Description
This paper examines the theory and practical realities of verification of international AI agreements. It makes four contributions: First, it outlines the reasons why states might want AI verification. Second, it provides a specific framing of the AI verification problem---namely, verifying that computations follow computational rules---which might be particularly tractable for both analysis and implementation. Third, it summarizes the technical prospects for implementing AI verification agreements, both now and in the near future. Fourth, it examines an array of political tradeoffs that are inherent in AI verification agreements. In sum, this paper summarizes why AI verification is desirable, provides a productive framing for the problem, and analyses the technical and political prospects for AI verification in the near future.
