Technology and Scienceupdated 51m ago3 min read
Anthropic Proposes AI Safety Bill Letting U.S. Agencies Block Models Over $1 Billion Damages Or 50 Deaths
This article is an automated aggregation of coverage from news outlets. Our editorials offer in-depth analysis curated by the NewsCord team.
Anthropic proposes federal AI safety bill to block models after 50 deaths or $1B. Would require court approval to suspend a model's deployment.
Judge Gate for Blocks
Anthropic’s proposed AI safety bill would let federal agencies block an AI model only when a “foreseeable and imminent risk” meets a catastrophic threshold of “more than $1 billion in damages” or “death or serious injury to more than 50 people.”
The bill would route enforcement through the U.S. Department of Commerce, which would then seek a federal judge’s order in the District of Columbia rather than stopping a model on its own.
The draft also targets specific threat categories, including “biological weapons” and “cyberattacks,” and it would cover situations where a model “evading its developer’s controls.”
Anthropic’s spokesperson told Politico that the initiative builds on the company’s “Advanced AI Framework,” which Anthropic has been pitching to lawmakers since June.
The draft’s structure would also require periodic evaluations by “independent organizations registered with the Department of Commerce,” with evaluators reporting imminent risks within 72 hours after an evaluation ends.
Evaluators, Payments, and Timing
The bill would use “independent evaluators” registered with the U.S. Department of Commerce to assess catastrophic scenarios, and it would allow AI labs to pay them as long as the “payment does not depend on the evaluation results.”
Critics raised concerns that the draft “does not set standards or licenses for those who audit the models,” and Charlie Bullock of LawAI warned that the absence of rules could let auditors compete by offering “the least burdensome and cheapest reviews.”
The draft would require an evaluator who finds an “imminent catastrophic risk” to notify the Department of Commerce within 72 hours, and the Department of Commerce would then have to seek a judge’s order to halt the model’s release.
The proposal would also “expressly prohibit any federal agency from stopping the distribution of the model on its own,” concentrating the stop mechanism in the judge-led process.
Anthropic’s spokesperson confirmed the draft to Politico, saying, “This bill represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been promoting to federal and state lawmakers since June,” according to one account.
Preemption and the Debate
The draft would limit state-level intervention by addressing risk disclosure, independent evaluations, and reporting of critical incidents, and it would “limit state interventions to the most severe scenarios.” One account said the bill would “exclude certain state rules on risk disclosure, third-party evaluation, and incident reporting,” while another account described the text as aiming to “limit the ability of American states to impose their own obligations on AI developers” in those same areas. The bill would also leave open how risks below the defined thresholds would be handled, because the framework would leave “pending the handling of risks that do not meet the defined thresholds but could multiply with the diffusion of models.”
The draft’s enforcement design would require the Commerce Department to present evidence to a judge, and it would place the burden of proof on the Secretary of Commerce, with the blocking order expiring after 45 days. As the draft stage continues, Anthropic’s proposal remains framed as a way to set safety obligations while restricting direct government interventions, with one account noting that the text “has not yet been officially filed with Congress.”