We’ve written an FAQ to answer some of the questions we've received about waterm

Anthropic's Claude text watermark works through word choice rather than hidden characters. Its FAQ explains that Claude normally picks among plausible next words, and the watermark keeps that variety while deriving each choice from a secret key and the preceding words. A detector holding the key can check whether a passage's choices match the resulting pattern.

Anthropic says the watermark carries no information identifying a person, organization or chat, and that the key can only indicate how likely it is that Claude partly wrote a text. It says detection is weaker on short samples and where few word choices exist, such as factual statements and code, and that light editing probably won't remove the watermark while a complete rewrite will. A detection result is therefore probabilistic evidence of Claude involvement, not proof of authorship.