Deepfake Detector
Paste a file and get the actual evidence: whether the file carries Content Credentials, what its metadata says, and how a detection model scores it. No single black-box verdict — every result shows what it is based on, so you can decide what to do next.
Check an image or video
Files are downscaled before analysis. Nothing here is a legal determination of authenticity.
Analysis is not live yet — leave your email and we'll tell you when it is.
What counts as a deepfake
“Deepfake” is used loosely online. In this tool it means media where a person's likeness, voice, or motion has been synthesized or altered by a generative model. The common cases:
- Face swap — one face replaced with another, usually frame by frame.
- Face reenactment — the original person stays, but expression or head movement is driven by someone else.
- Lip-sync manipulation — the mouth is regenerated to match new audio.
- Fully generated video — no source footage exists; the entire clip comes from a video model.
How the analysis works
We run three independent layers and report each one separately, because they fail in different ways and a conclusion built on only one of them is easy to overstate.
1. Provenance: C2PA Content Credentials
Many recent cameras, phones, and AI tools can attach a cryptographically signed manifest describing how the file was made and edited. If a manifest is present, we verify the signature and read the declared actions. A valid signature that names a generative model is the strongest single piece of evidence available — stronger than any classifier score.
Most files have no manifest. That is normal and proves nothing either way. Read a file's Content Credentials directly.
2. Metadata and file structure
We surface EXIF and XMP fields, including the IPTC TrainedAlgorithmicMedia
digital source type, editing history, and software tags. We also flag structural mismatches —
a container claiming a camera model it cannot have produced, a re-encode that wiped the original
encoder string, or timestamps that contradict each other.
3. Statistical risk score
A detection model scores the pixels or frames for artifacts typical of synthesis. We report it as a calibrated risk score with the uncertainty attached, never as a percentage that sounds like a diagnosis. A high score means “look closer”. It does not mean “proven fake”.
How to read the result
| Result | What it means | What to do |
|---|---|---|
| Provenance verified | A valid C2PA manifest was found and its signature verified. | Trust the declared history. Check whether the manifest itself names a generative step. |
| Elevated risk | Classifier and/or structural signals indicate likely synthesis. | Treat as unverified. Seek the original file or a second source before publishing. |
| Inconclusive | Signals conflict, or the file was re-encoded or screenshotted. | Do not conclude. Ask for the original upload or an unedited export. |
| No manipulation found | Nothing flagged. This is not proof of authenticity. | Note it as one data point, not a clearance. |
Where the method breaks down
Being explicit about limits is part of the product, not a disclaimer bolted on:
- Re-uploads erase evidence. Platforms strip metadata and re-compress; a screenshot keeps almost nothing.
- Classifiers are sensitive to compression. The same clip scores differently before and after a messaging app touches it.
- No watermark is not a signal. Most generators do not embed one, and watermarks can be removed.
- Short clips are weaker. Fewer frames means less to analyze.
Full write-up of each layer, including thresholds: our methodology.
Who uses this
- Newsrooms and fact-checkers documenting why a clip was flagged before publication.
- Trust and safety teams triaging user uploads with a retrievable audit trail.
- KYC and fraud review where liveness checks need a second opinion on submitted media.
- Marketplaces and second-hand platforms handling disputed photo or video evidence.
FAQ
How accurate is a deepfake detector?
There is no single accuracy number that holds across all media. Detection quality depends on the model, the generator used, resolution, and how many times the file was re-encoded. That is why we return the underlying evidence — provenance data, metadata traces, and a risk score — instead of a bare yes or no. A high score is a reason to look closer, not a verdict.
Does a missing C2PA credential mean the video is fake?
No. The vast majority of real cameras and editing tools still do not attach Content Credentials, and most platforms strip metadata on upload. Absence of provenance proves nothing. Presence of a valid signature that names an AI generator, on the other hand, is strong evidence.
Can you detect a deepfake in a screenshot or a re-uploaded clip?
It is significantly harder. Screenshots, screen recordings, and platform re-encodes remove EXIF and C2PA data and alter compression artifacts, which are two of the main signal sources. Detection then depends mostly on visual artifacts and the classifier, so results carry wider uncertainty and we label them as inconclusive more often.
Do you store the files I upload?
Files are processed to produce the analysis and are not kept as part of a library of your media. The file will be deleted within 24 hours.
What file types and file sizes are supported?
Images in JPEG, PNG, and WebP, and video in MP4 and MOV, up to 20 MB per file. Files are downscaled to 2 MB before analysis to keep response times predictable.
What is the difference between a deepfake detector and an AI image detector?
A deepfake detector focuses on manipulated people — face swaps, face reenactment, lip-sync, and generated video. An AI image detector covers any synthetically generated still image, including art, product shots, and scenery. Both live under the same evidence-first approach; the analysis pipeline is shared. See the full image forensic toolkit.
Run it on a file today
The web check is free. If you need this inside your own product — upload triage, moderation pipelines, or case records — see pricing for the API and forensic reports.