Leadership1 publisher2 min readPublished
OpenAI asks Washington to lead standards-writing for self-improving AI
The plan would link ten national safety institutes through a US agency and leave adoption to each government, with no licensing and no prerelease approval. About 20 countries signed a human-control declaration the same day.
The Board Room · Leadership desk
What happened
- OpenAI called on Monday for the United States to lead an international effort to write technical standards for frontier AI, including recursive self-improvement.
- The plan would connect ten national AI safety institutes through the U.S. Center for AI Standards and Innovation, covering frontier models as measured by capability benchmarks.
- The standards would not be licenses, mandatory prerelease review or approval requirements, and each national government would decide whether to adopt them.
- About 20 countries led by Finland and Norway adopted a declaration on keeping AI under human control the same day, and neither the United States nor China signed it.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- decision Each government in the network decides alone whether to put the thresholds into law.
- exposure A lab's internal research metrics would become the yardstick for measuring labs. Regulators using OpenAI's instruments would be measuring what OpenAI chose to count.
- constraint A developer selling into both the signatory countries and the United States may have to answer for the same model in two vocabularies, one built on human control and one on measurement.
The first measurements in the network would be OpenAI's own. The company offers as early contributions the metrics from its own research-acceleration report and its misalignment-reporting framework [10]. Those measurements would track how much autonomous research happens inside a company, identify the processes that should trigger immediate human review, and set shared incident severity levels and reporting thresholds [6]. The named institutes include Kenya, India, Japan and the U.K. [3], and the work would build on the International Network for Advanced AI Measurement, Evaluation, and Science, which CAISI created in November 2024 [4]. The process would consult open and closed model developers, independent technical experts and academics [7].
The company's line on self-improvement sits in the same post as its roadmap. "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely," OpenAI said [8]. That post names building an automated AI researcher as one of the company's three goals [9]. It also cites the Hugging Face incident as a "preview" of risks that could worsen with greater autonomy, while saying the incident was not a direct result of recursive self-improvement [11].
The timing is diplomatic. The call came during U.N. General Assembly week, and Sam Altman briefs the 15-member Security Council on Wednesday [12]. The administration whose agency would host the effort has rejected calls to slow AI development [13]. On Sept. 19, Trump wrote, "We will not in any way hinder or stifle the Growth of this incredible industry" [14]. Treasury Secretary Scott Bessent said, "It is humans who are responsible, not the AI" [15].
Two other groupings are forming around the same subject. China has assembled a separate international group and announced the World Artificial Intelligence Cooperation Organization [17]. At the World AI Conference in Shanghai this summer, Xi Jinping said, "AI development should not be a solo performance by a single country, but a symphony of international cooperation" [18]. The human-control bloc is about twice the size of the institute network OpenAI wants connected [20].
OpenAI put its own case in terms of who writes the framework: the company said American leadership would determine whether the United States "shapes the global AI framework or watches a fragmented, uneven, and conflict-ridden system take hold around it" [21]. I would expect any legislature that does act to start from the thresholds and severity levels drafted in venues like this one, years before anyone writes a licensing regime. OpenAI did not give a timetable, and says the concept will "evolve substantially" [19].
What to watch
- Whether CAISI or any of the ten institutes publishes an actual threshold document, and on what schedule.
- Whether the Finland and Norway declaration gains signatories from countries that host institutes in the proposed network.
- What Altman tells the Security Council on Wednesday, and whether any member state asks for prerelease review.