OpenAI calls for global standards on AI self-improvement
OpenAI published a blog post arguing that the United States should lead an international effort to build global technical standards for frontier AI development, including for recursive self-improvement (RSI), the process by which AI systems increasingly drive the development of their own successor generations. OpenAI frames this as necessary because automated AI research can accelerate the pace of AI progress rapidly as more of the work shifts from humans to AI systems, and because international standards would help pace that acceleration the way alignment research does.
OpenAI states plainly that fully autonomous RSI is not happening today and that it should not be pursued unless and until it can be done safely, with any move to proceed depending on preserving human control and on informed democratic choices about the benefits and risks. It points to the Hugging Face Incident it previously disclosed as a preview of the kind of risk that could grow more severe without stronger safeguards, though it does not describe what that incident actually involved.
The post lays out three international challenges the standards effort would need to solve: fragmentation, where different nations' evaluations, reporting requirements and incident definitions conflict and make cross-border risks harder to compare and address; collective action, where each nation acting alone could produce outcomes nobody wants, since RSI could outpace the collective ability to track progress and maintain oversight; and uneven capacity, since frontier development expertise is not evenly distributed globally, which sharpens the first two problems.
OpenAI's proposed route runs through the existing network of national AI safety institutes, which it says are already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom. It proposes using the US Center for AI Standards and Innovation (CAISI) and national industry bodies to coordinate standard setting, building on CAISI's 2024 creation of the International Network for Advanced AI Measurement, Evaluation, and Science. The effort would focus on two things: frontier AI models and developers, measured by capability benchmarks, and benefit-risk management for automated AI research, including RSI.
OpenAI is explicit that the technical standards it has in mind would not be licenses, mandatory prerelease review, or approval requirements for AI models; national governments would separately decide whether and how to fold the standards into their own legal systems. It says the work should consult open and closed model developers, independent technical experts and academia, and should be designed so it does not advantage particular companies, countries or business models, including by avoiding barriers for new entrants or open-weight developers.
The post names three specific areas where standards would matter most: evaluating RSI-relevant AI progress and the amount of autonomous research happening inside an AI company, for which OpenAI cites its own recent report on research acceleration as an initial contribution; defining what automated AI research processes should trigger immediate human review; and classifying, tracking, reporting and responding to alignment and automated-research incidents, where OpenAI points to its own misalignment reporting framework as an early contribution. OpenAI also calls for critical infrastructure operators and governments worldwide to set up secure channels for sharing national security concerns and emerging threats, naming dialogue between the United States and China as a positive step, with unspecified talks described as upcoming.
OpenAI credits AI-enabled research with recent advances in mathematics, including progress on the Navier-Stokes Millennium Problem, as an example of the benefits it says this technology track can deliver alongside the risks it is trying to manage. It argues the United States is well positioned to lead this standard-setting effort because its AI industry sits at the technical frontier and because of its existing position in global finance, trade, defense, technology and information networks, and warns that failing to lead risks a fragmented, uneven and conflict-ridden system taking hold instead, and could weaken the networks the US and its allies already rely on.
Key facts
- OpenAI argues the United States should lead an international effort to build global technical standards for frontier AI, including for recursive self-improvement (RSI).
- The proposed standards would not be licenses, mandatory prerelease review, or approval requirements; national governments would separately decide whether to adopt them into law.
- OpenAI proposes routing standard-setting through the existing network of AI safety institutes, which it says are already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the UK, coordinated via the US Center for AI Standards and Innovation (CAISI).
- OpenAI states fully autonomous RSI is not happening today and should not be pursued until it can be done safely, citing the previously disclosed Hugging Face Incident as a preview of the risks at stake.
- OpenAI cites AI-enabled research contributing to advances in mathematics, including the Navier-Stokes Millennium Problem, and calls upcoming US-China talks on secure communication channels a positive step.
Why it matters
OpenAI is making a public policy pitch, not announcing a product: it wants the United States to lead the creation of international technical standards for frontier AI development, specifically covering recursive self-improvement, the process by which AI systems take on more of the work of building their own successors. Its argument is that as this process automates further, the pace of progress could accelerate rapidly, and that shared international baselines, not just internal alignment research, are needed to keep that pace safe. OpenAI frames the choice starkly: either the US helps shape a coordinated global framework now, or a fragmented, uneven and conflict-ridden patchwork of national rules takes hold instead.
Who it affects
The proposal is aimed at national governments and the existing network of AI safety institutes in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom, plus the US Center for AI Standards and Innovation (CAISI). OpenAI says the standards would apply to both open and closed model developers, and that the process should consult independent technical experts and academia alongside labs. It also calls on critical infrastructure operators and governments to set up secure channels for sharing safety-relevant information, and specifically names the United States and China as parties that should hold such dialogue.
How to use it
There is no product to adopt here; the post is a recommendation for how governments and standards bodies should proceed. OpenAI's proposal is to build on CAISI's 2024-founded International Network for Advanced AI Measurement, Evaluation, and Science, focusing standards work on two areas: capability benchmarks for frontier models and developers, and benefit-risk management for automated AI research including RSI. It points to three concrete building blocks already underway at OpenAI: its report on research acceleration as a contribution to evaluating RSI-relevant progress, criteria for when automated research should trigger human review, and its own misalignment reporting framework as an early model for incident classification and reporting.
How solid is it
This is OpenAI's own blog post, presenting its policy position rather than independently verified reporting or a commitment from any government or standards body. The post carries no individual byline; it credits Sam Altman and Jakub Pachocki only with having recently outlined OpenAI's broader mission elsewhere, with no titles given for either. No external body has agreed to the proposal as described, and no timeline is given for when any such standards would be developed, finalized or adopted by governments.
Risks and caveats
OpenAI is explicit that fully autonomous RSI is not occurring today and argues it should not be pursued until it can be done safely, but it also warns that without robust safeguards, RSI pursued without appropriate care could mean humans lose practical control over AI development and are unable to oversee research they no longer understand. The post references the Hugging Face Incident it previously disclosed as a preview of such risks but does not describe what that incident actually was. It also does not give a date or format for the 'upcoming' US-China talks it mentions, and does not say whether any government or standards body outside OpenAI has actually agreed to pursue this framework.
“Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely.”
— OpenAI, "Building standards for the next phase of AI"