Independent evaluators could refer high-risk models to Commerce officials, who would seek a judge’s halt.
Anthropic is circulating a proposed AI-safety bill that could block an AI model’s release when evaluators identify a foreseeable, imminent catastrophic risk, TechCrunch reports. It remains a policy proposal under discussion, not an enacted rule or operating release-control system.
The reported triggers include a risk that could kill more than 50 people or cause over $1 billion in property damage. The danger would also need to involve creating or releasing a biological weapon, a model evading its developer’s controls, or a cyberattack, according to TechCrunch’s account of Politico’s reporting.
The process would begin with independent evaluators registered with the U.S. Commerce Department. If they determined that an imminent catastrophic risk existed, they would have 72 hours to contact the department.
Commerce officials would then ask a Washington, D.C., judge to halt the model’s release. Federal agencies could not independently stop distribution under the reported proposal; the court request would be the required route.
Anthropic’s plan reportedly would allow AI companies to pay evaluators, as long as payment did not depend on the evaluation’s outcome. That creates a practical oversight question: how would the system prevent financial ties from influencing judgments?
The reporting does not specify how evaluators would calculate the death or property-damage thresholds, or what evidence a judge would require. Anthropic confirmed the draft to Politico and said it reflects ideas in the company’s Advanced AI Framework.
The proposal remains under discussion, and the available reporting does not show that it has become law. Its next test is whether lawmakers can define the thresholds and evaluator safeguards clearly enough for the process to work.