Measure Utility, Gain Trust: Practical Advice for XAI Researcher

Shift focus from trust to utility: The paper argues that XAI research has been too narrowly focused on building trust in AI systems, when researchers should instead measure whether explanations are actually useful for their intended purpose. This distinction matters because usefulness is more objectively measurable than trust.

Five practical use cases for explanations: The authors identify five broad scenarios where explanations serve concrete purposes, and recommend designing specific experiments with measurable outcomes for each use case rather than pursuing vague trust-building goals.

Empirical rigor is essential: The paper emphasizes that XAI research needs "objective empirical measurements and falsifiable hypotheses"—the same scientific standards used in established fields studying human behavior—rather than relying on intuition about what makes explanations trustworthy.

Historical conflation of explanation and trust is problematic: The authors note that explanation and trust became intertwined for "historical reasons," but this connection has led the field away from rigorous scientific methods that could produce more defensible knowledge about how explanations actually work.

Pseudo-experiments with measurable outcomes: The paper recommends XAI researchers design controlled experiments that test specific, measurable outcomes rather than attempting to quantify abstract concepts like trust, making it possible to determine whether an explanation truly serves its intended function.