Anthropic has detailed how it plans to watermark text generated by its Claude models, becoming one of the first major AI providers to publish implementation specifics after agreeing to comply with the EU’s Code of Practice under the AI Act. The regulation requires AI companies serving the EU market to mark AI-generated content so it can be identified.
The watermark will not be visible to users and involves no hidden characters or added tokens. According to Anthropic, internal testing found no measurable impact on the creativity, readability, or content of Claude’s output, and generation speed is affected only negligibly.
How the watermarking works
Anthropic’s approach is based on Google DeepMind’s SynthID-Text method. Rather than modifying finished text or inserting hidden markers, it alters the source of randomness Claude uses when selecting among multiple reasonable next tokens during generation.
Instead of relying on an arbitrary random number generator, the system uses a secret key combined with preceding words to determine token selection. Individual choices remain statistically random and look normal to readers, but across a sufficiently long passage they form a detectable pattern. Anyone holding the key can analyze a text sequence and calculate the probability that it was produced by a watermarked Claude model, without needing computationally expensive access to the underlying model itself.
Exceptions for facts and code
Watermarking is not applied where there is only one correct answer. Anthropic cited factual completions, such as an arithmetic sum, as cases where no equally valid alternative token exists, meaning the watermark’s subtle influence would not be applied. The same logic extends to code, where substituting one term for another could break functionality, so code generally receives less watermarking than other text types.
Rollout timeline
Anthropic is applying watermarking globally at launch rather than restricting it to the EU, stating it currently lacks a durable way to scope the feature by region. Future Claude models will generate watermarked text by default. Models released before August 2, 2026 fall under the EU’s transition period, and Anthropic says it is working to extend watermarking to those existing models over the coming months.