Anthropic revealed earlier this week that it would begin watermarking text generated by its Claude models to comply with new transparency rules under the EU AI Act. The announcement prompted questions from users about how exactly the process works, whether it affects output quality, and whether the mark can be removed. As a response, the company published a detailed blog post explaining the method.
Anthropic explains that Claude generates text one word at a time, selecting each next word from a list of plausible candidates. In many cases, several options would fit equally well, and the final choice is made through randomness. The watermark changes the source of that randomness.
Instead of using an ordinary random number generator, the model relies on a secret key combined with a few preceding words to make the selection. To a reader, the choices still appear random and natural. But anyone with the correct key can later examine the sequence and estimate the likelihood that Claude produced the text. Anthropic's approach is based on Google DeepMind's SynthID-Text technique, published in Nature in 2024.
The company says the process has no visible effect on writing quality, readability, or creativity. It adds no extra tokens, does not slow the model down, and does not increase costs. The watermark also contains no personal information and cannot be traced back to a specific user or account.
The watermark only works where there are genuine low-stakes choices between words. In factual sentences where only one answer is correct, such as "Isaac Newton's most famous work was called Principia Mathematica," there is no alternative to choose, so nothing is marked. The same applies to code that must be exact, though comments within code can carry marks.
Proofreading is also largely unaffected. If Claude only corrects grammar and punctuation in human-written text, very few words are Claude's own, leaving little room for the watermark. Short passages, heavy edits, paraphrasing, or full rewrites may weaken or remove the signal entirely.
For files in supported formats such as images, Claude will attach signed metadata using the C2PA open standard. This indicates that Claude helped produce the file but does not change the file itself.
Anthropic says it will soon offer a detection API so users can check whether a piece of text likely carries a Claude watermark. The company is also working to bring the feature to older models released before August 2, when the EU rules took effect. Because Anthropic currently has no reliable way to limit marking by region, the watermark will be applied globally.
