Fact-checking
Last updated: July 26, 2026
Fact-checking is testing individual claims against independent evidence and returning a verdict for each. Automated fact-checking does this at scale: it extracts each claim from an answer, grounds it against sources, and labels it supported, contradicted, disputed, or unverifiable — rather than judging the whole text at once.
What is automated fact-checking?
Fact-checking is the process of testing whether specific claims are true by comparing them against independent evidence, and reporting a verdict for each. The unit is the claim, not the document — you don't grade an essay, you check its statements one by one.
Automated fact-checking applies this at machine scale. It extracts atomic claims from an answer, grounds each against independent sources, runs a judge over what the sources return, and assigns each claim a verdict. The output is a per-claim map of what held up and what didn't.
This claim-level approach is what makes verification actionable. A single 'this looks wrong' about a paragraph tells you little; a verdict on each sentence tells you exactly what to fix.
How claim-level fact-checking works
A modern automated fact-checker runs a pipeline:
- Extract the answer into atomic, individually checkable claims.
- Ground each claim against independent channels — retrieval, live web, structured knowledge like Wikidata and Wikipedia, and tools.
- Judge the retrieved evidence with a tiered LLM judge to decide what it supports.
- Verdict each claim: supported, contradicted, disputed, or unverifiable, and roll up a risk score.
The honesty rule throughout: when the evidence is merely absent, the claim is unverifiable, not contradicted — which keeps the false-alarm rate low. See the full how-it-works page.
A worked example
An AI answer says: "The Great Wall of China is visible from space with the naked eye, and it was completed in 1644." Fact-checking splits this into two claims. The first — visible from space — is grounded against independent sources and returned contradicted (a widely debunked myth). The second — completion date — is checked and returned disputed, since the wall was built and rebuilt over centuries with no single agreed completion.
One sentence, two claims, two different verdicts. That granularity is the value of claim-level fact-checking. Try it on the verifier.
Frequently asked questions
How is fact-checking different from an AI content detector?
A fact-checker tests whether claims are true. An AI content detector guesses whether text was machine-written. Those are unrelated: human text can be false, AI text can be accurate. Detecting authorship says nothing about truth. See AI detector vs fact-checker.
Can automated fact-checking be wrong?
Yes. A verifier can miss an error or flag a true statement, so its own precision and recall must be measured and reported. The honest design returns unverifiable rather than guessing, which keeps false alarms low. See how accurate it is.
Why check claims individually instead of the whole answer?
Because real answers are mixed — mostly right with one bad claim. Judging the whole text hides which part failed. Checking each claim tells you exactly what is contradicted, disputed, or unverifiable, so you can fix the one thing rather than distrust everything.
Verify what your AI just told you.
Paste any AI answer and Mephistopheles checks each claim against independent sources — no sign-up to try.