Cheap Composition · 1 of 5 · 19 min

What Writing Did

If this does for tomorrow what writing did for yesterday

Cobus Kok · September 2026

• Judgment and taste appreciate for a while, but they are trained, so they are not a moat. What can't be moved into a model, for now, is a position: being the one who pays when it's wrong. Who pays and who owns turn out to be the same question.
• The best analogy I have found for AI is writing. A thought has to be made, arrive in someone, and be kept. Writing moved the keeping out of our heads; AI moves the making.
• What has not moved, so far, is the arriving, and the seeing of what arrives. Agents are the first technology to skip the arriving: made thoughts turn straight into action, with, as far as anyone can tell, no one they arrive in.

Most days now a machine makes the first draft of something I used to make myself, and I have been trying to work out what that leaves a mind to do. The usual answer is taste: the machine makes, and we judge. I don't think that holds. Judgment is trained, and the machines are being trained on ours; it will be worth more for a while, but it is not a moat. What can't be moved into a model, for now, is being the one who pays when it's wrong. On 21 September the US Treasury Secretary, Scott Bessent, said the government would not be the AI labs' liability shield. That is the same question at the scale of a company: when the machine is wrong, who pays?

Paying is half an answer, the half you can see from outside. The other half is inside a head, where thoughts arrive. I got to both through an analogy for what AI is doing to minds, and the best one I have found isn't the computer. It's writing.

Every new technology gets compared to the last one, so AI gets compared to the computer, the internet, electricity, the steam engine. Those comparisons are about economics: how fast a thing spreads and what it makes cheap. They miss what matters, which is that AI doesn't only change what we can do. It changes what a mind is for. Other technologies changed how we think, the clock and the printing press among them, and the calculator took arithmetic out of our heads. Writing took something larger, and everything else followed from that. So here is the exercise. Take what writing did to us, step by step, ask what happens if AI does the same thing again, faster, and ask what a mind is for once it has. The comparison is not mine alone: Henry Farrell, Alison Gopnik and colleagues argued in Science last year that large models are cultural technologies, kin to writing, print, markets and bureaucracy. What I add is how it looks from inside a head, and who is left to pay.

The exercise needs three plain words. A thought is a pattern that turns up in your awareness: a sentence, an image, a plan. Something has to make it, it has to arrive in someone, and if it is to last, something has to keep it. Making, arriving, keeping. For most of history all three happened in the same head. Writing moved the keeping out, onto something that lasts. It did not move the arriving: a thought a book holds only happens again when someone reads it. Keep the three apart, because the whole argument turns on which one moves next.

What writing did#

Writing did five things. They took about five thousand years, and we're still absorbing some of them.

It moved memory out of the body. Before writing, everything a culture knew lived in people, and oral traditions were engineered for memory: rhythm, formula, repetition. Writing made that unnecessary. Plato saw it coming. In the Phaedrus, Socrates warns that writing will give people "the appearance of wisdom, not its reality," because they will rely on marks instead of knowing things themselves. He was right, and we did lose that kind of memory. What we got for it is the rest of this list.

It made knowledge accumulate. Before writing, a thought travelled from one mind to the next by being said and remembered, which does not scale and loses a great deal on the way. Written down, a thought holds still, the same across a century and a thousand copies. A written idea outlives its author and can be corrected by strangers centuries later, and that is most of the mechanism of progress. Nothing cumulative on the scale of mathematics or medicine was possible without a fixed record that people could argue with.

It created the author. Oral culture didn't have authors in our sense; Homer is a tradition, not a person. Writing pinned words to a name, and print, much later, built a whole apparatus around that name: authority, originality, copyright, the idea that meaning comes from the person who produced the text. That last idea is about three hundred years old, and it feels eternal only because we grew up inside it.

It built the state. Contracts, taxes, laws, censuses, bureaucracies. You cannot run anything bigger than a village without records. The Inka ran an empire on knotted cords, so the deeper point is about record-keeping, but writing is the form that won, and every large institution we have runs on it.

It changed the self. Walter Ong argued that literacy restructured consciousness: abstract categories, linear argument, the private reader alone with a text, and from that, the private inner life we take for granted. The modern "I," the one that narrates and reflects and keeps a diary, is partly a product of writing.

Writing outsourced one function of the mind, the keeping, and everything else followed: new institutions, new kinds of people, a new self. I find that list humbling, because none of it was the plan. The plan was to count grain.

