Anthropic has been quietly circulating a draft federal bill on Capitol Hill under which the government could stop the release or use of an advanced artificial intelligence model only if it threatened mass casualties or more than $1 billion in damage, and only through a court order sought by the Commerce Department, Reuters reported, citing Politico. Politico first reported the effort Friday. For companies that build or buy frontier AI, it is a rare look at the rules a leading developer is willing to accept, including a provision that would bar every other federal agency from suspending a model.
What Happened
According to Politico's account, which cited three people with knowledge of the effort, Anthropic representatives have pitched the draft to lawmakers. Reuters, which summarized that reporting, did not name the lawmakers involved.
The draft would let the government block a model's release or use only when it posed a "foreseeable, imminent risk" of more than $1 billion in damages or of killing or seriously injuring more than 50 people. The risk would have to stem from a biological weapon, a model escaping its developer's controls, or a cyberattack.
Developers would have to test frontier models, the most capable AI systems, disclose what they find and submit to independent evaluation. Evaluators would register with the Commerce Department and periodically assess labs and their models. Labs could pay evaluators as long as payment did not depend on the result, and evaluators would have 72 hours after an assessment to report any imminent risk in a listed category to Commerce.
If Commerce concluded a model posed such a risk, it could ask the U.S. District Court for the District of Columbia to suspend it. The language would explicitly bar any other federal department, agency or officer from suspending a model's development, deployment or internal use, Reuters reported.
An Anthropic spokesperson told Reuters the bill reflects the company's Advanced AI Framework, published in June, and would give the government the "ability to block or deter dangerous deployments of the most powerful models." The Commerce Department did not immediately respond to Reuters.
What the Framework Adds
The June framework, written mainly for U.S. federal policymakers, fills in details Reuters did not report. It would cover developers whose models take more than 10 to the 25th power computing operations to train and that either earn more than $500 million a year from AI or spend more than $1 billion a year on AI research. It lists four risk categories: biological weapons, offensive cyber operations, loss of control and automated research that could speed the other three. Developers would publish risk reports at least every six months and report critical safety incidents to the regulator within 15 days.
Among its options for checking regulators, the framework lists court enforcement, under which an agency would have to sue to impose remedies, and limiting the agency to specific violations rather than its own judgment of risk. It also says Congress should not override state AI laws unless it enacts a federal regime at least as strong, and then only narrowly. Reuters did not say whether the draft bill addresses state laws.
Why It Matters
For boards and policymakers, the high thresholds and single-agency design are the core of the debate. They would give developers a predictable, court-supervised process, but they would also limit when and how government could intervene. Anthropic is separately contesting in court the Pentagon's designation of it as a supply-chain risk; Reuters did not connect the two.
The draft joins other proposals. Sen. Maria Cantwell, D-Wash., has released a framework under which covered models could not be released until an independent audit confirmed compliance, as we reported Oct. 9.
What to Watch
- Whether any lawmaker introduces the draft, and whether its text is published.
- How the Commerce Department responds to being named the sole agency able to seek a halt.
- Whether the bill would preempt state frontier AI laws, as Anthropic's framework allows only under strict conditions.
