Congress Finally Wrote the Preemption Down: Three Years, Development Only. Sacramento Keeps the Rest.

For a year, federal preemption of state AI law has been a threat without a text. On Thursday it got one. Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a 269-page discussion draft of the Great American Artificial Intelligence Act, the most complete federal AI framework Congress has produced. The headline provision freezes state laws that specifically regulate the development of AI models for three years. The detail that matters just as much: it leaves state laws on the use and deployment of AI alone.

I wrote last week that with no federal standard above them, California's roughly 30 surviving AI bills were about to set the US floor. This draft is the federal standard showing up, and it is narrower than the preemption fight everyone was bracing for. It does not erase Sacramento. It splits the stack and claims the bottom half.

What the Draft Actually Says

The bill is a discussion draft, not introduced legislation, which means it is a position paper with bill formatting. But it is a bipartisan one, with co-sponsors on both sides: Suhas Subramanyam (D-VA), Scott Franklin (R-FL), Scott Peters (D-CA), and Erin Houchin (R-IN). Obernolte co-chaired the House AI Task Force. Trahan is one of the few members with a working technical background. This is the text the preemption fight will be argued over for the rest of the year.

The core provisions:

Preemption, scoped and sunset. State laws and regulations "specifically regulating the development" of an AI model are preempted. The preemption expires after three years, a forcing function that obliges Congress to revisit the framework rather than set it and forget it. Laws governing the use or deployment of AI models are explicitly not preempted.

CAISI gets statutory legs. The draft formally establishes the Center for AI Standards and Innovation, the renamed successor to the AI Safety Institute, and appropriates $100 million per year from 2027 through 2029. Its standards remain voluntary.

Obligations for frontier developers. Large frontier developers would be required to write and implement risk management plans before releasing new models, and to report critical safety incidents to CAISI.

Whistleblower protections. Frontier labs would be barred from retaliating against employees who report potential violations, critical safety incidents, or other risks related to advanced model development.

If that obligation list sounds familiar, it should. Risk plans before release, incident reporting to a named agency, whistleblower protections: that is the skeleton of California's SB 53, federalized. The draft does not so much repeal the state model layer as nationalize it, with CAISI standing in for the California Attorney General and voluntary standards standing in for enforcement. That last substitution is where the fight will be.

Which State Laws Live or Die

The development versus deployment line is the entire ballgame, so here is how it cuts through the state laws we have been tracking. Trahan's office published an accompanying document naming two California casualties directly.

Read that table against the Sacramento crop and the result is almost backwards from the rhetoric. Public Citizen called the draft a bill that strips states of authority to protect consumers, workers, and children. But the consumer, worker, and child-facing bills in California are nearly all deployment-layer bills, and the draft leaves them standing. What it actually freezes is the transparency and governance regime aimed at the labs themselves: training data disclosure, watermarking at the model level, and very likely SB 53's frontier governance framework, the one statute the labs have already built compliance documents against.

That last point deserves emphasis. OpenAI published a Frontier Governance Framework mapped to SB 53 and the EU AI Act eight days ago. Anthropic maintains its Frontier Compliance Framework against the same statute. If this draft becomes law, the named state anchor under those documents dissolves for three years and is replaced by a federal risk plan requirement with voluntary CAISI standards underneath it. The compliance documents will survive; the enforcement teeth behind them change owners.

The Week Washington Reversed Itself

The timing is not subtle. On May 28 the administration pulled a federal AI review executive order hours before signing, after calls from David Sacks, Elon Musk, and Mark Zuckerberg. We covered it as the federal government handing the rulebook to Sacramento and Brussels. Five days later, on June 2, a version of that order was signed after all: AI developers can voluntarily submit frontier models for up to 30 days of federal review before release, focused on cybersecurity and national security risks. Thirty days instead of ninety, voluntary instead of expected, but signed.

Two days after that, this draft landed. In one week, Washington went from no federal model-layer policy to a signed review order plus the most serious preemption text ever circulated. The federal vacuum I described on May 31 is filling in fast, and it is filling in precisely at the layer where the labs wanted relief: the model layer, where a 50-state patchwork is most expensive to comply with and where California was setting the de facto national standard.

Our Take

The development versus deployment split is a better-engineered compromise than I expected from this Congress. It gives the labs the one thing they have lobbied hardest for, a single national rulebook for building models, while letting states keep legislating where AI touches their residents: chatbots, hiring, healthcare, schools. The three-year sunset is the cleverest part. It converts preemption from a permanent land grab into a rental with a renewal vote, which is what makes the bipartisan sponsor list possible.

The weakness is the same one every voluntary framework has. SB 53 carries a $1 million per violation penalty enforced by a state AG. The draft replaces that with risk plans checked against voluntary CAISI standards and an incident reporting duty with unspecified teeth. For three years, the strongest binding obligations on frontier model development in the United States would get weaker, not stronger, even as the obligation list looks similar on paper. Whether that trade is worth a uniform national standard is the actual debate, and it is now a concrete one with section numbers instead of a Twitter fight.

Discussion drafts die more often than they pass, and this one faces an election year, a crowded calendar, and governors in both parties who like their AI laws. But the preemption argument now has an anchor text, and every state bill we track gets read against it from here on. Watch three things over the next ninety days: whether the draft gets formally introduced with a hearing date, whether California's July 2 sprint accelerates to get bills signed before any federal freeze can attach, and whether the labs publicly endorse the draft or stay quiet and let the trade associations carry it. The quiet path tells you they think they can get preemption without the federal obligations attached.