By the end of 2025, 89 percent of scientific articles show traces of AI

A new statistical methodology developed by an international team of researchers has enabled an accurate assessment of the involvement of Large Language Models (LLM) in biomedical scientific publications. According to ixbt.com, it has been revealed that by the end of 2025, 89 percent of all open-access biomedical articles were edited or written with the help of AI. This figure differs sharply from the lower-bound estimates in previous studies. This is reported by Ixbt.com reports .

The new approach is based on a deep analysis of over 1.194 million full-text scientific articles published in the PubMed Central database between 2017 and 2025. Experts examined the frequency of words in 379 services and methodological settings, which saw a sharp increase in usage after the emergence of ChatGPT at the end of 2022. For each marker-word, a counterfactual trend was constructed, and data from 2018–2022 were extrapolated to 2023–2025.

Which parts of scientific articles are being edited the most?

According to the research results, it was found that AI is used with varying frequency in different parts of the text. As of December 2025, the discussion section of the articles recorded the highest result at 68 percent. Abstracts ranked second with a 67 percent share.

At the same time, the methodology sections remained the least affected area, with a figure of 32 percent. However, when the full text of this section was analyzed, the figure rose to 54 percent. This indicates that scientists are using AI primarily for interpreting results rather than describing procedures.

Geographic disparities and language changes

The authors of the study also identified significant differences across countries. In 2025, South Korea (85%), China (82%), and Taiwan (80%) are leading in the use of LLM technologies. The lowest figure was recorded in the UK, at 28 percent.

Predominantly English-speaking countries showed a 72% versus 37% disparity compared to the rest of the world. While the frequency of marker-words in the texts of scientists from non-English speaking countries was significantly lower until 2023, by 2025 they had reached the level of native English speakers.

These figures align well with the results of a survey conducted by Wiley. Specifically, in a survey conducted in mid-2025, more than 70 percent of respondents confirmed that they used LLM assistance in writing drafts.

However, researchers are warning the academic community about emerging risks. On one hand, language models reduce the language barrier for non-English speaking researchers, but on the other, they pose risks such as the appearance of fabricated citations, the homogenization of scientific language, and the weakening of the link between writing and critical thinking.