Check-worthiness is the task of identifying which factual claims deserve fact-checking attention. It addresses a fundamental bottleneck in fact-checking workflows: with an infinite stream of potential claims, fact-checkers must prioritize claims that are (1) verifiable, (2) claim-able to be important to the public, and (3) likely to have measurable impact if checked.
Check-worthiness detection differs from general fact extraction: a sentence can be factual but not check-worthy (e.g., "Next Tuesday is Election Day"), and conversely, a claim can be worth checking even if it's vague or only partially verifiable. The task requires understanding both linguistic features (presence of numbers, modal language, entity types) and social context (who said it, in what forum, what's at stake).
Key papers
- Hassan et al. (2015) — The Quest to Automate Fact-Checking — Introduces ClaimBuster and formalizes check-worthiness as a three-class classification task (non-factual, unimportant factual, check-worthy factual); identifies cardinal numbers (POS tag CD) as the strongest discriminative feature for check-worthy claims in political debates; achieves 79% precision on 2015 GOP debate with validation against professional fact-checkers.
- Amatya et al. (2026) — Calibrating Small Language Models for Claim Check-Worthiness Detection — Proposes post-hoc calibration for small-model check-worthiness predictions via nearest-neighbor residual correction; achieves 12–33.80% F1 gains across model scales on ClaimBuster benchmark, enabling efficient production deployment.
Related topics
- claim-detection (broader detection framework)
- fact-checking (larger fact-checking pipeline)
- computational-journalism (journalistic applications)
Notes
Check-worthiness is inherently context-dependent and arguably subjective: what's worth checking depends on audience, stakes, and available resources. The task represents the first filter in fact-checking pipelines, reducing downstream computational load by orders of magnitude.