// writing

2026-06-08 // 14 min // ai, llms, language, voice

no one in particular

I read something this morning that said nothing, and said it well. The sentences were clean, the shape was confident, and when I reached the end I could not have told you a single thing it claimed. I have read that piece a hundred times this month under a hundred names, and so have you. The writing keeps getting better and it keeps getting more alike, and those turn out to be one fact rather than two. I want to lay out where the sameness comes from, and why it is starting to happen to people and not only to the machines.

everything reads the same

The em-dash is not the tell, whatever the internet has decided. It is a punctuation mark skilled writers have always loved, and treating it as a confession only teaches good writers to fear their own dashes.¹ The neat triad and the turn that sets one idea up only to knock it down are ordinary moves of decent prose that an LLM overuses, and neither catches it on its own.

What marks the machine is the sameness itself. Read enough of the stuff and you stop noticing the individual habits and start seeing the thing under them. It all bends toward one register, the same handful of moves in the same order, whatever the subject or the supposed author.

As a way to catch a particular text, this fails, and it fails in a direction worth sitting with. When Liang’s group ran the popular detectors over essays by people who had learned English as a second language, the detectors flagged most of them as machine-written.² A smaller vocabulary and a flatter syntax produce the very signal the detectors were built to call artificial. Sameness describes what a text is like. It cannot tell you who wrote it, and used as though it could, it mostly libels people writing carefully in their second language. Keep it as a smell and never as a verdict.

The reason it is worth naming anyway is that the register has stopped staying inside the machine. Kobak’s group tracked word frequencies across more than fifteen million biomedical abstracts and caught a sharp jump after 2022 in a specific band of style words.³ It was large enough to estimate that at least 13.5% of 2024 abstracts had passed through a language model, and that figure is a floor rather than a tally. Careful, edited writing. The number that should bother you more is the spoken one. Yakura’s group went through hundreds of thousands of hours of unscripted talks and podcasts and found the same machine-preferred words climbing in how people actually speak, after the LLM that favors them shipped.⁴ A manuscript you can sand toward the machine on purpose. Nobody edits their own conversation in real time, so the speech is people drifting without choosing to.

no one in particular

Why does it converge? Because an LLM read an enormous amount of writing and walked no particular path through any of it. It never loved a few books and missed most of the rest the way a person does, never built one corner it knew cold while the rest stayed dark. With no place to stand, it writes toward the middle of everything it saw. Ask it for a sentence and you get the likeliest sentence, which is the average of the sentences it read.

Chiang found the image for this a couple of years ago. An LLM is a blurry JPEG of the web.⁵ Asked to reproduce an exact passage, it hands back an interpolation, its best estimate of what fills the gap from whatever sat on either side. Fluency, built out of averaging.

It is measurable and not only a figure of speech. Doshi and Hauser had people write short stories, some of them working from ideas an LLM suggested.⁶ The assisted stories scored higher one at a time, and came out more like each other across writers. The collective range of what got written quietly narrowed. Every writer a little better off, the whole field a little flatter. And the fluency has no vantage standing behind it. The stochastic-parrots paper named that years ago: a system stitching forms together by probability with no reference to meaning.⁷

So the line under the title. An omniscient writer is no one in particular. A view of everything from everywhere is a view from nowhere, and nowhere is the address the average lives at.

I want to be careful not to overstate the cause, because it is a default and not a law. Yun’s group showed that most of the flatness comes from the scaffolding, the chat templates and the instruction-tuning habits, and that asking plainly, stripped of the format, brings much of the range back.⁸ The pull toward the average is real, and it is the path of least resistance, but it is a setting you can lean against. Call it a strong tendency and stop there.

a voice is a path

If the average has no vantage, then the thing worth protecting is whatever the average destroys, and what it destroys is the particular. A voice is the residue of one route through the world. You know a few things well and almost nothing else at all. You happen to have read these books and kept these arguments while losing most of the rest, in a pattern that is yours. The exact shape of it is a vantage nobody else stands in.

Haraway said the thing I keep returning to. Against what she called the god trick, the fantasy of seeing everything from nowhere, she argued that the only way to a larger view is to be somewhere in particular.⁹ Knowledge has a location or it is not knowledge, and the same holds for a voice. It comes from a position, which means it comes from everything the position leaves out.

