Anthropic’s watermark doesn’t swap letters. It doesn’t hide characters in whitespace. It nudges word choices you’d never notice, and by the time you’ve read a few paragraphs, the detector already knows.
The company laid this out in a blog post Friday, expanding on the watermarking system it announced earlier this week to comply with the European Union’s Artificial Intelligence Act. The short version: it’s about to get harder to pass off Claude’s writing as your own.
Some users didn’t wait for the explanation
The post may have landed too late for a chunk of Anthropic’s paying customers, who by all appearances have already hit cancel.
Posts on X show plenty of supposed Claude subscribers announcing their exit. Math and AI influencer John Ennis posted a screenshot of his cancellation Saturday and pointed to Anthropic’s “ridiculous watermark idea” as the reason.
He’s got company. “This is bullsh*t” wrote one user. Another replied to an Anthropic post about watermarking by calling the company the r-slur. “Why should I, being a non EU citizen watermark my work generated by a paid subscription of Claude?” asked another.
How the thing actually works
If you assumed text watermarking meant crude substitutions, swapping some instances of “S” for “$” and calling it a day, Anthropic’s explanation is worth your time. That’s not what’s happening here. “The difference between watermarked and un-watermarked text will not be distinguishable to readers,” the company claims.
Anthropic notes the system comes out of the 2024 SynthID paper, famous in certain circles, which means it’s built on the same underlying technology behind Google’s SynthID watermarking. Google’s own video summary of the text implementation covers it fast. Maybe a little too fast.
Here’s the mechanism in more detail. Claude’s outputs get watermarked through token choices weighted by a secret key, a string of characters Anthropic compares to pi for complexity and randomness. The system leans in when the stakes are low. Take the sentence “The weather today was cold and…” where the next token could just as easily be “grey” or “overcast.” Now take “Paris is the capital of…” There’s exactly one right answer, “France,” so watermarking is unlikely to kick in at all.
Why you can’t spot it by reading
The key pushes those low-stakes choices toward statistical preferences, and the reason a human can’t catch them is that the preferences shift with context. “Overcast” might be favored in one spot. Somewhere else the key might prefer “grey,” and there’s no way for a reader to know which applies when.
Stack up enough of those preferred tokens across a long passage and you have a watermark the detector can confirm came from Claude. Short texts are a different story. So is certain coding work where there’s little ambiguity to exploit, and the signal may not surface clearly for the detection system, which is set to be available as an API. Anthropic says generating the watermark has a “negligible” impact on speed and token cost.
The editing loophole isn’t a loophole
The most consequential part of the blog post covers editing, and it should give pause to anyone who figured light-touch Claude usage stays invisible. Even when Claude is only editing human-written text, the watermark may show up.
“Depending on the length of the text and how heavily Claude has edited it, those changes might not be enough to make Claude’s involvement detectable,” the post says. Read that carefully and the flip side is obvious. Caveat emptor to every AI-using “editor” out there.
Reporting on the cancellations has surfaced Claude users who openly admit the problem: they don’t want their AI-generated work detected by clients and school faculty who are paying for or grading human output.
Whether any of this dents Anthropic
Take the cancellation wave with some salt. This might be a blip, or less than one, rather than a real shift in how consumers feel about Anthropic.
People complain about Anthropic and other AI companies on X constantly, and claiming to cancel is part of the ritual. Right now some apparent Claude users announcing cancellations are citing a completely different reason: a recent Wall Street Journal article about CEO Dario Amodei’s wife’s past business practices.
And Anthropic’s own numbers don’t back the narrative. The company said cancellations haven’t increased since watermarking was announced.
If you’re a Claude subscriber weighing your options, the practical question isn’t whether the watermark is fair. It’s whether your work is long enough to carry the signal. Short outputs and low-ambiguity code may not register with the detector. Everything else probably will.
