People Are Not Coins: Morally Distinct Types of Predictions Necessitate Different Fairness Constraints

Michele Loi, Corinna Hertweck

Abstract: This paper responds to the claim that common group fairness metrics are not genuine fairness requirements. It argues that predictions about people based on data from similar people are morally different from predictions based only on information about a single individual. Because most machine-learning decision systems are human-group-based practices, group fairness metrics can remain morally relevant for real-world prediction-based decision-making.