OpenAI pushes US-led global standards for frontier AI and RSI

OpenAI published a proposal calling for the United States to lead international technical standards for frontier AI — including common ways to measure increasingly autonomous AI research and keep humans in control as recursive self-improvement (RSI) risks grow.

GOVERNMENT desk — as labs automate more of their own R&D, OpenAI is asking Washington to set the ruler everyone else measures with.

What the post means by recursive self-improvement. Automated AI research can still have a person in the loop. As AI systems take on more of the work of building the next generation of AI, OpenAI says they can drive recursive self-improvement, shortened to RSI, even while people remain involved. RSI means each generation does a larger share of the research that produces the one after it. OpenAI says that as that work becomes more automatic, progress could speed up quickly. The company also says fully autonomous RSI is not happening today, and that it should not be pursued unless and until it can be done safely while people keep meaningful control. Whether to proceed, OpenAI says, has to depend on that control and on informed democratic choices about the benefits and the risks. Without that care, the post says, people could lose practical control of AI development and be unable to oversee research they no longer understand. File that limit as OpenAI’s. This card does not say autonomous RSI has started.

Why OpenAI wants a shared ruler. The post says international safety and security standards for frontier development may matter as much as alignment research for pacing the frontier. Alignment means keeping a system’s behavior in line with what people intend. Pacing, in OpenAI’s words, is not a preset speed limit. It means the research on alignment, and the use of that research, stay ahead of what the systems can do. OpenAI names three problems a common standard could ease, and says they apply to both open and closed models. Open means the model’s weights can be downloaded. Closed means the lab keeps them. Fragmentation: one country’s tests, reporting rules, and incident definitions could clash with another’s, so evidence is hard to compare. Collective action: each country acting alone can produce a result no country wants, because RSI could speed research past the ability to understand it and keep a human in charge. Uneven capacity: the labs and the expertise are not spread evenly around the world, which makes the first two problems worse. File those three as OpenAI’s.

What the standards would cover, and what they would not. OpenAI says they would be a common technical base for measuring capabilities, assessing risk, and judging whether safeguards are good enough. Safeguards are the limits meant to keep a system from doing harm. For automated research and RSI, the post adds three more items: a way to measure RSI-related progress and how much of a company’s research is being done by automated systems; a rule for which of those automated research processes should trigger an immediate human review; and a shared way to classify, track, report, and respond to alignment failures and automated-research incidents, including common severity levels and the threshold that decides when an event must be reported. Severity is how serious it is. A threshold is the line that trips a report. OpenAI calls its recent research-acceleration report and its misalignment reporting framework early contributions to those measurements. They are earlier documents. They are not this Sep 21 proposal, and they are not a government form.

Who would hold the pencil. OpenAI does not ask for a single international regulator that approves models before they ship. It says these technical standards would not be licenses, mandatory pre-release review, or approval requirements. A national government would decide whether, and how, to write them into its own law. The path OpenAI sketches runs through national AI safety institutes and, in the United States, CAISI — the Center for AI Standards and Innovation — plus industry bodies. The post names institutes or related bodies in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, and the United Kingdom. The focus it names is frontier models and the developers who build them, scored by capability benchmarks, and the benefits and risks of automated AI research, including RSI. A capability benchmark is a test that scores what a model can do. OpenAI says the United States can build on a network CAISI created in 2024, the International Network for Advanced AI Measurement, Evaluation, and Science. That 2024 line is OpenAI’s. This desk did not audit the membership list. The post also says the work should include open-model developers, closed-model developers, independent experts, and academics, so the standards do not favor one company, one country, or one business model. It names ISO, the Frontier Model Forum, the Agentic AI Foundation, the Open Secure AI Alliance, and the Appia Foundation. ISO is the International Organization for Standardization. Appia, OpenAI says, is writing practical specifications that connect international standards to real assessments. A name on that list is not a contract this desk saw, and it is not a statement that CAISI has already ordered these standards.

