California State Capitol at dusk

Last week, Anthropic co-founder Jack Clark told the IMF that AI needs “unusual and serious rules.” Now we know what Anthropic thinks those rules should say — because the lab wrote them itself.

On October 16, POLITICO reported that Anthropic has been privately circulating a draft federal AI safety bill among US lawmakers — obtained by POLITICO as a confidential text, with three people familiar with the talks confirming the effort. It is the most detailed attempt yet by a frontier lab to write its own regulation. No congressional sponsor has been identified, and it has not been formally introduced.

What’s in the draft

The core is a deliberately high threshold for government intervention. Under the draft, Washington could block a model’s release or use only for a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people — and only when the risk stems from one of three sources: a biological weapon, a model slipping out of its developer’s control, or a cyberattack. The language closely mirrors California’s SB 53.

The mechanism is equally specific:

  • Only the Commerce Department could ask a federal judge — in the D.C. District Court — to halt a model. The draft explicitly bars every other department, agency, or officer from suspending a model’s development, deployment, or internal use.
  • Independent evaluators registered with Commerce would periodically assess frontier developers. Labs could pay the evaluators, provided compensation isn’t tied to findings. An evaluator finding an imminent covered risk must notify Commerce within 72 hours.
  • Labs would need safety frameworks, whistleblower protections, a published transparency report for each new frontier model, and incident reporting to Commerce.
  • The bill would preempt parts of state AI law — including California’s — covering public risk disclosure, third-party evaluation, and critical incident reporting.

Why the preemption matters most

The preemption clause is the sharpest part of the story, because it represents a reversal. In June, Anthropic wrote that without a strong federal law, states should retain full authority to legislate. Four months later, its draft overrides the states that did exactly that — California, New York, Illinois — replacing their disclosure and evaluation rules with a single federal standard.

Anthropic describes the draft as the legislative version of its Advanced AI Framework, which it says it has advocated since June. The company’s spokesperson framed it as giving the government the ability to “block or deter dangerous deployments of the most powerful models.” Charlie Bullock of the Institute for Law & AI called it “pretty reasonable” but weaker than the competing FRONTIER Act introduced in the House this summer.

Notice what the draft leaves out: the enforcement mechanism covers only three catastrophic risk categories. Bias, misinformation, and job displacement — the harms most people actually encounter — are not covered by the halt authority at all. And the judge-not-agency structure means even qualifying interventions move at the speed of federal litigation.

Reading the strategy

This is the second act of a play that opened at the IMF meetings in Bangkok, where Clark’s $44 trillion GDP projection argued that AI’s upside demands “unusual and serious rules.” The lab arguing for dramatic growth and the lab drafting the rules constraining it are the same lab — and the draft’s design choices all point in one direction: a federal ceiling, not a floor.

A high, narrowly defined intervention threshold plus state preemption plus judge-only enforcement equals regulatory certainty for the labs: one known standard, nearly impossible to trigger, displacing the patchwork of stricter state regimes that was starting to form. That may be good policy or bad policy — but it is unmistakably a policy written by the regulated.

The honest version of Anthropic’s position has two layers. The safety layer is real: third-party evaluators with publication rights, 72-hour reporting, and a halt mechanism for genuine catastrophes are more than any other frontier lab has put in legislative text. The self-interest layer is equally real: when you write the rules, you get to define “serious” as $1 billion and 50 bodies — a bar almost nothing short of catastrophe clears.

Congress hasn’t passed a federal AI law, and current policy leans on voluntary pledges. Whether lawmakers adopt a bill the industry’s most safety-conscious lab wrote for itself — or whether the FRONTIER Act’s broader standard wins — will say a lot about who actually gets to define AI risk in the United States.

Sources: POLITICO’s Oct 16, 2026 exclusive (via resultsense, dev.to, trendintech); TradingView/Seeking Alpha; plainnewsbyai. Draft terms are as reported by POLITICO from the confidential text; characterization of the strategy is the author’s analysis.