Content Credentials / C2PA
C2PA is the file-level half of AI content provenance: a manifest attached to an image, video, audio file or document saying how it was made and by what. Where text-watermarking hides a signal in a model's token choices, C2PA writes provenance into the container — PNG chunks, JPEG APP11 segments, ISOBMFF boxes, PDF XMP, DOCX document properties. Vendor checkers read it back: Claude's file checker, OpenAI's verification page and the Gemini app all report C2PA content credentials watermarks-remover.
Location is what makes it a governance problem. A token-level mark lives in the content; a container-level mark lives next to it, and metadata can be dropped.
Hard and soft binding
- Hard binding — the manifest travels in the file's metadata. Removing the metadata removes the provenance.
- Soft binding — an imperceptible in-content watermark that lets a remote Content Credentials manifest be re-linked after the metadata is gone.
watermarks-remover strips hard-bound C2PA across about 25 formats and states that soft bindings survive it, which is the honest version of what a metadata cleaner can do. The industry framing it cites is a two-layer one: C2PA plus an imperceptible watermark, precisely because either alone is removable or unreadable.
Why a labeling mandate leans on it
china-ai-regulation's 2025 AI-Generated Content Labeling Measures require on-screen disclosure plus embedded provenance metadata plus watermarking for AI text, image, audio and video. The embedded-metadata limb is C2PA-shaped, and it is the limb a free MIT-licensed tool removes in one command. The duty binds the producer; nothing in the mechanism binds whoever handles the file next.
The asymmetry:
| Marking | Removal | |
|---|---|---|
| Who acts | The producer, under a duty | Anyone downstream, under none |
| Cost | Vendor engineering | A pip-free Python script |
| Verifiable | By whoever holds the key or the checker | Metadata removal is checkable; nothing certifies a detector will now fail |
Open
Whether the EU AI Act's transparency obligations for AI-generated content name a machine-readable marking duty of this shape — and what, if anything, they say about removing one — is unread here; the spoke's eu-ai-act page does not yet cover it. That is growth edge material, not a claim.
Related
watermarks-remover · text-watermarking · claude-text-watermark · china-ai-regulation · eu-ai-act · ai-governance · audio-deepfake