Shapley Values are a go-to method for explaining AI model decisions, but new analysis reveals they can mislead when features are correlated. A recent guide demonstrates how introducing redundant features dilutes the attribution, making a primary driver appear less important. This challenges the reliability of the industry-standard approach for debugging and trust.

The issue stems from the method's Symmetry Axiom, which splits credit equally among dependent features. In a controlled test, duplicating a key feature 100 times spread its contribution thinly across all copies. This doesn't reflect the ground truth, where only the original feature holds real influence. It's a common problem with derived variables like lags or averages.

Fixing this requires moving beyond standard Shapley calculations. Practitioners must acknowledge that correlation-dependent explanations may not always align with practical needs. The guide suggests identifying where dependencies cause dilution and seeking alternative attribution methods. As AI models grow more complex, ensuring explanations are both mathematically sound and intuitively accurate remains a critical, unsolved challenge.