Framework to Mitigate AI-Enabled Extreme Risks

• Authority: Federal government
• Jurisdiction: United States
• Collections: Editors' Picks;Miscellaneous documents
• Most recent activity: Enacted
• Most recent activity date: 2024-04-16

Risk subdomain coverage

Good coverage — Comprehensive governance measures or mitigations with clear procedures relating to the risk subdomain are explicitly described.

• AI system security vulnerabilities
• Cyberattacks, weapon development or use
• Governance failure
• AI possessing dangerous capabilities

Minimal coverage — Brief mention but no specific mitigations, recommendations or governance measures relating to the risk subdomain are described.

• Disinformation, surveillance, and influence at scale
• Competitive dynamics
• AI pursuing its own goals
• Lack of capability or robustness

All other risk subdomains had no coverage in this document.

Risk Coverage Summary

The document has good coverage of approximately 8-10 subdomains, with strong focus on malicious actors (4.1, 4.2), AI system security (2.2), competitive dynamics (6.4), governance failure (6.5), and AI safety failures (7.1, 7.2, 7.3). Coverage is concentrated in security, misuse prevention, and AI safety domains, specifically addressing extreme risks from frontier AI models.

Sector coverage

Good coverage — Comprehensive governance measures or mitigations with clear procedures relating to the sector are explicitly described.

• Scientific Research & Development Services
• Public Administration (excl. National Security)
• National Security

Minimal coverage — Brief mention but no specific mitigations, recommendations or governance measures relating to the sector are described.

• Information

All other sectors had no coverage in this document.

Sectors Summary

This document does not govern AI use within specific economic sectors. Rather, it establishes a cross-sectoral federal oversight framework for frontier AI model development and deployment to mitigate extreme risks (biological, chemical, cyber, nuclear). The governance applies to AI developers and hardware providers regardless of the sector in which they operate.

Legislative Status

Other

Legislative Status Summary

This is a policy proposal document that outlines a potential framework for federal oversight but does not constitute binding law or formal regulation. It uses conditional language ('would', 'could') throughout and presents options rather than mandates.

Actors

Proposer

• Entity Types: AI_Governance_Actor
• Summary: The document is framed as a proposal to Congress for establishing federal oversight, suggesting it originates from within the federal government or policy advisory context.

Target

• Entity Types: AI_Developer, AI_Infrastructure_Provider
• Summary: The document explicitly targets developers of frontier AI models and hardware providers who sell or rent large-scale computing resources for AI development.

Enforcer

• Entity Types: AI_Governance_Actor
• Summary: The document proposes four potential federal oversight entities that would enforce the framework, though none is definitively selected.

Monitor

• Entity Types: AI_Governance_Actor
• Summary: The document specifies that the oversight entity would monitor compliance and study emerging risks, with expertise from biosecurity, chemical security, cybersecurity, and nuclear security domains.

AI Lifecycle Stages

Lifecycle stages covered by the documents, as defined in the OECD Framework for the Classification of AI Systems

• Build and Use Model
• Verify and Validate
• Deploy
• Operate and Monitor

AI Lifecycle Summary

The document covers multiple AI lifecycle stages with primary focus on Build and Use Model, Deploy, and Operate and Monitor stages. It addresses development notification requirements, pre-deployment licensing and evaluation, and ongoing monitoring of frontier models.

Technical Scope

Terms relating to technical scope that are mentioned or implicitly referenced in the document

• AI Models
• Frontier AI
• Compute Threshold
• General Purpose AI
• Open Weight / Open Source
• Task-specific AI

Technical Scope Summary

The document explicitly focuses on 'frontier models' defined by compute threshold (10^26 operations) and capability criteria. It addresses both general-purpose broadly-capable models and task-specific models intended for bioengineering, chemical engineering, cybersecurity, or nuclear development. No explicit mention of open-weight models, generative AI, or predictive AI.