Reproduce “Not All Explanations are Created Equal: Investigating the Pitfalls of Current XAI Evaluation” on OpenResearch

About this paper

Joe Shymanski, Jacob Brue, Sandip Sen

Researchers at The University of Tulsa demonstrated that conventional XAI evaluation methods, which rely on user satisfaction compared to a 'no explanation' baseline, are insufficient because users are equally satisfied by non-informative (placebic) explanations as by genuinely helpful ones. The study found that while placebic explanations initially boosted user confidence, only actionable explanations led to improved task performance and knowledge retention in a chess domain.

Key insights

• User satisfaction and perceived explanatory power are unreliable metrics for evaluating XAI systems, as users were equally satisfied by placebic and actionable explanations, and could not discern higher explanatory power in the actionable explanations.
• The 'placebic effect' was empirically validated: non-informative explanations can initially boost perceived understanding and satisfaction but lead to a false sense of knowledge and a significant drop in actual performance during independent tasks.
• Actionable explanations that provide concrete, domain-specific information effectively enhance user understanding and lead to measurably improved task performance and knowledge retention, unlike placebic or no explanations.

Summary by alphaXiv.