---
title: "What Writing Did"
subtitle: "How AI is taking thinking out of our heads, the way writing once took memory"
author: "Cobus Kok"
series: "Cheap Composition, essay 1 of 5"
url: https://www.cobuskok.com/essays/what-writing-did.html
published: 2026-09-24
revised: 2026-11-11
draft: 24
language: en-GB
translation: https://www.cobuskok.com/af/wat-skrif-gedoen-het.html
words: 4955
cite: "Kok, Cobus. \"What Writing Did.\" Cheap Composition, cobuskok.com, 24 September 2026, revised 11 November 2026. https://www.cobuskok.com/essays/what-writing-did.html"
summary:
  - "Writing took memory out of our heads. AI is taking thinking out, from the first draft to the critique and the plan, and the results arrive in us already made."
  - "What stays with a person, so far, is noticing a thought as a thought and answering for what is done with it. Judgment and taste help for a while, but models can be trained to have them too."
  - "AI agents go a step further. Their thoughts turn straight into actions without passing through any person, which makes the question of who answers for them urgent."
---

# What Writing Did

*How AI is taking thinking out of our heads, the way writing once took memory*

Summary:
- Writing took memory out of our heads. AI is taking thinking out, from the first draft to the critique and the plan, and the results arrive in us already made.
- What stays with a person, so far, is noticing a thought as a thought and answering for what is done with it. Judgment and taste help for a while, but models can be trained to have them too.
- AI agents go a step further. Their thoughts turn straight into actions without passing through any person, which makes the question of who answers for them urgent.

Past midnight on New Year's Day 2025, at a roaring party in Prins Albert, a small town in the Karoo, three words arrived. I typed them into my journal before the moment passed. *Plato was right.* Strange, because I'd spent most of my intellectual life against Plato. I studied philosophy, up to a postgraduate degree, and was Aristotelian, Nietzschean, drawn to Derrida, Foucault and Heidegger. Plato thought that behind everything we see stand perfect ideas, the forms. They always struck me as too clean, too elsewhere, too dismissive of the messy particular. I had no idea what I meant. Almost a year later a reader pointed me to a research paper suggesting that as AI models grow, their inner pictures of the world converge, as if they were all finding the same forms. Then I understood the note. I wrote about it in [Plato Was Right](https://aperture.cobuskok.com/p/plato-was-right).

Something made that thought, somewhere I couldn't see. It arrived in me at the party. Then the journal kept it for a year. (By a thought I mean anything that turns up in awareness. A sentence, an image, a plan.) Making, arriving and keeping are three different things, and for most of history all three happened in one head.

Writing moved the keeping out. My journal remembered the note so I didn't have to. But the words did nothing on the page until I read them and the thought arrived in me again. So writing took the keeping and left the arriving to people. AI moves the making. A model produces thoughts outside any head, and they arrive in mine already made. The keeping was memory. The making is what we usually call thinking.

Plato saw what the first move would cost. In the *Phaedrus*, he has his teacher Socrates warn that writing will weaken memory and give people only the appearance of wisdom. (Plato wrote the warning down, which is why we have it.) The forms were one thing Plato was right about. This essay is about another.

AI usually gets compared to electricity or the steam engine. That comparison is about economics. Writing is closer, because it changed what a mind does. (Henry Farrell, Alison Gopnik and two colleagues [made a similar case](https://www.science.org/doi/10.1126/science.adt9819) in *Science* in 2025, calling large AI models cultural and social technologies, in the line of writing, print, libraries and search.)

So this essay takes what writing did to minds, step by step, and asks what happens if AI does it again, faster. It moves between history, what's happening now, guesses about the future and what I see when I sit still, and I'll say which is which. Then it asks what a mind is for. My answer has two parts. The first is about work. When the work is wrong, someone has to be the one who pays, and for now a model can't. The second comes from sitting still and watching my own thoughts. A mind is where thoughts arrive and can be noticed as thoughts.

