AI Image Forensics: Check a File and See the Evidence Behind the Result
Upload an image, a video or a PDF. We read the file's own record first — C2PA Content Credentials, EXIF, XMP and the IPTC source type — then run statistical detection on the pixels, then tell you what the evidence supports and what it does not. No green check without a reason attached to it.
Check a file
Five free checks a month. No credentials in the file is a normal result, not an error.
Analysis is not live yet — leave your email and we'll tell you when it is.
Three layers, in this order
Most tools give you one answer. We give you three, because they differ in how much they are worth, and you cannot judge the result without knowing which one you are looking at.
| Layer | What we read | What it can prove |
|---|---|---|
| Provenance | C2PA manifest, signature chain, EXIF, XMP, IPTC source type | How the file says it was made, and whether that statement has been altered since |
| Detection | Generator artifacts, upsampling residue, noise inconsistency, blending edges | That traces consistent with synthetic or edited origin are present — a probability statement, not proof |
| Explanation | Which indicators fired, which did not, and what each one means | The part you can actually defend when someone questions the answer |
The reasoning behind each layer, including where it breaks down, is written out in full on our methodology page.
Tools on this site
| Tool | What it answers |
|---|---|
| C2PA Viewer | What this file's Content Credentials actually say: claim generator, edit history, source type, ingredients, and whether the signature verifies against a trusted signer. |
| Deepfake Detector | Whether a video or a face image shows signs of synthesis or a face swap, and which frame-level evidence it rests on. |
| Methodology | How the three layers work, how the score bands are defined, and every known failure mode. |
| Pricing | Free tier, Pro, API access, and per-case signed forensic reports. |
An AI image checker that shows its work
People arrive at a checker like this with three different questions, and they need different levels of detail. Some want to know whether a single photo was generated by a model. Some want to know whether a photo they already published was edited after the fact. Some are looking at a file attached to a contract and need an answer that survives being argued with.
All three start the same way here: the file is read before it is judged. That is also where most AI photo detectors stop being useful — they return a percentage and a colored bar, with nothing underneath it you can point at in a conversation. This one returns the manifest fields, the indicator list, and the statement of limits, whether or not the news is bad for us.
When a file has no credentials
This is the most common outcome and the most misread one. No C2PA manifest does not mean the file is fake, and no detection trace does not mean the file is real:
- Many cameras and editing applications still attach no manifest at all.
- Uploading to most social platforms strips metadata and embedded manifests.
- A screenshot or screen recording destroys the manifest — the pixels survive, the record does not.
- Detection models trained on a given model family have little to say about a model released after them, and release cycles are now measured in weeks.
We treat every one of these as "no information", not as a verdict in either direction.
Built for people who may have to defend the answer
- Journalists and fact-checkers — verifying a user-submitted image before publication, with a record of what was checked.
- Trust, safety and moderation teams — routing decisions that need a defensible trail, not a black-box score.
- Marketplaces and dating platforms — checking profile and listing photos for synthetic or stolen origin.
- Legal and compliance — evidence summaries that name the hash, the manifest and the timestamp.
- Researchers — reproducible findings, with the method published rather than described in marketing copy.
Questions
Why don't you just answer yes or no?
Because a yes or no would be wrong often enough to be useless. A file with no credentials and no traces is not proven clean; a file with a valid signature is not proven honest. We report the layers separately and state what the evidence cannot establish.
What do you check first, the metadata or the pixels?
The metadata, always. A signature that resolves to a trusted signer is stronger than any classifier's output. Pixels come second because detection is conditional — good inside its training distribution, weaker outside it.
What file types are supported?
JPEG, PNG, WebP, MP4, MOV and PDF, up to 20 MB per file, with batch upload on paid plans.
Do you keep my uploads?
Files are deleted 24 hours after analysis. The report — indicators, manifest fields, hashes, timestamps — is kept so the result can be reproduced later. We never train on uploaded files.
Is it free?
Five checks a month are free. Paid plans start at $19 a month; per-case forensic reports are $149. Details are on the pricing page.