A threshold, not a licence
Anthropic has been quietly pitching members of Congress on its own draft bill to regulate frontier AI, according to reporting by Politico that was relayed in detail by Tom’s Hardware on 18 October. The draft’s defining feature is where it sets the bar: the government could move to block a model’s release or use only if the model poses a risk of causing more than $1bn in damages, or of killing or seriously injuring more than 50 people.
That is a threshold rather than a licensing regime. Nothing in the draft, as reported, requires permission before a model ships. The state’s power attaches only when a specific category of catastrophic risk is in view.
Three categories, and a 72-hour clock
The risk has to come from one of three things: the potential creation or release of a bioweapon, cyberattacks powered or enabled by AI, or a laboratory losing control of a model or its agents. If an evaluator judges that one of those is about to happen, the draft requires a report to the Commerce Department within 72 hours.
What happens next is the part that distinguishes this from most proposals on the table. Commerce cannot act alone: it has to take the evidence to a federal judge in the District Court for the District of Columbia, who decides whether the release is halted. And the draft forecloses every other route — no other department, agency or officer of the United States could order a laboratory to stop releasing a model.
Who does the evaluating, and who pays
The evaluation itself would be done by an independent third party registered with the Commerce Department, periodically assessing both the laboratory and its models. The draft allows the laboratory to pay the evaluator, with the condition that payment must not affect the evaluator’s independence — a clause that will carry most of the weight of the scheme, and the one that will attract the most scrutiny.
Anthropic does not dispute authorship. “This bill represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June,” a spokesperson told Politico. “These are sensible policies we believe Congress should act on.”
Where it sits in the field
The draft arrives into a crowded Congress. Separate proposals would mandate kill switches on the most capable models, and Senator Bernie Sanders has proposed a twenty-year prison term for developing AI superintelligence, putting it on a par with illicit nuclear weapons work. Against that, a bill whose trigger is fifty casualties and whose only enforcement path runs through one district court reads as the narrow end of the range — which is both the case for it and the case against it.
It also lands in a month when Anthropic has been arguing with Washington from the other direction, backing chip export restrictions that other laboratories have stopped pushing for.
What to watch
Three things will determine whether this draft matters. The first is whether any member of Congress introduces it, because a circulating draft is not a bill. The second is the payment clause: an evaluator paid by the laboratory it evaluates is the mechanism the whole structure rests on, and “must not affect independence” is a standard someone has to enforce. The third is the exclusivity provision — stripping every agency except Commerce of the power to stop a model is a large concession to ask of a Congress that has spent the year expanding, not narrowing, the list of regulators with a view on AI.
