How to read Unvail results

It is not "how AI is this". It is how confident we are in the verdict beside it, so the number always moves in the same direction as the word: higher means more sure, whichever way the answer went.

The cutoff is the halfway point. Above it we say likely AI, below it likely human — and a result near the middle is genuinely uncertain, not a quiet AI verdict.

Scores are bounded to 1–99, everywhere, on purpose. Certainty is not something a detector can offer: a 99% score and a 1% score both still need your judgement. A round 100% would tell you we had stopped estimating and started asserting, and we would rather not be able to say it than rely on remembering not to.

We sample up to twelve frames as a video plays, not the whole runtime, and take the middle value rather than the worst one — so a couple of odd frames cannot swing an otherwise clean video, and a consistently generated one is not rescued by a few clean frames.

That means a green tick is never a statement about the entire video. Hover the tick after a scan and it tells you exactly which stretch was checked, and how many frames it used. If the part you care about was not covered, scan that part separately.

A track scan listens to about thirty seconds, not the whole song. The result says which seconds it covered — worth checking before reading it as a judgement on a four-minute recording, since a track can change character partway through.

Short passages, edited drafts and text written in a plain factual register are all harder to judge, and a human writer with a spare, consistent style can read as generated. Treat a text result with more caution than an image or a video, and never as the sole basis for a decision about a person.

A text result also shows what the text is likely made of: how much reads as AI, as AI-assisted and as human, and how many passages the engine read. A short text is often read as one passage, so the whole of it shares one verdict. When a passage looks like AI writing that was reworded to get past detectors, the result says so and gives the strength of that signal. That is a suspicion, not proof.

Detection asks how something was made. A claim check asks whether what it says holds up: we take the statements in an article or a video, search for reporting on each one, open the pages we find, and look for the sentence itself. What you get back is a description of what reporting exists — not a ruling on what is true.

The card opens with a rating out of four. That is the share of claims that stood up, in quarters, capped by how the headline itself held up — a piece whose headline the sources dispute cannot score well on the strength of a good claim list. Underneath, every claim carries its own word:

Each claim lists the sources behind it and says what we could confirm on each page. Where a page could not be opened we say so, rather than quietly counting it as agreement you can check.

On a YouTube video, every claim shows the moment it was said — click the time to jump there — and the card says which part of the video was read or watched. A passage you highlighted is checked on its own, so its rating describes that passage, not the whole page.

Some numbers change week to week: win rates, rankings, prices, follower counts. When no source covers the same period, the claim is tagged Live figure and left out of the rating instead of being counted against the piece. A card made only of them says Not rated.

Corroboration is not truth. A claim every outlet repeats can still be wrong, and one nobody has covered can still be right. Read the sources — they are on the card for that reason — and never treat a claim check as a finding about a person.

If something has already been checked — by you or by anyone else — you get that answer instantly, free, and it does not count against your limits. A cached result is the same result: it was produced by a real scan of the same content.