Anthropic Details How Claude's Invisible Text Watermarking Works

Anthropic began adding machine-readable watermarks to text generated by new Claude models starting August 2, 2026, and has since published more detail on how the system works. The move is a direct response to transparency requirements under Article 50 of the EU AI Act.

The company is using the SynthID Text approach Google DeepMind outlined in 2024: during generation, the model embeds an imperceptible signal directly into the output text that users won't see during normal reading, and Anthropic plans to release a separate watermark detection API.

The watermark survives copy-paste and, per Anthropic, light editing — but a complete rewrite where every word is replaced will remove it. That's a meaningful limitation: it identifies unedited or lightly-edited Claude output, not text that's been substantially reworked afterward.

Coverage at launch is scoped to Claude Sonnet 4.6, Claude Haiku 4.5, and other models released on or after August 2, 2026, and applies across Anthropic's access channels — the Claude platform API, Claude, Claude Code, Claude Cowork, and Claude Tag.

Anthropic emphasized the watermark carries no user-identifying information: nothing in the mark or its key would let anyone recover details about a specific user, their organization, or their conversation history — it only signals that the text came from a Claude model.