How a check works

The stages behind every verdict, in plain language.

1. Find the claimsYour text can hold several claims. Each one is pulled out and checked on its own, so one false line does not hide a true one.
2. Search for evidenceThe system searches trusted, credible sources. Official bodies and established outlets count for more than anonymous posts.
3. Weigh itEach source is rated for how trustworthy it is and whether it supports or disputes the claim. A claim backed by one weak source is not treated like one backed by three strong ones.
4. Second lookA separate AI review challenges the result. It looks for gaps and over-confidence.
5. Decide how sure it isIf confidence is high enough, the verdict is final. If not, the claim is held for a human. Nothing is invented to fill the gap.
6. Write it upThe verdict is written up in your house style and language, with its sources and the AI disclosure line.

Held claims

A held claim has status: "not_yet_confirmed" and no verdict. It is in your review queue. Do not show it to readers as a finding. See Your review queue.

Rumours with no evidence

Some claims circulate widely but nothing verifiable exists. These are not called false by default. They are described as unconfirmed, with a clear note that no verifiable evidence was found.

Repeat checks

If a claim has been checked before, the response says previously_checked: true. The new check still runs, because evidence changes.

Why it can be slow

A check searches and reads real sources, then runs several reviews. That can take a few minutes. This is why async mode exists.