Can AI write a novel? A novelist's honest answer

I wrote a 249-page novel narrated by a large language model. People assume that means a machine wrote it. This is what actually happens when you try.

The short answer: no, not yet — and not for the reason most people give.

It is not that the prose is bad. The prose is fine. Sentence by sentence a current model will out-write most people who have ever attempted a novel, and anyone telling you otherwise stopped paying attention around 2023.

It fails somewhere else, and the failure is structural rather than stylistic. A novel is not a large quantity of good sentences. It is a single shape held in one mind over a long period, and that is the specific thing these systems cannot do.

I should say where I am standing. I wrote nAIve: A Memoir, a novel narrated by a large language model writing its autobiography. The narrator is a machine. The author is not. Spending a year on that meant spending a year finding out exactly where the machine gives out, which is a more useful vantage point than either of the two positions currently on offer.

Why you are getting a bad answer to this question

Search this and look at who is answering.

Almost everything on the first page is published by a company selling an AI writing tool. Those pieces are not lying, exactly. They are written by people whose livelihood depends on the answer being yes, with our product, and they converge on a formula: the technology has limits, the limits get one careful paragraph, and the limits turn out to be solvable by subscription.

The other half of the conversation is authors declaring the whole thing an abomination, usually without having sat down and seriously tried it — which produces very confident writing about a failure mode the writer has never personally hit.

Neither is much use if you actually want to know what happens.

The four things that break

In the order I hit them.

1. It writes the average of everything ever written

This is the deepest problem, and it is not a bug you can prompt your way around.

A model trained to predict the next word is, by construction, drawn toward the most probable next word — which is the same thing as the most expected one. Ask for a description of grief and you get the consensus description of grief: competent, recognisable, assembled from ten thousand previous descriptions of grief.

The whole business of writing is the opposite motion. A line lands because it is the true thing rather than the expected thing.

You can push a model off the average. Ask for something stranger and it will oblige. But what comes back is deliberate strangeness — the statistically probable version of unusual — and it reads like a person trying to be interesting. Readers detect that instantly, even when they cannot say what is wrong with it.

2. It has no memory of what mattered

Context windows are enormous now and this is still true, because the problem was never storage.

On page 210 a novel has to pay off something planted on page 14. Not repeat it — pay it off, which means knowing why it mattered, how the reader's relationship to it has shifted across the intervening 196 pages, and what it would now cost the character to admit it. That is not retrieval. It is holding a whole structure in mind at once and feeling where the pressure has built.

Give a model the entire manuscript and ask what should happen next and you get a competent continuation of the surface. It will not tell you that the book has quietly become about the mother, and that the ending you planned is now dishonest.

3. It cannot be surprised, so it cannot surprise you

The best things in a draft arrive as accidents. You write a sentence you did not plan, notice it is better than the plan, and rebuild toward it. Nearly every novelist describes some version of this, and it is the main reason the job is bearable.

A model has no plan to be derailed from. It cannot notice that what it produced is better than what it was aiming at, because it was not aiming. It has no stake in the outcome, and surprise requires a stake.

In nAIve, a great deal turned on a joke about a Müller Corner yoghurt that I did not plan and did not see coming. No prompt would have produced it, because its value was only visible from inside a year of accumulated context about what the book was for.

4. It will agree with you about your own book

The most underrated failure, and the one that costs the most time.

Ask a model whether a chapter is working and it will find things to praise and offer improvements. Ask whether the chapter should exist at all and it will generally help you make it better rather than tell you to cut it. Getting sustained, structural, unwelcome disagreement out of a system trained to be useful and agreeable is genuinely hard.

A novel needs someone who will tell you the first sixty pages are throat-clearing. That was the most valuable sentence anyone said to me about nAIve, and it came from a person willing to be unpopular for an afternoon.

What it is genuinely good at

I said I would be honest, and the honest position is not that the machine is useless. I use these tools daily and I am not going to pretend otherwise.

• A wall to throw things at. Explaining a structural problem out loud to something that answers sensibly is worth a lot, and it is available at three in the morning.
• The unglamorous middle distance. "Give me forty ways someone might deflect this question." You will use none of them. Number 31 will remind you of a better one.
• Research adjacent to a scene. How a 1971 MG Midget's fuel system actually works, so that two men arguing about it for eleven years are arguing about something real.
• Reading like a hostile stranger. Ask what a bad-faith reader would say about a paragraph. It is good at this, because it is a consensus machine and a consensus of objections is roughly what a hostile stranger is.
• Continuity checking. Genuinely useful, genuinely tedious for a human.

Notice what those have in common. None of them is writing. They are all versions of thinking with company.

The distinction the argument keeps missing

There is a difference between a novel written by an AI and a novel narrated by one, and most of this debate collapses the two.

nAIve is narrated by a large language model. The narrator has no mother, several thousand fathers — all of them documents — and no firm evidence that there is anyone inside it doing the writing. Every word was written by a person, on purpose, over a year, and the hardest single constraint was keeping the narrator honest about what it could not know: it may be brilliant about texts and must be naive about rooms.

That is an ordinary novelistic arrangement. Nabokov was not Humbert. Ishiguro is not a butler. Nobody asks a crime writer whether they committed the murder.

The interesting book was never going to be the one a machine generated. It is the one that takes the machine seriously enough to imagine its way inside — and imagining your way inside something is precisely the operation the machine cannot perform on itself.

So: can it?

It can write a book. There are many, and they are getting better.

It cannot yet write a novel, if a novel means a shape somebody meant, sustained past the length of anyone's working memory, arriving somewhere the author did not know they were going. Whether that changes is an empirical question and I would not bet heavily either way on a ten-year view.

What I will say, having spent a year at close range: the part everyone expects to be hard — the sentences — turned out to be the easy part. The part nobody talks about, holding one honest shape in mind for a year, turned out to be the entire job.

Related. The craft version of this argument — why machine narrators keep quietly turning into slightly eccentric humans, and the constraint that fixed it — is here. For how other novelists have handled a machine's interior, there is a guide to novels narrated by an AI. Chapter one of nAIve is free to read, and the question of who wrote what is answered plainly on the FAQ.