AI Fact Checker: Verify Claims & Cite Sources

Paste a draft or point it at a published article. Every factual claim comes back with a verdict, a 0-100 score, and the source URL behind it — not a confidence number with nothing under it.

Built for the moment before you publish, when the text reads fine and you have no idea whether the numbers in it are real.

What it does

1. Reads every sentence. Statistics, quotes, named people, dated events, and legal, medical, or financial claims are pulled out. Opinion, advice, and predictions are skipped, because no public source could settle them.
2. Searches twice per claim. Once for the claim as written, and once for what is actually the case, phrased so it can come back negative. Searching only your own wording finds whatever agrees with it.
3. Weighs the sources. Encyclopedias, journals, and government records outrank news, which outranks forums and social posts.
4. Never asserts without a source. If nothing can be found, it says so. It will not call a claim false because the search failed — those are different answers, and confusing them is how a fact checker loses your trust.

Example output

Verdict	Score	Claim	Source
True	98	The Berlin Wall fell in 1989.	en.wikipedia.org
False	95	Microsoft acquired Figma for $22 billion in 2024.	cnbc.com
False	90	Nvidia was founded in 1993 by Jensen Hwang.	en.wikipedia.org

That third row is the one that matters. The date is right, the company is right, and the founder's name is misspelled — so the spellingFix field says sources write "Huang", not "Hwang". Search silently corrects spelling, which means a spellchecker and an AI reviewer both wave it through.

Who uses this

Content and SEO agencies — audit an article before it ships, or run a client's back catalogue and find the claims that have quietly gone stale.

Anyone publishing AI-assisted copy — language models produce fluent, specific, wrong numbers. A plausible statistic with a fabricated source is the failure mode that survives every other review step.

Newsrooms and researchers — check a press release or a submitted piece against the record before it runs.

Developers — put it in a publishing pipeline. One claim per dataset row, ready to gate a deploy on.

Input

Field	What it does
text	Paste a draft, article, or press release.
startUrls	Article URLs to fetch and check. The readable body is extracted; navigation and boilerplate are dropped.
maxCharsPerDocument	Truncates long documents. Lower it to control cost, since billing is per claim.
onlyProblems	Return only claims that are contradicted, unresolved, or misspelled. You are billed only for rows returned, so this is also the cheapest way to run it.

Output

One row per claim:

claim, verdictLabel (True / Probably true / Unclear / Disputed / Probably false / False), likelihood (0-100), verdictStrength, explanation, sourceUrl, sourceTitle, citations, needsAttention, spellingFix, disputed, noTrace, source.

likelihood is always "how likely this is true". verdictStrength restates it as the strength of the verdict, so a claim scored 5 shows as False 95 rather than True 5.

Pricing

Pay per event. $0.005 per run plus $0.015 per claim verified — $15 per 1,000 claims.

You are charged per claim rather than per document because a document with thirty claims costs thirty times what a document with one costs. Claims that fail to check are never charged. With onlyProblems enabled you pay only for the claims that need your attention.

A typical 1,000-word article yields 8 to 15 claims, so roughly $0.12 to $0.23 per article.

FAQ

How do I fact check an article automatically? Give the URL in startUrls. The readable body is extracted and every factual claim in it is verified against live web sources, one dataset row each.

Can it fact check AI-generated content? That is what it is for. Language models are fluent and specific about things that are not true, and the resulting sentences pass every stylistic check. This verifies them against sources instead.

Does it detect AI writing? No, and deliberately. AI detection guesses at how text was produced; this checks whether what the text says is true. A human can write a false sentence and a model can write a true one.

What happens when nothing can be found? The claim comes back Unclear, with noTrace set when the surrounding subject is well documented and the claim itself appears nowhere. That combination is the strongest signal of a fabricated fact, but it is still reported as unverified rather than false.

What if credible sources disagree? The claim is marked Disputed and the citations field carries sources on both sides, so you can see who says what and decide with your name on it.

Why is my true claim marked Unclear? Usually because the sources hedge where the claim does not, or because reputable sources genuinely differ. Check the citations field — for a contested claim, that disagreement is the answer.

Does it store my text? No. Only a hash, which is what makes re-checking the same text free.

---

Powered by the Claim Check verification API.