What We Know About Anthropic's Proposed AI Safety Bill

Anthropic has drafted a federal AI safety bill that would let regulators ask a judge to block the release of a dangerous AI model, but only after the company's own hired evaluators flag the danger first, according to TechCrunch, which cited an earlier Politico report on the draft legislation.

Anthropic has been circulating the bill on Capitol Hill, hoping lawmakers will introduce it, TechCrunch reports. A company spokesperson confirmed the draft to Politico, saying it reflects proposals the company has pushed since June under what it calls its Advanced AI Framework.

What does the bill actually do?

The draft creates a legal process for pausing an AI model's public release if it poses what the bill calls a "foreseeable, imminent risk," according to TechCrunch's reporting on the Politico story. That risk has to come from one of three things: a biological weapon, a model that escapes its developer's control, or a cyberattack. If an independent evaluator flags one of those risks, the company or the government would still need a federal judge's sign-off in Washington, D.C., before the model's release could be stopped. No agency could act on its own, the draft states.

What counts as a "catastrophic risk"?

The bill sets numeric thresholds. A risk qualifies as catastrophic if more than 50 people could die or if property damage could top $1 billion, TechCrunch reports, citing Politico. Those same numbers appear in California's Senate Bill 53, a state AI safety law passed this year that defines catastrophic risk the same way. Anthropic's federal draft borrows that language directly.

Who decides if a model is dangerous?

Under the proposal, "independent evaluators" registered with the U.S. Commerce Department would assess whether a model crosses one of those thresholds, according to TechCrunch. The bill allows AI companies to pay those evaluators directly, as long as the payment isn't tied to the outcome of the evaluation. TechCrunch notes it isn't clear how that rule would actually stop payments from shaping results, since the evaluators would still depend on the companies they're reviewing for income. DeepMind chairman Demis Hassabis has separately floated a self-regulatory model similar to FINRA, the financial industry's self-policing body, though that idea isn't formally part of Anthropic's draft.

Why can't federal agencies just block a model themselves?

This is the part of the bill drawing the most attention. If an evaluator spots an imminent catastrophic risk, it has 72 hours to notify the Commerce Department, TechCrunch reports. From there, Commerce doesn't get to halt the release on its own. It has to ask a federal judge in Washington for a court order. The bill explicitly bars any federal agency from having discretion to stop a model's distribution without that judicial step. Supporters of this structure would likely argue it prevents regulators from acting on incomplete information or political pressure. Critics would likely counter that building in a court delay, on top of a 72-hour evaluator notice window, slows down the one moment when speed might matter most — when a model is about to ship and a risk has already been identified.

How does this compare with existing state law?

California's SB 53 already uses the same 50-death and $1 billion thresholds to define catastrophic risk, and it became law this year. Anthropic's federal bill would apply those same numbers nationally, but it adds the court-order requirement that SB 53 doesn't include in the same way. That's the structural difference worth watching if the bill moves to a congressional sponsor: whether lawmakers keep the judicial-review step or give an agency like Commerce more direct authority to act.

What happens next?

Anthropic has not said which lawmaker, if any, will introduce the bill, and the company did not respond to TechCrunch's request for comment beyond the spokesperson's statement confirming the draft to Politico. Nothing in the reporting indicates a hearing date or a sponsor has been set. For now, the proposal is a lobbying document, not pending legislation, which means no monitoring changes take effect for consumers who use AI chat assistants or writing tools on their phones and laptops this week.

Readers who rely on everyday AI tools, including apps like Smorgi's mindfulAI Keyboard, won't see any immediate effect from this draft, since it targets future model releases from major labs, not products already on the market.

The debate TechCrunch frames is less about whether 50 deaths or $1 billion in damage is the right bar, and more about who gets to pull the trigger once that bar is met — a company-funded evaluator, a federal agency, or a judge.