Abstract

When predictions inform algorithmic policies, they are not mere forecasts but have a causal impact on outcomes. Performativity entangles pragmatic and epistemic issues, making notions of accuracy ambiguous. Proposals to address performativity, such as “endogenizing” performative effects or steering outcomes, are misguided. By distinguishing actual and counterfactual predictions, pragmatic and epistemic issues can be separated, enabling machine learning models to be both decision- and discourse-supportive.

About the Speaker

Sebastian Zezulka works on methods for the normative evaluation of algorithmically informed policies, performativity in algorithmic fairness, as well as problems of evidence-based policymaking and philosophy of (social) sciences as they arise in machine learning. He has studied Philosophy & Economics at the University of Bayreuth, Philosophy of the Social Sciences at the LSE, and Psychometrics, Econometrics, and Machine Learning at the University of Tübingen.