What a Low AI Signal Score Does Not Prove
You ran a video, the score came back low, and you want to know whether you can share it. The honest answer is narrower than the one you were hoping for. A low score means no public signal of synthesis was found. It is not a finding that the video is authentic, and treating it as one is the most likely way this tool gets you into trouble.
The reason is structural. Of the four families of Video Authenticity Signals, every one that carries real weight can only fire in one direction. A creator disclosure, a platform label, a generator name left in the description, a watermark hit: any of these can show that a video was machine made. Nothing in the set does the reverse. So a high score is evidence, and a low score is an absence, and those are different kinds of object even though the interface renders them as two ends of one bar.
A low score is the ordinary result, which is why it says so little
Disclosure is voluntary. Watermarking requires the generator to have cooperated and a matching detector to recognize it. Provenance manifests, where they exist at all, are routinely stripped by the re-encoding every platform performs on upload. The predictable consequence is that most video in circulation carries no positive signal whatsoever, whether it is real or fabricated. A result that the large majority of inputs produce cannot, by itself, separate the videos you are worried about from the videos you are not.
Read the score against the size of the record it read
This is the part that changes what you should do next, and it is the thing the number alone will not tell you. Two videos can return the same low score and mean entirely different things.
A thin record. A three day old channel, an empty description, eleven views, four comments, no reuploads anywhere. There was almost nothing to read, so "no signals found" is closer to "no signals were available to find". The informational value here rounds to zero, and a low score on material like this should never be quoted to anybody as reassurance.
A dense record. A long description, an established publisher, a large and active comment section, an upload old enough to have been reposted and argued about. A low score here is worth meaningfully more, because a synthetic clip traveling at that scale tends to accumulate signals whether its maker wants it to or not: a viewer asks, someone identifies the generator, the creator responds, a label gets applied. Silence across a large audience is weak evidence, but it is evidence.
The result page lists its inputs beside the number for exactly this reason. Before you rely on a low score, look at what the analysis actually had in front of it.
A low score has a timestamp, and it expires
Descriptions are editable after upload. Platform labels can be applied days later. An original can surface next week that reframes everything. A read is therefore a statement about the specific copy you submitted, at the minute you submitted it, and it can move in either direction afterward. If a decision of any consequence is resting on a low score from last month, run it again before you act.
Three things you can do in five minutes that the score cannot
- Find an earlier copy. The version that reached you has usually been through several pipelines and is the least examinable copy in existence. Locating a higher generation upload changes what evidence exists, not just what you conclude from it.
- Check whether your doubt is about the footage or the caption. Real footage with a false frame around it, a genuine protest relabelled with another country, a real speech redated, will return a clean read every time and still be a lie. No authenticity score of any kind resolves that, because the file is not the thing in dispute.
- Look for a second angle. If the event was public, someone else had a camera. One corroborating upload from an unrelated account is worth more than most technical reads and costs a fraction as much to obtain.
Where this page's own advice fails
The record density heuristic above is itself gameable. Comments, view counts and reuploads can be manufactured cheaply, so a dense record only raises the value of a low score when the audience behind it is genuinely independent, and you cannot always tell. Worse, the heuristic collapses entirely in the one case that matters most: a newly uploaded clip of a real event has no record yet, and neither does a fresh fabrication. For the first hours, when a video spreads fastest and the pressure to say something is highest, a genuine video and a synthetic one are equally unreadable by this method. That is not a limitation we expect to engineer away. It is what a public signal read is, and saying so is cheaper than being caught claiming otherwise later.
What genuinely moves a low score is looking inside the footage: the voice, the frames, the edits, none of which leave a trace in public metadata. That is the forensic layer, and its limits are stated separately. If you need a firmer answer than the public record can give, ask for one rather than reading certainty into an absence.