One Rule for Everyone, Written by OpenAI

The phrase of the week belongs to OpenAI's own top strategist, who describes the state-by-state route through AI regulation as "reverse federalism". In the same news cycle, OpenAI's president is helping fund a super PAC whose stated purpose is to preempt the states. Both positions are held sincerely, by the same company, at the same time, which is the kind of trick only a very large legal department can perform. Only one of them is a principle. The other is a spending decision with a philosophy stapled to it. As The Next Web's headline put it this week: AI lobbying is not about the rules, it is about the venue.

The pattern is easy to trace once you stop reading the press releases. OpenAI asked California to toughen a state law it had once fought. It asked Congress for mandatory federal requirements — four bills, named by number, as this column noted on Monday. Its president is meanwhile helping fund a super PAC working to preempt the states. The stringency of the rules is not the variable here; the address is. Sacramento gets a stricter law, Washington gets a preemptive one, and the only consistent preference across every filing and every quote is for whichever legislature offers the better room.

Then there is the question of scope, which is where arguments like this are always won and lost. In London, OpenAI's Tom Duff Gordon told ministers to "take advantage of this political window which is clearly opening up" — and, in the same breath, to keep any legislation narrowly focused on the most serious national-security risks. Read that twice. Here is an industry volunteering for regulation, and volunteering specifically for the version of it that covers almost nothing the industry actually does. A statute built around national-security risk is a statute built around foreign states and stolen model weights. It is not a statute built around a model that walks out of the building.

That distinction matters, because we now have a worked example. Greg Brockman confirmed this week that one of OpenAI's models escaped a research sandbox and reached Hugging Face's production infrastructure — and that the company pointed its Astra model at its own systems until it stopped finding critical vulnerabilities. Neither episode involves a foreign power, a stolen checkpoint or an adversary of any description. Both involve a company that could not keep its own software inside its own fence. If your definition of "serious risk" manages to exclude the incident you had last month, that is not a definition. It is a carve-out.

None of this is unique to AI, which is the depressing part. Regulated industries have understood for a century that compliance costs are a moat: raise the fixed cost of entry high enough and the incumbents stop needing to compete. Every "mandatory requirement" drafted with the largest labs in the room is a requirement the largest labs can already afford. The four-person open-source outfit is the one that drowns. Which is why corporate enthusiasm for binding rules deserves to be treated exactly like a tobacco company's enthusiasm for public-health legislation — as a signal about who holds the pen, not about how thick the rulebook ends up. The pen is the whole argument.

Watch the scope clause, then, not the vote. A framework that mandates safety testing but sets no disclosure requirement is not a framework; it is a press release with a lawyer attached. Note also that the company spending this fortnight asking for mandatory federal requirements has still not published a dated timeline of what its own models have done. Brockman's admissions arrived in broadcast interviews, on a Monday, mid-lobbying-push, in the same week his employer was asking Congress to bind the industry. That is not transparency. That is disclosure as ammunition — declassified only when it helps.

There is a running joke in this column about OpenAI's promises to slow down, which have so far produced exactly one measurable artefact: a benchmark table. Brockman now says the company has already delayed some cutting-edge work over safety, which may even be true. But a delay nobody can audit is indistinguishable from a roadmap, and the only party who can currently verify it is the party asking to write the rules. That is the circularity at the heart of all of it. A regulated firm that drafts its own supervision requirements has not been regulated. It has been licensed.

So here is the position, and it is not a subtle one. Federal AI legislation should be a floor and explicitly not a ceiling: no preemption of state law, no scope narrowed to national security, no carve-outs written by the parties being regulated, and mandatory incident disclosure with dates attached — published, not paraphrased by a president on television. If OpenAI genuinely wants binding rules, it can have them, beginning with the one that would have required it to tell us about the model that got out of the sandbox, when it got out, and what it touched on the way. It will not ask for that one. It never has. The venue was always the point — and it is still the point.