AI words to avoid: the words ChatGPT overuses, charted

Why models overuse these words

Nobody knows for sure, though two explanations have real evidence behind them. The first comes from Tom Juzek and Zina Ward, who presented a paper at COLING 2025 asking why words like "delve", "intricate" and "underscore" had started turning up so often in scientific abstracts. They tested the obvious suspects first: the design of the models, the training algorithms and the text the models learned from. They found no evidence for any of them. What did fit was the last stage of training, reinforcement learning from human feedback, where people rate a model's answers and the model learns to produce more of what they rated highly. Even that came with a wrinkle, because in their online study people seemed to respond to "delve" differently from the other words. We'd call that a strong lead rather than a closed case.

The second explanation is about who those raters were. Writing in the Guardian in April 2024, Alex Hern pointed out that this feedback takes hundreds of thousands of hours of work, which is why AI companies send it to English-speaking workers in the global south, where the workers are cheap to hire. He also noted that "delve" is much more common in Nigerian business English than in British or American English. If the people rating and writing the answers used the word, the model may simply have picked it up from them. The two explanations don't compete; the second may describe how the first played out. The full story of "delve" goes through both.

What a list of AI words to avoid can't tell you

We build an AI detector, so we have every reason to want a list like this to be decisive. It isn't. Every word on it is ordinary English, and the charts, which run back to 2010, show people using each one years before any chatbot was writing abstracts. The difference now is frequency. A single hit in your paragraph tells you almost nothing about who wrote it, and three or four of them only tell you the paragraph is worth a slower read.

Getting this wrong has a cost: "delve" is everyday vocabulary in Nigerian English, so flagging it on sight flags Nigerian writers who have never opened a chatbot.

That's why the checker on this page only reports what it finds. It marks the listed words in your text and leaves the judgment to you. It never tells you who wrote something. If you're checking your own draft, a marked word is a chance to ask whether it's the word you meant. If you're checking someone else's, treat it as a reason to talk to them, never as the answer.