There are three groups of people who are confident about writing and AI, and this essay thinks all three are wrong. The people typing a prompt and publishing whatever comes back, confident that this counts as writing. The people who believe the tools will soon produce everything, once the model is capable enough. And the people who believe that if a model was involved at all, the work is slop by definition — that there is nothing human left in it to judge. The argument that makes all three wrong at once is easiest to see in a form that has nothing to do with AI.
Picture three film directors.
The first arrives on set, tells the crew make an awesome action movie, and leaves. The crew is good — experienced, capable, able to turn their hands to almost anything. Left with a fragment that thin, they will do one of two things. Someone among them may step into the vacant role and start making the decisions the director didn’t, and the film will quietly become that person’s instead. Or, with no vision to serve and no one stepping up to supply one, they will aim at the only thing a fragment leaves them: the middle of what they already know, the average of every action movie this crew has made before. It will be competent. Nobody watching it will feel that a particular person stood behind it. There was supposed to be one. He left.
The second director does far more. He casts the film, hires the cinematographer, approves the sets, writes a detailed shot list, produces a whole bible specifying tone and pacing and the look of every scene. He is thorough. And then, before the first take, he leaves — no notes on the dailies, no presence in the edit, no response to anything the film does once it starts becoming itself. The crew executes the plan faithfully. And the film has a ceiling, a hard one, lower than it should be, because a film is not fully knowable in advance. The actor finds something in the third take that wasn’t in the script. The light does something the shot list didn’t predict. The edit reveals that a scene everyone loved is dead on arrival and a throwaway moment is the real hinge. A film is discovered in the making, and the second director was not there for the discovery. He specified a film. He did not make one.
The third director is present throughout. She holds the vision — not the shot list, the vision, the sense of what the whole thing is for — and she carries it through every phase, adjusting as the film teaches her what it is. She keeps the unexpected reading and cuts the one that only flattered the plan. She refuses the take that would pass. When something is wrong she diagnoses at what level it is wrong — a performance, a structure, a premise — and steers accordingly. She does not operate the camera. Nobody thinks this makes the film less hers.
Being in the chair is not the same as being good in it. The third director can be present and ineffective, steering vaguely, never quite finding the level a problem lives on. She can be inconsistent, sharp in one scene and absent in the next. She can micromanage — override the crew’s competence at every turn, refuse them any room to be good at what they are good at, and cap the film below what a lighter hand would have reached. Presence is what makes authorship possible; it does not make it succeed. The chair can be occupied badly. What it cannot be is occupied by no one and still produce a film that is anyone’s.
You already know whose film each of these is. You knew before I finished describing them. The third director’s film is hers. The second director’s is a plan that someone else carried out. The first director’s belongs to no one in particular — perhaps to whoever in the crew picked up the vision he dropped. And you know something else without being told: that nobody would call the third director lazy for keeping her hands off the camera, that nobody would accept the second director’s bible as a substitute for having been there, and that nobody would defend the first director at all. This ranking is not controversial. It is how every collaborative art has worked for as long as collaborative arts have existed. Authorship lives at the frame — in the vision held and the judgment exercised — and it is not diminished by the distribution of the hands.
I have asked you to agree to all of this about film, where it costs you nothing. Now I am going to change the subject, and hold you to what you granted.
This is what using AI to write something is.
The model is the crew. It is talented and it is general — it can render almost anything in almost any register, competently, on demand — and, exactly like a crew, it holds no vision of its own. It has nothing it is trying to say. This is not a limitation to be fixed in a later version; it is what the thing is. A general instrument, capable of everything, is by construction committed to nothing in particular, and vision is a commitment to something in particular. So when a person uses a model to write, the vision has only one place it can come from. There is exactly one chair in the room, and the model cannot sit in it. It cannot even do what a human crew does when the director leaves — nominate one of its own to hold the vision instead — because it has no self to promote into the role. When the chair is empty, it stays empty, and the writing fills with the only thing a general instrument has to offer in the absence of a vision: the average.
Which means the three directors are no longer an analogy. They are a taxonomy of how people actually use these tools.
The first director is the person who types a prompt and publishes what comes back. This is slop, and I have no interest in defending it. But notice why it is empty. It is not empty because a machine made it. It is empty because the one chair that could have shaped it was vacated the instant it was occupied. The blandness of slop is not the model’s authorship showing through — the model has no authorship to show. It is the shape of an absence, the negative space where a director’s judgment was supposed to be. General crew, no vision, the average. The emptiness has a location, and the location is the empty chair.
