Does Claude Watermark Text?

Anthropic describes weaving an imperceptible watermark directly into generated text, without changing the meaning, quality or readability of the response. Instead of arbitrary randomness, the key plus the preceding words settle which word the model picks — and Anthropic identifies this as a version of the SynthID-Text approach published by Google DeepMind in Nature in 2024.

The scope is broad. Anthropic states it is applying watermarking globally at launch rather than scoping it to the EU, in the context of transparency obligations under the EU AI Act. Models launched before 2 August 2026 were to be updated over the following months. Separately, supported file types can carry cryptographically signed C2PA provenance metadata.

Anthropic is unusually explicit here: nothing is added to the text and there are no hidden characters. This matters commercially, because a whole category of tools claims to remove Claude watermarks by stripping invisible Unicode. Against this mechanism, that does nothing.

It also means the watermark survives copy and paste, plain-text conversion and reformatting. Those operations preserve the words, and the words are the watermark.

Anthropic's own description is the clearest statement available of what actually affects this class of watermark.

• A complete rewrite in which every word is replaced will remove it; light editing probably will not.
• Text Claude only proofread carries very little, because nearly all the words are the person's.
• Factual passages carry a sparser mark, because there are fewer choices to encode into.
• Code generally carries less, since it often has to be exact.
• Detection does not work well on small samples.
• Translations do carry the watermark, because every word is chosen by Claude.

Anthropic states that a watermark can only determine that Claude was likely involved with the content at some point, and cannot distinguish "Claude wrote this" from "Claude heavily edited this". It says nothing about ownership or authorship and does not change a user's rights.

Anyone treating a watermark hit as proof that a person did not write something is asserting more than the mechanism supports — and, given that proofreading leaves almost no mark while translation leaves a full one, the mapping between the signal and human effort is genuinely loose.

Not yet, in practice. Anthropic says it will soon offer a watermark detection API and is working out implementation details. Until that exists publicly, neither we nor any third-party tool can measure whether a specific passage still carries the mark, and we will not pretend otherwise.