{Going Beyond Model-Centric XAI towards Human-Centric Explanations }

Model-Centric Explainable AI (XAI) techniques are most commonly used, although they typically do not focus on how useful, meaningful, or actionable the explanations are for humans. In this paper, we argue that this model-centric paradigm is insufficient for building trustworthy AI systems, particularly in high-stakes domains where human decision-makers must interpret and validate AI recommendations. We propose a shift toward human-centred XAI through an interactive visual analytics workflow and interface that integrate rule-based explanations with feature attribution methods, allowing users to dynamically explore and compare multiple explanation techniques side by side. By making agreements and contradictions between methods visible and explorable, the interface allows domain experts to identify inconsistencies and validate model behaviour against their own knowledge, and helps model developers assess which explanations to trust under which conditions. We illustrate this approach through a case study on fishing vessel movement classification, where the hybrid visualisation reveals discrepancies between the rule-based explanation, TreeSHAP, KernelSHAP, and LIME that would remain hidden when using any single method in isolation. This conceptually demonstrates the value of interactive visualisation as a bridge between model-centric explanations and human-centred understanding.