OpenAI published a proposal on 21 September asking the United States to lead an international effort to write technical standards for frontier AI, with recursive self improvement, AI systems doing more of the work of building their successors, as the central object.3 The Next Web 2026-09-23 Published 21 Sep; Hugging Face incident cited but not RSI; misalignment reporting framework offered; US China dialogue welcomed; Bessent: no liability shield; Security Council briefing 23 Sep. Open source The company states that fully autonomous self improvement is not happening today and should not be pursued until it can be done safely.1 Techzine 2026-09-22 US led standards for autonomous AI research; full RSI not pursued until safe; open and closed models; no licensing; builds on safety institute network, ISO, Frontier Model Forum; administration rejecting slowdown calls. Open source The standards would measure how much research is done autonomously inside AI companies, define which automated processes should trigger immediate human review, and set shared incident severity levels; they would not be licenses, mandatory prerelease reviews or approval requirements.2 Implicator.ai 2026-09-22 Ten institutes via CAISI; autonomy measurement, human review triggers, incident severity levels; no licenses or prerelease review; no timetable; automated researcher remains an OpenAI goal; 20 country declaration without US or China. Open source We assess with high confidence that this is a bid to define the measurement regime before others define a control regime, and with moderate confidence that it will shape the vocabulary of the coming debate more than its outcome.

What is actually being proposed

The architecture is light. Ten national AI safety institutes, among them those of Kenya, India, Japan and the United Kingdom, would be connected through the US Center for AI Standards and Innovation and national industry bodies, building on the measurement network CAISI created in November 2024.2 Implicator.ai 2026-09-22 Ten institutes via CAISI; autonomy measurement, human review triggers, incident severity levels; no licenses or prerelease review; no timetable; automated researcher remains an OpenAI goal; 20 country declaration without US or China. Open source OpenAI suggests roles for ISO, the Frontier Model Forum and the Open Secure AI Alliance, and says the standards should cover open and closed models alike.1 Techzine 2026-09-22 US led standards for autonomous AI research; full RSI not pursued until safe; open and closed models; no licensing; builds on safety institute network, ISO, Frontier Model Forum; administration rejecting slowdown calls. Open source It offers its own misalignment reporting framework as an early contribution, recommends secure channels for governments and infrastructure operators to share threats, and says US and China dialogue on these issues would be positive.3 The Next Web 2026-09-23 Published 21 Sep; Hugging Face incident cited but not RSI; misalignment reporting framework offered; US China dialogue welcomed; Bessent: no liability shield; Security Council briefing 23 Sep. Open source No timetable is given, and OpenAI says the concept will change substantially.2 Implicator.ai 2026-09-22 Ten institutes via CAISI; autonomy measurement, human review triggers, incident severity levels; no licenses or prerelease review; no timetable; automated researcher remains an OpenAI goal; 20 country declaration without US or China. Open source

The rationale is the Hugging Face incident, which OpenAI cites as a preview of the risks autonomy brings while saying it was not a direct result of self improvement.3 The Next Web 2026-09-23 Published 21 Sep; Hugging Face incident cited but not RSI; misalignment reporting framework offered; US China dialogue welcomed; Bessent: no liability shield; Security Council briefing 23 Sep. Open source The warning is that research done without care could leave humans unable to oversee processes they no longer understand.1 Techzine 2026-09-22 US led standards for autonomous AI research; full RSI not pursued until safe; open and closed models; no licensing; builds on safety institute network, ISO, Frontier Model Forum; administration rejecting slowdown calls. Open source

The tension inside the document

The same post that calls for caution on self improvement lists building an automated AI researcher as one of OpenAI’s three goals.2 Implicator.ai 2026-09-22 Ten institutes via CAISI; autonomy measurement, human review triggers, incident severity levels; no licenses or prerelease review; no timetable; automated researcher remains an OpenAI goal; 20 country declaration without US or China. Open source That is not a contradiction so much as a position: OpenAI wants to keep building toward automation while establishing that the milestone worth regulating is full autonomy, a line it says no one has crossed. Standards that measure progress toward that line, administered by institutes rather than regulators and adopted at each government’s discretion, are standards a frontier developer can live with.

