Anthropic explains how the watermark in Claude's text works
Anthropic published an official explanation of the mechanism that writes a watermark into text Claude generates. With the new European labeling rules, the subject stops being a technical curiosity and enters compliance territory, because it offers a way to establish the origin of content even when whoever publishes it does not say so.
On August 14, Anthropic published answers to questions about how the watermark in text Claude generates works. The mechanism derives from SynthID-Text, which comes out of Google DeepMind. The timing is not accidental, since European regulation has required transparency since August 2.
How it works
No hidden characters are inserted into the text, and the response is not altered afterward. What changes instead is the source of randomness the model uses for part of its word choice.
Those choices are individually meaningless, but a long text contains many of them, and together they leave a pattern. A reader does not notice it, and anyone holding the key that pattern was written with can recognize it.
Why now
Labeling that depends only on the goodwill of the publisher falls away the moment someone moves the text into another system. A watermark written into the text itself tries to make origin travel with the content.
Anthropic thereby joins OpenAI and Google in explaining how it intends to meet the European requirements. According to the support page, new Claude models released in the Union since August 2 support labeling from the start.
What this does not solve
The watermark survives copying but not rewriting. If a person or another model substantially reshapes the content, the pattern is lost. This is a means of establishing origin, then, not proof that something was not made with AI.
Sources
Related
How to label AI content under Article 50, and which part of it is not your job
Article 50 of the EU AI Act has applied since August 2, 2026, and it binds anyone serving people in the Union, wherever the server is. Most of the panic is about the machine-readable marking requirement, which for a site owner who calls somebody else's API is somebody else's obligation. Here is what is actually yours: a chatbot that says what it is, published text that either carries a name or carries a label, and a deepfake that admits it.
Evropska komisijaverified
Pew puts a number on it: 35% of pages published after ChatGPT show AI authorship
Pew Research ran nearly half a million pages from Common Crawl through a detector. Among pages published after November 2022, more than a third came back with significant signs of AI writing, and .com domains showed it at ten times the rate of .edu and .gov.
TechCrunchverified
Gemma passes a billion downloads and 100,000 derived models
On August 20, Google announced that the Gemma family of open models has passed a billion downloads. In two years the community has published more than 100,000 derived models, and the most recent Kaggle competition drew more than 1,600 entries. An Awesome Gemma repository launched with the announcement, meant as an index of vetted projects, fine-tuned models, tutorials, and tools. The figure matters to someone building sites too, because a model you run on your own server stops being exotic, which makes text processing without sending data to a third party workable.
Googleverified