I tested ChatGPT's PDF Content Credentials: what survives a re-save, a print and an edit

ChatGPT-generated PDFs can embed Content Credentials, which are digitally signed statements about the document's origin. These are structured data, not simple text, adhering to the C2PA specification. The C2PA manifest contains claims and assertions, secured with a COSE signature, designed to attest to the document's provenance and the tool that created it.

A key finding was that if ChatGPT modifies a PDF it previously generated, the new file might still carry the original signed manifest. However, the critical component for verifying authenticity is the `assertion.dataHash.match`. If the document's content changes, even subtly, the hash of the current file will no longer align with the hash recorded in the original manifest, despite the manifest's continued presence and valid signature.

This means that while the manifest's presence indicates a tool once made a document, validating it by matching the hash to the current file confirms whether the tool made this specific version. Without this hash match, the credential, though present and signed, does not vouch for the current iteration of the document, limiting its utility for verifying content integrity after an edit. The test was based on ten documents from one account over two days, not a broad survey.