A theoretical economics paper by researchers from Princeton, the University of Washington, and other institutions challenges the assumption that AI will improve scientific research. The study suggests that while AI saves time, it could lead researchers to spend less effort on each project, resulting in lower-quality work. Language models are expected to speed up science at every stage, from developing hypotheses to analyzing data to writing papers. However, the study argues that the time saved may not translate into better research, as scientists may prioritize starting new projects over thorough analysis. The paper's model builds on the concept of 'opportunity cost,' where time becomes more valuable when AI takes over routine tasks, leading researchers to allocate less time to deep-dive work.
The researchers built a mathematical model based on optimal foraging theory from behavioral ecology, adapted to simulate how scientists distribute their labor across projects. In the model, a research project unfolds in two phases: first, checking if an idea is viable, then deciding whether to abandon it or push forward. The voluntary part of the process, such as running extra experiments or polishing prose, is what gets sacrificed when time becomes scarce. The study outlines three scenarios depending on where AI is applied in the research process. In two out of three cases, thoroughness drops, as time saved is better spent on new projects rather than refining existing ones.
The paper argues that AI's impact on research is not uniform and depends on which phase of the process gets sped up. The study highlights that while AI may increase productivity, it can also shift bottlenecks to later stages, such as validation and maintenance. The authors warn that the perceived time savings do not automatically lead to deeper analysis, and that the fallacy of saved time can result in less rigorous research. The study calls for institutional responses that are discipline-specific, as AI's effect on research varies depending on the stage of the process it accelerates.
Source: thedecoder
