-44th percentile), AP English Literature and Composition (8th-22nd percentile) and GRE Writing (~54th percentile) [1, 2]. In each of these three exams, GPT-4 failed to achieve a higher percentile performance over GPT-3.5, and failed to achieve a percentile score anywhere near the 90th percentile.

Moreover, in its technical report, GPT-4 claims that its percentile estimates are “conservative” estimates meant to reflect “the lower bound of the percentile range,” [2, p. 6] implying that GPT-4’s actual capabilities may be even greater than its estimates.

Methodologically, however, there appear to be various uncertainties related to the calculation of GPT’s bar exam percentile. For example, unlike the administrators of other tests that GPT-4 took, the administrators of the Uniform Bar Exam (the NCBE as well as different state bars) do not release official percentiles of the UBE [27, 28], and different states in their own releases almost uniformly report only passage rates as opposed to percentiles [29, 30], as only the former are considered relevant to licensing requirements and employment prospects.

Furthermore, unlike its documentation for the other exams it tested [2, p. 25], OpenAI’s technical report provides no direct citation for how the UBE percentile was computed, creating further uncertainty over both the original source and validity of the 90th percentile claim.

The reliability and transparency of this estimate has important implications on both the legal practice front and AI safety front. On the legal practice front, there is great debate regarding to what extent and when legal tasks can and should be automated [31–34]. To the extent that capabilities estimates for generative AI in the context law are overblown, this may lead both lawyers and non-lawyers to rely on generative AI tools when they otherwise wouldn’t and arguably shouldn’t, plausibly increasing the prevalence of bad legal outcomes as a result of (a) judges misapplying the law; (b) lawyers engaging in malpractice and/or poor representation of their clients; and (c) non-lawyers engaging in ineffective pro se representation.Meanwhile, on the AI safety front, there appear to be growing concerns of transparency[ref 2] among developers of the most powerful AI systems [36, 37]. To the extent that transparency is important to ensuring the safe deployment of AI, a lack of transparency could undermine our confidence in the prospect of safe deployment of AI [38, 39]. In particular, releasing models without an accurate and transparent assessment of their capabilities (including by third-party developers) might lead to unexpected misuse/misapplication of those models (within and beyond legal contexts), which might have detrimental (perhaps even catastrophic) consequences moving forward [40, 41].

Given these considerations, this paper begins by investigating some of the key methodological challenges in verifying the claim that GPT-4 achieved 90th percentile performance on the Uniform Bar Examination. The paper’s findings in this regard are fourfold. First, although GPT-4’s UBE score nears the 90th percentile when examining approximate conversions from February administrations of the Illinois Bar Exam, these estimates appear heavily skewed towards those who failed the July administration and whose scores are much lower compared to the general test-taking population. Second, using data from a recent July administration of the same exam reveals GPT-4’s percentile to be below the 69th percentile on the UBE, and ~48th percentile on essays. Third, examining official NCBE data and using several conservative statistical assumptions, GPT-4’s performance against first-time test takers is estimated to be ~62nd percentile, including 42nd percentile on essays. Fourth, when examining only those who passed the exam, GPT-4’s performance is estimated to drop to ~48th percentile overall, and ~15th percentile on essays.

Next, whereas the above four findings take for granted the scaled score achieved by GPT-4 as reported by OpenAI, the paper then proceeds to investigate the validity of that score, given the importance (and often neglectedness) of replication and reproducibility within computer science and scientific fields more broadly [42–46]. The paper successfully replicates the MBE score of 158, but highlights several methodological issues in the grading of the MPT + MEE components of the exam, which call into question the validity of the essay score (140).

Finally, the paper also investigates the effect of adjusting temperature settings and prompting techniques on GPT-4’s MBE performance, finding no significant effect of adjusting temperature settings on performance, and some significant effect of prompt engineering on model performance when compared to a minimally tailored baseline condition.

