OpenAI Calls for US-Led Global Technical Standards for Frontier AI

OpenAI has published a policy proposal arguing that the United States should lead an international effort to write technical standards for frontier AI development, with a specific focus on automated AI research and recursive self-improvement. The post, filed under the company's Global Affairs series, sets out two components it considers essential: a mechanism linking national and international standards work, and shared protocols for measurement and incident reporting. It is a position paper, not a binding commitment, and OpenAI is explicit that it expects the concept to change substantially over time.

What OpenAI proposed

The document builds on a framing the company has used before: its stated priorities are to navigate the next period of AI progress by building an automated AI researcher, to deliver the scientific and economic benefits of capable machines, and to give individuals access to a personal AGI. OpenAI says AI is already accelerating its own research and engineering, and points to AI-assisted advances in mathematics, including work on the Navier-Stokes Millennium Problem, as evidence.

The core of the proposal concerns recursive self-improvement, which OpenAI abbreviates as RSI. The company describes this as a process in which AI systems take on more of the work of building successive generations of AI, potentially accelerating the pace of progress. Its stated position is that fully autonomous RSI is not happening today and should not be pursued unless and until it can be done safely, with the decision depending on the ability to preserve human control and on informed democratic choices. OpenAI references a previously disclosed incident involving Hugging Face, which it says was not a direct result of RSI but illustrates the category of risk that could become more severe without robust safeguards.

From there the argument turns to standards. OpenAI frames international standards as potentially as important to pacing frontier development as alignment research itself, because they can establish shared definitions of high-quality evidence and agreed baselines for how rigorous technical safeguards need to be. The company names three problems it says standards would address: fragmentation, where conflicting national evaluations and incident definitions make evidence hard to compare; collective action, where nations acting independently produce outcomes none of them want; and uneven capacity, since frontier development expertise is concentrated in a few places. It says these apply to open and closed models alike.

Notably, OpenAI grounds its rationale partly in avoiding the concentration of power. Standards, it argues, give stakeholders outside the labs a say in how the technology develops and provide a visible set of principles that does not depend on any particular company's internal practices. The post also states that any lab pursuing automated AI research must take accountability for doing so safely under existing laws.

What it would change in practice

The first proposed mechanism would work through the growing network of AI safety institutes. OpenAI names institutes already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom, and suggests using them alongside the US Center for AI Standards and Innovation and national industry bodies to set standards. It points to CAISI's 2024 creation of the International Network for Advanced AI Measurement, Evaluation, and Science as an existing foundation to build on. The scope would cover frontier models and developers, as measured by capability benchmarks, and benefit-risk management for automated AI research.

The most important limiting sentence in the proposal is about what these standards would not be. OpenAI states plainly that they would not function as licenses, mandatory prerelease review, or approval requirements for AI models. National governments would decide whether and how to write them into their own legal systems. That is a meaningful design choice: it places the proposal closer to the aviation and financial stability models the company cites, where countries agree on common technical baselines without surrendering national authority, and further from a licensing regime.

The second component covers measurement and reporting. OpenAI suggests standards for evaluating RSI-relevant progress and the share of autonomous research happening inside a company, for human oversight of automated AI research including what should trigger immediate human review, and for incident classification with common severity levels and reporting thresholds. It offers two of its own publications as early contributions: a report on research acceleration and a misalignment reporting framework. Beyond standards, it calls for secure communication channels between governments and critical infrastructure operators, and says dialogue between the United States and China on these topics would be a positive step, noting upcoming talks.

Who has to act

Very little of this is within OpenAI's unilateral power, which is the main thing to understand about the document. The proposal asks the US government to lead, asks other governments to participate through their safety institutes, and asks standards bodies to do the technical work. OpenAI names ISO, the Frontier Model Forum, the Agentic AI Foundation and the Open Secure AI Alliance, along with the Appia Foundation, which it describes as developing practical specifications that connect international standards with real-world assessments.

The company also sets conditions on how the work should be done: transparently, with consultation of open-model and closed-model developers, independent technical experts and academia, and designed so that no company, country or business model is advantaged. It specifically flags that standards should not make it harder for new entrants or open-weight developers to compete. That is a direct acknowledgement of the most common criticism of incumbent-backed safety regimes.

Open questions

The obvious tension is that a large frontier developer is proposing the rules under which frontier developers would be measured. OpenAI addresses this partly by insisting standards should be developed transparently and by non-lab institutions, but the proposal does not describe who would arbitrate disputes, how compliance would be verified, or what happens if a developer simply does not participate. Without licensing or prerelease review, adherence is a matter of national implementation and reputational pressure.

There is also a structural question about the pacing argument. OpenAI says pacing is about ensuring alignment research stays ahead of capabilities, but the proposal does not define how anyone would know whether that condition is being met, or who would decide that it is not. The suggested evaluations of autonomous research intensity inside a company would require disclosure that no current standard compels.

Finally, the geopolitical framing cuts both ways. OpenAI argues the United States is well positioned to lead because its industry is at the technical frontier and it holds a central position in finance, trade, defense and information systems. Presenting standards partly as a route to preserving US influence may make the proposal more persuasive in Washington and less persuasive elsewhere, particularly among countries the company is simultaneously asking to join the effort.

What to watch

• Whether CAISI or the International Network for Advanced AI Measurement, Evaluation, and Science takes up any of the proposed work programme, and on what timeline.
• Whether the named safety institutes in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the UK respond publicly, and whether they treat the proposal as a starting point or a vendor position.
• Whether other frontier developers publish comparable proposals, and how far their definitions of incident severity and autonomous research overlap with OpenAI's.
• Whether ISO, the Frontier Model Forum, the Agentic AI Foundation or the Open Secure AI Alliance announce concrete workstreams on RSI evaluation.
• Whether the US-China talks OpenAI alludes to produce any agreement on secure channels for AI safety and security information.