From future models on, every Claude text carries a statistical mark: invisible to the reader, zero extra tokens, untraceable to a person. The reason is the EU AI Act - since 2 August, providers in Europe must mark generated content.
- On 14 August Anthropic explained how the text watermark in future Claude models works.
- The method is DeepMind's SynthID-Text: the choice between equally good words follows a key rather than chance; to the reader the text is indistinguishable.
- Zero extra tokens or cost, nothing hidden in the characters, not traceable to a person or chat; other major providers are doing the same under the EU Code of Practice.
Hold on - this is not a mark that gets stamped. It is a mark that gets rolled like dice.
Let's be honest about what it is for. Not to catch the student - there are cheaper ways for that. It is so that someone with the key can say 'this is probably from a model'. Probably. Not certainly. Anthropic itself writes that the signal weakens in text a human has edited.
We are a publisher. We already say when a machine has touched a visual. Text under our editorial responsibility is ours and carries our name - and no watermark changes that rule. But it is worth knowing: from here on, anyone who copies Claude output one to one carries the mark with them.
Read their explanation. It is short and honest about its limits - a rare quality on this subject.