Keeping Your Voice: Editing with AI Without Losing It

Why editing with AI flattens your tone

A language model does not preserve your sentence; it moves the sentence toward the most probable way of saying it. The text gets cleaner and more even, and your voice — the particular logic by which you see the subject — drains out of it.

It helps to separate tone from voice. Tone is the surface — formal or casual, warm or dry — and a few short samples are enough to imitate it. Voice is the layer underneath: what you treat as important, what you find ridiculous, where you start and where you pause. A model copies the first and never builds the second.

Three mechanisms that smooth voice away

• Averaging. The model drifts toward the more probable phrasing, and any sentence outside the average pattern gets pulled back.
• Correcting on purpose. A short punchy sentence, a deliberate repetition — the model reads these as mistakes and repairs them, because it cannot tell they were chosen.
• Evening out the rhythm. Sentence lengths converge. The prose becomes musical, but on the average rhythm, not yours.

What you should not delegate at all is the subject of the complete guide to smart editing; here we only deal with the method for keeping voice.

Signals that the voice is slipping

Voice does not vanish in one dramatic paragraph. It leaks. Check a finished chapter against these:

• The spread of sentence lengths has narrowed; every sentence is about the same size.
• Connectives have appeared that you never use.
• Your key words have been swapped for synonyms — and the synonym turns into something else two chapters later.
• Your irony, your edge, your joke has been made polite.
• Your own comparisons have been replaced with general description.
• A sentence you deliberately left unfinished has been finished.
• Nothing in the passage surprises you any more.

The first three together mean this is no longer a copyediting question. Standardising half-spaces is correct work and does not touch voice; standardising vocabulary — which looks like the same job — does. The difference is set out in Copyediting Persian Text with AI.

The anchor-sample technique

The anchor sample is straightforward. Before you edit anything, choose two or three paragraphs of your own writing that best represent your voice — not the best paragraphs in the book, but the ones that would identify you to someone who knows your work.

Repeat that sample in every prompt, and ask the tool to describe its features first, then edit with those features as its brief. That description is itself a style sheet for voice: sentence length, adverbs, how you address the reader, how you build an argument.

Write the voice profile yourself

Do not stop at the model’s description. Write a few lines in your project file: who you are talking to; what you never say; how much of yourself you put in; where you refuse adverbs and adjectives.

If names and terms also have to stay fixed across the book, keep the voice profile next to the style sheet: it says what is correct, the voice profile says what sounds like you.

Three constraints you put in the prompt

A prompt without constraints gives the tool permission to do anything. Write these three every time.

1. Do not change vocabulary. No synonym substitutions unless the word is wrong or clashes with the style sheet. Your lexical field is part of your voice.
2. Do not move sentence boundaries. No merging, no splitting. If a sentence is unclear, flag it; do not rebuild my rhythm for me.
3. Flag, do not fix. Wherever you see ambiguity, repetition or a likely fault, report it in a separate list and leave the text alone.

A prompt skeleton that does the job:

Edit the text below against the anchor voice sample. Three constraints: do not change vocabulary; do not move sentence boundaries; where you are unsure, flag it rather than fix it. Return the edited text, a list of questions, and a list of the changes you made.

The third constraint matters most: it takes the decision back from the tool. In dialogue it stops a model from rewriting how your characters speak — see Editing Dialogue and Quotations with AI.

The before-and-after comparison test

Three quick tests, none of which produces a number. The measure is your own ear.

• Read both versions aloud. The edited version is usually smoother; the question is whether your places to breathe are still where they were.
• Alternating paragraphs. One paragraph from the original, one from the edited version, back to back. If you can hear the seam, the tool has gone too far.
• The reverse test. Hand the edited version back and ask the tool to restore the original from the anchor sample. If it cannot, voice information was lost in the edit.
• The distance test. Put the edited version aside for a day and then read it. In that state you see the text itself, not your memory of the changes.

If you are still unsure, accept suggestions one at a time rather than in bulk. Rejecting one good edit costs less than accepting ten mediocre ones.

When to roll the whole thing back

You do not always have to fight for it. Sometimes reverting the whole text is the cheapest move available:

• The edited version has changed your position — not the sentences, the argument.
• You read a paragraph and do not recognise it.
• Two edits have drifted apart and neither sounds like you: your prompts carried no constraints.
• The tool has “explained” one of your comparisons and killed it in the process.
• The sentences are correct but the music of the text is dead.

When that happens, do not rewrite on top of the damage. Go back to the anchor sample and rebuild those paragraphs, and always keep an untouched copy. If the problem has outgrown the sentence — if the book is in the wrong shape — returning to structural editing is cheaper than fiddling with sentences.