How FactWatch Works

Automated agents scan trusted public sources — political news, statements from elected officials, campaign claims, speeches, government announcements, and congressional statements — three times per day. We prefer APIs, RSS feeds, and permitted public sources. We never bypass paywalls, security protections, or robots.txt rules.

Multiple specialized AI agents work together: a News Scout finds stories, a Claim Extractor separates factual claims from opinions and satire, a Researcher gathers evidence, a Source Evaluator judges quality and independence, a Fact Checker assigns a verdict, a Context Analyst finds omitted context, a Citation Validator confirms citations support their statements, an Editor writes the article, a Duplicate Detector avoids repeats, and a Quality Controller does a final check. Any agent can flag a case for human review rather than inventing an answer.

We prioritize primary and authoritative sources: federal and state agencies, Congressional records, court documents, election agencies, BLS, Census Bureau, BEA, CBO, FEC, DOJ, DHS, CDC, FDA, the Federal Reserve, official legislation, executive orders, and official transcripts. Secondary sources are reputable journalism and expert institutions. We never treat a single news story as definitive proof when primary evidence is available, and we aim for at least three independent sources.

Verdicts use eleven categories — True, Mostly True, Misleading, Needs Context, Mostly False, False, Unverified, Insufficient Evidence, Opinion / Not Fact-Checkable, Satire, and Developing. We do not force every claim into True or False. If evidence is conflicting or incomplete, we use Unverified or Insufficient Evidence. The AI never fabricates certainty.

Every fact check includes a confidence score from 0–100% based on source quality, number of independent sources, primary-source availability, agreement between sources, recency of evidence, whether statistics can be independently verified, and whether authoritative records exist. We do not show high confidence when evidence is weak.

Fact checks go to human review when confidence is below the publishing threshold, evidence conflicts, sources are insufficient, the claim is high-risk (criminal accusations, election fraud, corruption, violence, sexual misconduct, national security), or wording is ambiguous. Admins can publish, edit, reject, request more research, change verdicts, or add and remove sources.

Every article has a Report an Error link. When a verdict changes, we never silently overwrite it — we show the correction date, original verdict, updated verdict, the reason, and the evidence that caused the change. Full revision history is maintained.

During rapidly developing stories we use Developing or Unverified rather than a definitive verdict, and we revisit the story in later crawler cycles. We never prematurely assign certainty.