Framing Computational Fairness and Non-Discrimination

This article outlines a new framework of computational fairness for the analysis of data-driven statistical discrimination. It begins by identifying the limitations of the still dominant paradigm of fairness through unawareness . Building on this critique, the article proposes an approach that distinguishes concerns of procedural fairness from questions of distributive justice when evaluating potentially discriminatory procedures. In contrast to a widespread view, it argues that procedural computational fairness rests on a single formal criterion. A computational procedure is non-discriminatory, it is claimed, if it achieves equal predictive values across demographic groups. In addition, two non-procedural requirements of distributive justice are advanced for selective procedures: the proportionality of individual burdens and their compatibility with a just overall social distribution of benefits and burdens.