Confirmed in a primary source 21 Sept 2026 Governance

OpenAI published its design for global AI standards, and the design has no licence and no approval step in it.

On September 21 OpenAI published a document called Building standards for the next phase of AI, proposing that the United States lead the writing of technical standards for frontier models and for AI that does AI research, through the existing safety institutes of ten countries plus industry bodies. The document says in terms that these standards would not be licences, mandatory prerelease review, or approval.

The caveat that matters is that this is a proposal about other people's future rules: it binds OpenAI to nothing today and creates no obligation on any government to take it up.

Until now

OpenAI's public position was that it could live with outside inspectors in statute, which read as the industry conceding the principle of mandatory oversight.

Now

Read the fight as being over the instrument rather than the principle, and watch whether pacing arrives as a standard you adopt or a permission you need.

Read the two documents next to each other and the move is easy to see. On September 15 OpenAI's frontier policy submission told Congress it could accept independent verification organisations written into federal law, and asked the British government to legislate. That was read at the time, here included, as a lab conceding that oversight should be compulsory. On September 21 the company published the architecture it actually wants: ten national safety institutes, a coordinating American standards centre, ISO, and four industry consortia, writing measurements for recursive self-improvement and a common incident reporting format. The architecture is detailed and much of it is useful. What is not in it is any point at which somebody who is not OpenAI can say no.

That is not a hidden implication, it is a printed sentence: these technical standards would not be licenses, mandatory prerelease review, or approval. And the document's own claim for itself is large, that standards may matter as much to pacing the frontier as alignment research does. Both things can be true. A shared way to measure how fast a model is improving itself is genuinely worth having, and the labs are the only people who currently have the data to build it. But a measurement regime with no gate is a speedometer, not a brake, and the company publishing the design is the one being measured. Read the fight as being over the instrument rather than the principle, and watch whether pacing arrives as a standard you adopt or a permission you need.

How it got here

Sep 12
Anthropic and OpenAI say they will give outside inspectors desks, badges and company laptops. Carded in Edition 020.
Sep 15
OpenAI tells Congress it could accept independent verification organisations written into federal law and asks Britain to legislate. Carded in Edition 023, the parent of this card.
Sep 18
Anthropic names its own embedded evaluator and picks a consulting firm rather than a safety institute. Carded in Edition 026.
Sep 21
OpenAI publishes its standards design: ten safety institutes, ISO and four industry bodies, with licensing, mandatory prerelease review and approval expressly excluded.
Who has to react
The ten named safety institutes, which have just been handed a job description written by a company they test; the Center for AI Standards and Innovation, named as coordinator on the day the White House is calling AI safety concern a hoax; the European Commission, whose AI Act already contains the approval step this document rules out; Anthropic and Google, who now have to say whether they want the same architecture; the four industry consortia named, whose membership lists are about to matter; and the members of Congress who read the September 15 submission as a concession.

The evidence

ObservedMonday September 21, 2026. OpenAI published Building standards for the next phase of AI on its own site.

The whole observation, as recorded

What it proposes, in two parts. Part one, a mechanism for complementary national and international standards for frontier models and automated AI research, built on the AI safety institutes already operating in Australia, Canada, France, Germany, India, Japan, Kenya, Korea, Singapore and the United Kingdom, coordinated through the United States Center for AI Standards and Innovation and national industry bodies. Part two, common measurements and incident reporting: evaluation of progress in recursive self-improvement and of levels of autonomous research, human oversight standards for automated AI research that trigger immediate review, and a single classification and reporting scheme for incidents. Bodies named in the document: the Center for AI Standards and Innovation, the International Network for Advanced AI Measurement, Evaluation and Science, ISO, the Frontier Model Forum, the Agentic AI Foundation, the Open Secure AI Alliance and the Appia Foundation. The exclusion, in OpenAI's own words: These technical standards would not be licenses, mandatory prerelease review, or approval. The sentence the document rests on: International standards for safety and security practices in frontier AI development may be as important to pacing the frontier as alignment research itself. The context that makes it a development rather than a restatement: on September 15, carded in Edition 023, OpenAI told Congress it could accept independent verification organisations written into federal law and asked Britain to legislate. Six days later the published design routes the same function through voluntary standards bodies and rules out the approval step.

Established

OpenAI published a standards proposal on September 21, 2026 naming the safety institutes of ten countries, the Center for AI Standards and Innovation, ISO and four industry bodies.
The document proposes common measurements for recursive self-improvement and autonomous research, human oversight standards triggering immediate review, and a unified incident classification and reporting scheme.
The document states that these technical standards would not be licenses, mandatory prerelease review, or approval.
It states that international standards for safety and security practices may be as important to pacing the frontier as alignment research itself.

Not established

That OpenAI opposes binding oversight. Its September 15 submission said it could accept verification organisations in federal law, and the two positions are not strictly contradictory.
That any government will adopt this. It is a proposal to institutions that did not write it and owe OpenAI nothing.
That the measurements are wrong or unserious. A shared metric for recursive self-improvement does not exist today and would be worth having whoever proposes it.
That this changes any obligation now in force. The EU AI Act, the Digital Services Act designation and California's audit law all still apply unchanged.

Why it weighs this much

Signal strength

CostLow

CommitmentMedium

ReversibilityModerate

PersistenceYears

Cost is low. Publishing a policy document costs staff time and nothing else, and this one asks other institutions to do the work. Commitment is medium rather than low because the exclusion is explicit and in writing: having published that standards should not be licences or approvals, OpenAI cannot quietly support a licensing bill without the document being held up. Reversibility is moderate; positions can be revised and this one is not a contract. Persistence is years, because standards architectures, once national bodies start building on them, outlive the companies that proposed them. We score a published position honestly: this is words, but words that foreclose something, which is why it is on the deck rather than in Undercurrents.

What to watch

strengthens Two or more of the named safety institutes, or CAISI, adopt or formally consult on a frontier standards workplan matching this architecture with no approval or prerelease step, by June 30, 2027.
weakens OpenAI publicly supports a specific bill or instrument containing a licensing, approval or mandatory prerelease review requirement for frontier models, by June 30, 2027.

The words

Recursive self-improvement: AI systems improving AI systems, including their own successors, with each round faster than the last. It is the specific capability this proposal wants measured.
Technical standard: An agreed way of measuring or doing something, written by a standards body. It has force only where a government or a contract requires it.
Prerelease review: A rule that an outside body must examine a model before it can be released. OpenAI's document rules this out of what it is proposing.
AI safety institute: A government-funded body that tests frontier models. Ten of them are named here, and they currently test by invitation rather than by right.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.