How to tell if a video is AI-generated, and where the checks fail

There is no single test that settles whether a video is real, and the people who need the answer most, courts, newsrooms and platforms, are now working from three different checks that catch three different kinds of fake.

The first check is provenance, and it is the only one that can prove something rather than estimate it. C2PA, the Coalition for Content Provenance and Authenticity, publishes an open technical standard that records a file's origin and its edits in signed metadata, the kind of data cameras, editing apps and AI generators attach as Content Credentials. Where a credential is present and verifies, it says which device or tool produced the file and what was done to it afterwards.

The limit is structural. A credential only exists if the thing that made the file chose to attach one, so its absence proves nothing. A video with no Content Credentials may be an untouched recording from a phone that does not sign its files, or a fabrication from a tool that strips metadata on export. C2PA's own framing is about establishing origin and edits, not about detecting fakes.

The second check is a detector model, and the research on it is more careful than the marketing around it. ReStraV, a method published at NeurIPS 2025, does not look for visual glitches. It measures how straight a video's trajectory looks when its frames are projected into the representation space of a pretrained vision transformer, on the theory that real footage moves through that space in straighter lines than generated footage does. The authors report 97.17% accuracy and 98.63% AUROC on the VidProM benchmark, which is a strong result on one benchmark and explicitly framed as an answer to the generalization problem: detectors trained on one generator's output tend to fall over on another's.

That problem is the whole difficulty. An ICLR 2026 paper built a dataset of more than 140,000 videos from 15 open-source and commercial generators precisely because older detection sets no longer resemble what current models produce. Its finding is a warning about how these tools are used: the resizing and cropping that most pipelines apply before analysis discard the high-frequency artifacts and spatiotemporal inconsistencies that detection depends on. A detector fed a re-encoded social-media copy of a clip is working on a different problem from the one it was trained for.

A survey of 221 detection papers published in the Findings of ACL 2026 reaches the same conclusion from the other direction. Its authors write that the evolving realism of AI-generated video "is rapidly rendering traditional artifact-centric detection insufficient, necessitating a paradigm shift from low-level inspection to high-level semantic verification." In plain terms: stop looking for a warped hand and start asking whether the event shown could have happened and whether anyone else recorded it.

That is where the third check lives, and courts have started to say what it requires. On September 30, 2026, the Arizona Court of Appeals vacated a manslaughter sentence because the sentencing judge had heard and relied on an AI-generated video of the dead victim, scripted by the victim's sister and built from his photograph and a voice profile. The court did not ban AI from the courtroom. It applied a reliability standard: evidence used to decide a sentence has to have at least minimal indicia of reliability, and a video that presents someone's guess at what a dead man would have said, wearing his face and voice, does not.

The part worth carrying out of that opinion is the court's treatment of disclosure. The video opened by telling the court it was an AI recreation, and the panel still held that the disclaimer did not fix it: "No disclaimer could cure that error." A label says what a file is. It does not make the content true.

Put together, the practical order of checks is the reverse of the intuitive one. Look for a verifiable credential first, because it is the only check that can settle the question. Treat a detector's score as a lead, not a verdict, and never run one on a resized or re-encoded copy if you can get the original. Then go looking for the source: who published it first, whether anyone else was present with a camera, whether the event it depicts has any other record. That last step is the one that actually works, and it is the same test a court applied when a video of a dead man walked into a sentencing hearing.

The same month produced a reminder that the verification can fail in the other direction. On October 21, 2026, The Portland Focal reported that the Oregon Journalism Project retracted its report on a viral clip of a Washington County commissioner, after questions arose about the image it had published to show the clip had been altered. The clip's authenticity was never independently established in either direction. A deepfake claim is a claim, and it needs the same sourcing as the video it accuses.

For creators, the practical version is a habit: keep the original file and its metadata, export with Content Credentials where the tool supports it, and treat any clip you did not record yourself as unverified until you have found who did.