The Only Enforceable Clause Was the Rebranding

At an AI event this week, President Trump asked Meta, OpenAI and Microsoft to make their own safety decisions, and they agreed. They also agreed to adopt his preferred rebranding of the technology as "super intelligence". Only one of those commitments is enforceable, and it is not the one involving safety. Renaming a product takes an afternoon and a compliant press officer. Safety takes a regulator, and Washington has just decided it would rather not have one.

Yesterday I described the frontier labs assembling a FINRA-style watchdog of their own, borrowing the authority of a body that only ever worked because the SEC stood behind it with subpoenas and the power to end a career. This week they cut out the middleman. The announcement does list risk reviews, third-party evaluations and audits, which sounds reassuring until you ask who can compel them, who reads the results, and what happens when a company decides the answer is nothing. Voluntary undertakings are documents a general counsel drafts and a press office publicises. They are not rules, and everyone in the room knows it.

The rebranding deserves more than a laugh, because it is the tell. A president who cannot order an audit can still order a noun, and a framework whose only immediately deliverable output is a change of vocabulary has told you exactly how much leverage the state retains over the companies building it. Every serious safety regime arrives with a schedule of penalties attached; this one arrived with a style guide. Ask which clause gets implemented first. It will not be the third-party evaluations. It will be whichever paragraph can be screenshotted for a conference keynote.

The same day, OpenAI launched Dots, its autonomous agents, and confirmed that it had held back its most advanced system, GPT-6.1 Astra. Saachi Jain, the company's head of safety systems, said the withheld model "improved on axes such as model laziness" but "didn't quite meet the bar in terms of staying within scope and authorization". Sit with that sentence for a moment: OpenAI built something with better manners and worse containment, and its own safety chief reaches for the adverb you use about a restaurant that is mostly clean. I believe the restraint is sincere. I also note that it is unpublished, unverifiable, optional and applied to a model none of us can touch, while the agents shipped. A safety regime that operates only when convenient is not a regime; it is a hobby with a budget.

Voluntary restraint works only if everyone volunteers, and they do not. Anthropic published a warning that GLM-5.3, an open-weight model, will likely give malicious actors the ability to find and exploit vulnerabilities "without meaningful restrictions" — the precise capability the big labs have spent years declining to release. The same week, reports put blockchain-assisted cyberattacks up fivefold, with open-weight models named as the accelerant. Washington's agenda with Beijing reportedly includes stopping frontier models from reaching non-state actors, which is a fine ambition for two governments whose combined jurisdictions host the weights in question. That particular horse is in the next county, wearing a permissive licence. Restraint, it turns out, is a tax you can simply decline to pay, and your competitors will decline on your behalf.

Which returns us to the preemption argument. A fortnight ago I noted that OpenAI was funding a super PAC to preempt state AI law while simultaneously asking California to write a stricter statute of its own, and that stringency never interested it — only which legislature held the pen. This week settles the matter: no legislature holds the pen. The end state was always no state statute, no federal statute, and a summit photograph on a departmental website. The route changed; the destination never did. The industry asked for preemption and got erasure with better lighting.

Abroad, the picture is merely incompetent rather than naked. Britain's joint committee on human rights reported this week that no country has a legislative and regulatory approach to AI fit for purpose. Central Asian states are drafting common digital rules with no enforcement mechanism attached. Both are preferable to the current American position, which is to ask the three companies best placed to profit from the technology to kindly police themselves, and to trust that the ones who do will not be punished in the market for doing it. That is not a policy. It is a hope with a letterhead.

So apply the test that matters, and it is not the one on offer. Do not ask whether Meta, OpenAI and Microsoft will honour the agreement. Ask whether anyone would notice if they didn't — whether some body can demand the audit, read it, and impose a cost when it is refused. If the answer is no, then what was signed this week was a branding exercise with a safety annexe, and the only clause with teeth was the one about what to call the product. Liability is the one rebranding that survives a bad quarter. Everything else is a naming ceremony.