How to Verify a Video Is Real in the AI Era (2026)

In 2026, a convincing thirty-second video takes one sentence to generate. The reflex that carried us for a century — "I saw it on video, so it happened" — is gone. And the problem is no longer only about spotting fakes: it is now equally about proving that your own video is authentic when it has to stand up in front of an insurer, a landlord or a court. This guide covers what actually works to verify a video's authenticity, what no longer works, and what to do when you are the one filming.

Why the human eye is no longer enough

The tells popularised in 2019 — six-fingered hands, missing blinks, smeared face edges — belong to a generation of models that is now obsolete. The measurements available on this are unambiguous.

In its Deepfake Blindspot Study, biometric verification company iProov showed 2,000 UK and US consumers a series of authentic and manipulated items. Only 0.1% of participants classified every item correctly — even after being told to look for fakes. Two findings matter directly for video:

Participants were 36% less likely to spot fake videos than fake images. Motion, sound and duration saturate attention: you watch the action, not the artefacts.
Over 60% of people were confident in their ability to spot a fake — regardless of how they actually scored. Overconfidence is the real risk factor.

Put plainly: if your verification method amounts to "look closely", its value is close to zero. You have to move to technical signals.

Écran d'ordinateur affichant des fenêtres d'inspection de fichier et de métadonnées — Verifying a video starts with inspecting what it carries: container, metadata, provenance manifest.

Three families of usable signals

There are three categories of technical signal today, and they answer different questions. Conflating them is the number one source of error.

1. Signed provenance (C2PA / Content Credentials)

The C2PA standard (Coalition for Content Provenance and Authenticity) attaches a cryptographically signed manifest to the file: which tool produced the content, when, and what edits followed. This is what the public knows as Content Credentials. The manifest is publicly checkable, notably through the Content Authenticity Initiative's Verify tool.

One point is crucial and consistently misunderstood: C2PA does not detect deepfakes. It does not answer "is this content true?" but "what signed history does this file carry?". A fully AI-generated video can carry a perfectly valid C2PA manifest — one that says, precisely, that it was AI-generated. It is a declaration of origin, not a verdict on truth.

Its practical weakness lies elsewhere: platforms strip metadata on upload at scale. A video reshared on social media, re-encoded, cropped or passed through a messaging app usually arrives bare. So the absence of Content Credentials proves nothing at all. To go further, the full specification is published by the C2PA.

2. Generation watermarks (SynthID and peers)

SynthID, built by Google DeepMind, embeds an invisible watermark at generation time into text, images, audio and video produced by Google models, including Veo. Unlike an overlaid label, it survives cropping, compression and format changes, and is checked through the SynthID Detector.

Its limitation is structural: SynthID only recognises what Google models produced. A video from another ecosystem will not show up — not because it is authentic, but because the watermark was never applied. We break this down in our analysis of SynthID.

At OpenAI the logic is twofold and worth understanding: videos generated with Sora carry both a visible watermark and C2PA metadata. OpenAI states explicitly that the visible watermark is not the machine-readable provenance signal — it is a cue for the eye, trivially cropped away. The usable signal is the manifest.

3. Statistical detectors

Finally there are the detectors that analyse the signal itself: temporal inconsistencies, compression artefacts, facial micro-movements, lighting coherence. They are useful, but none returns a reliable binary answer, and performance collapses as soon as they meet a generator model newer than their training set. A score of "87% likely AI" is an indication, never proof. The same caution applies to still images, which we documented in how to detect an AI-generated photo.

Verifying a video in practice

When a video's authenticity matters, here is an effective order of operations — from most conclusive to most fragile.

Trace the source. By far the highest-yield step, and the most neglected. Who published the video first, and when? An account created last week, with no history, posting "exclusive" footage of a major event: whether it was AI-generated is almost a secondary question.
Look for earlier versions. Extract two or three key frames and run them through reverse image search — one of our 7 tools and methods for verifying a photo. This is how you catch the most common case — far more common than deepfakes: a genuine video taken out of context, presented as recent when it is three years old and from another country.
Inspect the original file. If you have the file rather than a link, examine its metadata and any C2PA manifest. Always ask for the original, never a screen recording or a messaging-app forward, both of which destroy the information. The principle mirrors still photos, covered in our guide to EXIF metadata.
Check what is checkable inside the frame. The setting, the weather on the claimed date, the language on signage, the position of the sun, licence plates. A generated video invents a plausible world, rarely an exact one.
Run detectors — last, and with reservations. Their output adds to the body of evidence; it does not replace it.

