Somewhere today, a writer is being told a sentence sounds like ChatGPT. Shakespeare wrote that sentence four hundred years ago.
“The fault, dear Brutus, is not in our stars, but in ourselves.” “Not that I loved Caesar less, but that I loved Rome more.” “I come to bury Caesar, not to praise him.” Three of the most quoted lines in Julius Caesar use the same move: tell the reader what the thing isn’t, then tell them what it is. Post those lines in a writing group today without the attribution and see how long it takes for somebody to call them AI slop.
On its face, it’s a perfectly normal way to write. A lot of nonfiction is built that way, by human beings who never touched a chatbot. It’s a shame that a valid construction got used so heavily by AI that anything written with it now gets an “oh my God, that’s AI” from people who should know better. I don’t think it’s a bad habit. I think it’s a tool that got worn out by a machine that can’t stop picking it up.
What is the “It’s not X, it’s Y” AI tell?
It’s a sentence that defines something by first saying what it isn’t. The Atlantic ran a long piece on the construction in July 2026 and called it perhaps the best-known tic of AI writing.
It comes in a few flavors. In one, the second half adds to the first: the result was a win for the department, and a win for the whole company. In another, the second half replaces the first, the way a corrected witness statement would. In AI-written fiction there’s a third version that shows up constantly: no bag, no things, no armor, just me.
Nobody even agreed on a name until recently. Academics have terms like antithesis that cover some versions and miss others. A product manager at OpenAI called it “contrastive phrasing.” The name catching on is negative parallelism.
Did writers use “it’s not X, it’s Y” before AI?
All the time. Vince Lombardi made “Winning isn’t everything; it’s the only thing” famous in the 1960s. In the 1990s, a frozen pizza brand built an entire ad campaign on “It’s not delivery. It’s DiGiorno.” Speechwriters, preachers, coaches and copywriters have leaned on it for centuries because it works. It sets up an expectation and then knocks it over, and readers remember the knock.
The best-known line from Martin Luther King Jr.’s “I Have a Dream” speech uses it too. His hope that his children would be judged “not by the color of their skin but by the content of their character” is the same move. When the construction came up on Reddit, one commenter guessed the top reply to that line today would be “Okay ChatGPT.” Another pointed out that Faulkner built most of Absalom, Absalom! out of the same move, and nobody would mistake Faulkner for a chatbot.
Chatbots learned it for the same reason. They were trained on an enormous pile of human writing. That pile is full of the construction, from Shakespeare down to the back of the cereal box.
Shakespeare used “not this, but that” in three of the most quoted lines in Julius Caesar. Somebody today would run them through an AI detector. – Richard LoweShare on X
Why do chatbots overuse negative parallelism?
Nobody knows for certain. Apparently that includes the companies that build the models. The explanations researchers offered The Atlantic come down to three.
Start with training. After a model learns from raw text, human reviewers grade its answers, and the model drifts toward whatever scores well. Tuhin Chakrabarty, a computer science professor at Stony Brook University who studies AI writing, said it’s plausible reviewers rewarded the construction because it looks like insight. The machine appears to reason its way from a weak description to a better one, even when there was no reasoning involved.
The second is mechanical. A chatbot writes one chunk of text at a time, always weighing the obvious next word against the clever one. Once it starts a sentence describing something, a negation is a safe next step. The boring descriptor slots in after it, and the punchier one closes the sentence.
It’s a hedge dressed up as a flourish. One Reddit commenter called it the “uhhhm” of an LLM, except the model is forced to turn it into a whole sentence.
Another commenter, who builds AI characters for a living, offered a related explanation. Train a model to be agreeable, safe and clear for millions of different people, and the blandest, lowest-risk phrasing wins. The tidy rule of three and the “it’s not X, it’s Y” framing are what the safe middle sounds like.
The last explanation is the ugliest. New models train partly on text written by older models, and more of the internet is machine-written every month. Elyas Masrour, a founding engineer at the AI-detection company Pangram, said that once something gets into these models, it’s very hard to pull out. Chakrabarty called it a vicious loop. The machines are learning to write from their own leftovers.
How common is the construction in AI writing?
Common enough to measure.
Barron’s, working from AlphaSense’s database of press releases, earnings reports and filings, found the additive version of the construction more than quadrupled in corporate documents, from about 50 uses in 2023 to more than 200 in 2025. Pangram estimates that same additive version shows up three times as often in AI writing as in human writing, and Masrour said every major chatbot leans on it to some degree.
