On Artist-Driven AI

by Scott Henderson, The Name in the Stone

The question

I live in an old limestone house on King Street in Kingston. Cut into its exterior stone is a single word — Alfred — that I have never been able to trace to anyone who lived there. No record I have found accounts for it. The question of who he was would not leave me alone. Over a long while it became a novella, The Name in the Stone: the invented 1875 diary of a man named Henry Collier, set in a Kingston of real streets and real stone. Its diarist watches my house from the bank windows across the street.

Writing was the skill my whole working life was built on. But I had never considered writing a book. The book began as an experiment with a new technology, and it took a year. Somewhere in the first months it stopped being an experiment and became an odd collaboration, and months later that collaboration became a book. The further I went, the more I realized it had become something substantial: a thing others might enjoy, and a way for readers to come to love Kingston the way I do. I made it with artificial intelligence, and I hope to explain what that means.

This essay is my position on what I did, and the story of how I came to hold it, because the position did not arrive ready-made. As the project advanced, I went searching for understanding about where I had found myself: reading objections to AI at their full strength, following the discourse in the artist community, looking back at how artists have used assistance in the past, and at how technology has collided with art-making over and over again. What follows is that search, in roughly the order I made it, and where it has landed so far.

I could make my life easier by saying The Name in the Stone is “AI-assisted.” It’s a common phrase, and it sounds modest and safe. But it isn’t quite true in my case. “Assisted” implies that I wrote the sentences and a machine tidied them afterward. That’s not what happened, and calling what happened light help would be a lie. Saying the AI wrote the book would be a lie too, and the easier one to tell, because then none of its failures would be mine. It would also leave out everything I did, which was to direct the drafting, every line of it. The machines — including ChatGPT and Claude — produced the prose I asked for. Then I went through it, sentence by sentence, until it was right, and where a line would not come right, I wrote it myself.

Here is one sentence, and how it got there. Major Irwin is the officer who questions Henry on the Bank steps. The machine’s line for him was simply “Did she say this to anyone else?” — a question anyone might ask, in any decade. As part of my historical research for the book, I went looking for the real man. The commandant of the School of Gunnery at Kingston in 1875 was De la Cherois Thomas Irwin, born in County Armagh, Ireland. So I wrote in the margin: “Irwin is irish, but upper class irish.” After several more back-and-forths we arrived at the line in the book: “Tell me, is it to anyone else she said this?” A change of a few words, but a truer picture of how Irwin would speak. I did not type the revised dialogue. But the words that got it there were mine, and it is there because I found out that a soldier in Kingston in 1875 had been born in Armagh.

I wasn’t always comfortable with any of this. Early on I felt embarrassed, even guilty. The doubt was that I was somehow cheating, that I wasn’t actually doing the work that others do. And the technology itself frightened me, considering the artist community’s valid concerns about AI. So instead of running away from it, I decided to embrace it and see where it went. I also decided to be as transparent about it as I could.

When people ask whether I wrote the book or whether the AI did, they are folding together two questions that deserve separate answers. The first is mechanical: whose hand produced the words on the page? For The Name in the Stone that answer is easy, and I have already given it: the machine drafted the sentences, to my direction. The second question is older. Whose vision is it? Who decided? Who answers for what it finally says? I knew what I wanted the answer to be, but I did not yet know whether I was entitled to give it, or even what to call the thing I had done.

When I began my search for understanding, I couldn’t answer that. What I had was an intention. There is a version of AI technology that sets out to replace the artist, and a version that is an instrument in the artist’s hands. I was interested only in the second. Put more bluntly, a book made with AI is either directed or dumped: shaped by a person who answers for every line, or sent into the world as it fell. Mine would be directed.

