C2PA is the open standard for recording where a file came from and how it was changed. Content Credentials are that record, signed and embedded in the file. This guide explains what they prove, why they disappear from AI images, and how to keep them intact from generation to delivery.
What is C2PA?
C2PA (the Coalition for Content Provenance and Authenticity) publishes an open technical standard for attaching a tamper-evident history to images, video, audio and documents. Adobe, Arm, the BBC, Intel, Microsoft and Truepic founded it in 2021, and it now runs under the Linux Foundation's Joint Development Foundation.
The standard defines a manifest: a signed bundle of statements ("assertions") about a file. Typical assertions say which tool created it, which edits were made, and whether AI produced it.
What are Content Credentials?
Content Credentials are C2PA manifests as people meet them: the "cr" pin, the panel in a verify tool, the provenance row in an asset library. Same data, friendlier name. Three things set them apart from ordinary metadata:
- Signed. Any change to the file or the record breaks the signature, so tampering shows.
- Chained. When a credentialed file is edited, the new record links the old one as an "ingredient", so history accumulates instead of being overwritten.
- Standard. Any C2PA-aware tool can read them, not only the app that wrote them.
What C2PA records about an AI image
For AI work, the assertions that matter most are these:
| Assertion | What it tells a viewer |
|---|---|
c2pa.created with digitalSourceType: trainedAlgorithmicMedia | The image was made by a generative model |
softwareAgent | Which tool generated or edited it |
c2pa.edited | It was changed after creation, and by what |
Ingredients | The earlier files it was built from, with their own credentials |
What C2PA does not record by default: your prompt, seed or workflow. Those live in the tool's own metadata (for ComfyUI, the embedded workflow JSON). See ComfyUI workflows.
Why Content Credentials disappear
Most credentials never reach the viewer. The usual causes:
- The generator never wrote them. Many image tools embed no C2PA manifest at all.
- Platforms strip metadata. Most social platforms re-encode uploads and drop embedded data.
- Exports rebuild the file. Resizing, format conversion and "save for web" often write a fresh file with no manifest.
- Edits break the chain. An editor that isn't C2PA-aware saves a new file with no link to the credentialed original.
How Numonic keeps the record with your files
On import, Numonic reads what's already there. It detects C2PA manifests in PNG, JPEG, MP4 and MP3 files and shows the claim generator, the recorded actions and the digital source type on the asset.
On export, Numonic signs. Single exports, bulk exports and published collections in JPEG, PNG and WebP (up to 50 MB) leave with a signed C2PA manifest. The manifest records:
- that Numonic handled the file (
c2pa.edited), - when Numonic identifies the file as AI-generated, a
c2pa.createdassertion withtrainedAlgorithmicMediaand the generating tool, - the AI-disclosure fields written into the file's IPTC metadata,
- the source file's own credentials as an ingredient, so an upstream signature stays linked.
If signing can't run, you'll know. The file still exports, marked unsigned, instead of failing silently. Published assets also serve their C2PA manifest publicly, so anyone can inspect it.
- Numonic reads Content Credentials on import and signs JPEG, PNG and WebP exports.
- Upstream credentials stay linked as ingredients, so the chain survives.
- The signing certificate is self-signed today: signatures validate, the issuer shows as unrecognized.
How to check a file's Content Credentials
- Open Content Credentials Verify (Content Authenticity Initiative).
- Drop in the file.
- Read the chain: each signer, what it did, and any ingredients.
A Numonic export shows Numonic's credential with the original file's credential linked beneath it.
C2PA and the EU AI Act
Article 50(2) of the EU AI Act requires providers of generative AI systems to mark outputs as AI-generated in a machine-readable format. C2PA's trainedAlgorithmicMedia source type is one of the formats built for that. Numonic doesn't generate images, so the provider duty sits with your generator. What Numonic does is keep that marking, and its own record, attached through the files you deliver. Details: EU AI Act Article 50 guide.