The proposal, in brief

OpenAI's Global Affairs team published a proposal on September 21, 2026 calling for international technical standards for frontier AI, with a particular focus on recursive self-improvement, the process by which AI systems increasingly take on the work of developing successive generations of AI. The document, titled "Building standards for the next phase of AI," argues that such standards "may be as important to pacing the frontier as alignment research itself."

The proposal arrives at a moment when AI companies are racing to automate research itself. OpenAI frames the question in practical terms: if AI systems begin driving their own improvement, how do people outside the labs, including governments and independent experts, get a shared way to judge whether that process is safe?

This article separates what the document actually says from the predictions it makes and from any judgment about whether the proposal is good policy. The document is a company's argument, not a law, a regulation, or an adopted standard. Nothing in it binds anyone, including OpenAI itself.

What OpenAI actually proposes

The document proposes two things it calls essential. The first is a mechanism that facilitates complementary national and international frontier standards. OpenAI suggests leveraging the emerging network of AI safety institutes, naming institutes already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, and the United Kingdom, to facilitate standard setting through the Center for AI Standards and Innovation (CAISI) and national industry bodies. The effort would focus on two areas: frontier AI models and developers, as measured by capability benchmarks, and benefit-risk management for automated AI research, including recursive self-improvement.

OpenAI points to CAISI's creation of the International Network for Advanced AI Measurement, Evaluation, and Science in 2024 as a foundation to build on. The National Institute of Standards and Technology confirmed in a February 2026 update that the International Network was founded by CAISI in November 2024 and comprises government bodies from ten members, including Australia, Canada, the European Union, France, Japan, Kenya, the Republic of Korea, Singapore, the United Kingdom, and the United States. In February 2026 the Network published consensus areas and open questions on practices for automated AI evaluations. Note a small discrepancy in membership lists: the NIST page lists the European Union as a member while the OpenAI document lists India and Germany among institutes it names. Both are reported here as each source states them.

The second essential element is a set of common measurements and incident reporting protocols for what OpenAI calls better collective action. The document lists three concrete areas: evaluation of RSI-relevant AI progress and the amount of autonomous research happening within an AI company; human oversight over automated AI research, including what kinds of automated research processes should trigger immediate human review; and incident classification, tracking, reporting, and responding for alignment and automated AI research issues, such as common incident severity levels and reporting thresholds.

The document also calls for secure communication channels between critical infrastructure operators and governments worldwide to share national security concerns, vulnerabilities, and best practices, and says dialogue between the United States and China would be "a positive step."

What is explicitly ruled out

The proposal is notable as much for what it rules out as for what it asks for. The document states plainly: "These technical standards would not be licenses, mandatory prerelease review, or approval requirements for AI models. National governments would decide whether and how to incorporate these standards into their own legal systems."

That framing matters for builders and new entrants. The document says the work should support a competitive AI ecosystem by consulting open model and closed model developers, independent technical experts, and academia. It states that common standards should be developed transparently and "designed so that they do not advantage particular companies, countries, or business models, including by making it harder for new entrants or open-weight developers to compete."

These are commitments readers can track over time. If standard-setting through CAISI and national institutes later produces requirements that only large closed-model labs can satisfy, that would contradict the document's stated intent. Conversely, if open-weight developers and academic experts are genuinely consulted and the resulting benchmarks are public, the commitment holds in practice. The document itself provides the yardstick.

The paper says its approach can draw lessons from aviation and financial stability, where countries have developed common technical standards and trusted channels for cooperation without giving up national authority, and names existing and newer standards bodies, including ISO, the Frontier Model Forum, the Agentic AI Foundation, and the Open Secure AI Alliance, as well as the Appia Foundation, which it describes as developing practical specifications connecting international standards with real-world AI assessments.

Predictions versus facts on recursive self-improvement

The document devotes a section to recursive self-improvement, abbreviated RSI. It describes automated AI research as a process that can involve varying degrees of human supervision, and says that as AI systems take on more of the work of developing successive generations of AI, they can increasingly drive recursive self-improvement "even while people remain involved."

