On August 14, 2026 Anthropic published a technical explanation of the text watermark it is applying to Claude’s output. The method works on the model’s sampling step. When Claude is choosing among words that are all about equally viable, the watermark replaces the ordinary randomness source with a function of a secret key and the few words immediately preceding, so the choice is deterministic to anyone holding the key and looks unremarkable to anyone who does not. Anthropic credits Google DeepMind’s SynthID-Text method as the basis, and traces the idea to a 2022 proposal by Scott Aaronson.
On quality, Anthropic states there is no impact of watermarking on the content, level of creativity, or readability of Claude’s text, and cites Google’s testing finding no statistically significant differences from the unwatermarked model in user ratings. Coverage is forward-looking: the watermark applies to Claude models launched after August 2, 2026, with rollout to older models over the following months. A watermark detection API is described as forthcoming rather than available.
Anthropic is explicit about the limits, and they are the part a deployer needs. Watermarks work poorly on small samples and are sparser on factual passages, because a factual answer offers fewer equally viable word choices to encode into. The watermark cannot identify a specific user or organization, cannot certify that text was written by a human, and cannot detect watermarks from other AI systems. It is a signal that a passage probably came from Claude, not a general-purpose AI detector.
The timing is regulatory, not coincidental. August 2, 2026 is the day the EU AI Act’s transparency obligations became applicable, including the requirement that AI-generated or altered content carry machine-readable marks. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, alongside roughly 190 other organizations. This publication is what compliance with a machine-readable marking duty looks like in practice for a text model, and it makes clear that the honest version of the claim is probabilistic: provenance marking gives auditors and platforms a detectable signal on long passages, not a reliable verdict on a paragraph.