Anthropic’s AI shutdown proposal sets a 50-person threshold

A high bar for intervention

The proposal would apply only if a model could lead to one of three threats:

An independent monitor overseen by the Department of Commerce would assess the risk. If the monitor found that a future model could cause one of those threats, it would have 72 hours to alert department officials. They would then decide whether to present the evidence to a federal judge, who would make the final call on whether the model could be released.

Politico reports that three people familiar with the talks described Anthropic’s role in shaping the bill. The company told the outlet the proposal reflects its document, “Frontier AI Framework,” and said it has advocated for the measures at the federal and state levels since June. Anthropic considers them reasonable and believes Congress should pass the bill.

The negotiations come as AI companies spend heavily on politics ahead of US midterm elections. Known AI companies have spent at least $55.7 million, while executives at leading labs, including OpenAI and Anthropic, have directed tens of millions of dollars to political groups. Their specific positions on regulation had remained unclear.

What the threshold leaves out

I think the most consequential feature is not the power to halt a model, but how tightly the proposal defines when that power can be used. The 50-person and $1 billion thresholds, combined with three specified threats and a judge’s decision, make this a constrained intervention mechanism—not broad government authority over AI releases.

The announcement is quiet about how an independent monitor would assess the risks, or what evidence would be enough to persuade a judge. Those details could determine whether the process is a meaningful safeguard or a high bar that rarely changes what labs can ship. Anthropic’s proposal offers a clear boundary for intervention; its practical force depends on how that boundary is applied.