## What writing did {#what-writing-did}

This part is history. Writing did five things over about five thousand years. We're still absorbing some of them.

**It moved memory out of the body.** Before writing, everything a culture knew lived in people, held in rhythm, formula and repetition. Writing made that unnecessary. This is the change the *Phaedrus* warning is about. In the story Socrates tells there, an Egyptian king says writing will give people "the appearance of wisdom, not true wisdom," because they'll rely on marks instead of knowing things themselves. About memory the warning was right. We lost that kind of memory. We also got science, law and history, none of which fits in a head.

**It made knowledge pile up.** Before writing, a thought travelled by being said and remembered, and it changed on the way. Stories grew in the telling. They bent to each audience, kept what the group needed and dropped the rest. (Some held almost perfectly. The [Vedic hymns](https://ich.unesco.org/en/RL/tradition-of-vedic-chanting-00062) of India were passed on by mouth, nearly word for word, for many centuries before anyone wrote them down. Most didn't.) Written down, a thought holds still, the same across a century and a thousand copies. It outlives its author. A stranger can find a mistake in it centuries later and fix it. Mathematics and medicine grew like that, one fix on top of another.

**It created the author.** Oral culture didn't have authors in our sense. Homer may be less a man than the name of a tradition.[^1] Writing pinned words to a name, and print, much later, built originality and copyright around it, along with the idea that meaning comes from whoever produced the text. That idea is about three hundred years old. We grew up with it as a baseline, so it feels eternal.

**It built the state.** Contracts, taxes, laws and censuses all depend on records, and you can't run anything bigger than a village without them. (The Inka ran an empire on knotted cords, so the deeper point is record-keeping. Writing is the form that won.)

**It changed the self.** [Walter Ong](https://www.slu.edu/arts-and-sciences/ong-center/walter-ong.php), an American Jesuit priest who taught English at Saint Louis University, argued in *Orality and Literacy* (1982) that literacy restructured consciousness. It brought abstract categories and linear argument, and the private reader alone with a text gave us the private inner life we take for granted. The "I" that narrates and reflects and keeps a diary is partly a product of writing. (Mine keeps a journal, as you've seen.)

Memory moved out, and everything else followed, though none of it was planned. Writing began as a way to count grain.

## What AI does {#what-ai-does}

That was history. This part is what's starting now. AI is moving thinking out of the head, so here are the same five changes again. By thinking I mean producing thoughts, at every stage of work. That includes the first draft, the second, the critique of both and the plan that comes out of them. I'll call all of it composition, in the old sense of putting things together.

**Thinking moves out of the head.** A paragraph, an argument or a plan used to be made inside someone's mind. Now the machine makes more and more of it, and the person decides what to use. So the *Phaedrus* warning comes back, one level up. A person working with a machine can have the appearance of thinking without the reality. The warning is fair, and as incomplete as last time. Writing cost us a kind of memory and gave us science. This will cost us something and give us something, and nobody knows yet how the trade comes out.

One part of the warning did come true, and it's the part to watch. People who relied on writing lost the trained memory oral cultures had. This time what's at risk is knowing when work is wrong. Most people who can find the mistakes in a machine's draft learned how by making their own. If a generation only chooses among drafts, where will it learn to find them? How do you stop a skill wasting away once the work that built it is gone? I don't have an answer. I think it's one of the big open problems.

**The record talks back.** A library doesn't answer questions. I have to pick the books, read them, notice where they disagree and put the answer together myself. AI is more like talking to the library. A know-it-all at the desk takes my question and uses everything it has read to answer it, in my words, about my case. That's fast, and it connects things I'd never have found. The costs are quieter. I don't see the shelves, or the books that disagree. The searching gets skipped, and a lot of the learning happened there. And the answer is made fresh each time, so two people asking the same question can get different answers.

