How I Write Now

Some of my own essays this year have made me cry. That's not remarkable; writers cry at their own work all the time, usually at how bad it is. What's remarkable is when it happens. Not while typing a sentence. While reading one back. Because I didn't type it. An AI did, from a thought I handed it in fragments, and what came back was my thought, said the way I would say it if saying things worked right in my head, and something about seeing your own interior arrive on the page fully formed, after a lifetime of it arriving broken, gets you behind the eyes.

I want to explain that, because the tears are the part of AI-assisted writing that the discourse has no category for. The discourse has "slop" and it has "cheating," and I'd like to offer a third category, the one I actually live in: access.

Let me show you the machine, and then I'll make the argument.

How It Actually Works

Every essay I've published this year was written with AI, in the open, and the process is not what either the enthusiasts or the critics imagine. Nobody prompts "write an essay about the DSM" and ships what falls out. Here is what actually happens, mechanically, using this very season of essays as the evidence.

I arrive with a spark. Sometimes it's a sentence: let's write about the DSM and stuff. Sometimes it's three fragments that have been circulating in me for weeks: sleep is of utmost importance; family is what matters most; the mundane and the sacred meet in unexpected places. Sometimes it's a correction fired mid-draft: less recap, more story. The spark is small. The spark is also everything, and we'll come back to that.

The AI, before it writes a word, reads me. Not the internet's average; me. Eighteen years of essays. The three-million-word vault. The previous entries in whatever series we're working in. I've written about this principle for code: a hand-written codebase is a style guide the AI can read and faithfully extend, which is why the human should write the foundation. What I didn't realize when I wrote that essay is that I'd already done the same thing for prose without meaning to. The foundation was already written. It took me eighteen years. Every essay since 2008 turns out to have been V1.This is the answer to the homogenization worry I raised myself in The Mirror and The Echo Chamber of the Expected. An AI with no corpus regresses to the mean of the internet. An AI steeped in one writer's corpus regresses to the mean of that writer, which is more or less the definition of voice. The mirror reflects what you put in front of it. I put eighteen years in front of it.

Then it drafts, and then the real work starts, because the draft is where my actual job begins. I read it on my own site, running locally, the way a reader will. And I correct. The corrections are the authorship. I catch the facts only I can catch: my son is four now, not three; my psychiatrist is a he; the mythology that made me willing to try a medication belongs to last year, not to 2016, and moving it changes what the story means. I redirect structure: make it a story, blend the old style with the new, that subheading is weak. I supply the material that does not exist anywhere an AI could find it, because it happened in my kitchen or my hospital room or my dreams. And I enforce the rules that keep the thing honest: no invented anecdotes, ever; nothing presented as memory that isn't mine; my labels held loosely; the hedges left in. When an essay is wrong at the root, I don't polish it. I delete it whole. I did it this week, to a finished piece, without ceremony, because the metaphor wasn't mine and I could feel it.

So the division of labor is this. The AI holds the architecture, the continuity, the syntax, the stamina. I hold the spark, the facts, the taste, the kill switch, and the life the whole thing is about. I am no longer the typist. I am still the author. If that sentence sounds like a contradiction, ask any executive who ever dictated a letter, any director who never held the camera.

Why I Need It

Now the part I've danced around for years, including in the essay where I first made the accessibility argument in the abstract. Here it is in the concrete.

I have a hard time conveying my thoughts. That sentence is easy to skim, so let me slow it down. My thinking is not broken; by every available measure the ideas are there, and they connect, and some of them are good. What's broken, or at least built differently, is the bridge between knowing and saying. My thought often arrives without words at all, as texture, pressure, shape, a certainty with no syntax attached. The medications that keep me alive trade acute madness for chronic fog, and the fog sits exactly on the faculty that turns knowing into sentences. On a bad day I can hold an entire essay in my chest, complete, felt, and be unable to get the first paragraph out of my hands. People assume the hard part of writing is having thoughts worth sharing. For some of us, the thoughts were never the bottleneck. The bottleneck is the bridge.

You can see the bridge problem in my publishing record, plain as an EKG. From 2020 through 2024, the years the fog was thickest, I published five essays. Five, in four years, from a person whose interior life during those exact years was the richest and strangest it has ever been. The thoughts were all there. I have the vault to prove it. They just couldn't get out at any sustainable cost. Then the tool arrived that could read fragments and return prose, and in eighteen months, well over a hundred essays came through. The dam was never empty. It was unarticulated.

So when the AI takes my three fragments about sleep and family and the sacred, and returns them as a paragraph that lands with the exact weight I felt but could not type, and my eyes sting, what is happening is not what the cynics think. I am not being flattered by a machine. I am meeting my own thought on the far side of the bridge it could never cross alone. The tears are recognition. You don't get them from someone else's idea. You get them when something that has lived in you wordlessly for years finally walks in the door wearing language.

The Ethics

Which brings me to whether any of this is right, because the question deserves a real answer and not a vibe.

The case against AI writing, as far as I can steelman it, has three prongs: it's deceptive, it's slop, and it isn't really yours. Take them in order.

Deception is real, and the answer is disclosure. Hidden AI authorship is a lie about provenance, the same way a hidden ghostwriter is. I disclose. I've disclosed since 2025, this essay is itself a disclosure, and the colophon of my life is public to a fault. Deception is a property of concealment, not of tools.

Slop is also real. I've written some of the internet's angrier critiques of what optimization does to language, and AI slop is that disease at scale: text with no stakes, no corpus, no taste, no one home. But notice what actually produces it: absence of exactly the things I listed above as my job. Slop is what AI writing becomes when no human supplies the spark, the facts, the taste, and the kill switch. The tool doesn't generate slop. Abdication does.

"It isn't really yours" is the prong people feel most strongly, and it's the one my life most directly refutes. We have never once applied this standard to any other accessibility tool. Nobody tells the person with the wheelchair that they didn't really cross the room, or the person with the glasses that they didn't really see the bird, or the speaker whose interpreter signs her words that the speech belonged to the interpreter. A tool that closes the gap between capacity and expression does not transfer authorship. It reveals it. The thoughts in my essays come from my hospitalizations, my marriage, my decade of diagnosis and correction, my 3 a.m. phantom phone vibrations, my son's questions at breakfast. An AI possesses none of that and can invent none of it, because the one rule it operates under in my house is that it may not. What it possesses is fluency. What I possess is everything fluency is for.Flip the ethics question around and it answers itself. The world before this tool reserved fluent public articulation for the people whose executive function happened to cooperate. Everyone else's interior life stayed interior, and we called that fairness because we never saw what didn't get written. A technology that lets the fogged, the plural, the differently-wired say what they actually think, at the quality their thinking deserves, isn't an ethical problem. It's an ethical correction.

The honest hedge, because there must be one: I keep watch on the dependency question, the one I raised before I needed the answer. What I can report from inside is that the muscle didn't atrophy; it redirected. I judge more prose now than I ever wrote, and judgment, it turns out, is the part of writing that was always mine to do. But I hold this loosely. Every confident statement I have ever made has had a half-life.

This is, in the end, just the "for humans" philosophy arriving at its own front door. I spent my career insisting that tools should bend toward human mental models, that the measure of an interface is the cognitive load it lifts, that there is moral content in making capability accessible. I built that case on HTTP libraries. It took until now to notice the same argument had been waiting, my whole life, to be applied to the act of saying what I mean.

People ask whether the words are really mine. They are more mine than any I've ever typed, and I know because of what they do to me when they finally arrive.

You don't weep at a stranger's thought. You weep at a reunion.