Anthropic now watermarks everything Claude writes: an invisible signature that travels with the text

Europe's AI transparency rules kicked in on August 2 and half the industry signed the code of practice: Google, Meta, Microsoft, OpenAI and friends. So far, mostly paper. Now Anthropic has become the first major AI lab to actually flip the switch: its models are already embedding an invisible watermark in every text Claude generates. And it isn't limited to Europe — it applies worldwide.

Confirmation landed on August 11 through the company's support pages, picked up by TechCrunch and Fortune. There is no off switch: the signal is applied at the model level, before the text ever reaches you.

The signature survives copy-paste, but not a deep rewrite

You can't see the mark. It's a statistical, machine-readable signal woven into the model's word choices, one that travels with the text when you copy and paste it elsewhere and that, per Anthropic, "may persist through some editing". As Fortune explains, a heavy rewrite or a translation degrades it to the point of erasure; touching up a couple of sentences does not. It isn't a stamp at the end of the document — it's spread across the whole text.

The rollout covers Claude, the API, Claude Code, Cowork and Claude Tag, and applies to models released since August 2, with a promise to extend it to older ones. Files take a different route: generated PNG, JPG and SVG images ship with C2PA provenance metadata, the same open standard Sony and Leica cameras already use to certify photos.

It only proves Claude touched the text, not who wrote what

Here's the fine print, and Anthropic is the first to underline it: detecting the mark only indicates the content "may have been processed by Claude". A feature written by a human, with three paragraphs polished by the model, will carry the signature just like something churned out whole in ten seconds. And the reverse holds: no watermark is no proof a human wrote it. That's the complaint echoing loudest among people who use AI as a proofreader — whoever supplies the ideas fears wearing a label that suggests otherwise.

The nuance matters with AI slop piling up by the month: a flat label lumps together someone publishing five hundred automated articles and someone who reviews every line. Detection tools for third parties are "coming", the company says, with no date attached. Until they arrive, telling AI text apart remains a matter of instinct.

Broken code, distillers out of business and a watermark eraser already on GitHub

Reactions took hours, not days. The most serious technical worry is about code: a watermark works by nudging the model towards some words over others, and in prose that's harmless: there are a thousand ways to say the same thing. In a program, there aren't. When logic dictates the next token, forcing the alternative can slip in a bug or a vulnerability — and in a contract, it can change the legal meaning. Anthropic hasn't yet detailed how it avoids this in Claude Code, of all things its flagship product for developers.

The second reading is more mischievous: the signature also seeps into the datasets of anyone training rival models by distilling Claude's outputs, a practice as widespread as it is denied that already splashed Kimi K3 and China's open-weight models. If the mark holds, every batch of lifted data ships with the owner's name inside. That a rule sold as transparency ends up shielding the biggest players doesn't look like a coincidence.

And to nobody's surprise, a repository already doing the rounds on GitHub claims to erase the mark by rewriting the text — though without a public detector not even its authors can prove it works. It looks like a bad move to me: the signature won't stop whoever mass-produces junk — they'll rewrite and move on; it shines a light on people using AI well and in the open, and hands a perfect surveillance tool to universities and bosses eager to point fingers. If the price of working with Claude is carrying a snitch inside every text, more than a few will move to Chinese models, which neither mark nor ask. The real answer arrives when Anthropic ships the detector; that's the day we find out how much of the internet's "human" text stops being human.