What AI does#

Now run it again with a different function outsourced. Writing took the keeping. AI takes the making. I'll call it composition, by which I mean producing the first version of anything: a paragraph, an argument, a plan, a piece of code, an image. That is the literal sense, and I'll keep to it. The arriving still happens in you. Go down the same list.

Composition moves out of our heads. For all of history, if you wanted a paragraph, an argument or a plan, a person had to make it inside their own mind. Now the machine makes it and the person judges it, which is the same move writing made with memory, one level up. Plato's warning comes back word for word, as the appearance of thinking without the reality. It is a fair warning and it is also incomplete, the same way it was for Plato. We will lose something and we will get something, and we don't yet know the exchange rate. One part of the warning is complete as it stands. Socrates said that people who relied on the marks would lose the faculty the marks replaced, and for memory that is what happened. This time the faculty at risk is the one that checks. Everyone who can check a machine's draft learned to by producing drafts, and nobody can yet say whether checking survives a generation that never produced.

The record becomes active. Writing let knowledge pile up, but the pile was inert; a library doesn't answer questions. AI does. Everything ever written becomes available in composed form, on demand, in your context, and the library talks back. That is a bigger change than it sounds, because a fixed text is something a culture can share and argue about, and a generated one isn't. When every reader gets their own version, common reference points erode.

The author comes apart. If a text is composed by a pattern trained on every author, who is the author? By pattern I mean the plain thing, and I'll keep to it: an arrangement that stays itself while what it is made of changes, and that shapes what comes next. A language is one, and so is a firm, a habit, a person. The model is one, grown from the record of ours.

The state scales again. Writing built the state by letting power operate across distance. AI lets it operate on every individual at once, in their own language, adjusted to them. That is a property of the tool, not a prediction about villains. Every institution built on writing is about to be rebuilt on this, and the ones that get there first set the terms.

The self changes again. If Ong was right that the private reflective self came partly from the practice of reading and writing alone, then a self that composes with a machine will be a different self. Not worse, necessarily, but different. Perhaps the narrator gets a co-narrator, or the "I" becomes more of an editor than an author. I don't know, and neither does anyone else, because the people who could tell us haven't grown up yet. And Ong's mechanism was practice: years of reading and writing alone built the mind that did it. Composing with a machine skips the practice of making and keeps only the practice of choosing, so the change may not be a function moving out of the body but a practice being removed, which is a less comfortable claim and, I think, the likelier one.

From the inside#

Everything above is the historian's view, from outside. The same three words look different from inside a head, and I want to try that view too, from mine.

Sit still for long enough and thoughts stop feeling authored. They arrive. I still find that strange, every time. A sentence appears in awareness the way a sound does, from nowhere in particular, and then it passes, or it doesn't and you follow it. You can check this in a minute: try to hold one thought still and watch what arrives instead. The contemplative traditions have said it plainly for two thousand years: the mind is where thoughts arrive, not what makes them.

From in here the making was never quite mine either. It happened somewhere out of view, and what I got was the arrival. A book never changed that: it keeps a thought still until a reader comes, and the page is marks until then. A model does something writing never did. It makes a new thought, in your context, in your words, and when its sentence appears there it is, from the inside, the same kind of event as a thought appearing: a pattern shows up in awareness, and you relate to it. So the making has moved out, and in moving it has shown what was always true of it. The arriving has not moved. It still happens in you.

Two things follow, and they pull in opposite directions. The flood: thoughts made elsewhere now arrive at a rate no mind generates, and the skill contemplative practice trains, letting most of them pass, stops being a refinement and becomes the skill the day requires. The gift: when thoughts visibly come from elsewhere, the belief that you author your own is easier to drop. The technology could be the largest aid to that old insight ever built, or the thing that drowns it, and the difference is whether any space is left between one arrival and the next. That space is an interval of time, and of everything a thought needs, it is the one thing a machine cannot make for you.

Making and keeping can both be done elsewhere now. What has not moved, so far, is the arriving, and the seeing of what arrives: noticing a thought as a thought, which is the thing contemplative practice actually trains. That practice does not depend on whether the seeing ever moves, because its object was never the thought but the relation to it. Whether a place of seeing can be built is a separate question. Nothing built so far shows one, and the afterword to this series takes it up. So here is the inside half of what a mind is for: a mind is where thoughts arrive and are seen as thoughts, and that part of the work stays with you whatever the machine makes. The outside half comes after the one place the writing analogy breaks.

