Concerning Uncertainty -- A Systematic Survey of Uncertainty-Aware XAI

A team of researchers led by Helena Löfström has published a comprehensive survey, 'Concerning Uncertainty -- A Systematic Survey of Uncertainty-Aware XAI,' analyzing the nascent field of uncertainty-aware explainable AI (UAXAI). The 21-page paper, posted to arXiv, systematically examines how uncertainty quantification is incorporated into AI explanatory methods and how these methods are evaluated. The survey identifies three dominant technical approaches for quantifying uncertainty: Bayesian methods, Monte Carlo techniques, and Conformal methods. It also maps three distinct strategies researchers use to integrate this uncertainty into explanations: assessing the trustworthiness of a model's output, constraining the model or its explanations based on uncertainty levels, and explicitly communicating uncertainty to the end-user.

The survey's critical finding is that evaluation practices across UAXAI research remain fragmented and overly focused on the model itself, with limited attention paid to the end-users who consume these explanations. The authors note inconsistent reporting of crucial reliability metrics like calibration (how well predicted probabilities match actual outcomes) and explanation stability. They observe a recent trend toward distribution-free techniques and a growing recognition that the variability of the explainer itself is a central concern. The paper argues that for the field to progress, it must develop unified evaluation principles that explicitly connect technical uncertainty propagation, explanation robustness, and real-world human decision-making processes. It highlights counterfactual explanations and calibration-focused approaches as particularly promising avenues for building AI systems where interpretability aligns with reliability.