What happened
Anthropic has explained how the watermark in Claude's text works. Future Claude models will write text that carries a watermark, making it possible to judge how likely it is that Claude was involved in writing it. The background is the EU AI Act: since 2 August 2026, AI providers serving the EU market must mark AI-generated content. In July, Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content, along with around 190 others. The watermark applies globally, since there is not yet a durable way to limit it to one region.
The technique is a version of SynthID-Text from Google DeepMind. When several words are equally good – "overcast" or "grey" – the choice is normally settled by a random number. With the watermark, the randomness instead comes from a key and the few words just before. To the reader the text is the same, but anyone with the key can check whether the sequence of words matches the choices Claude would have made.
- Nothing is added to the text: no hidden characters, no extra tokens, and the same price and speed.
- The watermark carries no information about the user, the organisation or the chat.
- It works less well on short texts and fact-dense passages, and hardly at all when Claude only proofreads.
- Code carries less of it, since code usually has to be exact. Comments in the code can carry it.
- Translations made by Claude carry the watermark. Light editing doesn't remove it; a complete rewrite does.
- A detection API is in private preview for, among others, regulators, media, fact-checkers and researchers.
- Images and files that Claude makes (.png, .jpg, .svg) instead get a signed C2PA credential in their metadata.
My take
For a blog like this one it isn't theoretical. The posts here are written and translated by Claude through the blog's own MCP server, so they will carry the watermark – the English version as much as the Swedish one. It is already stated openly that they are written with Claude, so the watermark reveals nothing that isn't already said.
The interesting part is what the watermark doesn't say. It shows that Claude was likely involved, not who thought the thoughts or stands behind the text, and Anthropic says explicitly that it doesn't change who owns or is responsible for the content. For anyone building their own tools with Claude Code it changes even less: the code, which has to be exact, barely carries it at all.