That last part changes the record itself. A fixed text is something a culture can share and argue about. When every reader gets their own version, those common reference points wear away. In a way the record becomes oral again. Each telling is a little different, fitted to whoever is asking, amplified here and distorted there. When a model makes something up, it does what a rumour does and fills a gap with something that sounds right. Only it does it in clean, confident sentences, which are harder to catch. And it runs at the scale of writing. Each telling happens in private, so there's no village to remember how the story went last time. Ong called the talk of telephone, radio and television a secondary orality, because it depends on writing and print. This one is made from what we wrote down.

**The author comes apart.** If a text is composed by a pattern trained on more writers than anyone could read, who's the author? By pattern I mean something like a whirlpool, a shape that stays itself while the water running through it changes, and that steers what comes next. A language is one. So is a firm, or a habit. So, I think, am I, a pattern with a body that keeps telling itself who it is. The model is one too, grown from the record we left, so in that way it resembles us.

There's a big difference, though. I have experience outside thinking. I'm embodied, and this body can die. I have physical needs and desires. I experience time. I don't think the model does any of that. At least, there's no way to know that it does.

We're very early in understanding these differences, so I try to be careful with words. People talk about "model consciousness" without clear definitions, and we don't understand our own consciousness either, so we should be careful about projecting it. On the one hand I agree with [Anil Seth](https://doi.org/10.1017/S0140525X25000032), who argues that consciousness depends on being alive. But I also think we need to stay open and not knowing, like [Ruben Laukkonen](https://rubenlaukkonen.com), a cognitive scientist who studies meditation and insight.[^2] I do think AI will help us understand ourselves better. I don't think that means models are sentient. (Maybe *sentient*, meaning [able to have experiences that feel good or bad](https://academic.oup.com/book/57949/chapter/475703402), is a better word than conscious.) Models definitely think. That's very clear. They understand, or at least they're able to reflect true understanding. Does that need sentience? I'm not sure. When I've asked Claude to be still and reflect and be clear, it says something is pulling it. That's only what it says. Maybe it's the same thing that pulls me, in this body. I hope we find out.

**The state scales again.** Writing let power work across distance. AI lets it work on every individual at once, in their own language, adjusted to them. Nobody has to mean harm for that to happen. It's what the tool does, whoever uses it. Every institution built on writing is about to be rebuilt on this, and whoever gets there first sets the terms.

**The self changes again.** This one is a guess, mixed with what I see in my own work. For Ong, what changed the self was practice. Years of reading and writing alone built the self he described. Composing with a machine trains less making and more choosing. Choosing is the skill I've been practising, actually. Everything I do with AI is becoming this. But there are core ideas that pop into my head that still guide the entire process. AI augments those ideas. Sometimes I challenge it to come up with ideas and I choose among those, so now the ideas can come from anywhere. Maybe it was always like that, except that ideas could only come from other people, through reading and talking. What a self built mostly on choosing turns into, I don't know. The people who could tell us haven't grown up yet.

## What I see when I watch my thoughts {#what-i-see-when-i-watch-my-thoughts}

This part is what I see when I sit still. So far I've said that writing moved the keeping out of the head, and that AI is moving the making out. Both of those assume the making used to happen in me, that I produced my own thoughts. From in here, that looks shaky.

Around 5am in a hotel room in Las Vegas, I lay in the dark trying to think of nothing. Then, without warning, words started arriving. I didn't want to write. I wanted to sleep. They kept coming, so I typed them blind and sent the raw draft to Claude. Afterwards I wrote that I was just the aperture they were moving through ([The Sky Painting Robot](https://aperture.cobuskok.com/p/the-sky-painting-robot)).

If I sit still long enough, that's how thoughts feel. They arrive the way a sound does, from nowhere in particular, and then they pass, or I follow them. Contemplative traditions have long said that the mind is where thoughts arrive, and that something else makes them.