Where I went looking

The search began with the objections, at their full strength. The first came from the columnist Bret Stephens, who not long ago, in The New York Times, asked his readers never to write with a machine at all. He wasn’t arguing that it was cheating, and he allowed that before long nobody would outwrite the machine. His worry was that writing is how a person learns to think, and that handing the sentences to a machine lets that faculty go slack — an escalator where the mind is better served by the stairs. His worry is aimed squarely at what I did, because the sentences in The Name in the Stone did not come from my hand.

But I can tell you what the work asked of my own head, and it was not less. I wrote the rules of Henry’s voice down and argued them with myself. One of them read: would a bank clerk in 1875 write this sentence in a private diary? If not, rewrite it. Another: he thinks in ledgers, columns, debts and balances, and his metaphors come from that world. A third: when he is at his worst, his prose gets plainer, not more artful. Harder than any of them was holding the line on what a man of his position and his year could think. The machine kept offering me a simpler Henry, a more modern one, who would say plainly what a clerk of 1875 would more likely have kept to himself. No rule about prose could keep that Henry out; that took a judgment about a particular person in a particular time.

There is something in writing the rules down that is worth naming, because it cuts both ways. A novelist working alone never has to say what the voice is. She can write in a voice without having to explain it. A machine cannot infer any of that. So everything I would ordinarily have held silently had to be put into words, and in a form specific enough that a line could be judged against it. I’ve never had to state my own taste before. I’m still not sure whether writing it down made it sharper or thinner. But it was work, and not the faculty going slack. I had the judgment before the tool existed, and every line the machine offered had to pass it.

The next place I looked was the past. The fear Stephens raised, that a faculty is being lost, is an old one. Painting was supposed to die when photography arrived, and practical effects when CGI did. Neither did. But the nearest precedent I found runs deeper than paint or film. It is the act of writing itself. Some 24 centuries ago, Plato wrote a dialogue in which Socrates warns that writing would weaken memory and give people the appearance of wisdom without the substance. We know this because Plato wrote it.

Gutenberg’s press was 40 years old and printed books were everywhere when a German Benedictine abbot named Johannes Trithemius wrote a tract defending the hand-copying of books, and the discipline it demanded of a man, against the printed page. He had it printed, and once again the case against the new tool survives only because the new tool carried it. When the novel took hold in the 18th century, it was treated as a contagion — a “reading fever” accused of filling the young with false ideas and idling the hours of women — with Goethe’s The Sorrows of Young Werther (1774) being blamed for a wave of imitation suicides.

Each time, a new way of making or meeting words was accused of weakening something held to be essentially human: memory, then discipline, then virtue. Now it is composition. Three old panics came to nothing, and that does not make the current one foolish; new tools can weaken old skills even as they open others. But in none of these cases did the threatened faculty disappear. Books drafted with AI will sit on the shelf beside books written without it (they already do), and the choice will land where it has always landed, with the reader, who decides what they prefer.

I don’t claim to know where AI will end up, and I haven’t met anyone who does. But the look back at history settled this much for me: a faculty survives a new tool, changed but not gone. What history could not tell me was what to call a book made the way mine was, or whether the person who made it gets to call it his own.

So I went looking for someone doing something like what I had done. The comparison people reach for first is the film director, who makes the picture without ever touching the camera. It’s a fairer comparison than most people realize, because there’s a long tradition of treating the director, not the screenwriter, as the true author of a film. Think of the sci-fi film Arrival: it is called a Denis Villeneuve film, but Eric Heisserer wrote the screenplay, adapting a story by Ted Chiang, and was nominated for an Academy Award for it. Most people who loved that film could not tell you either of their names.

In my mind, I had two of the three primary roles on this book: director, commanding the performance I wanted, and editor, deciding from the raw takes what the book would finally be. The third role, the writer’s, was not mine. A reasonable person could look at that and say I directed this book rather than wrote it, and I have no quarrel with that description. What the description can’t do is settle whether the book is mine, because a book is not a film. In a film the directing is there on the screen for anyone to see. On a page, all a reader can see is the sentences.

