Anthropic has published a detailed explanation of how text watermarking will work in future Claude models, primarily to comply with the EU AI Act. The watermarking method is a version of Google DeepMind’s SynthID-Text approach, which relies on biasing the random choices Claude makes between equally plausible words during generation. This bias is controlled by a cryptographic key and the preceding context, creating a detectable pattern that is invisible to readers and does not affect output quality, creativity, or readability. In internal testing and the SynthID-Text paper’s experiments (including A/B tests on Gemini traffic), no statistically significant differences in quality were found between watermarked and unwatermarked text.

The watermark operates only on low-stakes word choices where multiple options are equally valid. For factual passages or code where exact outputs are required, the watermark is not applied because there is no meaningful choice. The same applies to proofreading: if Claude only lightly edits a human-written text, the watermark may be too sparse to detect. Longer passages provide more information for detection, while short passages may not be reliable. The watermark does not slow down the model or increase token usage, so serving costs remain unchanged.

Anthropic also clarifies that the watermark carries no identifying information about users, organizations, or specific chats. It cannot determine whether text was written by another AI model, nor can it prove human authorship. The company will soon offer a watermark detection API. For images and other file types, Claude will attach a C2PA content credential (a signed metadata note) rather than a watermark. The EU AI Act requires AI providers serving the European market to mark AI-generated content, and Anthropic is rolling out watermarking globally because regional scoping is not yet practical. Older Claude models will receive watermarking over the coming months.

The article also distinguishes watermarking from AI detection software like Pangram, which relies on stylistic tells rather than cryptographic keys. Watermarking can be evaded by a complete rewrite, but light editing is unlikely to fully remove it. Translation carries a watermark because every word is chosen by Claude. Finally, the watermark does not change ownership or legal responsibility for outputs.