Claude now puts a hidden watermark in the text it writes. You can't see it, it doesn't change a word of meaning, and it doesn't identify you. But it's there in every longer reply, it follows the text when you copy and paste it, and there's no setting to turn it off. So if you use Claude to help with emails, letters, or reports, it's worth 5 minutes to understand what the Claude watermark is, and who can actually check for it.
The short version: nothing about how you use Claude has to change. But Claude's words, sent out exactly as written, are no longer quite as anonymous as they used to be.
What the Claude watermark actually is
Forget a logo stamped in the corner. This watermark lives in the word choices themselves.
Any time Claude writes a sentence, there are lots of ways to say the same thing. "Quick" or "fast." "Help" or "assist." Normally the pick between them is a bit random. With the watermark, those picks follow a hidden pattern that only Anthropic's checking tool knows how to look for. A person reading it sees ordinary writing. The checker sees the pattern.
A few plain facts about it:
- It's already on. Anthropic started marking text this summer, and today's main Claude models carry it, including the newest Sonnet and Opus models.
- It's everywhere Claude is. The website, the phone and desktop apps, and Claude inside other companies' tools. Worldwide, not just in Europe.
- Images get a label too. Pictures Claude creates, like PNG and JPEG files, carry a digital label called Content Credentials, a sort of signed note in the file that says where it came from.
- Translations count. If Claude translates a letter for you, Claude picked every word, so the translation carries the mark.
And it was driven by a new European law that requires AI companies to mark AI-made content so it can be identified. Anthropic chose to apply it everywhere rather than try to sort users by country.
Does the watermark identify you?
No. This is the part most people worry about, and Anthropic is clear on it: the mark carries no identifying information. It can't be traced to a specific person, company, or chat. It says, roughly, "Claude probably wrote this." Not who asked, and not when.
Anthropic also tested whether the watermark makes Claude's writing worse. People rating watermarked and unwatermarked answers couldn't tell a meaningful difference.
Who can check for it?
Almost certainly not your boss. Not your neighbor. Not just anyone with a website.
Right now, the tool that detects the text watermark is in a private preview. Only approved organizations can use it: regulators, law enforcement, news outlets, fact-checkers, researchers, and educational organizations. That list may grow. Anthropic hasn't said whether it will ever open up to the public.
Images are different. Anthropic offers a free checker that anyone can use to read the Content Credentials on a file.
The catch, and what's still unclear
- You can't turn it off. There's no opt-out on any plan.
- Light editing won't remove it. Fixing a few words leaves most of the pattern in place. A full rewrite in your own words will.
- Short text is hard to check. A 2-line reply doesn't give the pattern enough room. The same goes for very factual passages, like a list of dates, where there aren't many ways to say it.
- It can't tell "wrote" from "edited." If Claude heavily polished something you wrote, the checker may still flag it. All it can really say is that Claude touched it.
- Older models are catching up. A few older Claude models don't carry the mark yet. Anthropic says they will over the coming months.
What to do now
For most people, nothing. Keep using Claude the way you do.
But here's the habit worth building anyway. Treat Claude's draft as a draft. Read it, change what doesn't sound like you, add the detail only you know, and cut what you wouldn't say. Your email gets better, it sounds like a person wrote it, and the question of a watermark mostly stops mattering, because the final version really is yours.
And where honesty matters, like a class assignment, a job application, or anything with your signature on it, check the rules first. If AI help is allowed, say so when you're asked. A hidden mark is a good reason not to be the one who said otherwise.
Turning a Claude draft into something that sounds like you takes about 2 minutes once you know the moves. I walk through exactly that, one real email at a time, on YouTube at @DoWithClaude.