Anthropic has provided more details about the text watermarking system it is introducing for Claude, explaining how the invisible marker will work and what it can reveal. The company says the system will rely on word-selection patterns rather than adding characters or other visible elements to generated text. The watermark will not affect the quality, meaning or readability of Claude’s responses and will not require additional tokens. Anthropic is implementing the technology as part of its compliance with the EU AI Act.
Anthropic Details Claude’s Text Watermarking System and Its Limits
The watermark will change the source of randomness Claude uses when choosing between words that are similarly suitable, according to an Anthropic press release. These choices can create a pattern across a passage that a detection system can identify using a key. The method does not force Claude to use unusual words or favour specific terms.
Anthropic is using a version of Google’s SynthID-Text method, which Google DeepMind published in 2024. The company says its internal testing found no effect on the content, creativity or readability of Claude’s responses. Google DeepMind also found no statistically significant difference in user ratings between watermarked and unwatermarked Gemini responses.
The watermark is designed to show how likely it is that Claude contributed to a passage, rather than establish its complete origin. It does not reveal the user’s identity, organisation or conversation, and cannot determine whether another AI generated the text. Detection is also less reliable with short passages because they provide fewer word choices for the system to analyse.
The technology has similar limitations with certain types of content. Factual writing leaves fewer opportunities for alternative word choices without affecting accuracy. Proofreading can also produce a weak watermark because most of the text remains unchanged from the original. Code is likely to contain less watermarking because it often requires exact outputs, although the system can still apply to flexible elements such as comments.
Anthropic says the watermark will have a negligible effect on Claude’s performance and will not increase usage costs because it does not generate additional tokens. The company is also developing a detection API that will allow users and third parties to check text for the watermark.
Light editing may not remove the watermark, while a complete rewrite can eliminate it. Translations generated by Claude will also carry the watermark because Claude selects all the words. Anthropic says this approach differs from conventional AI detection tools, which look for stylistic patterns rather than a model-specific watermark.
Anthropic will also use C2PA content credentials for supported files created or processed by Claude, including PNG, JPG and SVG files. The cryptographically signed metadata will show that Claude was involved without revealing user information. C2PA is an open standard already used by camera manufacturers and image-editing software.
The company is applying the watermark globally because it does not currently have a reliable way to limit the feature by region. Anthropic is doing so to meet EU AI Act requirements after signing the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside several other major AI providers and around 190 signatories.
Anthropic is also working to add the watermark to models launched before August 2, 2026. Those models are covered by a transition period under the EU rules, with the company planning to introduce the feature over the coming months.
