For every claim we first look for published fact-checks by recognized fact-checking organizations (ClaimReview). If none exist, we search the web for independent, reliable reporting and official records, and an AI model assesses the claim using only that evidence. The outlet that published the claim never counts as confirmation.
Verdicts: True — the evidence supports the claim. False — the evidence contradicts it. Misleading — partly accurate, exaggerated, or missing important context. Unverified — the evidence is too weak or conflicting to decide; we do not publish these.
Confidence reflects the evidence, not the model’s self-assurance: it is capped by the number of independent publishers (one source caps it at 60%, two at 75%, three or more at up to 95%) and drops when reliable sources disagree.
A fact-check is published automatically only when confidence is at least 85%, at least three independent sources (or two fact-checking organizations) support it, and no evidence conflicts. Everything else is reviewed by an editor first. Automatically published articles are labelled as such.
We re-check claims when they reappear, and when a verdict changes we say so on the article and on our corrections page.