C2PA (Coalition for Content Provenance and Authenticity) is an open standard for attaching cryptographically signed provenance metadata — called Content Credentials — to content.

What are Content Credentials?

Content Credentials are the C2PA manifests that record where content came from and what was done to it. Because they're signed, tampering breaks the signature.

Can C2PA metadata be removed?

Yes, technically — re-encoding, resizing, screenshots, or format conversion can drop the manifest. But removing it is itself detectable, and it defeats the point of the transparency mechanism.

C2PA & Content Credentials

The open standard for proving where content came from. Here's how signed provenance works, why Claude uses it for files, and what it means when you see a credentials badge.

What C2PA actually is

C2PA (Coalition for Content Provenance and Authenticity) is an open technical standard, backed by companies including Adobe, Microsoft, Google, OpenAI, and Anthropic. It defines a way to attach a cryptographically signed manifest — called Content Credentials — to content, recording where it came from and what was done to it.

Unlike a text watermark, C2PA doesn't hide a signal in the content. It attaches provenance as container-level metadata. Because the manifest is signed, any tampering breaks the signature, and any C2PA reader can say so.

Why Claude uses it

When Claude generates a supported file type (.svg, .png, .jpg), it attaches signed C2PA provenance metadata. This is the file-side half of Anthropic's marking rollout, complementing the embedded text watermark.

How to check a file's provenance

In browsers that support Content Credentials, a small badge or icon appears on images with valid manifests.
verify.contentauthenticity.org — the official web verifier from the Content Authenticity Initiative.
c2patool — the open-source command-line inspector for reading manifests.

The honest caveats

No manifest proves nothing. A screenshot, a re-encode, or a CDN optimization silently drops C2PA metadata. Its absence says nothing about whether AI made the image.
Stripping is detectable. Removing the manifest defeats the transparency mechanism — that's the whole point of signing it.
Presence means "AI may be involved." A signed manifest records the chain of edits; it doesn't by itself judge whether a human or model is the author.

Related: the text side

Claude's embedded text watermark covers generated words; the zero-width scanner cleans the invisible characters that slip into copied text.