Abstract

A common objection to the use of artificial intelligence in decision-making is the concern that it is often difficult to explain or understand how AI systems make decisions. There is a growing body of technical AI research developing techniques for making AI more “explainable” or “interpretable”. However, it is still not well understood why this is an important property for an AI system to possess, or what types of explanations are most important. While there are empirical studies of which types of explanations individuals subjected to AI decision-making find satisfactory, psychological evidence suggests people’s sense of understanding is often unreliable and easy to manipulate. In this paper, I argue that a pragmatist account of explanation provides a fruitful framework for exploring the problem of AI Explainability, which allows us to combine normative and empirical perspectives on user values.