OpenAI has published its position on standards for the next phase of artificial intelligence, and it is worth reading closely, because the argument is better built than the thing it is asking for.
The document sets out three stated goals. Build an automated AI researcher and work with it iteratively on alignment, keeping people inside the loop of self improvement. Make the benefits of the resulting scientific and economic progress available. And give every person their own personal AGI.
That third item sits in the same list as the alignment problem and human control of self improving systems, as though it belonged to the same category of thought. It does not. It is a product roadmap. Filing a distribution target alongside the question of whether the technology stays under human control is not an accident of drafting.
What the standards are not
The core ask is that the United States lead a joint international effort to develop global technical standards for frontier AI, including recursive self improvement. The document proposes building on the emerging network of AI safety institutes already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom, working through CAISI and national bodies, and cooperating with ISO, the Frontier Model Forum, the Agentic AI Foundation, the Open Secure AI Alliance and implementation focused organisations such as the Appia Foundation.
Then comes the sentence that governs the whole proposal. These technical standards would not constitute licences, mandatory pre release reviews, or approval requirements for AI models. National governments would decide whether and how to write them into their own law, and the document is explicit that any lab doing this work remains responsible for operating safely under existing law.
What is left inside the standards: measurement, benchmarks, shared definitions of high quality evidence, common severity levels for incidents. All of it useful. None of it able to stop anything.
Nothing here forbids a state from regulating, and that is worth stating clearly. The objection is not that the document bans anything. The objection is what it declines to build.
The analogy does the arguing
The document reaches for aviation and financial stability as precedents, sectors where countries built common technical standards and reliable channels of cooperation without surrendering national authority. The comparison is flattering and it is also the weakest part of the text.
In civil aviation, an aircraft type does not enter service without prior certification by a regulator with the power to refuse. In banking, capital requirements are binding, supervised and enforced, and the whole architecture exists because the sector demonstrated what happens when it is not. Both regimes work precisely because of the element this proposal rules out in advance.
You cannot borrow the prestige of a licensing regime while excluding the licence. Or rather you can, and it is being done here in a single paragraph, but it should be named.
Who signs off on “not yet”
On recursive self improvement the document is more candid than most. It states that fully autonomous RSI does not exist today and should not be pursued until it can be done safely. It warns that if it were carried out without due care, humans could lose effective control over AI development and find themselves unable to supervise research processes they no longer understand. It cites the Hugging Face incident the company disclosed as a preview of risks that could become far more serious without robust safeguards.
That is a serious paragraph. It is also unfinished, because it never says who decides when “safely” has been reached. It says the decision should depend on the ability to preserve human control and on informed democratic choices. Democratic choices require an institution capable of choosing otherwise, and the power to refuse a release is precisely what these standards are drafted not to carry. It is left to whatever each national legislature may or may not enact, later, separately, in as many versions as there are parliaments. That is not the same as prohibiting regulation. It is also not the same as building it.
What remains is a company stating that it will not do the dangerous thing until it judges the dangerous thing to be safe. That is not a standard. That is an internal policy with an international accent.
The problems described are real
None of this makes the diagnosis wrong. Fragmentation is real: conflicting evaluations, reporting obligations and incident definitions across countries do make it harder to compare evidence and respond to cross border risk. Collective action failure is real: if every country moves independently, everyone gets an outcome nobody wanted. Uneven capability is real, and it worsens the other two.
Shared reporting protocols for alignment incidents, common severity thresholds, measurement of how much autonomous research is happening inside a given company: all of that is worth building, and the document is right that it should cover open and closed models alike. Secure channels between critical infrastructure operators and governments are worth building. Dialogue between the United States and China on these questions is worth having.
The objection is not to the content. The objection is to the shape.
The part that is honest
The document also says something unusually direct about its own motivation: that support for standards comes from a desire to avoid the concentration of power, and that standards let more actors outside the labs contribute to decisions about how the technology develops, offering a visible set of principles independent of any one lab’s practices.
That is exactly right, and it is the argument against the version being proposed. Outside actors contribute to decisions when they can affect outcomes. A shared measuring instrument that cannot be used to withhold anything does not distribute power. It distributes vocabulary.
Governing the pace of AI development, the document says, does not mean maintaining a predetermined speed. Agreed. It means somebody other than the driver has a say in the speed, and this proposal is carefully drafted so that the standards themselves never say it. Somebody else can, in theory, one country at a time. The instrument being offered here is not that somebody.
Raffaele Di Marzio
All my “insane” books on cybersecurity and governance are here 👇 https://cyberium.limited/bookshelf.html