Ask three people to get a model to write something. One wants a launch post, one wants a note to a manager, one wants a product page. Different subjects, different models, different prompts. Then read the three drafts side by side.
They rhyme.
The same verbs show up. Something is always being unlocked or streamlined or empowered. The sentences run to about the same length, one after another, for paragraphs. Each piece opens by telling you what it is about to say, and then says it. None of the three is bad. All three are recognizable, and once you have noticed it you cannot go back to not noticing.
That sameness is what people mean when they say a draft sounds like AI. It is not error. Models are good at grammar and they are good at facts. What they are not good at is being anyone.
Where the sameness comes from
A base model trained on a large pile of text can write in a great many ways, including badly. That range is a problem for a product, so the model gets tuned. People rate outputs, the ratings train a reward model, and the reward model pulls the base model toward whatever the raters liked.
Think about what a rater does. They read two answers in twenty seconds and pick one. Under those conditions people reliably pick prose that is clear on the first pass, evenly paced, hedged rather than committed, and organized so the point arrives up front. Those are good instincts for a rater. They are also a narrow target.
Do that at scale, across every kind of writing, and the model converges on a house style. Not a bad style. A style with no floor and no ceiling. The words it reaches for are the ones that survive that filter: safe, formal enough for any context, slightly elevated, never wrong.
So utilize beats use, because it reads a little more careful. Robust
beats a number, because a number can be checked. It's important to note
survives because it makes a sentence sound considered. Every one of those is a
small win in a twenty-second comparison, and a small loss in a real document.
What that leaves behind
We keep a list of what survives. It sits in one file, it has a version number, and the site prints it in full: 97 words in seven categories, eight sentence shapes, twelve stock openers.
The words are the easy part. Anyone can search and replace delve.
The shapes are harder, and they last longer. A model loves a contrast it never
earned. It's not just a tool, it's a way of working. The question isn't whether, it's when. Strip the construction out and see what is left, and most
of the time what is left is one plain claim wearing a costume. Those eight
shapes survive a vocabulary pass, which is why a cleaned-up draft can still
read generated.
Then there are three things you cannot see in a single sentence at all.
Em dashes per thousand words. Sentence length variance. The share of sentences under ten words. None of those is visible in a line. All of them are obvious across a page, and the second one does more work than the other two put together. Human writing is bursty. Short sentence, short sentence, then one long one carrying a list, then a short one that lands. Model writing is metronomic. Twenty-two words, twenty-one words, twenty-four words, for six paragraphs, and the effect on a reader is closer to a hum than to a voice.
You can write every sentence cleanly and still fail on rhythm alone. Most cleaned-up drafts do.
Why this is a credibility problem
Nobody you work with runs a detector on your email. They read it, and something lands wrong, and they do not always know why. Then they read the next thing you send with a little less attention.
That is the actual cost, and it is worth being precise about it. It is not that you used a model. Most people are using a model. It is that the draft arrived carrying somebody else's habits, under your name, in a document where your name was the point.
What to do about it
Read your draft out loud and mark every sentence over about twenty words. Then cut half of them in two. That single pass does more than any word swap, because rhythm is the tell you cannot fix with search and replace.
After that, take out the openers. Anything that spends a whole sentence getting ready to say something can go, and the sentence underneath is unharmed. Then go after the words, which is the part everyone starts with and the part that matters least.
Keep the facts exactly as they were. A draft that admits something hard should still admit it when you are done. A rewrite that quietly softens a commitment because a shorter sentence read better is worse than one that sounds synthetic, and it is a much harder problem to notice later.
The words change. The list is version one and it will be wrong within a year, because the next round of tuning will land somewhere else and take a different set of habits with it. The measurements will still be there. Whatever a model is trained to sound like, it will sound like it consistently, and consistency is the thing a person does not have.