Definition
C2PA is C2PA is an open standard for attaching signed provenance data, called Content Credentials, to images, video and audio, recording who made a file, with what tool, and how it was edited.
C2PA stands for the Coalition for Content Provenance and Authenticity. It is an open technical standard for attaching a tamper-evident record of origin to a digital file. The user-facing name for that record is Content Credentials.
Think of a shipping label that is sealed with a signature. When a camera, an editing app or an AI image tool creates or changes a file, it can write a small bundle of facts into it. That bundle is called a manifest. It can say which tool made the file, when, and what edits followed, including whether the file was generated by AI.
The manifest is digitally signed by whoever created it. Cryptographic hashes tie the signed record to the exact contents of the file. If someone changes the pixels afterward without updating the record, the check fails. A viewer, browser or checking tool reads the manifest, verifies the signature, and shows you the history.
C2PA proves who signed a claim. It does not prove the claim is true. A signer can lie about how a file was made, and the trust you place in a record depends on how much you trust the signer.
The record also lives in the file's metadata, so it can be stripped. Re-saving a screenshot, uploading to a service that removes metadata, or converting formats can drop it. A missing credential therefore means "unknown," not "fake." That is why provenance records are often paired with AI watermarking, which hides a signal in the content itself and is harder to remove by accident.
If you generate images or video with AI, the tool may attach Content Credentials automatically. If you publish those files, you may want to keep them intact so readers and platforms can see how the file was made. If you receive a file, a valid credential is a useful signal, but only one signal. Treat a missing or broken credential as a reason to look closer, not as proof of anything.
Rules in some regions are pushing AI providers to mark generated content, and C2PA is one of the formats they use for images, video and audio. Text is harder, since plain text has no metadata to carry a record, which is why text marking relies on watermarking instead.
You generate a product image with an AI tool. The file carries a manifest that says an AI model made it and names the tool. You then crop it in an editing app that supports C2PA. The app adds a second signed entry for the crop. A viewer later shows both steps in order. If you instead take a screenshot of the image, the new file has no manifest, and the history is gone.
C2PA is a provenance record attached to a file. AI watermarking embeds a signal in the content. Both aim to make AI-generated material identifiable, and neither is proof on its own.
Content Credentials is the public name for C2PA provenance data: a signed record attached to a file that shows who made it, with what tool, and how it was edited.
Yes. Because the record is stored in the file's metadata, re-saving, screenshotting or uploading to a service that strips metadata can remove it, so a missing credential means unknown, not fake.
Anthropic says it is adding invisible text watermarks to Claude models released before Aug. 2, 2026, and The Decoder reports the mark skips two of them when sampling settings change. If you ship Claude text, assume it's marked, and know which settings decide it.