Artificial Intelligence Seminar - Dr. J. Schoeffers, University of Texas

Title: Explanations, Fairness, and Appropriate Reliance in AI-Assisted Decision-Making

Abstract:

Explanations are often seen as crucial for ensuring fair AI-assisted decision-making. However, there are widespread misconceptions regarding how exactly explanations can help humans enhance fairness. In this talk, I delve into the requirements for explanations for promoting fairness and share insights from an empirical study focused on the task of occupation prediction. Our findings suggest that the effect of feature-based explanations on fairness is contingent on whether the underlying AI system utilizes sensitive information, such as gender. However, previous research has debunked the idea of "fairness through unawareness," which deems an AI system fair if it abstains from utilizing information that is evidently associated with demographic attributes. This raises doubt on the reliability of feature-based explanations as a means to enhance fairness. I discuss the implications of these findings for the field of responsible AI and suggest potential paths forward.