Free · runs in your browser
A practical order of checks, from provenance data to the source of the clip, with an honest note on what each one can and cannot prove.
There is no reliable one-click test that tells you whether a video is AI-generated. The strongest evidence is provenance: a signed Content Credentials (C2PA) manifest or a maker’s watermark that says an AI tool made the clip; after that come the file’s other metadata, what you can see and hear, and above all where the video came from.
Each of those checks can confirm an AI origin when it finds something, and none of them can prove a clip is real when it finds nothing. Most videos you meet online have been re-uploaded, and platforms strip metadata as a matter of routine, so a genuine phone clip and an AI clip downloaded from social media often look identical at the file level.
This site can help with the metadata part. Open a video here and the Provenance section shows any Content Credentials, the AI-generation assertion if there is one, encoder and software tags, and signs of re-saving. It reads the file in your browser and uploads nothing. It does not judge pixels, does not verify signatures and cannot read invisible watermarks; this guide says where to go for those.
Work through these in order. Stop when you have strong evidence either way, and treat “nothing found” as “unknown”.
Trace the clip back to whoever posted it first, and get that file if you can. A copy of a copy has lost most of its metadata; the original upload, or the file from the person who recorded it, is worth far more.
Open the file at contentcredentials.org/verify, in the Gemini app, or in Google Photos under More → About (“How this was made”). A valid manifest that names an AI tool, or says the content was created by a generative model, is the strongest evidence available.
For Google AI tools, ask the Gemini app whether the video was made with Google AI; it checks for the SynthID watermark in clips under 100 MB and 90 seconds. Some generators also add a visible watermark or logo that crops and re-uploads may have removed.
Check for camera make and model, location, recording dates and the encoder tag. A camera original with consistent device data is a point in its favour; an editor or generic encoder tag only shows the file was re-saved.
Look for physical and continuity errors, listen for odd audio, and compare the claims made about the clip with other reporting of the same event. Weigh everything together rather than trusting any single sign.
| Check | If it finds something | If it finds nothing | How easy to defeat |
|---|---|---|---|
| Content Credentials (C2PA) naming an AI tool | Strong evidence, once the signature verifies | Nothing: most real and AI clips carry none | Easy to strip; very hard to forge a valid signature |
| Invisible watermark (such as SynthID) | Strong evidence for that vendor’s tools | Nothing about other vendors’ tools | Designed to survive compression; may fail after heavy edits |
| Visible logo or watermark | Suggestive; logos can be added to anything | Nothing; it may have been cropped | Very easy |
| Camera make, model, location, dates | Weak evidence of a real camera; tags can be typed in | Nothing; uploads strip them | Easy, with any metadata editor |
| Encoder or software tag | Shows which app last saved it | Nothing | Easy |
| Visual and audio tells | Suggestive, sometimes conclusive (impossible physics) | Nothing; good generators avoid obvious errors | Improving every few months |
| “AI detector” websites | A probability score of unclear accuracy | A probability score of unclear accuracy | Compression and edits move the score |
| Source and context | Often decisive: an original uploader, other angles, news reports | Unverified, not disproved | Hard to fake convincingly across many sources |
No metadata means no conclusion. A clip with no camera data and no credentials is exactly what both a re-uploaded phone video and a re-uploaded AI video look like.
Content Credentials are a signed record inside the file that says how it was made. Several AI video tools now attach them by default, and a few cameras sign what they record, so this is the check most likely to give a firm answer when it finds something.
These are the vendors’ own statements, checked on 7 November 2026. Open-source and self-hosted video models usually add nothing at all, so the absence of credentials is common for AI clips too.
The Content Credentials viewer here reads the manifest and explains it, including the digitalSourceType, where trainedAlgorithmicMedia means “created by a generative AI model”. It does not check the signature, so a manifest could in principle be forged. For that, use contentcredentials.org/verify or the Gemini app, which verify signatures against a trust list.
Without credentials, the ordinary metadata gives weaker hints. It is still worth a look, mainly because a genuine camera original tends to be rich and internally consistent, and because the encoder tag tells you what last touched the file.
| What you see | What it suggests | Why it is not proof |
|---|---|---|
| Make, model, lens and software of a phone, with a location and a local time zone | Probably recorded on that device | All of it can be typed in with a metadata editor |
| Telemetry tracks from a drone or action camera | Very likely a real recording from that kind of device | Rarely faked, but rarely present after sharing |
An editing app or a generic library such as Lavf in the encoder field | The file was re-saved or converted | Every platform and most editors re-save real footage too |
| No device data, a round duration, no audio track | Consistent with a generator’s output | Also consistent with any re-upload or screen recording |
| Dates that disagree, or a 1904 or 1970 default | The file was rewritten by software | Common on innocent files sent through apps |
The video forensics guide goes deeper into dates, encoder tags and re-encoding signs. The file info report lists every field, and where was this video taken shows any location on a map.
Generated video has improved fast, and the classic tells (extra fingers, melting faces) are much rarer than they were. Look for errors that come from the model not understanding the world:
On this site, Frames shows a filmstrip and Save frame keeps a still as a PNG, which helps you study details and run a reverse image search. Find scene changes in AI insights marks the cuts. Those tools describe what is in the video; they do not detect AI and do not give a verdict.
This is slower than uploading a clip to a detector, but it is how professional verification teams work, and it is the one method that gets better, not worse, as generators improve.
Detector sites run a classifier over the pixels and return a percentage. They can be a useful extra signal, but their accuracy on new generators and on heavily compressed social media copies is rarely published in a way you can check, and real footage can score high. Never treat a percentage as a verdict, and be careful about uploading private videos to them.
No. Ordinary metadata can be edited by anyone, and a missing camera tag is normal after sharing. The closest thing to proof is a camera original with a valid Content Credentials signature from the camera maker, plus a trustworthy source.
The signer states the content was made by a generative AI model (trainedAlgorithmicMedia) or mixes AI elements into other material (compositeWithTrainedAlgorithmicMedia). Once the signature verifies, that is strong evidence about where the file came from.
No. Messaging apps and social platforms strip most metadata from every video, real or generated. No metadata means you need other checks, nothing more.
Platforms such as TikTok and LinkedIn say they read Content Credentials attached by AI tools, and some also apply labels when creators disclose AI use or when the platform’s own AI effects were used. A label means one of those signals was found; no label means none was found.
It reads provenance data (Content Credentials, encoder and software tags) and explains it. It does not analyse pixels for AI, cannot read invisible watermarks such as SynthID, and does not verify signatures.
They can strip Content Credentials, which are metadata, and crop a visible logo. Invisible watermarks such as SynthID are built into the pixels and sound and are designed to survive that. A missing label is therefore not a clean bill of health.
No. The file is read in your browser and nothing is sent. Third-party verifiers and detector sites are separate services with their own terms, so check before you upload anything private to them.