1. Home
2. News
3. Anthropic

Anthropic says it will revise the preemption language in its draft federal AI bill after backlash

Its federal affairs lead says states should keep a say on shutting down models, and that the 50-death threshold may need another look.

By Think Facility · October 20, 2026 · Anthropic AI policy Save

On October 18, 2026, Anthony Cimino, who leads federal affairs at Anthropic, answered the people attacking a draft federal AI safety bill his company wrote. He did it in one long post on X. Anthropic will rewrite the section that overrides state AI laws, he said, and it's probably time to look again at the harm threshold that decides when Washington can stop a model.

So the company is backing off a bit, at least on paper. The draft had been passing around since a version reached the press, and Cimino says it only ever went to "a small number of congressional offices" working on AI. The first angry posts went up on the evening of October 16.

1. Oct 16, 7:30 pmShooster Shooster objects
2. Oct 16, 9:06 pmCalvin Calvin's thread
3. Oct 18, 5:29 pmcimino Cimino replies
4. Oct 19, 12:06 pmWiener Wiener weighs in

Eastern time, read off each post on X.

What the critics say the draft does

We're not quoting the draft itself. The only copy we found sits on an anonymous file-sharing site (linked from a newsletter, with no word on who uploaded it), and we can't confirm it's the version Anthropic sent. Everything below comes from people who read it, checked against Cimino's own account.

Nathan Calvin went furthest (he writes that parts of it made him angry). In his reading, the federal government could restrict a frontier model only when there's an imminent risk of catastrophic harm: more than 50 people killed, or more than a billion dollars in damage. Even then a judge decides. And the judge, he says, is told to weigh the harm to the developer with nothing set against it on the public's side.

His other objection is about the states. Calvin says the draft would override laws like California's SB 53 and also stop states enforcing their own orders to halt a model. That would leave the power with Washington alone.

Some people didn't bother with a thread. Ketan Ramakrishnan posted this a couple of hours after Calvin.

What Anthropic says now

Cimino takes the complaints one at a time. Since 2024, he writes, Anthropic has said Congress should override state AI safety law only in return for a very strong federal standard, and any override should be "narrow, even scalpel-like." The draft still leaves states room on child protection and whistleblower laws.

Then, to be fair, he gives ground. Anthropic will rewrite the preemption language so states keep a role in standards for independent evaluators and in regulating whether a lab can shut a model down. He's frank about why the override is there at all. Some members of Congress (he doesn't name them) won't back a bill that leaves regulation unconstrained at both the state and federal level.

The 50-death threshold? Lifted straight from California law, he says, and it doesn't let a company off liability. He defends a real court standard for blocking a model, since that's a big power and could be abused. But the bar can't be so high it never gets used.

Alex Bores went after the timing instead.

Cimino says the draft did move. A court could halt a dangerous model a lab is only using internally, as well as one on public release. Labs would have to report when a model breaks containment and gets into outside systems. He promises tighter incident reporting soon. Over the next few months Anthropic will also look at how the government could slow the whole industry when the risks call for it.

What it means for state laws

Most of this is a fight over which governments get to act. California's SB 53 was written by state senator Scott Wiener and signed by Governor Gavin Newsom on September 29, 2025. It makes frontier AI companies publish safety frameworks, gives people a way to report critical safety incidents to the state's Office of Emergency Services, and protects whistleblowers. New York has its own rules (Anthropic's page doesn't name that law). Anthropic's framework page says both states already require safety frameworks and system cards.

Preemption decides whether rules like those survive a federal law. Anthropic's own framework page sets the test: no preemption unless the federal law is "at least as strong as the framework we are proposing today." Wiener doesn't think the draft passes it. I think that's the hardest version of the complaint to wave away, since it uses Anthropic's own yardstick.

His complaint is specific. The draft takes SB 53's definition of catastrophic risk, which was meant for disclosures, and uses it to narrow when the federal government can stop a deployment. He calls his law "the first step, not the final step." Brad Carson found a different hole: under the draft, he says, developers write their own safety standards, can loosen them later, and hire evaluators who certify themselves.

If you live in California or New York, the question is simple enough. Could your state still step in on its own? Calvin reads the draft as saying no, and Cimino says the rewrite will leave states a role on shutdowns. Granted, I can't settle that until the new language is out.

What happens next

Cimino's post gives no date for the revised text. Wiener, for his part, told lawmakers to look at stronger proposals now being drafted. None of the saved sources shows a member of Congress introducing Anthropic's draft as a bill.