For three years, no state would be allowed to write a new law governing how AI models are built. That is the single most consequential line in The Great American Artificial Intelligence Act, a 269-page discussion draft unveiled on June 4, 2026 — and it is the line that turned a bipartisan safety bill into a fight.
The framework comes from an unusually broad coalition: Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.), joined by Suhas Subramanyam (D-Va.), Scott Franklin (R-Fla.), Scott Peters (D-Calif.), and Erin Houchin (R-Ind.). It lands just days after President Trump signed an executive order establishing voluntary federal reviews of new frontier models. On paper, the bill does more than the executive order: it creates real mandates, real penalties, and real audits. The question dividing Washington is whether the price of those federal rules — silencing the states — is worth paying.
What the bill actually requires
Strip away the preemption fight and the Great American AI Act is one of the more substantive federal AI proposals to date. Its core mandates target "large frontier developers" — companies with more than $500 million in gross revenue in the prior calendar year.
Those developers would have to:
- Publish a frontier AI framework describing whether a model could pose a "catastrophic risk" — defined as a foreseeable, material risk of death or injury to more than 50 people, or more than $1 billion in property damage — and the thresholds used to judge that risk.
- Document how they secure nonpublic model weights, respond to critical safety incidents, and manage risks from internal use, including a model circumventing oversight.
- Retain an Independent Verification Organization (IVO), licensed through NIST's Center for AI Standards and Innovation (CAISI), to perform semi-annual compliance audits.
- Report critical safety incidents within 15 days of discovery and imminent risks within 24 hours.
The teeth are financial: developers face liability of up to $1 million per day for violations of the safety requirements, and the same penalty for ignoring the audit regime or making material misrepresentations. The bill also formally establishes CAISI in statute, authorizes $100 million per year for fiscal 2027 through 2029, and adds whistleblower protections for employees who report violations.
On its own merits, this is a credible safety architecture: mandatory frameworks, independent audits, incident reporting, and whistleblower cover. The objection is not to what the bill adds federally. It is to what the bill takes away from the states.
The preemption clause
Here is the contested mechanism. The bill would preempt state laws "specifically regulating the development" of an AI model, with a three-year sunset. Preemption would not apply to laws governing the use or deployment of AI — so a state could still regulate how a hospital or a landlord uses an AI system. What states could not do, for three years, is pass new rules about how the models themselves are trained and built.
A FAQ from Trahan's office named specific casualties. California's AB 2013, which requires developers to publicly post high-level summaries of their training data, would be preempted. So would the watermarking portion of California's SB 942, the AI Transparency Act. Frontier-safety laws in California, New York, and Illinois would be "federalized" — folded into the federal standard, with state attorneys general retaining enforcement authority.
Why the safety community revolted
The backlash was immediate, and notably it came from the AI-safety camp rather than industry.
Brad Carson, president of Americans for Responsible Innovation and a former Democratic congressman, called the preemption provision a "generational mistake." His argument: "This bill takes the current floor on state AI legislation and turns it into a federal ceiling, preventing state lawmakers from addressing emerging AI harms in an era of fast-moving technology." ARI even launched an ad campaign in Massachusetts pressuring Trahan to drop the ban.
Brendan Steinhauser, CEO of the Alliance for Secure AI, praised the bill's bipartisanship and its focus on catastrophic risk — then opposed the preemption anyway: "A national AI standard should protect at least as much as it preempts." His point is that several states already do more on child safety and consumer protection than this federal draft would.
The structural worry is timing. Three years is a long freeze in a field where capabilities shift every few months. If a novel harm emerges in 2027 that the federal framework didn't anticipate, states — historically the faster-moving regulators — would be locked out of responding to its development.
The case for a single federal standard
The sponsors are not making a deregulatory argument, and it's worth taking their reasoning seriously. A patchwork of 50 state regimes for how models are built is a genuine compliance problem: a developer can ship one model nationally, but cannot easily train fifty versions to satisfy fifty different training-data or safety rules. Trahan's office argues that auditing, transparency, and whistleblower standards "belong and can be better implemented at the federal level," with state AGs still enforcing them.
There is also a coherence argument. Frontier-model development is inherently interstate and international; a California-only training rule governs a process that doesn't stop at California's border. Centralizing the development layer while leaving the deployment layer to states is at least a principled line — even if reasonable people disagree about where it should fall.
Industry, for its part, isn't fully satisfied either. NetChoice's Patrick Hedger backed the bill's goals but flagged the "aggressive auditing regime and data-sharing requirements" as a risk to trade secrets — a reminder that the preemption clause is the loudest fight, not the only one.
What happens next
This is a discussion draft, not introduced legislation, and both sponsors have explicitly invited changes. That makes the preemption clause a negotiating position as much as a final demand. The likeliest paths forward are a shorter sunset, a carve-out preserving state authority over child-safety and consumer-protection harms, or a narrower definition of what counts as "regulating development."
The deeper tension won't resolve cleanly. With binding federal AI law stalled for years, states became the de facto regulators — and any federal framework now has to decide whether to build on that foundation or override it. The Great American AI Act chooses override, for three years, in exchange for the first real federal safety mandates. Whether that trade is wisdom or overreach is precisely what the coming months of markup will test.
The Bottom Line
The Great American AI Act is two bills wearing one title: a serious federal safety regime, and a three-year ban on state-level AI development laws. Supporters see a long-overdue national standard that ends an unworkable patchwork; critics see a federal ceiling slammed down on the states that have done the most actual protecting. Both can be right. The real test is whether Congress can keep the audits, penalties, and whistleblower protections while narrowing a preemption clause that, as written, asks Americans to trust that Washington has anticipated every harm the next three years will bring.
