Writing With AI Without Losing Your Voice

The complaint about AI-assisted writing is rarely that it is wrong. It is that it is bland — fluent, well-organised, and completely anonymous. Every paragraph is defensible and nothing sounds like a person.

That matters if writing is part of how you build a reputation. Here is what I have found actually preserves a voice.

Why the default output is flat

A language model produces the statistically typical continuation. Voice is, almost by definition, the atypical continuation — the unexpected word, the sentence that runs long because the thought does, the willingness to state a position without hedging.

Generated prose regresses toward the middle. That is not a defect to be prompted away; it is what the tool does.

What actually works

Write the argument yourself, always

The claims, the structure, the position — those are the writing. If you outsource what you think, the result cannot sound like you, because there is no you in it.

Where assistance helps is after the thinking: tightening a paragraph that will not sit right, finding the word on the tip of your tongue, cutting three hundred words to two hundred.

Use it as an editor, not an author

The most valuable prompts are diagnostic rather than generative. Which paragraph is weakest? Where am I repeating myself? What would a sceptical reader object to here?

This gets you the benefit of an outside reader without handing over the prose.

Reject the tells

Certain constructions appear constantly in generated text and rarely in good human prose. Sentences built on “not only X but also Y.” The reflexive triad where two items would do. Opening with “In today’s rapidly evolving landscape.” Closing paragraphs that summarise what was just said without adding anything.

If you keep AI-assisted passages, strip these deliberately. They are the fingerprint.

Keep your specifics

What makes writing yours is usually the particular: the project that went wrong, the number you measured, the thing you believed until it turned out otherwise. Models cannot supply these, and generated text papers over their absence with generalities.

When a paragraph feels generic, the fix is almost never better phrasing. It is a concrete detail only you have.

Read it aloud

The most reliable detector available. Anything you would not say out loud does not belong. This catches the subtly-not-you sentences that survive every other check.

The failure mode to avoid

The insidious pattern is not wholesale generation. It is accepting small suggestions repeatedly — a smoother transition here, a tidier sentence there — until after twenty edits the text is grammatically better and no longer sounds like anyone.

Each individual change was an improvement. The aggregate was a loss. Guard against this by re-reading whole sections rather than judging edits one at a time.

Where I draw the line

Administrative writing: use it freely, nobody’s voice matters in an ethics application. Technical documentation: use it, clarity beats personality. Anything published under my name where the point is what I think: write it myself and use AI only to find weaknesses.

The test I apply before publishing anything: could someone else with the same brief have written this? If yes, it is not finished.