Mean Opinion Score as a New Metric for User-Evaluation of XAI Methods

User opinions don't always align with automated metrics: The study found only limited correlation between what automatic evaluation systems say about AI explanations and what actual users think, suggesting that current automated testing methods may miss important aspects of explanation quality that matter to people.

Mean Opinion Score offers a practical way to evaluate explanations: By adapting a metric commonly used to assess image quality, researchers demonstrated that gathering direct user feedback through structured surveys can effectively measure whether AI explanations are actually helpful to people—not just whether they pass technical tests.

Different explanation methods perform differently for users: Among the three feature attribution methods tested (Grad-CAM, MLFEM, and FEM), MLFEM showed the strongest alignment with both user opinions and automatic metrics, indicating that some explanation approaches may genuinely work better for human understanding than others.

There's a gap between technical and human-centered evaluation: The limited correlation between automatic metrics and user opinions highlights an important problem in responsible AI: we need better ways to ensure that explanations are actually understandable and useful to the people who rely on them, not just mathematically sound.