Predictions in the document are attributed clearly to OpenAI. The company predicts that automated AI research will yield models that directly enhance people's lives, that it can bring down the cost of advanced intelligence, and that automated researchers could substantially improve alignment because "an automated AI researcher can also be an automated AI safety researcher." These are the company's expectations, not established outcomes.

The document also states risks in OpenAI's own words: "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely." It warns that RSI done without appropriate care could result in humans losing practical control over AI development, unable to oversee research processes they no longer understand. It references the Hugging Face incident OpenAI disclosed, saying it was not a direct result of RSI but "is a preview of the kinds of risks that could become much more severe without robust safeguards and alignment."

Readers should treat the statement that fully autonomous RSI is not happening today as OpenAI's own current assessment. The document does not predict when, or whether, it will happen, and neither should any coverage of it. Claims about what RSI will produce remain predictions either way, in both the hopeful and cautionary directions.

The problems the document says standards would solve

The document identifies three challenges it says international standards are needed to solve. The first is fragmentation: evaluations, reporting requirements, and incident definitions from different nations could conflict, making it harder to compare evidence, understand emerging capabilities, and respond to risks that cross borders. The second is collective action: each nation acting independently can produce outcomes no nation wants. The third is uneven capacity: frontier development and related expertise are not evenly distributed across the world, which the document says accentuates both other problems.

It states that these challenges apply to both open and closed models, and that any AI lab pursuing automated AI research or other advanced capabilities must take accountability for doing so safely, in accordance with self-responsibility principles and existing laws.

The rationale the document gives for standards is, in its own words, "rooted in avoiding the concentration of power, and producing better practical outcomes." It argues standards give stakeholders outside the labs a say in how the technology unfolds and a visible set of principles that can be relied on regardless of any particular lab's practices. The document also argues the United States should lead the effort, citing its AI industry's position at the technical frontier and its network position in finance, trade, defense, technology, and information systems.

What remains unresolved

Several questions are left open. The document does not specify who would measure compliance with any standard, how adherence would be verified across companies and countries, or what happens when a lab's internal measurements conflict with an external assessor's. It names capability benchmarks as the measure of frontier developers but does not say which benchmarks, who maintains them, or how disputes are resolved.

Nor does it say how safety of recursive self-improvement would be demonstrated. The document says pursuit of fully autonomous RSI should wait "unless and until it can be done safely," but the evidence that would establish safety, and who would judge it, is not defined. That gap is where much of the practical debate will likely land.

There is also a structural tension worth naming, as analysis and clearly attributed opinion: this is a leading frontier lab proposing standards that would govern frontier work, including its own. OpenAI's stated answer is that standards avoid concentration of power and give outsiders a say. A skeptic could note that the same document argues the United States should lead, which aligns with the competitive position of US labs. Both readings can be true at once, and the test is whether the standard-setting process actually includes independent experts, open-weight developers, and non-US institutions in a way that can constrain incumbents, OpenAI included.

Finally, the document's pledge that standards should not disadvantage new entrants or open-weight developers is a commitment, not a guarantee. Standard-setting processes historically can be captured by incumbents who write rules around their own architectures. Whether this one avoids that outcome is an empirical question that only the process itself will answer.

Why this matters for access and what to watch

For readers who build with AI, the trackable items are concrete: whether CAISI and the International Network publish measurable RSI-relevant benchmarks and how open the consultation process is; whether incident severity levels and reporting thresholds are published in a form small developers can adopt; whether any standard ends up functioning in practice like a pre-release review despite the document's explicit disclaimer; and whether open-weight developers report being consulted and treated on equal terms.

The document says national governments would decide whether and how to incorporate standards into law. That keeps the access stakes local. A technical standard that is voluntary in one country can become a legal requirement in another, and the difference will decide whether these frameworks widen or narrow who gets to build.

The Expectancy's own position, stated as opinion, is that shared measurement and transparent incident reporting are constructive goals, and that the access commitments in this document, if honored, would serve builders and users alike. Whether they are honored is now a matter of public record for anyone to check.