So if a machine made the sentences first, what exactly is left that is mine? It was the choosing, refusing, and keeping them that was the making of the book, not a polish after it. Those choices were my own, and they are what made it this book and not some other book.

The closest thing to what I did is the example everyone knows, and the one I had reached for myself, long before this essay. A ghostwritten memoir says “by” the person whose life it is, and nobody thinks the cover is lying. The story is theirs. The ghost does the writing: takes the story down, drafts it, produces the sentences. The person on the cover reads what comes back, argues with the ghost, says what is right and what is wrong and what is missing, and sends the pages back until they say what happened. That is how this book was made. The machine was my ghost, and my collaborator.

The comparison holds all the way to the question of who the ghost is. A ghostwriter is a person, with a name and a life of her own. The machine is no one. It handed me sentences with nobody behind them, and that is where this book parts company with a ghostwritten memoir. It is also where the objections to a book made this way begin: nobody behind the words for a reader to find, sentences that cost nothing to make, and writers the machine learned from who were never asked. I will come to each.

Every comparison to what I had done stopped somewhere like that, and that is about where I quit looking for a comparison that fit. I had been trying to show that making a book this way is the same as writing one unaided, and it isn’t.

The most clarifying thing I found in all my reading was a line the art director Dan Littlewood wrote on Substack in 2025: when creation is effortless, meaning is the only thing that matters. I believe that’s true. A machine can produce a thousand plausible paragraphs in an afternoon, so in that world the sentences themselves have become cheap. What has not become cheap is deciding which of them belong in a book, and that decision depends on the one thing the machine does not have: something to say.

If meaning is what matters, then I should say where the meaning in this book came from. What the book means — who Alfred was, what Henry could not bring himself to say — I brought to it, and the machine never supplied it. What it supplied was prose. Some of it I had not asked for and could not have described in advance — a turn of phrase, a way into a scene — and some of that was right, and I kept it. Recognizing the right one is real thinking, taste, even.

What I actually did

If no comparison would tell me what this was, the year of making this book might. So I went back to what I had actually done. I learned the machine can generate a sentence, but it cannot decide whether that sentence belongs in a book, or feel that a chapter has gone soft in the middle, or that a picture two pages earlier has quietly changed what a later one means. Those judgments were mine, and I made thousands of them, and every one of them was measured against a world the machine had no way of knowing. Before it offered me a line, there was already a Kingston in my head — a bank, a house, a year — and a Henry Collier, who could notice only what a bank clerk of 1875 would see.

You can see what that judgment is worth by looking at books made without it. There are people generating books by the hundreds: a life story assembled overnight, a history of a bridge no one asked for, each one drafted whole by a machine and pushed into the world under a face that does not exist. The world has lately settled on a name for this — slop. I would like to say the difference between those books and mine is that someone decided what went into mine. But that difference cannot be seen from the outside. When I published, Amazon asked whether the book had been generated by AI. I answered yes. The person selling a hundred bridge histories answers yes as well. The label is the same for both, and it tells a reader nothing about what happened between the generating and the publishing. From where a reader stands, I am one of them until I show otherwise, and I can’t blame anyone for starting there.

Run my pages through a detector and it will say a machine wrote them, and it will be right. What no detector can see is what happened to the sentences afterward. Even Substack, when it handed its readers a detector to run on its writers’ posts, conceded the point: the scan can tell you a tool was used — not, in its own words, “whether great human care went into creating it.” Here is one thing no scan can see: the morning I spent deciding whether Henry would say “macadam” or simply “the road.”

I ran one test on myself, over and over: whether I could hear the machine in my own pages. Right now, it’s not hard to hear when a machine is writing on its own, and that sound is the first habit I set myself against: the tidy symmetries, the false little reversals, the even march of sentences that never change length, and, running the other way, a fussy refusal to use the same word twice, where a person would simply repeat himself.

