AI's Favorite Writing Trick Baffles Even Its Creators

The construction "It's not X; it's Y" is ancient. Shakespeare deployed it repeatedly in Julius Caesar: "The fault, dear Brutus, is not in our stars, but in ourselves," "Not that I loved Caesar less, but that I loved Rome more," and "I come to bury Caesar, not to praise him." Vince Lombardi used it in the 1960s ("winning isn't everything; it's the only thing"), and DiGiorno ran it in 1990s pizza commercials ("It's not delivery. It's DiGiorno").

Yet in the age of generative AI, this rhetorical device has become inescapable—and unmistakably artificial. Barron's documented the shift: appearance of the pattern in corporate communications more than quadrupled from 2023 to 2025. Researchers at Pangram, which builds AI-detection software, estimate it appears three times as often in AI writing as in human writing. All major chatbots—ChatGPT, Claude, Gemini, and various open-source models—rely on it to varying degrees, according to Elyas Masrour, a founding engineer at Pangram. The horror novel Shy Girl, pulled by its publisher this year, drew accusations of AI authorship partly because of phrases like "No bag, no things, no armor, just me." (The author denied using AI.) The construction is now the best-known tic of AI writing, so recognizable that a tech writer observed Shakespeare might have faced similar accusations had he debuted this year.

Yet no one—not even the companies that created these models—appears to know exactly why the pattern took root or how to uproot it. Laurentia Romaniuk, a product manager for model behavior at OpenAI, confirmed that ChatGPT "turns to it too often," making output feel formulaic, and said the company is working on ways to broaden its repertoire. She suggested users try "custom instructions" as a workaround. Anthropic and Google did not respond to interview requests. The term itself has no agreed-upon name; researchers and engineers now call it "negative parallelism," though academic terms like "antithesis" and "metalinguistic negation" capture only some forms of the construction.

Three main theories explain why models became so dependent on it. The simplest: negative parallelism was present in the vast text used to train large language models—books, academic papers, patent filings, and especially the internet—so models learned the pattern. But during reinforcement learning (the process where human reviewers grade responses), reviewers may have favored responses using the construction because it gives an impression of nuance: the AI seems to reason from a subpar descriptor to a more apt one. A second, stranger theory points to how language models work. They are fundamentally text-prediction machines generating one "token" (chunk of text) at a time, seeking a balance between clever and obvious word choices. When a chatbot has begun a characterizing sentence, the path of least resistance is to say what the subject isn't (X) first, then what it is (Y). After "This is not just," the choice of X (the obvious, negated descriptor) becomes likely, setting up the final choice of Y (the punchier descriptor). This statistical pathway may make negative parallelism both more probable and safer than alternatives.

But the hardest barrier to fixing the problem may be structural. "When something gets into these models, it's very hard to pull it out," Masrour explained, because newer AI models train on text generated by older ones—many of which are saturated with negative parallelism. As an ever-growing share of the internet becomes AI-generated, that infected text enters the training data for the next generation of models. Adding to the loop, some AI labs now use AI systems instead of human reviewers in post-training. Tuhin Chakrabarty, a computer-science professor at Stony Brook studying AI writing, warned this risks "model collapse," where AI reinforces its own biases so thoroughly that it loses touch with the human data meant to ground it. "It's a very vicious loop," Chakrabarty said. "There's already negative parallelism in the text, and then AI is preferencing negative parallelism—it comes to a point where it just cannot write without that."

There is a silver lining: the prevalence of chatbot clichés makes AI writing easier to distinguish from human writing. Masrour noted that although specific markers of AI writing keep changing, Pangram's detection software hasn't found it harder to detect. Ironically, the very persistence of negative parallelism may be one reason detection remains effective. But for human writers, the cost is high. A once-potent rhetorical device is now so associated with bots that using it raises suspicions. Some writers have found themselves insisting they did not use AI—that's just how they write. A recent study by German researchers adds a troubling wrinkle: AI's writing tics are now appearing in spontaneous human conversation. If that trend continues, negative parallelism could eventually lose its status as a reliable AI-writing tell. The irony, Shakespeare might appreciate, is that we could end up talking and writing like machines, not because we chose to, but because we've been exposed to enough of it—a self-fulfilling prophecy hidden in the fault lines of language itself.