Methodology
How the analysis works.
RealOrAiVideo does not look at pixels. It does not run a secret model over the footage and hand you a verdict you have to take on faith. It does something quieter and more honest: it reads the public signals that surround a video — the words the creator chose, the labels they applied, and what thousands of viewers said in the comments — and turns those observable signals into a single, transparent number: an AI Signal Score.
That score is a read of the public record, not a forensic determination — a starting point for judgment, not a substitute for it.
Two layers, kept separate
Authenticity work has two distinct layers, and we are deliberate about not blurring them:
- The signal layer (what this tool does today). Reading the public record — disclosures, metadata, and viewer reports — to produce an AI Signal Score. This is what the YouTube data can honestly support.
- The forensic layer (human-reviewed). Determining whether the footage itself is synthetic — voice, frames, edits. No metadata API can do this; it requires analysis and expert review. That is the deeper analysis we offer separately.
A platform that pretends the first layer is the second is not an authenticity tool — it is the very kind of overclaim we exist to push back on.
What we read
The free analysis evaluates the signals that are publicly observable around any YouTube video:
- AI generation indicators. Named AI video, voice, and animation generators referenced in the title, description, or tags — Sora, Runway, Kling, Veo, Synthesia, HeyGen, ElevenLabs, and others.
- Creator disclosures. The video's own "AI-generated," "made with AI," or YouTube "altered or synthetic content" labels. When a creator tells you, we listen.
- Viewer signals. How many of the top comments independently call a clip AI, fake, deepfake, or manipulated. A crowd is not a verdict, but a strong consensus is a signal worth surfacing.
- Metadata & source. The channel behind the clip, the publication context, and the provenance limits of the platform itself.
The parts that require looking inside the footage — voice consistency, visual artifacts, editing anomalies — are reserved for a human-reviewed forensic analysis. We do not fake those signals with a public-data tool.
How the score is built
The score is a transparent, weighted sum of observable signals, not a black box. A creator's own synthetic-content disclosure carries the most weight. A named AI tool in the description carries substantial weight. The proportion of viewers flagging the clip adds to it; a clear consensus that the footage is real subtracts from it, gently. There is a small baseline, because in the synthetic-media era the absence of a label is not evidence of authenticity.
Two properties matter most. First, it is deterministic: the same video returns the same percentage every time, because it is computed from stable, observable inputs — never a random guess. Second, it is bounded by honesty: the score never reaches 0% or 100%, because no public-signal read can rule AI fully in or fully out.
A high number means the signals point toward AI. A low number means those signals are absent — which is not the same thing as proof a video is real.
What we cannot claim
This is the part most tools skip. We will not.
- We never claim certainty. No "100% AI," no "guaranteed fake," no "perfect accuracy." Those claims are dishonest, and they are exactly what an authenticity platform should refuse to make.
- Absence of signals is not innocence. The most convincing synthetic clips are the ones nobody labeled and nobody questioned. A low likelihood means we found nothing — not that there is nothing to find.
- Crowds are wrong in both directions. Viewers call real footage fake and fake footage real. We weight comment signals carefully and show you the count, so you can judge for yourself.
- Platforms strip provenance. Content Credentials (C2PA) and watermarks are increasingly attached at the source, but YouTube re-encoding often removes them. We tell you when provenance cannot be read, rather than pretending it was authentic.
When a clip genuinely matters — a piece of evidence, a viral claim, a reputational risk — the responsible next step is a frame-level forensic review by an analyst. That is the deeper analysis we offer, and it is reviewed by people, not asserted by a script.