Anthropic's AI safety bill: 50 deaths or $1B threshold

Anthropic has proposed an AI safety bill with a notably high activation threshold: it applies only when 50 deaths or $1 billion in damages are at stake. For anyone following AI governance, that detail matters as much as the headline.

The proposed process runs through reporting. Evaluators would flag any imminent risk to the Department of Commerce within 72 hours. The Department would then bring the evidence before a federal judge and await an order. Under this structure, Commerce would effectively be the only agency able to stop an AI model.

This raises questions that go beyond one company's proposal. How do you set a threshold that catches real danger without slowing legitimate development? How much speed does a 72-hour reporting window give evaluators, and how much judicial process is enough? For professionals building, auditing, or regulating AI systems, these design choices will shape compliance work, risk assessment, and how teams document model behavior.

Where do you think the line should sit between early intervention and due process?