Human-Centered Evaluation of XAI Methods

Multiple explanation methods achieve similar human understanding: Despite using different technical approaches, three leading XAI methods (Prototypical Part Networks, Occlusion, and Layer-wise Relevance Propagation) all enabled users to understand and categorize images with roughly equivalent effectiveness, suggesting that the specific technical approach matters less than previously thought.

Different methods highlight different image regions: The study found that these explanation methods identify varying "pivotal pixels" to justify AI decisions, yet users still gained comparable understanding—indicating that there may be multiple valid ways to explain the same prediction rather than one "correct" explanation.

Human-centered testing is essential for XAI evaluation: Rather than relying solely on mathematical metrics, this research directly measured whether explanations actually help people understand AI decisions, highlighting the importance of testing explainability methods with real users before deploying them in practice.

Explanations should mirror human reasoning: The research leverages the fact that humans naturally explain image classifications by pointing to salient features, suggesting that effective AI explanations should be designed around how people naturally think and communicate about visual decisions.

Transparency methods have practical value: By demonstrating that these explanation methods help users efficiently understand and categorize images, the study provides evidence that XAI methods can meaningfully enhance AI transparency and trustworthiness in real-world applications.