The potential functions of an international institution for AI safety. Insights from adjacent policy areas and recent trends

International coordination on AI safety research is currently fragmented. The paper identifies that while global institutions like the OECD, G7, and UNESCO have begun developing ethical AI frameworks, these efforts lack institutionalized international processes specifically designed to identify and assess potentially harmful AI capabilities—a gap that existing initiatives alone cannot fill.

Three functional domains could structure international AI safety work. The paper proposes organizing an international AI safety institution around: (1) technical research and cooperation, (2) safeguards and evaluations, and (3) policymaking and governance support—drawing lessons from how other international bodies handle safety in adjacent policy areas.

Existing national AI safety institutes provide a blueprint. The UK and US have recently established national AI safety institutes; analyzing their structures and approaches offers concrete models for what functions could be performed at the international level, rather than requiring entirely new institutional designs.

International governance models from other sectors offer practical lessons. The paper examines how existing international institutions have addressed safety considerations in adjacent policy areas, suggesting that proven governance structures and mechanisms can be adapted for AI rather than invented from scratch.

Creating new institutions is not the only solution. The paper emphasizes a "modular perspective," suggesting that understanding existing international bodies' structures can help identify and deploy available tools for AI safety governance, implying that coordination and adaptation of current mechanisms may be as important as establishing new organizations.