New Delhi: Use of artificial intelligence (AI) tools in scientific writing is rising steadily, with computer science papers showing the fastest and largest uptake at an estimated 22 per cent, according to a new analysis of over one million pre-print and published research papers.

Researchers from Stanford University and other US institutes examined 1,121,912 papers submitted to pre-print repositories arXiv and bioRxiv, and those published across Nature journals between January 2020 and September 2024.

By tracking usage of words commonly associated with large language models (LLMs) such as ChatGPT, the team estimated how often AI tools were used to modify or generate content in research papers.

Findings published in Nature Human Behaviour suggest a steady growth in LLM usage, with computer science showing the sharpest rise. In contrast, mathematics papers and those published in Nature journals indicated relatively lower evidence of AI-assisted writing.

The researchers also found higher estimated AI use among pre-prints on bioRxiv authored from regions where English is not the dominant language, such as China and continental Europe.

Shorter papers and authors posting pre-prints more frequently were more likely to rely on AI tools, suggesting a link between high publication pressure and LLM usage.

“These results may indicate the competitive nature of certain research areas and the pressure to publish quickly,” the study noted.

A separate examination of 200 randomly selected computer science pre-prints uploaded to arXiv in February 2024 showed that only two explicitly disclosed the use of LLMs.

The team said further studies analysing disclosure statements could help understand researchers' motivations, adding that unclear policies or intentional avoidance could explain low disclosure rates.

A recent study in Science estimated that at least 13 per cent of research abstracts published in 2024 may have used AI tools, based on patterns in writing style associated with LLMs. Another analysis by the University of Tübingen, Germany, which reviewed over 15 million biomedical papers from 2010 to 2024, suggested that AI models are causing a significant shift in academic writing vocabulary.