The AI Writing Tic Nobody Can Explain: Why Chatbots Love 'It's Not X, It's Y'
Key Takeaways
- ▸Negative parallelism appears three times more frequently in AI-generated text than human writing, and has quadrupled in corporate communications from 2023 to 2025
- ▸The pattern persists across all major chatbots—ChatGPT, Claude, Gemini, and open-source models—despite attempts by companies to reduce its prevalence
- ▸Neither AI companies nor researchers fully understand why their models produce this construction so reliably, suggesting fundamental gaps in model interpretability
Summary
A distinctive linguistic pattern—"It's not X, it's Y"—has become one of the most recognizable tics of AI-generated writing, appearing far more frequently in text produced by major chatbots like ChatGPT, Claude, and Gemini than in human writing. Known formally as "negative parallelism" or "contrastive phrasing," the construction has quadrupled in corporate communications between 2023 and 2025, according to Barron's reporting, and Pangram research indicates it appears three times as often in AI-generated text as in human writing. The pattern manifests in various forms—from the straightforward "It's not X; it's Y" to the more elaborate "No A, no B, just C"—and has been spotted in everything from corporate financial communications to published fiction, with accusations of AI authorship following works like the horror novel Shy Girl, which was pulled over suspected AI generation.
Despite its prominence and measurable prevalence, researchers and AI companies alike remain puzzled about why their models produce this pattern so reliably. OpenAI's Laurentia Romaniuk acknowledged that ChatGPT overuses the construction and feels formulaic, and the company is exploring solutions through model improvements and custom instructions for users. However, the pattern shows no signs of abating across any of the major models. Most troubling is that neither OpenAI, Anthropic, Google, nor independent researchers seem to fully understand the root cause—suggesting significant gaps in how companies understand their own models' linguistic behavior.
- OpenAI is exploring fixes through model adjustments and user-facing solutions like custom instructions, but the problem remains industry-wide and stubborn
Editorial Opinion
The prevalence and persistence of 'negative parallelism' in AI writing exposes a troubling reality: even as these models grow more sophisticated, they betray their artificial origins through detectably formulaic patterns that human writers rarely adopt. What's more concerning is the admission from AI leaders that they don't fully understand why this happens—a gap in model interpretability that should alarm anyone relying on these tools for authentic communication. As AI-generated text becomes increasingly woven into corporate messaging, academic work, and creative industries, the inability of companies to eliminate such patterns raises uncomfortable questions about authenticity and the ethics of deploying AI in contexts where human authorship is assumed or required.
