The Disagreement Problem in Faithfulness Metrics

Current faithfulness metrics for AI explanations frequently disagree with each other, meaning that different measurement tools can rank the same explanation method as either reliable or unreliable depending on which metric is used—creating confusion for practitioners trying to choose trustworthy explanation methods.

The field of explainable AI lacks consensus on how to evaluate whether explanations are actually faithful to how models make decisions, despite significant innovation in developing new explanation techniques, suggesting that progress in XAI has outpaced our ability to validate these methods.

Users implementing AI systems cannot reliably determine which explanation method is most trustworthy when available metrics contradict each other, potentially leading to false confidence in explanations that may not accurately reflect the model's actual decision-making process.

The disagreement problem is particularly acute for tabular data classification tasks, which are common in real-world business and healthcare applications, indicating that practical AI deployment scenarios may be especially vulnerable to this validation gap.

Standardizing how we measure explanation faithfulness is essential before organizations can responsibly deploy explainable AI systems, as the current fragmented landscape of metrics undermines the core goal of XAI—providing trustworthy insight into automated decision-making.