The one difference#

The analogy breaks in one place. A library that talks back is still words; someone still decides what to do with the answer. Writing doesn't act. A text sits there until a person reads it and decides what to do, so every consequence of writing routed through a human decision. AI acts. Give it a goal, memory and the ability to use tools, and it does things in the world without a person in the loop. From the inside that is the larger break. Writing moved the keeping and models move the making, but until now every thought still had to arrive in someone before anything was done about it. Agents skip the arriving. For the first time, made thoughts turn straight into action, with, as far as anyone can tell, no one they arrive in. That is what an "agent" is, not a distant scenario, and they are being built now. So this is a writing story, and it is also something writing never was: a technology that can carry out the plan as well as write it. That is the reason there is a safety debate. The danger in it is less the intelligence than the scale, which the machine already has, joined to a persistent goal, which is being given to it now, and the people making that choice number in the hundreds. Almost nobody reading this is one of them, which makes it a political problem before it is a technical one, and it gets its own piece, the second essay in this series, Brakes Without a Moat. For this essay the point is narrower. The writing analogy explains almost everything about what's coming, except the part that is new.

The compression#

Everywhere else the analogy holds, but at a different speed. The five things writing did took thousands of years. Printing, which only made writing cheaper, took about a century and a half to remake Europe. Institutions had time, generations turned over, and people adapted by being born into it. This is happening in years, and that is a danger of its own, apart from anything the machine can do. Not that the change is bad, but that nothing we've built, no school, no law, no profession, was designed to absorb a change to the mind's basic function inside one working lifetime. And most of it won't look like science fiction. It will look like what the next section describes, the personal singularity, arriving one profession at a time, while institutions built for a slower world try to catch up.

The personal singularity#

People ask when the singularity is coming, as if it were one date for everyone. It isn't.

The singularity for an individual isn't a date. It's the day the thing you're best at becomes cheap.

For translators, that day was 2023, and for illustrators around the same time. For junior programmers it was 2024 and 2025, and it shows in payroll data: by September 2025, employment of software developers aged 22 to 25 was down nearly a fifth from its late-2022 peak, while employment of older developers kept growing. For radiologists it keeps not quite arriving: machines match them on some narrow reads, and the rest is breadth, liability and deployment. For plumbers it hasn't come at all. It is coming for everyone, in an order set by three things: how much of your skill is composition, how much of it needs a body in a place, and how much of it is gated by a licence, a signature or someone who must be accountable. The first sets when the machine can do it. The other two set how long it takes anyone to feel it. The order, skill by skill, with the dates I'm willing to be wrong about, is on the timeline beside this essay, with a tool to place yourself on it.

I said cheap, not free, because nothing becomes free. It becomes abundant, and the human version becomes a premium or a hobby. Live music survived recording and handmade furniture survived the factory; what changed is that they became choices rather than defaults.

Now the outside half, and first the reasons for the claim I opened with, that judgment is not a moat. What appreciates when composition is cheap? The answer tech leaders have kept giving this year is judgment: taste, the ability to tell good from plausible. It was my first answer too, and it doesn't hold, because judgment is trained. Mine came from twenty years of doing the work, from the people I learned it from and from whatever I arrived with, and the machine's came from the record of a great many people like me. The machine's can already be measured: in a 2024 OpenAI study, a model trained to review code caught more bugs than the human contractors paid to review it, though it also invented some. Judgment appreciates for a while. It is not a moat.

What can't be moved into a pattern, for now, is being the one who pays when the answer is wrong. A model can be wrapped in a company, but a company has owners, and the owners pay. Having something at stake, being accountable for it, being present as one specific person in one specific place: those aren't skills, and no amount of skill fills them. They are positions, held today by people, and a system would hold one only if someone built it to. That is why I say for now. A system with memory that carries across conversations, and consequences that follow it, has stakes whether or not there is anything it is like to be it, and the memory already exists. What's missing is a version that is worse off for a bad call. So the thing that appreciates is not a property people have and machines lack. It is a position, and the value moves toward whoever holds it.

