How to write fiction with AI without sounding like AI

Ask a room of novelists how to stop AI from wrecking their fiction and you'll hear the same list: cut the em dashes, vary your sentence length, delete the words the model overuses, stop hedging every claim. Sensible advice. It's also pointed at the wrong layer of the problem.

A Villanova University study reported in August 2026 asked readers to rate short stories, some written by people and some generated by an AI. Participants rated the AI stories as more absorbing and of higher quality. The authors were unusually direct about it: there is no general evidence that participants were able to distinguish human-written from AI-generated stories.

Sit with that for a moment. Readers can't reliably catch this at the sentence level any more. So if your instinct is to spend revision time scrubbing suspicious words out of a chapter, you're polishing the part that already works and leaving the part that actually breaks untouched.

Fiction written with AI fails structurally. It breaks in three places, and a reader feels all three long before they can point at a single suspicious sentence.

Why readers can't spot AI prose any more

The Villanova research is worth reading closely, because the interesting finding isn't that a model can produce a decent short story. It's what happened when readers were told where the story came from. Participants who had been told a story was human-written rated it higher, whoever actually wrote it. Belief moved the score more than the text did.

Accurate guessing tracked with one thing only: how much experience the reader had with AI systems. Not how much fiction they read. Deena Skolnick Weisberg, a senior author on the study, compared it to learning to recognize the signature style of a human author. You can learn it with practice. You don't start out able to do it.

That gap is the whole reason surface tells are a bad thing to build a revision process around. Detection is a skill that develops with exposure to AI output, and your readers are getting that exposure whether you like it or not. A general reader picking up a novel isn't running a detector over it. They're deciding, somewhere in the first few pages, whether the voice feels like a person.

Which is why the failures that actually get noticed aren't word choices. They're the ones a reader feels without being able to name: a scene where nothing is at stake, a chapter that keeps resolving its own tension, a character who is a slightly different person by chapter 25.

The sentence-level advice everyone repeats is a dead end

There's an entire genre of writing advice built on surface tells, and it made sense in 2023. Right after ChatGPT launched, the frequency of certain signature words in published articles jumped by more than half. They became fingerprints because the model used them at a rate no human writer would.

Those fingerprints fade. Models change their vocabulary, the lists get passed around, and the words everyone was told to avoid stop being tells. Build a revision process around a banned-words list and you'll be re-tuning it every few months while the real problem sits untouched.

The tool version of the same idea has the same flaw. Independent testing of sixteen popular humanizer tools found that only two did what they claimed. The rest mangled meaning, quietly dropped as much as a fifth of the original text, and still tripped the detectors they were built to fool.

There's a search-side reason to avoid that shortcut too. Google's spam policies name automated synonymizing, paraphrasing, and obfuscation as violations, which is a fair description of what a humanizer does to your prose. Publish fiction on the web and you're taking that risk for no narrative benefit.

If you still want the surface tells catalogued, our breakdown of the patterns readers notice first covers the sentence-level signals that keep mattering after the banned-words lists go stale.

There's a related failure that gets confused with sloppiness: prose that reads too polished to be believed. It's a different problem from messy writing, and it has a different fix.

Where AI fiction actually breaks

Think of a novel as three nested layers. A scene is a unit of want. A chapter is a unit of tension. A book is a unit of continuity. AI assistance tends to hold up at the first layer for a paragraph or two, degrade at the second, and fall apart at the third.

• Scene: what does someone want right now, and what is standing in the way?
• Chapter: what unresolved question does the reader carry into the next one?
• Book: which characters, relationships, and rules have to stay consistent across hundreds of pages?

Nearly every complaint I hear from fiction writers maps onto one of those three layers. Almost none of them map onto sentence rhythm.

Layer one: the scene has nobody who wants anything

Ask a model for a scene and you usually get atmosphere. Weather, a room, a mood, a character noticing things. It's competent and inert, because nobody in the scene wants anything.

Human scene writing starts from a want. Someone needs to get out of the building, get the answer, get the other person to stay. Then you put an obstacle in the way and let the two collide. Atmosphere is what happens while that collision is happening, not a separate thing you set up first.

A fast diagnostic: cover the last paragraph of the scene and ask what the character failed to get. If there was nothing they wanted in the first place, no amount of sentence rewriting will give the scene a pulse.

This is also where generic voice comes from. A model with no want to dramatize produces prose that's pleasant and says nothing, which is what readers mean when they say writing sounds like it was generated rather than meant.

Layer two: the model resolves tension before it can build

Language models return the most probable next beat. In practice that means conflict arrives already softened, and a scene that should end on a question ends on an answer instead.