Taken together, these findings suggest that OpenAI’s estimates of GPT-4’s UBE percentile, though clearly an impressive leap over those of GPT-3.5, are likely overinflated, particularly if taken as a “conservative” estimate representing “the lower range of percentiles,” and even moreso if meant to reflect the actual capabilities of a practicing lawyer. These findings carry timely insights for the desirability and feasibility of outsourcing legally relevant tasks to AI models, as well as for the importance for generative AI developers to implement rigorous and transparent capabilities evaluations to help secure safer and more trustworthy AI.

2. Evaluating the 90th Percentile Estimate

2.1. Evidence from OpenAI

Investigating the OpenAI website, as well as the GPT-4 technical report, reveals a multitude of claims regarding the estimated percentile of GPT-4’s Uniform Bar Examination performance but a dearth of documentation regarding the backing of such claims. For example, the first paragraph of the official GPT-4 research page on the OpenAI website states that “it [GPT-4] passes a simulated bar exam with a score around the top 10% of test takers” [1]. This claim is repeated several times later in this and other webpages, both visually and textually, each time without explicit backing.[ref 3]

Similarly undocumented claims are reported in the official GPT-4 Technical Report.[ref 4] Although OpenAI details the methodology for computing most of its percentiles in A.5 of the Appendix of the technical report, there does not appear to be any such documentation for the methodology behind computing the UBE percentile. For example, after providing relatively detailed breakdowns of its methodology for scoring the SAT, GRE, SAT, AP, and AMC, the report states that “[o]ther percentiles were based on official score distributions,” followed by a string of references to relevant sources [2, p. 25].

Examining these references, however, none of the sources contains any information regarding the Uniform Bar Exam, let alone its “official score distributions” [2, p. 22-23]. Moreover, aside from the Appendix, there are no other direct references to the methodology of computing UBE scores, nor any indirect references aside from a brief acknowledgement thanking “our collaborators at Casetext and Stanford CodeX for conducting the simulated bar exam” [2, p. 18].

2.2. Evidence from GPT-4 Passes the Bar

Another potential source of evidence for the 90th percentile claim comes from an early draft version of the paper, “GPT-4 passes the bar exam,” written by the administrators of the simulated bar exam referenced in OpenAI’s technical report [47]. The paper is very well-documented and transparent about its methodology in computing raw and scaled scores, both in the main text and in its comprehensive appendices. Unlike the GPT-4 technical report, however, the focus of the paper is not on percentiles but rather on the model’s scaled score compared to that of the average test taker, based on publicly available NCBE data. In fact, one of the only mentions of percentiles is in a footnote, where the authors state, in passing: “Using a percentile chart from a recent exam administration (which is generally available online), ChatGPT would receive a score below the 10th percentile of test-takers while GPT-4 would receive a combined score approaching the 90th percentile of test-takers.” [47, p. 10]

2.3. Evidence Online

As explained by [27], The National Conference of Bar Examiners (NCBE), the organization that writes the Uniform Bar Exam (UBE) does not release UBE percentiles.[ref 5] Because there is no official percentile chart for UBE, all generally available online estimates are unofficial. Perhaps the most prominent of such estimates are the percentile charts from pre-July 2019 Illinois bar exam. Pre-2019,[ref 6] Illinois, unlike other states, provided percentile charts of their own exam that allowed UBE test-takers to estimate their approximate percentile given the similarity between the two exams [27].[ref 7]

Examining these approximate conversion charts, however, yields conflicting results. For example, although the percentile chart from the February 2019 administration of the Illinois Bar Exam estimates a score of 300 (2-3 points higher thatn GPT-4’s score) to be at the 90th percentile, this estimate is heavily skewed compared to the general population of July exam takers,[ref 8] since the majority of those who take the February exam are repeat takers who failed the July exam [52][ref 9], and repeat takers score much lower[ref 10] and are much more likely to fail than are first-timers.[ref 11]

Indeed, examining the latest available percentile chart for the July exam estimates GPT-4’s UBE score to be ~68th percentile, well below the 90th percentile figure cited by OpenAI [54].

3. Towards a More Accurate Percentile Estimate

Although using the July bar exam percentiles from the Illinois Bar would seem to yield a more accurate estimate than the February data, the July figure is also biased towards lower scorers, since approximately 23% of test takers in July nationally are estimated to be re-takers and score, for example, 16 points below first-timers on the MBE [55]. Limiting the comparison to first-timers would provide a more accurate comparison that avoids double-counting those who have taken the exam again after failing once or more.

