Anthropic Extends Invisible Watermarks to Earlier Claude Models

Anthropic is extending invisible text watermarks to earlier Claude models as the European Union’s AI Act applies Article 50 transparency obligations to providers of generative AI systems.

The watermark is created by adjusting word choices during generation. Readers cannot see the watermark in the response. Because the signal is part of the text, it can remain when someone copies the response elsewhere and may survive some editing. The method is designed not to change the answer’s meaning, quality or readability.

Anthropic began using the same technique with newer Claude models in August. The current expansion covers earlier models. The watermark is applied at the model level, so it travels with Claude text across the company’s products. It does not identify the person or organization that generated the text.

Article 50’s transparency obligations began applying Aug. 2, 2026. They require providers of generative AI systems to make outputs such as text, images, audio and video machine-readable and detectable as AI-generated. Anthropic has signed a related code of practice, which is voluntary; the transparency obligations in the AI Act are legal requirements.

Readers cannot identify a watermarked response by looking at it. Anthropic’s text detector is in a private preview for groups including regulators, law enforcement, media organizations, fact-checkers and researchers. Educational institutions, civic groups in the European Union and companies with similar compliance duties are also included. The detector is a separate tool for checking text; the watermark itself does not display an alert to the person reading a response.

Anthropic plans to widen access to its detection tools and publish additional technical guidance. Anthropic plans to extend the watermark to earlier Claude models on Nov. 3.