A new study conducted by a group of researchers from the US and Germany has revealed that artificial intelligence is spreading rapidly and silently in the academic publishing world. The team, which examined publications in the biomedical field in particular, found that at least 13.5 percent of articles published in 2024 used AI-supported tools in the writing process. The findings show that this use can be traced through structural changes in written language.
AI use peaks in academia
The researchers examined more than 15 million biomedical abstracts in the PubMed database. The study compared the period before and after major language models such as ChatGPT became publicly available.
The texts were searched for traces of AI based on changes in word usage, sentence structure, and word types. Inspired by the “superstandard death” methodology applied during the COVID-19 pandemic, the team adapted this method to academic writing as “superstandard word usage.”
According to the analysis results, while the vast majority of non-standard words in academic texts before 2024 were nouns, this trend has changed sharply after 2024. Now, verbs and adjectives are prominent. In texts published after 2024, 66 percent of the words used above the norm were verbs and 14 percent were adjectives.
In addition, a noticeable increase was detected in the number of ornate expressions in the texts. This change shows that the traditional academic style is being deviated from and that the writings are increasingly being produced by artificial intelligence.
The study examined not only the prevalence of artificial intelligence use, but also its geographical and sectoral distribution. It was determined that the LLM (Large Language Model) usage rates were at the highest level in biomedical articles published in the USA, China and India. In Turkey, the artificial intelligence usage rate remained at 4 percent. This rate only covers biomedical articles; no data was shared on other scientific fields.
The findings reveal that the formal structure of scientific publications is rapidly changing and that large language models are becoming effective not only in content production but also in the presentation of scientific thought. However, the consequences of this change in terms of reliability continue to be a matter of debate.
Large language models are known for occasionally producing incorrect information. For this reason, the effects that AI-supported writing processes can have on scientific validity are being closely monitored by both universities and publishing houses.
