Why AI writing doesn't sound like you
By Jack Stovell X · Instagram · LinkedIn · published 2026-09-20 · updated 2026-09-30
Because it does not sound like anyone. A model has no voice; it has a distribution, and every draft regresses towards the middle of it. The middle has a recognisable shape: sentences that all land between twenty and forty words, "furthermore" and "delve into", three items in identical grammar, a "not X but Y" somewhere in the second paragraph, an em dash where a comma would do. Readers feel it before they can name it. Detectors count it.
You, meanwhile, have habits. You contract or you do not. You open with a question one paragraph in four. You never use a semicolon and you put asides in brackets twice a page. None of that is in the prompt "write like me", so none of it survives.
Count the tells. Paste any draft below. It is scanned in your browser against the list ScriptGrain's own generator is forbidden to write, with every finding shown in its sentence.
The three reasons, in order of how much they matter
1. The target is a description, not a measurement. "Friendly, direct, a bit dry" describes a thousand writers. Your contraction rate, your sentence-length variation and your pronoun mix describe one. A model can hit a number; it can only guess at an adjective.
2. Nothing checks the draft. A human editor reads a draft and says "that's not us". A score does the same with numbers: formality 7.5 against your 4.8, sentence variation 0.35 against your 0.61. Without a check, drift compounds: each draft is edited a little less than the last, and the average wins.
3. The model never learns from what you changed. Every edit you make says something about your voice. In most tools that information evaporates. A profile that reads the difference between the draft and what you published gets closer with every piece.
What actually fixes it
Measure the voice (45 attributes from your own writing), hand the measurement to the assistant you already use, score every draft against it, and feed the edits back. The free tools on this site do the first two steps without an account: the checker below for the tells, the prompt builder for the measured prompt, the off-voice check for a score between two pieces.
What a model does when you say 'write like me'
It fills the gaps. That's the whole story.
Give a model the instruction "write like me, friendly and direct" and it treats that as a short list of constraints. Everything the list leaves out gets settled by the model's own defaults. Sentence length, how often you contract, whether you open with a question, where the punchline sits in a paragraph. None of that was specified, so the model picks what it always picks: the middle of its own distribution.
A description covers maybe four or five features of your writing. A measured profile covers 45. The other 40-odd attributes still have to come from somewhere, and in a bare prompt they come from the model.
An example makes it plain. Say you write short, blunt paragraphs and close on a dry one-liner. The model reads "direct" and produces a tidy, confident paragraph of four evenly sized sentences, then a neat summary sentence. Direct, technically. Yours, not at all. The adjective was honoured; the habits were ignored.
And to be fair, the model isn't being lazy. It's doing exactly what you'd expect from something with no memory of how you actually write. You asked for a vibe. It supplied one, plus a great deal of its own furniture.
Why a longer prompt does not fix it
So you add more. "Friendly, direct, a bit dry, uses British spelling, avoids jargon, short sentences, occasional humour." Better? Slightly. Then you paste in three samples and say "match these."
That helps, and it's worth doing. Samples carry real information that adjectives can't. But there's a ceiling, and you hit it quickly.
Adjectives stay vague however many you stack. "Short sentences" can mean eight words on average or fourteen, and a reader feels the difference even when you can't name it. Samples are better, but the model still has to infer the pattern from a handful of examples, and it will infer the obvious ones (topic, tone) while missing the statistical ones (how much your sentence length swings, how often a paragraph opens with a fragment). Those are the features detectors and readers both pick up on.
Then there's drift. A long prompt is read once, at the top. By the third paragraph of the draft, the model's defaults have crept back in. You can reinforce the instructions, but you're now maintaining a document whose only job is to approximate something that could simply be measured.
More words in the prompt means more guesses about you. A number isn't a guess.
How to test whether a draft sounds like you
Don't trust your ear alone. You've read the draft too closely, and by the second pass it always sounds fine.
Score it. A voice-match score runs from 0 to 1 against your measured profile. On-voice drafts typically land between 0.85 and 0.95. Partial matches sit between 0.5 and 0.8. A clearly different voice comes in under 0.45. One number tells you whether to keep editing or start again.
The number is only half of it, though. The useful part is the per-feature notes: for each attribute, the draft's value set against yours. That turns "this feels off" into something you can act on. A typical read might be:
- Sentence length: the draft averages longer than you do, and with less variation.
- Contractions: well below your rate, which is why it reads stiff.
- Formality: pitched higher than your profile, so the tone is slightly stern.
Each line points at a fix. Shorten, loosen, add a fragment or two. Re-score, and watch the gap close.
Do this and you stop arguing with a draft about how it feels. You fix the three numbers that are wrong, and the rest tends to follow.
Questions
Is this the same as an AI detector?
No. Detectors estimate whether a machine wrote the text. The checker counts specific phrases and sentence shapes that machines overuse, so you can remove them. Both can be wrong in the same way: a heavily edited draft reads clean on both without sounding like you.
If I remove the clichés, will it sound like me?
It will sound like nobody in particular, which is an improvement on sounding like a model. Sounding like you needs a measured target and a score against it, which is what a profile adds.
Does ChatGPT's memory or custom instructions solve this?
They store what you tell them about yourself, in prose. The habits that make a voice recognisable are mostly ones you cannot articulate, so they are not in the box. A measured profile puts them there.