Study argues AI makes scientists do more research — but worse, not better
A new theoretical paper argues the real risk of AI in science isn't that models make mistakes — it's that they make researchers' time too valuable to spend polishing anything. Even a perfect LLM, the authors say, could push scientists toward shallow output.
A team from Princeton, the University of Washington and elsewhere has built a mathematical case that saving scientists' time with LLMs may degrade the quality of science even when the models work flawlessly.
Their paper on arXiv deliberately treats large language models as error-free and essentially free tools, isolating the economic effect of time alone. That effect comes down to opportunity cost: once AI removes the tedious stages of a project, a researcher's remaining hours become more valuable, and every hour spent refining an existing paper is an hour that can't go toward launching the next one.
The model borrows optimal foraging theory from ecology to simulate how researchers allocate effort. Depending on where AI slashes costs, the outcomes differ. If LLMs speed up idea evaluation, researchers get pickier about which projects to start but give each one less thorough treatment. If AI accelerates the publishing mechanics — writing, formatting, analysis — weaker projects suddenly become worth finishing, flooding the system with shallower papers. In two of the three scenarios modeled, individual publication quality drops. Only when AI speeds up the deep-dive phase, the voluntary extra experiments and careful analysis researchers have always cut under time pressure, does quality actually improve.
The counterintuitive headline — "do more, less well, rather than the same amount, better," as the authors put it — has real-world echoes even if the felt speedups aren't always real. A METR study the paper cites found experienced developers using AI tools took 19 percent longer to finish tasks while feeling 24 percent faster. The strain is already visible in the publication system, too: submissions are climbing in fields where LLMs accelerate writing, ArXiv has tightened penalties for AI-bungled papers, and Sakana AI's AI Scientist-v2 pushed a fully machine-generated paper through an ICLR workshop.
None of that means AI can't help science — it means institutions may need discipline-specific answers, because the effect depends on which phase of the research lifecycle gets sped up.
What to watch: whether labs design AI tools around the third scenario — targeting the deep-work phase — or simply shovel productivity gains into volume.
If saved time doesn't flow into deeper research, what should institutions do to steer LLMs toward quality over quantity? Tell us in the comments.