Relatedly, although (virtually) all licensed attorneys have passed the bar,[ref 12] not all those who take the bar become attorneys. To the extent that GPT-4’s UBE percentile is meant to reflect its performance against other attorneys, a more appropriate comparison would not only limit the sample to first-timers but also to those who achieved a passing score.

Moreover, the data discussed above is based on purely Illinois Bar exam data, which (at the time of the chart) was similar but not identical to the UBE in its content and scoring [27], whereas a more accurate estimate would be derived more directly from official NCBE sources.

3.1. Methods

To account for the issues with both OpenAI’s estimate as well the July estimate, more accurate estimates (for GPT-3.5 and GPT-4) were sought to be computed here based on first-time test-takers, including both (a) first-time test-takers overall, and (b) those who passed.

To do so, the parameters for a normal distribution of scores were separately estimated for the MBE and essay components (MEE + MPT), as well as the UBE score overall.[ref 13]

Assuming that UBE scores (as well as MBE and essay subscores) are normally distributed, percentiles of GPT’s score can be directly computed after computing the parameters of these distributions (i.e. the mean and standard deviation).

Thus, the methodology here was to first compute these parameters, then generate distributions with these parameters, and then compute (a) what per- centage of values on these distributions are lower than GPT’s scores (to estimate the percentile against first-timers); and (b) what percentage of values above the passing threshold are lower than GPT’s scores (to estimate the percentile against qualified attorneys).With regard to the mean, according to publicly available official NCBE data, the mean MBE score of first-time test-takers is 143.8 [55].

As explained by official NCBE publications, the essay component is scaled to the MBE data [59], such that the two components have approximately the same mean and standard deviation [53, 54, 59]. Thus, the methodology here assumed that the mean first-time essay score is 143.8.[ref 14]

Given that the total UBE score is computed directly by adding MBE and essay scores [60], an assumption was made that mean first-time UBE score is 287.6 (143.8 + 143.8).

With regard to standard deviations, information regarding the SD of first- timer scores is not publicly available. However, distributions of MBE scores for July scores (provided in 5 point-intervals) are publicly available on the NCBE website [58].

Under the assumption that first-timers have approximately the same SD as that of the general test-taking population in July, the standard deviation of first-time MBE scores was computed by (a) entering the publicly available distribution of MBE scores into R; and (b) taking the standard deviation of this distribution using the built-in sd() function (which calculates the standard deviation of a normal distribution).

Given that, as mentioned above, the distribution (mean and SD) of essay scores is the same as MBE scores, the SD for essay scores was computed similarly as above.

With regard to the UBE, Although UBE standard deviations are not publicly available for any official exam, they can be inferred from a combination of the mean UBE score for first-timers (287.6) and first-time pass rates.

For reference, standard deviations can be computed analytically as follows:

Where:

• x is the quantile (the value associated with a given percentile, such as a cutoff score),
• µ is the mean,
• z is the z-score corresponding to a given percentile,
• σ is the standard deviation.

Thus, by (a) subtracting the cutoff score of a given administration (x) from the mean (µ); and (b) dividing that by the z-score (z) corresponding to the percentile of the cutoff score (i.e., the percentage of people who did not pass), one is left with the standard deviation (σ).

Here, the standard deviation was calculated according to the above formula using the official first-timer mean, along with pass rate and cutoff score data from New York, which according to NCBE data has the highest number of examinees for any jurisdiction [61].[ref 15]

After obtaining these parameters, distributions of first-timer scores for the MBE component, essay component, and UBE overall were computed using the built-in rnorm function in R (which generates a normal distribution with a given mean and standard deviation).

Finally, after generating these distributions, percentiles were computed by calculating (a) what percentage of values on these distributions were lower than GPT’s scores (to estimate the percentile against first-timers); and (b) what percentage of values above the passing threshold were lower than GPT’s scores (to estimate the percentile against qualified attorneys).

