What Claude's Text Watermark Actually Does,and What It Cannot Do

Two things people repeat about Claude's text watermark are wrong: that web crawlers will pick it up, and that retyping the text into Word removes it. Here is what the watermark is, who can read it, and why it changes nothing about how you work.

What the room believed

I sat in a room of thirty people at an Ask Us Anything About AI forum in Greenville where both of those were said, and I let them stand.

The room had heard it lands in November, that crawlers would pick it up, and that the fix was to generate the text in Claude and retype it into Word. Someone said they had better generate all their content now, and someone else said, you have got a deadline.

I added that it was there to satisfy EU rules and that no US law asks for it. That part was right.

The next morning I read Anthropic's own page, dated August 14.

There is no hidden mark

The room was picturing an invisible mark buried in the code.

There is none.

Nothing is added to the text. When two words are equally good, overcast or grey, a key settles which one Claude picks. Over a few paragraphs, whoever holds that key can read the pattern.

Nobody else.

So retyping does nothing. The mark is the words Claude chose, and copying them keeps it. The page says light editing probably will not remove it, and that "a complete rewrite where every word is replaced" will, at which point you wrote it.

No crawler can read it

Not Google's either, even though the method is a version of Google DeepMind's own SynthID-Text, published in a Nature paper in 2024.

Reading it takes Anthropic's key. As of September 1 that key sits behind a private-preview API for regulators, media, researchers and other listed groups, including companies with their own EU obligations.

The EU code of practice Anthropic signed calls that restriction temporary, with an interoperability deadline of February 2, 2027.

Why it exists, and why it reaches you anyway

The reason is the EU AI Act. The reach is everywhere.

Anthropic's page is blunt about it: "We're applying watermarking globally at launch because we don't yet have a durable way to scope it by region."

A business in Greenville with no European customers gets the same watermark.

The deadline joke was right, and I had it backwards

On August 27 no Claude model carried the watermark yet, as far as Anthropic's release notes show. Opus 5 came out July 24, Sonnet 5 on June 30, and Fable 5 on June 9 - all before the August 2 threshold. The page says those older models will get it "over the coming months."

The first watermarked text came from Fable 5.1 and Mythos 5.1, two models released on September 1. Five days after that joke.

I had that backwards until I read the release notes.

What it can actually prove about your writing

If someone runs a detector on your text, the most it can tell them is that Claude was probably involved somewhere.

Anthropic's page says it "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'"

It cannot prove you did not write something.

I am not changing anything I do because of it.

What do you think of the watermark? Would you stop using Claude because of it?