Driving the default voice out is not a matter of striking a line and asking for another one. I tried to break the symmetries. I tried to knock the rhythm off whenever it fell into step. I put a word back where the machine had varied it for the sake of variety, and cut the ones it had reached for because they sounded like writing. I recrafted sentences until they said the thing and not the near thing, and sent them back again when the paragraph around them moved.

There is a kind of writing in all of this that has yet to be named, and it is where most of my own words went: into the margins of the drafts. Every draft came back to me and every one went back marked, and the marks are sentences — sentences about sentences, written to be read by a machine and then thrown away. Directing prose is done in prose, and a vague instruction gets a vague line back. The instruction has to say exactly what is wrong with a line and what a right one would have to do, without writing the line itself, and the words chosen to say it decide what comes back.

In March I wrote against one draft that the fractured, broken prose we were using to show Henry’s decline was not the same thing as the unpolished diary of an educated bank clerk, and that the book needed a bible to keep the two apart. The rules that governed every draft after that came out of that sentence. There is a skill in this writing, and maybe, in time, an art: to say what a sentence must do, precisely enough that the next one does it. It is certainly not the same skill as writing the sentence, but it is writing, and it is the part of this book that is all my own hand.

The machine also creates a world by default. Asked for 1875, it drifts toward Victorian-in-general: a fog and a gaslamp and a street that is quietly London. What makes this book Kingston is the limestone and the lake and the cathedral bells, as someone who lives here sees and hears them, and each of those details is a correction against that drift. The machine did not know that the courthouse was being rebuilt in 1875, or that the Assizes had to sit in City Hall while it was.

But the drift does not stop at atmosphere; it reaches into fact, and historical fiction shows that more plainly than any other kind of book could. The Name in the Stone is set in a real place with records and artifacts, and a sentence about it can be flatly false. A machine writes a false sentence in the same steady voice as a true one. This is where the difference between directed and dumped stopped being a phrase for me. A book that is dumped has no one standing between plausible language and the truth of what it claims. A person will not catch every error, but nothing inside the writing can catch any, and someone outside it can at least try.

So I treated what was coming back as questions rather than answers, and went to the archives and to the buildings themselves for what was true. And the facts were only a start, because a true Kingston is not yet a story. The record can tell me that the Bank of Montreal left its King Street premises in 1906. It cannot tell me that a clerk’s diary was walled up in an old vault and came to light when the building was cleared.

All of that is what the making was. Whether it counts for anything with a reader is another matter, and the objection that troubles me most on that score comes from a novelist. Tom Rachman wrote in The Globe and Mail this summer that the instant he discovers what he is reading was generated, he stops caring. That’s a real concern for this book. It’s not about the prose at all, and making the prose better would not answer it. It says a reader reads with one eye on the meaning and the other searching for the person behind it, and when the second eye finds a machine, it closes. “The words dim before me,” he wrote. “It’s an off-switch.”

I can only answer that with the record. A reader searching for the person behind the words is searching for someone who cared, and the choosing is where that shows: the hundreds upon hundreds of drafts, the rules I wrote for Henry’s voice, the arguments I had with the machine and with myself. Nobody does all that unless the book matters to them. The critic Spencer Kornhaber, writing in The Atlantic, put the same thought in a way I have not been able to improve on: “When media are infinite, the details poured into any individual work matter more, not less.” That care is what I have to show, and I hope knowing how this book was made does not dim the reader’s enjoyment of it.

The book is illustrated, and the pictures raise the same question, though it runs differently for them. The plates were not decoration. Part of what I wanted was to let a reader see the Kingston of 1875 — not a Victorian city in general but this particular one, with the streets they walk now as they stood 150 years ago. Prose does a great deal of that but cannot quite do all of it. The book also presents itself as a diary serialized in 1906, and plates belonged to that form. The harder question about the pictures is what they cost.

