An AI company is asking Congress for the power to have AI systems switched off. Anthropic, maker of the Claude models, is promoting a bill that would let US authorities suspend the release and use of AI systems judged to pose a serious threat. Companies usually ask for lighter regulation, so this is a slightly unusual request, though a company that builds one of the most powerful model families is also well placed to know what it wants to be regulated.
Subhead: A $1 billion damage threshold, a registry of outside experts, and a court order as the off switch.
What the bill would let the government do
The core mechanism is intervention. If an AI system presents a serious enough danger, the government could move to pause its release or its continued use. The reports on the bill do not describe a general licensing regime. They describe a trigger, a review process and a court.
The trigger has two parts, a size of harm and a type of threat.
- Size of harm. Potential damage of $1 billion or more, or death or serious injury for more than 50 people.
- Type of threat. The risk must be tied to the creation of biological weapons, a model escaping the control of its developer, or a cyberattack.
The list is narrow. A chatbot that gives bad legal advice or writes a mediocre sonnet does not qualify. The bill, as described, is aimed at the scenarios that appear in safety papers and congressional hearings.
Who decides, and how
The judging would not be left to the developers. The bill proposes that independent experts, registered with the US Department of Commerce, assess the potential danger of a given system. On the strength of their findings, the department could go to court and ask for a given project to be suspended.
That is a three-step chain: outside assessors, an agency, then a judge. It is slower than a regulator simply pulling a plug, and that is probably the point. A suspension order against a frontier AI project would be expensive and contentious, and a court is where the argument over evidence would take place.
The reports do not say how experts would be chosen for the registry, how they would be paid, or what standard of proof a court would apply. Those details matter a great deal, and they are not in what has been published so far.
What developers would owe
The bill also places duties on the companies themselves. Developers would have to:
- Set up internal safety units.
- Publish reports on how their AI models work.
- Promptly tell the regulator about serious failures.
Separately, the bill would protect employees who disclose violations from retaliation. Whistleblower protection is an old idea in regulated industries, and it tends to appear in bills like this one for a practical reason. The people who first notice that a model is behaving dangerously are usually the people employed to build it.
Not written on a blank page
According to the reporting, several provisions echo the California AI laws that took effect in 2025. That suggests a familiar pattern. A state tests a framework, an industry player decides the framework is tolerable, and then argues for a federal version that is at least predictable. Whether Anthropic sees it that way is not stated. What the company does say is that the proposed measures are a reasonable way to keep the most powerful AI systems under control.
What we do not know
The sourcing here is thin, and the gaps are worth stating plainly.
- No sponsors or bill number. The reports describe Anthropic promoting the bill, but do not say which members of Congress have introduced it, or whether anyone has.
- No timeline. There is no word on hearings, markup or votes.
- No reaction from other labs. We do not know what competitors, or the Commerce Department itself, think of the proposal.
- Unclear definitions. Terms such as a model going out of its developer's control are doing a lot of work, and the summary does not say how they would be defined in law.
The wry part
There is an old observation that incumbents sometimes welcome regulation, because compliance departments are cheaper for the big than for the small. Nothing in the reports says that is happening here, and the bill's thresholds (a billion dollars of damage, more than 50 casualties) are aimed at catastrophic risks rather than at smaller developers as such. Still, a rule requiring internal safety teams and published reports is easier to meet if you already have both.
The more interesting reading is the plain one. A company that sells powerful models is saying, in public, that there should be a legal way to stop powerful models. Whether Congress takes that offer up, and whether the rest of the industry agrees, will decide if this is a policy proposal or a press release with a bill number attached.