OpenAI published a Global Affairs policy position on September 21, 2026, arguing that the United States should lead an international effort to set technical standards for frontier AI — including automated AI research and recursive self-improvement (RSI). The company frames standards as a complement to alignment research, not a substitute, and says shared baselines are needed as AI systems take on more of the work of building successive generations of AI.
The post opens from OpenAI’s mission framing, citing recent writing by Sam Altman and Jakub Pachocki, and states explicitly that fully autonomous RSI is not happening today and should not be pursued unless and until it can be done safely, with human control and democratic choices about benefits and risks. As a preview of how severe risks could become without robust safeguards, OpenAI points to the Hugging Face incident it previously disclosed.
What the position asks for
OpenAI’s concrete ask is that the U.S. convene work with other countries on global technical standards for frontier AI, including RSI. It proposes leveraging the emerging network of AI safety institutes — naming Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, and the United Kingdom — and routing standard-setting through the Center for AI Standards and Innovation (CAISI) and national industry bodies. The focus would be frontier models and developers measured by capability benchmarks, plus benefit-risk management for automated AI research.
The post says those standards should cover capability measurement and evaluation, risk assessment, safeguard sufficiency, human-oversight triggers for automated research, and incident classification and reporting. OpenAI cites its research-acceleration report and misalignment reporting framework as early contributions. It also argues for secure communication channels between critical infrastructure operators and governments worldwide — including dialogue that can involve major competitors.
OpenAI is explicit about what the standards would not be: licenses, mandatory pre-release review, or approval requirements. National governments would decide whether and how to adopt them. The work should consult open- and closed-model developers, independent experts, and academia, and be developed transparently without advantaging particular companies, countries, or business models. Alongside CAISI’s International Network for Advanced AI Measurement, Evaluation, and Science (noted as created in 2024), the post points to bodies such as ISO, the Frontier Model Forum, the Agentic AI Foundation, and the Open Secure AI Alliance.
Why it matters
OpenAI is tying its RSI and automated-researcher agenda to a U.S.-led measurement and incident-reporting stack — shared definitions that would make labs’ claims about autonomy and safeguards more comparable across borders. That is a governance play as much as a safety one: whoever sets the baselines shapes how “responsible” frontier development is scored. The hard line that fully autonomous RSI is off-limits until it can be done safely is the load-bearing caveat; without it, the standards pitch reads as cover for racing.
Our take
This is OpenAI asking Washington to industrialize the yardsticks that would make RSI-era progress auditable — while insisting those yardsticks must not become licensing walls that freeze out open-weight builders. The country list and CAISI/network pointers are real scaffolding; the missing pieces are timetable, convenor, and how China or a reluctant EU would sit in a U.S.-led room. Treat the post as a policy ask with a clear RSI brake, not as evidence that international standards already exist.
