SHAP scores fail pervasively even when Lipschitz succeeds

SHAP scores can systematically mislead decision-makers across multiple model types: The paper demonstrates that SHAP scores—one of the most popular explainability tools in AI—produce "thoroughly unsatisfactory" explanations not just in artificial cases, but for Boolean classifiers, regression models, and even models designed to be mathematically well-behaved (Lipschitz continuous).

The problem is more widespread than previously understood: While earlier research identified issues with SHAP scores, this work shows the problems are fundamental and pervasive rather than limited to edge cases, occurring "arbitrarily many" times across different model classes and mathematical properties.

Mathematical guarantees don't ensure reliable explanations: Even when regression models satisfy Lipschitz continuity—a mathematical property increasingly used to ensure AI robustness—SHAP scores still fail to provide trustworthy explanations, suggesting that technical safeguards alone may not solve explainability problems.

Current XAI tools may create false confidence in AI decisions: Since SHAP is ubiquitously used for explaining AI systems to stakeholders and decision-makers, the pervasive failures documented here suggest organizations may be relying on explanations that actively mislead them about how their AI systems actually work.

Responsible AI deployment requires caution with popular explanation methods: Organizations using SHAP scores for high-stakes decisions (hiring, lending, healthcare) should be aware these explanations may not reliably reflect what their models are actually learning, warranting additional validation before trusting them for critical decisions.