Oxford Martin AI Governance Initiative report concludes verifying international AI agreements appears possible without new technology
A report finds many international AI agreements could be monitored using existing technology, while some activities remain difficult to verify.
Summary
A report by a large group of authors, whose list begins with Ben Harack and Robert F. Trager, looks at whether international agreements about artificial intelligence could be verified. The report starts from the observation that AI's growing effects are pushing states to think about such agreements, and that whether they are politically feasible can depend on how far each party can tell if the others are complying.
The authors examine several possible agreements and how each might be checked. To make their conclusions sturdier, they deliberately assume pessimistic conditions for both the technical and the political side of the verification problem.
They reach three main findings. The first is that verifying many agreements looks achievable without speculative breakthroughs in verification technology. Some could be checked with hardware that already exists, though others would need heavy investment in new verification infrastructure. Overseeing how AI is developed and deployed in data centers seems possible within a few years if serious effort is made. One approach would combine special-purpose verification hardware fitted in data centers with a mutually verified data center able to run privacy-preserving computations. The second finding is that verifying certain AI-related activities would face combined technical and political obstacles that limit chances of agreement. Fine-grained rules for mobile AI-enabled devices used in sensitive fields, weapons being an example, are singled out as facing severe political difficulty.
The third finding is that steps taken soon, in research and development and in state policy, can improve the outlook for later verification deals by lowering costs and easing security worries. The report concludes that it describes workable ways of verifying international AI agreements and shows that investment in verification now can shape what is politically possible in the future.