Overview
- Anthropic announced that Claude models released on or after Aug. 2, 2026 will include an imperceptible, machine-readable watermark in generated text and C2PA-signed provenance metadata in supported image files.
- The watermark is applied at the model level so marked text appears across Claude products and travels with copied text, and Anthropic is working to retrofit older models during the EU transition period.
- Anthropic says a detected mark signals possible Claude processing but does not prove sole authorship, and the company warns that heavy editing, short excerpts, or file conversions can erase marks.
- The firm pledged to publish technical documentation and detection tools for third parties, but experts and users say public detectors and the watermark technique—which biases token selection to create statistical patterns—can be imperfect and open to circumvention.
- The move follows the EU’s Article 50(2) Code of Practice on transparency, has prompted user backlash over proofing and privacy workflows, and joins other major AI providers in adopting machine-readable provenance as a compliance and tracking measure.