Theoretical Behavior of XAI Methods in the Presence of Suppressor Variables

Most popular XAI methods can mislead users by highlighting irrelevant features: The paper demonstrates that a majority of widely-used explainability methods will assign high importance to "suppressor variables"—features that have no actual relationship to what the AI is predicting—when those features are correlated with noise in the data.

The problem is fundamental, not accidental: Using mathematical analysis on simple test cases, researchers showed this misleading behavior occurs across many different XAI methods, suggesting it's a systematic issue rather than a bug in individual tools. This confirms earlier empirical findings by Wilming et al. (2022) that the problem is widespread.

Current XAI methods lack formal validation standards: The paper highlights that the XAI field has created many explanation tools without establishing clear definitions of what "correct" explanations should look like or rigorous testing standards. This means organizations using these tools for quality control or transparency cannot reliably verify whether the explanations are trustworthy.

Users should be cautious about trusting XAI outputs at face value: Since these methods can confidently highlight features that are actually irrelevant to predictions, stakeholders relying on XAI for decision-making (such as in hiring, lending, or healthcare) should not assume the highlighted factors are genuinely important without additional validation.

The field needs theoretical foundations before widespread deployment: The authors argue that XAI methods require formal problem definitions and correctness criteria before they can be responsibly used for high-stakes applications requiring transparency and accountability.