Anthropic has quietly circulated a draft federal AI safety bill among U.S. lawmakers, seeking to establish the first congressional framework governing frontier model deployment3,4.
Representatives for Anthropic pitched the draft legislation to Capitol Hill offices, according to Politico. The effort aims to fill a gap in federal legislation concerning risks associated with advanced AI, with Congress yet to pass any comprehensive AI law.
The bill would empower federal authorities to block the release or use of an AI model, but only under narrow conditions: an imminent risk of death or serious injury to more than 50 people, or losses exceeding $1 billion. Any such action would require a judge's approval.
An Anthropic representative told AFP that the initiative would give the U.S. government the ability to "block or restrain dangerous deployments of the most powerful models". The draft also provides for developer-run testing of frontier models, public transparency regarding test results, independent evaluation, and maintenance of a robust safety program.
Politico's framing noted that the bill would limit the circumstances in which the government could block a model's release. ANALYSIS The dual emphasis, granting the government a blocking power while constraining when that power applies, positions the draft as a middle path between unregulated deployment and broad premarket licensing.
U.S. AI policy under President Donald Trump has largely relied on voluntary industry commitments and executive orders. Congress has not passed a federal AI law and will be in recess until the conclusion of next month's midterm elections.
ANALYSIS Anthropic's decision to draft and circulate its own legislative text, rather than wait for a congressional committee to originate one, places the company in the unusual position of a regulated entity authoring the rules it would operate under. The narrow trigger thresholds, 50 or more casualties or $1 billion in losses, and the judicial-approval requirement would set a high bar for government intervention, effectively shielding routine commercial deployments from regulatory action while preserving a safety valve for catastrophic scenarios.