Seminars

Abstract: Artificial intelligence (AI) tools have become commonplace in workplace settings and commercial interactions. These systems typically communicate through terse recommendations (``police reported ahead,'' ``engine service required'') that convey general information about an underlying state without revealing the algorithm's uncertainty about the state. Despite their ubiquity, we still know remarkably little about how ordinary decision makers update their beliefs based on these opaque AI recommendations. This paper provides evidence on this question using a controlled online experiment and finds that a tractable and parsimonious non-Bayesian model of belief revision fits the data better than leading alternatives.