Anthropic Pushes AI Safety Bill With a 50-Death Threshold

Anthropic, the company behind the Claude models, is pushing an AI safety bill in the US Congress. The bill would let authorities block a model release only in the case of a "foreseeable, imminent risk." That means 50 or more deaths, or damage above a billion dollars. The danger has to come from biological weapons, from a model evading its own restrictions, or from a cyberattack. No single federal agency can make that call alone. Independent experts assess the risk first. They are registered with the Department of Commerce and paid regardless of what they conclude. If they find a catastrophic risk, they have 72 hours to notify the Department of Commerce. After that, the Department of Commerce must go to a federal judge in Washington to approve any halt. The bill sets no deadline for this last step.

For AI labs, this draws a clear line tied to actual harm, not to model size or media noise. No agency can block a release without a judge's prior approval. One point stays open. The tech outlet TechCrunch notes that it is unclear whether payment from the labs could still sway the experts' decisions. Their pay is not tied to the outcome, but it still comes from the very labs being reviewed. If the bill passes as written, companies get a predictable, narrow procedure for a rare catastrophic case. That beats living under an open-ended threat that any new model release could be delayed without explanation.