How this works

The mechanics behind every verdict on claimpit — and how to dispute one.

What claimpit is

claimpit fact-checks YouTube videos, podcasts, and articles. Every checked source gets a public page listing the individual claims made in it. Every claim carries a verdict, the verbatim quote it was drawn from, and cited sources — each with the quoted sentence that bears on the claim and a stance: Supports, Refutes, or Background (a real, relevant page that is silent on the claim itself).

Checks are made by an AI model

The checker is an AI model, not a human editorial desk. It reads the transcript or article, extracts the individual claims, and judges each one against live web sources it retrieves at check time. That is worth being direct about, and it is why every check page says “Checked by an AI model against live web sources” under its title.

Machinery exists specifically to keep the model honest:

The five verdicts

Accurate

Supported by the evidence: the page cites sources and quotes the line that backs the claim.

False

Refuted by the evidence: the page names what makes the claim false and quotes a refuting source. A sweeping claim with a well-defined test ("every X…") can be False on a single counterexample.

Misleading

The claim creates a false impression by leaving out a key fact. This is the most accusatory verdict, so it carries the highest bar: the page must name the omission and quote a source carrying it, or the verdict is not published.

Unverifiable

Couldn't be settled with available evidence — or isn't precise enough to test (a superlative with no defined measure, for example). Unverifiable means "nothing to weigh", not "probably wrong": it never counts against a speaker's accuracy record.

Contested

The terms are defined and the evidence is abundant, but named institutions have publicly reached opposing conclusions and no body with jurisdiction has ruled. This verdict is being phased in: a claim that earns it is currently recorded as Unverifiable, with a note on the claim naming both sides. Like Unverifiable, it never counts against a speaker.

What the confidence number means

The percentage beside a verdict is the model's own confidence in that verdict — not a calibrated probability that it is correct. A hedged or twice-checked-and-disagreed claim has its confidence capped at 50%. Treat it as the checker's self-report, and treat the quoted sources — which you can open and read — as the actual evidence.

What this can get wrong

AI checks fail in AI ways: a claim can be extracted more strongly than it was spoken (which is why the verbatim quote is shown beside the model's restatement), a source can be misread, and a verdict can simply be wrong. The gates above reduce this; they do not eliminate it. If you find an error, say so — see below. Corrections are the point of a receipts site.

Who runs this

claimpit is an independent, single-operator project — not affiliated with any publisher, party, or platform, and funded by nobody. The operator reviews attributions and curates topics; verdicts come from the pipeline described above. Contact: hello@claimpit.com.

Dispute a verdict, or ask for removal

Email hello@claimpit.com with a link to the claim. If you can point at evidence the check missed, the check is re-run or corrected. If you are a person named on this site and want a page corrected or removed, the same address works and removal requests are honored without argument — as the privacy policy already promises.

← claimpit