The first time I asked a model to polish a draft, it handed me back something technically cleaner and noticeably less mine. Shorter sentences. Fewer em-dashes. A weird politeness in places where I had been deliberately blunt. Nothing was wrong with it. Nothing felt like me.

I have since worked out what happened. The model trained on an enormous range of human writing, and it learned what "correct" and "clear" look like across that range. Aggregate correctness is smooth. Smooth is not the same as good. It is the same as average — and average writing, however technically sound, has no character.

The solution is not to stop using the model for editing. It is to use it more precisely.

The three rules

Rule one: never ask to "improve." The word "improve" is where your sentences go to die. Improve toward what? The model's answer is: toward the aggregate. Its corrections pull every sentence toward the center of the distribution. Some of your sentences live at the center and should go there. The ones that don't are the ones that make your writing yours. They will be smoothed first.

Ask instead to "preserve voice, fix only errors of fact or grammar." Be specific about what you want caught: missing antecedents, comma splices, claims you can't support. Not "clarity." Not "flow." Those are invitations for the model to rewrite in its own voice.

Rule two: paste examples before you ask for anything. One or two short passages of your own writing, from something you like, at the top of every editing session. The model mimics what's in front of it before it mimics its defaults. Three paragraphs of your best work shifts the reference frame toward you. This sounds too simple to work. It works reliably.

Rule three: read every output aloud before accepting it. Your ear catches the beige before your eye does. There is a particular cadence to model-edited prose — a kind of measured brightness, sentences that flow without surprise — and you can hear it before you can articulate it. If you are reading back something that sounds like it could have been written by anyone, the model has done too much. Go back to the previous draft.

The best use of AI in writing is the same as the best use of an editor: fewer words, same voice.

What "voice" actually means

Voice is not style. Style is learnable — sentence length, paragraph rhythm, the choice between active and passive. Voice is what remains when you write about something you genuinely understand or care about. It is the byproduct of specificity and conviction, not of technique. You cannot install it. You can only bring it.

The model cannot give you conviction. It can give you structure, clarity, and volume. When conviction is already there, the model becomes genuinely useful: it can find the overlong paragraph, the repeated word, the section that belongs in a different essay. When conviction is absent, the model fills the vacuum with its own defaults, and those defaults are beige.

This is why the most important editing decision you make is not about the model at all. It is about whether you had something to say before you opened the tool. If you did, the model is a fast and useful editor. If you didn't, it will produce something technically proficient and completely forgettable.

A practical kit

Beyond the three rules, a few specific habits have helped:

Never let the model touch the opening sentence. The opening is where voice is most concentrated and most fragile. Edit it yourself, always. Once the opening is right, the model is less likely to drift the rest.

Ask for cuts, not rewrites. "What could be removed without losing the argument?" is a much safer instruction than "make this tighter." Cutting preserves voice. Rewriting replaces it.

Keep a "voice file." A document with five or ten short passages that sound unmistakably like you. Paste it at the top of any session that requires significant editing. The model will use it. Update the file every few months, as your voice evolves.

Run a diff. When you have the model's edit, compare it to your original side by side. Every sentence that changed is a choice the model made. Review each one. Reject the ones that smoothed something intentional.

The model is a powerful editor. It needs a strong writer to edit. Be that writer first, and the model becomes very useful. Expect it to make you a writer, and it will make you sound like everyone else.