Illustration has not always come from the hand of the person whose name is on it: the plates of Henry’s own era were an illustrator’s drawing, an engraver’s block and a printer’s press, and the credited artist may never have touched the surface the reader saw. One objection, though, lands harder on the pictures than on the prose, and it deserves a direct answer. Illustration is a living trade. The usual way a book gets its plates is that a person is paid to draw them, and a machine in that chair looks like a hand put out of work. Before I ever turned to the machine, I looked for a living illustrator, in Kingston and beyond — someone working in pen and ink who could carry off the wood-engraved look of the 1870s — and I could not find one free to do it in the time and budget I had.

As for the style, I asked for an idiom and not for an artist. The plates imitate the wood-engraved illustration of the 1870s, a technique many hands learned and no one owns, and no artist’s name went into any instruction I gave. I asked for nobody’s hand, but I can’t tell you whether somebody’s is in it. The machine’s fluency in that old manner was not built only out of old work. Whatever it knows about pen and ink it learned partly from people who are alive and drawing now, and no instruction of mine could separate the one from the other.

And where a living hand could be put to this book, I put one there. The cover, the interior and the design of the object a reader holds are the work of my book designer. A book is a made object as well as a written one, and the making of it is an art. That art was not mine, and neither were the typefaces — another artistic judgment of the same order, and the designer’s. The book carries four voices across 150 years, and a reader has to feel the distance between them before a word is read. One type choice stands for all of them: the 1906 newspaperman is set in Perpetua, a face drawn from monumental lettering — letters cut into stone, in a book called The Name in the Stone. I would not have thought of it.

What none of this settles

None of what I have said about who made this book touches the things people are right to worry about. The largest of them is this: AI-generated prose was built by swallowing the written work of a civilization without consent, payment, or credit. It is now sold back to us at a profit, and the people whose work was taken are said to be undercut by it. The factual core of that is true, and my enthusiasm for the tool settles none of it. The training was extractive, the data centres behind these systems draw power and water from somewhere, the legal questions are unsettled and still being fought over as I write, and the people building these systems are themselves arguing in public about how fast is too fast.

I didn’t assemble the training data and I can’t defend how it was gathered. Being a user rather than a maker limits my share of that, but it doesn’t make me a bystander. I can’t keep my distance from it, because the sentences in this book came out of a machine that learned to write by reading authors who were never asked: novelists, historians, the writers of every kind of book, including the kind I have just made.

The cost that is hardest to dodge is the one paid by other writers. The writing that pays people’s rent is scripts and songs, ad copy, corporate communication, technical writing — the hundred kinds of commissioned work that let a person with a facility for language make a living out of it. That is the work most exposed, and the fear of it is real and it is everywhere in the trade. My feeling is that the work still needs a person to direct it, but that is an answer about the craft, and they are asking about the rent. Someone who spent a career learning to write well for hire is watching the market for that skill be threatened, and nothing I have said here helps them. I know that trade, and I worry about the ones who come after me.

I put little weight on the argument that the AI revolution is coming whether anyone likes it or not. Plenty of things arrive that way and are worth refusing. But the tool will not be uninvented. Whether a person ought to use it at all is one question, and I have answered it here only for myself. A second question sits beside that one, and it is being answered every day by default: whether the people who do use it bring anything to it. The writing is being done, in enormous quantity, and the tool makes it very easy to bring nothing to it.

Where it landed

To come back to the question I went looking to answer: the search never did turn up a precedent, but it did leave me with four words. They are ordinary words. I had to pin them down myself, because the one word everyone uses — wrote — covers all four at once. Generation is the production of sentences. Composition is deciding which sentences a book will contain. Responsibility is who answers when the book is wrong. And authorship is whose vision governs all three. The machine did the first one, and I did the other three.

“Writer” and “author” are used interchangeably in English, but they are worth looking at separately, because by the older meaning of the words, the machine can write, but it cannot author. A writer makes the sentences. An author is the one the work comes from, and the word says so: the Latin auctor means not a scribe but an originator, who vouches for the work and can be held to it.

