Overview
- Anthropic began applying invisible, machine-readable watermarks to Claude models launched on August 2, 2026, and says it will backport the feature to earlier models and roll it out across its API, apps and cloud deployments.
- Text outputs will carry an imperceptible watermark encoded in the model’s token sequence while generated files will include signed provenance metadata following industry standards such as C2PA.
- Anthropic warns the marks are not foolproof because edits, format conversion, screenshots or short passages can remove or obscure the signal, so absence of a mark does not prove content was not processed by Claude.
- Detecting the marks requires compatible tools or vendor detection systems, and Anthropic plans to publish detection utilities in the coming weeks to help verify when content passed through Claude.
- The move follows similar commitments from other big AI firms and is likely to set a de facto global standard for provenance, affecting how publishers, platforms and users verify AI-generated content and manage trust.