Anthropic has explained how it plans to watermark text generated by Claude, following concern from some users after the company said it would comply with the EU AI Act’s Transparency Code. The system is intended to make AI-generated content identifiable without changing how it reads, and Anthropic says it will not reduce output quality.

The company will use Google DeepMind’s 2024 SynthID-Text approach, embedding detectable patterns through low-stakes word choices such as alternative descriptions of weather. Light editing is unlikely to remove a watermark fully, while a complete rewrite can; human-authored text merely proofread by Claude may contain little or none. Watermarking will have a negligible effect on code, apart from flexible wording such as comments, and Anthropic plans to release a detection API.