Most AI drafts arrive the same way: clean grammar, sensible structure, and a voice like a careful stranger who has never met you. The grammar was never the problem. Little in the draft could only have come from you.
The fix is a division of labor we call the 90/10 rule: ChatGPT handles the 90 percent that's sentences and structure, and your effort goes to the 10 percent that makes a piece yours. This post covers where the robot sound comes from, how to brief the draft, and the five edits that close the gap.
Where the Robot Sound Comes From
ChatGPT writes by predicting what plausibly comes next. Plausibly is the operative word: the most likely next sentence is also the most average one, and that average is what the robot voice is made of. A request to "sound professional" tends to pull the draft toward familiar language, not away from it.
Researchers have measured it. One 2024 analysis of PubMed abstracts found that words like "delve" spiked so sharply after ChatGPT's release that the surge couldn't be explained by the words' earlier trajectory (Kobak et al., 2024). Millions of prompts, pointed at the same training data, keep reaching for the same favorite words.
A tone instruction doesn't fix this. It can make a draft warmer, shorter, or more formal, but it can't add the experience behind the piece. Specific writing comes from specific material, which is what the next two sections cover.
The 90/10 Rule for AI Writing
The 90/10 rule: ChatGPT produces the first draft, roughly 90 percent of the words. You supply the 10 percent that turns it into your writing: the specifics only you know, the decisions only you can make, and the edits that make it sound like you.
The exact ratio matters less than where the effort goes. Drafting got cheap; judgment didn't. What to say, what to leave out, what sounds true: that part is still yours.
How to Brief the Draft
A generic draft usually starts with a generic brief: "Write a LinkedIn post about our new product." That gives ChatGPT almost nothing specific to work with, so it invents from the average. Hand it the rough material instead:
- What happened, in plain words ("we rebuilt onboarding after watching five users get stuck on step 3")
- Who's reading it and what they should do after ("candidates deciding whether to apply; reply to the post")
- Anything you'd say out loud ("honestly, the old flow embarrassed us")
Bullet points, half sentences, even a rambling voice-memo transcript work. The more of you goes in, the less averaging the model does on the way out. Our post on how to brainstorm with ChatGPT covers the same move from the other side: have ChatGPT interview you before it produces anything.
Edit the 10 Percent
The draft comes back, and the five passes start.
-
Add what only you know. Numbers, names, dates, what surprised you, what it cost. ChatGPT fills those gaps by inventing plausible details, so the specifics are your job. Compare "We learned that onboarding matters" with "We rebuilt onboarding after watching five users get stuck on step 3." The second carries a decision, a number, and a moment of failure, and no tone instruction could have added them, because they came from the work.
-
Cut the echoes. AI drafts open by restating the topic and close by restating everything: "In conclusion," "Ultimately," "It's important to note." Delete the summary paragraph unless it adds something the body didn't.
-
Replace the stock phrases. When a sentence could appear in anyone's draft ("In today's fast-paced world"), replace it with how you'd say it to one person. Reading aloud often catches these faster than scanning, because your mouth stumbles where the words aren't yours.
-
Break the rhythm. Robot drafts have uniform sentences: medium, medium, medium. Combine two. Split one in half. Varied sentence length is one quick way to make a draft feel less mechanical.
-
Add one opinion. "The free tier is fine for this" or "I'd skip the video version." A stance is something the average draft tends to flatten.
What About AI Detectors?
Worth knowing, briefly: detectors are unreliable in both directions, so chasing them is the wrong target. OpenAI released its own AI-text classifier in early 2023 and retired it that July, citing a "low rate of accuracy" (OpenAI, 2023). A study in the journal Patterns found that popular detectors misclassified over half of essays by non-native English writers as AI-generated while passing US eighth-graders' essays with near-perfect accuracy (Liang et al., 2023).
The practical version: don't write to defeat scanners. If a school or employer has rules about AI use, follow them and disclose where required. Write so the piece contains you, and the detector question tends to shrink on its own, because idiosyncratic detail gives detectors less of the pattern they're designed to identify.
FAQ
Should I ask ChatGPT to imitate my writing style?
You can paste writing samples and ask it to match your tone, and it helps a little. Tone is the surface half of voice, though: a mimicked tone wrapped around average content tends to read generic. Specifics do more work than style matching.
How much of an AI draft should you rewrite?
There's no fixed percentage. A draft is ready when every claim is true, the specifics are yours, and nothing in it could have appeared in anyone else's version. Low-stakes text might need light edits; public writing can mean restructuring most of the piece.
Is it cheating to use ChatGPT for writing?
For your own work, it's a tool, like asking a colleague to sketch a first draft. For assignments or contests with originality rules, check those rules first and disclose where they require it. Keeping the judgment with you is what keeps it on the tool side of the line.
Will ChatGPT replace writers?
It has already absorbed part of the job: the typing. What it hasn't replaced is deciding what's worth saying, verifying claims, and editing for a specific audience. Writers who draft with it and keep the judgment tend to produce more, while generic, interchangeable writing is the part likely to lose value as drafts get cheaper.
Want to go further? How to brainstorm with ChatGPT (and get good ideas) covers the interview-first technique, and how to write better ChatGPT prompts without prompt engineering covers the conversation-first approach to briefs.