No one can verify what a frontier lab says about its own safety

Europe now enforces a law the United States never wrote

On 2 August 2026 the European Union's AI Office gained the power to enforce the AI Act's rules for general-purpose AI.³ It can now demand a developer's documentation, run its own evaluations of a model, and order that model off the European market. It can fine the company up to 3% of its worldwide annual turnover, or fifteen million euros where that figure is larger.⁴ The duties themselves had been applicable since 2 August 2025. What changed this month is the enforcement.³

The United States has no such law. The only enforceable domestic regime for frontier models is California's SB 53, the Transparency in Frontier Artificial Intelligence Act, in force since 1 January 2026. It reaches any developer that trains a model above 10²⁶ computing operations and binds those whose affiliates clear five hundred million dollars in yearly revenue.⁵ One state has a disclosure statute. The federal government has nothing. That gap, not the European enforcement date, is the thing Washington now has to answer.

The answer is not to copy Brussels. It is to take the part of SB 53 that travels, its disclosure core, and set it as a national floor.

What a disclosure floor asks of a frontier lab

For the developers it covers, SB 53 is a disclosure statute and little else. A large developer must publish a frontier AI framework setting out how it assesses and mitigates the risk that one of its models materially contributes to a catastrophe. It must publish a transparency report before it deploys a covered model. It must report a critical safety incident to the state within fifteen days, and within twenty-four hours where the incident carries an imminent risk of death or serious physical injury. It must protect employees who raise safety concerns. Break a framework commitment and the Attorney General can seek a civil penalty of up to one million dollars per violation.⁵ The statute even fixes what counts as the catastrophe worth reporting: a single incident that kills or seriously injures more than fifty people, or does more than one billion dollars in property damage.⁵ Anthropic, describing the duties it endorsed, lists the same four acts: publish the framework, publish the assessment, report the incident, protect the whistleblower.²

None of that is the European regime, and a federal floor should not reach for it. The EU AI Act governs the whole supply chain, providers and deployers alike, where SB 53 touches only the frontier developer.⁶ It presumes systemic risk once training compute passes 10²⁵ operations, an order of magnitude below SB 53's line.⁷ It demands model evaluation and documented adversarial testing.⁸ And it asks for one thing SB 53 never does: a conformity assessment before the model reaches the market.⁶ That last step is licensing by another name. A disclosure floor requires a lab to say what it does and how it checks itself. It does not require the government's sign-off to ship, and it should not.

The floor closes a gap that good faith cannot

Grant the labs everything. Assume every developer now signing the European Code of Practice means it, tests its models honestly, and writes down what it finds. The case for a federal floor does not weaken, because the problem it solves is not dishonesty. The problem is that a safety judgment kept inside one company cannot be checked or compared by anyone who has to act on it.

A regulator cannot audit a framework it has never seen. A competitor cannot match a safety practice it cannot read. An engineer choosing between two frontier models cannot weigh two catastrophic-risk assessments written to no common standard and released on no schedule. The developer holds information that no amount of its own good faith transfers to the people downstream of its decisions. Good faith produces a sound private judgment. It does not produce a public, comparable record. Only disclosure does that, and only structured disclosure, published on a fixed trigger and open to a third party's reading, produces a record two labs can be held to on the same terms.⁵

This is why the floor serves the honest developer as much as it constrains the careless one. A lab that does the work now has a way to prove it did, in a form a regulator and a customer can check against its rivals. The asymmetry is the target, and the asymmetry is indifferent to intent.

Licensing aims at a different target and misses. To approve a model before release, a regulator must out-reason the frontier about what the model will do, on the company's evidence, before anyone outside has seen it. A disclosure floor asks only that the evidence be on the table. The first design needs the government smarter than the developers it oversees. The second needs the developers to show their work. Between a rule that must predict the frontier and a rule that must only make it legible, the legible one is the rule a serious country can actually run.

Little Tech's case that a floor entrenches the incumbents

The strongest case against a federal floor is not made by people who want frontier labs unwatched. It is made by people who want the watching not to crush everyone smaller than an incumbent. Matt Perault, who runs AI policy at Andreessen Horowitz, argues that "compliance with a 50-state patchwork of state transparency obligations would present massive challenges for companies with small engineering, law, and policy teams," and that the mandates' "long-term effect will be to increase concentration in AI markets."⁹ His colleague Martin Casado argues the prior question is unsettled. He calls the marginal risks of frontier AI over existing software "research questions and hypotheticals," and says policy should "only depart from the existing regulatory regime ... once we understand the marginal risks of AI relative to existing computer systems."¹⁰ Meta's Joel Kaplan, refusing to sign the European Code, calls it a set of "measures which go far beyond the scope of the AI Act" and says "Europe is heading down the wrong path."¹¹

On the patchwork, the critics are right. The burden they describe is also the case for a national floor: a single federal standard is the one thing that dissolves a fifty-state patchwork instead of adding a fifty-first tile to it. Perault's own strongest objection, the "massive challenges" of complying with fifty regimes at once, is the objection a preemptive federal law answers and state-by-state drift does not.⁹ Concentration meets the same answer from the other side. SB 53's duties attach only above five hundred million dollars in revenue and 10²⁶ operations, so a startup training beneath those lines owes nothing and cannot be entrenched against by a rule it never triggers.⁵ Casado's point is the sharpest, and a disclosure floor is built to grant it: a floor bans nothing and approves nothing. It produces the very evidence he says is missing, the record of what frontier developers test and find, so a later and heavier rule can "depart from the existing regime" on data rather than angst.¹⁰

Two facts cut against the floor. The first is that a paper mandate with no enforcement was just ignored. The EU already requires developers to publish summaries of their training data. In the window before the AI Office could act, the largest developers did not comply: leading labs "have failed" to publish usable summaries, while a smaller open-source developer did the work.¹² That is the best evidence available on whether disclosure alone changes behavior, and it says: not without teeth. It is an argument for enforcement, not against a floor. The failure it describes is a floor no one was enforcing, which is exactly why the national version needs the Attorney General's penalty and a real trigger, not a request for voluntary candor. The second fact is that the floor's own best advocate has outgrown it. Dario Amodei, who backed SB 53, wrote in June 2026 that "it is time to go beyond transparency to more serious and binding regulation of AI."¹³ Take him at his word. A floor is a minimum, not a ceiling. Setting a national disclosure baseline forecloses none of the third-party testing Amodei wants next; it builds the shared record any such regime would stand on. That the strongest advocate wants to go further is a reason to lay the first course, not to skip it.

What would change this paper's judgment

Stripped to what the evidence carries, the claim is narrow. Two regimes are live, no federal law exists, and the United States should nationalize SB 53's disclosure core and stop at its edge. The argument for doing so rests on the asymmetry a floor repairs, not on any showing that disclosure makes a model safer. That showing does not exist in this record. No source here demonstrates that a transparency floor reduces catastrophic risk, and the honesty of the case depends on saying so.

If a national floor, fully enforced, produced only better-formatted safety reports and no better public knowledge of what frontier systems can do, the case would fail. The warning is already on the page: the EU's training-data duty, ignored until a regulator could act, is what a floor without teeth looks like.¹² Evidence that mandated, enforced disclosure still changes nothing a developer does would collapse this argument into the case for the licensing regime it was built to avoid. Until that evidence arrives, the reverse holds. A country with one state law and no federal one should set the floor, enforce it, and leave the ceiling for the argument Amodei has already started.¹³