What is C2PA, and what is a Content Credential?
C2PA (the Coalition for Content Provenance and Authenticity) is an open technical standard for attaching signed origin information to a media file — the group behind it describes Content Credentials as the equivalent of a nutrition label for digital media. Where this site deals entirely in hidden characters inside plain text, C2PA operates one layer up: it marks image, video, audio, and PDF files with metadata about where they came from.
The coalition was announced in February 2021 by Adobe, Arm, the BBC, Intel, Microsoft, and Truepic, merging Adobe's Content Authenticity Initiative with the BBC/Microsoft-backed Project Origin. Its steering committee has since grown to include Amazon, Google, OpenAI, Meta, and Sony, and the specification itself is developed in the open under the Linux Foundation's Joint Development Foundation — anyone can read the current C2PA technical specification without a login or license fee.
How does a Content Credential actually work?
A C2PA manifest is a digitally signed data structure that travels with a file. Inside it are assertions — discrete, machine-readable claims such as "captured by this camera model," "edited in this application," or "generative AI produced this image from this prompt." Each assertion can be added by a different tool in the chain: a camera signs the capture, an editing app signs its changes, and an AI generator signs the fact that it generated the content, building an append-only history rather than a single tamper-proof stamp.
Assertions
The individual claims: device, editing actions, AI model and prompt, timestamp, location.
Signature
A cryptographic signature over the manifest, so any edit made after signing invalidates it.
Chain
Each new tool in the workflow can append its own signed manifest, building a linked history.
That signature is what makes a Content Credential meaningfully different from ordinary EXIF metadata: EXIF fields can be edited freely with no one noticing, while altering a C2PA-signed file after the fact breaks the signature and marks the credential invalid.
Does a Content Credential survive being posted online?
Often, no — and this is the standard's best-documented limitation. A Content Credential sits alongside the pixels, not inside them, so any process that re-encodes the file can leave the manifest behind without corrupting the visible image at all.
| What happens to the file | Credential survives? |
|---|---|
| Taking a screenshot | No — a screenshot is a brand-new file with no manifest |
| Re-saving or converting format | No, unless the saving tool explicitly re-signs it |
| Uploading to most social platforms | Usually no — recompression during upload strips metadata |
| Forwarding via messaging apps | Usually no — re-encoding on send removes it |
| Sharing the original file unmodified | Yes — the manifest and signature stay intact |
The practical result is a paradox: the content most likely to matter — something reshared widely enough to need a provenance check — is exactly the content most likely to have lost its credential somewhere along the way. A missing Content Credential is therefore not proof a file is fake; it may simply have been screenshotted or reposted through a service that strips metadata.
How does this compare to the watermarks this site covers?
C2PA is a third, distinct approach to marking AI-influenced content, alongside the two this site already documents in depth on what an AI text watermark actually is:
Character-level (text)
Hidden Unicode inside the text itself, like the narrow no-break space (U+202F). Travels with copy-paste; deterministically removable — this is what CleanPastedText does.
Statistical (text)
Word-choice bias, like Google's SynthID-Text. Invisible and survives copy-paste; no character cleaner can touch it.
C2PA (media files)
Signed metadata beside an image, video, or audio file. Cryptographically tamper-evident, but stripped by any re-encoding — the opposite trade-off from a statistical text watermark.
One driver behind faster C2PA adoption is the same one behind statistical text watermarking: the EU AI Act's Article 50(2) marking requirement, which took effect August 2, 2026 for generative AI systems. Anthropic's own August 2026 watermark launch for Claude, for instance, paired a SynthID-style statistical text watermark with C2PA metadata for the files Claude generates — two different techniques for two different content types, adopted for the same regulatory reason.
CleanPastedText's own honesty rule applies here too: this tool cleans plain text, never image or video files, so it has no effect on a Content Credential one way or the other. For the AI-text side of provenance, the AI watermark checker reports exactly which hidden characters a piece of text contains without altering it.