Large Language Models (LLMs) have become increasingly popular tools for generating text in various domains, including scientific writing. However, the challenge lies in detecting whether a piece of writing has been generated using an LLM. A group of researchers has recently developed a novel method to estimate the usage of LLMs in scientific writing by analyzing vocabulary changes over time.
The researchers examined a large set of scientific writing, particularly abstracts published between 2010 and 2024, to identify “excess words” that became more prevalent during the LLM era (2023 and 2024). They discovered that at least 10 percent of the abstracts from 2024 were processed using LLMs. This analysis involved tracking the frequency of specific words and comparing their expected versus actual usage before and after the widespread adoption of LLMs.
The study revealed that certain words experienced a significant surge in usage after the introduction of LLMs. Words like “delves,” “showcasing,” and “underscores” saw a considerable increase in frequency in 2024 compared to previous years. Additionally, common words such as “potential,” “findings,” and “crucial” also showed a notable uptick in usage post-LLM era. These changes in vocabulary were unprecedented in both quality and quantity, indicating a significant shift in scientific writing style.
While language evolution is natural and words may go in and out of style over time, the researchers noted distinct differences in the post-LLM era. Prior to the introduction of LLMs, significant spikes in word usage were typically associated with major world health events like the Ebola outbreak or the Zika virus. However, in the post-LLM period, a wide range of words experienced sudden increases in scientific usage, independent of external events.
The researchers classified these sudden vocabulary changes as “marker words” that are indicative of LLM usage in scientific writing. These markers predominantly consisted of style words such as verbs, adjectives, and adverbs, as opposed to nouns that were more prevalent during the Covid-19 pandemic. By highlighting hundreds of these marker words, the researchers were able to identify patterns that signal the influence of LLMs in text generation.
Through statistical analysis of marker word appearance in individual papers, the researchers estimated that approximately 10 percent of post-2022 papers in the PubMed corpus were likely written with some level of LLM assistance. However, this number could be higher as the researchers acknowledged the possibility of missing LLM-assisted abstracts that did not contain the identified marker words.
The impact of LLMs on scientific writing is evident through the shifts in vocabulary usage and writing style observed in recent years. By developing innovative methods to detect LLM-generated text, researchers can gain valuable insights into the evolving landscape of automated content creation in academia.