Anthropic explains how Claude text watermarking works

Anthropic announced that text generated by future Claude models will include watermarks to help determine the likelihood that Claude was involved in writing the text, in order to comply with the EU AI Act. The watermark uses the SynthID-Text method published in Nature in 2024 by Google DeepMind. The authors say it has no practical effect on output quality, creativity, or readability; readers cannot distinguish watermarked text from watermarked text. It does not add tokens or increase costs.