Anthropic writes its own AI bill: a stop only above 50 victims or a billion dollars

The draft applies to what it calls covered developers: companies that train a model with more than 10^25 FLOP, a measure of the computing used in training, and that also earn more than 500 million dollars a year from AI or spend more than 1 billion dollars a year on AI research. Those are the same thresholds as in the Advanced AI Framework Anthropic published in June. Anthropic says the draft reflects the proposals in that framework.

Those companies would have to publish a safety framework for catastrophic risks, publish reports on what their models can do and how they were tested, report serious safety incidents, and protect whistleblowers. Independent evaluators, registered with the Commerce Department, would periodically review the labs and their models. The labs may pay them, but the payment may not depend on the outcome. If an evaluator finds an imminent risk in one of the three categories, it must tell Commerce within 72 hours.

Commerce can then ask a federal judge in the District of Columbia to suspend the model. The reported text bars every other US department, agency or officer from suspending a model's development, deployment or internal use. An Anthropic spokesperson calls the proposals sensible policy.

The draft would also set aside state laws on public risk disclosure, outside evaluation and incident reporting. That touches laws such as California's SB 53, which Anthropic backed in September 2025, and New York's RAISE Act. Anthropic's own framework from June says Congress should only do that with a federal regime that meets or exceeds the strongest measures in that framework.

What falls outside the stop. Read literally, the trigger leaves out every risk below 50 victims or a billion dollars, however widespread. It leaves out every risk outside the three categories: fraud and scams, disinformation and election interference, discrimination in hiring, lending or policing, privacy and surveillance, chemical, radiological and nuclear risks unless they fit under biology, job losses and market manipulation, and smaller-scale hacking. It leaves out risks that are building up slowly, or that nobody can yet call foreseeable and imminent. And it leaves out smaller labs and models below the thresholds.

That does not mean those harms are unregulated. The disclosure and evaluation duties still apply to the covered labs, and other law still exists. It means that this bill's stop does not reach them.

The proposals that were already on the table are in the superintelligence ban and the off switch. A week before Politico, Axios reported that the companies are preparing for the law that would follow a first catastrophe. Until now, American AI regulation ran mostly through the states, and that is exactly the level this draft would push aside.

What it means for Europe is a reading, not something any source reports. A US federal law does not change what applies here. A model offered in the EU answers to the AI Act regardless.

The AI Act uses the same 10^25 FLOP line, but as a presumption that a general-purpose model carries systemic risk (article 51), without a revenue filter. Those providers have had to evaluate their models, assess and mitigate risks, and report serious incidents to the AI Office since August 2025. Since 2 August 2026 the Commission can enforce that. Under article 93 it can, after an evaluation that raises serious and substantiated concern about a systemic risk, order mitigations or restrict, withdraw or recall a model. Under article 101 it can fine up to 3 percent of worldwide turnover or 15 million euro. The text has no floor in victims or dollars, and no list of three categories.

So on paper Brussels can act on a model earlier than Washington could under this draft. The reverse also holds. The AI Act is market law and does not reach a model that is only used inside a company. The July break-in came from exactly such an internal prototype, see what the AI Act does not do. The American draft does name internal use.

A second reading. If the draft passed with its preemption, California and New York would lose rules that look most like the European ones. Brussels would then be the strictest regulator these labs answer to, with an AI Office that Lawfare describes as underresourced for its task. The public reports the draft requires would also be readable in Europe, and Europe currently knows mostly what the companies choose to say.

Whether Washington would use such a law to push Brussels towards the same narrow trigger is in no source. Anthropic said in July 2025 that it intends to sign the EU's code of practice for these models.

The draft text is not public. I have not read the Politico article itself. Everything about the draft rests on Politico's copy, through Reuters and secondary coverage. The thresholds for covered developers I checked only against the June framework, not against the draft.

I read Zvi Mowshowitz and the LessWrong analysis in excerpts, not end to end. The FRONTIER Act text I have not read.

I found no reaction from the European Commission, the AI Office or a Belgian government body as of 18 October.

Whether the incidents in this dossier would have reached the trigger cannot be settled from here. For most of them there is no public damage figure in dollars. For the French supermarkets, NOS wrote many millions, not a billion. They are not attributed to an AI model either.