I asked ChatGPT to help me write about something that’s been bothering me when I read.
I keep noticing phrases that sound AI-generated, then second-guessing myself because people have always written that way. The draft came back with this heading:
We learned the accent, not the species.
A little further down:
Style is evidence. It isn’t provenance.
It also contained an entire section complaining that AI writing tries to make every paragraph end with a quotable line.
I replied:
look at how hard you do it in this blog post
It agreed, identified the patterns it had overused, and suggested making that failure part of the article. So that’s what we’re doing. ChatGPT is helping draft this version too.
I don’t particularly care whether someone used AI to write a post. Some of the habits I associate with it still irritate me, especially when a thin observation is made to sound important. Good writers use the same techniques, though, and AI might help someone express an idea better than they could have on their own. I’m not always sure I would have objected to the same sentence if I hadn’t known where it came from.
I might also be learning something about writing myself. ChatGPT gives me several versions of a sentence, and I start noticing what changes: the rhythm, the emphasis, the way a contrast works. Maybe I used to read straight past these techniques. Now I recognize them elsewhere and think “AI,” because that’s where I started paying attention.
The patterns have names
The construction I notice most is some variation of this:
Great engineering isn’t about writing more code. It’s about solving the right problems.
That came from the earlier draft. Sometimes I’m halfway through an article and realize I’ve spent the last few paragraphs inspecting the prose rather than following the argument.
The device is called antithesis: putting contrasting ideas alongside each other, often in a parallel structure.
Another passage in the draft went:
Every paragraph has a lesson.
Every lesson has a contrast.
Every contrast comes in balanced clauses.
Repeating the beginning of successive clauses or sentences is anaphora. Here, the model was using a conspicuous rhetorical pattern to complain about conspicuous rhetorical patterns.
Before large language models
Take the opening of A Tale of Two Cities:
It was the best of times, it was the worst of times, …
It makes sense that models trained on human writing would pick up these techniques. Their presence alone doesn’t explain why I like one sentence and find another annoying.
Sometimes I wonder whether what I’m calling AI writing is just someone using English well. AI may also have made some good writing harder for me to enjoy. A contrast might help explain an idea and still annoy me because I’ve seen the construction so many times.
The engineering advice could be useful in a conversation about measuring productivity. On its own, it doesn’t explain what the “right problems” are or how to choose them. I’m left with a confident-sounding sentence and nothing specific to use. Being able to say that feels more useful than just calling it AI writing.
That complaint has a history too. In his 1946 essay Politics and the English Language, George Orwell criticized the habit of reaching for ready-made phrases instead of choosing words for their meaning. He also acknowledged doing it himself, an admission that feels familiar after the exchange that started this post:
I have again and again committed the very faults I am protesting against.
When readers hear AI
In a July 2026 Hacker News discussion, dang described readers developing a sensitivity to language that sounds generated, then discounting a piece once they make that association. Further down, one participant accused another of posting an AI-written comment. The other person denied it. I don’t know who was right.
Some readers can be very accurate. A 2025 study selected five experienced users of language models who performed well on an initial detection task. Their majority vote misclassified just one of 300 articles across tests that included paraphrasing and humanization. The experiment tested nonfiction articles; it doesn’t establish the same result for a comment that a person wrote and a model helped edit.
I don’t have a running score for my own guesses. Usually, I never learn whether I was right. I recognize the irritation in that thread. By the time I notice I’m looking for more evidence, I’ve sometimes lost track of the argument.
What the draft added
The model also supplied interpretations that went further than what I meant. One heading announced:
There is no human writing style.
I had come in saying that I kept getting distracted while reading blog posts. We were now apparently settling a question about all human writing. The heading made the argument look more settled than it was.
Making that heading more casual wouldn’t have fixed it. An “I think” could make it sound tentative while leaving the same sweeping idea in place. I had to decide whether the claim belonged in the article at all.
That bothers me more than an em dash. I could remove every familiar tell and still publish something that makes my thinking look tidier than it actually is.
The less-AI version
Raymond Queneau retold the same incident ninety-nine ways in his 1947 book Exercises in Style.
Here are three ways the model framed the problem, all generated for this post.
AI isn’t just changing how we write. It’s changing how we read.
Then a more explanatory version:
Since I started using ChatGPT, I spend more time wondering whether the things I read were written with it. Sometimes I notice a phrase and lose track of what the article was saying.
And a casual one:
idk if anyone else does this but i keep getting halfway through a post and wondering whether it’s AI. then i remember people have always written like that and i have no idea anymore
The first turns the observation into a general claim. The second stays with what happens while I’m reading. They aren’t quite saying the same thing, although I could easily treat the choice between them as a matter of style.
The casual version sounds close to how I first described the problem. I really was unsure, and the model is expressing that uncertainty in words I might have used. I can recognize myself in it without having written those words. I still have to decide whether they explain the experience well.
More believable
When I complained about the draft, ChatGPT proposed a plainer opening and described it this way:
That already feels more believable to me because it isn’t desperately trying to earn its place in a quote card.
“More believable” bothered me. I wanted help explaining something I’d experienced, and we were discussing how to make it sound as though I’d experienced it.
When I prefer the plainer version, I’m not always sure what I’m responding to. It might explain the experience better, and I might also be relieved that a phrase I’ve learned to distrust is gone. I can compare what the sentences say, but that doesn’t tell me how much each reaction influenced the edit.
The next time I write without ChatGPT, I might choose a contrast because it helps me explain something. I could understand why the sentence works, be glad I wrote it, and still get annoyed at how much it sounds like ChatGPT.