Can AI write a novel?
It can produce something novel-shaped. It cannot produce a novel worth reading, and the gap is not small. AI is useful for the scaffolding around fiction: brainstorming, naming, working through a plot snag, drafting a synopsis you’ll rewrite. But ask it to actually write the book and you get prose that’s competent sentence by sentence and hollow everywhere it counts. The stakes don’t pay off, the characters all sound like the same person, the structure repeats itself, and the whole thing reads like it was assembled rather than told. Fiction depends on exactly the things a prediction engine can’t do: genuine causality, real loss, a human voice that belongs to one person. Used as a tool under a writer who knows its failure modes, AI helps. Handed the keys, it writes the kind of book readers abandon by chapter three. For the publishing and copyright side, see my
AI and Your Book FAQ, and I’ve written at length about working with these tools in my
AI-Enhanced Writing handbook series.
Why is AI bad at the bones of a story?
Because it has no model of cause and effect, only of what tends to follow what. So AI plots move because the story needs them to, not because of what came before. The clue surfaces exactly when the protagonist is stuck. The right person walks in at the right moment. These are conveniences, and readers feel them even when they can’t name them. The deeper problem is that AI plots are safe: the protagonist faces obstacles but never truly loses anything that matters. Real stories make the character pay, something real is gone by the end that was there at the start, a relationship, a belief, an illusion about themselves. AI also defaults to a clean three-act shape at every level no matter what structure you asked for, so the story collapses toward setup, confrontation, tidy resolution even when the material wants something else. Strong fiction earns every turn and leaves some things broken. That’s the part AI can’t fake.
Why do AI characters feel flat?
A few specific reasons, and once you see them you can’t unsee them. AI gives every significant character a clean arc: they all learn something, change, and arrive somewhere, when real people often end exactly where they started and that’s the truer thing. AI secondary characters are purely functional, they show up to deliver information or opposition and vanish when the protagonist doesn’t need them, instead of having their own agendas and lives that continue offstage. AI antagonists are conveniently wrong, legible villains who confirm the hero is right, rather than people with a coherent worldview the reader could imagine holding. And worst of all, AI characters blur: different names and roles but they think alike, speak alike, notice the same things, because the model’s single voice leaks into all of them equally. Real characters are each a different person, you should be able to delete the dialogue tags and still know who’s talking. AI almost never clears that bar on its own.
Why do AI chapters and scenes feel the same?
AI builds to a template and repeats it. Scenes summarize instead of dramatize, telling you what happened rather than putting you in the room while it happens. They lack subtext, the characters say what they mean instead of maneuvering around it. And in a group scene, everyone is politely focused on the same topic, when a real scene has four people running four private agendas at once: one watching the door because they need to leave, one short-tempered because they woke up wrong, one throwing covert jabs to see what lands. AI also lets the middle sag, it pours attention into the hook and the ending and treats the middle as filler, which is exactly where most readers quit and exactly where character is actually built. The fix is rebuilding scenes so each one dramatizes, carries subtext, and does a job only it does. That’s hands-on structural work, not a prompt.
What are the prose-level tells that AI wrote it?
There’s a fingerprint at the sentence level, and it’s consistent. Cognitive process theater: “I noted this,” “I registered,” “I found myself,” “I was aware that,” narrating the act of noticing instead of just noticing. The computational tell of a character “filing it away,” which is how a machine describes its own memory, not how a person thinks. Stock body-language tics on repeat: the breath he didn’t know he’d been holding, the quirked eyebrow, constant nodding, swallowing when nervous. Hedged gestures that won’t commit, the “almost smiled,” the “something like a laugh.” The relentless “this, then this, then this” cadence where every sentence runs the same length and shape until the prose flatlines. Triads everywhere because three feels complete to the model. Qualifier pileups, “very,” “just,” “somewhat,” softening every line. And em dashes scattered through the narration. Any one of these is a habit. All of them stacked at machine density, page after page, is a signature.
What about AI detectors like Originality.ai and GPTZero?
Treat them as a weak signal, never a verdict. The irony is built in: these tools use AI to detect AI, and they are wildly inaccurate in both directions. They flag human writing as machine-written and pass machine writing as human, routinely. A positive result does not prove a person used AI, and it should never be used to condemn a piece of work on its own. Plenty of people write in ways that happen to trip these detectors: clean, plain, well-organized prose, or writers who learned in a structured corporate or academic setting, or simply anyone whose natural style is tidy. Non-native English speakers get falsely flagged at especially high rates. There is some use in running a detector as one input among many, a prompt to look closer at a passage, but the moment you treat the score as proof you are trusting a tool that is wrong often enough to ruin someone unfairly. The reliable read on whether something was machine-drafted comes from a human who knows the actual tells and checks the facts, not from a number a detector spits out.
Can you tell when fiction was written by AI, and can you fix it?
Yes to both. I read for the fingerprint at both levels: the story bones (unpaid stakes, conveniences, clean arcs on every character, legible antagonists, the sagging middle, the default three-act collapse) and the prose (cognitive process theater, “filed it away,” the stock body-language tics, hedged gestures, the “this then this then this” rhythm, triads, qualifier pileups, em dashes). Then I rebuild: stakes that actually cost something, characters who each sound like a different person, scenes that dramatize instead of summarize, a middle that earns its pages, and prose with varied rhythm and a real voice. The goal isn’t to disguise that AI touched a draft. It’s to make the book read like a person wrote it, because in every way that reaches the reader, a person did. If you have a manuscript that came out of AI and feels off but you can’t say why, that off feeling is the fingerprint, and it’s fixable.