This essay argues that the writer working with AI is not facing a collaborator, a master, or a threat. The relationship is closer to that of a tired writer and a thoughtful bartender in an otherwise empty bar. A companion essay, The Cathedral and the Word Palace, traces why the named-author arrangement is contracting. This piece takes up the working discipline that survives the contraction.
· · ·
The volume problem is worth investigating before moving forward. Four million books published in one country in one year is not a sign of a thriving literary culture. The number as large as it is doesn’t take into account the likely volume of millions of other books that are partially or wholly written but not published. This vast reservoir of writing is a sign of an industry where entry costs have collapsed and where quality controls have not been replaced. Amazon caps writers at three book uploads per day, which is a fact worth letting sink in. Three a day. Per writer. The cap exists because some writers (we don’t know the precise number) were uploading more than three books a day. Most of the uploads are extrusion, a fancy word to describe a manufacturing process where material is pushed through a die.
Visualize toothpaste squeezed out of a tube. When you squeeze it, the shape is determined by the opening. Now visualize books manufactured by the same process. Amazon decided to limit the number of squeezes per day per author, otherwise their system was running the risk that their warehouses would be filled with toothpaste-like books. Many of these manufactured books are now produced with the help of large language models. This process stokes what the polemicists most fear: a faster way to fill inventory that no one will read. This is the dark version of the present moment. It is not a hypothetical and it is not a strawman. It is what most uses of these tools by most writers most of the time will look like, and any honest essay about AI and writing has to begin by saying so.
The volume problem is not new in kind. Earlier versions of this publishing treadmill exist. The pulp magazines of the 1920s and 1930s ran the same engine at a smaller scale. The self-publishing wave of the 2010s, before any machine could write a paragraph, was already producing inventory at a rate the readership could not absorb. The arrival of AI accelerates the curve without changing its direction. What is new, and what the polemicists are right to notice, is that the tools carry the burden of what used to be the writer’s labour. The book a writer used to produce in a year can be produced in a week. The production assembly line accelerates.
The book that took a week can be produced in an afternoon. The afternoon book is mostly worthless, in the sense that no one will read it, no one will remember it, and the writer who produced it will not have learned anything from the production. But the presence of afternoon books is real. They exist in a quasi-zombie state. Not really alive, not entirely dead. The book enters the inventory vault and like Schrödinger’s Cat is neither alive nor dead. It thickens the pile through which any other book has to be visible. That is a real cost and it is paid by every writer working seriously, whether or not she uses the tools herself. The lesson of this glut returns us to the toothpaste metaphor: how many times in one day do you brush your teeth.
The older mentorship arrangement could not have absorbed this volume. It was never built to support large numbers. Barney Rosset took on a small number of writers in his lifetime. Stirling Silliphant taught a small number of younger writers. The mentor relationships I had were rare in their own time and would be rarer now. The arithmetic does not work. Four million writers cannot be mentored by the small number of available mentors in the old style, and most of the four million are not writing the kind of book that a serious mentor would take on anyway. The honest version of the mentorship argument is that the old arrangement served a tiny fraction of the writers who benefited greatly from it, and the larger fraction of writers made do without. What is changing is the size of the larger fraction. The mentorship gap is not new. The gap is widening, and the writers inside it are now reaching for what is at hand.
Fifty-four years ago John Berger taught a generation how to see. The BBC series Ways of Seeing aired in 1972 and the book of the same name followed shortly after. Berger’s argument was that the act of looking at art is not innocent. The viewer brings a history, a class position, a set of habits, and an inherited apparatus of interpretation, and the work of seeing well requires attending to all of these at once. The argument extended past painting. Berger was teaching how to read images in a culture saturated with them, how to notice what the image was doing and what it was hiding, how to recognise the difference between a picture that opened the world and a picture that closed it. He taught a generation of artists, critics, and ordinary viewers how to look without being captured by what they were looking at. The lesson is now fifty-four years old and is more relevant than when he taught it, because the volume of images and texts the modern viewer is asked to process has grown by orders of magnitude, and the apparatus of interpretation has not.
