First off, here’s the official statement from Anthropic as to how it works: https://www.anthropic.com/news/claude-text-watermark 

It affects models released after 8/2/2026, and will be retroactively added to models released before that date. They’re doing this to comply with EU regulations, so, if you have a problem with this development, I suggest you direct your disapproval towards the regulators. Anthropic’s loud and early compliance is of course very smug and very Eddie Haskell and very tiresome. But virtue-signaling about the awfulness of some particular corporation is also very tiresome, and I don’t feel like encouraging it. I do feel like trying to make sense of the mechanics of what they’re trying to do, in this particular case.  

At their most basic level, LLMs are an exercise in using statistically determined connections between words to construct an answer satisfying to the “User.” (Yes, I am using Tron-speak in an attempt to be funny.) Anthropic claims that its watermark tweaks the probabilities to favor certain words, in situations where the probabilities do not strongly favor a particular word. In situations where the probabilities strongly favor a particular word or symbol, the watermark does not apply. They use an example at the link that kind of straddles the prose and coding use cases:“2+2=” ends in 4 in any base-10 mathematical context, and also ends in 5 in certain George Orwell contexts, and in neither case are there other options to consider.  

It’s unclear how this watermark would work in coding, which in general behaves a lot more like the example above (“IF ORWELL: 2+2=5; ELSEIF: 2+2=4”) than like prose in general. There’s been a lot of whinging among the coding bros about Claude’s tendency to comment out every step of its coding, so I assume that’s where the watermark is going to show up. 

In the case of Claude editing an existing piece of text, the arrangement of words is already largely settled by the writer. Claude is fixing spelling, punctuation and homophones, maybe doing search engine optimization if the User prompts for it. But there isn’t much in the way of word choice probability for the watermark to fiddle with, if you see what I mean. Maybe you’re giving the watermark an opening to do its thing if you’re handing Claude something really illiterate and incoherent to work with, but if you’re that desperate for help, even watermarked Claude is going to be an improvement. 

Dictation cleanup with Claude is a little bit more of a gray area. There may be chunks of transcribed speech so messy and incoherent that, again, you’re handing the AI watermark room to operate, but in my experience, if the transcript is as bad as all that, you’re going to see Claude’s interpretation of it and go: “No, Claude, you got it ALL WRONG” and go away and heavily rewrite the cleaned-up version. There’s a reason I sometimes call the dictated version a zeroth draft.  

It’s also worth noting that shorter passages, whether AI-drafted or AI-edited, are not going to be statistically significant samples, and therefore are probably not going to show the watermark reliably. This is not really a threat to anyone who already discloses AI use, and not much of a threat to anyone who’s using Claude for support tasks. ESL speakers who are using Claude to translate their thoughts into English and people on the spectrum who use it to translate into Normie-Speak are already being bullied about the obvious Claudisms in their prose; this may or may not intensify it. Basically, what this does is label AI prose for people who are not sharp enough to catch the tics for themselves.

I don’t sell AI-drafted works, only post them for free. This is because I haven’t been able to get them good enough to sell so far, and because what interests me is usually not commercial enough to be worth the hassle of publishing unless I really love it. There are people who can get the AIs to draft/generate works they consider to be of commercial quality. They’re mostly working in niches with voracious readers who demand that authors honor a certain formula and a certain set of tropes and don’t care so much about originality or fancy prose. And that’s fine. I’m not down with some of those genres myself, but there’s nothing wrong per se with wanting, or supplying, formulaic brain candy. If writing were my full time job, instead of a hobby or a side hustle, I’d probably be out there selling Claudefics in hungry, formulaic niches with the best of them. With all those disclaimers out of the way, what can someone who drafts with AI do about this watermark? 

First off, anything you can successfully do to constrain Claude’s writing style in a certain direction is going to reduce the open probabilities that the watermark can fiddle with. I’ve seen knowledgeable people who say that a good “voice print” with grammar and vocabulary structures based on your own writing might do a lot to dampen the watermark effect. Since I’m only starting to fool around with stylometry and similar ideas, and probably won’t post any free claudefics using voiceprints before November 2026, I can’t speak to how well that would work. 

Another option is to rewrite Claude’s output with a different AI, but keep in mind that since watermarking is now a regulatory requirement, all the big AI shops are going to be doing something like this sooner or later. Local models, which run outside any lab’s compliance controls, should help with this, at least if they are based on models released before the regulations went into effect. We shall see.