In a policy paper published on 21 September, OpenAI proposed that the United States lead an international effort to build technical standards for frontier AI systems. At the centre of the proposal is the prospect of models running research increasingly on their own, and recursive self-improvement (RSI).

The clearest line in the paper: fully autonomous self-improvement is not happening today, and should not be pursued unless and until it can be done safely.

What the proposal is, and is not

Some coverage described the announcement as a call for mandatory safety rules. Two separate statements need to be kept apart:

  • 9 September: OpenAI asked for mandatory, capability-based national safety requirements in the United States, aimed only at the handful of well-resourced labs building the most capable systems.
  • 21 September: This newer paper proposes international technical standards. As CNBC reports, they would not be licences, mandatory pre-release review or approval requirements; each government would decide whether to write them into its own law.

The first is binding rules; the second is a shared measurement language. They are not mutually exclusive, but they are not the same thing either.

What the standards would cover

The proposed framework aims at a common language in four areas:

  • How model capabilities are measured.
  • How autonomous research inside a company is tracked.
  • Which safeguards are applied.
  • At what severity threshold, and how, incidents are reported.

For coordination it points to an existing base: the worldwide network of AI safety institutes, including bodies in Australia, Canada, Germany, France, Kenya, Japan, South Korea, Singapore, India and the UK. OpenAI wants to build on these rather than create a new institution.

The timing

The paper landed during the week of the UN General Assembly. As Implicator reports, Sam Altman was scheduled to brief the UN Security Council. A day earlier, the UN scientific panel on AI had published its first report, writing that there is no assurance humans will keep control over AI agents.

The criticism comes from inside the same document. Implicator points to the contradiction: OpenAI says fully autonomous self-improvement is not happening today, while listing an "automated AI researcher" among its own goals. When the company proposing the standard is also building the capability the standard is about, questions about motive follow.

Why it matters to readers

The debate can look abstract, but it has a practical edge. Without a common measurement language, every company publishes results from tests it chose itself, and the numbers cannot be compared. Two major model launches this week, each arriving with its own chart, are the concrete example. Alongside safety, what is at stake in the standards debate is comparability.