TDF5

Algorithmic predictions are used to allocate social goods such as healthcare, job training, and education. Despite efforts to apply fairness frameworks and participatory approaches, practical outcomes remain problematic as recent investigations have shown. This talk examines standard approaches to ‘fair machine learning’ through three cases: (1) health programs, (2) long-term unemployment, and (3) school dropout. It critically assesses their limitations and normative assumptions. Two key distinctions clarify the debates: fairness-focused versus welfare-focused methods on the one hand, and whether predictions are instrumentally or communicatively rational on the other. The latter distinction stresses whether algorithms serve effective implementation or facilitate collective evaluation of policy goals.