Challenges in Evaluating Explanation Methods for Static and Evolving Data
2026-08-06 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors talk about problems with how we check if Explainable AI (XAI) really works well. They show these problems using DetoxAI, a system that finds bias in images and learns to ignore unwanted patterns. They also give an example where real people help test explanations from image classifiers. The paper discusses how to update explanations when the data changes over time and how counterfactual explanations can be adapted. In the end, the authors highlight the difficulty of keeping explanations accurate as data and AI models both change together.
Explainable AIDetoxAIBias detectionConcept unlearningHuman-grounded evaluationImage classificationConcept driftCounterfactual explanationsModel-data co-evolution
Authors
Jerzy Stefanowski
Abstract
This paper addresses the limitations of Explainable Artificial Intelligence (XAI) with respect to insufficient evaluation. They are illustrated through the DetoxAI image recognition system for bias detection and concept unlearning. Then, an example of a human-grounded evaluation of methods for explaining image classification is presented. The paper further explores methods for adapting explanations to evolving data streams with concept drift. Experiences with adapting counterfactuals for this problem are discussed. Finally it is related to the challenges of tracking the co-evolution of data, models, and explanations.\footnote{This paper has been accepted for a publication in J.Nalepa (ed) Explainable AI in Space. Proceedings of EASi 2026 Workshop at IJCAI-ECAI 2026 Bremen, Springer CCIS vol 3107 (2016).}