And one principle running through all of it: the absence of a signal is not a signal. No Content Credentials, no SynthID, no metadata — the correct conclusion is "I don't know", not "it's authentic".

Personne filmant une scène avec son smartphone tenu à bout de bras — Most videos that end up as evidence are shot on a phone, in a hurry — that is the moment everything is decided.

The inverse problem: proving YOUR video is real

Everything above concerns someone else's video. But the fastest-growing use case is the mirror image, and far more concrete: you film a water leak, the state of a flat at handover, a damaged delivery, defective work on a building site — and six months later you have to convince an insurer, a landlord or a judge that the footage was neither retouched nor fabricated.

That problem is structurally different, and much easier to solve. Detecting a fake after the fact is a race you lose against models improving faster than detectors. Sealing the real thing at the moment it is created is a problem cryptography solved long ago.

This is the principle behind certification at capture: instead of analysing after the fact, everything is frozen at the moment of the shot — the file's cryptographic hash, the timestamp, the location, the device. Any later modification, down to a single pixel, breaks the seal. The claim is not "this video was not AI-generated": it is "this exact file has existed in this exact form since this exact date, and nobody has altered it since".

At Truth-Check, video certification works like this: the raw video is never kept on our servers. We extract a sequence of frames at a regular interval, sealed into a single animated file, plus a transcript of the audio. It is that reduced content — publicly checkable via a short code — that is certified, timestamped and produceable. It then sits alongside your other items in a chronological evidence file, next to the documents, emails and photos tied to the same dispute.

Main tenant un iPhone dont l'application appareil photo est ouverte — Certifying at capture: the hash, the time and the place are sealed before the file ever circulates.

What EU law changes in 2026

Since 2 August 2026, the transparency obligations of Article 50 of the EU AI Act apply. Two points bear directly on video.

First, providers of systems generating synthetic content must ensure their outputs are marked in a machine-readable format and detectable as artificially generated. Second, anyone deploying an AI system that produces a deepfake — defined in Article 3(60) as AI-generated or manipulated image, audio or video content resembling existing persons, objects, places or events and that would falsely appear authentic — must disclose it, clearly and perceptibly, at the latest on first exposure. The obligation applies even without intent to deceive.

Two caveats, though. A Commission "omnibus" proposal contemplates targeted relief for some of the Article 50(2) marking obligations; as of this article, the final timetable is not settled. And more fundamentally, a legal obligation does not mechanically produce marking on content circulating outside the Union, nor on output from open models run locally. On what generative AI is already doing to image authenticity, see generative AI and visual authenticity.

Intérieur d'une salle d'audience vue en contre-plongée — In front of a judge, what counts is not a detector score but a set of dated, consistent items.

And in court?

Under French law a video is not "valid" or "invalid" evidence in itself: it is material submitted to the judge's sovereign assessment. What weighs is the coherence of the whole — where the file came from, its date, its integrity, how it squares with the rest of the file. An isolated video with no verifiable date, produced by the party relying on it, carries little weight. The same video timestamped at capture, with a checkable hash, set inside a documented timeline alongside letters and reports, becomes solid. We develop this in digital evidence in court.

FAQ

Is there a reliable tool to tell whether a video is AI-generated?

No — no tool gives a reliable binary answer today. Provenance checks (C2PA, SynthID) are conclusive when the signal is present, but their absence proves nothing: platforms strip metadata, and each watermark only covers its own ecosystem. Statistical detectors give probabilities, not verdicts.

Is a video without Content Credentials suspicious?

Not at all. The vast majority of authentic videos carry none, either because the device does not generate them or because a platform stripped them on re-encode. A missing manifest means "information unavailable", nothing more.

Does a visible "Sora" or "veo" watermark guarantee a video's origin?

No, in both directions. It can be cropped out in seconds, and it can just as easily be added to a video that did not come from there. OpenAI states explicitly that the visible watermark is not the machine-readable provenance signal: the usable signal is the embedded C2PA metadata.

Does Truth-Check detect deepfakes?

No, and that is deliberate. Truth-Check does not analyse third-party videos and makes no claim about whether content is AI-generated. It does the opposite: it certifies your captures at the moment they are taken, so you can prove their authenticity and their date. Detecting fakes and proving what is real are two distinct problems; only the second has a solution that does not expire.

Does Truth-Check keep the original video?

No. The raw video is never stored on our servers: only an extracted frame sequence and an audio transcript are certified and kept, bound to the hash and timestamp of the capture.