Mitigating the judicial human–AI fairness gap

Abstract

Machine adjudication is perceived as less fair than human judging. We study whether hybrid procedures, in which a human consults an algorithmic recommendation but retains final authority, mitigate this human-AI fairness gap. In two experiments ($N = 7{,}651$), participants evaluate sentencing, bail, and arbitration decisions made by a human, an algorithm, or a human reviewing an algorithmic recommendation. The hybrid procedure eliminates the gap even when human scrutiny is minimal. The effect is complete among White and conservative participants but partial among Black liberals. Insensitivity to review intensity raises the risk of human-washing: nominal oversight may legitimize substantively flawed algorithmic processes.