How Claude’s text watermark works

Anthropic has announced that future Claude models will embed a watermark in the text they generate, a measure intended to comply with the EU AI Act. The watermark works by altering the source of randomness used when the model selects words. Instead of drawing from an arbitrary random number generator, the model uses a secret key and a few preceding words to decide which word to pick. The resulting pattern is invisible to readers but can be detected by anyone who has the key. Anthropic stresses that the watermark does not bias the model toward particular words, nor does it push the model to choose words it would not normally consider. In internal tests, the company reports no impact on content, creativity, or readability, and a study by Google DeepMind found no statistically significant differences in user ratings between watermarked and unwatermarked Gemini traffic. The watermark is most effective on longer passages with many word choices and is less useful on factual or code passages where choices are limited. It also has negligible impact on model speed and cost. The watermark is designed to indicate the likelihood that Claude was involved in writing a text, but it cannot confirm human authorship or detect other AIs’ watermarks. This development reflects a broader industry trend toward embedding traceability mechanisms in AI-generated content to enhance transparency and regulatory compliance.