Why AI Writing Sucks
It's not the butt. It's the ass.
Here’s a claim on AI writing I found online: “Unfortunately, I think that in the near future, not using LLMs to write for you will be like someone refusing to use Google Maps for directions in a new city. A bizarre idiosyncratic choice that’s just completely incomprehensible to the vast majority of people.” It’s probably more accurate than not, although it depends on the cost of inference going forward. If AI companies continue to raise prices and restrict access, we’ll see how readily marketing departments use it for email copy. AI has reminded me of how little people care about writing in general; as Joe notes later in the thread, it is viewed as an efficiency problem. AI meets a B+ writing standard, so it’s good enough for a X Dot Com The Everything App post about fitness hokum. It’s not only efficient to produce, but to juice engagement, too, because it’s the prose equivalent of someone talking over a Subway Surfers clip. The “it’s not X, it’s Y” construction is the prime example. It’s microscopic tension and release, over and over. No patience required. You’re drawn along by a trail of candy.
Sometimes writing is an efficiency problem, it’s true. Sometimes you need to crap out an email for your boss to skim. I might text my wife “ok i l meet you ther” and that’s fine. I had Claude compose a postmortem on a technical issue at work because the issue and the report are boring as fuck. I don’t want to spend more than 10% of the time to report as I did to investigate1. It can’t all be pearls! I have more to say about AI technical writing, but for now I’ll say Claude gets the job done. That is, it can write a reliable record of a technical issue, its diagnosis, and its solution for the one or zero people who will need to know about it in the future.
AI writing sucks, though. I’m sure I don’t need to enumerate its stylistic tics, they’ve been covered. I think they’ll probably get ironed out as long as people keep complaining about and mocking them. Remember when GPT wouldn’t shut up about goblins2? However, I think the worst part is that it produces nonsense. It’s vague, imprecise, and lazy. It’s lazy in the sense that it gestures toward a point without making it concrete. Here’s a weird and distracting example:
https://x.com/HunterBiden/status/2069797401078939851
This was about the results of the recent elections in New York. Look, I don’t think Hunter Biden was spinning gold before. I don’t think anyone who posts AI prose on X DOT COM was writing anything good otherwise. I do not insist Hunter Biden labor over his political analysis. I don’t even disagree, but the first point upset me so much that I first complained to friends, and now I’m writing a blog post about it. But LOOK.
“Authenticity is measurable. Voters can smell a focus group from a mile away.” Measurable by whom? Was the election outcome the measurement? Was it measured by exit polling? What does that have to do with focus groups? I understand what he means, which is simply that voters like authenticity and they responded to authentic candidates. What it is, literally, is fucking nonsense. I got the point, so it’s not useless, but its constituent elements do not add up to its intended meaning. AI writes in pre-cliche. It comes out of the extruder already tired and deflated. It’s worn out to the ear, despite its unfamiliarity. I guess it’s hard to say what you mean, especially when “you” don’t “mean” anything.
For contrast, here is an excerpt from one of my favorite essays, “E. Unibus Pluram” by David Foster Wallace, originally published in 1993 by The Review of Contemporary Fiction.
It’s widely recognized that television, with its horn-rimmed battery of statisticians and pollsters, is awfully good at discerning patterns in the flux of popular ideologies, absorbing them, processing them, and then re-presenting them as persuasions to watch and to buy. Commercials targeted at the eighties’ upscale boomers, for example, are notorious for using processed versions of tunes from the rock culture of the sixties and seventies both to elicit the yearning that accompanies nostalgia and to yoke purchases of products with what for yuppies is a lost era of genuine conviction. Ford sport vans are advertised with “This is the dawning of the age of the Aerostar”; Ford recently litigates with Bette Midler over the theft of her old vocals on “Do You Wanna Dance”; claymation raisins dance to “Heard It Through the Grapevine”; etc. If the commercial reuse of songs and the ideals they used to symbolize seems distasteful, it’s not like pop musicians are paragons of noncommercialism themselves, and anyway nobody ever said selling was pretty. The effects of any instance of TV absorbing and pablumizing cultural tokens seem innocuous. But the recycling of whole cultural trends, and the ideologies that inform them, are a different story.
I like that it breaks rules. “And anyway” is a nice colloquial flourish that helps reinforce its sentence as a folksy, practical rebuff to the preceding argument. The last sentence employs the “it’s not X, it’s Y” form but in a way that isn’t fucking stupid. It also happens to describe the digestive and reproductive process of AI training. It is exactly the automation of the “pablumization” of culture, an amoeba that turns centuries of thought, as well as air purifier recommendations on Reddit, into something clean, sterile, professional, work-appropriate, if incoherent in the specifics.
If you don’t care enough about your idea to develop it, why should I read it?