How can we trust opaque systems? Criteria for robust explanations in XAI

Explanations from AI systems need to be "robust" to be trustworthy: The paper argues that we should only trust an AI explanation if multiple different explanation methods produce the same result when analyzing the same decision. If different tools give conflicting explanations for why an AI made a choice, that's a red flag that something may be wrong.

Even agreement among explanation methods isn't enough: The authors warn that all available explanation tools could agree on an explanation that is still fundamentally incorrect. This means we need a two-step verification process: first checking that explanation methods themselves are reliable (explanation method robustness), and only then checking whether they agree with each other (explanatory robustness).

Current XAI methods have credibility problems: The paper cites "recent reviews" showing reasons for skepticism about existing explainability techniques, suggesting that popular tools used today to explain AI decisions may not be as trustworthy as commonly assumed.

Deep learning's accuracy comes at a cost to transparency: While deep learning algorithms make impressive predictions, their decision-making processes remain fundamentally opaque—even to researchers and experts. This opacity creates a trust problem that explanations alone may not fully solve.

We need formal standards for trustworthy AI explanations: The paper proposes a framework with specific criteria for evaluating whether explanations of AI systems are actually reliable, providing a structured approach rather than relying on intuition or individual tools.