Title:
Understanding discrimination - the purported epistemic advantage of algorithmic decision-making
Abstract:
Proponents of explainable AI (XAI) hope that it will allow us to reap the benefits of more accurate and efficient AI systems but also detect, prove, and mitigate algorithmic discrimination. Indeed, the hope is that AI systems become more understandable than human decision-makers, when such tools are used, and that algorithmic discrimination becomes easier to prove in court than direct discriminatory decisions where no AI systems were involved. However, such a position over-estimates how understandable AI systems are, even with XAI tools, and underestimates how understandable humans are. The same point holds for providing direct discrimination in court. In both cases, we must rely on background normative and empirical commitments to understand or prove discrimination.