Talk: Tim Henning (Mainz)

Title:

Why Algorithmic Justice must sometimes be Blind - or at least Partially Inaccurate

Abstract:

An increasing number of decisions (both private and public) are made on the basis of predictive algorithms. Recent work has shown that such algorithms cannot satisfy both of the following desirable criteria of fairness: They cannot be equally accurate indicators for different social groups, or they cannot have equal rates of errors (false negatives, false positives) for all social groups. This talk offers an interpretation of these results and argues for the controversial claim that in certain central cases, equal error rates are more important than accuracy. More specifically: In the context of punishment by the state, equal predictive accuracy will force us to take an arbitrary, and incoherent, view of certain delinquents' autonomy.