Anthropic published a detailed explainer on August 14 of how Claude’s text watermarking will work, confirming that future Claude models will generate text carrying a statistical pattern that can identify the likelihood the model wrote it. The company’s announcement says the change is being implemented to comply with the EU AI Act, which since August 2 requires AI providers serving the European market to mark AI-generated content.
Anthropic says it is not alone in making the shift: other major model developers signed the same EU Code of Practice and will implement their own watermarks. The company stresses that the technique differs from visible watermarks on banknotes or documents. Readers will not be able to tell the difference, nothing is added to the text, there are no hidden characters, and the watermark carries no identifying information that could be traced to a specific person, organization, or chat.
How the watermark works
Claude’s text watermark is a version of the SynthID-Text approach that Google DeepMind published in a Nature paper in 2024, part of a family of techniques that traces back to a 2022 proposal by Scott Aaronson. The core idea exploits the fact that language models pick the next word from a list of reasonable candidates. In low-stakes choices where several words work equally well, the selection is normally settled by a random number. Watermarking changes the source of that randomness: instead of an arbitrary generator, the model uses a key and the preceding words to decide. The result is a pattern across a passage that a detector holding the key can check, assigning a probability that Claude wrote the text.
Anthropic says internal testing found no impact on the content, creativity, or readability of Claude’s output. The technique also builds on Google DeepMind’s SynthID-Text experiments, which served watermarked text to a portion of Gemini traffic and found no statistically significant differences in thumbs-up and thumbs-down ratings, and on a controlled study in which human raters saw no quality difference between watermarked and unwatermarked answers.
What the watermark can and cannot prove
There are real limits. The watermark only answers one question: the likelihood that text was partly written by Claude. It does not confirm that text was human-written, and it cannot tell whether another AI wrote it, since a different provider’s watermark would use a different key. Detection also weakens on short samples, where there are fewer word choices to leave a pattern, and the watermark is sparser on factual passages where only one word is correct. Code gets less watermarking than prose, because exact outputs have little room for arbitrary choices, and lightly edited human text may carry too few Claude-chosen words to register.
The rollout has broader implications for how AI content is attributed. Because the watermark is tied to a hosted service’s key, models that users run on their own hardware carry no such marker at all, a distinction that matters for the local AI case we have covered. Anthropic has also been expanding its hardware ambitions, announcing earlier this month that it will design its own chips to power Claude. The watermarking explainer arrives the same week Google launched Gemini 3.7 Flash, underscoring how quickly the largest AI providers are moving on both model capability and compliance. The watermark will apply to future Claude models, with the exact rollout timeline left to the models’ release schedules.