Other tells come and go. Some of the chatbots’ favorite words faded as the companies tuned their models. This one hasn’t budged, and the detection companies are happy about it, because a stubborn habit makes their job easier.
Can negative parallelism get a writer accused of AI?
Yes, and it’s already happening.
When Hachette pulled the horror novel Shy Girl in March 2026, the AI suspicion had started with readers on Goodreads and YouTube, and lines in the “no bag, no things, just me” pattern were part of what they pointed at. The author, Mia Ballard, denied using AI herself and said an editor she’d hired had introduced the material. I don’t know what happened with that book, and neither does anybody yelling about it online.
A sentence pattern is evidence of nothing. Plenty of humans write that way on their own, and plenty of AI output never uses it. A stranger who reads one “not this, but that” sentence and announces the whole book came out of a chatbot is guessing, and calling a writer a fraud on a guess is a rotten way to treat a person. I’ve called that reflex AI Derangement Syndrome: the moment AI stops being one fact about a book and becomes the only fact. A sentence shape is a flimsier excuse than most.
Writers feel it already. In the Reddit thread, a man who said he’s been an avid reader for about sixty years reported being dismissed as AI for writing in whole, grammatical sentences. A professional writer with two decades in the business said he now runs his own drafts through an AI just to flag the devices readers might mistake for a bot, so he can decide whether to change them. He called it maddening.
I’d call it backwards.
Somebody else in the thread brought up an old tailoring tradition: only the worst apprentices were allowed to stitch the lapels, because perfect stitching looked machine-made to the customers. Writers are starting to do the same thing to their prose. Writers who rough up good sentences so a stranger won’t call them a bot have taken the craft somewhere ridiculous.
The detectors don’t settle it either. I’ve written about why AI detection can’t be treated as evidence, and a pattern a detector loves to flag is exactly the kind of thing that catches innocent writers.
Should writers stop using negative parallelism?
No. I don’t think it’s a bad habit, as long as it isn’t used too much.
Plenty of good nonfiction runs on it. A writer correcting a common belief, a memoir pushing back on the family story, a business book taking apart a popular piece of advice: those are places where telling the reader what something isn’t does real work. Readers came in believing the X. You have to knock it down before they’ll hear the Y.
The fiction version is a different animal. A string like “no bag, no things, just me” can land once in a novel, at a moment that deserves the drumroll. Three of them in a chapter and the rhythm turns mechanical, and a reader who has been reading machine prose will notice before you do.
Frequency is the problem with the chatbots. One of these in an essay is a choice. Five in a page is a machine with a tic. If you like the construction, keep it for the sentence where you’re correcting a real misunderstanding your reader probably has. Then check the rest of the page.
If it shows up twice more, those two are padding.
One commenter remembered a writing workshop that taught a simple rule: delete the “it’s not” half and keep the “it is” half. Use negatives sparingly.
That’s old advice, from long before chatbots, and it still holds.
A quick test works on most of them. Cover the “it’s not X” half and read what’s left. If the sentence still says everything you meant, the first half was throat-clearing, so cut it. If you need the contrast, keep it and don’t apologize. Lombardi didn’t.
I’ve thinned them out of my own site copy anyway. The construction is fine. Readers have been trained to flinch at it, and a writer who wants to be read has to know what makes readers flinch.
That’s a shame, and a lousy reason to give up a good tool.
If you’re cleaning up a draft that went through a chatbot, this construction belongs on your list along with the rest of the usual suspects. I laid out the whole process in the cleanup protocol for AI fingerprints, and what to do when your book sounds like AI covers the bigger picture.
Are people starting to write and talk like AI?
Some of them are. Researchers at the Max Planck Institute for Human Development analyzed more than a million YouTube videos and podcast episodes and found words chatbots favor, like “swift” and “inquiry,” turning up more often in spoken English after ChatGPT arrived. People pick up the habits of whatever they read most, and a lot of people now read a lot of machine text.
If that keeps going, the tells will stop being tells. Half the population will write “it’s not X, it’s Y” because they read it four hundred times last month, and the detectors will have to find something else to flag.
Leave Shakespeare alone
The construction didn’t change. The world around it did. A device that carried Mark Antony’s funeral speech is now a warning sign in a Facebook thread, and the people paying for that are human writers who learned to write before any of this existed.
Write the sentence when it’s the right sentence, and don’t write it six times a page.
And the next time you see a familiar shape in somebody else’s prose, read all of it before you decide a machine wrote it. The AI and writing hub has more on keeping your own voice when these tools are everywhere, and if you want help figuring out where AI belongs in your work, that’s what my AI services are for. The fault, dear reader, was never in the sentence.