What Berger gave the generation that watched him is what the next generation of writers will need to develop in their relationship to the new tools. The capacity to look at what the tool produces and see what it is doing, and what it is hiding, and what it is offering, and what it is taking away. The capacity to distinguish a sentence that opens the world from a sentence that closes it, when both are technically competent and both arrived in the same response. The sharpened skill to recognise when the tool is working with you and when the tool is working against you, when the assistance is strengthening an existing scene or character or identifying a plot gap, and when the assistance is a call to fill in a blank page from a three-sentence prompt. One is a substitution; the other is revision. None of this is taught in the current curriculum because the current curriculum does not yet know the tools exist in the form they exist. A reinvented mentorship would teach it. The writers who learn to see what the tool is doing are the ones who will use the tool well. The writers who do not learn will produce afternoon books and wonder why no one reads them.
The difference goes deeper than seeing. We do not only have different ways of seeing. We have a different way of constructing what we see, and the AI does its constructing under constraints we cannot see from inside our own.
The AI’s sentence and mine come out of different machines. I work from a continuous body. I work from a memory of the last forty years of reading and the last forty minutes of writing, and both are present in the same act. The AI works from a model trained on a corpus, called fresh for each turn, with no body and no continuous reading life. The output can resemble mine closely enough that the resemblance is the trap. The trap is parity. The illusion that because the sentences look comparable, the processes that made them are comparable, and the kinds of intelligence behind them are comparable.
They are not. The illusion is hard to dispel because the evidence on the page argues for parity. A competent paragraph is a competent paragraph. The reader cannot tell from the page which architecture produced it. What the reader can tell, sometimes, is when the paragraph has been produced by a process that does not know what it does not know. That is the architectural fingerprint. The writer working with the AI needs to recognise her own architecture as different from the tool’s, and to use the tool with that difference in view.
Go back to the letter. When I write to a friend, I am writing to a memory of her. I see her face. I hear the cadence she uses when she disagrees with me. I know which jokes will land and which will not. The words I choose are tuned to a person who exists in my mind as more than a name on the page. Forty years of conversations sit behind every line. The letter is not the friendship. The letter is the form the friendship takes when we are apart.
The AI has no equivalent of that memory. Each session begins with a wiped slate. The conversation we had yesterday is gone unless I bring it back into the room with me. Whatever the AI learned about my voice in last week’s draft is unavailable to this week’s draft. Oliver Sacks wrote about humans whose memory worked this way. The Man Who Mistook His Wife for a Hat described patients living in a perpetual present, unable to consolidate what they encountered into a self that extended forward. Sacks treated these cases as catastrophic. The patients had lost something essential to being a person.
The asymmetry between the writer and the tool is illustrated by the asymmetry between Sacks’s patients and their visitors, though the cases differ. Sacks’s patients had a damaged architecture inside a self that had once been whole. The AI has no self to damage. The frame is useful for the asymmetry it shows, not because the two cases are the same. The visitor remembers. The patient does not. The visitor carries the relationship. The patient cannot. The writer using the AI must hold the continuity for both of them, because the tool cannot hold it for itself. Your memory carries the micro-details that make the work personal. The tool’s memory does not exist between sessions. That is the deepest difference, and the most useful one to keep in view.
Hold that asymmetry in view and the question shifts. Not whether the tool resembles a partner, but what kind of room the tool occupies when the writer brings the continuity it cannot supply. There is a room being built, and we do not know yet what it is. The available words for it (workshop, agora, studio, atelier, salon, scriptorium) all import a history that does not quite apply. The relationship inside the room is not the relationship of a master to an apprentice, because the tool is not a master. It is not the relationship of an editor to a writer, because the tool has no editorial commitments. It is not the relationship of a reader to a text, because the tool participates in the writing rather than receiving it.
The closest description I can offer is that it resembles, on a good night, the conversation between a tired writer and a thoughtful bartender in an otherwise empty bar. The bartender has no stake in what the writer is working on. The bartender has read more than the writer expects and remembers less. The bartender will ask questions the writer was not yet ready to ask herself, and will sometimes, by accident, open a creative window the writer had not seen was there. The conversation is not the writing. It is the room the writing is done in. Whether it survives the morning, whether anyone is paying for the drinks, whether the bar will be there next week, are all questions for which no one has the answer yet.
