Writing with AI without sounding like a

AI writing tools produce text that is fluent, organized, and weirdly nobody’s. Readers have learned the flavor. The fix is not to avoid the tools. It is to change where in the process you use them, and to edit for the specific patterns that give machine drafts away.

Why AI drafts sound generic

A language model predicts the most statistically likely next words across everything it has read. The output gravitates toward the average of all writing on your topic, and the average has no opinions, no stories, and no stake. What an LLM actually is explains the mechanics; the practical consequence is simple. The model supplies competence. You have to supply the person.

Feed it yourself first

The single highest-leverage move: give the model two or three samples of your real writing before it drafts anything. Emails you actually sent, a page you are proud of, anything that sounds like you on a good day. Ask it to match the voice in the samples.

This works because the model imitates whatever context you hand it. Given nothing, it imitates the internet. Given you, it imitates you, imperfectly but usefully. Keep a small file of your best samples and paste it at the start of writing sessions. The briefing techniques in prompting fundamentals apply here in full.

Draft in pieces, not in one go

Asking for a complete finished piece produces the most generic result, because the model fills every gap with average choices. You get better material by splitting the job.

First ask for substance: the points worth making, the objections worth answering, the examples that might fit. Argue with that list. Cut half of it. Then ask for a rough draft of one section at a time, built from the points you kept. You stay the editor-in-chief instead of becoming a subscriber to whatever arrived.

Edit for the tells

Machine drafts share recognizable habits. Editing for them is mechanical once you know the list:

Inflated significance: everything “marks a pivotal moment” or “underscores the importance.” Delete the ceremony, keep the fact.
The rule of three, everywhere, turning every sentence into a tidy triplet of nouns.
Balanced contrasts like “it is not just X, it is Y,” which sound profound and say little.
A suspicious evenness: every sentence the same length, every paragraph the same shape.
Vague authority: “experts agree,” “studies show,” with no expert and no study attached.

Your editing pass replaces these with the things only you have: a specific customer, a number from your own books, an opinion you can defend, a sentence that runs long because you got interested. Those are the details a model cannot fake, and they are exactly what readers trust.

Keep a human-only layer

Some writing should stay yours end to end, and not for sentimental reasons. Apologies, condolences, anything resolving a conflict, anything where the reader is asking “did a person actually consider me.” People can absorb a clumsy sentence from a human. They do not absorb discovering that their complaint was answered by a template with feelings.

A useful rule: AI helps with writing about things, you write to people.

The honest workflow, end to end

Collect your raw material yourself. Brief the model with samples of your voice and the specifics of the situation. Draft in pieces, keeping the points you chose. Edit for the tells, adding the details only you know. Then verify anything checkable before it ships, exactly as laid out in the verification habit, because fluent and correct are different properties.

The time AI saves you on typing is real. Spend a third of it putting yourself back into the text, and the result is faster than writing alone and better than the machine alone. That is the whole trade.