‘If I don’t cheat, I’m going to be behind’: What it’s like to be at uni in 2026

A new study suggests that nearly all biomedical research papers published in December 2025 show evidence of artificial intelligence assistance, with 90% displaying signs of AI involvement. The findings, based on an analysis of over 10,000 papers archived in the PubMed Central repository, reveal that AI large language models (LLMs) are increasingly integrated into the writing process within the scientific community. The study, posted on the arXiv preprint server on 12 August and not yet peer-reviewed, indicates that 77% of papers published throughout 2025 and 52% from 2024 were influenced by AI tools. These numbers surpass earlier estimates, which suggested AI use ranged from 13.5% to 57%, depending on the methodology and scope of the analysis. The study focused on detecting patterns of AI-generated text by analyzing word frequencies common in LLM outputs. Researchers identified these markers in the introduction, discussion, and abstract sections of biomedical papers more frequently than in the methods and results sections. For instance, 78% of discussion sections in December 2025 publications showed signs of AI influence, while only 58% of results sections did. This discrepancy highlights a growing concern among scientists about the potential impact of AI on the integrity of scientific communication. Dmitry Kobak, a computer scientist at Ghent University in Belgium and one of the lead researchers, initially doubted the accuracy of the study's findings. He expressed surprise at the high percentage of AI-assisted content, stating that he believed there must have been an error in the calculation. However, after conducting additional checks, he confirmed that the data appeared valid. Kobak noted that the current study employed a more refined method for identifying AI-generated text, resulting in a significantly higher estimate of AI use compared to prior analyses. The study's approach, which directly estimates AI use instead of relying solely on lower-bound assumptions, led to a notable increase in the estimated rate of AI involvement. For example, the rate of AI use in 2024 abstracts rose from 13.5% to 31%. This finding aligns with a 2025 survey in which 71% of researchers admitted to using AI for writing assistance. Kobak and his colleagues argue that the actual prevalence of AI use may even exceed self-reported figures, as some researchers may understate their reliance on AI tools. While the study underscores the rapid integration of AI into scientific publishing, other experts caution against drawing definitive conclusions. Some researchers suggest that the high percentages observed in the study may reflect broader trends in AI adoption rather than an accurate representation of the entire scientific literature. Nevertheless, Kyle Siler, a social scientist at the University of Toronto, asserts that the trend is irreversible. “The toothpaste is out of the tube, and it’s not going back,” he remarked, emphasizing that AI-driven writing is becoming an established norm in academia. The implications of widespread AI use in scientific writing extend beyond mere efficiency. Concerns have emerged regarding the potential for AI to introduce biases or inaccuracies into research. Kobak warns that the use of LLMs to draft introductions and discussions could distort the direction of scientific inquiry, as any inherent biases in the training data of these models may be reflected in the final output. Additionally, the ability of LLMs to generate fabricated or misleading information, known as hallucination, raises ethical questions about the reliability of AI-assisted content in scholarly contexts. As AI continues to shape the landscape of scientific communication, the challenge lies in balancing its benefits with the need to maintain the integrity of academic discourse.