Definition
Video Authenticity Signals
Video Authenticity Signals are the observable indicators that raise or lower the likelihood that a video is what it claims to be. A signal is any readable fact about a video that moves an estimate without settling it: a creator's own disclosure, a platform label, a named generator sitting in the description, a watermark, a compression history, a second upload of the same event from another angle, a count of viewers independently saying the same thing. Each one is an input. None of them is an answer, and a page that treats one of them as an answer is why so much published authenticity work falls apart the moment somebody checks it.
The distinction that carries the whole definition is this. Provenance is what a file says about itself. A signal is what the world says about the file. Content Credentials, standardized by C2PA and surfaced to the public through contentcredentials.org, record what a cooperating tool chose to record at the moment of creation. When a manifest is present and validates, it is the strongest single thing available. It is also voluntary, routinely stripped by platform re-encoding, and absent from very nearly every video anybody actually stops to ask about. Signals are the discipline that has to work when that self report is missing, which is the normal case and not the edge case.
Every signal has three properties, and one of them is usually skipped
Before a signal can be weighed it has to be described, and a complete description has exactly three parts.
- Direction. Which way it pushes the estimate. Most useful signals push one way only. A named generator in the description pushes toward synthetic. Its absence does not push back toward authentic, and reading it as though it does is the single most common error in this work.
- Weight. How far it moves the estimate, which has almost nothing to do with how convincing it feels. A creator's own synthetic content disclosure is quiet and heavy. A commenter writing "this is 100% AI" is loud and nearly weightless.
- Failure mode. The specific, nameable way this signal is wrong. Not a disclaimer, a mechanism: this watermark detector only recognizes its own generator, this artifact disappears at higher bitrates, this comment count can be manufactured for the price of an afternoon. A signal whose failure mode cannot be stated is not a signal, it is a hunch with a percentage attached.
Practitioners tend to write the first two down and carry the third in their heads. The third is the one that decides whether a finding survives being challenged, so it belongs in the record beside the reading it qualifies.
Four families, sorted by who has to cooperate
Signals are usually grouped by technique. Grouping them by who has to participate for the signal to exist at all is more useful, because that is what determines whether an adversary can simply switch the signal off. The governing rule: the value of a signal is inversely proportional to how easily the person who made the video can withhold it.
1. Declared signals: the maker cooperates
Disclosures written by the creator, "made with AI" statements, generator names left in the title, description or tags, and the platform's own altered or synthetic content label. YouTube has required creators to disclose realistic altered or synthetic content since 2024, and the EU AI Act's Article 50 transparency obligations extend disclosure duties to deepfake content on a defined timetable. These signals are cheap to read, high in weight when present, and completely under the control of the one party with a motive to suppress them. They also decay in a way nobody expects: descriptions are editable after upload, so a declared signal read today is a statement about today, not about publication.
2. Embedded signals: the tool cooperates
Marks the generator writes into its own output for later detection. Google DeepMind extended SynthID to video in 2024. These are better than declared signals because they survive an uncooperative uploader, and worse because they require the generator's participation and a matching detector. A hit is strong positive evidence. A miss says only that this detector found nothing from the generators it knows about, which is a far smaller statement than "not AI" and gets quoted as the larger one constantly.
3. Intrinsic signals: nobody cooperates
The footage speaking for itself. Compression and re-encoding history, sensor noise, lighting and geometry consistency, temporal continuity across cuts, synchrony between lips and voice, and the familiar trouble spots at hands, teeth, hairlines, reflections, and the physics of cloth and water. This is the only family that still works against an adversary who opts out of everything else, which is what makes it the backbone of serious forensic work. It is also the most expensive, the most frequently faked in public, usually by running a color mapped filter over a compressed still and narrating the pattern, and the fastest to go stale: an artifact checklist is a snapshot of the model generation it was written against.
4. Circumstantial signals: the record cooperates without being asked
The earliest copy findable, the publisher's history, independent uploads of the same event from other positions, contemporaneous reporting, and viewer reports counted rather than characterized. Individually these are the weakest signals in the set. Collectively they are the most resistant to suppression, because no single actor controls them, and they are the only family that scales to the volume of video actually in circulation. A corroborating second angle from an unrelated uploader is worth more than most forensic reads and costs a fraction as much to obtain.
The families are worth keeping separate because they fail independently. A finding built on one of them is fragile in precisely the way that family is fragile, and the fragility is predictable, which means it is also avoidable.
The asymmetry that governs every honest read
Signals are far better at establishing synthesis than at establishing authenticity, and the two are not close. A validating credential naming a generator, a watermark hit, a creator disclosure: any one of these can show that a video was machine made. Nothing in the set does the reverse. The most convincing synthetic clips in circulation are exactly the ones nobody labeled and nobody questioned, so a clean read across all four families means no signal was found, never no signal exists.
Absence of a signal is the weakest result the method can produce, and it is the one most often reported as a clean bill of health.
