It’s always a little dangerous to write about any rapidly-developing technology, because chances are pretty good that whatever you say will be incredibly and obviously dated within a few months. But I’m going to plant my flag anyway, because even if nothing else changes — even if there’s no meaningful advancement in LLM performance beyond the state-of-the-art right now, in March 2025 — the potential disruption is already so enormous that you can think of it as a kind of Industrial Revolution for text.

Just like in the first one, we’ve figured out how to use machines to do a broad swathe of things people used to do, swapping energy and capital in for human labor. And just like in the first one, the output isn’t necessarily better (in fact, it’s often worse), but it’s so much cheaper in terms of human time and thought and effort that the quality almost doesn’t matter. Sometimes that’s wonderful: if you desperately need to put a roof for your barn right this moment, it’s a blessing to be able to slap on some corrugated tin instead of going to the effort of thatching. When you have to write your seventeenth letter to the insurance company explaining that no, they really ought to be covering this, it’s a relief to hand the composition off to Claude instead. But do that too much and you forget how to do it yourself — or more plausibly, you never learn.

The greatest risk of AI is probably “we all get turned into paperclips”, or maybe “someone uses it to design a novel and incredibly fatal pathogen”, but the most certain risk — the one that’s already here, at least on the edges — is a great deskilling. Just as the mechanization of physical labor lost us all those traditional skills that Langlands describes, the ability to automate cognitive tasks undermines their acquisition in the first place. Why pay any attention at all to word choice and metaphor and prosody when ChatGPT can churn out that essay in a few seconds? Why worry about drafting a convincing email when you’re pretty sure your recipient is just going to ask Grok for a summary?¹ Why learn to code when a machine can do it faster?

I was recently informed that someone — “not anyone you know, Mom, someone at another school” — used ChatGPT to write his essay about the causes of the Civil War. This was obviously deeply upsetting to the congenital rule-follower who reported it to me, on account of THAT’S CHEATING (you must imagine this in the whiniest she-touched-my-stuff voice possible), but it was a good teachable moment — for me, if not for the history teacher at another school. What’s the point of an essay about the causes of the Civil War, anyway? It can’t be that the teacher wants to know the answer: she can find a dozen books on the topic if she cares to look, each more cogent and thorough than anything a middle-schooler is likely to produce.² Heck, even the Wikipedia article will probably give her a better understanding. And if it’s not for the teacher’s benefit, it’s certainly not for the benefit of any other audience, since as soon as the essay is marked and graded it’ll probably be crumpled up and tossed into the recycling bin. No, it’s for the kid.

The point of writing an essay about the causes of the Civil War is not to have an essay about the causes of the Civil War, it’s to undergo the internal changes effected by the process of thinking through, planning, drafting, and editing the darn thing. Writing forces you to put your thoughts in order, to shape whatever mass of inchoate ideas is bouncing around in your head into something clear and reasoned you can pin to the page. The thinking is the hard part; putting words to it is simple by comparison. (This book review began life as about seven hundred words of stream-of-consciousness riffing, with only the vaguest kind of structure. When I experimentally pasted it into an LLM and asked for an essay, the result was terrible.) But even the putting of words is a valuable skill: what’s the right tone here? What’s the right word? Do I want to say “writing forces you to” or “when you write you have to”? How do they feel different? Asking a machine to do this for you is like bringing a forklift to the gym.

Of course, that kid who had ChatGPT write his essay was almost certainly thinking of the assignment not as one small step in the alchemical process of self-transformation that is education but as basically equivalent to an appeal letter to the insurance company: just another dumb hoop you have to jump through in your interactions with a vast impersonal machine that doesn’t particularly want to grind you to dust but wouldn’t mind it either. And since this was at another school, he might not even be wrong. Maybe the teacher was just pasting the rubric and the essays back into ChatGPT and asking it to assign a grade.³

But there’s an even bigger problem than lying about who (or what) has done the work, which is lying about whether the work has been done at all. LLMs make lying very easy indeed. Yes, yes, sometimes they hallucinate and tell you things that are patently untrue, and that’s a bigger danger for students and other people who don’t have the background to notice when something seems off — this is all true, but it’s not what I mean.

LLMs, when working exactly as intended, enable human falsehood — because our society relies on written records as proof of work. Until recently that was fine, because writing down lies actually used to be pretty hard: putting together a convincing false report from scratch — maintenance records for the airplane you’re about to board, say, or a radiologist’s report on your brain scan — was almost as time-consuming as actually checking the things that were supposed to checked and then documenting them, and the liar had to spend the whole time aware of their own dishonesty. (Not that this stops everyone, of course.) But now that it takes about two clicks to generate an inspector’s report for the house you’re considering buying, or the pathologist’s findings in your biopsy, how much are you going to trust that they actually looked?

LLMs can be useful tools,⁴ but all tools change what we make and how we make it. It’s often a good tradeoff! Sure, each individual example of simplification and automation in the name of efficiency is a tiny bit of alienation, removing the maker from the making, but it’s also a gift of time we can spend on other things: I couldn’t write this if I also had to sew my family’s clothes and wash our laundry by hand. And yet those bits pile up, and once it becomes possible to exist in the world without really needing to come into contact with it, once you can get by without ever really needing to make anything, some people just won’t. And that’s terrible! Being entirely without cræft — never bringing mind-body-soul into harmony with one another and then using them to master the world — means missing out on something deeply human.

Jane Psmith, “REVIEW: Cræft, by Alexander Langlands”, Mr. and Mrs. Psmith’s Bookshelf, 2025-03-24.

---

1. All the “AI written/AI read” communication begins to resemble Slavoj Zizek’s perfect date:

“So my idea of a perfect date is the following one. We met. Then I put, she puts her plastic penis dildo into my … “stimulating training unit” is the name of this product. Into my plastic vagina. We plug them in and the machines are doing it for us. They’re buzzing in the background and I’m free to do whatever I want and she. We have a nice talk; we have tea; we talk about movies. What can be — we paid our superego full tribute. Machines are doing — now where would have been here a true romance. Let’s say I talk with a lady, with the lady because we really like each other. And, you know, when I’m pouring her tea or she to me quite by chance our hands touch. We go on touching. Maybe we even end up in bed. But it’s not the usual oppressive sex where you worry about performance. No, all that is taken care of by the stupid machines. That would be ideal sex for me today.”

2. Well, okay, most of them.
3. See footnote one again.
4. Personally I’ve found them useful in three cases: (1) when I’m blanking on how to begin an email I will occasionally ask for a draft, which inevitably makes me so mad about how bad it is that I immediately rewrite it in a way that doesn’t suck; (2) when it’s Sunday night and I need a picture of a Japanese man in a business suit and a samurai helmet for a book review going up in the morning; and (3) when I can’t figure out the right search term for my question. (Turns out it was “sigmatic aorist”. Thanks, Claude.)