Are Objective Explanatory Evaluation metrics Trustworthy? An Adversarial Analysis

Popular XAI evaluation metrics can be fooled by adversarial explanations: The paper demonstrates that SHAPE, an adversarial explanation method, outperforms established techniques like GradCAM and GradCAM++ on standard evaluation metrics, despite not providing genuinely trustworthy explanations. This reveals a critical vulnerability in how we currently measure whether AI explanations are reliable.

Objective metrics alone are insufficient for evaluating explanations: The authors show that metrics designed to assess explanation quality can give high scores to explanations that don't actually reflect how a model makes decisions. This suggests that relying solely on automated metrics to validate XAI methods is risky and may create a false sense of security about AI trustworthiness.

Human evaluation is essential for responsible AI deployment: The paper explicitly calls for "human involvement for an overall better evaluation" of explanatory techniques, indicating that automated testing cannot replace human judgment when assessing whether AI systems are truly interpretable and trustworthy.

Different explanation methods frequently disagree with each other: The paper notes that various XAI strategies "only occasionally agree," meaning there's no consensus on what constitutes a good explanation. This fragmentation in the field makes it harder for organizations to confidently choose which explanation method to trust.

Current XAI evaluation frameworks may need fundamental redesign: By exposing weaknesses in causal metrics used to measure explanation robustness, the research suggests that the theoretical foundations for evaluating XAI methods require reconsideration before these tools can reliably support responsible AI practices.