Summary
On September 21, 2026, as world leaders gathered in New York for the UN General Assembly, OpenAI published "Building standards for the next phase of AI," a policy post arguing that "the United States should lead an effort to work together with countries around the world to develop global technical standards for frontier AI, including for RSI," or recursive self-improvement. The post proposed routing standard-setting through the US Center for AI Standards and Innovation (CAISI) and the network of national AI safety institutes, and named three areas for common standards: measuring how much autonomous research happens inside an AI company, defining which automated research processes should trigger immediate human review, and shared incident severity levels and reporting thresholds. It stated that such standards "would not be licenses, mandatory prerelease review, or approval requirements for AI models," and that national governments would decide whether to adopt them.
What Happened
The post, filed under OpenAI's Global Affairs section, opened by restating three company goals "as outlined recently by Sam Altman and Jakub Pachocki," the first of which is to "navigate the next period of AI progress, by building an automated AI researcher, iterating with it on the alignment problem, and finding ways for people to remain part of the self-improvement loop." It then defined the risk it was addressing. As AI systems take on more of the work of developing their successors, the post said, they "can increasingly drive a process of recursive self-improvement (RSI), even while people remain involved," and "the pace of AI progress could accelerate rapidly." OpenAI wrote that "fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely," and that without care it "could result in humans losing practical control over AI development, unable to provide oversight on research processes they no longer understand." It said the Hugging Face incident it had disclosed, "while not a direct result of RSI, is a preview of the kinds of risks that could become much more severe without robust safeguards and alignment."
The case for international standards rested on three named problems. Under "Fragmentation," the post said national evaluations, reporting requirements and incident definitions could conflict; under "Collective action," that "each nation acting independently can produce outcomes that no nation wants"; and under "Uneven capacity," that frontier development and expertise are concentrated in a few places. OpenAI said these problems "apply to both open and closed models," that its rationale for standards "is rooted in avoiding the concentration of power," and that standards "may be as important to pacing the frontier as alignment research itself."
The first of two proposed components was an institutional mechanism. OpenAI suggested using the AI safety institutes already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom "to facilitate standard setting through the CAISI and national industry bodies," focused on frontier models and developers "as measured by capability benchmarks" and on "benefit-risk management for automated AI research, including RSI," building on the international measurement and evaluation network that CAISI already convenes. The standards would cover capability measurement, risk assessment and "safeguard sufficiency," would be developed in consultation with open-weight and closed-model developers, independent experts and academia, and should not make it "harder for new entrants or open-weight developers to compete." The post cited aviation and financial stability as models of common technical standards adopted without countries "giving up national authority," and named ISO, the Frontier Model Forum, the Agentic AI Foundation, the Open Secure AI Alliance and the Appia Foundation as bodies to work with.
The second component was a set of common measurements and incident protocols: standards for evaluating "RSI-relevant AI progress and the amount of autonomous research happening within an AI company"; for "human oversight over automated AI research, including what kinds of automated AI research processes should trigger immediate human review"; and for incident classification, tracking, reporting and response, "such as common incident severity levels and reporting thresholds." OpenAI presented its own September research-acceleration report and its misalignment reporting framework as early contributions. Separately, it called for governments and critical infrastructure operators to set up secure channels for sharing national security concerns and vulnerabilities, adding that "dialogue between the United States and China in these areas would be a positive step, and upcoming talks are happening at an opportune time." The post described pacing as "not about maintaining a predetermined speed" but about ensuring "that alignment research and deployment of that research stay ahead of capabilities," and closed: "Leading now will determine whether the United States shapes the global AI framework or watches a fragmented, uneven, and conflict-ridden system take hold around it."
Semafor reported the same evening that OpenAI had called for a US-led international coalition "as discussion of the technology's risk and promise looms over this week's UN General Assembly," noting that Washington was pursuing tentative outreach with Beijing on AI, including a possible deconfliction hotline, and that President Trump had been reluctant to shape domestic AI safety standards. Bloomberg and Axios also reported the post on September 21, Axios under a headline tying it to the US-China talks.
Why It Matters
The post came fifteen days after OpenAI chief scientist Jakub Pachocki called for international coordination and "widely mandated safety bars for continued development," and it turns that call into a specific proposal aimed at automated AI research and recursive self-improvement. It is also narrower than Pachocki's essay: it names a venue (CAISI and the safety-institute network) and three measurable subjects, but it explicitly rules out licensing and prerelease approval and leaves any binding force to national governments. It also places the question of how much autonomous research a lab is doing, which Anthropic had begun reporting with its own index days earlier, inside a proposed shared measurement standard.
What the post does not settle is whether anyone with authority will take it up. The post names no government sponsor, timeline or first deliverable, and, as Semafor noted, President Trump had been reluctant to shape domestic AI safety standards. Whether Beijing would participate in a process explicitly framed around US leadership is also unaddressed; the post's reference to US-China dialogue concerns secure communication channels, not standard-setting. The categories it proposes, such as incident severity levels and triggers for human review, are defined only in outline, and the post says it expects "this concept to evolve substantially over time."
§ How to read the metadata
- Landmark
- Fundamentally alters the trajectory; 2–5 per year.
- Major
- Meaningfully shifts the landscape; 2–4 per month.
- Notable
- Worth documenting; significance can be upgraded later.
- Confidence
- High = primary sources corroborate. Medium = credible secondary only. Low = provisional. Disputed = credible sources disagree.
- Contestation
- Uncontested = no formal challenge. Contested = at least one challenge open. Superseded = replaced by a later entry. Unresolved = dispute still open.