Not All Explanations are Created Equal: Investigating the Pitfalls of Current XAI Evaluation

Simply adding any explanation improves user satisfaction, regardless of quality: The paper demonstrates a critical flaw in current XAI evaluation—most studies compare explanations against no explanation at all, but this comparison is misleading because users report higher satisfaction with virtually any explanation, even incorrect or low-quality ones. This means current evaluation methods cannot reliably distinguish between genuinely helpful explanations and poor ones.

Current XAI evaluation methods lack rigor and generalizability: The researchers found that existing evaluation techniques are "ad hoc at best and not generalizable," relying heavily on simple user surveys rather than objective, standardized metrics. This inconsistency makes it difficult to compare XAI systems across different studies or to have confidence that improvements in one context will transfer to others.

Actionable explanations should be prioritized over mere transparency: The paper argues that the field should shift focus toward explanations that enable users to take meaningful action, rather than simply providing information. In their chess-teaching experiment, actionable explanations proved more valuable than explanations that merely described what the AI was doing.

The field needs comprehensive evaluation standards beyond user satisfaction: The authors call for developing more rigorous, objective evaluation frameworks that can measure explanation quality independently of whether users feel satisfied. Without such standards, researchers cannot definitively prove their explanations are actually improving AI transparency and trustworthiness.