Overview
- Anthropic announced this week that Claude models launched on or after August 2, 2026 will embed invisible, machine-readable watermarks in text and attach C2PA-signed provenance metadata to supported image files.
- The text watermark is applied at the model level so it travels with copied text and may survive some editing, and the file metadata records whether a file was processed by Claude and if its provenance has been altered.
- Anthropic warns a detected mark only indicates Claude may have processed the content and does not prove original authorship because human-written material can be watermarked after proofreading, translation, or other edits.
- Technical limits include loss of a detectable watermark after heavy paraphrasing, short passages that lack enough signal, and metadata loss through format changes or screenshots, and Anthropic says it will publish detection tools and documentation.
- The decision is part of a wider industry move to meet EU transparency standards, has prompted user backlash over effects on legitimate workflows, and Anthropic is working to retrofit older Claude models during the EU transition window ending December 2, 2026.