Anthropic will implement text watermarking in future Claude models to comply with the EU AI Act and its Code of Practice on Transparency of AI-Generated Content. The company uses the SynthID-Text method published by Google DeepMind, which embeds a detectable pattern in generated text without affecting output quality or adding extra tokens.
- Watermarking alters the source of randomness used to select words among equally likely candidates, leaving a pattern undetectable to readers but verifiable with a key.
- The technique has no practical impact on content quality, creativity, readability, speed, or cost, as it produces no additional tokens.
- Detection is probabilistic and requires sufficient text length; it does not identify specific users, organizations, or chats.
- Watermarks are sparse in factual passages and code where word choices are constrained, but apply to translations and comments.
- For images and files, Claude attaches C2PA content credentials in metadata rather than using text watermarks.
This change ensures compliance with EU regulations requiring AI providers to mark generated content, while maintaining the indistinguishability of Claude's outputs from human-written text.