Writers worry that using AI will sand their voice down to a smooth, anonymous gloss. The worry is legitimate — but the cause is almost always the same, and it is fixable. Generic prose is what a model produces when nobody told it whose voice to write in. Voice is not something you rescue after the fact with a "humanizer"; it is something you specify before the first sentence and enforce on every draft.
Voice is a set of decisions you can name
It is tempting to treat voice as a mystical quality, but on the page it decomposes into concrete, repeatable choices. Who narrates, and how close are we to them? Past or present? Long, rolling sentences or short, clipped ones? Plain diction or ornate? Do characters speak in subtext or say what they mean? What images recur? And crucially — what do you refuse to do? Every one of those is a decision you can write down, hand to a collaborator human or otherwise, and check a draft against.
That is why the fix for "AI sounds generic" is not a better prompt like "write in my voice" — the model has no idea what that means. It is a specification. The eight-part contract above is that specification: it turns the intangible into eight instructions concrete enough to enforce.
A worked revision: one paragraph, held to the contract
Suppose the contract says: past tense, close third on Mara, short sentences, plain concrete diction, a recurring salt-and-tide image system, and a hard avoid on rhetorical questions and the word "suddenly." Here is a generic draft:
DRAFT: "Suddenly, Mara felt a wave of overwhelming emotion wash over her as she gazed out at the vast, glittering expanse of the ocean. What did the future hold for her now? Everything had changed, and she knew that nothing would ever be the same again."
Now the same beat, revised against the contract:
REVISED: "The tide was going out. Mara watched the water pull back from the rocks, leaving them slick and black. Her hand found the salt-crust on the rail. Whatever came next, it would not be this."
Every change traces to a rule, which is the point — nothing here is taste for its own sake:
- "Suddenly" is gone — it was on the hard-avoids list.
- The rhetorical question ("What did the future hold?") is cut for the same reason.
- "Overwhelming emotion" and "vast, glittering expanse" go — the contract asks for plain, concrete diction, so we show the tide, not the feeling.
- Salt and tide do the emotional work — that is the declared image system earning its place.
- Sentences are short and we stay inside Mara’s senses — close third, the declared distance and rhythm.
Notice what this is not: it is not a pass to "humanize" or disguise anything. It is craft — a draft measured against choices the author made, and revised where it fell short.
The honest take on "humanizing" AI writing
Search interest in "humanizing AI text" is usually really a search for this: writing that sounds like a person because a person made real choices in it. The reliable way to get there is the contract-and-revision loop above. It produces prose that is genuinely yours, not prose engineered to fool a detector — and it holds up to the only audience that matters, which is a reader.
This page deliberately stops there. Defining and enforcing your voice is craft worth teaching. Tools built to evade AI detection are a different thing with different aims, and they are not what this is about. Do the craft, and the "does this sound human" question mostly answers itself.
Measure voice drift with three passages, not one impressive rewrite
A voice contract is only useful if it survives different narrative jobs. Test it against three passages from the same book: a quiet interior beat, a high-pressure scene, and dialogue between characters with unequal power. Give the assistant the same contract and the minimum context needed for each task. Then compare the outputs against your own accepted pages, not against a generic idea of “good writing.”
Score observable features rather than asking whether the result feels human. Count point-of-view slips, tense changes, forbidden words, rhetorical questions, sentence-length bands, abstract emotion labels, repeated image families, and dialogue lines that state what the contract says should remain subtext. A rule that cannot be recognized in the prose is too vague to govern it; rewrite the rule with a positive example and a hard boundary.
Run the test again after several chapters. The risk in a long book is not one bad paragraph but slow normalization: sentences lengthen, diction becomes shinier, character speech converges, and the assistant treats a one-off flourish as a new standard. Preserve the original test passages and results as a baseline. When a later chapter drifts, revise the smallest failing passage and update the contract only if the author’s intended voice truly changed.
- Baseline: three accepted passages that represent different narrative pressures.
- Checks: tense, distance, syntax, diction, imagery, dialogue, and hard avoids.
- Decision: revise the draft when it breaks the contract; revise the contract only when the author changes the rule.
Keep the contract where the book applies it
A voice contract you write once and lose is a voice contract you stop enforcing. The reason voice drifts across a long AI-assisted book is that the standard lived in a chat that scrolled away, so nothing was holding chapter twenty to the choices you made in chapter one. The contract has to persist, and it has to be applied.
In BookWriter, an approved voice contract lives with the project as a rule the drafting can be held to, chapter after chapter. You keep writing in whatever conversation you like; the standard travels with the book, so the last page sounds like the same author who wrote the first.
Keep narrator voice and character voices from collapsing into one style
A book can maintain its narrative voice and still fail because every character speaks like the narrator—or like every other character. Build a small dialogue matrix for the recurring cast. Record what each person notices first, how directly they answer, typical sentence length, vocabulary range, preferred evasions, status behavior, and one hard avoid. The fields should describe choices visible in speech, not accents assembled from misspellings or a list of personality adjectives.
Test the matrix by removing dialogue tags from a short exchange. A reader should not identify every line with certainty, but the speakers should apply different pressure. One may answer questions with questions; another may give exact numbers; another may avoid names and speak through implication. If the only distinction is catchphrases, the voice will feel mechanical. Vary the underlying decision pattern instead.
Then check character speech under changing power. A person may sound expansive with a sibling and clipped before a judge while remaining recognizably themselves. Add a stable core and a context modifier to the matrix rather than demanding identical diction everywhere. This prevents a voice contract from flattening realistic adaptation into a continuity error.
During revision, compare the scene with two accepted samples for each active speaker and cite the specific mismatch. Propose the smallest repair—often one line or response pattern—without rewriting the whole exchange into smoother generic dialogue. Save a new sample only after the author approves it. Otherwise one accidental generation can become the reference that shifts every later scene.
Keep dialogue rules subordinate to character development. When a guarded character finally answers directly, the break can be the point of the scene. Mark the change as a deliberate, story-earned exception and identify the beat that caused it. A mechanical checker should flag the deviation with evidence, but the author decides whether it is drift or growth. If accepted, update the character’s effective voice state from that scene forward instead of rewriting earlier speech to match the new openness.
Run a final scene-level check for convergence. Highlight five lines from each speaker and compare syntax, image choices, emotional directness, and conversational tactics. If exchanging the names would leave most lines plausible, identify one decision pattern that should differ and revise only the lines where the overlap matters. Characters can share a setting, class, or family vocabulary without becoming interchangeable; the test is whether each one pursues the scene in a recognizably different way.
Read the revised exchange aloud without performance. Mark any line whose rhythm, vocabulary, or implied emotion requires the speaker label to make sense. Return those lines to the matrix and fix the decision beneath them rather than adding decorative verbal tics.
| Character voice field | Question to answer |
|---|
| Attention | What does this person notice before anyone else? |
| Directness | Do they answer, deflect, counter, or stay silent? |
| Rhythm | Short bursts, balanced clauses, or extended stories? |
| Status response | How does speech change above and below their power? |
| Hard avoid | Which phrase, disclosure, or rhetorical habit would break character? |