Twenty countries are calling for more international oversight of advanced AI, with a focus on how powerful models are tested before release and who gets to evaluate their risks.
The proposal arrived ahead of the United Nations General Assembly, where AI governance is expected to be a major topic. The U.S., U.K. and China did not sign the statement.
The plan centers on testing models before release, developing common safety standards and improving how serious incidents are reported. It also asks U.N. members to consider a new international body that could bring countries together when systems reach certain capability thresholds.
Why It Matters: Global rules for advanced models could add another layer to enterprise AI governance. Independent testing and shared reporting standards could affect what organizations require from vendors before models reach production, particularly for companies operating across multiple countries. Much of the proposal still needs to be worked out, including how common standards would be applied and what authority a new international body would have.
- Testing Would Start Before Release: Frontier AI developers would be expected to test systems for safety issues before deployment and follow transparent protocols that explain how those evaluations are conducted. Independent evaluators would also receive enough access to examine potential risks themselves, although who would conduct those evaluations and how much access they would receive have not been established.
- Common Standards Could Cross Borders: Governments and regional organizations would work toward common standards for evaluating advanced AI systems, while making the expertise and resources needed for those evaluations more widely available. The details would still need to be worked out between countries, including how the standards would apply across different regulatory systems.
- Serious Incidents Would Be Shared: Greater transparency would also apply when advanced systems bypass safeguards, exploit vulnerabilities or gain unauthorized access. The proposal calls for serious incidents to be reported and shared across countries, although it does not define what would qualify as a serious incident or exactly what information would need to be disclosed.
- Crossing the Threshold: Once advanced systems reach certain levels of capability, the proposal calls for a way to bring governments together to assess what comes next. That could include sharing information about new risks and deciding whether additional safeguards are needed. What those thresholds would look like, and which capabilities would trigger that response, have yet to be defined.
- A Global Body Could Take Shape: U.N. members are being asked to explore a new international institution that could set standards, support verification and bring countries together when capability thresholds are crossed. The proposal would build on existing U.N. efforts around AI science and governance, giving countries another forum for coordinating oversight of advanced models.
A Call for Control of Frontier AI Models – President of the Republic of Finland
