The Most Famous AI Writing Tic Is Also the Most Mysterious

Summary

A distinctive writing pattern has emerged as perhaps the most recognizable mark of AI-generated text: the "It's not X, it's Y" construction, now formally known as "negative parallelism." This rhetorical device, which presents contrasting statements to clarify what something is and isn't, appears across all major AI models including ChatGPT, Claude, and Gemini, as well as open-source alternatives. The pattern has measurably increased in prevalence—Barron's reported a more than fourfold increase in corporate communications from 2023 to 2025, and researchers at Pangram estimate that "Not just X but Y" sentences appear three times more frequently in AI-generated text than in human writing.

While negative parallelism can be effective when used judiciously—and has historical roots in Shakespeare and classical rhetoric—AI models have become overly reliant on it, creating a formulaic quality that signals artificial origin. OpenAI's product manager for model behavior, Laurentia Romaniuk, acknowledged that ChatGPT turns to the construction too often and disclosed the company is actively working to broaden its models' rhetorical repertoire. Similar variants, including the "No A, no B, just C" construction, have become so prevalent that users on forums trade techniques for removing these patterns from AI outputs. Despite widespread recognition of the issue, the fundamental reason why AI models favor negative parallelism remains unclear, with neither Anthropic nor Google responding to requests for comment on their understanding of the phenomenon.

• OpenAI acknowledges overreliance on the pattern and is working to expand their models' linguistic variety
• The underlying reason why AI models gravitate toward negative parallelism remains a mystery, even to the companies building them

Editorial Opinion

Negative parallelism represents a fascinating case study in how AI models inherit linguistic quirks from their training data and architecture, yet this apparent blind spot also reveals the limits of human understanding of these systems. While the pattern can be detected and even removed through user workarounds, the fact that even the creators of these models struggle to explain why it occurs raises important questions about AI interpretability and the hidden biases embedded in our largest language models. As AI writing becomes increasingly indistinguishable from human writing in casual use, developing stronger introspection into these systems' tendencies will be critical for both content authenticity and public trust.