LLMs’ impact on science: Booming publications, stagnating quality

Recent events have highlighted alarming trends in scientific publishing, particularly regarding the integrity of research papers. A series of retractions, with the most recent occurring just weeks ago, has drawn attention to the prevalence of AI-generated errors in academic articles. Puzzling terms such as "runctitional," "fexcectorn," and "frymblal" have raised serious concerns about the effectiveness of peer review processes in some journals. The question looms: Are these notable cases indicative of a broader issue? To investigate the influence of AI on scientific literature, researchers from Berkeley and Cornell undertook a significant study. They examined three major pre-publication archives, aiming to determine the extent to which Large Language Models (LLMs) have been utilized in scientific writing. Their findings revealed a paradox: while the adoption of AI technologies has led to a surge in the volume of published papers, it has concurrently resulted in a decline in the perceived quality of these submissions. The team analyzed abstracts from major preprint archives, including arXiv, which contributed 1.2 million documents, the Social Science Research Network with 675,000, and bioRxiv, offering an additional 220,000. This extensive dataset spanned various research disciplines and included documents submitted before AI models could produce credible text. The researchers trained a model based on human-generated abstracts from the pre-ChatGPT era to assess the likelihood of AI involvement in newer submissions. By comparing the productivity of authors before and after the introduction of LLMs into their writing processes, the researchers observed a notable increase in output. The study concluded that the integration of LLMs correlates with a significant rise in scientific publications across all three examined preprint repositories. However, this surge raises critical questions about the implications for research quality and the future of scientific discourse.