Overview
- Anthropic announced that Claude models launched on or after August 2, 2026 will embed an invisible, machine‑readable watermark in generated text and attach C2PA‑signed provenance to supported files, and the company says the marking will be applied worldwide across Claude products and API access.
- Text watermarks are implemented at the model level and are woven into the words so they travel with copied text and may survive some editing, while files such as SVG, PNG and JPG receive cryptographic provenance metadata under the C2PA standard.
- Anthropic plans to publish detection tools and technical documentation but has not yet released full details, and the company cautions that a detected mark only shows Claude may have processed content and does not prove Claude authored it.
- The watermarks have known technical limits because heavy editing, short excerpts, format conversions or screenshots can erase or weaken signals, and Anthropic is working to add marking to older Claude models during the EU Act’s transition window.
- The change prompted swift user and expert pushback focused on false attribution for human work, potential effects on proofreading and code workflows, and the likelihood of a new cycle of detection versus circumvention efforts across the industry.