A federal AI bill that cancels state rules on the day it passes, but asks nothing of AI companies for months, would leave a stretch in which the displaced state rules have no federal replacement. Zvi Mowshowitz, a writer who closely follows AI developments and AI risk, raised that objection against a draft federal AI bill that, according to the show, Anthropic wrote and that had leaked the previous week. As drafted, he said, the bill "would immediately nullify" the state laws it displaces, "but it would take many months before any implementation of the new requirement." The result, in his words: "no such regulation and a ban on such regulation by the states for like six months or more."
Mowshowitz made the case on The Cognitive Revolution, in a highlights episode published October 24, 2026, and hosted by Nathan Labenz and Prakash Narayanan. Narayanan took the other side and made what Mowshowitz called "an excellent steel man" of Anthropic's position.
What the draft trades
The central bargain is preemption: a federal law that overrides, or preempts, state laws on the same subject. AI companies would get one national rulebook in place of state rules, and in return they would accept federal obligations.
Labenz described the leaked draft as resembling the FRONTIER Act, a compromise bill that Mowshowitz calls well-meaning: federal transparency rules in exchange for preempting state AI laws. Labenz placed the FRONTIER Act in June, but the introduced text of H.R. 9925 is dated July 23, 2026. Labenz said the Anthropic draft drew heavy criticism, and the argument centered on what Washington should offer now in exchange for displacing the states.
The case for passing something now
Narayanan began by assuming everyone was acting in good faith. On that reading, he said, Anthropic wants to "get anything done as soon as possible." And if a bill has broad bipartisan support, it can be adjusted later: "There's no reason that you only have to pass one bill." Legislation often takes several sessions of Congress to fine-tune, he said, and Congress can move fast when it has to, as it did during COVID and the 2008 financial crisis. "When people are scared, things move," Narayanan said. In his view, Congress is slow only when there is no political consensus, and building that consensus "is a process on its own."
Why Mowshowitz thinks the bargain is bad
Mowshowitz said he could imagine Anthropic acting in roughly that spirit, caught in an old idea of what a realistic compromise looks like, with the same people in Congress and the sense that time is short. If so, he said, the company "misread the situation pretty severely." He believes support for regulation is "much stronger now" and that advocates can push for "quite a bit more" than the FRONTIER Act. He also said the draft is weaker than FRONTIER, "largely in a clearly self-centered way," with provisions that suit the company's particular interests.
The company's response frustrated him. It would make sense, he said, if Anthropic had explained that it was circulating many drafts or just putting out feelers. Instead, the messaging he heard sounded like endorsement, followed by apologies about "some of the details" rather than an admission that the bill was too weak for the current political moment.
Mowshowitz called Narayanan's last argument, that Congress can always pass another bill, "potentially the strongest one," and agreed it holds for a true emergency. But he said preemption is a bargaining chip that can be spent only once. Once it is given, "you can't offer it again." Some supporters would also already have what they cared about most, so the case for a second bill would be weaker. "I strongly believe that if something like this were passed it would make it harder to pass a bill in the new year," he said.
The gap between repeal and requirements
The debate then turned to whether a weaker bill might still reduce some risk in the months before a stronger one could pass. Mowshowitz said any real risk reduction would be a good thing, but as written, he believes the bill "would actually increase risk," because what it takes away by cutting off state laws outweighs the safety it adds.
The counterpoint in the discussion was that the interim period might last only a few months, and that state laws rarely bite at once. Many give companies three months, six months or a year to comply. Mowshowitz answered with the timing problem: preemption would take effect immediately, while the bill's own requirements would arrive many months later. He called the resulting stretch, with no such regulation and a ban on state rules of that kind, "obviously disaster from the perspective of moving quickly," and added that he considers it "almost a bug" and "clearly a mistake."
Emergency powers and the trust problem
Mowshowitz had listed the government's recent ad hoc moves against AI companies, including the takedown of Fable, which he cited as an emergency action against a model. Narayanan pushed back that if technology accelerates, "you're going to need an acceleration of governance to kind of keep up."
Mowshowitz agreed some improvisation will be necessary. Part of the reason to pass legislation, he said, is to make government action "much more formalized and bounded," with clear judicial review. The test he applied to the draft: if the government wanted to take down a model in an emergency again, how would it do it? Under Anthropic's bill, he said, officials would have to argue before a judge, with limited chances to sustain an order and high burdens of proof that they might lack the information to meet. He called it "a very onerous set of provisions." In a real emergency, he said, the government would skip the new process and fall back on its existing tools, such as export controls or general threats.
The FRONTIER Act takes a different approach. Its introduced text lets the Commerce Secretary issue a provisional restriction on a preliminary determination of imminent catastrophic risk, even before a technical assessment is prepared. Before issuing such an order, the Secretary must normally give the developer notice and an opportunity to cure the problem and be heard. Only if the Secretary determines in writing that the risk is too imminent for that can the order come first, in which case the developer gets the chance to cure promptly afterward. Provisional orders lapse after a set period, and the bill provides for expedited hearings and judicial review.
Mowshowitz traced the weakness to a conflict in Anthropic's position. The company is unwilling to grant reasonable emergency powers "to this administration," he said, because it does not trust the people running the federal government. Yet it wants to move authority from the states to that same government. Crippling the federal powers while preempting the states, he said, is "worse than just leaving it in the states." Anthropic, in his view, "is trying to cut it both ways."
"Meet the moment"
Asked what else was on his mind, Mowshowitz returned to policy. What upset him most, he said, is that the draft does not try to meet the moment. He pointed to Anthropic CEO Dario Amodei's call to pace the frontier, the theme of Amodei's essay We Must Pace the Frontier, which proposes embedded outside evaluators, pacing arrangements among democracies and, eventually, international coordination. According to the labs and people inside them, Mowshowitz said, a crisis is coming in which capabilities outrun alignment, forcing hard choices between commercial viability and catastrophic problems, "including potentially loss of control at some point in 2027." If a proposal cannot meet that, he asked, why not keep working until it can? At minimum, he said, the response has to "count as some form of facing the frontier."
Labenz then raised a recent public exchange. Nathan Calvin of Encode had listed nine Anthropic decisions he saw as putting the race ahead of safety, and Amodei had replied at length. In that reply, as reported by The Next Web on October 21, 2026, Amodei cited sacrifices such as holding back broad access to Mythos-class models for four months for cyber safeguards, and accepted responsibility for poorly handled policy communication.
Mowshowitz's assessment was mixed. Anthropic has "a much stronger safety culture although not adequate to the moment," he said, and is willing to make big sacrifices where it, or Amodei in particular, holds strong beliefs. But the company also competes, buys capacity and encourages more chip production and power supply, "acting like a normal corporation remarkably large percent of the time," and on policy it often falls short of its ideals. He added that he had recently heard views from various employees about what it would take to solve alignment problems that he found alarming.
Anthropic still seems to be the most responsible lab, he said, "but you know that's grading on a curve and I don't like grading on curves because reality doesn't have a curve, right? Reality just decides what happens."