Congress Finally Has a Real AI Bill. The Fight Is Over Three Words: States Can't Act.

Covered companies would have to publish a "frontier AI framework," a detailed safety plan covering how they identify and manage the most severe risks their models could pose, and submit to semi-annual third-party audits against it. The draft also authorizes $100 million per year for a Center for AI Standards and Innovation and adds oversight requirements for government AI adoption.

Worth pausing on: mandatory third-party audits of frontier labs, on a schedule, in a bipartisan bill. Two years ago that idea lived exclusively in think-tank papers. The Overton window has moved further than the discourse admits.

The preemption trade

Now the contentious part. In exchange for federal obligations, the draft bars states from enacting new laws regulating AI development for three years, with a sunset. States keep their authority to regulate the use of AI within their borders; what they lose is the ability to legislate how models are built.

The development-versus-use line is the entire ballgame, and both sides know it. Industry has spent two years warning about a fifty-state patchwork, and the draft hands them exactly the shield they asked for, time-limited. Consumer groups read the same text and see the reverse. Public Citizen's response, titled "Obernolte-Trahan Bill Strips States Authority to Protect Consumers, Workers, and Children", argues the development carve-out will be lawyered into covering anything a state tries to do, since every harmful deployment can be recharacterized as a development choice.

Both readings are plausible, which is what makes the drafting fight matter. Where exactly "built" ends and "used" begins is a boundary that will be litigated for the full three years if this passes as written.

Why the timing isn't an accident

The bill dropped during the most consequential month in AI's commercial history. Anthropic and OpenAI filed for IPOs within a week of the draft's release, and SpaceX, now fused to xAI, priced the largest listing ever. Companies heading into public markets crave regulatory predictability; a three-year federal preemption with knowable audit requirements is exactly the kind of thing you'd want in your risk factors section instead of "fifty legislatures, unknowable."

I don't think that's conspiracy, just convergence. The industry got big enough that everyone now needs the rules written down, including the companies being regulated.

The takeaway

Discussion drafts exist to take fire, and this one will. The audit and disclosure provisions have a real bipartisan lane. The preemption language, as written, does not; expect it to narrow, sprout exceptions for children's safety and biometric laws, or shrink to eighteen months before anything reaches a floor vote.

The thing to watch isn't the bill's fate this session. It's that both parties just agreed, on paper, that frontier AI development is a federally regulated activity. That premise survives even if this draft doesn't.