Rethinking XAI Evaluation: A Human-Centered Audit of Shapley Benchmarks in High-Stakes Settings

The deeper finding here isn't just that metrics are imperfect: it's that the entire benchmarking infrastructure for Shapley-based XAI may be optimizing for the wrong objective from the start, measuring mathematical properties of explanations rather than whether those explanations change analyst behavior in useful ways.

This connects most directly to the 'White-Box Signal-Subspace Probe' paper from the same day, which also challenges opaque learned representations by substituting interpretable signal components. Both papers are pushing toward a similar position: that interpretability tools need to be validated against what practitioners actually do with them, not just against internal consistency measures. The broader cluster of April 24 arXiv work on this site is otherwise focused on efficiency and control problems, so this paper and the WG-SRC piece form a small but coherent thread about the gap between what models expose and what humans can use.

Watch whether fraud detection vendors or financial regulators cite this study when revisiting model explainability requirements. If a regulatory body references human-centered XAI evaluation criteria in updated guidance within the next 12 months, this line of research will have found real institutional traction.