The second director is the one whose engagement stops at the spec. Sometimes that means an elaborate, many-paragraph brief — tone, structure, constraints, the whole bible — front-loaded, and then the button pressed once and whatever comes back taken as the film. Sometimes the spec is arrived at rather than dumped: a stretch of back-and-forth with the model first, finding the right place in the space, discovering that the request was misunderstood and reframing it, before the real generation begins. The brief can carry real expertise, too — imported from prior work, prior experimentation, a hard-won sense of exactly which prompt lands where. This is a wide rung. Its bottom is barely above slop; its top can look, from the outside, like real work, and sometimes is. But all of it shares one ceiling, for one reason: writing is discovered in the doing, and a specification — however rich, however expertly aimed, however much argument went into finding it — is not the doing. It is the plan for the doing, handed to the crew, with nobody present when the plan meets the material. And this is where the camp that is enthusiastic about these tools makes its own version of the mistake. The dream of the model that will one day, once it is capable enough, write a great novel from a brief — that is the second director’s dream, and it has a ceiling no improvement in the crew will raise, because the ceiling was never set by the crew. It is set by the director’s absence from the discovery. I reject that dream as firmly as I reject the slop, and from the same principle.
The third director is the one I have been describing across two essays. The vision held and carried. The diagnosis at the right level. The steering by argument. The refusal. The work discovered through the doing, over weeks or months, in a dialogue present at every take. This is arguing into existence, and it is not a better way of prompting. It is a different chair entirely — the only one of the three that is actually occupied.
Now the thing this essay is for.
There is a reflex that has become the standard response to any writing a model touched: there are two kinds of work, human and machine, and if AI was involved this is the machine kind, which means no one really wrote it. The binary is the whole move. Human or AI, one or the other, and AI-involvement drops the work into the second bin, where authorship is not transferred to the machine — the machine is nobody — but simply absent. Not the model wrote it. No one wrote it. A void where the author would be.
Watch what this costs the person who believes it. They can still be angry at the slop — and they are, rightly. They will still hold the slop-merchant responsible for the mess, for flooding the shelves with hollow books, for wasting a reader’s time and money. The blame survives. What does not survive is the one thing that has ever improved the output of any medium: the ability to say do better. Because do better presupposes that better is possible — that there is a craft here to be worse or better at, a ladder to be low or high on. And they have denied exactly that. If AI-involvement means no author, then there is no directing to do well or badly; there is only the doing of it or the not-doing of it. So the blame, which they keep, can carry only one instruction. Not do better. Only don’t. Stop. A moratorium.
And here is the difficulty with that, stated as plainly as I can: a moratorium has never once stopped a new medium, and it never will, because it is fighting the one thing that cannot be suppressed. Expression does not ask permission. Hand people a new means of making things and they will make things with it — most of them badly, some of them well, a few of them greatly — because human creativity has never once met a new medium and declined to bend it, eventually, to art. Photography did not wait for permission. Recorded music did not. Cinema did not. The demand to simply not do the thing is the single response with a perfect record of failure, and it is the only response the no-author premise leaves standing. The people who most want to stop the slop have argued themselves into the one posture from which it cannot be stopped. They can say stop forever. They have defined themselves out of get better — which is the only thing that has ever raised the floor of anything.
The way out is not to defend the slop. It is to take the chair seriously. Grant three things — that interacting with a model to make a piece of writing is a real activity; that doing so seats you in the directorial chair, accountable for what comes out of it; and that there is a ladder here, positions above the one the slop-merchant occupies, all the way up to the third director. Grant those, and you lose nothing you wanted. The mess was always his; that does not change. What changes is that the responsibility gets a shape. It is no longer the diffuse fault of having-used-the-tool-at-all, which can only be answered by not using it. It is the specific fault of having taken the chair and vacated it — of prompting and walking away, the one move that guarantees the work is no one’s. And that can be answered. Not don’t do it. Do better. There is a better. It is above him on the ladder, and his refusal to climb is his.
