Anthropic Details Claude Text Watermarking for EU AI Act Compliance
Anthropic revealed plans to incorporate a built-in watermark in upcoming Claude AI models to adhere to the EU's AI Act. This technique, inspired by Google DeepMind's SynthID-Text from 2024, maintains the quality and readability of outputs. It modifies the randomness in word choice, generating a recognizable pattern without the need for hidden characters or additional expenses. In July 2026, Anthropic committed to the EU Code of Practice on Transparency of AI-Generated Content. The watermark will be implemented worldwide, with a detection API expected shortly. For images and files, Claude will utilize C2PA content credentials. Watermarking will also be applied to older models during the transition for those released before August 2, 2026, and the method does not link text to users.
Key facts
- Anthropic will watermark text generated by future Claude models.
- The watermarking method is based on Google DeepMind's SynthID-Text approach.
- Watermarking does not affect output quality or readability.
- Watermarked text is indistinguishable from unwatermarked text to readers.
- No extra tokens are required, and there is no additional cost.
- The watermark cannot be traced to individual users or organizations.
- Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026.
- A watermark detection API will be offered soon.
- Claude will attach C2PA content credentials to image and file outputs.
- Older Claude models will receive watermarking over the coming months.
Entities
Institutions
- Anthropic
- Google DeepMind
- EU
- European Union
- Claude
- C2PA
- Suno
- Substack
- Pangram
- Black Forest Labs
- Meta
- Microsoft
- OpenAI
- Synthesia
Locations
- European Union