The Responsibility Gap in Explainable AI

Abstract

This paper critically examines the persistent responsibility gap in explainable artificial intelligence (XAI), arguing that current approaches fail to adequately bridge it. While XAI is often promoted as a remedy for the opacity of AI systems—the so-called “black box problem”—and as a way to enable human accountability in both high- and low-stakes AI applications, I contend that these hopes are premature. The paper distinguishes between legal and moral responsibility, arguing that regulation alone cannot address the full spectrum of responsibility gaps, especially in everyday, low-stakes scenarios. Then, I argue that XAI needs to tackle two separate but interconnected problems in its attempt to bridge the responsibility gap: an epistemological problem concerning the nature of XAI explanations, and a moral attribution problem regarding the distribution of blame and praise in situations where AI is involved in decision-making. Through detailed discussion of XAI methods and their limitations, I demonstrate that existing techniques often provide only post-hoc, technically focused explanations, which are insufficient for grounding moral responsibility. The paper concludes that without fundamental advances in both the design of explanations and our practices of responsibility attribution, the responsibility gap in AI will remain an unresolved and pressing issue for the field.