AI writing that doesn't sound like AI is what you get when the human does the parts that matter and the machine does the grunt work. Two different jobs. Most people have them backwards, and the readers who matter can tell: a 2025 study found that people who use ChatGPT a lot for their own writing spot machine text almost every time, a panel of five of them misclassifying only 1 article in 300 (Russell, Karpinska and Iyyer, ACL 2025), better than most commercial detectors. Those heavy users are reading your Kindle sample.
So the goal isn't to trick anyone. It's to keep the machine out of the places readers can hear it. The tells first, because you can't fix what you can't hear. Then a real sample, with the tells marked.
How readers catch it
The default voice of a language model is a specific, recognisable thing. Once you've clocked it, you see it everywhere: cover letters, newsletters, that suspiciously even-tempered Amazon review. Wikipedia's editors keep a catalogue, Signs of AI writing, from years of cleaning it out of articles, and most of it carries straight over to fiction (software is a worse judge than these readers: when we ran AI detectors on fiction, a free one missed four of six scenes from Mistral Small):
- Over-polished symmetry. Every sentence the same medium length. Every paragraph three sentences. Clauses balanced like a furniture showroom. Real writing limps and sprints. It has a bad knee.
- Hedging on everything. "Can be," "often," "may help," "in many cases." The model is trained to never be wrong, so it commits to nothing. A human writer plants a flag.
- The summary reflex. It restates what it just said, then tells you what it all means, then ties a bow on it. Nothing is allowed to just sit there and be unsettling.
- No real specifics. It'll say "a delicious meal" where you'd say "cold pad thai eaten over the sink at 1 a.m." The model reaches for the category. A writer reaches for the thing.
- The "not X, but Y" tic. Wikipedia's page calls these negative parallelisms: "it wasn't anger, it was grief." One per book, maybe. Models do one per paragraph.
- The rule of three, and the vocabulary. Three adjectives, three clauses, three examples, every time. And the words themselves: "pivotal", "crucial", "underscore", "showcase", and the verb for digging into a topic that the Kobak paper's own title winks at. That study of 15 million biomedical abstracts measured the spike in these words after 2023 and estimated at least 13.5% of 2024 abstracts had been through a model (Kobak et al., Science Advances, 2025). Scientists didn't all start writing "pivotal" at once. Their tool did. The same goes for advice: "write fights in short sentences" is repeated everywhere, and when we counted five classic fights the longest-sentenced one was among the fastest to read.
We counted the stock phrases ourselves in the AI Slop Index: 1,321 short scenes from 13 current models, checked against a public list of the phrases AI over-uses in fiction. Three frontier models used none of them in about 27,000 words each; Llama 4 Maverick used one every 200 words. The newer tics survived everywhere, though: every model wrote some "it wasn't X, it was Y" sentences. The same list explains why AI kiss scenes all sound alike: "heart skipped a beat" and "breath caught in her throat" are both in its top hundred, and our guide on how to write a kiss scene has the counts.
None of these is really about word choice. They're about stance. The model has no body, no grudges, no Tuesday it can't stop replaying. So it writes from nowhere, about everything, to no one in particular. That's the smell.
Write the messy human version first
Here's the single biggest shift, and it's about workflow more than prompting: don't let AI originate. Let it expand.
Ask a model for a scene from a blank prompt and you get the average of ten thousand scenes like it. The orphan is plucky. The rain is symbolic. The mentor dies in act two. This is measurable, not a vibe: the fine-tuning that makes chatbots helpful was found to "significantly reduce output diversity" (Kirk et al., 2023). A model hands you the centre of the distribution, and the centre is exactly where everyone else already is. The button in your word processor works the same way: Help me write in Google Docs drafts from your prompt, not from your book. Real ideas rarely start there; where five famous books got their inspiration shows most began as one odd scene.
So write the ugly version yourself first. Not a polished draft; a beat. A few rough lines carrying the things only you know:
She apologizes by overwatering his plants. He notices the dripping pot, says nothing, gets it.
Barely a sentence, and already more alive than anything a cold prompt produces, because it runs on a specific, slightly off emotional logic: her logic. Now AI can expand it. You stay the author of the idea. The machine upholsters it. That's why a beat-at-a-time workflow beats one big "write me a chapter" prompt: when you draft rough beats and then expand them one at a time, you decide what the scene is about before a word gets generated. The AI never drives. It navigates the parking lot you built.
What the machine actually did with that beat
Here is the beat above run through deepseek/deepseek-v4-flash-0731, the model KudoWrite's built-in KudoAI credits use, via OpenRouter, October 2026. A bare API call, one system line ("write prose only"), asked for about 150 words. No bible, no steer, no rewrite. 354 tokens, under a hundredth of a cent. Five of its seven sentences, verbatim:
She apologized by overwatering his plants. The fern on the windowsill sat in a saucer brimming with cloudy water, and the jade plant’s soil had turned to dark mud. He found the dripping pot on the kitchen counter, a slow, persistent plink against the tile, and paused.