The writer who holds the architectural difference in view is the writer who can use the tool without being used by it. The two halves of the discipline run together. Forget the difference and you start treating the AI’s sentence as if it carried the same authority as yours. Forget the use and you refuse a tool that can do work no other tool has done before. The writers who survive the present moment with their work intact will be the ones who hold both halves at once.
The London Underground says mind the gap when the train pulls in. The gap between platform and carriage is small. Step into it and you break an ankle. The same instruction works for prose. The writer’s job is to find the gaps before the reader does. Contradictions inside a character’s motivation. A repeated word three pages apart that the writer did not notice. A line of dialogue that the speaker would not say at that point in the story. A claim in paragraph six that paragraph two has already undercut. The brain that built the manuscript cannot always see the gaps it left, because the brain remembers what it meant and reads what it meant onto the page even when the page does not carry it.
The AI is good at this work. Run a draft through and ask it to find inconsistencies, repetitions, motivational holes, voice slips. It will find some. It will miss some. It will flag some that are not gaps but deliberate silences the writer left in for reasons the AI cannot see. The writer keeps the judgment. The AI does the patient sweep that the writer’s tired eye no longer performs.
The red team is the harder use. Hand the draft to the AI and ask it to attack the argument. Find the counter that the writer has not addressed. Find the line of meaning that contradicts the thesis. Find the reader who will close the book at page sixty because the writer ducked the question on page fifty-nine. The attacks the AI returns will not all be right. Some will be persuasively wrong, which is the harder kind of wrong because the wrongness buries itself in plausible sentences. The writer must red-team the AI as well, vetting every challenge it raises before deciding which to act on. Follow your own judgement. The AI is not an oracle. Trust but verify is the working phrase, and it applies to the AI the way it applied to the human counterparties that taught us to use it.
Calibration came out of coding. A programmer asks the AI to evaluate code against criteria, and the AI returns a judgment with reasons. The same function works for prose. The bartender of the previous section is the room. Calibration is one of the functions that happens inside it. The role the AI plays during calibration is closer to a senior editor than to a critic. The critic stands outside the work and assesses. The senior editor sits inside the work and asks whether each move earns its place. The senior editor does not deliver verdicts. The senior editor delivers questions and observations the writer can act on or decline.
Ask the AI to read the draft as a senior editor would. Does this paragraph carry its weight. Does this character motivation hold across the chapter. Does this image earn the page it occupies. The answers will not always be right. The writer has the final reading. But the writer who reads her own draft sees what she meant. The senior editor reading the same draft sees what is there. Calibration is the function that closes that gap.
This is not the same as the red team. The red team attacks. The senior editor stands beside the writer. Both are useful. Both are different. The writer who learns to use them separately gets a tool that no previous generation of writers had, and a tool that adds to the work she would have made alone.
What a tool of this kind cannot do is the part the polemicists are right to defend. It is useful to set out the constraints. Your AI cannot know what your life has cost you. It cannot read your sentence against the standard of its own life because it has no life as a standard. It will produce a competent paragraph unaware of what you need is the sentence that hurts, and it will not know the difference unless you tell it, and even then it can only approximate. It will not pursue an obsession across a decade. It will not, of itself, refuse the easy turn. If your aim is to fill tubes of toothpaste, it won’t question your goal, and will work alongside to manufacture a warehouse of toothpaste tubes. All of those refusals and limitations remain the work that a writer still has to do if he is going to be craftsman rather than a manufacturer.
The AI, as a tool, extends the writer’s reach as a painter of words. If you want to paint by numbers, it will help to accomplish that goal. If you want a handcrafted book that comes from your experience, reading, life and observation, it can help like a mentor at your shoulder. The tool does not replace the writer’s judgment. That is manufacturing, not writing. The person using AI to bypass judgment is just another blood bag working a keyboard. There is a boundary between the working room and the assembly line. You find the demarcation line is governed by the writer’s discipline, motivation, and creative talent and is found nowhere else.
This is the place where an essay that wanted to be honest about its own conditions would say what it had been made with. I am going to leave that question open, because the essay’s argument does not depend on resolving it. The reader who has been paying attention has already formed a view. The reader who has not formed a view is welcome to read the essay as straightforward cultural argument, and the argument stands either way. What I will say is that the new room exists, that the conversation inside it is possible, and that an argument in the service of the unworkable is a sign of desperation. The unworkable arrangement is not the named author. The named author will continue.
