Anthropic Is Now Watermarking Everything Claude Writes. The Detector Comes Later.
· CX Pulse
Claude now weaves an invisible watermark into the text it generates. It survives copy-paste. Public detection isn't available yet.
Anthropic Is Now Watermarking Everything Claude Writes. The Detector Comes Later.
Anthropic confirmed today that Claude embeds an invisible watermark in the text it generates. Not metadata attached to a file. A mark woven into the words themselves.
From their own help page: "When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won't see it, and it doesn't change the meaning, quality, or readability of Claude's response."
The part that matters for anyone publishing anything: "Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing."
Any Claude model launched on or after 2 August supports it at launch. It sits at the model level, so it applies the same whether the text came from the app, the API, or a tool built on top of one.
Why now
The EU AI Act's transparency rules took effect on 2 August and require AI-generated content to be marked in a way other systems can identify. Anthropic is complying, and the reporting on it is consistent that the marking applies everywhere rather than only to European users.
So a European regulation is quietly setting the default for text written anywhere.
The asymmetry is the actual story
Read the last line of Anthropic's page carefully. They are "working to enable users and other third parties to detect Claude's embedded watermarks and provenance metadata."
Working to enable. Meaning the marking is happening now and the detection isn't available yet.
Everything drafted with Claude since the start of this month is going out into the world carrying a mark that nobody can currently read, in a form that survives being copied and may survive being edited. The detector arrives later and gets pointed backwards at everything already published.
That is a different situation from anything AI-detection has offered before. The AI detectors people have been arguing about for three years are guessers, and they're wrong often enough that accusing someone on their output is reckless. A watermark isn't a guess.
What this means for you
Assume anything AI-drafted is identifiable. Customer emails, product descriptions, proposals, job ads, review responses, the About page. If Claude touched it after 2 August, treat it as marked. Not necessarily today, but at some point when detection becomes something a customer, a competitor or a journalist can run.
Know what it actually proves, because the distinction will get flattened. A positive result means the text carries a Claude mark. It doesn't mean a human never touched it, and says nothing about whether the thinking behind it was yours. In practice, expect that nuance to survive about five seconds in a public argument. Plan for the headline version, not the accurate one.
It changes nothing about whether the work is good. This is worth saying plainly, because the reflex will be to panic and stop using the tools. A well-researched, carefully edited piece drafted with AI is still a well-researched, carefully edited piece. The thing that will embarrass a business is publishing something it never read, and that was already true when nobody could prove it.
The uncomfortable part
"May persist through some editing" is doing a lot of work in that sentence, and Anthropic hasn't said how much editing removes it. Nobody outside the company can answer that yet, which means anyone selling you certainty about it this week is guessing.
Other providers haven't matched this, so text from other models isn't in the same position today. That is worth watching rather than acting on, because the regulation applies to them too.
The five minute version
Have a look at what your business has published in the last ten days and ask a simpler question than "is this detectable."
Ask whether you would be comfortable if it were labelled. Not embarrassed, not defensive, just comfortable. If the answer is yes for a piece, the watermark changes nothing about it. If the answer is no, the watermark was never the problem with that piece.
That is the useful sorting exercise, and it works today, without a detector.
Source: Anthropic, "How Claude marks AI-generated content," Claude Help Center, and TechCrunch, 11 August 2026.