What is C2PA, and what does it actually record?
If you have generated an image recently and noticed your file is larger than you expected, or seen a "Content Credentials" badge on a platform, you have met C2PA. It is worth understanding what it is before deciding whether you want it in your deliverables.
The short version
C2PA stands for the Coalition for Content Provenance and Authenticity — an industry body including Adobe, Microsoft, OpenAI, the BBC and others. It defines a standard for embedding a cryptographically signed record of how a piece of media was created and edited, directly inside the file.
Adobe markets the consumer-facing version of this as "Content Credentials". Both names refer to the same underlying manifest format.
What goes into a manifest
A C2PA manifest is a structured record that typically includes:
- The software that produced the file, including version — for example a specific generator and model.
- An assertion that the content is AI-generated, where the producer chooses to declare it.
- A creation timestamp.
- A chain of edit actions, if the file passed through other C2PA-aware tools.
- A cryptographic signature binding all of the above to the pixel data.
That signature is the interesting part. It means the manifest is not just metadata that anyone can forge — a verifier can check that the record genuinely came from the claimed producer and that the pixels have not changed since.
Where it lives in the file
The manifest is stored in a JUMBF box — a container format from the JPEG standards family. In PNG that box lives in a dedicated chunk type. In JPEG it sits in an APP marker segment. In both cases it is structurally separate from the image data, which is why removing it does not require touching your pixels.
This separation is the whole reason lossless scrubbing is possible. The manifest is a distinct region of the file, so it can be excised while the compressed pixel data is copied through byte for byte.
Why creators remove it
The reasons are usually mundane and professional rather than sinister:
- Client deliverables should not carry your internal tooling choices as embedded metadata.
- Some manifests, and the text chunks alongside them, include the full prompt you wrote.
- Agencies working under NDA often have contractual obligations not to leak process detail.
- Some platforms visibly badge C2PA-tagged media in ways that are not appropriate for the placement.
What removal does not do
Removing a manifest removes one declared signal. It does not make an image undetectable as AI-generated. Classifier-based detection works on the pixels themselves. Some generators additionally apply pixel-level or steganographic watermarking that survives metadata stripping entirely. And platforms increasingly keep their own server-side provenance records that never depended on your file in the first place.
Anyone selling you a tool that claims to make AI images "undetectable" is overstating what is technically possible. We would rather you know that up front.
Checking your own files
The fastest way to understand what your generator embeds is to look. Drop one of your own exports into the free scrubber — it lists every segment it finds, with sizes and readable previews, before removing anything.