Unavoidable Tensions in Explaining Algorithmic Decisions

Solon Barocas

Microsoft Research and Cornell University

Recent developments in methods for explaining the decisions of machine learning models have been widely embraced for their ability to provide transparency and accountability without limiting model complexity or compelling model disclosure. Yet applying these methods is far from straightforward and they rarely prove a cure all. This talk identifies a number of unavoidable tensions that decision makers must navigate as they seek to employ these methods—and the deeply subjective judgment that must go into these considerations.