Words ChatGPT Overuses (And Why)

What words does ChatGPT actually overuse?

Beyond character-level tells like the narrow no-break space or an unbroken run of em dashes, large language models have a measurable vocabulary fingerprint: a small set of words that show up far more often in AI-assisted writing than in ordinary human writing. Researchers found this by comparing word frequencies before and after ChatGPT's late-2022 release across millions of documents — a method that works precisely because it doesn't need access to the model itself, only a large enough before/after sample of text.

This is a trimmed selection of the 21 "focal words" catalogued in the academic literature, not an exhaustive list — treat it as a starting point for a find-and-replace pass, not a rulebook.

The research behind the list

The primary source is "Delving into ChatGPT usage in academic writing through excess vocabulary" (Kobak, González-Márquez, Horvát & Lause, arXiv:2406.07016, June 2024), which analyzed 14.2 million PubMed abstracts published between 2010 and 2024. Rather than trying to detect any single AI-written abstract, the authors measured "excess" usage: how far a word's frequency rose above its own pre-2023 trend line. Style words that a human population has no collective reason to suddenly start preferring — but that a widely adopted writing tool would inject uniformly — show up as sharp, synchronized spikes.

By that measure, "delve" stood out furthest: the paper's authors measured it appearing at roughly 25 times its expected pre-ChatGPT rate in 2024 abstracts, with "showcasing" and "underscores" close behind at roughly 9 times each. Using the combined excess frequency of words like these as a lower-bound estimator, the authors concluded that at least 13.5% of 2024 abstracts showed measurable LLM involvement — rising to roughly 40% in some sub-corpora broken down by country, journal, and discipline. A related preprint, "Why Does ChatGPT 'Delve' So Much?" (arXiv:2412.11385), extended the word list to 21 focal terms and investigated the cause more directly.

One of the more striking side findings: once "delve" overuse was widely reported in early 2024, later data showed its frequency measurably dropping in some corpora — people noticed, and edited it out. That's the practical use of a list like this one: awareness of a tell is often enough to stop relying on it by accident.

Why does ChatGPT keep saying "delve"?

No single cause has been proven, but the leading, actively studied explanation involves reinforcement learning from human feedback (RLHF) — the stage where human reviewers rank a model's candidate responses to teach it which ones sound best. The arXiv:2412.11385 authors ran comparative model tests whose results were consistent with RLHF annotator preferences playing a role, though they're candid that their own exploratory study didn't fully confirm the mechanism.

A specific version of that theory, raised publicly by AI researcher Tony Zador and covered by outlets including The Guardian, points to where that annotation work happens: much RLHF labeling is outsourced to English-fluent reviewers in countries such as Nigeria, where "delve" is a more common word in formal business English than in US or UK usage. If annotators in that pool rated "delve"-using responses slightly more favorably on average, that preference would get baked directly into the model during training — a plausible, human-shaped explanation for a pattern that otherwise looks like a mysterious quirk of the algorithm.