Title: Strategic Information Disclosure to Classification Algorithms: An Experiment (with Aurelien Salas).

Abstract: We experimentally study how individuals strategically disclose multidimensional information to a Naive Bayes algorithm trained to guess their characteristics. Subjects' objective is to minimize the algorithm's accuracy in guessing a target characteristic. We vary what participants know about the algorithm's functioning and how obvious are the correlations between the target and other characteristics. Optimal disclosure strategies rely on subjects identifying whether the combination of their characteristics is common or not. Information about the algorithm functioning makes subjects identify correlations they otherwise do not see but also overthink. Overall, this information decreases the frequency of optimal disclosure strategies.