Summary

  • Starting August 2, 2026, new Claude models launched in the EU will include a machine-readable watermark in all generated text, applicable globally.
  • This watermark will be integrated across all Claude platforms, including the API, Claude Code, and partnerships with cloud services.
  • Open-source solutions aimed at removing the watermark have already surfaced.

Anthropic has initiated the incorporation of an invisible watermark in all text produced by its latest Claude models. This feature is set to begin with models released in the European Union on August 2, 2026, and the company intends for it to be implemented worldwide.

The company detailed its watermarking strategy in a support article following its commitment to the EU AI Act's Code of Practice on transparency. This indicates that Anthropic is not merely doing this voluntarily. The watermark will be present across all functionalities of Claude, including the chatbot, API, Claude Code, and through cloud partners such as AWS, Google Cloud, and Microsoft Foundry.

Myriad: When might OpenAI unveil GPT-6? Share your prediction.

According to Anthropic, “When a supported Claude model generates text, it embeds an invisible watermark directly into the text. This watermark is undetectable and does not alter the meaning, quality, or readability of the output.” Furthermore, the watermark remains with the text when copied and may endure through certain edits.

This method is more intricate than typical user assumptions. As the model writes, it integrates an imperceptible watermark into the text itself, without any visible tags. This means the watermark survives through copy-pasting and, as Anthropic concedes, “may persist through some editing.” Additionally, files will have a second layer of verification through signed metadata using the C2PA open standard, akin to a digital manifest that details the file's creator and any subsequent alterations.

Watermarking Technique Remains Confidential

Anthropic has not disclosed the specifics of how the watermark is created. The support article describes it as model-level (trained within the model) and text-native (not reliant on external tools like metadata generators), but the precise detection methods and techniques remain undisclosed.

Researchers speculate it may involve a statistical signature, where the model subtly influences its word choices to create a detectable bias, similar to the method employed by Google in SynthID Text. However, this remains conjecture until Anthropic releases its detection tool.

This lack of transparency has not deterred privacy advocates; some experts are already developing techniques to circumvent Anthropic’s watermarking. For example, mikiane/claude-watermark-cleaner (106 stars on GitHub) eliminates invisible Unicode characters and rewrites text using a different model to disrupt the token patterns.

A larger initiative, guillaumemeyer/watermarks-remover (4.6k stars on GitHub), aims to remove watermarks from Claude text as well as C2PA and SynthID signals across various file formats like PNG, JPEG, SVG, PDF, and DOCX. The creators argue that a statistical watermark is “not a reliable method to prove origin” and often requires users to run their text through a second model for cleanup. No guarantees of removal can be assured until Anthropic provides its detection tools and establishes thresholds.