Anthropic has begun embedding invisible, machine-readable watermarks into text generated by Claude, enabling users and third parties to trace content back to the AI even after it’s copied elsewhere. The move was disclosed via an updated Claude Help Center article.
Claude models launched on or after August 2, 2026 will support this marking from day one. The watermark sits at the model level rather than the app level, meaning it travels with text across platforms and can survive light editing – without altering meaning, quality or readability.
The system spans Anthropic’s ecosystem: Claude, its API, Claude Code, Claude Cowork and Claude Tag, plus cloud platforms like AWS, Google Cloud and Microsoft Foundry. Anthropic ties the rollout to its commitments under Article 50(2) of the EU AI Act’s Code of Practice on Transparency of AI-Generated Content.
A parallel provenance system is being introduced for files – SVG, PNG and JPG formats will carry C2PA-based signed metadata indicating Claude involvement, though availability may vary by platform.
Crucially, Anthropic stresses the watermark isn’t a definitive AI-detection tool. Content run through Claude for proofreading or translation could carry the mark despite being user-authored – and heavy editing, paraphrasing or translation can make the signal undetectable. Very short passages may lack enough text to watermark reliably.
Anthropic frames this as a provenance signal, not a standalone detector, and says it’s building tools to let users and third parties actually check for these watermarks and file metadata.