Two positions that stay OpenAI’s, not a government decision. The post says operators of critical infrastructure and governments should have secure channels to share national-security concerns, new weaknesses, and safety practices. It says a U.S.–China dialogue on these topics would be a positive step, and that upcoming talks are well timed. “Would be a positive step” is the company’s view. It is not a readout of those talks, and it is not a claim that either government has adopted this proposal. OpenAI also writes that leading now will determine whether the United States shapes the global AI framework or watches a fragmented, uneven, and conflict-ridden system take hold, and whether advanced AI strengthens the networks the United States and its allies already have. File that sentence as OpenAI’s.

Do not mix the dates. A different OpenAI page, “Helping build shared standards for advanced AI,” is dated June 23, 2026. That page is about the Appia Foundation and earlier standards work, and it points at OpenAI’s blueprint for democratic governance of frontier AI. It is not this Sep 21 proposal. The 9 Sep 2026 note “The AI policy window is open,” already filed as openai-ai-policy-window, is the California-bill endorsement and the earlier call for mandatory national safety rules. This 21 Sep post says the technical standards it is asking for would not themselves be licenses or a mandatory pre-release approval. A later statute, a treaty, or a CAISI mandate would be a different event. None of those is on this card.

Same-day wire, not a second primary. Reuters, in a dispatch datelined the United Nations on 21 Sep 2026 and carried by MarketScreener at 1:01 p.m. Eastern, reported that OpenAI called on Monday for the United States to lead global technical standards for frontier AI, including recursive self-improvement. The wire quotes the “Leading now” sentence and the pacing sentence from the blog. Michelle Nichols reported it; Nia Williams edited it. File those lines as Reuters’ account of the same post. This desk read the wire on MarketScreener. It did not retrieve a reuters.com permalink. Reuters-only scene-setting — a UN General Assembly week, a planned Security Council briefing, and the administration’s separate refusal to pause development — stays in Sources. It is not a claim that a council adopted these standards.

Plain English for the rest of the card: frontier AI = the most capable systems, at the edge of what labs can build. RSI / recursive self-improvement = each generation of AI does more of the work of building the next one. fully autonomous RSI = that loop with no meaningful human control — OpenAI says this is not happening today. alignment = keeping a system’s behavior in line with what people intend. pacing = not a preset speed; alignment research and its use stay ahead of capabilities. open model = the weights can be downloaded. closed model = the lab keeps them. CAISI = the U.S. Center for AI Standards and Innovation. AI safety institute = a national body that tests and studies advanced AI. capability benchmark = a test that scores what a model can do. safeguard = a limit meant to keep a system from doing harm. severity = how serious an incident is. reporting threshold = the line that decides the incident must be reported. ISO = the International Organization for Standardization. R&D = research and development, the work of building the next system. This filing is the 21 Sep proposal. It is not a law, a treaty, or a CAISI order.

PRIMARY here: OpenAI’s 21 Sep 2026 post “Building standards for the next phase of AI” — Tier A PRIMARY, the company’s own page, the originating record. Reuters, carried the same afternoon by MarketScreener, is same-day corroboration of the Monday call and of two sentences from that page, not a substitute primary. The June 23, 2026 page “Helping build shared standards for advanced AI” is an earlier Appia post. It is cited only so the dates stay apart. The U.S.-led standards ask, the RSI definition, the not-happening-today line, the human-control condition, the three problems, the open-and-closed line, the capability / risk / safeguard list, the human-review trigger, the incident severity and reporting thresholds, the no-license and no-approval line, the institute list, the CAISI path, the 2024 network sentence, the named standards bodies, the Appia specifications line, the secure-channel line, the U.S.–China dialogue as a positive step, and the “Leading now” sentence are the Sep 21 post’s. The Monday wire, the byline, and the two matching quotes are Reuters’, via MarketScreener. NOT claimed: that any government has adopted the standards, a treaty, treaty text, that CAISI has mandated these standards, a model-approval regulator, that fully autonomous RSI is underway, that this desk joined the talks or audited the 2024 network, a stock tip, or investment advice. Distinct from the already-filed trump-ai-force-czar, un-ai-panel-agents, bessent-china-ai-dialogue, faa-smart-ai, and openai-ai-policy-window.