Metrics for Explainable AI: Challenges and Prospects

Explainability alone doesn't guarantee understanding: The paper distinguishes between presenting an explanation and users actually achieving "pragmatic understanding" of how an AI system works—simply having an XAI system doesn't automatically mean it's effective or that users comprehend it.

Multiple dimensions need measurement: Evaluating XAI requires assessing at least six distinct factors: explanation quality, user satisfaction, actual user comprehension, user curiosity, appropriate trust and reliance levels, and overall system performance—not just one metric.

User understanding is distinct from user satisfaction: The paper identifies that users can be satisfied with explanations without truly understanding the AI system, highlighting the importance of measuring comprehension separately from satisfaction ratings.

Trust calibration matters: The research emphasizes measuring whether users develop appropriate trust and reliance on AI systems, not just whether they trust them—misplaced confidence or excessive skepticism both pose risks in responsible AI deployment.

Psychometric evaluation is essential: The authors recommend grounding XAI evaluation in established psychometric methods from social science research, suggesting that rigorous measurement approaches from psychology should inform how we assess whether explanations actually work in practice.