What happened

Anthropic, the AI company that makes Claude, has pitched a draft federal AI safety bill to US lawmakers. TechCrunch reported the proposal after Politico described a confidential draft. Anthropic later confirmed that the draft reflects its Advanced AI Framework. The company says it has promoted that framework to federal and state lawmakers since June.

The proposal would let the government stop a powerful AI model only in a narrow set of cases. The reported threshold is a foreseeable and imminent risk of more than $1 billion in damage, or of killing or seriously injuring more than 50 people. The danger would also need to involve a biological weapon, an AI model escaping its developer’s controls, or a cyberattack.

How the proposal would work

The draft would use outside evaluators registered with the US Commerce Department. They would periodically examine an AI lab and its models. If an evaluator found one of the listed risks, the Commerce Department could ask a federal judge in Washington, DC, to suspend the model. The judge, not the department alone, would make the final decision.

A related report from Tom’s Hardware says the reported text would prevent other federal departments, agencies, or officers from independently ordering a model’s release or use to be stopped. That makes the proposal more than a list of safety tests. It is also a plan for who gets emergency power.

Why it matters

The idea has two sides. A single process could make government action easier to understand. Companies would know which risks matter, and a judge would review the evidence. Anthropic says the policies are sensible.

However, a single process could also slow action. A future risk may be severe without crossing 50 deaths or $1 billion in damage. Forecasting those numbers before an incident is hard. If only Commerce can seek a court order, another agency might see a danger but lack a direct way to pause the model.

The proposal also raises a conflict-of-interest question. The evaluator would be outside the lab, but reports say the lab could pay for the evaluation if that payment did not affect independence. The rule would need to make that independence believable to the public.

What is confirmed

The draft exists as a proposal presented to lawmakers. Anthropic has acknowledged its connection to the company’s framework. The 50-person and $1 billion thresholds resemble those in California legislation, so they are part of a wider US policy debate.

What has not happened is just as important. The proposal is not a federal law. It has no confirmed final text, vote, or date for taking effect. Reporting about it is based on a confidential draft, not a passed statute.

What remains unclear

The public still needs to see the complete language. It is not clear how officials would define an imminent risk, calculate likely deaths or damage, or judge that a model has escaped its developer’s control. It is also unclear how the system would interact with state laws and existing regulators.

What to watch next

The next clues are a bill number, a public draft, congressional sponsors, and responses from other agencies. The larger question is whether AI companies should help write the rules for the systems they sell. Any final law will need both a clear way to stop a dangerous model and a clear explanation of who can act before the damage becomes irreversible.