Anthropic Watermarks Claude Text Under EU Law

Anthropic is putting a hidden, machine-readable watermark in text written by its Claude chatbot to comply with the European Union's AI Act and its Transparency Code, which requires AI companies to make AI-generated content identifiable.

In a blog post published Friday, the company laid out how it works: readers won't see anything. There are no visible marks and no hidden characters — instead Claude uses a secret key to steer the low-stakes word choices it would otherwise make at random, leaving a statistical pattern only someone with the key can decode.

The rule is European. The effect is not. Anthropic said it is applying watermarks to all of Claude's output at launch because it has no reliable way to switch the behavior on by region.

How It Works

Large language models pick one word at a time from a list of plausible next words, choosing among equally valid options more or less at random. With watermarking on, Claude uses a key instead of a random number generator to make those picks — Anthropic's own example used the digits of pi. The result is a pattern in word choice that is invisible to a reader but detectable with the key.

"Watermarking does not impact the quality of Claude's output," the company said. "To a reader, a watermarked response is indistinguishable from an unwatermarked one."

The method is an adaptation of SynthID-Text, the approach Google DeepMind published in the journal Nature in 2024. Anthropic said watermarking does not slow generation, require extra tokens or raise costs, and that it plans to release a detection API that can check whether a block of text came from Claude.

Anthropic listed the limits itself. The watermark cannot say whether Claude wrote a text or merely edited it — an edit or even a translation gets marked too, and all it really proves is that Claude was involved at some point. Very short text, or text Claude only proofread, may carry too little watermark to detect. Light editing "probably won't remove the watermark completely," the company said; a complete rewrite in which every word is replaced will — at which point, Anthropic argued, it is debatable whether the text is AI-generated at all.

Code carries "generally less watermarking than some other forms of text," the company said, because working code leaves few arbitrary choices; where choices exist, such as in comments, the watermark can attach, with what Anthropic called a negligible effect on the code itself. Images get a different treatment: a cryptographically signed note in the metadata saying Claude generated them.