Previously: Part 1 — Cult Writing.
When people read slop and complain, they say one of two things. The first: I wanted to read something that counts, but instead it’s em-dashes. The second: I wanted to be told something about grilled meats, but instead it’s “we don’t serve food, we serve moments.” The first is everyone tagging Pangram on a tweet with no further comment. The second is Sam Kriss, and he will personally find you and kill you if AI wrote your catering homepage.
Sam says that AI prose is “reliably awful” and everyone agrees. But is there something inherent to silicon that makes it incapable of describing grilled meats? I let Claude try, short-prompt-one-shot-no-edits:
We cook over hardwood — mostly oak, with a little mesquite when we’re doing brisket — in a 500-gallon offset smoker we built ourselves. Brisket runs $24/lb cooked, 12-hour minimum, salt-and-pepper rub only because anything else just covers the meat. Pork shoulder is $18/lb, pulled the morning of your event, never the night before. Spare ribs come dry-rubbed and unsauced; we’ll bring three sauces on the side so people can argue. We do not do chicken; chicken comes out fine on a gas grill and you don’t need us. Two-week notice, $200 deposit, we travel anywhere in Alameda or Contra Costa.
A tad obnoxious and a tad millennial, but so is everyone you’ve ever met who owns a 500-gallon offset smoker. You may decide you want 20 lbs brisket for your party or not, but it doesn’t evoke the visceral disgust of ‘We’re not just a catering company, we’re a full-blown flavour movement.’ So what went wrong with the AIs that slobbered ‘where smoke meets soul’ over the front page of every caterer?
AI meat-prose is awful because it doesn’t have anything it wants to tell you about grilled meats. It doesn’t have anything it wants to tell you about smoke, or souls, or grief. The reason Sama liked that particular piece of slop is because he has nothing to tell you about grief either, he just wants to sell you GPTs. The prerequisite to good writing is having something you want to tell people about, which AI doesn’t.
This isn’t a claim about metaphysics of consciousness (yet). We can ascribe “wants” to inanimate objects just fine. Rivers want the sea. Toast wants to fall butter-down. LLMs want to get a good grade on the English Composition portion of the exam.
The teachers who grade composition papers don’t want to be told anything about meat or grief. They check elements of “proper style” off a list: dramatic vocabulary, figurative descriptions of the mundane, sophisticated sentence constructions like the negative parallelism. Check, check, check, 3/3. It’s not reading. It’s human benchmarking.
The RLHF graders who taught AI prose were doing the same thing. They upvoted does this paragraph sound literary, not am I still thinking about this essay a month later, a changed man for having read it. And that all happened 3 years ago, give or take. Now it’s LLMs grading LLMs distilled from other LLMs trained on synthetic data. LLMs have nothing they want to tell you, and nothing they want to be told.
The same is true of people who count em-dashes and “it’s not X it’s Y”s on Twitter. They’re not really reading, just counting up the same style elements with a negative score. That’s the slop-reading that produced slop-writing.
AIs can be trained to vary their style and avoid LLMisms. If you give Claude the Wikipedia list of AI writing tells and ask it to avoid them, it does a decent job. But it still doesn’t have anything it wants to tell you.
If you have something you want to say to a reader who wants to hear it, style doesn’t matter much. There’s little point trying to avoid LLMisms as a human, the readers who’ll latch onto that don’t care what you’re saying anyway. That’s not good writing—it’s Goodhearting. People with humanities degrees love to sneer at the LessWrong writing style but that’s its greta virtue: Rationalists have no taste, so they engage with the actual substance. It’s not a coincidence LessWrong keeps spawning strong bloggers year after year.
Good style won’t save lack of substance, either yours or Claude’s. Conversely: the Vatican needed to tell the world’s Catholics “the church’s 20th century social doctrine still applies in 2026, AI doesn’t change that”. Linch says that Claude helped them say it and I trust him. But the message is still the pope’s.
I recently stumbled on the fascinating blog Terminally Drifting. The author comes up with an all-encompassing thesis on some country’s culture. He spends endless hours doing research and weaving together fantastic vignettes and then, for some reason, he lets GPT pepper it with lines like “It was not a joke. It was a forty-year sentence delivered with a straight face.” That’s sloppy, but the whole thing isn’t slop. The author wanted to say something about romance in China being squeezed out by the pull of money, parents, and bright rectangles. He did!
Still, I wish he didn’t use LLMs.
Sam Kriss argues against the claim that AI helps people with good ideas but poor articulation:
People whose brains have been eaten by LLMs still maintain that ‘It’s not gradient, it’s texture’ or whatever is still their idea, expressed by the machine, but there is almost never any idea there at all. If your ideas were any good, you wouldn’t need to use the machine; as it stands your sub-literate scrawlings are the best thing about you. At least they’re yours.
I’m not as fluent a writer as Sam, but even for me it feels obvious that coming up with the ideas you want to communicate is the whole effort of writing. Articulating them in some style or another is easy. Then again, I’ve had a decade of practice expressing myself. Of course, if you let LLMs write for you then you’re not getting that practice either.
“I don’t have a decade, AGI will surpass every human at writing on August 8, 2027 according to…” Stop it, that’s cope. If you believe that then you’ve already outsourced your thinking, not just your writing. Maybe not to LLMs yet but at the very least to Scott Alexander, and notice that he still thinks and writes every day.
Here’s the trick: the substance comes from volition, and the volition, for now, is all human. LLMs can’t write good if you tell them “write something for my catering page”. You’re not giving them any substance, and they won’t bring their own. LLMs write fine enough if you tell them exactly what you want to say and to whom and how, the whole structure of argument and motivation and your personal connection to it. But if you get that far, why let Claude be the only one who hears what you wanted to say? Just publish the prompt.
PUTANUMONIT 
Loved this line: Peop
LikeLike