How Anthropic plans to mark future text generated by Claude

Future Claude models will include an invisible statistical watermark designed to detect their involvement in a text without adding hidden characters.

Future Claude models will include a text watermark designed to estimate the likelihood that the model was involved in writing a piece of content. Anthropic says it is implementing the mechanism to comply with European requirements for marking AI-generated content. The system will initially be applied globally, as the company says it does not yet have a sufficiently durable way to limit it by region.

The system is based on a variation of SynthID-Text, a method published by Google DeepMind. No hidden characters or additional elements are inserted into the text. Instead, the watermark operates when the model chooses between multiple words considered equally acceptable. A key changes the source of randomness used for those choices, gradually creating a statistical pattern that can be detected afterward.

Anthropic says this modification does not create any perceptible difference in content, creativity, or readability, and does not require any additional tokens. The watermark also contains no information that could be traced back to a specific user, organization, or conversation.

Detection remains probabilistic and depends heavily on how much text Claude generates. In a short, highly factual, or tightly constrained passage, the model has fewer possible choices and therefore leaves a weaker signal. The same applies to code, where many tokens cannot be arbitrarily replaced without changing or breaking the result. Comments and certain non-functional choices, however, can carry more of the watermark.

Editing is more nuanced. When human-written text is only proofread or lightly corrected, too few words may be changed for Claude’s involvement to become detectable. A more substantial rewrite leaves a stronger signal. Translations, however, are watermarked because every output word is chosen by the model.

Anthropic also acknowledges that the system can be weakened by editing the text. Minor changes may not necessarily remove the signal, while a complete rewrite can make it disappear. The watermark also cannot establish that a text was entirely written by Claude. It can only estimate whether Claude was likely involved at some point.

A detection API is expected to be offered later. For files such as PNG, JPG, or SVG, Anthropic will use a different approach based on C2PA Content Credentials embedded in the metadata and cryptographically signed. This mechanism indicates that a file was produced or processed with Claude without altering its visual content or including personal information.