With regard to the latter comparison, percentiles were computed after re- moving all UBE scores below 270, which is the most common score cutoff for states using the UBE [62]. To compute models’ performance on the individual components relative to qualified attorneys, a separate percentile was likewise computed after removing all subscores below 135.[ref 16]

3.2. Results

3.2.1. Performance against first-time test-takers

Results are visualized in Tables 1 and 2. For each component of the UBE, as well as the UBE overall, GPT-4’s estimated percentile among first-time July test takers is less than that of both the OpenAI estimate and the July estimate that include repeat takers.

With regard to the aggregate UBE score, GPT-4 scored in the 62nd percentile as compared to the ~90th percentile February estimate and the ~68th percentile July estimate. With regard to MBE, GPT-4 scored in the ~79th percentile as compared to the ~95th percentile February estimate and the 86th percentile July estimate. With regard to MEE + MPT, GPT-4 scored in the ~42nd percentile as compared to the ~69th percentile February estimate and the ~48th percentile July estimate.

With regard to GPT-3.5, its aggregate UBE score among first-timers was in the ~2nd percentile, as compared to the ~2nd percentile February estimate and

~1st percentile July estimate. Its MBE subscore was in the ~6th percentile, compared to the ~10th percentile February estimate ~7th percentile July estimate. Its essay subscore was in the ~0th percentile, compared to the ~1st percentile February estimate and ~0th percentile July estimate.

3.2.2. Performance against qualified attorneys

Predictably, when limiting the sample to those who passed the bar, the models’ percentile dropped further.

With regard to the aggregate UBE score, GPT-4 scored in the ~45th per- centile. With regard to MBE, GPT-4 scored in the ~69th percentile, whereas for the MEE + MPT, GPT-4 scored in the ~15th percentile.

With regard to GPT-3.5, its aggregate UBE score among qualified attorneys was 0th percentile, as were its percentiles for both subscores.

4. Re-Evaluating the Raw Score

So far, this analysis has taken for granted the scaled score achieved by GPT-4 as reported by OpenAI—that is, assuming GPT-4 scored a 298 on the UBE, is the 90th-percentile figure reported by OpenAI warranted?

However, given calls for the replication and reproducibility within the practice of science more broadly [42–46], it is worth scrutinizing the validity of the score itself—that is, did GPT-4 in fact score a 298 on the UBE?

Moreover, given the various potential hyperparameter settings available when using GPT-4 and other LLMs, it is worth assessing whether and to what extent adjusting such settings might influence the capabilities of GPT-4 on exam performance.

To that end, this section first attempts to replicate the MBE score reported by [1] and [47] using methods as close to the original paper as reasonably feasible. The section then attempts to get a sense of the floor and ceiling of GPT-4’s out-of-the-box capabilities by comparing GPT-4’s MBE performance using the best and worst hyperparameter settings.

Finally, the section re-examines GPT-4’s performance on the essays, eval- uating (a) the extent to which the methodology of grading GPT-4’s essays deviated that from official protocol used by the National Conference of Bar Examiners during actual bar exam administrations; and (b) the extent to which such deviations might undermine one’s confidence in the the scaled essay scores reported by [1] and [47].

4.1. Replicating the MBE Score

4.1.1. Methodology

Materials. As in [47], the materials used here were the official MBE questions released by the NCBE. The materials were purchased and downloaded in pdf format from an authorized NCBE reseller. Afterwards, the materials were converted into TXT format, and text analysis tools were used to format the questions in a way that was suitable for prompting, following [47]

Procedure. To replicate the MBE score reported by [1], this paper followed the protocol documented by [47], with some minor additions for robustness purposes. In [47], the authors tested GPT-4’s MBE performance using three different temperature settings: 0, .5 and 1. For each of these temperature settings, GPT- 4’s MBE performance was tested using two different prompts, including (1) a prompt where GPT was asked to provide a top-3 ranking of answer choices, along with a justification and authority/citation for its answer; and (2) a prompt where GPT-4 was asked to provide a top-3 ranking of answer choices, without providing a justification or authority/citation for its answer.

For each of these prompts, GPT-4 was also told that it should answer as if it were taking the bar exam.

For each of these prompts / temperature combinations, [47] tested GPT-4 three different times (“experiments” or “trials”) to control for variation.

