This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: Up until two years ago, using large language models was still considered a special skill that everyone was encouraged to learn. Now, it is ubiquitous, especially in writing research papers that eventually become part of the scientific knowledge we build around a subject.
For universities and publishers to create better policies or ethical guidelines for LLM use, they first need to know how widespread AI use really is. In a recent study, researchers estimated LLM use in academic writing by tracking more than 350 words that AI tools like ChatGPT use far more often than humans do. Rather than judging whether a single paper was written by AI, this study shifts the focus to the corpus level, where it analyzes trends across a massive collection of papers.
After reviewing 1.1 million English-language biomedical studies, they found signs of AI-assisted writing or editing in almost 90% of open-access papers by 2025. The prevalence of use, however, was not uniform across the paper. The strongest signs of AI use appeared in the discussion section, at about 78%, whereas only 54% of methods paragraphs showed AI use.
These findings are published on the preprint server arXiv. OpenAI released ChatGPT in November 2022, and within a year, its use in academic writing accelerated. For many students and academics whose first language is not English, LLMs became a tool that helped lower the long-standing language barrier.
AI chatbots can enhance writing skills by providing instant feedback on grammar and style, while also offering a space to brainstorm ideas. Despite its advantages, the increasing use of AI in this area is also raising concerns among experts, as LLMs are known to hallucinate information. AI models frequently invent facts or fake references and deliver them with a confidence that makes those errors surprisingly easy to miss.
This can lead to errors and even outright academic fraud. In addition, many academic authors ignore existing ethical guidelines for LLM use and often don't disclose their use at all, which could erode public trust in science itself. Standard AI detectors are unreliable at detecting actual use, and studies have repeatedly shown this.
Extract — continue reading at the source.
