Definition

Video Authenticity Verification

The canonical definition, the evidence it rests on, and the limits it must carry.

Video Authenticity Verification is the practice of establishing, from evidence, what is actually known about a video's origin, its edit history, and whether any part of it was generated or altered by a machine, then reporting that finding together with its confidence and its limits. It is a claims-and-evidence discipline rather than a verdict machine. It does not answer "is this real" with a yes or a no. It answers a narrower and far more useful question: what can be shown to be true about this specific file, on what evidence, and how much weight will that evidence bear.

The distinction that carries the whole definition is this. Detection asks a model a question. Verification builds a case. A detector takes a file, runs it through a classifier trained on the generators someone thought to train it on, and emits a number. Verification treats the file as one exhibit among several, gathers provenance, context, publication history and forensic reads, weighs them against each other, and produces a finding that another person can inspect, challenge and reproduce. The first is a component. The second is the job.

Four things it is routinely confused with

  • Deepfake detection. A single technique inside the discipline, useful and unreliable in known ways. It is not the discipline, and treating it as the discipline is why so many published verifications collapse when the generator changes.
  • Provenance. Content Credentials record what a cooperating tool chose to record. That is powerful evidence when present and completely silent when absent, which describes most of the internet.
  • Fact checking. Fact checking asks whether a claim is true. Verification asks what is true about the artifact carrying the claim. The two overlap and are not substitutes, and the gap between them is where most real deception lives.
  • Moderation. Whether a video should be shown is a policy question. Whether it is what it purports to be is an evidentiary one. Answering the second with the first is how authenticity work loses its standing.

The four layers of evidence

A verification is only as good as the layers it can actually reach. There are four, and they fail in different directions, which is precisely why a serious finding uses more than one.

1. Provenance

Cryptographically signed manifests attached at capture or at export, standardized by C2PA and surfaced to the public as Content Credentials. When a manifest is present and validates, it is the strongest evidence available, because it is tamper evident and it names the tool and the edit history. Adoption is real and growing: the Content Authenticity Initiative passed 3,700 members in 2024. Adoption is also voluntary, so its absence carries almost no information at all.

2. Watermarking

Signals embedded by the generator into its own output, detectable later by a matching detector. Google DeepMind extended SynthID to video in 2024. A watermark hit is strong positive evidence. A watermark miss tells you only that this particular detector found nothing from the generators it knows, which is a much smaller statement than "not AI".

3. Forensic analysis

Reading the footage itself: compression and re-encoding history, sensor noise patterns, lighting and geometry consistency, temporal continuity across cuts, lip and voice synchrony, artifacts around hands, teeth, hairlines and reflections. This layer requires the file and an analyst, and it is the only layer that can speak to footage carrying no metadata and no watermark. It is also the layer most often performed badly in public, usually by running a color-mapped filter over a JPEG and narrating the pattern. A technique that cannot state its own error behavior is not evidence, it is decoration.

4. The public record

Who published it, what the description and tags disclose, whether a platform label is attached, when the earliest copy appeared, what independent uploads of the same event show from other angles, and what viewers are independently reporting. 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. The decisive property of this layer is that it does not require the cooperation of whoever made the fake, which is exactly why it stays readable when the other three go quiet.

The asymmetry that governs every honest finding

Evidence of synthesis and evidence of authenticity are not equally obtainable, and a verification that treats them as symmetric is wrong before it starts. A validating credential naming a generator, a watermark hit, a creator disclosure: each of these can establish that a video was machine made. Nothing available establishes the reverse. The best synthetic clips are the ones nobody labeled and nobody questioned, so a clean read across all four layers means no signal was found, not no signal exists.

Absence of evidence of manipulation is the weakest finding in the discipline, and it is the one most often reported as a clean bill of health.

The practical consequence is that confidence bands must be asymmetric. A high likelihood can be stated firmly. A low likelihood must be stated as an absence, in words that survive being quoted without the paragraph around them.

The order to work in

Most failed verifications are failures of sequence, not of technique. This order is cheap first, expensive last, and it puts the step people skip at the top.

  • Fix the artifact. Hash the file, record where it was obtained, from whom, and when. A verification that cannot name its exhibit is an opinion about a memory.
  • Find the earliest copy you can. Reverse search keyframes and look for prior uploads. Every repost is a re-encode, and re-encoding is what destroys the evidence in layers one, two and three. Work on the highest generation copy available, and note that the earliest copy you find is a floor on the origin date, never a proof of it.
  • Establish the claim, not just the file. This is the step that decides whether the work was worth doing. Most video deception is authentic footage carrying a false caption: a real protest relabelled with another country, a real speech redated, a real accident attributed to a new cause. A pristine forensic result on a genuine file says nothing about a lie told in the sentence above it. Verify the assertion the video is being used to make.
  • Read provenance, labels and disclosures. Seconds of work, and occasionally the whole answer.
  • Read the public record. Publisher history, corroborating angles, contemporaneous reporting, viewer reports counted rather than characterized.
  • Run detection last, and weight it least. Its output is a prior, not a conclusion, and it should never be the only thing standing behind a published finding.
  • Write the finding with its limits attached. Claim, evidence, confidence band, what was not checkable, and the date. Then it can be reviewed, and reviewed again when the facts move.

How to write the output

A finding that cannot be re-run is not a verification. Four elements make it reproducible: the identified artifact and where it came from, the layers actually reached and the layers that were unavailable, a confidence band in plain language rather than a bare percentage, and a date. The date matters more than it looks. A verification is true as of the evidence available when it was made, and new provenance, a new detector, or an original upload surfacing later can all change it. Findings age. Saying so in the finding is what keeps the discipline honest as its own conclusions get overtaken.

The band language should also refuse the two dishonest endpoints. Nothing reachable by public evidence justifies zero or one hundred percent, and a scale that offers those numbers will eventually be quoted at them.

Where this approach fails

An honest definition names its own weak points, and this one has five.

  • Re-encoding. Platform processing routinely strips provenance manifests and degrades watermarks. The copy that reached you is often the copy least able to be verified, and there is no way to recover what the pipeline discarded.
  • Generalization. Detection models are trained on the generators that existed when they were built. Performance on an unseen generator is not the performance quoted in the paper, and the gap is largest exactly when a new model appears, which is exactly when verification demand spikes.
  • Voluntary adoption. Provenance and watermarking are opt in by design. An adversary simply opts out, and the honest untagged video is indistinguishable, on those two layers, from the deliberately untagged one.
  • Crowds are wrong in both directions. Viewers call real footage fake at least as readily as they call fake footage real. Comment consensus is a signal worth counting and never a verdict worth adopting.
  • The context gap. No layer described here can verify a caption. The most consequential deceptions in circulation need no synthesis at all, and a discipline that only inspects files will keep clearing them.

The standard this site holds itself to

RealOrAiVideo runs the public record layer 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. The frame level forensic layer is human reviewed and offered separately, because no public metadata read can honestly stand in for looking inside the footage. The two are kept apart on purpose, and the methodology page states which signals belong to which.

An authenticity platform that overstates its own accuracy is running the same play as the fabrications it claims to catch. So the rule here is simple and it costs us conversions: report indicators and likelihoods, never guarantees, treat an absent signal as unknown rather than innocent, and say plainly where the public record runs out. Verification earns its 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
  • Content Authenticity Initiative passes 3,700 members, Adobe, 2024. blog.adobe.com
  • 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

Measure your brand's pull.