Anthropic wants Washington to police dangerous AI, and it wants strict rules for when Washington can stop a model. The Claude maker has privately pitched lawmakers a safety bill that would put numerical thresholds, outside evaluators and a federal judge between regulators and a shutdown.
Under the reported draft, intervention would require a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people. Those are conditions for responding to a threatened catastrophe. The proposal being circulated to Congress is not law.
Who gets to pull the plug?
Politico reported the effort on October 16, citing three people familiar with it. Reuters subsequently reported the proposed terms, including a narrow list of qualifying threats: biological weapons, cyberattacks and an AI model escaping its developer's controls.
The Commerce Department could ask a federal judge in the District of Columbia to suspend a model's release or use. The reported language would bar other U.S. departments, agencies and officials from suspending its development, deployment or internal use.
Outside assessors, paid by the companies they inspect
Independent evaluators would register with Commerce and periodically assess frontier AI developers and their models. Companies could pay the assessors, provided the payment was not tied to the outcome of their evaluation.
If an assessment identified an imminent risk covered by the proposal, the evaluator would have to report it to Commerce within 72 hours of completing the assessment. That deadline runs from completion of the review.
This gives the evaluators a consequential role: their assessments would help determine whether the government has grounds to seek a suspension. Who qualifies to do that work, what they can inspect and how independently they can operate would matter as much as the reporting deadline.
The company has been spelling out its preferred limits
The public framework Anthropic published in June helps explain its approach, although that document is separate from the confidential legislative draft. It proposes testing, outside evaluations, risk disclosures and government accountability, alongside safeguards against regulatory overreach.
Among its enforcement options, the company describes agencies pursuing remedies through lawsuits. It also discusses expedited cases and provisional remedies for imminent catastrophic risks. The paper describes the tradeoff this way:
The remedies and safeguards specified above can be mixed and matched to achieve different tradeoffs between the risks of inadequately safeguarded models and the risks of overbroad regulation.
Its position on state authority is conditional, too. The June paper says Congress should not preempt state law without a rigorous federal regime matching or exceeding the strongest measures it proposes. Any override should be limited to specific functions Congress expressly takes over, such as catastrophic-risk testing, evaluator licensing and related reporting.
The same document adds:
Compliance with a federal regime should not itself confer immunity, a safe harbor, or a presumption against liability under otherwise applicable state law. Absent a strong federal law, states should retain full authority to legislate.
Those are the company's published policy preferences, with the final scope of any legislation still dependent on what Congress takes up and passes. Anthropic and Commerce did not immediately respond to Reuters' requests for comment on the reported bill.
Clark says the industry still needs guardrails
The lobbying effort comes alongside a fresh public appeal for oversight from cofounder Jack Clark. Speaking on October 16 at an IMF-World Bank annual meetings event in Bangkok, he compared the industry's regulatory situation to Formula One cars racing at night without brakes, according to AFP.
He put the gap between ordinary consumer rules and AI oversight more bluntly:
It's remarkable to me that the toys I buy for my children are regulated more than the literal super intelligence I work on at my day job.