From our other sites
The story
After AI company Anthropic introduced an “AI watermark” into text generated by its chat assistant Claude — one that’s detectable through quirks in writing style — some users began canceling their subscriptions over it. On 5ch, technical debate continued over how this invisible watermark actually works and how it differs from watermarks embedded in images.
Anthropic’s move to introduce a “watermark” into Claude-generated text is driving some users away from the AI assistant.
“It just keeps inserting watermarks no matter what.”
Richard Echols, an AI consultant based in Georgia, USA, revealed that he canceled his Claude Max subscription on the 12th, saying the watermark was a major reason behind the decision.
Source: businessinsider.jp / Original article here
What people said
The genius and the idiot are the same guy
Europe used to treat Galileo like an idiot too, back in the day
That's "Cloud." Claude is pronounced "Claude" (kurōdo, like the French name)
"Kurōdo" is also the word for a sake-brewery craftsman (a pun on how Claude is pronounced in Japanese)
Stick "Maki" in front of Claude and the reading becomes obvious (a nod to Japanese actor Maki Kurando)
It's expressed through quirks in sentence style and biases in word choice — it's not a watermark you can spot at a glance
Ah, I see.
> According to Anthropic, this watermark travels with the content even through copy-paste, and may survive some amount of editing.
This is presumably built for English, French, and the like. How would it even work for Japanese — some kind of vertical acrostic reading?
Anyone who skips it, or illegally strips it off, will probably end up facing penalties
Interesting angle — I bet pirated copies and government/military use will just quietly run unregulated versions
It's using the fluctuations unique to how AI generates text — the writing itself is the basis for detection
It's not a binary-level watermark being inserted — the point is that quirks in the writing let you identify a specific author. Like if someone always ends sentences with "da'cha" or "dattebayo," you can tell who wrote it (catchphrases from the characters Lum and Naruto). The idea is to embed something like that naturally. Some people have a distinctive way of placing punctuation too, right?
Strictly speaking it's not just sentence endings, but what happens if someone builds a generator that impersonates a specific person? Like if you ask it to "write this in Hiroyuki's style" (referring to 2channel founder Hiroyuki Nishimura's distinct blunt tone), that would clash with the AI watermark. I'd guess building a "Murakami Haruki-style generator" or a "Hiroyuki generator" would conflict with the watermark lol
That's just because you're assuming the watermark can only live in sentence endings lol
Makes me curious how the AI's telltale fluctuations would get embedded when generating text in Dazai Osamu's style lol
Whether you turn it into an image, compress it, or whatever, I think it's the same as long as it's still readable as text — the quirk is baked into the word choices themselves
Oh, sorry, were you talking about images?
That's a slightly different topic. Text watermarks and image watermarks work on completely different principles. A text watermark is about stylistic quirks, but an image watermark operates at the binary level.
Look up "steganography" and you'll see — there have long been ways to embed extra data inside image files.
Huh, I don't get that at all
When it does add them, just tell it not to and they stop — same with other models
"Hello, my name is Yamada Taro. Nice to meet you."
→ "Yo, I'm Yamada Taro! Stoked to meet ya!"
Think of it as the writing style shifting like that (the second line mimics Son Goku's rough speech pattern from Dragon Ball)
Background and Key Points
Anthropic is the AI lab behind Claude, sold in consumer tiers including the premium “Claude Max” plan that consultant Richard Echols cancelled. The “watermark” at issue is not a hidden tag or file-level code; it belongs to a family of LLM techniques (research such as the Kirchenbauer “green-list” method and Google DeepMind’s SynthID-Text) that quietly bias which words or tokens the model favors during generation, so a statistical detector — not the human eye — can later flag the text as machine-written. That is a different animal from image watermarking, which embeds binary-level data via steganography and can survive or be destroyed by recompression. The thread’s Japanese framing is worth noting too: posters compare the mechanism to catchphrases from Naruto and Urusei Yasura’s Lum, invoke Dragon Ball’s Son Goku, and reference Hiroyuki Nishimura, the blunt-spoken founder of 2channel/5ch, as a hypothetical “style” an AI generator might imitate — all local shorthand for “a recognizable verbal tic,” not literal claims about how the watermark works.
The actual disagreement in the thread was technical: some posters (123, 133) argued the watermark could be tied to identifiable stylistic quirks like sentence endings, while others (126, 135, 145) insisted it is purely statistical token-selection bias, unrelated to any single stylistic feature, and a separate sub-thread had to correct posters who conflated this text mechanism with image-file steganography.
What the thread never addresses is the more likely reason for the backlash: forcing a model to favor certain tokens over others can measurably narrow phrasing and reduce output quality, which plausibly explains complaints like “disrupting my work” better than objections to watermarking as a principle.
*This article is excerpted and summarized from the 5ch (News Express+) thread「🤖Users Begin Cancelling Claude Subscriptions… Backlash Against Automatic “AI Watermark” Feature: “It’s Disrupting My Work”」.
Leave a Reply