Explainable AI and IR

Neural information retrieval and generative AI improve ranking, retrieval, and content generation but make system behavior harder to understand. We examine explainability in information retrieval through medical AI, where predictions must support clinical knowledge and decisions. We clarify core concepts, including interpretability, transparency, robustness, bias, and uncertainty, and distinguish between global and local, intrinsic and post-hoc explanations. We compare interpretable classical machine-learning methods with deep neural networks and review techniques such as saliency maps, CAM, Grad-CAM, LIME, and concept-based explanations. Examples from medical imaging and histopathology show both their value and limitations. We argue that explanations should be evaluated by how well they support human decisions, not only by their visual plausibility. We conclude that trustworthy AI requires robust models, explicit uncertainty, user-adapted explanations, and rigorous evaluation in real application contexts.