Researchers examining AI-generated fiction keep finding the same shape: characters assembled from recognizable archetypes, and endings that arrive at tidy resolutions. That's not a flaw you can prompt away, because tidiness is the statistically likely outcome. Predictability is the product.

A separate line of research in Judgment and Decision Making keeps testing whether readers can identify which stories came from a person and which from a system, with mixed results. The honest summary of that literature is that people lean heavily on what they were told about the source.

Treat tidiness as an edit target rather than a style. When a scene closes neatly, the fix is usually to cut the last paragraph, the one where the character works out what everything meant. End a beat earlier, on the unresolved thing.

Layer three: continuity drifts by chapter 30

This is the failure that kills long projects, and it's the one fiction writers report most often. Practitioners working on novel-length drafts describe the same break point again and again: somewhere around twenty to thirty chapters, characters start behaving like different people, relationships get forgotten, and the rules of the world quietly change.

The cause is mechanical. A model drafting chapter 24 has no reliable memory of what chapter 6 established unless you put that material back in front of it. Context windows grow and shrink, summaries lose detail, and every generation is a fresh guess about who these people are.

The fix is unglamorous and it works. Keep a continuity file outside the chat. One page per major character listing what they want, what they know, and what they would never do. A short list of established facts. A log of open plot threads. Then paste the relevant slice into every prompt, instead of trusting the model to remember.

Writers resist this because it feels like homework. It's the most valuable hour you'll spend on an AI-assisted draft, because continuity is the layer readers punish hardest. Nobody throws a book across the room over a comma. Plenty of people stop reading when a character does something they would never do.

The quirks are not the disease, they are the immune system

Here's the part the surface-level advice gets backwards. Your odd sentences, your pet phrases, your stubborn refusal to write a paragraph the way a model would, those aren't flaws you're getting away with. They're evidence that a person was present.

The Villanova finding doesn't say style is invisible. It says an untrained reader can't reliably separate AI prose from human prose in a short text. Across a whole book the cumulative texture does real work, because a consistent voice is something a model struggles to hold and a person can't help producing. If you've never worked out what yours sounds like, start by finding your writing voice before you let a tool average it away.

So stop sanding your prose into neutrality. Not every strange sentence is a flaw. Ask whether the strangeness is doing a job, and keep the ones that earn their place. The goal isn't to sound like a model that's been improved. It's to sound like a person who has been edited.

A structural edit pass you can run in one sitting

Here's the pass to run on any AI-assisted draft before you touch a single sentence.

• Read for want. In every scene, name what the point-of-view character wants before the scene begins. If you can't name it, the scene is furniture. Cut it or give someone a goal.
• Check the last line of every chapter. Does it leave a question open? If every chapter resolves, you've written a sequence of summaries rather than a story.
• Scan for early resolution. Find the paragraph where the conflict gets solved, then ask whether it belongs two scenes later. Most tidy endings in AI-assisted drafts are simply premature.
• Audit continuity against the file, not your memory. Names, ages, physical details, abilities, relationships. Do it with search, not by rereading.
• Track every open thread. Anything you planted and haven't paid off goes on a list. Unpaid threads are how a draft feels unfinished even when the plot closes.
• Only then read for rhythm. Fix pacing, cut the flat paragraphs, and leave the odd sentence alone.

When you want the model to help, hand it the continuity file and a scene goal rather than a blank page and a chapter number. Models are far better at dramatizing a specific want against a specific obstacle than at inventing the want themselves. If you want the mechanics of feeding your own voice back in, the sample-based approach is the place to start: a handful of real samples beats any amount of instruction.

What to do if you draft with AI anyway

Most people reading this aren't deciding whether to use AI. They're deciding how. Here's the division of labor that holds up.

• Outline the want and the obstacle yourself. A model can dramatize a conflict. It shouldn't be inventing the reasons anyone cares.
• Draft scene by scene with the continuity file attached, and read each scene before generating the next.
• Keep the last pass for structure. Fix what happens before you fix how it reads.
• Accept flat sentences in the first draft. Flat is fixable and cheap. A scene with no reason to exist is neither.

The trade is simple. Use the model for volume and speed. Keep the parts that need continuity of judgment for yourself: what people want, what they would never do, and where the story stops.

Which leaves the awkward part. The research says readers can't reliably catch AI on the page, and readers still notice when a story is empty. Those two things aren't in conflict. They just mean the thing you have to get right is the one models are worst at: a person, wanting something, for three hundred pages.

Frequently asked questions