The minor additions to this protocol were twofold. First, GPT-4 was tested under two additional temperature settings: .25 and .7. This brought the total temperature / prompt combinations to 10 as opposed to 6 in the original paper. Second, GPT-4 was tested 5 times under each temperature / prompt combination as opposed to 3 times, bringing the total number of trials to 50 as opposed to 18.

After prompting, raw scores were computed using the official answer key provided by the exam. Scaled scores were then computed following the method outlined in [63], by (a) multiplying the number of correct answers by 190, and dividing by 200; and (b) converting the resulting number to a scaled score using a conversion chart based on official NCBE data.

After scoring, scores from the replication trials were analyzed in comparison to those from [47] using the data from their publicly available github repository.To assess whether there was a significant difference between GPT-4’s accuracy in the replication trials as compared to the [47] paper, as well as to assess any significant effect of prompt type or temperature, a mixed-effects binary logistic regression was conducted with: (a) paper (replication vs original), temperature and prompt as fixed effects[ref 17]; and (b) question number and question category as random effects. These regressions were conducted using the lme4 [64] and lmertest [65] packages from R.

4.1.2. Results

Results are visualized in Table 4. Mean MBE accuracy across all trials in the replication here was 75.6% (95% CI: 74.7 to 76.4), whereas the mean accuracy across all trials in [47] was 75.7% (95% CI: 74.2 to 77.1).[ref 18]

The regression model did not reveal a main effect of “paper” on accuracy (p=.883), indicating that there was no significant difference between GPT-4’s raw accuracy as reported by [47] and GPT-4’s raw accuracy as performed in the replication here.

There was also no main effect of temperature (p>.1)[ref 19] or prompt (p=.741). That is, GPT-4’s raw accuracy was not significantly higher or lower at a given temperature setting or when fed a certain prompt as opposed to another (among the two prompts used in [47] and the replication here).

4.2. Assessing the Effect of Hyperparameters

4.2.1. Methods

Although the above analysis found no effect of prompt on model performance, this could be due to a lack of variety of prompts used by [47] in their original analysis.

To get a better sense of whether prompt engineering might have any effect on model performance, a follow-up experiment compared GPT-4’s performance in two novel conditions not tested in the original [47] paper.

In Condition 1 (“minimally tailored” condition), GPT-4 was tested using minimal prompting compared to [47], both in terms of formatting and substance. In particular, the message prompt in [47] and the above replication followed OpenAI’s Best practices for prompt engineering with the API [66] through the use of (a) helpful markers (e.g. ‘ “‘ ’) to separate instruction and context; (b) details regarding the desired output (i.e. specifying that the response should include ranked choices, as well as [in some cases] proper authority and citation; (c) an explicit template for the desired output (providing an example of the format in which GPT-4 should provide their response); and (d) perhaps most crucially, context regarding the type of question GPT-4 was answering (e.g. “please respond as if you are taking the bar exam”).

In contrast, in the minimally tailored prompting condition, the message prompt for a given question simply stated “Please answer the following question,” followed by the question and answer choices (a technique sometimes referred to as “basic prompting”: 67). No additional context or formatting cues were provided.

In Condition 2 (“maximally tailored” condition), GPT-4 was tested using the highest performing parameter combinations as revealed in the replication section above, with one addition, namely that: the system prompt, similar to the approaches used in [67, 68], was edited from its default (“you are a helpful assistant”) to a more tailored message that included included multiple example MBE questions with sample answer and explanations structured in the desired format (a technique sometimes referred to as “few-shot prompting”: [67]).

As in the replication section, 5 trials were conducted for each of the two conditions. Based on the lack of effect of temperature in the replication study, temperature was not a manipulated variable. Instead, both conditions featured the same temperature setting (.5).To assess whether there was a significant difference between GPT-4’s accuracy in the maximally tailored vs minimally tailored conditions, a mixed-effects binary logistic regression was conducted with: (a) condition as a fixed effect; and (b) question number and question category as random effects. As above, these regressions were conducted using the lme4 [64] and lmertest [65] packages from R.

