Challenges in Evaluating Explanation Methods for Static and Evolving Data
A paper accepted for the EASi 2026 Workshop at IJCAI-ECAI 2026 addresses limitations in XAI evaluation, illustrated via the DetoxAI image recognition system for bias detection and concept unlearning. It presents a human-grounded evaluation of image classification explanations and explores adapting explanations to evolving data streams with concept drift, including adapting counterfactuals. The paper relates these to challenges in tracking co-evolution of data, models, and explanations.
This research highlights insufficient evaluation practices in Explainable AI (XAI). Using the DetoxAI system, it demonstrates issues in bias detection and concept unlearning. A human-grounded evaluation of image classification explanations is provided. The work further investigates adapting explanations, particularly counterfactuals, to data streams with concept drift, and discusses the broader challenge of tracking the co-evolution of data, models, and explanations.
The paper reveals that current XAI evaluation methods are inadequate for dynamic environments. It proposes adapting counterfactual explanations to concept drift, suggesting a need for explanation methods that evolve with data distributions. The DetoxAI case study underscores the difficulty of evaluating unlearning and bias mitigation in image recognition.
For AI systems deployed in changing environments, static explanation methods may become unreliable. This work signals a growing need for robust, adaptive XAI tools in sectors like autonomous systems, healthcare, and finance, where model decisions must remain interpretable over time.
Improved XAI evaluation can enhance trust and compliance in AI products, reducing regulatory risk. Adaptive explanations could lower maintenance costs for deployed models and support safer AI in critical applications.
Next signals include development of benchmarks for evaluating explanations under concept drift, integration of co-evolution tracking into MLOps pipelines, and increased focus on human-grounded evaluation standards. The DetoxAI system may inspire further research on verifiable unlearning and bias detection.