Interpretability is in the eye of the beholder: Human versus artificial classification of image segments generated by humans versus XAI

Interpretability is not one-size-fits-all: The study found that different explanation methods work better for different types of images—Grad-CAM performed well for indoor scenes and landscapes, while XRAI was better for object classification. This suggests that claims about which AI explanation method is "best" need to specify the context.

Humans and AI may agree on what's interpretable, but not always: Interestingly, humans and the AI model largely agreed on which visual explanations were easier to understand. However, they diverged significantly when evaluating explanations based on human attention patterns (eye movements and manual selections), suggesting that human-centered explanations don't automatically benefit both humans and machines equally.

Current metrics for evaluating explanations may be incomplete: The research highlights a gap between automated metrics and human-centered evaluation of AI explanations. Organizations developing responsible AI systems should not rely solely on one type of evaluation metric, as they may miss important aspects of how explanations actually work in practice.

Context matters more than the explanation method itself: The quality of an explanation depends on three factors working together: which explanation method is used, what type of image is being explained, and who (human or AI) is interpreting it. This complexity means responsible AI evaluation requires testing explanations across multiple realistic scenarios rather than general benchmarks.