OpenAI calls for US-led international standards on frontier AI

The proposal would create common approaches to capability testing, risk assessment, incident reporting and human oversight while leaving individual governments to decide how standards enter national law

OpenAI is proposing international technical standards for frontier AI covering capability evaluation, risk assessment, incident reporting and human oversight

OpenAI has proposed a US-led effort to develop international technical standards for frontier AI, including common methods for measuring advanced capabilities and overseeing increasingly automated AI research.

The company argues that national rules alone could leave countries using conflicting definitions, evaluation methods and reporting requirements as AI systems become more capable. Its proposal instead sets out a common technical foundation that governments could choose to incorporate into their own legal systems. 

Crucially, OpenAI says the proposed standards would not amount to licenses, mandatory pre-release reviews or approval requirements for new AI models. National governments would retain authority over whether and how the standards were used domestically. 

George Osborne, Head of OpenAI for Countries and former UK Chancellor of the Exchequer from 2010 to 2016, framed the proposal on LinkedIn around international coordination. 

“Our new proposal sets out how the US can work with countries around the world to lead the development of shared standards for frontier AI, building on OpenAI’s long-standing calls for global coordination,” he wrote.

Osborne added that the standards would provide “a common foundation for evaluating capabilities, assessing risks and safeguards, and maintaining human oversight.”

Automated AI research brings oversight into focus

A major part of the proposal concerns automated AI research and recursive self-improvement, or RSI.

OpenAI uses the term to describe a process in which AI systems take on increasing amounts of the work involved in developing subsequent generations of AI. People could remain involved, but as more of that research becomes automated, the company says the pace of AI development could accelerate.

OpenAI is explicit that fully autonomous RSI is not happening today and says it should not be pursued unless and until it can be done safely.

The concern is that increasingly automated research could eventually become difficult for people to understand or oversee. OpenAI says proceeding without sufficient safeguards could leave humans with less practical control over AI development. 

Its proposed international standards would therefore cover areas including the evaluation of AI progress linked to RSI, measurements of how much autonomous research is taking place inside an AI company and the types of automated research processes that should trigger immediate human review.

The company also wants common approaches to classifying, tracking and reporting incidents, including shared severity levels and thresholds for when incidents should be reported. 

Building on an existing international network

Rather than proposing an entirely new global regulator, OpenAI suggests using the emerging network of national AI safety and standards institutions.

Its proposal points to bodies already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the UK, alongside the US Center for AI Standards and Innovation, or CAISI. 

OpenAI says the standards process should include open and closed model developers, independent technical experts and academia, and should be designed so that it does not advantage particular companies, countries or business models.

The company also draws parallels with aviation and financial stability, where countries use shared technical standards and cooperation mechanisms while retaining national authority.

Osborne made the same point on LinkedIn, writing: “Aviation and financial stability offer lessons in how countries can cooperate without giving up national authority. Each government would decide whether and how to incorporate AI standards into its own legal system.”

OpenAI’s proposal also calls for secure communication channels between governments and critical infrastructure operators to share information on emerging vulnerabilities, security threats and frontier AI safety practices. It specifically identifies dialogue between the United States and China as one potentially useful part of that cooperation. 

The company says AI developers themselves remain responsible for pursuing advanced capabilities safely. Its argument for common standards is that governments, researchers and other stakeholders outside individual AI labs would have a clearer technical basis for assessing how frontier systems are being developed.