AI-isms: why AI text sounds robotic
Where “delve”, the em dash and the rule of three come from, and why they all turned up at once.
Somewhere around 2023, the word delve got a promotion. For centuries it had led a quiet life, mostly in the company of archaeologists and people describing handbags. Then, almost overnight, it was everywhere: in emails, in cover letters, in the abstracts of medical papers. Researchers who analysed more than 15 million biomedical abstracts found a sudden surge in a small set of style words after ChatGPT arrived, and estimated that at least 13.5% of 2024 abstracts had been through a language model.
Words don’t do this on their own. Something was pushing.
Why the average sounds like nobody
A language model writes by predicting the next word. Asked to describe a city, it reaches for the word that most often turns up in that kind of sentence, and the answer is bustling. This is a sensible strategy if you want to be right and a terrible one if you want to be interesting. Each choice on its own is reasonable. Stack a few hundred of them together and you get prose that reads as if it was written by a committee.*
Then the trainers had their say
After the predicting comes the polishing. Models are tuned on human feedback: people compare two answers and pick the better one, thousands of times over. Raters, being human, reward answers that look thorough. Headings. Bullet points. A friendly opener and a tidy summary. The model learns that structure gets applause, so structure is what you get, whether you asked for a shopping list or a eulogy.
It also picks up whatever quirks its raters had. One popular theory, put forward in the Guardian, is that delve came from data annotators in Nigeria, where formal English uses the word far more than British or American English does. Nobody outside the labs can prove it. It would explain a lot, though, and it’s a useful reminder that “neutral” writing always comes from somewhere.
The em dash problem
The em dash is the AI-ism people spot first. It’s a genuinely handy bit of punctuation. It can stand in for a comma, a colon or a pair of brackets, so a model that isn’t sure which to use can use it for all three and never be strictly wrong. It’s also common in the edited books and journalism that models learn from, where typesetters put them in properly. Most people, typing on a keyboard with no em dash key, don’t bother. So the dash became a tell, not because it’s wrong, but because it’s too well dressed for the occasion.†
Threes, and other habits
Models love lists of three (fast, flexible and fun), sentences built as it’s not just X, it’s Y, and questions that answer themselves. These are old rhetorical tricks. The trouble is how often they appear. A speechwriter uses the rule of three once, for effect. A model uses it every paragraph, because every time it worked before, somebody pressed the thumbs up.‡
So what do you do about it?
Swapping delve for “explore” doesn’t help. It just moves the cliché one square along the board. The fix is the boring one: read the draft, find the bits that sound like nobody in particular, and rewrite them the way you’d actually say them. That’s what Sounds like me is for. It points at the tells and leaves the rewriting to you, because you’re the only one who knows what you sound like.
- * A committee is the only creature known to have many heads and no opinions. It is also the natural habitat of the word vibrant. ↩
- † It got bad enough that in late 2025 Sam Altman celebrated on X that ChatGPT would finally stop using em dashes if you asked it to. Celebrating a machine for following instructions is an odd sort of victory, but it is a victory. ↩
- ‡ This is also how dogs learn to sit, and how some managers learn to say “circle back”. ↩