Atlantic: AI's 'It's not X, it's Y' tic proves hardest to shake

The most reliable way to spot AI writing right now is not the em-dash, and not the word 'delve.' According to The Atlantic, it is the sentence that tells you what its subject is not before it tells you what it is: 'It's not X; it's Y.' Once you notice the construction you start seeing it everywhere, from an annual review by Citizens Financial Group that called a growth number 'not just a win for the private bank—it's a win for the entire enterprise,' to the horror novel Shy Girl, which was pulled by its publisher this year after lines like 'No bag, no things, no armor, just me' helped fuel AI-writing accusations.

The most popular label for the tic is 'negative parallelism.' Laurentia Romaniuk, a product manager for model behavior at OpenAI, prefers 'contrastive phrasing.' Whatever you call it, the reporting's central claim is that no one is quite sure why chatbots are so enamored with the construction, maybe not even the companies that created them. The simplest theory is that it was heavily present in the human-written text the models were trained on. A related theory is that human reviewers rated responses using the construction highly, because it gives the impression of nuance and insight.

Why this matters if you are not a linguist. Earlier chatbot tells like the overuse of 'delve' reportedly came and went as AI companies honed their models and worked out kinks. This one has shown no signs of abating. For anyone paid to write, that turns a once-punchy rhetorical device into a public tell. Corporate comms, marketing copy and increasingly novels are being read for it. OpenAI's stated fix is that it is working on ways to broaden the chatbot's repertoire; in the meantime, users can try giving ChatGPT 'custom instructions.'

The honest caveat is that the piece is essentially a stylistic diagnosis, not a technical audit. It does not disclose what fraction of ChatGPT outputs contain the construction, what other frontier labs are doing about it, or which specific training-data or reward-model changes have been tried and failed. The forward-looking thought: the lab that first ships a model whose default voice does not lean on this crutch gets a marketing story you can hear in one sentence, and writers who deliberately avoid the pattern quietly inherit a new, low-cost credibility signal.