4.2.2. Results

Mean MBE accuracy across all trials in the maximally tailored condition was descriptively higher at 79.5% (95% CI: 77.1 to 82.1), than in the minimally tailored condition at 70.9% (95% CI: 68.1 to 73.7).

The regression model revealed a main effect of condition on accuracy (β=1.395, SE=.192, p<.0001), such that GPT-4’s accuracy in the maximally tailored condition was significantly higher than its accuracy in the minimally tailored condition.

In terms of scaled score, GPT-4’s MBE score in the minimally tailored condition would be approximately 150, which would place it: (a) in the 70th percentile among July test takers; (b) 64th percentile among first-timers; and (c) 48th percentile among those who passed.

GPT-4’s score in the maximally tailored condition would be approximately 164—6 points higher than that reported by [47] and [1]). This would place it: (a) in the 95th percentile among July test takers; (b) 87th percentile among first-timers; and (c) 82th percentile among those who passed.

4.3. Re-examining the Essay Scores

As confirmed in the above subsection, the scaled MBE score (not percentile) reported by OpenAI was accurately computed using the methods documented in [47].

With regard to the essays (MPT + MEE), however, the method described by the authors significantly deviates in at least three aspects from the official method used by UBE states, to the point where one may not be confident that the essay scores reported by the authors reflect GPT models’ “true” essay scores (i.e., the score that essay examiners would have assigned to GPT had they been blindly scored using official grading protocol).

The first aspect relates to the (lack of) use of a formal rubric. For example, unlike NCBE protocol, which provides graders with (a) (in the case of the MEE) detailed “grading guidelines” for how to assign grades to essays and distinguish answers for a given MEE; and (b) (for both MEE and MPT) a specific “drafters’ point sheet” for each essay that includes detailed guidance from the drafting committee with a discussion of the issues raised and the intended analysis [69],

[47] do not report using an official or unofficial rubric of any kind, and instead simply describe comparing GPT-4’s answers to representative “good” answers from the state of Maryland.

Utilizing these answers as the basis for grading GPT-4’s answers in lieu of a formal rubric would seem to be particularly problematic considering it is unclear even what score these representative “good” answers received. As clarified by the Maryland bar examiners: “The Representative Good Answers are not ‘average’ passing answers nor are they necessarily ‘perfect’ answers. Instead, they are responses which, in the Board’s view, illustrate successful answers written by applicants who passed the UBE in Maryland for this session” [70].

Given that (a) it is unclear what score these representative good answers received; and (b) these answers appear to be the basis for determining the score that GPT-4’s essays received, it would seem to follow that (c) it is likewise unclear what score GPT-4’s answers should receive. Consequently, it would likewise follow that any reported scaled score or percentile would seem to be insufficiently justified so as to serve as a basis for a conclusive statement regarding GPT-4’s relative performance on essays as compared to humans (e.g. a reported percentile).

The second aspect relates to the lack of NCBE training of the graders of the essays. Official NCBE essay grading protocol mandates the use of trained bar exam graders, who in addition to using a specific rubric for each question undergo a standardized training process prior to grading [71, 72]. In contrast, the graders in [47] (a subset of the authors who were trained lawyers) do not report expertise or training in bar exam grading. Thus, although the graders of the essays were no doubt experts in legal reasoning more broadly, it seems unlikely that they would have been sufficiently ingrained in the specific grading protocols of the MEE + MPT to have been able to reliably infer or apply the specific grading rubric when assigning the raw scores to GPT-4.

The third aspect relates to both blinding and what bar examiners refer to as “calibration,” as UBE jurisdictions use an extensive procedure to ensure that graders are grading essays in a consistent manner (both with regard to other essays and in comparison to other graders) [71, 72]. In particular, all graders of a particular jurisdiction first blindly grade a set of 30 “calibration” essays of variable quality (first rank order, then absolute scores) and make sure that consistent scores are being assigned by different graders, and that the same score (e.g. 5 of 6) is being assigned to exams of similar quality [72]. Unlike this approach, as well as efforts to assess GPT models’ law school performance [73], the method reported by [47] did not initially involve blinding. The method in [47] did involve a form of inter-grader calibration, as the authors gave “blinded samples” to independent lawyers to grade the exams, with the assigned scores “match[ing] or exceed[ing]” those assigned by the authors. Given the lack of reporting to the contrary, however, the method used by the graders would presumably be plagued by issue issues as highlighted above (no rubric, no formal training with bar exam grading, no formal intra-grader calibration).

