The Explanation Game -- Rekindled (Extended Version)

Current AI explanation tools have serious flaws: The paper confirms that SHAP scores, a widely-used method for explaining AI decisions, can provide misleading information to human decision-makers who rely on them to understand why an AI system made a particular choice.

The problem isn't with the underlying concept, but the implementation: Rather than abandoning Shapley values entirely, the researchers propose a corrected definition that addresses the fundamental issues with how these explanation scores are currently calculated and presented.

A practical solution is now available: The paper offers not just a theoretical fix but also a computationally efficient method for calculating improved SHAP scores, making it feasible for real-world AI systems to implement better explanations.

Experimental validation confirms the improvements: Preliminary tests demonstrate that the new approach overcomes the documented flaws of current SHAP scores, suggesting that more reliable AI explanations are achievable with this methodology.

Better explanations matter for responsible AI: This work addresses a critical gap in trustworthy AI—ensuring that when humans are asked to make decisions based on AI recommendations, the explanations they receive accurately reflect why the AI reached its conclusions.