Anthropic has announced that all Claude AI models released from August 2, 2026, onwards will embed an invisible watermark within the text they generate. This initiative aligns with transparency requirements set by the European Union’s AI Act, specifically Article 50.

The watermark is not a visible label or disclaimer but operates at a statistical level. When Claude produces text, it selects words from a range of possible options. The watermark subtly influences these choices following a hidden pattern, allowing detection tools to identify AI-generated content without affecting the natural flow of the writing.

Scope and Implementation

  • The watermark applies globally, not limited to European users.
  • It covers all Claude-based products including the API, chat app, Claude Code, Claude Cowork, and Claude Tag.
  • Access through major cloud platforms such as AWS, Google Cloud, and Microsoft Foundry is included.

Detection and Limitations

The watermark survives common text operations like copying and pasting, and may persist through minor edits. However, Anthropic has not released the watermarking algorithm or a public detection tool, leaving practical effectiveness uncertain.

Importantly, detecting the watermark only confirms that the text was processed by a Claude model at some point. It does not identify who authored the ideas or the extent of AI involvement. For example, human-written text lightly edited by Claude could still carry the watermark, while heavily rewritten or mixed texts might evade detection.

Older Claude models released before the August 2026 cutoff will receive the watermark through a future transition, with no specific timeline announced.

Watermarking for Digital Files

For supported image formats like PNG, JPG, and SVG, Anthropic is embedding provenance metadata using the C2PA open standard. Unlike the text watermark, this metadata exists outside the file content and can be removed by screenshots or file conversions.

Transparency, Not Authorship Proof

Anthropic emphasizes that these measures aim to enhance transparency about AI involvement rather than provide definitive proof of authorship. Detection tools for third parties are still under development, and the watermark should be viewed as an indicator of AI processing rather than a conclusive attribution.