Anthropic drafted an AI safety bill with court-enforced release blocks: credit for enforceable tests, scrutiny for who holds the pen

Anthropic has confirmed it drafted an AI safety bill that would let the US Commerce Department ask a judge to block a model release posing an imminent risk of 50 or more deaths or a billion dollars in damage. Whatever one thinks of the thresholds, the numbers in this debate are now coming from the people building the models.

That deserves some credit, because a test a court can enforce is far more useful than a page of voluntary principles. It also deserves scrutiny. A threshold for this proposed power to block a release is not a declaration that smaller harms are acceptable. It still raises the question of who should set the bar for intervention. Proposals written by the regulated party tend to set the bar where the proposer can clear it, and because the full draft has not been published, the detail could look quite different once it is.

The most consequential drafting in any statute is rarely the headline clause. It sits in the definitions, the burden of proof, the meaning of "imminent" and who is allowed to bring the application, and that is where this bill will be won or lost.

The question for Canberra as much as Washington is whether governments are comfortable letting the party that would be liable write the first draft of its own liability test, and if not, who should hold the pen?

Source: TechCrunch, "Anthropic pitches AI safety bill that may block a model if 50+ people could be killed" (20 October 2026), https://lnkd.in/gE47uZ35. Anthropic confirmed the draft to Politico; the full draft text has not been published.

#AIGovernance #Regulation #AI