In Defense of Writing With AI

Every week somebody announces that writing with AI is a failure of thinking. The writing was the thinking, the argument goes, so you’ve outsourced your own mind. What I notice is how little thinking the argument itself contains.

“Writing with AI” covers everything from pasting a two-sentence prompt and publishing whatever falls out, unread, to sitting with the machine for an hour and fighting over every paragraph: is the claim true, is it the claim I even believe, is there a plainer word. If answering those questions isn’t thinking, I don’t know what to call it. Treating the two as one activity is the actual laziness, and the people most certain they’re the thinkers in the room can’t be bothered to notice the difference.

How the raw words first hit the page is the least interesting fact about a piece of writing, and nobody can tell from the outside anyway. What matters is the ideas, and whether they’re any good.

Art school beat this into me. You generate ideas by the hundred and most of them are bad. You keep a folder on your desktop stuffed with other people’s work, saved one right-click at a time, and dig through it for the thing that sparks yours; we call that inspiration, and nobody calls the painting lesser for it. Half of what you make looks like everyone else’s regardless. So the real education isn’t in generating at all. It’s recognition, learning to spot the one live thing in a pile of dead ones, and that’s most of what taste is. Volume never guaranteed it. Beethoven wrote the Ninth after his hearing was gone. There are 11-year-olds writing music most adults never will. Typing every word yourself proves one thing: that you typed every word yourself.

Can you use the tool badly? Sure. You’ll build a bad house with a good hammer if you’ve never built a house, and nobody blames the hammer. Used well, it’s the best teacher anyone has ever had, the whole record of what people have written sitting right there, willing to argue with you for as long as you can stand it.

What actually sets people off isn’t quality, it’s contamination. A machine touched the words, so the words are dirty. I watched this happen to a writer whose piece began life as a voice memo. He transcribed it, cleaned it up by hand, ran it through a model to tighten the structure, posted it. A reader told him he couldn’t get past the opening because it read like AI. So he posted the raw memo instead, and the same reader liked that one better. Maybe it was. It was looser. A model will take something strange and alive and sand it down until it sounds like everything else, and if that’s what happened here, the writer trusted a bad edit. That says nothing about whether his thinking was real.

A smell can tell you the writing is bad, not where the ideas came from. And the smell is real. Something happens to the cadence when a model has been at the words, some rhythm underneath them, and I was trying to describe it to my co-founder the other day and got nowhere. We can both spot it instantly. Neither of us could say what it is. The closest I’ve got is bad CGI in a movie that meant to be realistic. You fall out of the thing it was asking you to believe. But the seam showing is a different complaint from a computer touching the frame, and nobody walks out over the CGI they didn’t catch.

And most writing is bad, was bad long before the machines showed up. The slop was always ours. People sneering at AI prose are comparing it to some imagined version of what humans write on their own, and that version is a fantasy. Ask the only question that counts: did the machine make your writing better or worse? Mostly it raises the floor and leaves the ceiling alone. If you live at the ceiling it has little to give you, and almost nobody does. Refusing it on principle usually just means insisting on your own worse draft.

None of this is new, either. Ghostwriters. Pen names. Editors who rebuilt half the book and got one warm sentence in the acknowledgments. Writing was never a pure mind alone in a room handing down finished sentences. People have always climbed in, moved the furniture, told you when you were full of shit. The name at the top never meant I made every sentence alone. It meant I stand behind this, and that survives AI intact. A developer whose AI-written code takes down production doesn’t get to point at the model. Neither does a writer. Pick the argument, check the claims, sign it, answer for it. The machine takes none of that away and, if anything, hands you more.

I get it with photos and video. When you watch footage of a thing, your default is that it happened, and the default was safe because a convincing fake cost real hours and real skill. Photoshop spent thirty years eroding the arrangement and AI vaporized what was left overnight. The fake costs nothing now and the default hasn’t caught up. Anyone can put a speech a politician never gave into their mouth by lunchtime, so the fear there is earned. Some of the flinch at AI writing is that fear splashing over, I think, from a tool that got famous for its worst uses. But text never had a default worth breaking. A lie has always been free to write down. Nobody ever believed a paragraph because of how it was produced. We believed the name on it, the outlet that ran it, the sources under it, whether it held when we pulled on it. That system is centuries old and it doesn’t care about AI in the slightest.

Which is why the labels confuse me. There are honest cases for disclosure. A school needs to know the student can write. A publication promised its readers no AI, and a promise is a promise. But those are agreements about process, binding because somebody made them. A label on the words tells you neither of the things a reader actually needs to know: whether it’s true, and whether it’s any good.

This essay began as a reply to one of those weekly announcements, a piece arguing that writing with AI is a failure of thinking. I handed it to a model, said I disagreed, and asked it to interview me about why. It did. And from those answers it wrote our first draft. And I went through what it wrote and spent hours over the course of a day multitasking on other things but also discussing and debating word choice and framing with multiple models, until it was something I’d put my name on, and that’s the version you’re reading. I never typed a word of it. You now know more about how this piece was made than about anything else you’ll read this week. While I was working on it, Substack shipped a classifier that scores posts for AI, and this essay gets a perfect 100. The argument, the writing, and whether to trust the name at the top don’t come with a meter.

Somewhere back up the page I said the only question that counts is whether the machine made your writing better or worse. For me it did neither, it made the writing exist. There was never a version of this essay I typed alone, no purer draft in a drawer. There was this or there was nothing. Whether this beats nothing is your call, it always was, and you make it the old way: on the words, with my name on them.