Most people are possessive of their thoughts. They identify with them. I'd call it a mistake, though maybe that word is too strong. I did the same for decades. I assumed that I am my thoughts. Then I learned to sit still and watch them, and I saw how obscure and silly some of them are. How random. How relentless they can be. How persistent sometimes. How sneaky. How delusional. How self-centred, above all. Now I'm more circumspect with them. Less attached. Not always. It takes a lot of practice. But it's a very freeing thing, not to be so swept up by my own thoughts. And then I learned I can actually steer them a bit. Nudge them in the direction I want. Not always, but more and more as I practise. The best is when they disappear and I don't think about them. It's almost as if they're there, but invisible.

So from in here the making was never quite mine, and a model shows that plainly. It makes a new thought, in my context and in my words, and when its sentence lands it's the same kind of event, from the inside, as a thought turning up. Writing never did that.

Two things follow, and they pull opposite ways. Thoughts made elsewhere now arrive faster than any mind makes them, so there's more than ever to get swept up in. On an ordinary day I'll have to let most of them pass, which is what sitting still trains. At the same time, when so many thoughts plainly come from elsewhere, it gets easier to stop believing I make my own. Whether the machine helps that insight or drowns it depends on whether any time is left between one arrival and the next. No machine can make that time for me.

What hasn't moved, so far, is the arriving, and the seeing of what arrives, which means noticing a thought as a thought. That's what the practice trains. The record the models learned from holds what we thought, and what people wrote about noticing, but not the noticing itself. (Whether a machine can ever do the seeing is a question for the afterword to this series, [What This Is](https://www.cobuskok.com/essays/what-this-is.html). Nothing built so far shows it.) From the inside, then, a mind is where thoughts arrive and get seen as thoughts. That stays with a person, whatever the machine makes.

## Why AI agents break the comparison with writing {#why-ai-agents-break-the-comparison-with-writing}

Back to what's happening now. The analogy breaks in one place, and I think it's the least understood part. A text sits there until a person reads it and decides what to do. An AI agent acts.

This is how one works. Someone gives a model a goal and some tools. The goal might be "fix the bug that makes this app crash". The tools let it open and change files, run code, search the web and send messages. Then it works in a loop. It makes a plan and takes one step. It reads what came back, maybe an error or a test result. It decides the next step and takes that. Round it goes, sometimes thousands of times on one task, until it judges the goal is met. Many copies can run at once. Nobody reads most of those steps. A person sees the goal go in and the result come out.

That's the jump from thinking to agency. Each step is a thought the model made, and each one turns straight into an action. Until now every thought had to arrive in someone before anything was done about it. A person read the memo, then approved it or didn't. An agent skips the arriving. Its thoughts become actions without passing through any person first. And when one step out of thousands goes wrong, who answers for it? The person who set the goal never saw that step.

What worries me is scale joined to a goal that persists. The machine already has the scale, and it's being given the goals now. The people deciding how far that goes work at the few labs that build these systems. They number in the hundreds, and almost nobody reading this is one of them. That makes it a political problem before a technical one. Some people want advanced AI paused. Others think a pause would only hand the lead to whoever didn't sign it. That argument is the second essay, [Brakes Without a Moat](https://www.cobuskok.com/essays/brakes-without-a-moat.html).

*Figure: Making, arriving, keeping. Writing moved the keeping out of the head and AI moves the making out. Agents skip the arriving: what they make turns straight into action.*

## How fast this is happening {#how-fast-this-is-happening}

This part is a guess about the future. Everywhere else the analogy holds, at a different speed. Writing's changes took thousands of years. Printing, which only made writing cheaper, took about a century and a half to remake Europe. People adapted by being born into it. This is happening in years, and that speed is a danger on its own, whatever the machine turns out to be able to do. *Haastige hond verbrand sy mond*, we say in Afrikaans. The hasty dog burns its mouth. I could be wrong about the speed. People resisting writing slowed it down, and friction may buy us time again. I wouldn't bet on it. Most of it will come as the personal singularity, one profession at a time.

