Anthropic pitches an AI safety bill that kicks in only at $1B in damage or 50 deaths

Anthropic is pitching lawmakers a draft AI safety bill that would let the government block a model only after it has caused more than $1 billion in damage or 50 deaths, according to a report by Pasquale Pillitteri.

The proposal turns an abstract argument into two numbers. Instead of asking what a model can do, it asks what a model has already done, and it sets the bar for intervention extraordinarily high.

That is notable coming from Anthropic, which has built its public identity around safety and controllability. A bill written this way concedes that frontier AI oversight may work better as after-the-fact accountability than as pre-deployment prevention.

For U.S. legislators, the value is that it creates something concrete to debate. Federal work on frontier AI has stalled repeatedly over who judges risk and how much risk is too much; a damage figure and a death toll give the argument an anchor.

The threshold is also easy to attack. Supporters can argue that only catastrophic outcomes justify the strongest regulatory power, and that a high bar keeps compliance costs off smaller developers. Critics will point out that $1 billion and 50 lives are themselves an enormous price to pay before anything happens.

The harder questions are technical. How is damage counted, how is a death attributed to a model, how much of the causal chain does the developer own, and who is liable when several models act together? Without answers, even a precise number is hard to enforce.

What to watch next is whether lawmakers adopt the text and whether other frontier labs endorse it or distance themselves from it. Anthropic has put a number on the table; the argument now moves to drafting language.