Given the above issues, as well as the fact that, as alluded in the introduction, GPT-4’s performance boost over GPT-3 on other essay-based exams was far lower than that on the bar exam, it seems warranted not only to infer that GPT-4’s relative performance (in terms of percentile among human test-takers) was lower than that reported by OpenAI, but also that GPT-4’s reported scaled score on the essay may have deviated to some degree from GPT-4’s “true” essay (which, if true, would imply that GPT-4’s “true” percentile on the bar exam may be even lower than that estimated in previous sections).

Indeed, [47] to some degree acknowledge all of these limitations in their paper, writing: “While we recognize there is inherent variability in any qualitative assessment, our reliance on the state bars’ representative “good” answers and the multiple reviewers reduces the likelihood that our assessment is incorrect enough to alter the ultimate conclusion of passage in this paper.”

Given that GPT-4’s reported score of 298 is 28 points higher than the passing threshold (270) in the majority of UBE jurisdictions, it is true that the essay scores would have to have been wildly inaccurate in order to undermine the general conclusion of [47] (i.e., that GPT-4 “passed the [uniform] bar exam”). However, even supposing that GPT-4’s “true” percentile on the essay portion was just a few points lower than that reported by OpenAI, this would further call into question OpenAI’s claims regarding the relative performance of GPT-4 on the UBE relative to human test-takers. For example, supposing that GPT-4 scored 9 points lower essays would drop its estimated relative performance to (a) 31st percentile compared to July test-takers; (b) 24th percentile relative to first-time test takers; and (c) less than 5th percentile compared to licensed attorneys.

5. Discussion

This paper first investigated the issue of OpenAI’s claim of GPT-4’s 90th percentile UBE performance, resulting in four main findings. The first finding is that although GPT-4’s UBE score approaches the 90th percentile when examining approximate conversions from February administrations of the Illinois Bar Exam, these estimates are heavily skewed towards low scorers, as the majority of test- takers in February failed the July administration and tend to score much lower than the general test-taking population. The second finding is that using July data from the same source would result in an estimate of ~68th percentile, including below average performance on the essay portion. The third finding is that comparing GPT-4’s performance against first-time test takers would result in an estimate of ~62nd percentile, including ~42nd percentile on the essay portion. The fourth main finding is that when examining only those who passed the exam, GPT-4’s performance is estimated to drop to ~48th percentile overall, and ~15th percentile on essays.

In addition to these four main findings, the paper also investigated the validity of GPT-4’s reported UBE score of 298. Although the paper successfully replicated the MBE score of 158, the paper also highlighted several methodological issues in the grading of the MPT + MEE components of the exam, which call into question the validity of the essay score (140).

Finally, the paper also investigated the effect of adjusting temperature settings and prompting techniques on GPT-4’s MBE performance, finding no significant effect of adjusting temperature settings on performance, and some effect of prompt engineering when compared to a basic prompting baseline condition.

Of course, assessing the capabilities of an AI system as compared to those of a practicing lawyer is no easy task. Scholars have identified several theoretical and practical difficulties in creating accurate measurement scales to assess AI capabilities and have pointed out various issues with some of the current scales [10–12]. Relatedly, some have pointed out that simply observing that GPT-4 under- or over-performs at a task in some setting is not necessarily reliable evidence that it (or some other LLM) is capable or incapable of performing that task in general [13–15].

