Learning from an Unknown DGP: Experimental Evidence on Belief Updating with AI Recommendations

We use a controlled experiment to study how beliefs are updated after receiving qualitative information (AI recommendations) from an unknown data-generating process (DGP). Across 60,252 pairs of prior and posterior beliefs, we document three behavioral patterns: updates close to zero when recommendations confirm extreme priors, larger updates when recommendations contradict extreme priors, and smaller updates for intermediate priors. These three behavioral patterns suggest four testable properties of belief updating, which we assess at the aggregate and individual levels. Finally, we examine how well updates are captured by three models of belief updating.