Claude’s file checker reads C2PA credentials

On September 2, Anthropic opened a free web tool that checks whether a file carries a Claude content credential, the signed provenance record Claude attaches to files it generates. The Claude Content Checker accepts images, video and audio up to 100 MB: JPG, PNG, GIF, WEBP, TIFF, HEIC, AVIF, SVG, DNG, JXL, MP4, MOV, AVI, WAV, MP3, M4A and FLAC.

The marking comes from the transparency obligations in Article 50 of the EU AI Act and the code of practice Anthropic signed under it. Anthropic did not limit the feature to European users. The same marks ship worldwide.

The credential follows C2PA, an open standard for provenance metadata. A C2PA manifest is cryptographically signed metadata bundled into the file. It records who or what created the file, when, and with which tool, and any tampering shows. Anthropic’s help center says that when Claude generates a supported file type, such as a PNG or JPEG, it attaches a signed Content Credential, and a signed label means Claude processed the file. The signature covers a SHA-256 hash of the image content and is made with Anthropic’s private key, so changing the pixels invalidates it.

The checker looks only at that record. “Your file stays on your device,” the page says. The tool reads the embedded credential, not the file itself, and the file is never stored or used for anything else.

The upload list is longer than the list of files Claude marks today, which has been reported as SVG, PNG and JPG.

The page lists the tool’s limits itself. A positive result shows the content was processed by Claude, and “it doesn’t tell you anything about its substance.” A missing signal doesn’t prove Claude wasn’t involved: the credential may have been removed, or the content may have come from a model, platform or feature that doesn’t support marking.

Coverage also depends on the model. In the help center’s table, Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 attach C2PA Content Credentials, while Opus 4.7 and Sonnet 4.6 do not. Anthropic says it is adding watermarks to outputs from models released before August 2, 2026, and that the rollout on Amazon Bedrock will be complete by October 17.

Text is handled separately. Claude’s text watermark is applied at the model level, so it appears in output from every Claude product, and the file checker doesn’t read it. Text watermark detection is in private preview. Under EU law, it is open only to eligible organizations: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups.