Anthropic’s watermark isn’t swapping letters around. There are no characters tucked into whitespace. What it does is tilt word choices in ways you’d never spot, and after a few paragraphs of reading, the detector already has what it needs.
Friday’s blog post from the company spelled this out, filling in detail on the watermarking system it announced earlier this week to satisfy the European Union’s Artificial Intelligence Act. Put simply: passing Claude’s prose off as your own is about to become a lot tougher.
Some users didn’t wait for the explanation
For a slice of Anthropic’s paying customers, the post appears to have arrived too late — every sign suggests they’ve already clicked cancel.
X is full of supposed Claude subscribers announcing their exit. John Ennis, an influencer in math and AI circles, shared a screenshot of his cancellation on Saturday and named Anthropic’s “ridiculous watermark idea” as his reason.
Others echoed him. “This is bullsh*t” one user wrote. A second responded to an Anthropic post on watermarking by hurling the r-slur at the company. “Why should I, being a non EU citizen watermark my work generated by a paid subscription of Claude?” a third asked.
How the thing actually works
Anyone who pictured text watermarking as clumsy character substitution — trading the odd “S” for a “$” and calling it done — should read Anthropic’s explanation. That isn’t the technique at all. “The difference between watermarked and un-watermarked text will not be distinguishable to readers,” the company claims.
By Anthropic’s account, the approach traces back to the 2024 SynthID paper, well known in certain circles, meaning it rests on the same foundation as Google’s SynthID watermarking. Google’s own video walkthrough of the text implementation runs through it quickly. Arguably a bit too quickly.
The mechanism, in fuller detail: a secret key — a character string whose complexity and randomness Anthropic likens to pi — weights the token choices that carry the watermark through Claude’s outputs. Where the stakes are low, the system asserts itself. Consider “The weather today was cold and…”, a sentence whose next token might plausibly be “grey” or “overcast.” Compare that with “Paris is the capital of…”, which has precisely one correct continuation, “France” — so watermarking probably won’t engage there at all.
Why you can’t spot it by reading
Those low-stakes decisions get steered by the key toward statistical preferences, and what keeps humans from noticing is that the preferences move with context. In one location “overcast” may be the favored pick. Elsewhere the key may lean toward “grey,” and a reader has no means of telling which rule is in force where.
Accumulate enough of these preferred tokens over a lengthy passage and the detector can verify a Claude-produced watermark. With short texts, the picture changes. The same goes for some coding work, where ambiguity is scarce, and the signal may fail to emerge clearly for the detection system — which is slated to be offered as an API. According to Anthropic, producing the watermark carries a “negligible” cost in speed and tokens.
The editing loophole isn’t a loophole
Editing gets the blog post’s weightiest passage, and it ought to worry anyone assuming that a light touch of Claude leaves no trace. The watermark can appear even where Claude has done nothing but edit text a human wrote.
“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 states. Parse that sentence closely and the inverse speaks for itself. Caveat emptor to every AI-using “editor” out there.
Coverage of the cancellations has turned up Claude users who concede the issue outright: they’d rather clients and school faculty paying for — or grading — human output not detect their AI-generated work.
Whether any of this dents Anthropic
A pinch of salt is warranted on the cancellation wave. It may amount to a blip, or something smaller still, rather than any genuine turn in consumer sentiment toward Anthropic.
Grumbling about Anthropic and rival AI firms is a constant on X, and declaring a cancellation is part of the ritual. At this moment, some ostensible Claude users publicizing cancellations point to an entirely separate grievance: a recent Wall Street Journal piece on the past business practices of CEO Dario Amodei’s wife.
The company’s own figures also cut against the story. Anthropic said cancellations have not risen since watermarking was announced.
For Claude subscribers mulling their next move, the useful question isn’t whether the watermark is fair. It’s whether what you produce runs long enough to carry the signal. Brief outputs and low-ambiguity code might slip past the detector. Nearly everything else won’t.