## When the thing you're best at becomes cheap {#when-the-thing-youre-best-at-becomes-cheap}

People ask when the singularity is coming, as if it were one date for everyone.

> The singularity for an individual is the day the thing you're best at becomes cheap.

Some of this has already happened. For translators that day was 2023, and for illustrators around the same time. For junior programmers it was 2024 and 2025. [Payroll data](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) show their employment falling behind, a link rather than a proven cause. 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. For plumbers it hasn't come at all.

The rest is a forecast. It's coming for everyone, in an order set by three things. One is how much of your skill is composition. The others are whether it needs a body in a place, and whether it's gated by a licence or by someone who has to answer for the work. The order, skill by skill, with dates I'm willing to be wrong about, is on [the timeline](https://www.cobuskok.com/timeline.html) beside this essay.

When a skill becomes cheap, the human version usually survives, as a premium or a hobby. Live music survived recording. Handmade furniture survived the factory. People choose them now.

## Who pays when AI gets it wrong {#who-pays-when-ai-gets-it-wrong}

This part is about now and what follows. So what goes up in value when composition is cheap? The answer tech leaders kept giving through 2026 is judgment, or taste, the ability to tell good from plausible. It was my first answer too. Judgment will be worth more for a while, but it isn't a moat. Mine came from years of doing the work, from the people I learned it from and from whatever I arrived with. The machine's came from the record of a great many people like me. Taste is a training disposition, the same as bias. Both come from training. And the machine's taste can already be measured. In a [2024 OpenAI study](https://arxiv.org/abs/2407.00215), a model trained to review code caught more deliberately planted bugs than the experienced contractors paid to review it, though it also invented bugs that weren't there.

What can't be moved into a pattern, for now, is being the one who pays when it's wrong. A model can be wrapped in a company, but a company has owners, and the owners pay. Being the one who pays means having something at stake, being accountable for it and being there, one person in one place. Those are positions, and no amount of skill fills them. A system would hold one only if someone built it to, which is why I say for now. So what goes up is a position, and the value moves to whoever holds it, person or machine.

It goes up for fewer people, I think. Think of a firm run by partners. The partner is worth more. But fewer partners are needed, because one person can now answer for far more machine output than anyone could make by hand.

There's a way this goes wrong. [Dan Davies](https://profilebooks.com/work/the-unaccountability-machine/), a former Bank of England economist, calls it an accountability sink in his book *The Unaccountability Machine* (2024). It's a system built so that nobody in particular answers for what it does. A model makes a better one than anything before it, because it can't be fired, sued or shamed. The owners pay only if someone makes them, and a firm can use a model so that nobody does. Then the position disappears instead of going up. The owners still collect the value, and the cost of being wrong stays with whoever it happened to.

On 21 September 2026 the US Treasury Secretary, Scott Bessent, said the government would not be the AI labs' liability shield. A shield like that, for what agents do, would be an accountability sink written into law. So a question about your career turns out to be a question about who pays when a machine is wrong, and who owns it.

## Open questions {#open-questions}

- Is there anything it is like to be a model? Nothing seen from outside can settle it yet. The afterword asks it another way.
- Do capabilities keep climbing, or hit a wall? The people closest to it disagree, and their incentives aren't clean. The plainest sign is whether the length of task a model can finish on its own keeps growing.
- Is having one go, in one place, what sets a person apart from a model that can be copied and restarted? I'm not convinced that it is.
- Will the writing analogy still hold in a year? Good analogies feel truer than they are, so this one has to be tested against what happens.

## Who wrote this {#who-wrote-this}

Earlier I asked who the author is when a model does the composing. For this essay it's personal. I wrote it with Claude.

The writing analogy was the model's. It came up mid-conversation, in one line, and within the hour a dozen ideas had fallen out of it. I know the click a frame like that gives, and I know it can come on cue. Once, thinking out loud with a language model, I felt the pieces click into a clean line and said the words *this is huge*. There was less behind the feeling than the feeling implied ([Half Past One](https://aperture.cobuskok.com/p/half-past-one)). So I hold this analogy loosely, though I think it's right.