The timing is pointed. The proposal arrived during UN General Assembly week, the day about 20 countries led by Finland and Norway adopted a declaration calling for AI to remain under human control, which the United States and China did not sign.2 Implicator.ai 2026-09-22 Ten institutes via CAISI; autonomy measurement, human review triggers, incident severity levels; no licenses or prerelease review; no timetable; automated researcher remains an OpenAI goal; 20 country declaration without US or China. Open source It arrived two days before Sam Altman was due to brief the Security Council alongside Dario Amodei, Yoshua Bengio and Clement Delangue.3 The Next Web 2026-09-23 Published 21 Sep; Hugging Face incident cited but not RSI; misalignment reporting framework offered; US China dialogue welcomed; Bessent: no liability shield; Security Council briefing 23 Sep. Open source And it arrived in an administration that rejects calls to slow AI development, in part over concern that China would benefit.1 Techzine 2026-09-22 US led standards for autonomous AI research; full RSI not pursued until safe; open and closed models; no licensing; builds on safety institute network, ISO, Frontier Model Forum; administration rejecting slowdown calls. Open source A voluntary, US led, measurement first proposal is the version of international governance that fits that administration.

The liability backdrop sharpens the incentive. The same week, Treasury Secretary Scott Bessent said the administration would not shield frontier labs from liability.3 The Next Web 2026-09-23 Published 21 Sep; Hugging Face incident cited but not RSI; misalignment reporting framework offered; US China dialogue welcomed; Bessent: no liability shield; Security Council briefing 23 Sep. Open source A lab that expects to answer for incidents has a direct interest in shared severity levels and reporting thresholds that it helped write, because those definitions will frame what counts as negligence when the next incident arrives.2 Implicator.ai 2026-09-22 Ten institutes via CAISI; autonomy measurement, human review triggers, incident severity levels; no licenses or prerelease review; no timetable; automated researcher remains an OpenAI goal; 20 country declaration without US or China. Open source

Who gains and who loses

OpenAI gains first mover position on definitions: whoever specifies how autonomy is measured influences where the thresholds sit. CAISI gains a coordinating role and a reason to exist internationally. Safety institutes abroad gain a seat in a structure the US would lead. Advocates of binding controls lose ground, because a credible voluntary offer from the largest lab makes mandatory testing harder to argue for. Open weight developers face mixed outcomes: the standards would cover them, but without licensing they would not be excluded.1 Techzine 2026-09-22 US led standards for autonomous AI research; full RSI not pursued until safe; open and closed models; no licensing; builds on safety institute network, ISO, Frontier Model Forum; administration rejecting slowdown calls. Open source

The counter case

The best argument against our reading is that measurement is where every serious regime starts, and that OpenAI is proposing the hard part. No government today can say how much of a lab’s research is automated. Shared incident severity levels would be the first common language for AI failures across borders. If the institutes adopt the measures and publish the numbers, the voluntary label matters less than the transparency, and a later binding regime would be built on them.

What to watch

  • The UN moves before Washington does. A Security Council session on loss of human control looks likely within days, probably convened by France. The Nordic Call for Control should pick up European and African endorsements there.
  • No US adoption. We expect no formal US government embrace of OpenAI's text this autumn.

What happened next (updated Oct 13, 2026)

This section was added after publication. The Security Council met on 23 September in a session convened by France and chaired by Jean-Noel Barrot, the first time the Council took up loss of human control over AI as a standalone subject.4 Unite.AI 2026-09-23 Council meeting 23 Sep, first standalone session on loss of control; Call for Control asks for mandatory pre deployment testing and shared incident reporting; France and Sierra Leone endorsed 23 Sep. Open source The Call for Control of Frontier AI Models launched by Finland and Norway on 21 September, which asks for mandatory pre deployment testing, common standards and shared incident reporting, gained endorsements from France and Sierra Leone on 23 September.4 Unite.AI 2026-09-23 Council meeting 23 Sep, first standalone session on loss of control; Call for Control asks for mandatory pre deployment testing and shared incident reporting; France and Sierra Leone endorsed 23 Sep. Open source We found no formal US government adoption of OpenAI’s proposal through 13 October.