Anthropic Embeds Invisible Watermarks into Claude Text

Artificial intelligence developers are striving to increase internet content transparency. Following the lead of image-based markers, Anthropic has begun labeling text generated by its Claude neural network using invisible watermarks.

What Are Claude Text Watermarks?

The technology involves embedding machine-readable, human-imperceptible signatures directly into generated text. Anthropic states that this process does not affect the quality, meaning, or readability of the model’s responses. The watermark persists through copy-pasting and remains intact after minor content editing.

How the AI Content Identification System Works

The system operates on all Claude models launched in the European Union as of August 2. This requirement is mandated by the EU Code of Practice on the transparency of AI-generated content. The mechanism triggers automatically regardless of the access method: Claude API, Claude Code, Cowork, Tag, or platforms such as AWS, Google Cloud, and Microsoft Foundry.

Limitations and Labeling Nuances

There are scenarios where labeling may fail: for instance, if the text is too short or has undergone significant manual editing. Notably, watermarks are added even when AI serves as a supporting tool, such as for translation, formatting, summarizing, or proofreading third-party text.

Future Rollout and Detection Tools

While the current mandatory implementation applies only to models within the EU, Anthropic plans to extend this practice to older versions of Claude and make it available globally. Additionally, special tools that will help users identify these watermarks online are currently in development.