OpenAI is calling for the United States to lead development of international technical standards for frontier artificial intelligence, including common approaches to measuring increasingly autonomous AI research and maintaining human oversight as systems become more capable.
The proposal focuses heavily on automated AI research and the possibility of recursive self-improvement, or RSI, in which AI systems could take on progressively larger portions of the work required to develop later generations of AI.
OpenAI said fully autonomous recursive self-improvement is not occurring today and argued that it should not be pursued unless developers can demonstrate that it can be done safely while preserving meaningful human control.
The company nevertheless sees automated AI research as an important part of its technical roadmap.
OpenAI said one of its goals is to develop automated AI researchers that can contribute to model development and alignment work while people remain involved in the self-improvement process.
OpenAI argues that automation could accelerate scientific and technical progress while also strengthening AI safety research.
The same systems capable of improving AI models could potentially be used to develop better alignment methods, defensive technologies and protections for critical infrastructure, according to the company.
The central concern is what happens as the amount of human supervision decreases.
OpenAI said more automated research could increase the speed of AI development to a point where people have greater difficulty understanding or overseeing how new systems are being created.
That possibility is one reason the company is advocating standards that extend beyond the internal safety policies of individual AI developers.
OpenAI’s proposal emphasizes international coordination because advanced AI systems are being developed and deployed across multiple countries.
The company identified three challenges that it believes common standards could address: fragmented evaluation and reporting systems, collective-action problems between countries and large differences in technical expertise and AI-development capacity around the world.
These concerns apply to both open and closed AI models, according to OpenAI.
The company argues that developers pursuing advanced automated research remain responsible for operating safely under existing laws, while shared standards could give governments, researchers and other stakeholders a clearer framework for evaluating frontier systems.
OpenAI’s proposed approach would not create a single international regulator with authority to approve AI models.
Instead, it envisions common technical standards covering areas such as capability measurement, risk assessment and the adequacy of safeguards.
Individual national governments would then determine whether and how those standards should be incorporated into domestic law.
One potential mechanism outlined by OpenAI would use the growing network of national AI safety institutes to coordinate technical standard setting.
The company identified institutes or related organizations in countries including Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom.
In the United States, OpenAI proposed working through the Center for AI Standards and Innovation, or CAISI, and relevant industry bodies.
The standards would focus particularly on frontier AI developers and models that cross specified capability thresholds as well as the management of benefits and risks associated with automated AI research.
OpenAI also said the framework should include open-model developers, closed-model developers, independent technical experts and academic researchers so that standards do not disproportionately favor particular companies or business models.
The company pointed to organizations including ISO, the Frontier Model Forum, Agentic AI Foundation, Open Secure AI Alliance and Appia Foundation as potential participants in a broader standards ecosystem.
OpenAI’s second major recommendation involves creating common measurement and incident-reporting protocols.
Those standards could establish ways of measuring progress toward increasingly autonomous AI research, including how much research activity inside an AI company is being performed by automated systems.
The framework could also specify when AI-driven research activities require immediate human review.
Another area would be common systems for classifying, tracking and reporting incidents involving alignment problems or automated AI research, including standardized severity levels and thresholds determining when events should be reported.
OpenAI additionally proposed secure communication channels connecting governments and critical infrastructure operators so they can exchange information about emerging security vulnerabilities, national-security concerns and frontier AI safety practices.
The company also said U.S.-China dialogue on these issues could be useful, presenting that as its policy position rather than an established government policy.
The September 21 proposal expands on OpenAI’s recent push for stronger national and international AI governance.
Earlier this month, the company said it supports mandatory U.S. capability-based safety requirements for frontier AI developers as well as common testing, independent assessment, cybersecurity and incident-reporting standards.
OpenAI has argued that those requirements should primarily target the relatively small number of organizations building the most capable systems rather than imposing the same obligations on startups and researchers working below the frontier.
The company’s international proposal follows the same general approach.
Rather than setting a predetermined speed for AI development, OpenAI says progress should depend on whether alignment methods and safeguards remain capable of keeping pace with increasingly powerful systems.
Its recommendation that the United States lead an international standards effort is a policy position from OpenAI, not an assessment that one particular regulatory framework has been adopted or agreed upon by governments.
As AI systems become increasingly capable of completing longer technical projects and participating directly in AI research, the company is positioning international measurement, reporting and oversight standards as part of the infrastructure needed to manage that transition.
For OpenAI, the broader objective is to create rules that allow technical development to continue while ensuring humans remain meaningfully involved in decisions involving increasingly autonomous AI research.
