Abstract
This talk discusses a basic problem in algorithmic fairness in criminal justice. There are several statistical definitions of algorithmic fairness, and well-known results show that an algorithm usually cannot satisfy all of them at once. So, whoever uses a risk-assessment tool in criminal justice has to accept some kind of unfairness and decide which one. The talk asks how such a choice can be justified to each person affected by it. The answer draws on the contractualist idea that a principle is wrong if someone could reasonably reject it. On this view, we should loaok at the objections of each affected person, such as a defendant who fears wrongful detention and a neighbour who fears crime, and ask whose objection is stronger. The talk also considers what is owed to those who lose out under the chosen tool.
