Regulating Artificial Intelligence

Advances in AI promise substantial benefits but pose societal risks. We analyze optimal AI regulation when there is uncertainty about societal costs, disagreement over their likelihood, and experimentation, via beta testing and red teaming, can reduce uncertainty. Pigouvian taxes fail to implement the first-best allocation because there is heterogeneity in expectations about societal risks, and the beliefs of AI developers are unobservable to the regulator. We characterize a regulation policy that implements the first best. This policy has two stages. First, the regulator decides whether to release the algorithm or require experimentation. Second, using the information obtained, the regulator determines whether the algorithm should be made available to the public or withdrawn.