Crowdsourcing Evaluation of Saliency-based XAI Methods

Automated metrics for AI explainability may not reflect human understanding: The paper found that automated evaluation schemes for saliency-based explanations (visual highlights showing which parts of images influenced AI decisions) produced different results than human evaluations, suggesting that high scores from automated systems don't guarantee that humans actually find the explanations helpful.

Human evaluation is necessary for trustworthy AI explanations: Rather than relying solely on computational metrics, the researchers developed a crowdsourcing-based evaluation method where real people assessed whether saliency maps (visual explanations) actually helped them understand AI decisions, highlighting the importance of testing explainability methods with actual users.

Current XAI evaluation practices may be misleading: The study demonstrates a gap between how well explanation methods perform on automated benchmarks versus how well they actually communicate to humans, which has implications for organizations deploying "explainable AI" systems that may appear validated but fail to build genuine human understanding.

Crowdsourcing can efficiently validate AI explanations: The researchers showed that crowdsourcing—leveraging many non-expert evaluators—can be an efficient and practical way to assess whether different explanation methods truly improve human comprehension of AI decisions, offering a scalable alternative to expert evaluation.