Fiction With AI: Tools, Limits and the Craft Question
11 min read · updated August 4, 2026
A language model supplies the most probable continuation of what it has been given. A novel is an attempt to produce a sequence of sentences nobody could have predicted and everybody accepts. Those two descriptions are in direct conflict, and almost everything true about fiction and these tools follows from the conflict rather than from anybody’s opinion about art.
What the machine actually contributes
Strip away the argument and the mechanism is not in dispute. Given a context, the model returns a distribution over next tokens and something samples from it. Averaged over a paragraph, what comes back is the central tendency of everything written that resembles your context.
For most writing that is a gift. A support article, a specification, a release note: you want the expected thing, expressed the expected way, because deviation costs the reader effort for no return.
For fiction it is the problem in its entirety. The reason a novel is worth reading is that a specific person made a series of choices no average would have produced — this detail and not the obvious one, this scene ended eight lines early, this character allowed to be unpleasant without being punished for it. Central tendency is the material fiction is made against.
This is why prompting does not resolve the question. You can push the distribution around with instructions, style exemplars and sampling parameters, and you are still drawing from a distribution over what text like this usually does. The failures visible in generated dialogue are the same property at the sentence level, and they are the clearest demonstration available.
Where that property is fatal
| Task | Description |
|---|---|
| Prose in the finished book | The sentences are the work. A paragraph drawn from the centre of the distribution reads as competent and belonging to nobody, which is the one thing a novel cannot survive. |
| Dialogue | Requires withholding, mismatch and refusal to resolve. All three are low-probability moves by construction. |
| Plot turns | A turn a reader could have predicted is not a turn. Generated plot proposals converge hard on the familiar shape of the genre, which is exactly the set you are trying to leave. |
| Voice | Voice is accumulated refusal — everything a writer will not say, and the constructions they will not use. A model has no refusals of its own, only a very good impression of everyone else's. |
| Deciding what the book is about | Not a text-generation problem at all. This is the part that takes years, and it is the part nobody can outsource, including to a human collaborator. |
Where it is irrelevant, and the tool is just useful
The same mechanism that makes the model useless at the above makes it genuinely good at the surrounding work, which is substantial and which most writers do badly because it is not what they came for.
- The continuity bible. Feed it the manuscript and ask for every physical description of a character, every date mentioned, every name of a place, with chapter references. Then look for the contradictions yourself. This is search over supplied text and it is reliable in a way generation is not.
- Timeline reconstruction. “List every stated or implied interval between events in chapters one to nine” catches the winter that lasted eleven months.
- Research questions, verified normally. What a 1930s tram conductor carried, what the harvest was in a given region, what an occupation calls its tools. Treat every answer as a lead to check, not a fact — the source-first workflow applies here exactly as it does to non-fiction.
- Line-level diagnosis of your own prose. Repeated words within a page, filter phrases (“she felt”, “he noticed”), sentences with three clauses in a chapter of two-clause sentences. Reporting, not rewriting.
- Dictation cleanup. Punctuating and paragraphing a transcript of your own speech changes nothing about what was said.
- The synopsis and the query letter. These are summaries of supplied text in a conventional register — exactly the central-tendency job the model is best at, and exactly the task most novelists find hardest.
- The deliberately bad version. Generate the obvious scene so you can see what the obvious scene is, and then write the other one.
The craft objection, at full strength
The objection is usually dismissed as sentimentality about human creativity. Stated properly it is not sentimental and it is not about the reader.
Writing is how a writer finds out what they think. The draft is the instrument, not the by-product: you discover that your protagonist would not do the thing you had planned by trying to write her doing it and finding that the sentence will not go. A draft you did not write contains no such discovery, so the book you end up with is the book you could already imagine before you started. Over a career the compounding loss is not prose quality but range — you never find the things you would only have found by writing badly for six months.
The second half of the objection is about the reader and is simpler. Fiction works partly because someone meant it. A reader who learns that a passage which moved them was produced without intention does not usually conclude that intention was unnecessary; they conclude they were mistaken about what they were reading. Whether that reaction is rational is beside the point. It is the actual response, and it is a fact about the market a writer operates in.
The reply worth taking seriously
Two replies are strong enough that dismissing them would be dishonest.
The first is that every writing tool has been resisted on these grounds and the objection has been wrong before. Word processors were said to produce flabby prose because revision became too easy; typewriters were said to remove the hand from the sentence. The difference this time is one of kind rather than degree — earlier tools changed the cost of recording and revising choices, and this one proposes the choices — but somebody making the continuity argument in 1985 also believed their case was different.
The second is accessibility, and it is the harder one. A writer with a motor impairment, a dyslexic writer, a writer working in their third language: for each of them the tool removes a barrier between an intention that is entirely theirs and a text on a page. The intention is the thing the craft objection is about, and it is intact. Any position that rules those uses out has confused the discovery argument with a purity argument, and only the first of those was ever any good.
Submission rules, awards and disclosure
Beyond craft there is a practical layer that decides what you may do with a finished manuscript, and it is moving quickly. Publishers, agents, magazines and competitions have adopted a range of positions: outright prohibition, mandatory declaration at submission, permission for assistive use only, and silence.
The ownership question sits underneath all of it. The more of the final expression is yours, the stronger your position when a contract asks you to warrant the work is original and to assign rights in it — which is the copyright half of the plagiarism question, and worth reading before signing anything.
A working rule
One line, and it follows from the mechanism rather than from a position on art: the machine may touch everything about the book except the sentences and the choices. Continuity, research, timelines, diagnosis, transcription, the synopsis, the admin. Not the prose, not the plot turns, not what the characters say.
The test for whether you have crossed the line is not how the passage reads. It is whether you could explain, specifically, why that sentence and not the obvious one. If the answer is that it came back that way and it seemed fine, the choice was not made, and the choices are the book.