Anthropic has begun embedding an imperceptible watermark in all text generated by its newest Claude models. The change took effect for models launched in the European Union on Aug. 2, 2026, and the company said it plans to apply the system worldwide.
The plan was described in a support article published after Anthropic signed the European Union AI Act’s Code of Practice on transparency. According to the company, the watermark reaches every Claude surface, including the chatbot, API, Claude Code, and cloud partners such as Amazon Web Services, Google Cloud, and Microsoft Foundry.
The watermark is built into the text itself
Anthropic said, 「When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won't see it, and it doesn't change the meaning, quality, or readability of Claude's response.」
The company added, 「Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing.」
That makes the system more involved than a visible label or a separate tag attached after generation. In Anthropic’s description, the marker is woven into the words produced by supported Claude models, with no visible notice attached to the output. Files receive a second layer as well: signed metadata under the C2PA open standard, which records who produced a file and whether it was altered afterward.
Anthropic has not disclosed the method
Anthropic has not explained how the watermark is created. Its support article describes the approach as model-level, meaning the model is trained with the capability, and text-native, meaning it is not an external tool such as a metadata generator. The detection documentation and the exact technique have not been released.
Researchers have inferred that the system may rely on a statistical signature, where the model nudges word choice toward a faint but detectable bias. That would place it in the same general family as Google’s SynthID Text. For now, that remains inference rather than confirmation, because Anthropic has not published a detector.
Developers are already trying to remove the marks
Privacy concerns have not slowed attempts to break the system. Some developers are already working on methods to disrupt or remove Anthropic’s undisclosed watermarking approach.
One GitHub project, mikiane/claude-watermark-cleaner, has 106 stars. The project removes invisible Unicode and then rewrites the text with a non-Claude model to disturb the token pattern.
A larger repository, guillaumemeyer/watermarks-remover, has 4.6k stars on GitHub. Its description says it strips Claude text marks as well as C2PA and SynthID-class signals across PNG, JPEG, SVG, PDF, and DOCX files. The authors argue that a statistical text watermark is 「not a reliable way to prove origin」 and mostly pushes users to run their own writing through a second model pass for cleanup. The report noted that no removal method can be guaranteed until Anthropic ships its detector and thresholds.
An earlier Unicode tracking dispute is adding to the backlash
Anthropic’s own history is making the reaction sharper. In March, the company removed a hidden Claude Code tracker after researchers found that it tagged some users’ location and proxy use through undisclosed Unicode markers. The report said that episode used the same kind of quiet marking technique now drawing attention in the watermark rollout.
The mark shows Claude touched a text, not that Claude wrote all of it
The watermark can indicate that Claude had a hand in a piece of text, but it cannot prove that the entire text was generated by Claude. Under that logic, an originally human-written passage with a small edit could be treated the same way as fully AI-generated content.
If a user asks Claude to proofread or translate a paragraph, the resulting output can still carry the signal. Anthropic also says heavy editing can strip the mark, while the absence of a mark does not prove a human wrote the text.
Similar ideas are appearing in U.S. policy discussions
The report also pointed to a U.S. bill, the COPIED Act, which pushes a similar concept: a standardized way to watermark AI content so platforms can trace where it came from.
Anthropic has not said when it will publish the detection tools that would allow outside parties to verify the watermark.