Large language models promise to accelerate scientific discovery by automating the drudgery of research. But a new mathematical model from a team including Carl Bergstrom of the University of Washington warns that the same speed gains could unintentionally erode the quality of scientific work, by making researchers less selective about what they publish and less thorough in how they refine it.
The preprint, posted on arXiv by researchers from the University of Chicago, North Carolina State University, and the University of Washington, formalizes what many scientists have intuitively worried about: that LLMs, by reducing the friction of research, may change the incentive structure of science in ways that harm rather than help.
“As a labor-augmenting technology, large language models have the potential to accelerate scientific activity across the research pipeline,” the authors write. But they caution that the effects are not uniformly positive.
The selectivity paradox
The model predicts that LLMs will have opposite effects on publication selectivity depending on how they are used.
In fields where LLMs serve primarily as tools for discovering promising projects, mining literature for hypotheses, identifying underexplored questions, suggesting experimental designs, researchers will become more selective about what they pursue. They can sift through more possibilities and focus their effort on the best ones.
But in fields where LLMs primarily facilitate the process of publishing existing data, drafting papers, generating figures, summarizing results, researchers will become less selective. The friction that once caused them to abandon borderline-negative results or skip a paper that was not quite ready will vanish, and the volume of published work will rise.
The second scenario, the authors note, is likely to dominate in many fields, and it carries risks for the reliability of the scientific literature. More papers, each less carefully vetted by its own authors, could overwhelm peer review systems and make it harder for genuine findings to stand out.
The opportunity cost of speed
The model’s second major prediction is more counterintuitive. By allowing scientists to work more quickly, LLMs raise the opportunity cost of researcher time. Every hour spent refining a paper is an hour not spent starting the next one.
“By allowing scientists to work more quickly, LLMs raise the opportunity cost of researcher time, creating incentives to refine papers less thoroughly before moving on,” the authors write.
The prediction directly contradicts the optimistic narrative that LLMs will free scientists from mundane tasks and give them more time for deep thinking. The model suggests that the freed time will be redirected into more output, not deeper engagement with any single piece of work.
“Enticing as it is to imagine that, by saving us time on mundane tasks, LLMs will provide us with more time to think deeply and develop projects completely, our results temper such hopes.”
A formal framework
The paper develops a simple mathematical model of researcher decision-making, treating time as a resource that can be allocated between discovery (finding new projects) and production (developing and publishing results). LLMs shift the relative costs of these activities, and the model predicts the equilibrium effects on publication volume, selectivity, and thoroughness.
The framework is intentionally simple, the authors acknowledge that real scientific communities are more complex, but it provides a formal basis for concerns that have largely been expressed anecdotally.
Bergstrom, a biologist known for his work on scientific publishing and the sociology of science, brings a track record of empirically grounded analysis to the question. In previous work, he has documented how publication pressure, metric-based evaluation, and perverse incentives have distorted scientific practice. The LLM model extends that framework to the newest technology reshaping research.
What it means for policy
The paper does not argue against using LLMs in science. It argues that the effects are context-dependent and that policymakers, funding agencies, and journal editors should consider how LLM adoption changes the balance of frictions that currently shape researcher behavior.
If the model is right, then simply adopting LLMs without adjusting reward structures, promotion criteria, grant review, journal standards, could accelerate a trend that is already underway: more papers, less reliability, and a system that measures productivity by the pound. The question is not whether to use LLMs in science, but how to design the incentives around them.