JK Says Anthropic Will Invisibly Watermark Claude Text Worldwide From Tuesday, With Detection Limited to Regulators and Big Companies
anthropic is about to watermark the text claude writes for you, worldwide, starting tuesday. the detector goes to regulators, journalists and big enterprises. everyone except the person who paid for the text. so your cold email, your linkedin post, the resume you rewrote three times can get flagged as claude by some hr tool or editor, and you can't even run the check yourself to see what they'll see. openai only does this in the eu. anthropic rolled it out everywhere. you pay them monthly and they build a snitch layer on top of your work.
Context
If text that Claude touched can be flagged by tools its users cannot run themselves, people who write with Claude may be judged by checks they never see.
Anthropic has published a page explaining that its newer Claude models carry an invisible text watermark to comply with the EU AI Act, and that its detection API is in private preview for groups such as regulators, media, fact checkers and researchers, plus enterprises with compliance duties. The Nov 3 start for more models comes from a customer email that Pete Sena (@petesena) described in a post JK quotes; Anthropic has not published that email. This is a separate story from xAI's plan to watermark Grok API text, also in this deck.
Deep dive
JK (@_junaidkhalid1) wrote on Oct 31 that Anthropic is about to watermark the text Claude writes, worldwide, starting Tuesday, and that the detector goes to regulators, journalists and big enterprises but not to the person who paid for the text. He also claims OpenAI only does this in the EU. Those comparisons are his own and are not confirmed.
His post quotes Pete Sena (@petesena), who wrote on Oct 30 that Anthropic had just emailed customers saying that starting Nov 3 Claude text carries an invisible watermark that survives copy and paste and some editing. That post passed 3.2 million views. Several outlets that saw the email as a screenshot report it covers four older models, Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, and says newer models such as Opus 5.5 already carry the same mark. Anthropic has not published the email itself.
Anthropic's own explainer confirms the basics. It says the watermark changes word choice patterns without adding characters or tokens, carries no information about the user or organization, and exists to comply with the EU AI Act. It says detection works poorly on short passages and only estimates the likelihood that Claude was involved, so a proofread done with Claude could also leave a signal. Its detection API is in private preview for eligible organizations named under EU law and for enterprises with their own compliance duties, and Anthropic says it plans to widen access.
The story is related to, but separate from, xAI's plan to watermark Grok API text by default from Nov 7, which is also covered in this deck.
What to expect next
Watch for Anthropic to publish the Nov 3 schedule itself, confirm which models and cloud platforms are covered, and say whether the watermark applies outside the EU in practice. The bigger open question is access: whether ordinary users and writers will ever be able to run the detector on their own text, and how often it flags text that a person wrote and Claude only edited. Expect other labs to face the same questions as their own EU watermarks roll out.