As it turns out, LLMs are partial to certain words. The team analyzed the abstracts from 14 million scientific papers published between 2010 and 2024 to track the frequency of each word used. They found an abrupt increase, “unprecedented in both quality and quantity,” in the frequency of “certain style words” after LLM tools became widely available in late 2022.
What words do AIs like to use too much?
Answer: “Delve,” for one.
It isn’t easy to determine if something was written by a large language model (LLM) these days. However, a group of researchers from Germany’s University of Tubingen and Northwestern University have discovered that analyzing the vocabulary of a piece of writing could help determine if its author was human or not.
As it turns out, LLMs are partial to certain words. The team analyzed the abstracts from 14 million scientific papers published between 2010 and 2024 to track the frequency of each word used. They found an abrupt increase, “unprecedented in both quality and quantity,” in the frequency of “certain style words” after LLM tools became widely available in late 2022.
Many of these words, which the team dubbed style words like verbs, adjectives and adverbs, were very uncommon in scientific abstracts before the beginning of 2023. For example, the word “delves” appeared 25 times more often in these papers in 2024 than in previous years. “Showcasing” and “underscores” were nine times more common, while the use of “potential” increased by 4.1 percentage points. A small sample of the words AIs seem to like includes “across, additionally, comprehensive, crucial, enhancing, exhibited, insights, notably, particularly, within.”
As it turns out, LLMs are partial to certain words. The team analyzed the abstracts from 14 million scientific papers published between 2010 and 2024 to track the frequency of each word used. They found an abrupt increase, “unprecedented in both quality and quantity,” in the frequency of “certain style words” after LLM tools became widely available in late 2022.
