OpenAI calls for global standards to govern frontier AI and recursive self-improvement

OpenAI has called for the United States to lead an international effort to develop technical standards for frontier artificial intelligence, including systems capable of conducting increasingly autonomous AI research and recursive self-improvement.

In a policy proposal published on 21 September, OpenAI said global standards will be needed as AI development accelerates across companies and countries, creating challenges that individual governments and organisations may struggle to address independently.

The company argued that standards could establish common approaches to measuring AI capabilities, assessing risks, implementing safeguards and reporting incidents, while allowing individual governments to decide how such standards are incorporated into national laws.

OpenAI highlights risks from automated AI research

A central focus of the proposal is recursive self-improvement (RSI), where AI systems increasingly contribute to the development of successive generations of AI systems.

OpenAI said automated AI research could accelerate the pace of AI progress while potentially reducing the cost of advanced intelligence and helping researchers improve AI alignment and security.

However, the company stressed that fully autonomous RSI is not currently happening and should not be pursued unless it can be conducted safely.

As AI systems become more capable of conducting research and development themselves, maintaining meaningful human oversight will become increasingly important, OpenAI said. Without adequate safeguards, it warned, humans could lose practical control over AI development and struggle to understand or oversee research processes.

Proposal calls for common AI safety standards

OpenAI identified fragmentation, collective-action challenges and uneven technical capacity as key reasons for greater international coordination.

Different countries could develop conflicting evaluation methods, incident definitions and reporting requirements, making it harder to compare evidence or respond to risks that cross borders.

The company proposed creating a mechanism that connects national and international frontier AI standards. It pointed to the emerging network of AI safety institutes in countries including Australia, Canada, Germany, France, Kenya, Japan, South Korea, Singapore, India and the UK as a potential foundation for such cooperation.

The proposed standards would focus on frontier AI models and developers, capability measurement and the management of benefits and risks associated with automated AI research and RSI.

OpenAI said these standards should provide a common technical foundation for capability measurement, evaluation, risk assessment and determining whether safeguards are sufficient. They would not themselves constitute licences, mandatory pre-release reviews or approval requirements for AI models.

National governments would retain authority over whether and how to incorporate the standards into their own regulatory frameworks.

Common incident reporting also proposed

The proposal also calls for common measurement and incident-reporting protocols to strengthen collective responses to AI risks.

Potential standards could cover the evaluation of RSI-related AI progress, the extent of autonomous research taking place within AI companies and the types of automated research that should trigger human review.

OpenAI also proposed common systems for classifying, tracking and reporting incidents involving alignment and automated AI research, including shared severity levels and reporting thresholds.

The company said critical infrastructure operators and governments should also establish secure channels for sharing information about national security concerns, emerging vulnerabilities and threats related to frontier AI.

It specifically said dialogue between the US and China on these issues could be beneficial.

US urged to take leading role

OpenAI argued that the US is positioned to lead international efforts because of the strength of its AI industry and its wider role in global finance, trade, defence, technology and information networks.

The company said the objective should not be to impose a predetermined pace of AI development, but to ensure that AI alignment research and its deployment keep pace with increasingly capable systems.

It also argued that strong national governance combined with practical international cooperation could help strengthen safeguards while maintaining innovation and access to AI's potential benefits.

The proposal called for standards to be developed transparently, with input from open and closed model developers, independent experts and academia, and designed to avoid giving an advantage to particular companies, countries or business models.