Linguistic and statistical analyses of scientific articles suggest that generative AI may have been used to generate a large amount of scientific literature.
Two academic papers claim that the analysis of word choice in the corpus of scientific publications reveals an increasing use of AI to write research articles.
A study, published in March by Andrew Gray, from University College London (United Kingdom), suggests that one percent of all articles published in 2023 were written, at least partially, by AI.
A second study published in April by a team from Stanford University (USA) claims that this figure could range from 6.3% to 17.5%, depending on the subject.
In both works, certain words that large language models (LLMs) commonly use, such as “intricate,” “fundamental,” and “meticulously,” were searched for. By tracking the usage of these words in scientific literature and comparing it to words that are not particularly favored by artificial intelligence, both studies claim that they can detect a growing dependence on machine learning within the scientific publishing community.
In Gray’s work, the use of control words such as “red”, “conclusion”, and “after” changed by a small percentage from 2019 to 2023. The same happened with other specific adjectives and adverbs until 2023 (referred to as the post-LLM year by Gray).
In that year, the use of the words “meticulous”, “commendable”, and “intricate” increased by 59%, 83%, and 117%, respectively, while their prevalence in scientific literature barely changed between 2019 and 2022. The word with the highest increase in prevalence after 2022 was “meticulously”, with a 137 percent.
The Stanford document found similar phenomena, demonstrating a sudden increase in the words “kingdom”, “exhibit”, “intricate”, and “fundamental”. The first two were used approximately 80 percent more than in 2021 and 2022, while the last two were used approximately 120 and almost 160 percent more respectively.
Researchers also took into account word usage statistics in various scientific disciplines. Computer science and electrical engineering ranked highest in terms of using the language preferred by AI, while mathematics, physics, and articles published in Nature only experienced increases of between 5% and 7.5%.
Stanford researchers also observed that authors who publish more preprints, work in more saturated fields, and write shorter articles seem to use AI more frequently. Their article suggests that the lack of time in general and the need to write as much as possible encourage the use of LLMs, which can help increase production.
The use of artificial intelligence divides the entire scientific community
Using artificial intelligence to assist in the research process is not something new, and many scientists are willing to use it to improve experiments and obtain better results.
However, using AI to write summaries and other parts of articles is very different, as the general expectation is that scientific articles are written by real human beings, not robots, and at least a couple of editors consider the use of LLM to write articles to be a violation of scientific ethics.
The use of AI models can be very risky, as they often produce inaccurate texts, which is precisely what scientific literature is not supposed to do. AI models can even falsify citations, something that has caused an infamous mess for two New York lawyers for citing cases that ChatGPT had invented.
“Authors who use text generated by LLM should feel pressured to disclose it or think twice before doing so, for the sake of the integrity of basic research,” said Gray, from University College London.
Stanford researchers also raised similar concerns, writing that the use of generative AI in scientific literature could create “risks to the security and independence of scientific practice”.
