Scientists who use large language models (LLMs) to support their research may spend less time improving existing work and move more quickly to new projects, according to a modelling study1.
The study suggests that widespread LLM adoption could lead scientists to “do more, less well — rather than the same amount, better”. However, the researchers say artificial intelligence is not solely responsible for this potential decline in research quality.
The findings highlight longstanding incentives in academia that reward publication volume more than scientific quality, says study co-author Carl Bergstrom, a biologist at the University of Washington in Seattle. Researchers face pressure to produce an ever-increasing number of papers, while LLMs could make it easier to meet those targets. “LLMs are rarely the problem themselves,” says Bergstrom. “LLMs hold up a mirror to problems that we already have.” The study was posted on the arXiv preprint repository on 19 July and has not yet been peer reviewed.
How AI could change the scientific research process
To estimate how LLMs might affect scientific productivity, the authors divided the research process into several stages. The first is discovery, when scientists develop hypotheses and conduct initial experiments to assess a project’s potential. This is followed by a two-part development stage: required tasks, such as creating figures and drafting manuscripts, and discretionary work, including follow-up experiments and improving the quality of scientific writing.
The researchers used principles from optimal-foraging theory — a framework that examines how animals maximize energy gains while conserving resources — to model how scientists might redistribute their time after adopting LLMs. The model assumed that LLMs perform at their highest potential: quickly, inexpensively and accurately.
AI linked to explosion of low-quality biomedical research papers
The model predicts that LLMs could accelerate every stage of scientific research, but that faster productivity would not necessarily produce higher-quality papers. More efficient discovery and completion of required tasks could allow researchers to publish more quickly. Yet because academic incentives prioritize publication counts, scientists may have little reason to spend additional time refining their analyses, writing or conclusions.
Source: www.nature.com