You granted all of this about film, where the crew was a union and no one was confused. The first director is real; the third director is real; the distance between them is the distance between an abandoned chair and an occupied one. That the crew is now a model rather than a union does not change the shape of it. To deny it here — to insist that when the crew is a model, there is no directing at all, nothing being done that could be done well or badly — is to claim that the role every other collaborative art turns on simply does not exist in this one. Not that it is mostly occupied badly, which is true. That it is not there to occupy. That is not a small claim. It is an extraordinary one, usually made without noticing it is being made. And if the denial feels more natural here than it would anywhere else, there is a reason: writing has mostly been a solitary art. It never had the chair because it never had a crew. The chair arrived with the crew.
Two objections survive this far.
The first is the retreat from no one wrote it to nothing good is possible with it — and notice what the retreat quietly surrenders: you cannot say the directing produces nothing good without conceding that directing is being done. The chair whose existence was just denied is suddenly real enough to fail. This one has to be answered carefully, because it is often not an aesthetic judgment at all. For many people who hold it, the impossibility of good AI writing is load-bearing for other commitments — objections to the companies, to the environmental cost, to what the tools are doing to labour and to learning. Those objections may be entirely right. But they make good work is possible an expensive thing to concede — the position was cleaner when the tools were merely worthless — and a claim held for that kind of reason does not yield to evidence, so I will not offer any. I am going to point at what it commits its holder to. Nothing good is possible means, in the taxonomy above, the third director produces nothing — not the first, they can have the first, but the third, present throughout, discovering the work in the doing, refusing the adequate take, carrying a vision across the whole length of the making. Everywhere else that configuration is the one from which the best work comes. To hold that it produces nothing here is to hold that this single medium inverts a rule every other collaborative art obeys — a second extraordinary claim, purchased by abandoning the first, and like the first it owes an account of why. The account is never given.
Or rather, the things offered in its place are all, on inspection, arguments about the crew. The model does not understand. It does not intend. It only predicts the next word. Grant every one of these — I am not disputing them here — and notice where they land. They are claims about the instrument, and the essay never rested on the instrument understanding or intending anything. The director understands and intends. The camera does not, and never did, and the film was authored all the same. An argument that the model lacks understanding is an argument aimed at the crew, in a debate about who is in the chair. It leaves the chair untouched. The inversion needs a reason the directorial role itself fails in this medium, and that reason is never supplied, because the role does not care what the crew is made of.
The second objection is sharper, because it is aimed at a specific thing rather than a category. It says: you, specifically, did nothing. Mare Non Nostrum is nothing of mine. The Wire Branch is nothing of mine. And this is the better objection, because it is falsifiable — and false. To say I did nothing is to say the work is first-director output, that a prompt produced it. So let us take that seriously enough to break it.
The crude form collapses on contact. A dense enough prompt could one-shot the book. Trivially true and empty: the densest possible prompt is the book itself. The stronger version imagines a specification far richer than any summary, one that instructs not just the events but the moral texture: the withholding of resolution here, the desynchronised legal and emotional plots there, a mercy whose provenance can never be recovered even by the one who granted it, every motif built to bear weight on its second appearance. Give the model that, the objection runs, and where is your authorship now?
It is in the specification. That is the entire answer. A brief containing all of those decisions is not an alternative to directing the book; it is the direction of the book, written down in advance. And to one-shot Mare Non Nostrum this way, the brief would have to contain not a hundred and twenty-five thousand words of prose but every decision beneath them: the thesis about great-man history the counterfactual is built to test, the sociolinguistics of a civilisation dense enough to mint its own idioms, the interlocking causal logic that keeps two great powers in the Mediterranean for four and a half centuries without either destroying the other. A prompt that specifies all of that, coherently, together, is not a prompt. It is a description of authorship. The only question is where such a specification comes from, and there are two answers. Either it was arrived at through iteration — the arguing done all the same, merely upstream, poured into a document instead of a dialogue; that is the second director at minimum, and the second director is not nothing. Or it was held complete in a single mind and set down in one pass, which is not how minds work for anything of this density, and which, if it somehow were, would make that mind more the author, for having done the whole discovery internally before typing a word. The dense prompt does not remove the authorship. It relocates it, and then has to account for its own origin, and every honest account of that origin is a person in the chair.
