How To Write A Novel With AI Without It Sounding Like AI

Readers can't always name what's wrong with a passage of AI-generated fiction, but they feel it fast. A narrator who hedges every claim instead of committing to one. Three characters who all argue in the same measured, well-organized way. A death scene that tells you the protagonist "felt a wave of overwhelming grief" instead of showing you what her hands did. None of these are fatal on their own. Stacked across sixty thousand words, they're the difference between a novel a reader finishes and one they put down at chapter four without quite being able to say why.

This is a genuinely fixable problem, not a reason to avoid drafting fiction with an AI ebook generator in the first place. The tells that make AI-assisted prose feel synthetic are specific, nameable, and — once you know what you're looking for — editable in a focused, repeatable pass. This isn't a piece about whether AI can write good fiction in some abstract sense. It's a working list of the actual moves that separate a manuscript that reads like a novel from one that reads like a very fluent summary of a novel, and the concrete techniques that close that gap.

A quick note on scope before we get into it: this piece is about craft and voice specifically — the sentence-level and character-level tells that make fiction feel machine-generated, and the editing techniques that remove them. It's a companion to, not a repeat of, a related post on writing in your own voice with AI, which covers the mechanics of personalizing nonfiction voice more broadly. Here, the focus narrows to something nonfiction doesn't need to worry about at all: making a cast of invented people sound like different people, and making a narrator sound like a person telling a story rather than a system describing one.

Why AI-generated fiction has a recognizable "sound" in the first place

It helps to understand why the tells exist before trying to edit them out, because it changes how you edit. Large language models predict likely next words across an enormous, averaged corpus of text. That averaging is exactly what makes them useful for structure and getting a blank page moving — and exactly what works against them at the sentence level in fiction, where the whole point is that a specific character says a specific thing in a specific voice, not the statistically most probable version of a sentence.

A few consequences follow directly from that mechanism, and they show up as recognizable patterns:

Hedging and false balance. A model trained partly on cautious, on-the-other-hand nonfiction prose will sometimes carry that instinct into narration, softening claims a confident narrator would just make. Fiction doesn't want balance — it wants a specific point of view, even an unreliable or biased one.

Averaged dialogue. Ask a model to write "a conversation between two characters" without heavy constraint, and it will often produce dialogue that's grammatically complete, emotionally legible, and interchangeable — the same measured cadence no matter who's talking, because nothing in the prompt told it two specific people with two specific histories were talking.

Telling over showing. Naming an emotion directly ("she felt betrayed") is a safer statistical bet than dramatizing it through action, because it's shorter and appears constantly in training data as a plain descriptive sentence. Showing an emotion through a specific physical detail is a narrower, riskier choice — exactly why skilled writers do it and unconstrained generation tends not to.

Repetitive sentence rhythm. Left alone, generated prose tends to settle into a comfortable medium-length, subject-verb-object cadence, sentence after sentence. Human prose — even unpolished human prose — usually has more irregularity: a fragment, then a long compound sentence, then two short ones in a row.

Reach-for clichéd figurative language. Metaphors and similes that show up often in training data ("her heart pounded like a drum," "a storm of emotions") are statistically safe completions. Fresh, specific figurative language is inherently less common in any corpus, so it's a less likely default output.

None of this means the tool is incapable of better prose — it means better prose requires deliberate intervention rather than a single unconstrained generation and an assumption the first draft is close enough. The rest of this piece is a set of concrete interventions, organized around the tells above.

Give every character a voice you could identify with the names removed

The single most diagnostic test for flat AI-generated dialogue is simple: take a page of dialogue, strip out every dialogue tag and every piece of surrounding action, and see if you can still tell who's talking. If you can't, the dialogue is doing the "averaged" thing described above — technically fluent, functionally interchangeable.

Distinct character voice comes from a small number of concrete levers, and it's worth treating each one as a deliberate decision for every major character rather than something that emerges on its own:

Sentence length as a personality trait, not just a stylistic accident. A nervous, deferential character might speak in short, incomplete sentences that trail off. A character who's used to being obeyed might speak in flat, complete declaratives and rarely ask questions. Decide this per character, in advance, and hold to it.