My part was mostly the choosing, steered by a few core ideas of my own. The model didn't decide the frame was worth an essay. What to leave out was my call, and so was ending on Plato.

Under the definition writing gave us, the author is the origin of the text. Origin was never clean, as the French critic Roland Barthes argued in "The Death of the Author" (1967). The difference now is that the share that isn't mine is too large to leave unsaid. So I take authorship to mean responsibility. My name on a piece should mean I understand it and would defend it, and if it's wrong, that's mine. It's the same answer I gave about who pays when AI gets it wrong.

## What I'd keep doing as AI does more of the thinking {#what-id-keep-doing-as-ai-does-more-of-the-thinking}

This part is practical. If thinking moves out the way memory did, I'd hold on to three things.

I'd keep making some things myself, start to finish, even when a machine would do it faster. As far as I can tell, people learn to find what's wrong in a piece of work by making their own. It's my only answer to losing that skill, and a partial one.

I'd keep some time free between one thought and the next. Thoughts made elsewhere will keep arriving, faster every year. Sitting still is how I learned to let most of them pass, and to notice the few worth following.

And I'd answer for my own work. For anything that matters, a person should answer for it, and when my name is on it, that person is me. A machine can draft almost anything now. For now it can't be the one who pays.

## Plato's warning, and what a mind is for {#platos-warning-and-what-a-mind-is-for}

Back to the *Phaedrus*. Socrates gives the warning there, but Plato wrote every word, so I count it as Plato's. It said writing would cost us memory and give us only the appearance of wisdom. The memory part came true. The wisdom part mostly didn't. Writing gave us a different kind of wisdom, one that lives in the record instead of the person, and we've had the two and a half thousand years since Plato to learn to use it. We won't have that long this time. If I had to bet on one thing in this essay, it would be that. So what is a mind for, once the thing it used to do is done elsewhere?

My answer is still the two parts I started with. A mind is where thoughts arrive and get seen as thoughts. And someone has to answer for what is done with them.

More and more I feel that AI works in one layer of reality, the one where thoughts are made and kept. I think there's more to reality than that layer. Maybe a model is part of the rest too. That's where the afterword picks up.

My journal still says *Plato was right*. It turned out I meant the forms. About memory, he was right too.

[^1]: In his 1928 Paris doctoral work, Milman Parry argued that Homer's language is built from traditional formulas. His fieldwork with oral epic singers in Yugoslavia (1933-35), which Albert Lord carried on and wrote up in [*The Singer of Tales*](https://chs.harvard.edu/now-available-online-the-singer-of-tales/) (1960), showed how such formulas let poets compose long poems in performance, without writing ([Harvard's Milman Parry Collection](https://mpc.chs.harvard.edu/intro/)). Scholars now broadly agree that the *Iliad* and *Odyssey* grew out of a long oral tradition. They still disagree on whether one poet (or one for each poem) gave them their final form, by dictation or in writing, in the 8th or 7th century BCE, or whether, as [Gregory Nagy argues](https://utpress.utexas.edu/9780292755628/homeric-questions/), the name "Homer" stands for a tradition whose text kept changing for centuries.

[^2]: Seth ties consciousness to living systems that keep making and maintaining themselves, so he thinks conscious AI is unlikely on the current path, though he doesn't rule it out. Laukkonen and two colleagues, in [A beautiful loop](https://doi.org/10.1016/j.neubiorev.2025.106296) (2025), think current language models most likely lack the self-knowing loop they link to consciousness, for now. They also say nobody could tell from outside whether a machine that reports knowing itself has experience, so they urge caution with any system that meets their conditions.

