C2PA Content Credentials, Explained

What Content Credentials are

C2PA is a standard for embedding a cryptographically signed manifest into a file: what created it, when, and what was done to it since. The signature uses certificates on a trust list, so a manifest can be verified rather than merely read.

In practice it means a file can declare “I was generated by this model on this date” in a way that cannot be quietly forged — a real improvement over ordinary metadata, where every field is editable with free tools.

Who actually applies it

As of 2026 the picture is uneven, and the gaps matter more than the coverage.

Applying it: OpenAI for DALL·E 3 and Sora, Adobe for Firefly (automatically on fully generated images), Google for Imagen, and Stability's hosted API.

Not applying it: Midjourney embeds no C2PA manifest — see our page on that. And every locally run open model, which is the bulk of volume generation, has nobody in a position to enforce signing at all.

So a missing manifest is completely uninformative. It might mean the image is a photograph, or that it came from Midjourney, or from a local model, or that it was screenshotted once.

SynthID: a different bet

Google took another approach alongside C2PA. SynthID embeds a watermark in the pixels rather than the metadata, designed to survive cropping, resizing, compression and filters.

That solves the fragility problem — but introduces another: verification runs through Google's own tooling. No third party can read it, ours included, which we state plainly on the Gemini page. A verification system only its issuer can check is useful, but it is not an open standard.

The structural weakness

Provenance systems answer “what does this file say about itself”. They cannot answer “what is this file”.

Screenshot a signed image and the manifest is gone. Re-encode it, upload it to a platform that strips metadata, pass it through a messenger — gone. Nothing about the pixels changed; the provenance simply did not survive the journey.

This makes C2PA excellent for the honest case — a publisher proving what their image is — and close to useless against anyone who does not want to be identified. Removing a manifest requires no skill whatsoever.

How to use it sensibly

Check it first, because it is free and occasionally decisive. Our EXIF viewer shows the Software field and surrounding metadata; dedicated C2PA viewers display the full signed chain.

Then ignore its absence entirely and analyse the pixels. That is the only route that works regardless of how the file travelled, and it is what our detector does across twelve methods.

The realistic outlook: provenance will become standard on professionally produced content, which is genuinely valuable. It will never cover the material that is deliberately misleading, because that material is produced by people with every reason to strip it.

Key points

• C2PA embeds a signed, verifiable record of how a file was created.
• DALL·E, Sora, Firefly, Imagen and hosted Stability apply it; Midjourney does not.
• Local open models have nobody to enforce signing — that is most volume generation.
• SynthID marks pixels instead of metadata, but only Google can verify it.
• A screenshot removes any manifest, so absence proves nothing.
• Check provenance first, then analyse pixels regardless of the result.

Frequently asked questions

Check an image yourself

Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.

More from the blog