The part most people have backwards is the forgetting. Letting things go is how the vantage forms in the first place. A brain drops most of what reaches it on purpose, keeping what will help it act next and clearing the rest.¹⁰ That pruning is what lets it generalize at all instead of drowning in detail. Borges wrote the limit case. Funes, after a fall, remembers everything, every leaf he has ever seen and the shape of every cloud. He is for that exact reason unable to think, because, in Borges’ line, to think is to forget a difference, to generalize, to abstract.¹¹ What you have forgotten is the negative space that gives the rest of you its edge.

There is a quick test for a voice. Give the same brief to two real writers and two different essays come back, because it lands in two heads with two different stores behind them. The LLM, asked twice, returns one essay twice; a thousand people leaning on it return a thousand near-copies of that one. Divergence under a shared prompt is the mark of a vantage, and the convergence everyone keeps clocking in machine prose is the mark of none.

This is also why a voice is not a coat of paint. You cannot bolt one on by asking for a few quirks, a folksy aside say, because they come back as more average with tics added. A voice is what is left when a particular mind works a particular problem with the store only it has. Prompt the LLM to sound human and you get the average in a costume, the slightly uncanny thing you have felt and maybe not named: fluent, and no one home.

Which lands the contrast the whole piece is built on. The LLM has read everything, forgotten nothing, and been nowhere, and that is precisely why it has no voice to speak in. A person is the other arrangement, a small and partial record of a single life, and the smallness is exactly the source of the vantage.

you can’t see it go

Machine writing being flat is the smaller problem. The larger one is that the flatness spreads, and you cannot feel it spreading in yourself.

The part of you that would notice your own register sanding down is the same part that goes quiet once you stop using it. I wrote about that in the reps you stop taking. Hand the work over often enough and the judgment that could tell a real voice from an average one is the first thing to thin. Use was the only thing keeping it alive. So you slide toward the mean, and from the inside the slide feels like nothing at all, which is what makes it worth naming before it reaches you.

I will not hang this on a number. The most-cited study claiming a measurable cognitive cost from leaning on a language model has been challenged hard on its methods, and the point does not need it. A capacity you stop exercising is one you stop having, judgment included, and nothing sends up a flare while it happens.

keep what can’t be averaged

Start with the concession, because it is real and it is large. The LLM is genuinely good at the average-shaped work, and it is fast. Noy and Zhang ran the experiment: people handed a language model finished their writing tasks quicker and scored higher for it, and the weakest writers gained the most.¹² When the job is to produce what a competent person would from a clear brief, the standard format or the section you have written a hundred times, hand it over. You lose nothing that matters.

The line falls where the vantage does. Give the LLM the work with no particular place in it. Keep the work only your position can do: the first real pass at a problem you are the one standing in front of, the sentence the argument actually rests on, the read that turns on something only you happen to know. That is the same call I keep landing on, in the brain drives the tool and a tool for thinking, reached this time from the side of the language.

Here is the thing no average reaches. Graciliano Ramos was a novelist of the Brazilian Northeast, of its droughts and its poor, and he had done time in one of Vargas’s prisons, so he brought a particular life to whatever he looked at. In 1946 he wrote to the painter Candido Portinari about a canvas of a mother holding her dead child.¹³ Partway through, the letter turns into a question he plainly did not set out to ask: whether art could survive in a world that had abolished suffering, and whether he would even want that world if its price were the art. He does not resolve it cleanly. He lands somewhere uncomfortable: he would not wish the misery away, even though wishing it away is the only decent thing to want, because without it there is nothing for the art to be made from. It is not a tidy thought, or a kind one. And it is unmistakably his, the kind of thing only that man could write, with that politics and that grief, from the one place he stood. Nothing trained on the average of every letter ever written arrives there. No amount of scale closes the distance, because what produced the letter was a life and not a distribution, and there is nothing in the middle of everything that was ever going to land on it.

someone in particular

This is the other half of a compression of a compression. There I argued that the structure of a thought collapses inside the writer before a word of it is recorded, so an LLM only ever learns the flattened trace and never the thing that cast it. Here is what survives the collapse that the average still cannot reach: the place a particular life sees from.

The LLM has read everything and been no one. You have read almost nothing next to it, and you have been someone, somewhere, the entire time. That difference is the whole game, and it is the one thing an average can never be made to produce.