Post

Caring about the quality of AI prose is the source of our collective confusion. I don’t expect watermarking to work well either. But that someone who claims he cares so much about writing would even consider using AI at all is simply shocking. That you called your critics “idiots” is just petty.

damn, maybe you should just develop and write your own ideas

That was hard to read. Never had the feeling AI produced „good writing“ anyway, so why cry about it? Also some regulation finally is wanted and desperately needed. Know this: the reason I read your blog is exclusively because you, a human, finds, selects, comments on important news by hand.

I'm perfectly OK with making plagiarism and AI laziness more readily detectable. Cynically, I think this is actually an effort to make it easier for AI companies to avoid ingesting generated content in their training data.

Actually, anything that makes it easier for us to point and laugh at losers who use this shit to "write" their slop is good, based and above all else, extremely funny.

Lolololol. It's funny seeing plagiarism app users cry about this. And the funniest thing about your complaint is that almost certainly 1 reason they're doing it is to protect their app from model collapse, i.e., being reduced to gibberish by being trained on its own vomit. bsky.app/profile/sjon...

Play in what sense? If you're a writer, you'll very much prefer an LLM with a higher temperature setting, with more random results. LLM with low temperature lacks imagination, is bad at many types of reasoning etc. But it's good for classification tasks for example. Which of them is "high quality"?

What does this have to do with your original argument? We’re saying adding watermarking isn’t , on aggregate, different from slightly raising temperature, and therefore can be done in a way that doesn’t affect the quality of output.

EXACTLY! I felt like I was takin' crazy pills reading John's essay. This isn't like it's fudging math or changing variable names in code. Writing is largely subjective anyway, and -- as you rightly note -- LLMs were never picking the "best" or "most precise" words!

They are based on probability and, hence, are non-deterministic. "A language model estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens." developers.google.com/machine-lear...

He is.

Is this supposed to be a bad thing? I don’t understand why this is bad.

wow, the replies here show me that blue sky is made up of the most smug, condescending and rageful people on the internet. Which is saying something.

They already don’t care about nuance or shades of meaning or meter or anything else. If this watermark makes it easier to catch these people then awesome, fuck them, they should choose their own goddamn next word and not let the robot do it

Is bad, shitty regulation worth it just because we don’t like who it happens to?

No but for real I predicted a reflexive response but not something that unhinged

For example he bitched about eu usb c enforcement from a technology progress position and never once conceded the practical and environmental benefits But the length and tone of his latest screeds betray something else. There is a detachment from reality evident

I mean, they are comically foolish when it comes to privacy, E2EE, technology in general. Much like most other governments around the world, especially the morons in charge of the US right now.

Claude is merely a tool and isn't inherently good or bad, aside from works stolen for initial training. Its very useful for generating automated software testing regimens and for finding bugs in complex codebases.

“Clarity, coherence, meaning and quality,” he says! As if the “tool” actually does ANY of those things!!! Hahahahahahahaha!

But the tool knows nothing of clarity, coherence, meaning, quality, et cetera…

he's not saying they aren't wanted he's saying they are trivial to defeat and have the additional problem of making LLM output less optimal for valid uses of the technology (like code reviews, not having it be your girlfriend!).

All the time! My previous home is right along the city boundary, but Google Maps has the boundary wrong and says it's in the neighboring city. Considering how easy the official maps are to find, I'm sure that's not an accident!

You will note that Anthropic doesn't provide before / after examples. Nor do they yet provide a watermark detector. Google apparently has been using this in Gemini for a while and don't provide a detector either.

In most cases *including the examples John gives*, the diff is largely in preference, not meaning, since the contextual intent itself generally can't be inferred. And even if I'm wrong there, if John really prefers bananas vs. pineapples, *he can just substitute that text if/when he repurposes it!*

And is watermarking really going to bias those percentages away from the “right” word, more than the system prompt, or the model’s training?

Awww poor baby, do you have to write things yourself now?

And that says something Not anything good in my view But something

My god, the responses to your position here are unhinged.

That is ridiculous. Corollary- no startup should spend money on green energy because that wastes VC money. You give up clarity every time you watch streaming video or don’t listen to vinyl. People put secret messages in text for millennia. At least it’s not an 88.

Unless, apparently, that “tool” (😉) is a human being sacrificing his or her remarkable, unique and unpredictable abilities by inserting the output of a mindless word sequence generator into ‘their’ work.

That’s the craziest post I ever read.

how tf does one live through Opus 4.7 through 5 and only afterwards, when watermarks are being introduced suddenly get concerned about clarity. holy shit

Came here to say what everyone else is saying: it seems weird to have a take based on the purity of writing when that writing has been ceded to a machine under someone else’s control

If watermarking degrades the output, does that really differ from a junior writer as your assistant? Philosophically I guess yes, they have differing goals. But AI generated is AI generated and should be treated as such.

Truly unhinged. I thought this was a joke, but it’s too long

If you're letting the tool "do the writing for you" are you really writing at all?

Alas, I have to say the Belgians make the best fries.

Multiple ironies here with a writer outraged that an AI company is pledging to make text slop imperceptibly sloppier. 😜

"Sacrificing clarity, coherence, meaning, quality" etc. is in the sense described inherent to prompting an LLM, and not in any way added to by the watermarking process. You do not have more control as the end user over a RNG than a PRNG.

But my objection does not apply to all LLM writing. LLM output can be better or worse. I am more convinced today than a week ago that SynthID watermarking makes it worse. Maybe only slightly worse. But it has to be worse. That’s how it works.

Meant to thread this here; bsky.app/profile/elto...

Especially of you are paying for a premium upgrade.

@snell.zone In addition to the excellent points you made, it’s also chilling echo of the printer tracking dots used to put Reality Winner in prison. Now your “trusted assistant” is secretly ratting you out, encoding covert messages like some kind of hostile double agent.

That said, I do agree with you that this EU legislation is very ill-advised and unlikely to be a net positive for society.

I bet Claude, as bad as it is, understood the article better than almost everyone responding here. Jumping on a band wagon is more important than reading comprehension, it seems.

Claude itself is already a perversion of writing—my writing, that I’ve written and published over thirty years, that it stole from me—and if it is to be allowed to exist at all, its loser users should forever be forced to bear the mark of shame. What it makes isn’t “writing” at all, it’s plagiarism

I'm fine with AI slop users being smoked out. Just damn fine. ❤️

(That said, the Declaude type of workaround means it may not actually work.)

Claude is a perversion of writing.

Fair; that was the claim. Which led to an ecosystem of every con game ever invented. Because they all start with “trust me…”

i'm putting it out there: whatever. watermark that.

The obvious answer is you actually read the output and decide if the word choice is appropriate. I guess you aren't already doing that? I guess you only care now when that output comes with receipts.

But did you really never in a print publication compromise your sentence structure or word choice… in order to spell a hidden word from the first letters down the edge of the page? It's super fun… I even have some in my dissertation. 😃