Like Julius Caesar had it started this year, William Shakespeare might have been accused of writing it by AI. A certain questionable rhetorical device appears again and again in the play. It’s in Act I, Scene ii: “The fault, dear Brutus, is not in our stars, but ourselves.” In Act III, Scene II: “Not that I loved Caesar less, but that I loved Rome more.” And later in the same event: “I have come to bury Caesar, not to praise him.”
These famous lines include what has become the most well-known tick of AI writing – the sentence that tells you the topic. it is not as well as are: Not X; is Y. Once you start seeing construction, you see everywhere. In one version, Y is complementary: It highlights, strengthens, or expands X. Citizens Financial Group’s annual review reported that growth in its private banking division “wasn’t just a win for the private bank—it’s a win for the entire business.” In another variant, Y replaces X as the preferred explanation. “The target was never a person. The target was the truth,” Michael Flynn, a former adviser to Donald Trump, wrote in Post of March X.
Then there is construction like No A, no B, only Cwhich seem to be popping up in AI-generated fiction. Lines like “No bag, no stuff, no weapon, just me” helped add fuel charges of AI to write in a horror novel The Shy Girlwhich was pulled by its publisher this year. (The book’s author denied using AI to write it. Citizens Financial Group has previously said that its communications team “uses technology in several areas.” Flynn did not respond to a request for comment.)
The prevalence of this device isn’t just a myth—it’s measurable. (Sorry.) of Barron reported that his appearance in company communications more than four times from 2023 to 2025. Researchers at Pangram, which develops an AI detection tool, estimate that Not just X but Y sentences appear three times as often in AI writing as they do in human writing. Elyas Masrour, a founding engineer at Pangram, told me that all major chatbots—including ChatGPT, Claude, Gemini, and various open-source models—are dependent on each other to varying degrees.
Many other well-known chatbots such as the use of investigate-have come and gone as AI companies improve their models and work on the bad stuff. Last autumn, ChatGPT was haunted by goblins and gremlins, which led to another intervention: OpenAI removed the “nerdy” personality of OpenAI, whose connection to mythical creatures had infected its other structures. Not yet Not X; is Y has shown no signs of slowing down.
Before ChatGPT came along, the construct was so vague that it had no agreed name. Now there is a scramble for what to call it. Terms from scholars, such as the opposite and Metalinguistic denialcapturing some construction types but not others. In an email, Laurentia Romaniuk, product manager for prototype behavior at OpenAI, referred to it as “conflicting rhetoric.” Despite its subtlety, the most popular name I’ve seen is “negative parallelism.”
When used wisely, bad balance can be difficult. But ChatGPT turns to it too often, Romaniuk admitted, which can feel formulaic. So the company is working on ways to expand the chat track. At the same time, he added, users can try to give ChatGPT “special instructions.” On a Reddit forum about AI writing, users exchange tips for analyzing negative equity from chat text. One he suggested pasting Claude’s output into another AI chat and telling it to act like a copy editor that has a strict ban on “negative pairings” like “it wasn’t X, it was Y.”
One obstacle to a more comprehensive adaptation is that no one seems to know for sure why AI models are so interested in negative equity in the first place—perhaps not even the companies that created them. (Anthropic and Google did not respond to my requests for an interview.)
The simplest theory is that humans trained them that way. Large language models are built by first identifying patterns in an infinite number of human-written texts: books, academic papers, patent filings, and especially the Internet. Negative correlation was present, of course, in the original training data. Shakespeare aside, there are many famous examples: In the 1960s, legendary football coach Vince Lombardi popularized the saying that “winning is not everything; it is the only thing.” In the 1990s, ads for the frozen pizza brand insisted: “It’s not delivery. It’s DiGiorno.”
But the training data also included a lot of bad text that AI companies don’t want their chatbots to imitate, Tuhin Chakrabarty, a computer science professor at Stony Brook University who studies AI coding, told me. So they also undergo “reinforcement training,” a process in which human examiners mark examples of their responses. Through trial and error, chatbots are steered away from inappropriate responses (making things up, giving illegal advice, insulting the user) and towards those rated as useful. Chakrabarty said that it was true that human reviewers tended to give higher marks to answers that included them. Not X; is Y. That may be because negative parallelism gives the impression of nuance and insight: AI seems to be reasoning its way from the smallest to the most relevant details.
That still may not be enough to explain what construction looks like in the main types of AI. Several experts I spoke to pointed me to other, even more mysterious, explanations.
Although chatbots have advanced greatly in their research and reasoning capabilities, they are still essentially text prediction machines. It gives answers one “sign” – or piece of text – at a time, based on what came before. Each successive word choice contributes to the statistical probability of that word following in the sequence, based on patterns in the original training data, and the probability that it will result in an overall overestimated answer. In other words, models are always looking for a balance between a reasonable choice of words and an obvious one.
When chatter uses negative symmetry, according to this theory, it is essentially oscillating between the two. When starting a sentence whose function is to indicate something, the path of least resistance is to say that thing first it is not (X), and only then what is the object it is (Y). Put another way: For a sentence that begins with “This is,” following it with “not only” is more likely and safer than most options for directly specifying its subject. And after “This is not just,” the rest of the sentence becomes simple as well. The next word might be X—a boring and obvious descriptor that gets refuted—which sets up the last option of Y, a somewhat punchier descriptor.
Even if researchers could figure out exactly why chatbots embrace negative balance, there’s another reason that could make it so hard to fix: “Once something gets into these structures, it’s very hard to get it out,” Masrour, the Pangram engineer, said. That’s because one of the main ways AI models have continued to evolve is by learning about text generated by other robots. That AI text is probably full of negative symmetry, which puts it in a newer format. Now imagine that a growing part of the writing on the Internet is also powered by AI. This, too, becomes training data for future generations of AI.
On top of that, some AI labs are also using AI instead of, or in addition to, human reviewers in the post-training process, Chakrabarty said. Without intervention, there is a risk of “fall for example,” where AI reinforces its biases to the point that it loses touch with the human data it was meant to underpin.” “It’s a very bad loop,” Chakrabarty said. “There’s already negative symmetry in text, and then AI prefers negative symmetry—it gets to the point where it can’t write without it.” The AI language is eating its tail.
Chatter terms can be interesting, but there’s a flip side: They make it easier for AI writers to differentiate themselves from the human variety. Masrour said that even though the special AI writing symbols keep changing, it’s actually not getting any harder for the Pangram software to detect. The stubborn persistence of structures such as negative equity may be one reason.
The trade-off, for human writers, is that a powerful irony device is now clichés that make you sound like a robot. That has put some people in the awkward position of insisting that they don’t use AI—that is just the way they write. Before you scoff at them, consider that you too may soon find yourself speaking and writing more like a machine: A recent study by researchers in Germany suggested that AI writing techniques are now more spontaneous. human conversation. If that continues, perhaps negative parallelism will eventually lose its status as an AI author after all. The fault, dear readers, will not be in our chatbots, but in ourselves.