He said nothing. He simply righted the pot, tipped the excess into the sink, and ran a finger along the cool, damp leaf of the fern. It was her way of asking for forgiveness, and his way of granting it was to leave the plants exactly where they were, thriving under the weight of her regret.
Credit where it's due: the fern, the jade plant, the cloudy saucer and the mud are real specifics. Keep those. Now the tells.
The first sentence is the beat handed back with the tense changed. Cut it; the pictures do the job. The sentence left out of the quote (it carried an em dash, a tell in itself) explains the subtext outright: he knew she had "let her guilt spill over in lieu of words". The beat said "says nothing, gets it" precisely so the reader would get it too. Cut. "Simply" is filler. "Righted, tipped, ran a finger" is the rule of three. "Cool, damp leaf" and "slow, persistent plink" are paired adjectives, twice in six lines. And the last sentence is the bow, "her way of asking for forgiveness, and his way of granting it", a symmetrical summary ending on a metaphor that tells you what to feel. Delete it whole.
What survives is the saucer, the mud, the plink on the tile, the pot righted, a finger on the leaf. Around fifty words, ending on the gesture instead of the moral. That's the edit, every time: keep the furniture, cut the explanation.
Rewrite line by line, never accept the paragraph whole
The lazy move: generate a paragraph, skim it, decide it's "good enough," paste it in. Do that across a novel and you get a book that's mostly good-enough. Good-enough is the exact texture readers have learned to distrust.
Better: treat every generated line as a suggestion you're arguing with. Read it out loud. Where does it go slack? Where did it reach for the obvious word? Kill the hedges. Cut the sentence that summarises the previous three. Swap one generic noun for the real one only you would know to use. You'll often keep the model's structure, the order of images, the rhythm of an exchange, while replacing half the words. Fine. That's the work. Generate two or three versions of a beat, steal the best line from each, and rewrite the stitched result by hand until it sounds like you on a slightly better day. The final sentences passed through your judgment one at a time, and a reader feels that even when they can't say why.
Feed it your voice, and the specifics it can't invent
A model defaults to a voice: competent, neutral, faintly corporate. If you want yours, you have to show it. Paste in a few paragraphs of your best prose and tell it to match the cadence, the sentence shapes, the way you handle dialogue. Keep a "voice sample" somewhere you can reuse. A good story bible, the context notes the AI reads before it writes, should hold more than plot and character facts. It should hold a sample of how the book actually sounds. Then the expansions come back closer to your wavelength than to the default one. (Feeding a long draft back in is its own problem: how to get AI to continue your story.)
And guard your specifics like the last matches in a cold cabin, because they're the one thing the machine genuinely cannot produce: the exact wrong thing your character says at the funeral, the smell of the apartment that tells us the marriage is over before anyone speaks, the regional word, the inside joke, the texture of a real job, an opinion the narrative actually holds. The model will paper over a missing specific with a pleasant generality every time, and that generality is precisely the AI smell. When a generated passage gives you "she felt overwhelmed," you delete it and write what being overwhelmed looked like for this person, in this kitchen, on this day.
Do the hard human pass at the end
However careful you were beat to beat, read the whole thing through once more with a red pen and a foul mood. This is where you catch the drift the line-level edits missed: the chapter gone suspiciously smooth, the three paragraphs that all open the same way, the opinion that quietly leaked out and left everything sounding fair and dead. An AI beta reader can flag where a chapter sags, but it reads for structure, not for whether the voice is still yours.
Hunt for the tells from the top of this piece. Search your draft for "began to," "a sense of," "it was clear that," "couldn't help but," and for the "not X, but Y" shape. Orwell's first rule in Politics and the English Language is the whole AI problem in one line: never use a figure of speech "which you are used to seeing in print". A model has seen all of them in print. Read dialogue aloud. AI dialogue is too articulate, too on-topic; nobody interrupts or misunderstands. Add the friction back. Let someone answer the wrong question (a before/after is in how to write dialogue that sounds human).
On a long project the same instincts scale: what fixes one paragraph keeps a whole manuscript sounding human across an entire novel. One caveat: a tool that works beat by beat, KudoWrite included, makes the human pass easier but doesn't do it for you. Nothing in the app rewrites your voice back in. You do.
The throughline
Think of the model as a power tool. A table saw doesn't design the cabinet; it cuts faster and straighter than you can by hand, and it'll happily cut the wrong board if you let it.
So keep your hands on it. Write the strange specific seed, let the machine expand it, then rewrite the expansion until it bleeds your voice and not its own. A beat at a time. Then the human pass no amount of compute will do for you.
Get it right and nobody ever asks whether AI touched your book. They're too busy reading it.