Anthropic is watermarking every word Claude writes, and it still can't tell you who wrote it

Anthropic has begun stamping an invisible watermark into everything its Claude chatbot writes, and published a post on Friday explaining why users should not panic about it.

The system leaves a pattern in the words the model picks. A reader cannot see it. Anyone holding the key can detect it.

It is built on Google’s SynthID Text, and it is being applied worldwide rather than only in Europe, where the EU’s AI Act requires companies to label AI-generated text.

"We're applying watermarking globally at launch because we don't yet have a durable way to scope it by region," the company said.

Users pushed back hard when the change was announced, arguing it would degrade Claude’s writing and put a scarlet letter on work they had produced with it.

Anthropic says internal testing found no effect on content, creativity or readability. It compared the change to playing Monopoly with digits of pi instead of a die — still random, still no difference to the game, but traceable afterwards if you knew what to look for.

The company also spelled out what the watermark cannot do, and that list is longer than the pushback.

The detector can only assign a probability that text came from Claude. It cannot confirm a human wrote something. It cannot tell whether a different AI produced it, because another company’s watermark would use a different key or a different method entirely.

It does not work on short passages, factual writing or most code, because there are too few word choices available to hide a pattern in.

And it carries nothing about the user.

"There's nothing in the watermark, or its key, that would allow anyone to recover any information about the user, their organization, or their chats with Claude," the company said.

Which matters, because the institutions most desperate for an answer are schools and universities, and the tools they already have do not work.

Research from James Cook University found detection software was wrong in both directions — flagging students who wrote their own work, and missing work that a machine produced.

"You get people trying to write like they're not AI, which then adds whole layers of stress and existential dread," researcher Wayne Bradshaw told AAP.

"You have people inserting artificiality into their work to make it sound less artificial."

The tells everyone repeats — "delve", "tapestry", the em dash — are not evidence of anything.

"They're words which were frequently used by people before they were used by AI," Bradshaw said.

"And the reason they're used by AI is because they were used by people."

Students with autism and students from linguistically diverse backgrounds get flagged more often than others.

James Cook University does not use the software at all. Deputy Vice Chancellor Mitch Parsell said it sets the institution against the student.

"Our approach to securing learning is to focus on the learning design, not on detection," he said.

On Monday, NSW Deputy Premier and Education Minister Prue Car asked the state’s exam authority to consider suspending unsupervised take-home assessments for HSC students, and to develop a common approach for identifying inappropriate AI use.

The best version of that technology now exists. It answers one question — probably ours, or probably not — and it was never built to name a writer.