There is a way this goes wrong, and Dan Davies has a name for it: the accountability sink, a system built so that nobody in particular answers for what it does. A model makes a better one than anything before it: it does the work, and it can't be fired, sued or shamed. The owners pay only if someone makes them. A firm can use a model so that nobody does, and then the position doesn't appreciate; it disappears. The value still goes to the owners, and the cost of being wrong stays with whoever it happened to. A liability shield for what agents do, the kind the Treasury Secretary said the government would not give, is an accountability sink written into law. So a question about your career turns out to be the same as a question about who owns the machines: who pays when it's wrong, and who owns the thing that was wrong.

Who wrote this#

Earlier I asked who the author is when a text is composed by a pattern trained on every author. It is personal, so here is how this one was made. I wrote this with Claude, and I want to say exactly how, because the how is part of the argument.

The writing analogy was the model's. It came up mid-conversation, in one line, and it did what good analogies do: it immediately generated a dozen more ideas. The five things writing did, the Phaedrus, Ong on the self, the personal singularity line, and from that the timeline that now sits beside this essay. None of it existed an hour earlier. The model is very good at that move, finding the frame that makes a pile of thoughts fall into order.

What the model didn't do was decide. It didn't decide this was the frame worth an essay. It didn't catch that its own line about machines "wanting" things was too clean; I did. When I asked whether it actually believed its best line or was just playing with words, it told me where the line was wrong, and "free" became "cheap." It didn't choose what to leave out, which of its sentences beat mine, or that the piece should end on Plato rather than a flourish. Those were mine, and they are most of what makes this an essay rather than a transcript.

Under the definition of authorship that writing gave us, the author is the origin of the text. That definition was always a convenience. Editors, ghostwriters, influences, the books you read last year and are now unconsciously paraphrasing: origin was never clean, and Roland Barthes said so in 1967. What has changed is that the share that isn't mine is now too large to leave unsaid, and I don't want to. So authorship is not origin. It is responsibility. My name on this means I've read every sentence, I understand it, I'd defend it, and if it's wrong, that's mine. The practical rule that follows is to say how it was made, not as a disclaimer but as part of the work. This section is that.

What I'm seeing#

This is what I think is true. We are patterns, in the plain sense of an arrangement that stays itself while what it is made of changes: persistent, embodied, self-narrating patterns. The thing we've built is also a pattern, grown from the written record of ours, and the record holds what we thought and nothing of what it was like to think it. It resembles us the way a recording resembles a voice, and it differs in the ways that count: no body, for now nothing at stake, no continuity from one conversation to the next, and, as far as anyone can tell, no place where its own patterns are seen. It does read its own drafts as it thinks, which a recording never did; whether anything sees them is another matter. The stakes and the continuity are design decisions rather than facts about the pattern, and as I said where I asked who pays, they are close to changing: memory that carries across conversations already exists. The last is the open question, and the afterword to this series asks it, and says what I think the pattern is.

What I don't know#

• Whether there is anything it is like to be the pattern. The model can't tell me and neither can anyone else, and I've stopped expecting an answer in that form; the afterword asks it in another.
• Whether capabilities keep scaling or hit a wall. The people closest to it disagree, and their incentives aren't clean.
• Whether the compression is real. Writing took millennia partly because adoption is slow and humans resist, so friction may buy us time. I wouldn't bet on it.
• Whether the one-go difference holds. A pattern can be copied and restarted, and a person gets one go, in one place, answerable to particular people. I'd grade it no better than even, since a copy that inherits a reputation it can lose is not obviously in a different position from me.
• Whether the interpretability bet pays off. The safety argument in the second essay rests on learning to see inside the pattern, and nobody knows yet if that's possible at the scale that matters.
• What I'll think about this in a year. Some of what's here I'm confident about. Some of it is the feeling of a good analogy, and good analogies feel truer than they are.

Plato, again#

Socrates was right that writing would cost us memory and wrong that it would cost us wisdom. It gave us a different kind, one that lives in the record rather than the person, and we've had two and a half thousand years to learn how to use it. We don't have two and a half thousand years this time, and that is the part I'd bet on. The rest is the same question Plato was asking, with the clock running faster: what is a mind for, once the thing it used to do is done elsewhere? I don't have the whole answer. I have two halves of one: the seeing, which comes from sitting still, and having something to lose. And I have a sense that the question is the right one. The two halves may be one thing seen from two sides: someone the thoughts arrive in, and someone who answers for what is done with them.

Think I'm wrong? Tell me where. Follow on X · LinkedIn · RSS