Evaluation of Saliency-based Explainability Method

Saliency maps lack rigorous human validation: While saliency-based explainability methods are widely used to show which parts of an image a CNN focuses on, most evidence supporting their value is anecdotal rather than based on systematic testing with actual users.

Human experiments are essential for evaluating XAI: The paper demonstrates that computational tests alone are insufficient—the researchers conducted three human subject experiments to properly assess whether these visualization methods actually help people understand how AI models make decisions.

End-user testing should be standard practice: Since humans are the intended audience for explainability tools, the paper emphasizes that XAI methods must be evaluated through direct human studies rather than relying solely on technical metrics, highlighting a gap in how explainability research is typically conducted.

Visual explanations need empirical validation: The research suggests that intuitive-seeming explanations (like highlighting important image regions) require rigorous testing to confirm they genuinely improve human understanding and trust in AI systems, rather than assuming they work as intended.