A comprehensive study on fidelity metrics for XAI

Current XAI evaluation methods are unreliable: The study found that existing fidelity metrics—tools used to measure whether AI explanations are actually accurate—show concerning disagreements with each other and lack proper verification, making it difficult to trust which explanation method is best.

Even the best metrics fail significantly: When tested against decision trees (a transparent model with "perfect" explanations), the best-performing fidelity metric still showed a 30% deviation from expected values, indicating that current metrics cannot be trusted for real-world applications where accuracy matters.

There is no objective way to verify XAI explanations: The paper highlights a fundamental problem in the field: there is no established "ground truth" to determine if an AI explanation is actually correct, which is why selecting the right explanation method remains an open question without clear answers.

New evaluation standards are urgently needed: The authors conclude that current fidelity metrics are "not reliable enough to be used in real scenarios" and call for the development of new, more trustworthy metrics to properly assess whether AI explanations are genuinely faithful to how the AI system actually works.

Transparent models can serve as benchmarks: The study demonstrates that using decision trees as reference points with known-perfect explanations provides a practical way to objectively test and compare different XAI metrics, offering a path forward for improving the field.