The Case for Supervising Frontier AI Developers

‍
GovAI research blog posts represent the views of their authors rather than the views of the organisation.

‍
Introduction 

Improvements in the capabilities of frontier AI systems present policymakers with a classic challenge: how to preserve the benefits of innovation while mitigating attendant risks. Certain jurisdictions are beginning to regulate or consider regulating frontier AI development. Such regulations must avoid two perils: unduly stifling innovation and failing to genuinely reduce risks. This raises the question: "What approach to regulation is likely to most reduce risks without creating unnecessary burdens?". In a new article, I argue that government supervision is an especially promising approach to regulation.

Government supervision is an approach to regulation in which public servants oversee certain entities and can impose consequences if those entities breach legal requirements. At its core, supervision works by combining two elements: robust information-gathering powers for public servants, and discretionary authority to act on that information.

This combination has proven particularly suitable for industries where regulators face uncertainty, and therefore need flexibility and the opportunity to build expertise. This makes supervision particularly suitable for complex, quickly-evolving technological fields – like frontier AI.

Supervision offers the hope of providing effective oversight of the AI industry without stifling innovation. The government can build expertise and set up early warning systems, but hold off on imposing rigid rules for now. Oversight could then be quickly scaled up later if serious risks or evidence of wrongdoing emerges.

Of course, the supervisory approach isn’t perfect. There are risks that supervising regulators may become captured by industry interests, particularly given the stark differentials in technical expertise and compensation; that the scope of supervision will expand over time, potentially growing the regulatory burden and diluting focus on the most serious risks; that information sharing will create security vulnerabilities; and that supervisory discretion will be used arbitrarily or capriciously. 

These challenges demand careful attention, none is strong enough to undermine the promise of supervision as a tested regulatory approach that can scale with technological advancement.

In the rest of this post, I summarize my argument in favour of the supervisors model. For more detail, see the full paper.

‍

What is supervision?

Supervision combines two key elements: information gathering and discretionary authority. These elements emerged from supervision’s roots in overseeing banks. Bank supervisors don't just enforce a static set of rules. They have broad authority to:

• Review internal documents and risk assessments
• Interview employees at all levels
• Evaluate new products or practices before they go to market
• Require changes when they spot concerning patterns

This dynamic approach allows regulators to build deep expertise over time while responding flexibly to new developments. Rather than trying to anticipate every possible risk in advance, supervisors can adjust their oversight based on what they learn.

Information gathering means supervisors can access non-public information from companies outside the normal investigatory process. They can review internal documents, observe meetings, interview employees at all levels, and request additional data when needed. This goes beyond occasional audits or compliance checks and helps regulators build deep expertise over time. It also helps supervisors spot patterns and potential risks that might not be visible from public information alone.

Discretionary authority means supervisors can exercise judgment about when and how to intervene based on what they learn. Rather than just enforcing predetermined rules, they might offer informal guidance, require changes to business practices, delay product launches, or escalate to more serious enforcement actions. This flexibility is crucial for emerging technologies where risks may not be fully understood in advance and where overly rigid rules could stifle beneficial innovation.

When properly structured, supervisory discretion serves several crucial functions, including:

• Allowing oversight to evolve alongside technology. Rather than waiting for new regulations to be written, supervisors can adjust their approach as they learn more about emerging risks and opportunities.
• Enabling proportional responses. Not every concern requires heavy-handed intervention. Supervisors can start with informal guidance and escalate only if needed.
• Creating incentives for cooperation. Companies know that being forthcoming with information and responsive to concerns can lead to lighter-touch oversight.

Together, information gathering and discretionary authority allow regulatory oversight to evolve with the technology it governs.

Supervision isn't without drawbacks, including:

• The extensive discretion given to supervisors can create tension with rule-of-law principles, and lead to arbitrary or biased decisions. 
• Supervisors can become too deferential to the industry they oversee - a weak form of regulatory capture. This risk is particularly acute when regulated entities can offer substantially higher salaries than the public sector. 
• Supervision can make industry uncertain. The flip side of regulatory flexibility is that it may not always be obvious how to comply.
• Supervisors could use their powers overly vigorously and so impede beneficial innovations
• Supervisors’ access to company’s information may make supervisors a cybersecurity target.

‍

Avoiding these potential pitfalls demands care in designing the supervisory regime, as well as in who is hired to act as supervisors. How well supervision goes depends both on institutional design and human capital.

‍

‍

Supervision and AI

Supervision may be particularly appropriate for frontier AI regulation because the field's rapid evolution makes traditional rigid regulations likely to become quickly outdated or irrelevant. Supervision allows oversight to evolve alongside the technology. 

The potential risks from AI development – from misuse of powerful systems to unintended consequences of deployment – require early warning systems and rapid response capabilities that supervision can provide. At the same time, many of these risks – especially the most serious ones – are currently speculative. Supervision allows governments to monitor these risks and act if needed without prematurely constraining technological development and its economic benefits.

Supervision is well-established as a modality of regulation when private entities have more expertise than government, including in aviation safety, nuclear power, food and drug safety, and environmental protection. As in AI development, these domains each involve complex technical systems, the potential for serious harm from failures, ongoing technological change, and significant information asymmetries between industry and regulators.

‍

For frontier AI developers, supervision would rest on a legislative prohibition on developers performing certain activities – such as training a frontier model using more than a threshold amount of computing power – unless they submit to supervision. Supervisors would then gather information from these developers, first through regular reports on developers’ activities, safety measures, and testing results.  Supervisors could request additional information about specific systems or practices that raise concerns, observe internal meetings, and review documentation about model capabilities and limitations. 

‍

With the information supervisors gather, supervisors can act to reduce risks. Initially, they could help spread safety-related best practices between different developers and thereby raise the safety floor. If, later, supervisors identified serious risks, they could delay deployments, require modifications to systems, or even revoke licenses in extreme cases. This oversight would be particularly valuable for monitoring societal-scale risks that might be difficult to address through traditional liability frameworks, such as potential misuse for weapons development. Supervisors could also facilitate appropriate sequencing of AI deployment, ensuring that safety-critical institutions have time to adapt to new capabilities before they become widely available.

‍

The information-gathering capacity can also enable supervisors to detect and oversee non-public deployments of technologies. Supervisors will be able to find out about capability increases that are known inside frontier developers but have not yet been publicised and to monitor how developers are using AI tools themselves. The combination of insight and power changes the deployment decision for risky frontier AI: rather than be a question for a company according to its own wisdom (or a Safety Commitment), supervisors put the public interest in-the-loop.

‍

‍

Implementation Considerations

Making supervision work for AI requires careful attention to several factors. 

First, supervisory agencies need sufficient technical expertise to engage meaningfully with AI developers. This means competing for talent with industry. While governments often are unwilling to compete with industry on raw salaries, government agencies can offer talented individuals the chance to serve the public interest and have an impact on the world by wielding decision-making authority.

Second, the scope of supervision must be clearly defined. A natural starting point would be companies developing the most capable AI systems, as measured by compute usage or other technical metrics. The supervisory mandate should focus on the most serious risks while avoiding mission creep into areas better handled by other regulators.

Third, information security will be crucial, as supervisors likely need access to sensitive technical details. Robust systems must be in place to prevent both accidental leaks and targeted attacks seeking to exploit supervisors' access to company information.

The path forward isn't simple, but supervision offers a tested regulatory model that can scale with AI advancement – responding proportionally to societal risks while preserving the benefits of innovation.

‍

‍