Anthropic launches Claude file provenance checker using C2PA metadata

Anthropic’s Claude file checker reveals provenance via C2PA metadata

Takeaway: Anthropic now offers a browser‑based tool that reads cryptographically signed C2PA metadata embedded in supported media files (images, video, audio) to tell you if Claude was involved in creating or processing the file, addressing emerging regulatory demands for AI‑generated content transparency.

What the tool does

The tool identifies only the content credential, not the content itself.

  • When Claude creates a supported file (e.g., PNG, JPG, SVG, MP4), it attaches a small C2PA‑compatible note in the file’s metadata.
  • The checker reads that note locally in the browser; the file never leaves the user’s device.
  • If the credential points to Claude, the UI displays a positive result.
  • The tool does not analyze the visual or textual content of the file, nor does it reveal who generated the file.

"Your file stays on your device. The checker only reads the embedded credential, not the file itself. Your file is never stored or used for any other purpose." – Anthropic documentation

How it works under the hood

  • Claude embeds a C2PA (Coalition for Content Provenance and Authenticity) credential, the same industry standard used by camera manufacturers and photo‑editing software.
  • The credential is a cryptographically signed statement that the file was made or processed with Claude.
  • Any C2PA‑aware reader can verify the signature; the Claude checker ships a WebAssembly version of the c2pa-rs Rust library to perform verification client‑side.

"This thing works using WebAssembly – https://claude.com/_next/static/media/c2pa_verify_wasm_bg… – which is a WASM compiled version of this Rust library: https://github.com/contentauth/c2pa-rs" – @simonw

Supported formats and limitations

  • The UI lists JPG, PNG, GIF, WEBP, TIFF, HEIC, AVIF, SVG, DNG, JXL, MP4, MOV, AVI, WAV, MP3, M4A, FLAC – up to 100 MB.
  • Formats such as PDF, Excel, or other document types are not supported, which limits the tool’s usefulness for checking academic papers or reports.
  • The metadata can be stripped easily by re‑saving the file, but forging a Claude credential is infeasible without Anthropic’s private signing key.

"You can resave the file and the ‘made with Claude’ signal disappears, but you cannot make a random file pass as Claude‑made without Anthropic's signing key. So the useful guarantee is one‑way. No signature means almost nothing." – @coffeecoders

Why the checker matters now

  • Regulatory pressure: The EU AI Act and California’s AI Transparency Act require large AI providers to embed provenance information in generated content. Anthropic’s text watermarking and this file credential are direct responses to those mandates.

"As far as I can tell, this and the recent change to add watermarking to text outputs … is to become compliant with the EU AI Act and CA's AI Transparency Act…" – @csmoak

  • Model training hygiene: By flagging Claude‑generated media, Anthropic can discourage inadvertent inclusion of its own outputs in future training data.

"I think this is primarily done so as they don't get AI generated content in their training data." – @kelvinjps10

Community observations and edge cases

  • Metadata injection path: Claude’s web UI generates some images by first creating an SVG, then rasterizing it. The C2PA credential appears on the final download, but disappears if the file is repackaged (e.g., placed in a ZIP) before download.

"When Claude Code CLI needs a media file … the downloaded JPEG had C2PA metadata, seemingly injected by the artifact‑presentation thing. … The extracted JPEG had no C2PA metadata." – @Retr0id

  • False negatives: Users have reported cases where a Claude‑generated figure was not flagged, suggesting the tool may miss credentials in certain pipelines.

"I just made a figure with Claude and fed it to the tool, which told me there was no evidence it was made with Claude. So… this utility appears to be pretty worthless." – @cmiles8

  • Ease of removal vs. forgery: Stripping C2PA data is trivial (re‑save the file), but adding a convincing Claude credential without the private key is practically impossible.

"Stripping the C2PA data is easy, but faking it is hard. No signature means almost nothing." – @coffeecoders

  • Scope of watermarking: The tool only covers media files; Claude’s text watermarking remains a separate, private‑preview API for EU‑required detection.

"To detect watermarks embedded in text we have a Detection API which is currently in private preview…" – Anthropic documentation

How to use the checker

  1. Visit https://claude.com/check-content.
  2. Drag‑and‑drop or click to select a supported file (max 100 MB).
  3. The browser‑side verifier reads any C2PA credential and reports whether Claude is listed as the provenance source.
  4. No data is uploaded; the verification runs entirely locally.

Comparison with other providers

  • OpenAI, Google DeepMind, and Gemini also expose provenance tools (e.g., OpenAI’s Verify, DeepMind’s SynthID). Each uses its own metadata scheme, but all rely on industry‑standard signatures to avoid easy spoofing.

Open questions and future directions

  • Broader format support: Adding PDF, DOCX, and spreadsheet formats would address the most common academic and business use cases.
  • Robustness against stripping: Techniques such as embedding credentials in multiple metadata fields or using perceptual watermarks could make removal harder.
  • Public detection API for text: Wider availability of the text watermark detector would complement the file checker and provide a unified provenance ecosystem.
  • Policy implications: As more jurisdictions require provenance, the balance between transparency and user privacy (e.g., not revealing who generated the file) will be critical.

Anthropic’s Claude file checker is a concrete step toward verifiable AI‑generated media, leveraging the open C2PA standard to provide on‑device provenance while respecting user privacy. Its current limitations—format coverage and susceptibility to metadata stripping—are acknowledged by the community, and future enhancements will likely broaden its applicability as regulatory demands increase.

Sources

Related