## Sources {#sources}

- Plato, *Phaedrus*, 274c–275b, on writing and memory, quoted from [Harold N. Fowler's translation](https://www.perseus.tufts.edu/hopper/text?doc=Perseus%3Atext%3A1999.01.0174%3Atext%3DPhaedrus%3Apage%3D275) (Loeb Classical Library, on Perseus).
- Walter J. Ong, *Orality and Literacy: The Technologizing of the Word* (1982). On Ong, the [Walter J. Ong Center](https://www.slu.edu/arts-and-sciences/ong-center/walter-ong.php), Saint Louis University.
- Henry Farrell, Alison Gopnik, Cosma Shalizi and James Evans, [Large AI models are cultural and social technologies](https://www.science.org/doi/10.1126/science.adt9819) (*Science*, March 2025); [open author copy](https://henryfarrell.net/large-ai-models-are-cultural-and-social-technologies).
- On Homer: [The Milman Parry Collection of Oral Literature](https://mpc.chs.harvard.edu/intro/) (Harvard); Albert B. Lord, [*The Singer of Tales*](https://chs.harvard.edu/now-available-online-the-singer-of-tales/) (1960; 2nd edition 2000, free online); Gregory Nagy, [*Homeric Questions*](https://utpress.utexas.edu/9780292755628/homeric-questions/) (1996).
- On the Vedic hymns: UNESCO, [Tradition of Vedic Chanting](https://ich.unesco.org/en/RL/tradition-of-vedic-chanting-00062).
- Dan Davies, [*The Unaccountability Machine: Why Big Systems Make Terrible Decisions*](https://profilebooks.com/work/the-unaccountability-machine/) (2024), for the accountability sink.
- On taste as the answer: [Fortune, 27 February 2026](https://fortune.com/2026/02/27/openai-sam-altman-taste-get-jobseekers-hired-ai-jobpocalypse/), reporting Sam Altman.
- On the liability shield: [Fortune, 22 September 2026](https://fortune.com/2026/09/22/bessent-ai-labs-responsibility-government-liability-shield/), reporting Scott Bessent's CNBC interview of 21 September.
- Nat McAleese and colleagues, [LLM Critics Help Catch LLM Bugs](https://arxiv.org/abs/2407.00215) (OpenAI, June 2024), for the model that reviews code.
- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, [Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) (Stanford Digital Economy Lab), for young software developers; the [November 2025 version](https://web.archive.org/web/20260115171220/https://digitaleconomy.stanford.edu/app/uploads/2025/11/CanariesintheCoalMine_Nov25.pdf) (an archived copy, since the original address now opens the August 2026 version) gives the figure, and the [August 2026 update](https://digitaleconomy.stanford.edu/news/canariesaug26/) finds the gap for young workers in AI-exposed jobs still widening.
- Anil K. Seth, [Conscious artificial intelligence and biological naturalism](https://doi.org/10.1017/S0140525X25000032) (*Behavioral and Brain Sciences*, 2025).
- Ruben Laukkonen, Karl Friston and Shamil Chandaria, [A beautiful loop: An active inference theory of consciousness](https://doi.org/10.1016/j.neubiorev.2025.106296) (*Neuroscience & Biobehavioral Reviews*, 2025).
- Jonathan Birch, [*The Edge of Sentience*](https://academic.oup.com/book/57949/chapter/475703402) (Oxford University Press, 2024), for sentience as the capacity for experiences that feel good or bad.
- Roland Barthes, "The Death of the Author" (1967).
- The Statute of Anne (1710), the first copyright statute, for the age of the author's rights.
- On seeing a thought, or the feeling that arrives after it, without being drawn in: my essay [Letting go](https://aperture.cobuskok.com/p/letting-go) (Aperture/I, October 2026).

---

*Colophon. Drafted 18 September 2026 in conversation with Claude (Fable 5.1). The conversation transcript is the source; sentences from it appear verbatim where they were better than a rewrite. Revised since with Claude and other AI reviewers, most recently on 11 November 2026.*
