OpenAI on US AI Safety Governance and Reverse Federalism

US AI Safety Governance through Reverse Federalism

OpenAI is advocating for a national AI safety framework grounded in "reverse federalism," where aligned legislation across multiple states creates a de facto national standard. This approach aims to ensure that critical frontier safety decisions are made by democratic governments rather than solely by AI labs, establishing a foundation for a US-led global AI framework.

State-Level Alignment and the De Facto National Standard

In the absence of a formal federal framework, OpenAI argues that states can establish a shared direction by passing mirrored laws. This prevents a "patchwork of regulations" that could hinder innovation or divert resources from startups and small companies toward regulatory compliance rather than safety.

According to OpenAI, California, New York, and Illinois have already advanced frontier safety legislation that aligns on three core elements:

• Documented Safety Frameworks: Requirement for risk assessments for frontier models and the public disclosure of those assessments and results.
• Incident Reporting: Mandatory reporting of serious safety incidents.
• Governance and Accountability: The use of independent, objective audits to ensure compliance.

OpenAI notes that California established the core disclosure framework, New York demonstrated the approach's cross-jurisdictional viability, and Illinois added requirements for independent verification of disclosures.

Federal Integration and National Security

OpenAI asserts that a federal framework is essential because frontier AI involves national security risks and technical reviews that exceed the capacity of individual states. The organization argues that states should not be tasked with managing national security decisions or conducting highly technical reviews that require access to classified systems.

Federal Testing and the Role of CAISI

OpenAI proposes that the federal government lead the testing and evaluation of the most advanced systems. Specifically, they suggest strengthening the Center for AI Standards and Innovation (CAISI) to provide the durable federal capacity needed to prevent harm before it occurs, rather than relying on post-incident accountability.

Cyber Evaluations and Government Access

The Trump Administration is currently working with experts to develop a framework for US government testing of capable AI models specifically regarding cybersecurity. OpenAI expects this framework to be in place by early August. A consistent, repeatable testing approach is deemed necessary to get advanced AI tools into the hands of government, critical infrastructure defenders, and allies quickly enough to counter malicious actors.

Legislative Progress and the Frontier Safety Blueprint

OpenAI views current proposals from members of Congress, including Representatives Jay Obernolte and Lori Trahan, as productive steps toward a federal framework. To guide this legislation, OpenAI's "frontier safety blueprint" outlines three primary requirements:

1. Federal Leadership in Testing: The federal government must lead evaluation of advanced systems due to the most significant national security and public safety risks.
2. Company Requirements: Developers of the most capable systems must adhere to independent audits, incident reporting, strong security standards, and whistleblower protections.
3. Mutual Reinforcement: Federal and state efforts should complement each other, with states continuing to act as "laboratories of democracy" for non-frontier safety issues such as AI education, youth protection, and environmental policy.

Moving Toward a Global AI Framework

OpenAI argues that a cohesive national standard is a prerequisite for leading an international framework for AI standards. This vision includes a US-led international forum to establish accepted standards, provide impartial analysis of risks and capabilities, and make technology available to nations and companies that follow the rules. This global approach is intended to be grounded in democratic values and a democratic vision for AI.