What are Content Credentials (C2PA)?
What the signed record contains, why present and authentic are different things, and why its absence tells you almost nothing.
Was this image made by AI?
No file can prove how it was made — and nothing here pretends to.
Some generators write a record of their work into the file they produce: the tool that made it, the settings, sometimes the whole prompt. Image tools do it most thoroughly, but a generated song carries the name of the program that made it in the same field a recorder writes its own name into. This page reads that record, from five different places, entirely in your browser.
When a file carries one, you see exactly what it says. When a file carries nothing, that is not evidence of anything — and this page will not pretend it is.
Drag & drop your file here
or paste from the clipboard
Reads metadata in: JPEG · PNG · WebP · HEIC/HEIF · AVIF · MP4 · QuickTime MOV · WebM · PDF · MP3 · WAV · M4A
Your file never leaves your device. Processing happens entirely in your browser.
What this reads
Five places, gathered into one list. Four of them are ordinary metadata that most viewers show without naming; the fifth is a signed record that has to be decoded.
- Content Credentials (C2PA) — a signed record naming the producing tool, the actions taken on the file, who signed it and when, and sometimes a declaration of how the content originated.
- PNG text chunks — where several image generators write the prompt, the model, the seed and the sampler, in plain text.
- The IPTC digital source type — a standard field for stating how a file came about.
- The EXIF Software field — what the last piece of software to write the file called itself.
- The XMP creator tool — the same idea, in a different metadata block.
Three things to know before you read a result
- Traces are voluntary. Some generators write them and some do not, and two products from the same company can differ. A file carrying no record is unremarkable.
- Traces are easy to remove. Re-encoding an image, screenshotting it, or uploading it through an app that recompresses will drop them without anyone deciding to — and this very site removes them on request.
- So absence proves nothing. A file with no record has not been shown to be anything at all. The only question this page answers is what a file carries.
- And presence is not proof either: this reads what a record says, it does not check that the record is true. Establishing that means verifying a signature against a trusted certificate, which is a different job and not one this build does.
What this actually finds, on real files
Every line below is a result from a real file, not an illustration. Together they show why a single rule about “AI images” does not hold.
- A video made with Sora
- Carries a Content Credential naming Sora as the producing tool, signed and dated, and declaring the content as produced by a generative model.
- An image made with Adobe Firefly
- The same: a credential naming the tool, with the same declaration of generative production.
- An image made with ChatGPT
- Carries a credential naming ChatGPT — but it records an "opened" action and declares no source type at all. A file from a generative product need not say it was generated.
- A PNG from ComfyUI
- No credential, but the complete generation prompt, the model, the seed and the sampler sit in its text chunks, written there by the tool.
- A PNG from a "stealth" AUTOMATIC1111 setup
- The same kind of generation parameters — hidden in the image’s alpha channel instead of a text chunk. Nothing reading metadata can see them, and this page reports the file as carrying no traces.
- A photograph from a Nikon Z 9
- Carries a credential too, signed by NIKON CORPORATION. Credentials are not an AI marker: cameras write them as well.
- A photograph edited in Lightroom
- No credential at all, yet the file still names its tool through the EXIF software field. Provenance information is not only in credentials.
Why there is no "scan the pixels" button
Classifiers that guess from the pixels do exist. They are probabilistic, they disagree with each other, and they are confidently wrong often enough that a percentage next to an image reads as authority it has not earned.
Invisible watermarking is the more promising idea, and it is not available to a tool like this one: the detector for the best-known scheme is not public, and the detection services that do exist are cloud APIs that would require sending your file to a company server. That would trade the one property this site is built on for a guess.
So this page does the part that can be done honestly and completely: it reads what the file states, and tells you when it states nothing.
Your file does not leave the device
Most tools that read or strip metadata do the work on a server, which means uploading your file to a company you now have to trust — with how long it keeps the file, who can reach it, and what it does with what it learned. That is a policy you have to read and believe.
This one has no such policy to read, because there is nothing to keep. The parser is JavaScript that runs in your browser: the file is read into memory on your own device, processed there, and handed back to you. Open your browser’s network tools and analyse a file — you will see no request at all.
It is not a promise, it is a property of how the thing is built, and the test suite asserts it: a full analyse, clean and edit cycle produces zero requests to any external origin.
Read further
Metadata vs content provenance
Why metadata is not evidence, and what a provenance system adds that a plain EXIF field never could.
Frequently asked questions
Can this tell me whether an image was generated by AI?
No, and neither can anything else with certainty. It tells you what the file states about how it was made. On a file that carries a record, that is often a clear and useful answer. On a file that carries none, the honest answer is that nothing is known.
The file says it was made by a generative model. Is that reliable?
It is a claim the producing tool wrote and signed. It is far more likely to be true than invented — writing one serves no purpose if the claim is false — but this build reads the record without verifying its signature, so treat it as a strong lead rather than proof.
The file carries nothing. Does that mean a person made it?
No. It means the file carries no record. Records are optional, and ordinary handling removes them: a screenshot, a re-encode, or an upload through an app that recompresses will all drop one. Most files in circulation carry nothing.
Is my image uploaded to check it?
No. Everything happens in your browser. A referenced credential stored on another server is shown to you by its address and never requested, because requesting it would tell that server which file you are looking at.
Can I remove what it finds?
Yes, under Remove metadata. Generation parameters go with an ordinary clean. Content Credentials are a separate choice, because erasing a file’s record of where it came from is a different act from stripping its GPS coordinates.