On the Moral Justification of Statistical Parity

Michele Loi, Corinna Hertweck, Christoph Heitz

Abstract: This paper analyzes the moral basis of statistical parity, also called independence, as a fairness criterion in machine learning. It examines whether statistical parity is justified when group differences in predictive features arise from unjust social disparities or measurement errors. Through counterexamples, the paper argues that the moral case for statistical parity cannot be settled only by asking whether observed feature differences are just or unjust.