Anthropic is said to propose federal AI safety rules with limits on government

Anthropic (ANTHRO) is circulating a draft AI safety bill among congressional lawmakers that would establish federal oversight of advanced models while narrowly defining when officials could seek to stop their release or use. The proposal, obtained by Politico, would also override certain state requirements, giving the Claude developer a direct role in shaping potential regulation of its industry.

Under the proposal, government intervention to block a model would require a foreseeable, imminent risk of more than $1 billion in damage or death or serious injury to more than 50 people. Qualifying threats would be limited to biological weapons, cyberattacks or a model escaping its developer’s controls. The language resembles provisions in California’s AI safety laws.

For investors, the proposal illustrates how federal legislation could affect AI companies’ compliance costs and ability to bring products to market. A national framework could simplify overlapping state obligations, while restrictions on government intervention could make deployment decisions more predictable. Independent evaluations and mandatory reporting would nevertheless introduce oversight expenses, and the draft’s prospects remain uncertain amid congressional disagreements.

According to the proposal, independent evaluators registered with the Commerce Department would periodically assess frontier developers and their models. Companies could pay those evaluators, provided compensation didn't depend on their findings. An evaluator identifying an imminent risk covered by the legislation would have to notify Commerce within 72 hours of completing the assessment.

Commerce could then ask a federal judge in Washington, DC, to halt a model’s release. The draft would explicitly bar other federal departments, agencies and officials from suspending its development, deployment or internal use. That approach follows export restrictions that, according to the report, shut down two advanced Anthropic models over the summer and heightened industry concerns about abrupt government decisions.

The bill would also require company safety frameworks, transparency reports for new frontier models, whistleblower protections and disclosure of critical incidents to Commerce. Its federal requirements would supersede some state rules governing risk disclosures, outside evaluations and incident reporting—a politically contentious step given Anthropic’s previous support for state safety legislation.

Anthropic (ANTHRO) says the draft reflects a framework it has promoted since June. Its lobbying comes alongside competing House and Senate initiatives, with lawmakers divided over how much authority Washington should have to review or restrict advanced AI. The eventual balance between industry flexibility and government oversight remains unsettled.