All of this brings me back to where this essay started, at the wall of the house. The book began with a name cut into it, and a question only I was standing close enough to ask. There is one part of this that no collaborator, human or machine, could have supplied, and it is the part Littlewood’s line points at: having something to say. A model can draft a chapter on command, but it cannot ask who Alfred was, and every decision in this book runs back to that question.

A friend of mine posted a single line on Facebook: “AI is the end of imagination.” I understand what she’s afraid of. But my own experience ran the other way: I have never felt my imagination more awake. For the year I spent writing this book, I lived in 1875: the smell of the Works and the sounds of horses on the macadam, the light on the limestone, what Henry would notice from the bank windows and what he would never permit himself to say, and who a boy called Alfred might have been.

The machine did none of that imagining. It kept pace with mine. That matters, because the distance between imagining a world and getting it down is where most books die. For me, each draft the machine returned was a set of questions. Is this the world you meant? Is this what Henry would notice, and is this how he would say it? Would a man of 1875 know this word? Answering them over and over forced me to imagine the world of The Name in the Stone more precisely.

So this is where the search ended, and it is not where I expected. I started out trying to show that a book made this way was the same as any other book, and every comparison I tried told me it wasn’t. Nevertheless, I believe making a book by directing a machine to draft it is creative work. It takes invention, and judgment that never rests. It’s a different thing, and it does not have to be the same thing.

People ask why I did not write the sentences myself: why someone who has made his living by writing would give that to a machine. I had never set out to write a book. It began as an experiment with a new technology, and the experiment became a book; I never considered writing it the ordinary way, because without the machine there would have been no book. The search answered the rest: nothing about how it was made makes the book less mine. So the question turned around. Why wouldn’t I?

No machine, no book. I keep turning that over, and it is closer to wonder than to argument. There is no shortage of stories. Everyone is carrying one or two, and most of them never get set down. The people carrying them never think of themselves as someone who writes books, as I hadn’t. A boy’s name cut into a wall, noticed by the one person who lives behind it, would have stayed a name in the stone. Readers have only ever had the stories that got written. The act of writing let a story outlast the person who told it, and the printing press let it reach people that person would never meet. Now the machine lets a few more of them get told. Not all of them will be good, and none of them will be anyone’s without a person deciding what goes in. But more of them will reach a reader than would have, and that is the part I wonder at.

Authorship has taken more than one form before now — the memoir with a ghost behind it, the film that is its director’s — and now there is another, newer than those and bound to be argued over. It needed a name, and partway through the search I found that others had already reached for one. A growing number of people working with these tools in visual art and film call it artist-driven AI. I have come to understand that it describes what I did. It names a form of authorship, in which the machine writes and a person authors, and the book remains that person’s.

Which answers the question I started with, the one I wanted to answer and wasn’t sure I was allowed to. Yes, the machine produced the words on the page, to my direction. But whose vision was it? Whose decisions? Whose responsibility? They were mine.

Making anything asks a kind of daring: a willingness to put something into the world that may be bad, or wrong, or unwelcome, and to have your name on it either way. That is the ordinary condition of art, not a special virtue, and a book made this way is no exception. I may turn out to be wrong about all of it, but I don’t think I was wrong to try.

One more thing. I have said how this book was made, in the book itself and on this page, because I don’t think a position you have to hide is worth holding. So I should say that this essay was made the same way the book was: drafted by the machine to my direction and re-drafted until it said what I meant. Every first-person sentence in it — that I felt guilty at the start, that I can’t defend how the training data was gathered, that I have never felt my imagination more awake — was drafted by the machine from what I told it, and each one is true, and I answer for it the same way I answer for the book. I ran it through Pangram, the AI detector Substack uses. The scan suggests it’s 88% machine-written. My name remains on it. I am the author.

Yes, the machine produced the words on the page, to my direction. But whose vision was it? Whose decisions? Whose responsibility? They were mine.