Anthropic’s Claude is going to start watermarking the text it generates by making subtle word choices which, over the course of a long document, may reveal if it was largely AI generated. This initiative started due to a European Union regulation designed to help people understand if creative works are AI generated.

The watermarking made Daring Fireball’s John Gruber very angry. He’s right to say it’s embarrassing that Anthropic won’t restrict this feature to EU citizens only. And Anthropic’s first article about the topic was vague and unhelpful. Beyond that it turns into a weird, irrelevant defense of the concept of word choice in writing.

M.G. Siegler is equally incensed, linking approvingly to Daring Fireball’s article. Spyglass’ article is titled “You Gotta Fight For Your Write”, implying that the specific pseudorandom number generator Anthropic used previously is a human right. Or am I reading it wrong, and that right is as inconsequential as your right to party?

I’m tempted to fsck choice bits of Daring Fireball’s article, but my knowledge of LLMs isn’t up to snuff. Given the gravity of the terms he uses:

• “a perversion of writing”
• “sacrifice an iota of clarity, coherence, meaning, quality”
• “adulterate and corrupt the semantics of the text”
• “anything other than my needs should factor into the generation of text for me is patently offensive”
• “make it worse, on purpose, for purposes that do not benefit me in any way”
• “it’s only slightly worse and that everyone is either too stupid to notice or too indifferent to care”
• “Poisonous is exactly the right word”

you would think this was a tremendous mangling of the text. The examples he cites are indeed relevant — mangoes are not bananas, a sky can be grey or overcast for different reasons — but that’s not the context of how this feature works. This is for long blocks of text, many paragraphs, where the model calculates the word choice isn’t important. If you’re having Claude write a story about Nomango Noway, the man who is allergic to mangoes, it would definitely be a problem if the model picked mangoes as part of its watermark when discussing his favorite tropical fruit. But that’s not a trivial word choice! If you belive Anthropic, that’s not what’s going to happen.

Again, this isn’t my forte. See posts by Alex Hern and (sigh) conputer dipshit for a little more info. I found this on Techmeme where it’s funny to see the difference in opinion between X posts (how dare the EU/Anthropic alter their AI’s word choices) and BlueSky posts (what’s the big deal, something needs to be done about people passing AI writing off as their own).

Beyond the fact that he covers technology in general, I’m baffled that Gruber chose to drill into this topic and defend the honor of writing here. Is it because duty calls – the same reason I’m writing this?

Anthropic isn’t a literary critic or fellow author disparaging the idea that word choice is important. It’s an AI lab responding to a regulation that addresses public concern over AI output being disguised as human output. Images and videos are possible to watermark on the file in a manner invisible to the user. Text is harder, sure, but the cost of one word being slightly different per paragraph is worth a stronger signal that people are copying AI content without a disclosure.

It’s called a collective action problem, the thing governments are meant to solve. You might have to make a tiny sacrifice to make a better world for everyone.

Were I a commercial blogger I’d be delighted by this. Writing produced by Claude and the other AI labs will be worse, making my own writing look better. It only applies to large blocks of text, which is not something I’m in the habit of copy-pasting into my own work. Other writers who use AI writing in a deceptive manner will have a stronger signal to have their poor choice exposed. Even if I’m not regurgitating Claude’s writing verbatim – if I were foolish enough in my use of AI for these specific word choices to affect how I absorb the output – I might be relying on it too much.

I feel like John’s distaste for EU regulations in general is leading him down this weird path of being annoyed at something that exclusively highlights him as a good writer. It also gives him a chance to call people other than him “stupid,” another favorite pastime of Daring Fireball. I agree that it’s embarrassing that Anthropic can’t geofence this, but that’s one paragraph, not 181.

I don’t know if writing needs writers to stick up for its honor. But if so, writing is the best method.