77% of Biomedical Papers Now Use AI, 2025 Data Reveals

According to a report by Solidot, a study of preprints posted on arXiv analyzed all English-language papers included in PubMed Central for the full year of 2025, revealing that the usage rate of large language models (LLMs) reached 77%. This figure represents a significant increase from 52% in 2024. Successive reports of survey results indicate that the integration of AI in paper writing is more widespread than previously thought.

Survey Details and Numbers

The research team targeted all English papers published on PubMed Central in 2025, detecting traces of AI-assisted writing. The analysis found that 77% of all papers exhibited characteristics indicating some form of LLM usage. This finding is consistent with a separate survey conducted in 2025, where 71% of researchers reported using AI assistance in writing their papers. Notably, for papers published in December 2025 alone, papers with traces of AI-assisted writing reached approximately 90%. Compared to the annual average of 77%, this shows a trend of usage increasing further toward the end of the year.

Bias in Usage Location

The latest study also shed light on which parts of the paper AI was used for. The results showed that AI-assisted writing was used more frequently in the abstract, introduction, and discussion sections. In contrast, its use in the core methods and results sections of the research was relatively low. This may suggest that LLMs are well-suited for summarizing existing knowledge, explaining context, and describing interpretations. The findings also support the notion that human creative involvement remains indispensable for a study’s unique experimental procedures and the data itself.

Industry Impact and Challenges

The 77% usage rate of AI has posed several challenges for the academic community. First, concerns have been raised about the homogenization of paper quality. The style and structure generated by AI tend to be similar, risking the loss of diverse academic expression. Second, the boundaries of intellectual property rights and plagiarism become blurred. AI-generated text makes it difficult to distinguish which information is cited from original sources and which is created by the LLM. Third, the reliability of the peer review process is questioned. Traditionally, reviewers have evaluated the logical development and expressive habits of human thought, but new evaluation criteria are now required in a world where AI is involved.

Editorial Opinion

The results of this survey vividly demonstrate that the use of AI in academia is rapidly transitioning from “assistance” to a “standard method.” In the short term, this will likely accelerate the introduction of AI detection tools and the establishment of rules mandating the disclosure of AI use by publishers and academic societies. Major journals, including Nature, have already strengthened their disclosure policies regarding AI use, and this trend is expected to strengthen further.

Looking at the long term, it is possible that the very criteria for evaluating papers themselves will be forced to shift. Some view that as AI takes on the creation of abstracts and introductions, the essence of research will become concentrated solely in the ingenuity of methods and the originality of interpreting results. Furthermore, how will an environment where researchers are trained to treat AI-generated content as their own “intellectual findings” impact the speed of scientific evolution a decade from now? Universities and research institutions are being asked how they will embrace AI, not merely as a tool for efficiency but as a force that could fundamentally alter the thinking process itself.

References

References

• “九成生物医学论文有 AI 辅助写作痕迹”, by (author unknown) — Solidot, 2026-08-24T07:26:30.000Z (ARR)
• Source URL: https://www.solidot.org/story?sid=85173