OpenAI calls for global standards on frontier AI

OpenAI has called for international technical standards covering increasingly capable AI systems and the development of AI that can automate parts of its own research.

The company argues that shared standards could help countries use common approaches to measuring AI capabilities, assessing risks and evaluating safeguards. It identifies fragmented evaluation requirements, different reporting practices and uneven technical capacity as challenges that could be addressed through greater international cooperation.

A key part of the proposal concerns recursive self-improvement (RSI). OpenAI uses the term to describe AI systems taking on an increasing share of the work involved in developing later generations of AI. In principle, this could accelerate the pace of AI development.

OpenAI says fully autonomous recursive self-improvement is not happening today. It also says such systems should not be pursued unless they can operate safely while maintaining meaningful human oversight.

The company proposes common standards for measuring progress towards more automated AI research and assessing how much of this work is being carried out autonomously. It also suggests defining points at which automated research should trigger human review.

Another part of the proposal concerns evaluation and reporting. OpenAI calls for shared methods for measuring capabilities, assessing whether safeguards are adequate and classifying incidents involving AI alignment or automated research. Common reporting thresholds could make serious incidents easier to compare across countries.

OpenAI proposes that the United States lead an international effort to develop these standards. It points to cooperation between national AI safety institutes, the US Center for AI Standards and Innovation (CAISI) and national industry bodies as one possible approach.

The proposed standards would not themselves create licences or require governments to approve AI models before release. Instead, OpenAI envisages them as a common technical framework that individual governments could choose to incorporate into their own laws and regulatory systems.

The company also calls for secure communication channels between governments and operators of critical infrastructure. These channels could be used to share information about emerging vulnerabilities, national security concerns and frontier AI safety practices. OpenAI specifically identifies dialogue between the United States and China as an area where greater cooperation could be useful.

What are frontier AI standards?

In this context, they would be common technical methods for assessing the capabilities and risks of the most advanced AI systems. Rather than creating one global regulator, the proposal focuses on agreeing on how systems are tested, how risks are measured and when certain developments should receive additional human oversight.

The proposal comes alongside broader efforts to develop international approaches to frontier AI. Its focus is on creating comparable technical evidence and safety practices across countries, while leaving governments to decide how those standards should be reflected in their own regulatory frameworks.