Explaining the role of evidence in data-driven fact-checking

Fact-checking is crucial for combating misinformation, and computational methods are essential for scalability. The most effective approaches leverage neural models that use domain-specific evidence to validate claims. However, these models often act as black boxes, providing labels without explaining the rationale behind their decisions. In this work, we introduce a comprehensive evaluation protocol to assess how factchecking verifiers attribute decisions to individual evidence pieces. We evaluate whether existing explainable AI methods, such as LIME and SHAP, can be adapted to perform evidence-level attribution and to classify evidence relevance across four established datasets. Our findings show that post-hoc attribution methods can support the analysis of how verifier predictions change under evidence perturbations, thereby improving transparency by highlighting patterns and potential issues in model behavior. This was achieved through the evaluation of five fact-checking systems, showing that for verifiers amenable to perturbation-based analysis, post-hoc attribution methods often provide higher-quality evidence-level signals than native explainers.