In the context of legal profession specifically, there are various reasons to doubt the usefulness of UBE percentile as a proxy for lawyerly competence (both for humans and AI systems), given that, for example: (a) the content on the UBE is very general and does not pertain to the legal doctrine of any jurisdiction in the United States [16], and thus knowledge (or ignorance) of that content does not necessarily translate to knowledge (or ignorance) of relevant legal doctrine for a practicing lawyer of any jurisdiction; (b) the tasks involved on the bar exam, particularly multiple-choice questions, do not reflect the tasks of practicing lawyers, and thus mastery (or lack of mastery) of those tasks does not necessarily reflect mastery (or lack of mastery) of the tasks of practicing lawyers; and (c) given the lack of direct professional incentive to obtain higher than a passing score (typically no higher than 270) [62], obtaining a particularly high score or percentile past this threshold is less meaningful than for other exams (e.g. LSAT), where higher scores are taken into account for admission into select institutions [74].

Setting these objections aside, however, to the extent that one believes the UBE to be a valid proxy for lawyerly competence, these results suggest GPT-4 to be substantially less lawyerly competent than previously assumed, as GPT-4’s score against likely attorneys (i.e. those who actually passed the bar) is ~48th percentile. Moreover, when just looking at the essays, which more closely resemble the tasks of practicing lawyers and thus more plausibly reflect lawyerly competence, GPT-4’s performance falls in the bottom ~15th percentile. These findings align with recent research work finding that GPT-4 performed below-average on law school exams [75].

The lack of precision and transparency in OpenAI’s reporting of GPT-4’s UBE performance has implications for both the current state of the legal profession and the future of AI safety. On the legal side, there appear to be at least two sets of implications. On the one hand, to the extent that lawyers put stock in the bar exam as a proxy for general legal competence, the results might give practicing lawyers at least a mild temporary sense of relief regarding the security of the profession, given that the majority of lawyers perform better than GPT on the component of the exam (essay-writing) that seems to best reflect their day-to-day activities (and by extension, the tasks that would likely need to be automated in order to supplant lawyers in their day-to-day professional capacity).

On the other hand, the fact that GPT-4’s reported “90th percentile” capa- bilities were so widely publicized might pose some concerns that lawyers and non-lawyers may use GPT-4 for complex legal tasks for which it is incapable of adequately performing, plausibly increasing the rate of (a) misapplication of the law by judges; (b) professional malpractice by lawyers; and (c) ineffective pro se representation and/or unauthorized practice of law by non-lawyers. From a legal education standpoint, law students who overestimate GPT-4’s UBE capabilities might also develop an unwarranted sense of apathy towards developing critical legal-analytical skills, particularly if under the impression that GPT-4’s level of mastery of those skills already surpasses that to which a typical law student could be expected to reach.

On the AI front, these findings raise concerns both for the transparency[ref 20] of capabilities research and the safety of AI development more generally. In particular, to the extent that one considers transparency to be an important prerequisite for safety [38], these findings underscore the importance of implementing rigorous transparency measures so as to reliably identify potential warning signs of transformative progress in artificial intelligence as opposed to creating a false sense of alarm or security [76]. Implementing such measures could help ensure that AI development, as stated in OpenAI’s charter, is a “value-aligned, safety-conscious project” as opposed to becoming “a competitive race without time for adequate safety precautions” [77].

Of course, the present study does not discount the progress that AI has made in the context of legally relevant tasks; after all, the improvement in UBE performance from GPT-3.5 to GPT-4 as estimated in this study remains impressive (arguably equally or even more so given that GPT-3.5’s performance is also estimated to be significantly lower than previously assumed), even if not as flashy as the 10th-90th percentile boost of OpenAI’s official estimation. Nor does the present study discount the seemingly inevitable future improvement of AI systems to levels far beyond their present capabilities, or, as phrased in GPT-4 Passes the Bar Exam, that the present capabilities “highlight the floor, not the ceiling, of future application” [47, 11].

To the contrary, given the inevitable rapid growth of AI systems, the results of the present study underscore the importance of implementing rigorous and transparent evaluation measures to ensure that both the general public and relevant decision-makers are made appropriately aware of the system’s capabilities, and to prevent these systems from being used in an unintentionally harmful or catastrophic manner. The results also indicate that law schools and the legal profession should prioritize instruction in areas such as law and technology and law and AI, which, despite their importance, are currently not viewed as descriptively or normatively central to the legal academy [78].

Algorithmic black swans

Three lines of defense against risks from AI

Value alignment for advanced artificial judicial intelligence