Anthropic is circulating proposed federal AI safety legislation that would establish independent oversight of advanced models while tightly limiting the circumstances in which the government could block their release or use. The draft would require a foreseeable, imminent threat of more than $1 billion in damage or death or serious injury to more than 50 people before officials could seek a court-ordered halt.
Politico first reported the proposal, citing three people familiar with Anthropic’s discussions with lawmakers and obtaining the draft. The legislation would create a national framework for frontier AI developers, combining safety assessments and reporting duties with restrictions on federal intervention. It remains a proposal, not an enacted law.
The threshold concerns the risk of future harm, rather than requiring deaths, injuries or financial losses to have already occurred. Qualifying threats would be limited to biological weapons, cyberattacks and AI systems operating beyond their developers’ control. That would reserve the proposed blocking mechanism for a narrow set of potentially catastrophic dangers, rather than every type of harm associated with artificial intelligence.
Under the draft, independent evaluators registered with the Commerce Department would periodically examine advanced AI developers and their models. Developers could pay for those reviews, but evaluators’ compensation could not depend on their conclusions. An evaluator finding an imminent danger that met the legislation’s criteria would have to notify Commerce within 72 hours of finishing the assessment.
Commerce could then ask a federal judge in the District of Columbia to halt the model’s release or operation. The department would not have unilateral authority to suspend a model through that process. Politico’s reporting also describes provisions barring other federal departments, agencies and officials from independently stopping model development, deployment or internal use under the framework.
The distinction between oversight and a shutdown is central to the proposal. Companies would face ongoing obligations even when their systems did not meet the threshold for a court-ordered suspension. Those obligations would include maintaining safety frameworks, publishing transparency reports for new frontier models, protecting whistleblowers and disclosing critical incidents to Commerce.
The draft would also supersede certain state requirements covering risk disclosures, independent evaluations and incident reporting. A federal framework could reduce overlapping compliance duties, but replacing state rules would also change which authorities set and enforce those safeguards. The proposed limits on intervention therefore sit alongside a separate question about how responsibility should be divided between Washington and the states.
Frontier AI oversight generally focuses on the most capable systems and whether their capabilities could enable serious misuse or undermine developers’ ability to control them. Independent evaluations are intended to test those risks separately from a developer’s own assessments. The proposed payment restrictions would address one aspect of that relationship while still allowing developers to finance the reviews.
Anthropic, which develops the Claude assistant, says the draft reflects a safety framework it has promoted since June. Its proposal joins competing congressional efforts to regulate advanced AI, with the scope of federal authority and the treatment of state safety laws still unresolved. The assessment requirements, harm thresholds and judicial enforcement process would need congressional approval before becoming binding federal law.
Sources: PC Gamer, Daily Beirut, PYMNTS.com, TradingView