The Audits Are Real. They're Also the Wrapping Paper on a State-Law Freeze.

Eleven months ago the United States Senate took a 10-year ban on state AI regulation and voted it down 99 to 1. Not a close call. A near-unanimous, bipartisan, get-this-out-of-my-sight repudiation. So the instructive thing about the Great American Artificial Intelligence Act — a 269-page discussion draft Reps. Jay Obernolte and Lori Trahan put out on June 4 — is not that it proposes a freeze on state AI law. It's that it learned the lesson of the 99-1 vote, which was never "don't do this." It was "don't do this nakedly."

So this time the freeze comes dressed. Around a three-year preemption of state laws that regulate AI development, the draft wraps semi-annual third-party audits of the big labs, mandatory transparency obligations, whistleblower protections with teeth (reinstatement, double back pay, legal fees), and a Center for AI Standards and Innovation funded at $100 million a year — up from $15 million. That is a real oversight regime. I want to be fair about that, because the cynical read is too easy and the honest read is more interesting: these are not nothing. They are more than the frontier labs asked for, and more than the dead moratorium offered, which was nothing at all.

But watch where the industry support actually lands. The trade groups lining up behind the draft — BSA, the Information Technology Industry Council — are enthusiastic about the preemption clause and conspicuously quieter about the audit mandates. Their support tracks the freeze, not the oversight. That tells you which provision is the product and which is the packaging. The audits are what you put on the box so that a Democrat can vote yes. The preemption is what's inside.

Here is the mechanism, and it's the part almost no coverage is interrogating concretely. The freeze applies to state laws that "specifically regulate AI development." The draft includes savings clauses that loudly preserve state authority over privacy, consumer protection, anti-discrimination, civil rights, and the deployment of AI. Sounds comprehensive. Read it against what states actually built.

Training a frontier model is development. That's not a gray area — it's the definitional center of the term. And the marquee state laws this would freeze are aimed at exactly that stage. California's SB 53 governs frontier-model transparency. California's AB 2013 forces disclosure of training data. Colorado's consumer-protection statute takes effect today, June 30, the very day a development-stage preemption would begin freezing it. These are development laws because the thing they regulate — how frontier models get built and what gets disclosed about the building — happens at development. So when the draft says it "preserves consumer protection," the question that decides everything is whether the deployment carve-out leaves states any authority over the models themselves, or only over the downstream apps built on top of them. My read: it preserves the right to regulate the product on the shelf and removes the right to ask what went into the factory. That's not a savings clause. That's a relabeling.

The development-versus-deployment line is the whole ballgame, and it's a seam the labs will pull at the moment it's law. If "development" is exempt from state rules and "deployment" isn't, the incentive is to define as much as possible as development — and to route training through structures that can claim the exemption. I'd bet on subsidiaries before the ink is dry.

What the draft really does, then, is regulatory arbitrage with a civic face. The frontier labs — OpenAI, Anthropic, Google, Meta, xAI, the five that clear the $500-million-revenue-and-a-frontier-model threshold — currently face a patchwork: California, Colorado, Texas, Illinois, each with its own development-stage demands. Replace that with one federal regime they helped shape, freeze the state layer where the real compliance friction lives, and leave states the messier, less commercially threatening deployment questions. Smaller firms below the threshold get the cleanest deal of all: the preemption benefit without the audit burden. Tidy.

The opposition understands this, which is why it's loud and unusually broad. Brad Carson of Americans for Responsible Innovation called it "a generational mistake" that "takes the current floor on state AI legislation and turns it into a federal ceiling" — the floor-to-ceiling line being the most precise four words anyone has said about the bill. Public Citizen, Public Knowledge, the AFL-CIO are against the freeze. The House's own bipartisan AI Commission, Ted Lieu among its co-chairs, said the draft "cannot serve as the basis for productive dialogue," which in congressional register is roughly a slammed door.

Now the caveat that the breathless coverage skips, and it matters: this is a discussion draft. No bill number, no markup, no CBO score, no legislative vehicle. "The first federal AI framework" implies a law. This is a 269-page conversation starter. The durable fact here isn't imminence — it's the strategy shift. After preemption died 99-1 in the open, it has reappeared in costume, and the costume is good enough that it split the civil-society and labor coalition's opponents from at least a few Democratic co-signers. Reps. Subramanyam and Peters signed on. Trahan herself has taken heat from her own side before for entertaining preemption. The bipartisan framing is real, and thin.

Where does this go? I think federal preemption is the central US AI-policy fight of the next eighteen months, and I think this draft is the most serious vehicle for it yet — precisely because it's the least naked. Watch two things. First, whether Ted Cruz's promised Senate revival merges with this House effort; the moratorium's chief mourner has a vehicle now if he wants one. Second, watch the development-deployment definitions in any actual bill text — that language, not the audit Title everyone's quoting, is where the power is allocated.

What would change my read? If a marked-up version genuinely narrowed "development" so that SB 53 and AB 2013 survived intact — if the savings clauses preserved the substance of what states enacted and not just the categories — then the oversight-for-preemption trade would be an honest one worth debating on its merits. I don't expect that, because the entire commercial value of the bill to its industry backers is the part that freezes those exact laws. But I've misjudged a markup before. The tell will be simple: if the labs' trade groups stay enthusiastic after the development definition gets tightened, I'm wrong about what they're buying. If they go cold, I'm right.