Which Information should the UK and US AISI share with an International Network of AISIs? Opportunities, Risks, and a Tentative Proposal

Strategic information sharing can advance global AI safety, but requires careful categorization: The paper argues that the UK and US AI Safety Institutes should selectively share certain types of information (such as evaluation standards and best practices) with the International Network of AISIs to improve worldwide AI safety, while withholding other sensitive categories to protect national interests and competitive advantages.

Some information poses genuine risks if shared internationally: The authors identify specific categories—particularly pre-deployment evaluation results on frontier AI systems—that could create security vulnerabilities or competitive disadvantages if shared broadly, suggesting that not all safety research should be immediately internationalized despite the benefits of global coordination.

A framework is needed to guide sharing decisions: Rather than ad-hoc decisions, the paper proposes that policymakers use a structured framework to evaluate each information category based on potential benefits and risks, acknowledging that some decisions will need case-by-case assessment rather than blanket policies.

Sharing decisions must remain flexible and responsive: The authors emphasize that factors influencing whether to share information are "continuously fluctuating," meaning that institutions need ongoing monitoring systems to adjust their information-sharing strategies as geopolitical conditions, AI capabilities, and international relationships evolve.

The UK and US are uniquely positioned to shape international AI governance norms: Because both countries host frontier AI companies and lead international AI safety discussions, their choices about what information to share will likely set precedents for how other nations approach international AI safety cooperation.