Claude adds invisible text watermarks, with detection limited to selected parties
A Reddit post citing Anthropic's help page says Claude's statistical text watermark is extended globally to several models from November 3, and detection is restricted.
Claude's text output is set to carry an invisible statistical watermark across more models starting November 3, according to a Reddit post that cites Anthropic's help page. The post says the mark extends to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, and that newer models are already covered. Detection is limited, the post says, to selected regulators, media, researchers, educational institutions and some compliance-heavy enterprises, not ordinary users.
The restriction on who can check matters most in practice. The watermark is invisible, so ordinary text reads normally to a person, but a detector can find it. If only a short list of parties can run that detector, a reader who receives a Claude-written document has no way to check it. The verification power sits with the institutions the post names. The post doesn't say how those parties get access or what they can do with a result.
Commenters on the thread raised the reliability of detection. One said a 2020 master's thesis was flagged as 56% likely AI-written. That is a score rather than a verdict, and the concern was about false positives.
The post also raises edge cases. A commenter described Opus 5.5 editing an email and changing one period and two words, and asked whether a mark would survive that kind of light edit. The post argues that a paraphrase through another model could remove a lexical or statistical watermark. It also says third-party detectors might learn the scheme through watermark-stealing attacks.
A commenter drew a line between a mark that says content was made by Claude and a mark that identifies a specific user, such as a user ID. The first is a provenance label for the content. The second would tie text to an individual account, which the commenter considered unacceptable. The post doesn't say which kind of information the watermark carries.
What changed
Claude text output is set to carry an invisible statistical watermark across more models, with detection limited to selected parties, according to a Reddit post citing Anthropic's help page.
Type: Other · Applies to: Mythos 5, Opus 4.7, Opus 4.6, Sonnet 4.6
Why it matters
Companies that publish or review Claude-assisted text may face attribution questions they cannot check themselves. They need clear internal rules on how AI-assisted content is labeled.
What you can do
Legal and compliance teams can review whether Claude-assisted documents need provenance records, since end users may not be able to verify watermarks.
Original source
We have no link to the original source for this story yet.
Where we saw it
Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works