Vocabulary register. A surgeon and a dockworker don't reach for the same words even when they're saying functionally the same thing. This doesn't mean leaning on heavy dialect spelling (which usually reads as a caricature and ages badly) — it means controlling word choice: formal versus casual, precise versus vague, technical versus plain.

What a character avoids saying. Real people rarely answer the question they were actually asked. They deflect, they answer a different question, they change the subject when a topic gets close to something they don't want to discuss. A character who always answers directly, completely, and helpfully is one of the fastest ways to make dialogue read as generated — it's the same instinct that makes a model want to be maximally helpful and complete in every response, which is exactly wrong for a guarded or evasive character.

A verbal tic used sparingly. One character who always answers a question with another question, one who never uses contractions, one who trails into silence instead of finishing a thought under pressure. A single well-placed tic, used only occasionally, does more than an over-applied accent.

A useful discipline, echoed by working novelists and craft resources alike, is to read dialogue aloud, one character's lines at a time, straight through a scene, ignoring the other characters' lines entirely. Writer's Digest's guide to writing authentic dialogue makes this the first recommendation for a reason: prose that looks fine on the page often reveals itself as stilted or wrong for the character the moment it's spoken aloud, because your ear catches rhythm problems your eye skips past. If two characters' dialogue sounds interchangeable read aloud back to back, that's the signal to rewrite one of them with the sentence-length and vocabulary levers above. This is also where a chapter-by-chapter AI ebook creator workflow actually helps rather than hurts: revising one character's dialogue in a targeted pass, chapter by chapter, is far more manageable than trying to fix voice consistency across a whole manuscript in one sitting at the end.

A hypothetical before-and-after makes this concrete. Imagine a generated first draft where two characters, a detective and a suspect, both speak like this:

"I understand your concern, but I need you to walk me through exactly what happened that night," the detective said.

"I appreciate that you're trying to help, but honestly, I don't remember all the details clearly," the suspect replied.

Both lines are complete, polite, grammatically careful, and roughly the same length and register — that's the averaging problem in miniature. A revised version, deliberately unbalanced, might look like:

"Walk me through it. That night." Flat. No hedge.

"I— it's not — look, I wasn't really paying attention, okay?" The suspect wouldn't meet his eyes.

Nothing about the second version is more "correct" than the first — it's more specific, and specificity is exactly what a statistically-averaged first draft tends to sand off.

Stop the narrator from hedging and hand it an actual point of view

Narrative voice has its own version of the dialogue problem. A close third-person or first-person narrator is supposed to have opinions, blind spots, and a particular way of noticing the world — not deliver a balanced account of events the way a careful nonfiction summary would.

Watch specifically for these narrator-level hedges, which tend to survive an unedited generation:

• Qualifying language around a character's internal state ("she seemed to feel," "it was almost as if he") where a confident narrator would just commit: "she felt," "he was."
• Both-sides framing inside a single character's head — presenting a decision as perfectly balanced pros and cons when real internal conflict is usually lopsided and irrational.
• Narration that steps back to explain the significance of something the scene has already dramatized, functionally repeating itself in a more abstract register. If the scene already showed the reader that a marriage is falling apart, the narrator doesn't also need to state "their marriage was falling apart" a paragraph later.

The fix here is a specific edit pass: read through a chapter looking only at how the narrator handles interiority, and delete every sentence where the narrator explains a feeling the scene already dramatized, or hedges a claim the point-of-view character would actually be certain of. A single point of view, held consistently and confidently — even when that character is wrong or biased — reads as far more like fiction than a narrator trying to be fair to every possible interpretation of a moment.

Show the specific thing, not the category of thing

"Telling" is the tell that shows up most often in generated fiction, and it's also the most mechanical to fix once you're looking for it. The pattern is almost always the same: a sentence names an emotion or a quality directly, in the kind of language you'd use in a summary, instead of dramatizing it through a specific, concrete, sensory detail.

A hypothetical pair makes the pattern obvious:

Generated, told: "Maria was nervous as she waited for the results."

Revised, shown: "Maria had read the same sentence in the waiting-room pamphlet four times without absorbing a word of it."

