Anthropic Is Now Watermarking Claude's AI Text, Here's How It Works

Anthropic has shared a deeper look at how future Claude models will carry watermarks in their text output, a change the company says is required under the EU AI Act's transparency rules. The watermark is created by subtly influencing how Claude chooses between equally reasonable words, such as picking "overcast" instead of "grey" when describing the weather. The resulting pattern is imperceptible to readers, but anyone with the appropriate key can estimate the likelihood that Claude generated or processed a passage. Anthropic also plans to release a watermark detection API.



The company published the additional details after its earlier announcement prompted debate among Claude users over the new watermarking requirement. According to Anthropic, the approach has no practical impact on output quality, adds no hidden characters, requires no extra tokens, and contains no identifying information that could be traced to a specific user, organization, or chat.


Claude uses a version of the SynthID-Text method published by Google DeepMind in Nature in 2024. The technique has a negligible impact on speed, does not increase cost, and differs from traditional AI detection software, which looks for stylistic patterns rather than checking for an encoded key. The company also says other major model developers that signed the EU Code of Practice will implement their own watermarking systems.

The system works best on longer passages, where Claude has more opportunities to make discretionary word choices. Light edits probably will not remove the marker completely, but a full rewrite where every word is replaced will. Anthropic stresses that the system can only estimate whether Claude was involved in creating or editing a passage, not prove authorship or distinguish between AI-generated and heavily AI-edited text. Likewise, if Claude only proofreads or lightly edits human writing, there may be too few Claude-selected words for reliable detection.

Code generally carries less watermarking than other forms of text because there are fewer opportunities to choose between equally valid alternatives, though comments and other discretionary text can still be marked. Translations produced by Claude also carry a watermark because every word is chosen by the model.

For images and other supported file types, Claude will instead attach C2PA content credentials in the file's metadata. Anthropic is applying the change globally because it does not yet have a durable way to limit watermarking by region. The company says older Claude models will gain watermarking over the coming months as part of the EU AI Act's transition period.