$1B damage or 50 dead before Washington can act

Anthropic's draft frontier AI bill drew immediate backlash from safety advocates over its threshold: the government could block a model only when a foreseeable, imminent risk exceeds $1 billion in damage or more than 50 people killed or seriously injured. Critics argued the bar is set so high that harms below it — and harms already occurring today — fall outside the government's power to act, and that a lab writing its own thresholds is self-serving.

The draft's preemption clause drew the sharpest objections. It would override state rules on safety disclosures, third-party evaluation and incident reporting, including California's SB 53, which Anthropic had previously endorsed; Senator Scott Wiener, SB 53's author, said the proposal falls far short of protecting public safety. Other critics noted the draft sets no minimum standards for evaluators, who are paid by the labs they assess, and Anthropic said within days it would revise the state-law section.