Anthropic explains how Claude's AI text watermark will work

Anthropic has provided more details on how it plans to watermark text generated by future versions of Claude, saying the system is designed to identify the model’s involvement without changing how the text looks or reads.

The company said its watermarking method does not add hidden characters, extra tokens or identifying information to the output. It also said the process will not increase the cost of running Claude and will have a negligible effect on model speed.

How Claude’s watermark will work

Anthropic said Claude generates text by selecting the most appropriate word from several possible options. In cases where multiple words would work equally well, such as choosing between words with similar meanings, the model has some flexibility in its selection.

The watermarking system uses those low-impact choices to create a statistical pattern. Instead of relying on an ordinary source of randomness when selecting between possible words, Claude will use a key along with preceding words to determine the choice. Someone with access to the key can then examine a passage and calculate the likelihood that Claude generated it.

Anthropic said readers will not be able to distinguish watermarked text from unwatermarked text. Its internal testing also found no effect on the content, creativity or readability of Claude's writing.

The company is using a version of Google DeepMind's SynthID-Text approach, which was published in 2024. Anthropic said the system has limitations and cannot establish with certainty that a passage was written by Claude. It instead indicates the likelihood that Claude was involved. Shorter passages are also harder to assess because they contain fewer word choices from which a watermark can be detected.

Editing may weaken the watermark

Anthropic said minor editing is unlikely to completely eliminate the watermark, while replacing the entire text can remove it.

The watermark also behaves differently when Claude is used to edit existing human-written material. If Claude only makes limited grammar or punctuation changes, there may be too few words selected by the model for a detectable signal to emerge. More extensive editing gives the watermark more material to work with.

The company stressed that detecting a watermark does not establish who authored the content. It can only indicate that Claude was likely involved at some stage, whether through writing or substantial editing.

Code will carry less of a watermark

Anthropic said code generally presents fewer opportunities for watermarking because many programming decisions have to be exact. Changing a particular term can make code incorrect or cause it to fail.

Where there is flexibility, however, watermarking can still be applied. Anthropic specifically pointed to comments in code as an area where arbitrary wording choices may provide room for the system to operate. The company said this should have a negligible effect on the code itself.

The same principle applies to factual passages, where there may be only one accurate answer. Anthropic said such passages tend to contain less watermarking.

No user information in the watermark

Anthropic said the watermark cannot be used to identify a particular person, organisation or conversation. The signal applies to Claude's output rather than to the user who generated it.

The company is introducing the system as part of its compliance with the EU AI Act and said it will apply the watermark globally because it does not currently have a durable way to restrict the system by region. Other major AI developers that signed the EU Code of Practice are also expected to implement their own watermarking systems.

Anthropic said it will soon offer a watermark detection API, although details of its implementation are still being worked out.

The company also clarified that translations produced by Claude will carry a watermark because the model selects the words in the translated output. Older Claude models launched before August 2, 2026 are covered by a transition period, with Anthropic working to add watermarking to them over the coming months.