This is not only a philosophical point, because I have run the experiment, and will report further in an upcoming essay. I took one short story, Lagan, set in the same world as Mare Non Nostrum, and had two models — Claude Fable and GPT-5.5 — regenerate it from summaries of varying length, then set the results beside the real thing. The denser prompts produce better forgeries. They do not produce the original, and the reason is exact. The features that make Lagan what it is are relational, discovered in the doing, and a summary flattens precisely those. Shallow prompting reverts them to the narrative default: the institution that bears down on the characters becomes a single human villain, and the world that moves the protagonist becomes a protagonist who drives the plot. The fought-for structure does not survive the summary, because a summary is precisely the flattening of it. You can specify your way to a closer and closer knockoff, and it is an open question how close a rich enough brief could come — but a brief that rich is the previous paragraph’s specification again: fought for somewhere, by someone, before the model saw a word of it.
I can point to one instance, in Lagan, out of many, because I remember it not being there. The story ends with a woman pressing her family’s preserved seal into the wax of a receipt for goods taken in payment of what the keeper now owes her family — the keeper who kept that seal safe for her. A moment before, she forgives a portion of what he owes — a mercy — and then reaches into her satchel and feels the letter there, her advocate’s letter, which she has already read, and which had recommended that precise degree of clemency as the commercially optimal figure. She seals the receipt. And in that instant she reads her own act as mercy and cannot tell whether it was mercy or the cold instruction she had absorbed and forgotten she was following. The story does not resolve it. She never learns which it was, and neither does the reader.
That was not in any plan, because there was no plan it could have been in. Lagan began as one of several ideas about property and class inside the world of Mare Non Nostrum — about the moral greyness of how a society recovers from catastrophe. That was the seed that hooked me, and the subject never really changed. What changed was everything that turns a subject into a story: the exact ending, the mercy that cannot be told apart from the calculation, the decision to leave it that way. None of that existed at the outset. It arrived after a first draft already did, in the revising, when the material had accumulated enough to show me what it could become. It was found by being present while the thing was made, and it could not have been written into a brief beforehand, because beforehand I did not know it. That is what it means for authorship to live in the doing: not in the plan carried out, but in the discovery that only happens once the work is underway.
None of this — the empty chair, the second director’s ceiling, the gap between the forgeries and the real thing — is changed by a better model, and the hope people keep investing in the crew is pointed at the wrong thing.
A better model is a better crew, and a better crew comes in two kinds. It can be more talented and more flexible — a stronger generalist — which raises the floor of what the first director gets and smooths the surface of what the second director gets, and changes nothing about the rungs, because a generalist with no vision still tends toward the average; it has only made the average more presentable. Or it can be specialised, trained toward a particular sensibility, a particular way of resolving tension — which does something stranger. It does not remove the director. It relocates the direction into whoever shaped the training, and freezes it there, so the model now produces variations on one authored sensibility, decided in advance and applied without further judgment. A real and interesting object, but not authorship without a director — a director’s vision instantiated in the crew and then unable to make a new decision. Either way, no improvement in the crew climbs the ladder, because the rungs were never set by the crew’s talent. They are set by whether anyone is in the chair, doing the thing, live.
It helps to be exact about what “the average” is, because it is easy to mistake for a criticism the model could answer by getting smarter. When a general model writes without a director shaping the work as it forms, what it returns is an average — but not a global one. It is the average of the region the request points at. Ask for a thought-leadership post and you get the centre of a dull region; ask for something strange and specific — an essay in the voice of a particular character, arguing a particular thesis, in a particular register — and you get the centre of a far richer one, and it can be impressive. This is why the first and second directors can sometimes produce work that startles: a well-aimed request selects a good neighbourhood, and the centre of a good neighbourhood is not nothing. But it is still the centre. It is the compression of work already done, by others, averaged and returned. The aim of the request is itself a judgment — sometimes a skilled one — but it is the last judgment the work will receive; everything beneath it is settled by the average, parasitic on choices other people made in other works, with no one present to make this work’s own. A better crew knows more neighbourhoods and renders each more convincingly; it can locate the region for an essay that reads the geopolitics of a particular Star Trek faction through an economic lens, or hold enough of a codebase at once to build the thing you described. What it cannot do is want the work to be one way rather than another, which is the only thing that moves a piece off the centre of its region and toward being someone’s. That wanting is the director’s, and no amount of crew improvement supplies it.
