John Gruber calls Claude’s text watermark a perversion of writing
John Gruber has turned into the sharpest voice against Anthropic’s decision to watermark everything Claude writes. The Daring Fireball blogger, who co-created the Markdown markup language, called the plan a “perversion of writing” in a post on 16 August. His objection isn’t really about EU compliance, and he says he’d have written the same thing without it. What bothers him is that watermarking gives Claude a reason to pick a word other than the best one.
Anthropic had published its own explanation two days earlier, on 14 August, and it argues the quality worry is unfounded. The company says nothing gets added to the text and there are no hidden characters. What changes is where the randomness in word selection comes from.
Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick.
Anthropic, via its own explainer
The worked example is a sentence beginning “The weather today was cold and…”, where “gray” and “overcast” both fit. Those are the low stakes decisions the mark rides on, and Anthropic says its testing found no impact on content, level of creativity, or readability. That’s the entire defence: if the swaps are genuinely free, Gruber’s complaint has nothing to attach itself to.
He rejects the premise rather than the engineering. Once the key influences a single decision point, he argues, something other than the reader’s needs is shaping the prose. He wrote that “the exact words we choose when writing matter”, a line Business Insider picked up as the core of the complaint.
The idea that anything other than my needs should factor into the generation of text for me is patently offensive.
John Gruber, Daring Fireball
ZDNet’s reply was that the premise was never true. Claude is a product owned by a company, and it was never optimised solely for you. Gruber’s own framing does concede one thing, though: he calls it unacceptable for a tool to sacrifice “an iota of clarity, coherence, meaning, quality” for provenance clues, which makes the whole argument turn on whether any iota is actually lost.
Removal tools showed up within hours
But while the argument ran, other people were simply routing around it. Developer Guillaume Meyer published an override within four hours of Anthropic confirming the rollout, Wired reported. His code has since drawn more than 100 contributors on GitHub and more than 20,000 bookmarks on X.
| Who | Method | Detail |
|---|---|---|
| Guillaume Meyer | Rewrites via a non-watermarking model | Swaps synonyms and reorganises content |
| Erik Hughes | Character and sentence shuffling | Built in 15 minutes, using Claude |
| Leon Chlon, Oxford | Round-trip translation | Condense, translate to a dialect like Arabic, translate back |
None of it is confirmed to work, because Anthropic hasn’t shipped the detector yet. A spokesperson told Wired the company plans a text-detection API so users can check for themselves. Until that lands, nobody outside Anthropic can score their own removal attempt, and third-party AI detectors are a poor substitute.
The pressure behind all of this is regulatory, which is why the deadline matters more than the discourse. Around 190 organisations signed the EU transparency code of practice, OpenAI, Microsoft and Meta among them, and providers who don’t mark synthetic output face fines of up to 3 percent of annual turnover. Meyer told Wired he’s “not against transparency” and is “all for content attribution”, but thinks watermarking carries too much false-positive risk.
Still, not everyone reads the change as a loss. Writing on 21 August, Zvi Mowshowitz called the technique free and good, saying it “has no practical impact on outputs” and that “the marginal cost of doing this is very close to zero”. He points out that Google shipped the same idea years earlier, with a public detector.
That gap is what to watch. Anthropic told Business Insider it hasn’t seen a spike in cancellations, so the commercial damage so far looks smaller than the noise. The argument only resolves when the detector ships and someone can measure both claims, Gruber’s about quality and the coders’ about removal, against something other than assertion.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
