Annie Liang

Abstract

Algorithm designers increasingly optimize not only for accuracy, but also for the fairness of the algorithm across pre-defined groups. We study the tradeoffs between fairness and accuracy induced by a given set of inputs to the algorithm. We formalize these tradeoffs as a fairness-accuracy frontier, defined as the set of outcomes that cannot be simultaneously improved upon in both fairness and accuracy. Our results show that the shape of the frontier is determined by a simple property of the inputs, which we call group-skew. Group-skewed inputs inherently advantage one group, resulting in lower errors for that group even when the algorithm is optimized for the other. We show that decreasing accuracy for both groups in order to increase fairness can be justified by fairness considerations if and only if inputs are group-skewed. We further study an information design problem where a designer flexibly regulates the inputs, but another agent chooses the algorithm. The optimal regulation of inputs generally depends on the designer's preferences, but when inputs are not group-skewed then two implications hold across all designer preferences: (1) banning group identity is strictly suboptimal, and (2) if group identity is available, then banning any informative input is strictly suboptimal.