Skip to content

Evidence conflict in fact-checking

Evidence conflict refers to the challenge of resolving contradictory evidence in automated fact-checking systems. In real-world social media verification, retrieved evidence frequently presents opposing stances on the same claim: one source may interpret scientific findings as supporting a claim while another disputes the causal link, or one headline may support a claim while fact-checkers and epidemiologists who co-authored the original study refute it.

This splits into two types: evidence-level conflict (different documents present contradictory stances on a single atomic fact), and aggregation-level conflict (sub-claims within a complex claim yield incompatible verdicts that must be reconciled into a final verdict).

Traditional fact-checking pipelines aggregate all retrieved evidence equally or apply simple heuristics (majority vote), which fails when evidence quality and stance genuinely conflict. Modern approaches prioritize evidence via source credibility signals, metadata-derived quality scores, and LLM-assisted reasoning about which evidence is most authoritative or factually grounded.

Key papers

  • Zhang et al. (2026) — CoVer: Conflict-Aware Claim Verification — Three-stage pipeline for resolving conflicts: evidence schema normalization (standardizing directionality, attribute alignment, source reliability, informativeness), factual consensus (LLM-assisted scoring of evidence stance and quality), and support verification (validating that selected evidence directly entails the claim); ContraNote dataset provides 88K instances of real conflicts from X Community Notes