How Claude’s text watermark works

Claude models will soon generate watermarked text as part of Anthropic's compliance with the EU AI Act, which requires AI providers serving the EU market to mark AI-generated content. The watermark is designed to determine the likelihood that Claude was involved in writing a given piece of text, and other major AI providers have signed the same Code of Practice to implement their own watermarks.

The method works by exploiting low-stakes word choices in generation. When Claude picks the next word, it selects among plausible candidates (e.g., 'overcast' vs. 'grey' after 'cold and'). Normally this is settled by a random number; with watermarking, the randomness is instead derived from a secret key and a few preceding words. This creates a pattern across many choices that is undetectable to readers but can be checked by anyone holding the key to estimate the probability the text was AI-generated.

Anthropic emphasizes that watermarking has no practical impact on output quality or content, and readers cannot distinguish watermarked from unwatermarked text. Nothing is added to the text (no hidden characters), no extra tokens are required, and it does not increase cost. The watermark also carries no identifying information and cannot be traced to a specific person, organization, or chat, and it will not be unique to Claude—other providers will have their own implementations.