How Claude's text watermarking works

Anthropic says future Claude models will start generating text with a watermark, a hidden pattern that lets someone estimate whether Claude helped write it. The company says the change is about EU compliance, not product tinkering. And it insists the reading experience won’t change at all.

The trick is in the model’s random choices. Large language models pick words one at a time, and many of those choices are basically interchangeable: overcast or grey, for example. Anthropic’s watermark nudges those low-stakes selections using a key, so the sequence of words ends up consistent in a way that can be checked later. To the reader, though, the text should look identical.

That matters because Anthropic says the watermark adds nothing visible to the text, no hidden characters, no extra tokens, and no extra cost. It also says the watermark carries no identifying information and can’t be traced back to a person, organization, or chat. The company says its internal testing found no hit to quality, creativity, or readability, and it points to Google DeepMind’s SynthID-Text work, which reported no statistically significant difference in ratings.

The watermark is not a magic lie detector. It can say whether text was likely partly written by Claude, but it cannot prove human authorship, identify a different AI, or work well on short samples. It also gets thinner in places where the wording has to be exact, like factual passages, proofreading, and much of code. In those cases there are fewer choices to hide a pattern in.

Anthropic says the rollout is tied to the EU AI Act, which requires providers serving the market to mark AI-generated content as of August 2. The company says it is applying watermarking globally at launch because it doesn’t yet have a durable way to limit it by region. It also says a watermark detection API is coming soon, while files such as .png, .jpg, and .svg will use C2PA content credentials instead of a text watermark.