Regulating artificial intelligence

Advances in AI offer substantial benefits but also pose societal risks. We analyze optimal regulation under uncertainty about societal costs, differing expectations regarding risks, and opportunities to reduce uncertainty through beta testing. Pigouvian taxes fail to achieve the first-best outcome due to heterogeneous beliefs about risks and the regulator’s inability to observe developers’ expectations. We propose a two-stage optimal policy: first, deciding between immediate release or sandbox experimentation; second, using gathered information to determine whether to publicly release or withdraw the algorithm. This approach achieves the socially optimal outcome.