Refutation of Shapley Values for XAI -- Additional Evidence

Shapley values have fundamental limitations for explaining AI decisions beyond simple cases: The paper extends previous critiques beyond Boolean (true/false) classifiers to show that Shapley values fail as an explanation method across broader categories of AI systems, including those with continuous features and multiple classification outputs.

Shapley values may identify irrelevant features as important: The research demonstrates that features flagged as important by Shapley value analysis don't necessarily correspond to features that actually matter for how an AI system makes decisions, raising concerns about the reliability of these explanations for understanding model behavior.

Current XAI methods may give false confidence in understanding AI systems: If Shapley values—a widely-used explanation technique—systematically misidentify which features drive AI decisions, organizations relying on these explanations may believe they understand their systems better than they actually do, potentially masking problematic decision-making.

The gap between mathematical theory and practical AI explanation remains unresolved: While Shapley values have strong theoretical foundations in game theory, this paper adds to evidence that these mathematical properties don't translate into reliable explanations of how real AI systems actually work, suggesting the field needs alternative approaches.