The Limits of Value Transparency in Machine Learning

Abstract Transparency has been proposed as a way of handling value-ladenness in machine learning (ML). This article highlights limits to this strategy. I distinguish three kinds transparency: epistemic transparency, retrospective value and prospective transparency. corresponds different approaches transparency ML, including so-called explainable artificial intelligence governance based on disclosing information about the design process. discuss sources ML—problem formulation, inductive risk, specification gaming—and argue that is only well-suited for dealing with first, while third raises serious challenges even