The Bright-Line Rule for Using AI to Write Without Sounding Like AI

If you're using AI to draft anything you'll put your name on, the workable rule fits in one sentence: write it yourself first, then use AI only to flag what's broken — never to supply the fix.

What changed and why it matters

On September 17, Simon Willison highlighted an essay by Thomas Ptacek with a rule blunt enough to actually follow: "You may not use a single word an LLM suggests to you." Willison said he holds himself to the same rule on his own blog. The workflow is narrow on purpose — write the draft yourself, hand it to AI to flag mechanical problems like passive voice, filler, or a paragraph doing two jobs, then fix those problems in your own words.

Two days later, a Hacker News post arguing professionals should "almost never use AI to write anything substantive" hit 316 points, with commenters split over whether outsourcing the writing outsources the thinking behind it. The same week, The Neuron reported a bind that makes the rule more than a style preference: managers increasingly distrust polished writing on sight, suspecting AI even from people who write well on their own. AI has made average writing look better and good human writing look suspicious, in the same motion.

Research published in Science on September 18 gives the strongest reason to hold the line if you do let AI write for you: the same techniques that make AI-generated arguments more persuasive make them systematically less accurate. Persuasive and truthful moved in opposite directions in the study, not together.

What to do differently Monday morning

None of this argues for keeping AI out of your writing process. It argues for exactly where in the process it belongs — after a draft exists, not before it. Write first, however rough. Then ask AI a narrow question: where does the argument sag, where's a sentence hiding behind the passive voice, where's a paragraph trying to do two jobs at once. Fix what it finds yourself, in your own sentences. What you never do is paste its rewrite back in.

A communications manager at a 60-person nonprofit had been sending every board update through ChatGPT for a full rewrite before each meeting, then wondering why a board member started asking if she'd "had someone write" her reports. Nothing in them was wrong — the sentence rhythm was just a shade too even, the phrasing a shade too generic for a reader used to her voice. She stopped asking for rewrites and started asking only "what's unclear here" and "where does this run long" — flags, not fixes. The next report still took her the usual forty minutes. It came back sounding like her again, and nobody asked about it a second time.

A sales director at a 25-person agency had the opposite worry: his cold outreach had gotten flatter since he'd stopped leaning on AI to punch it up. His account of what changed once he switched to flag-only use: he asked AI to point out where a sentence buried the actual ask, then rewrote that one sentence three or four ways himself until it felt right. The emails got shorter, not more polished, and his reply rate held steady over the following month rather than sliding the way it had when he was accepting AI's suggested phrasing outright.

The one thing

The Science finding is the sharper warning for the opposite habit: if you let AI phrase the persuasion for you, the more convincing draft is the less accurate one, and that trade doesn't show up on a read-through.