The unworkable arrangement is the twentieth-century apparatus that promised a livelihood to a much larger number of named authors than the readership could ever have supported, and the writers most invested in the apparatus are the ones most likely to spend the next decade arguing for its preservation. The arguments will not save it. Nothing will save it. Letting go of what cannot be saved is the hallmark of the great artist, and the writers who can do that letting go are the ones who will still be writing in twenty years, on terms that do not yet have names.
The cathedral is closing. The masons who finished its last bays will go on calling themselves masons and will be right to do so, because the work they did was real and the cathedral they built will stand for as long as anything stands. But no one is starting a new cathedral. The next buildings will be smaller, quieter, harder to find from a distance, and made by hands that work in rooms whose names have not been agreed yet. Some of those hands will be helped by tools that did not exist when the cathedral was rising. The help will be real and the limits will be real and the work will be different work. A handful of the buildings will last. Most will not. No one makes a fifty-year commitment to construct a hand-made cathedral. This was the age when projects might take multiple generations to complete. It gave continuity, predictability and purpose. Without those three elements, we very quickly become unanchored. Lost. Vertigo. The case is strong that all three are no longer stable.
The unease felt by writers is real. It’s not going to go away. If anything, the discontinuity, unpredictability and lack of purpose will grow. That is the deal: don’t chase the AI to manufacture your book. Write it yourself. Remind yourself that the book manufacturers are seeking an easy way out of the hard work.
Readers know this.
They don’t buy or read such books. Believe in yourself. Your way of seeing can’t be replicated by a machine or tool. It belongs to you. You’ll be fulfilled by the process of writing where you reveal yourself to yourself. No machine can do that for you. Promise and deliver your way of seeing. The deal has not changed because the tools have. The writers who understand this will be at peace with the change before the polemicists are, and the work they make in the meantime will be the work that survives the burning of the rest.
I enjoyed both essays. I also recognise the grief you describe.
I felt something similar when digital photography largely displaced film. It wasn't simply the loss of a technology. It felt like the loss of a craft, a way of working, and a professional identity. Looking back, film never disappeared, but the dominant arrangement changed.
Where I would extend your argument is that I think this process began long before AI.
AI is not the first technology to transform writing. It is the latest in a long sequence of innovations that progressively reduced the cost and skill required to produce and distribute written work.
Phototypesetting replaced much of the traditional compositing trade. Desktop publishing put professional page layout onto ordinary computers. Word processors accelerated drafting and revision. The internet removed many distribution barriers. Self-publishing platforms allowed authors to bypass traditional publishers. Print-on-demand removed the need for large print runs and warehousing, making it economically viable to publish books that would previously never have reached the market. Each innovation removed a different bottleneck in the publishing process.
None of these developments destroyed literature. What they collectively did was remove successive bottlenecks. Each innovation lowered the cost of producing and distributing text, increased the volume of material available to readers, and lowered the barriers to publication. The result was not necessarily lower-quality literature, but a much larger proportion of mediocre writing competing alongside excellent work for the same finite attention.
AI is therefore better understood as another step in a much longer structural trend rather than an entirely new disruption.
One place where I remain slightly unconvinced is the economic framing. Historically, publishing has always produced a highly skewed distribution of rewards. A handful of authors earned extraordinary incomes, but most earned relatively little. The arrangement may therefore have promised the possibility of a livelihood more than a livelihood itself.
What particularly resonated with me in your second essay was the distinction between using AI as a substitute for judgement and using it to strengthen judgement. Using it as a red team, an editor, or a calibration tool feels fundamentally different from using it to manufacture books. That strikes me as one of the most valuable distinctions in the current discussion, because the presence of AI tells us remarkably little on its own. What matters is how it has been used.
Reading both essays, I found myself thinking that AI seems to be reducing the cost of producing information far more rapidly than it reduces the cost of producing judgement. If anything, judgement becomes more valuable as information becomes more abundant.
To me, that is the deeper structural change. Every technological innovation has increased the supply of written material, but none has increased the number of hours people have available to read. The scarce resource has never really been the ability to produce words. It has always been human attention. As the cost of producing words approaches zero, the value of producing sound judgement may only increase.