Anthropic adds invisible watermarks to text from Claude

Anthropic has started inserting invisible watermarks into text generated by Claude. The signal is embedded directly into the generated words so that content can be identified as created or processed by artificial intelligence, even after being copied and pasted into other applications such as Word, Gmail or social networks.

The measure applies globally to models released since 2 August 2026 across products, APIs and compatible cloud services. Anthropic stated that the signal does not alter the meaning, quality or readability of the output. In addition to text, compatible files such as images can incorporate cryptographically signed provenance metadata through the C2PA standard.

The rollout follows the application of transparency rules under Article 50 of the European AI Act on 2 August 2026. Anthropic is among approximately 190 organisations that signed the associated voluntary Code of Practice by late July, alongside Google, Meta, Microsoft, Mistral and OpenAI.

Anthropic is developing detection tools for users and third parties and plans to release technical documentation. The company noted that detecting a mark indicates that Claude processed the content at some point, including editing or translation, rather than confirming the text was written entirely by AI. The mark may also disappear if text is heavily rewritten, paraphrased, translated, mixed with other copy or too short.