Frontier AI Regulation: Managing Emerging Risks to Public Safety

Abstract: Advanced AI models hold the promise of tremendous benefits for humanity, but society needs to proactively manage the accompanying risks. In this paper, we focus on what we term "frontier AI" models: highly capable foundation models that could possess dangerous capabilities sufficient to pose severe risks to public safety. Frontier AI models pose a distinct regulatory challenge: dangerous capabilities can arise unexpectedly; it is difficult to robustly prevent a deployed model from being misused; and, it is difficult to stop a model's capabilities from proliferating broadly. To address these challenges, at least three building blocks for the regulation of frontier models are needed: (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. Industry self-regulation is an important first step. However, wider societal discussions and government intervention will be needed to create standards and to ensure compliance with them. We consider several options to this end, including granting enforcement powers to supervisory authorities and licensure regimes for frontier AI models. Finally, we propose an initial set of safety standards. These include conducting pre-deployment risk assessments; external scrutiny of model behavior; using risk assessments to inform deployment decisions; and monitoring and responding to new information about model capabilities and uses post-deployment. We hope this discussion contributes to the broader conversation on how to balance public safety risks and innovation benefits from advances at the frontier of AI development.

• The paper proposes a regulatory framework for frontier AI models addressing unpredictable capabilities and significant public safety risks.
• It emphasizes rigorous pre-deployment testing, continuous monitoring, and dynamic safety standards to tackle deployment challenges.
• The paper recommends a multi-stakeholder compliance process with mandatory disclosures to manage risks from rapid AI proliferation.

Frontier AI Regulation: Managing Emerging Risks to Public Safety

With the increasing capabilities of AI models, there arises a pressing need to manage and mitigate potential risks to public safety and global security. This paper outlines a regulatory framework for "frontier AI" models, which are defined as advanced AI systems that possess capabilities that could potentially be dangerous. These systems necessitate unique regulatory approaches due to the unpredictability and spontaneity of dangerous capabilities, difficulties in ensuring safe deployment, and the potential for rapid proliferation.

The regulatory infrastructure proposed in this paper includes a combination of standard-setting, increased regulatory visibility, and mechanisms to ensure compliance. The goal is to balance the innovation benefits of AI with robust public safety protocols.

Regulatory Challenges with Frontier AI Models

Defining Frontier AI Models

Frontier AI models are characterized as highly capable foundation models that could result in hazardous capabilities, potentially causing significant risk to public safety. Such models can span capabilities involving biochemical weapon design, disinformation propagation, offensive cyber capabilities, and evasion from human control.

Figure 1

Figure 1: Example frontier AI lifecycle.

Key Regulatory Challenges

1. Unexpected Capabilities Problem: Dangerous capabilities can emerge unpredictably and may not become evident until after deployment. This unpredictability requires rigorous pre-deployment testing and continuous post-deployment monitoring.
2. Deployment Safety Problem: Ensuring that deployed AI models consistently operate securely and as intended is complex due to the difficulty in specifying comprehensive behavior controls. This includes preventing adversarial exploitation and addressing dual-use capabilities.
3. Proliferation Problem: Frontier AI models can quickly proliferate, especially if open-sourced or leaked, making broad regulatory accountability challenging. This calls for a framework that considers the entire lifecycle of AI development and deployment.

Figure 2

Figure 2: Certain capabilities seem to emerge suddenly.

Building Blocks for Frontier AI Regulation

Development of Safety Standards

The establishment of dynamic and robust safety standards is crucial. Multi-stakeholder processes involving industry, academia, and civil society should lead this effort, informed by empirical assessment methods to operationalize these standards effectively.

Increasing Regulatory Visibility

Regulatory authorities need comprehensive insights into AI development processes. This can be achieved through mandatory disclosure regimes, audits, and protections for whistleblowers. Ensuring high information security for sensitive disclosures is essential to mitigate risks of adversarial access.

Ensuring Compliance with Standards

Regulatory approaches should scale from voluntary guidelines to mandatory compliance through supervisory authorities or licensing regimes for especially high-risk developments and deployments. This dual strategy ensures current safety without stifling innovation unnecessarily.

Initial Safety Standards for Frontier AI

Risk Assessment and External Scrutiny

Conduct thorough risk assessments for dangerous capabilities and control robustness prior to deployment. Engage third-party experts to independently evaluate and scrutinize models, ensuring comprehensive coverage of potential risks.

Deployment Protocols Based on Risk Assessment

Deploy models following standardized protocols based on assessed risk. These protocols should be regularly reviewed and adaptable in light of new discoveries or enhancements in AI capabilities.

Monitoring and Responding to New Information

Maintain continuous oversight of deployed models, adjusting risk assessments and deployment strategies as new information becomes available. This includes adapting to post-deployment enhancements such as fine-tuning or tool usage expansions.

Figure 3

Figure 3: Computation used to train notable AI systems. Note logarithmic y-axis. Source: Various.

Conclusion

The proposed regulatory framework seeks to address the emergent risks associated with frontier AI models through comprehensive regulation that supports safety while enabling innovation. Regulatory measures, when well-conceived and implemented, can ensure AI advances contribute positively to society while safeguarding public trust and security.

Figure 4

Figure 4: Scaling reliably leading to lower test loss.

For meaningful implementation, international collaboration will be crucial, leveraging collective insights to establish norms and frameworks that preempt potential safety and ethical challenges posed by advanced AI capabilities.