Signal Paper Examines Limitations of Explainable AI Evaluation Using Bias-Detection Case Study

Summary

A preprint by Jerzy Stefanowski takes up the problem that Explainable AI, or XAI, is often evaluated insufficiently. That limitation is illustrated concretely through the DetoxAI image-recognition system, which is used for bias detection and concept unlearning. The paper then presents a case of human-grounded evaluation for methods that explain image-classification decisions, checking whether explanations actually align with human understanding rather than relying only on automated metrics. It further explores how explanation techniques can be adapted as data distributions shift over time, extending the discussion beyond static datasets. This approach grows out of a concern that current XAI evaluation practice leans too heavily on static, automated measures.