Anthropic recently announced that it is adding watermarks to text generated by future Claude AI models. The Dario Amodei-led firm says the move is being made to comply with the European Union’s AI rules, which require providers to mark AI-generated content.
The watermark will not appear as visible text or symbols. Rather, Anthropic says it will create a pattern in the way Claude selects certain words. This pattern can later be checked using a specific key to estimate whether Claude was involved in generating the text.
Anthropic said, “Future Claude models will generate text that contains a watermark.” It added that it is making the change alongside other major AI providers that have signed the EU’s Code of Practice on AI-generated content.
How Claude’s AI Watermark Will Work
Large language models choose the next word from several possible options while generating a response. In many sentences, there can be multiple words that make sense without changing the overall meaning.
Anthropic says its watermarking system will use these low-stakes word choices to create a detectable pattern. The words themselves will not contain any added characters or visible marks.
According to the company, “Nothing is added to the text and there are no hidden characters.”
The system changes the source of the randomness used when Claude selects between possible words. A person reading the response would not be able to identify the watermark, but someone with the relevant key could analyse the sequence of words and calculate the likelihood that Claude generated the text.
Anthropic says the watermark is not designed to change how Claude’s responses look to users. The company states that “The difference between watermarked and un-watermarked text will not be distinguishable to readers.” It also says its internal testing found no impact on the content, creativity or readability of Claude-generated text.
However, the watermark does not work equally well in every situation. It depends on Claude having enough opportunities to make choices between different words.
For factual statements where only one answer is correct, there may be little opportunity for the watermark to operate. Anthropic gives the example of factual information about Isaac Newton, where changing a specific word could make the statement incorrect.
What Happens With Code And Editing?
The watermark will also have limited scope when Claude is working with code. Anthropic says the system generally does not apply its watermark where an exact output is required and changing a word could cause the code to fail. However, the watermark can appear in areas such as comments within code, where there may be more freedom in word selection.
The same applies when Claude edits human-written text. If Claude only makes minor grammar or punctuation changes, there may not be enough Claude-generated words for the watermark to be detected reliably. Anthropic says, “The more Claude writes, the more decisions it has to make, and the more space there is for a watermark.”
Watermarking Will Not Add To The Cost
Anthropic says adding the watermark will not require additional tokens. The company states, “Watermarking has a negligible impact on the speed of models, and because it produces no extra tokens, the model is the same price to serve and use.” This means users are not expected to pay more simply because their Claude-generated text contains a watermark.
Anthropic says models launched before August 2, 2026 are covered by a transition period under the EU rules. The company is working to add watermarking to those models as well, with the rollout expected to happen over the coming months. The watermarking approach described for text is separate from how Claude-generated files will be marked.