The second version never uses the word "nervous." It doesn't need to — a reader infers the state from a specific, concrete action, which is both more convincing and more memorable than the label. This is the core mechanism behind the classic "show, don't tell" craft advice, and it's worth understanding why AI-generated first drafts skew toward the "told" version by default: naming the emotion is a shorter, lower-risk completion, while inventing a specific behavioral detail requires committing to a concrete, invented particular that has to actually fit this character in this situation.

A practical technique for this pass: go through a chapter and highlight every sentence that names an emotion, a personality trait, or a quality directly with an adjective or an abstract noun — nervous, angry, beautiful, exhausted, cold, in love. For each one, ask what specific, physical detail would let a reader infer that state without being told. Not every instance needs to go; naming an emotion plainly is sometimes exactly the right economical choice in a fast-moving scene. But if a chapter has a dozen of these and none have been questioned, that's a signal the draft leaned on the model's default instead of a deliberate authorial choice.

The same principle applies to setting and description, not just emotion. "The old, dusty library" is a generic category — it could be any library in any book. "The library where the returns slot still had a hand-lettered sign from 1987" is a specific, invented, ownable detail. Concrete, specific nouns do more work than generic adjectives stacked in front of a generic noun, and they're one of the fastest ways to make a scene feel like it belongs to your book rather than a book.

Vary sentence rhythm on purpose, in a dedicated pass

Sentence-rhythm monotony is one of the harder tells to catch by reading normally, because any single sentence, read in isolation, is usually fine. The problem is cumulative — paragraph after paragraph of medium-length sentences with similar internal structure, none of them wrong individually, all of them tiring together.

This is worth treating as its own editing pass, separate from a content edit or a line edit for word choice — you're reading for shape and cadence, almost ignoring meaning. A few concrete techniques:

Read a page and mark sentence length in the margin, or just eyeball where the periods fall. If four or five sentences in a row are roughly the same length, that's the pattern to break. Real prose rhythm tends to look uneven on the page — a short sentence next to a long one next to a fragment — even before you read a word of it.

Deliberately write — or rewrite — some sentences as fragments. Not every sentence needs a subject and a verb. "She didn't answer. Just walked." does something a grammatically complete pair of sentences doesn't: it mimics the clipped rhythm of someone who's actually upset, rather than describing that person from a calm remove.

Combine two short, choppy sentences into one longer one occasionally, and split one long sentence into two occasionally — in the opposite direction from wherever the draft's current pattern sits. If a paragraph is all short sentences, look for one that would benefit from being extended, complicated, or run together with a comma splice a copyeditor would flinch at but a reader wouldn't. If a paragraph is all long, flowing sentences, find the one moment that deserves a hard stop.

Read the chapter aloud, or use text-to-speech, and listen for where you run out of breath or where your voice flattens out. This is a genuinely different way of catching monotony than silent reading, because rhythm is fundamentally an auditory pattern, and a flat rhythm is much more audible than it is visible.

None of these techniques are exotic — they're the same instincts a human writer applies during a normal revision pass. The difference with AI-assisted drafting is that this pass needs to be deliberate and scheduled, because an unedited first generation has no particular reason to vary its own rhythm; it will happily produce a hundred structurally similar sentences in a row if nothing pushes it away from that default.

Build and use a banned-phrase list

Certain words and constructions show up disproportionately often in AI-generated prose because they're statistically safe, broadly applicable completions — which is exactly what makes them feel generic rather than specific to your book. Wikipedia's own internal guidance for editors on identifying signs of AI-generated writing catalogs many of these patterns in detail: an inflated sense of significance attached to ordinary events, vague attribution ("some say," "it is widely believed"), and a small cluster of words and phrasal patterns — including certain stock transitions and hedging constructions — that recur across generated text regardless of topic. That guidance is written for encyclopedic prose, not fiction, but the underlying mechanism — a model reaching for the statistically safest, most broadly applicable phrasing — produces the same kind of genericness in narration and description, just in fiction-specific forms.