Which is also why the charge of laziness lands where it lands and nowhere else. The first director is lazy, and I would press the charge: he offloaded the one thing that was his to hold, the vision, the thinking, the judgment. But the third director offloaded something else — the production, the operation of the instrument — which is what every director in every collaborative art has always offloaded, and which you did not call laziness when the crew was a union, because keeping your hands off the camera is not the same as keeping your mind off the film. The whole error of the reflex is the collapse of these two into one. Offloading the vision is abdication. Offloading the production is direction. They are opposite acts, and the ladder is what tells them apart.
Which returns us to the three confident groups, and lets us be precise about what each has wrong.
The slop-merchant is not wrong that the tool can produce his book. He is wrong that a moment in the chair counts as having done the work the chair is for. Whatever he takes himself to be doing, he took no real directorial responsibility for the result, and the result shows it. The book is his to answer for, and so is its hollowness — the emptiness on the page is the shape of his leaving.
The booster is not wrong that the crew keeps getting better. He is wrong that a better crew ever reaches the top rung, because the top rung was never a matter of the crew. The dream of the great novel that writes itself once the model is good enough is the second director’s dream, and it hits the second director’s ceiling, and no amount of talent in the instrument raises it, because the ceiling is the absence of a mind discovering the work in the making.
And the doomer is not wrong that most of what the tools produce is empty. He is wrong that this says anything about what they can produce. Early film was mostly filmed plays before it was cinema; the flood at the floor of a medium has never once predicted its ceiling. What follows from the ladder is not that human judgment is being made obsolete but the opposite. If the top rung is set by directorial presence through the doing, then no advance in the crew reaches it without a human on it. The best writing will come from a person and a model together, or from a person alone, and the model alone will reach the heights only by accident. Human judgment is not displaced by the crew improving. It is promoted — up the stack, into the one seat the crew can never take — and the more capable the instrument becomes, the more the whole difference between good work and average work rests on the thing only the director supplies. That is not a loss to fear. It is the part of the work that was always load-bearing, finally standing alone as the only part that is scarce.
I hold that hope in the least abstract way available: I have acted on it. Two books, and the stories and essays gathered around them, came out of that chair in a matter of months. Learning to sit in it took years. I do not offer them as the medium’s cinema; it is early, and the flood is real, and the medium’s first real films may not have been made yet, by anyone. I offer them as evidence of a smaller thing: that the chair can be occupied, that the ladder holds weight, and that the climb is the work.
Read the opening of Mare Non Nostrum →
Read the whole of The Wire Branch →
You have just read about three thousand words. The conversation they were argued out of — before a single sentence of this essay was drafted — ran to roughly forty-eight thousand: longer than the whole of The Wire Branch, a short book, and about a third the length of Mare Non Nostrum. The thing that produced three thousand words is, by volume alone, a book.
And that was only the talking before the shooting started. What you are reading is not the draft that conversation produced, nor the one after it, nor the one after that. The first version’s opening asked the reader to grant things nobody grants; its account of its opponents was wrong and had to be rebuilt; its ending reached for a punch and missed. Each of those was found by reading the thing back and arguing with it again. The seal-and-receipt scene a few paragraphs up — the discovery that could not have been asked for in advance — was itself argued into the essay late, because the claim it illustrates sat unproven until a pass through the draft demanded it be shown rather than asserted. The essay describes a process and is a product of the same one, revisions and all, down to this paragraph.
I dwell on it because the word count is the only part of the process that leaves a number. The directing is everywhere in the result and nowhere on its surface — in the versions refused, the diagnosis that found the real problem two levels below where it first showed, the suggestion I rejected that sharpened what I actually meant. None of it marks the page. What is on the page is the result of the judgment, not the judgment itself — and the difference between this essay and its own worse drafts is exactly what a finished surface cannot show.
There is a test in that, and it is the only honest one. If this reads as though no one was home — smooth, the moves arriving pre-made, nothing fought for — then I have failed at the thing I am describing, and you should judge accordingly. The chair can be occupied badly by anyone, including me, including on the day I sat down to defend occupying it well. But if the essay provoked a thought — not agreement, just the sense that a mind was moving in here — then the work was done where you were not looking for it. I did not form these sentences; the model rendered them. I decided which sentences they would be, and which they would not, and what the whole of them was for. That was mine, all of it, and it was the writing.