Two practical consequences follow. Confidence bands have to be asymmetric: a high likelihood can be stated firmly, a low one has to be written as an absence, in words that survive being quoted with the surrounding paragraph stripped away. And a scale must never offer zero or one hundred percent, because a scale that offers those numbers will eventually be quoted at them.
Signals decay, and the copy you were sent is the worst one
Every repost is a re-encode, and re-encoding is what destroys signal. Platform processing strips provenance manifests, degrades watermarks, and flattens the compression history that intrinsic analysis reads. The copy that reached you has usually been through several pipelines and is the copy least able to be examined, which means the first useful minute of work is not analysis at all: it is finding a higher generation copy, because that changes what evidence exists rather than what you conclude from it. The earliest upload you can locate is a floor on the origin date and never a proof of it. The full order of operations belongs to the practice built on these signals, and it is set out on the verification page.
Signals are adversarial in both directions
Most treatments assume the adversary wants a fake to pass as real. The opposite move is cheaper and increasingly common: getting genuine footage dismissed as fabricated. Robert Chesney and Danielle Citron named this the liar's dividend in the California Law Review in 2019, and it lands directly on this method, because an accusation is a signal that costs nothing to produce. Any weighting that treats "many viewers called it AI" as strong evidence can be manufactured on demand. Count viewer reports, publish the count, and never let them dominate a reading. The same caution applies upstream: metadata is writable, a generator name in a description is a claim rather than a measurement, and a disclosure can be added to real footage as easily as it can be withheld from synthetic footage.
The signal the file cannot give you
Every family above examines an artifact. The most consequential video deception in circulation needs no synthesis at all: authentic footage with a false frame around it, a real protest relabelled with another country, a real speech redated, a real accident attributed to a new cause. All four families come back clean and the video is still a lie. Verify the assertion the video is being used to make, not only the file carrying it. This is the boundary between authenticity work and fact checking, and it is crossed without anyone noticing, in the direction of a pristine forensic result that answered a question nobody was asking.
How to weigh signals without lying with a number
A score is a communication device, and it is honest only under conditions. Six of them, and each one is a place where public scoring usually breaks.
- Deterministic. The same inputs return the same output every time. A number that drifts between runs cannot be argued with, only believed.
- Published weights. If a reader cannot reconstruct why the number is what it is, the transparency is decorative.
- No double counting. A creator disclosure, a platform label, and a generator name in the description are frequently one fact wearing three hats. Summing correlated signals as if they were independent is how a single observation becomes an overwhelming case.
- Bounded away from certainty. No public read justifies zero or one hundred.
- Missing is unknown, never zero. Unmeasured and measured absent are different findings, and rendering the first as the second quietly converts ignorance into reassurance.
- Dated. Signals age. A new detector, a surfaced original, or a credential appearing later can all move the reading, so a finding without a date has no way of being overtaken and therefore no way of being trusted.
Where this framing fails
An honest definition names its own weak points, and this one has four that matter.
- Absence dominates the population. Most video carries no declared and no embedded signal at all, so most real reads run on the two weakest families. The method is at its thinnest on exactly the material people bring to it.
- Intrinsic signals expire. Artifact lists describe the failures of the generators that existed when the list was written. Performance against a model released afterward is not the performance anyone measured, and the gap is widest at the moment a new generator appears, which is the moment demand for reads spikes.
- Suppression is cheap and one sided. The two families with the highest weight are the two an adversary can simply decline to produce. That is not a flaw in the implementation, it is a property of voluntary systems, and no amount of adoption removes it.
- A single number is quoted without its band. Compressing several distinct kinds of uncertainty into one percentage makes the finding portable, and portability is exactly what strips the qualifications off it. The number will travel further than the sentence explaining it, so the sentence has to be built into how the number is written.
The standard this site holds itself to
RealOrAiVideo runs the declared and circumstantial families in the open and publishes the weighting: paste a link and the analysis reads disclosures, named generators, platform labels and viewer reports, and returns a deterministic AI Signal Score with its inputs shown beside it. The intrinsic layer is human reviewed and offered separately, because no public metadata read can honestly stand in for looking inside the footage. The methodology page states which signals belong to which, and what the score will never claim.
An authenticity platform that overstates its own accuracy is running the same play as the fabrications it exists to catch. So the rule here is plain, and it costs conversions: report indicators and likelihoods, never guarantees, treat an absent signal as unknown rather than innocent, and say where the public record runs out. Signals earn their authority by being checkable, which means being willing to be checked.
Analyze a Video How the analysis works
- C2PA, the technical standard for content provenance. c2pa.org
- Content Credentials, the public-facing implementation of C2PA. contentcredentials.org
- SynthID watermarking extended to video, Google DeepMind, 2024. deepmind.google
- Disclosing altered or synthetic content, YouTube, 2024. blog.youtube
- Transparency obligations for providers and deployers of certain AI systems, EU AI Act, Article 50. artificialintelligenceact.eu
- Chesney and Citron, "Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security", California Law Review, 2019, which introduces the liar's dividend. californialawreview.org