Building a banned-phrase list is a cumulative, project-specific habit, not a one-time checklist you complete before you start. Some categories worth watching for from the start:

• Stock physical descriptions of emotion that show up constantly because they're safe defaults: heart pounding/racing, breath catching, a chill running down a spine. These aren't wrong exactly — they're just so common a reader's eye slides past them without registering anything specific.
• Narrator hand-holding phrases that step outside the scene to explain what it means: "little did she know," "in that moment, everything changed," "it was a turning point."
• Vague sensory filler that describes a category rather than an instance: "the air was thick with tension," "a heavy silence fell."
• Overused connective tissue between actions: "with a heavy sigh," "a small smile playing at her lips" — fine used once in a whole manuscript, a glaring tell used once per scene.

The practical use of a list like this isn't to run a find-and-replace and refuse every instance outright — some of these phrases are genuinely the right choice in a specific sentence. It's to search for them at the end of a chapter and treat every hit as a place to stop and ask whether a more specific alternative exists for this exact moment, this exact character, this exact room. Over several chapters the list gets longer and more personal to your book's own recurring defaults — whatever phrase you're adding this week is probably the phrase the model, or your own tired end-of-chapter drafting, is currently reaching for by default.

Treat the first generation as a draft to correct, not a draft to accept

Every technique above assumes the same underlying discipline: iterative correction, not one-shot acceptance. A single unconstrained generation of a chapter is a starting point for revision, the same way a human writer's own rough first draft is a starting point — not a finished product to lightly proofread and move past.

In practice, that means treating chapter generation as a loop rather than a single event: generate, read critically against the tells above, regenerate or manually rewrite the passages that fail the test, and repeat until the chapter holds up to the read-aloud and highlight-the-adjectives tests described earlier. This is the same logic behind a dedicated AI editor built into a novel-writing workflow — going back into an already-generated chapter in conversation, tightening a scene, or asking specifically for more concrete sensory grounding, rather than starting over from a blank prompt every time something reads flat. Regenerating an entire chapter when only one scene's dialogue feels generic wastes the parts that were already working; a targeted revision pass on just the flat section is faster and produces a more consistent chapter overall.

It's also worth being honest about where this loop has diminishing returns. A generic phrase swapped for a fresher one, a hedge removed from the narrator, a character's dialogue rewritten to sound less like everyone else's — these compound. But no amount of regeneration substitutes for an actual line-edit read-through of the finished manuscript, ideally aloud, checking specifically for the rhythm and telling-versus-showing patterns above rather than just for typos. Craft resources on dialogue point to the same discipline from a different angle — Reedsy's course on writing dialogue frames distinct character voice and subtext as skills built through deliberate practice and revision, not something that arrives correct on a first pass, whether that first pass came from a human writer or a generation call. The tool changes how fast a draft gets to the page. It doesn't change the fact that a novel gets good in revision.

Putting it together across a full manuscript

None of the techniques above are meant to run once, on one chapter, and then be considered handled. A novel is long enough that tells creep back in after you've caught them once — a banned phrase eliminated in chapter three reappears in chapter fourteen because a new generation call didn't know about your list, a character whose voice was sharp in the opening chapters drifts back toward the generic middle because eighty thousand words is a long distance to hold a deliberate choice without reinforcement.

That argues for building these checks into a recurring pass rather than a one-time fix: read dialogue aloud per character every few chapters, not just once at the start. Re-scan your banned-phrase list against new chapters as they're generated, not only at final edit. Do a dedicated rhythm pass on each act rather than assuming the fix from chapter one held for the whole book. This is more work than treating a novel-writing tool as a single button that outputs a finished manuscript — but it's also exactly the work that turns a technically complete draft into a novel a reader can't tell was drafted with help at all. Used this way, an AI novel generator is closer to a fast, tireless first-draft collaborator than a finished-book vending machine — genuinely useful for getting several chapters of raw material onto the page quickly, and entirely dependent on a writer who knows what to listen for once that material exists.

The tells covered here — hedge-y narration, interchangeable dialogue, told-not-shown emotion, flat rhythm, and reach-for clichés — are not mysterious. They're specific, they're checkable, and every one of them has a concrete, repeatable fix. A writer who knows to look for them, who reads dialogue aloud per character, who runs a dedicated rhythm pass, who keeps a growing banned-phrase list, and who treats every generated chapter as a draft to interrogate rather than a draft to accept, ends up with prose a reader experiences as a specific voice telling a specific story — which was always the actual goal, independent of what tool got the words onto the page first.