AI Tools Drive More Papers With Less Polish
Doing More, Less Well
Scientists who use large language models - the AI tools trained on vast amounts of text to generate and refine language - will spend less time perfecting their work, according to a new modelling study. They will instead jump quickly to fresh projects. The authors write that adopting these tools will cause scientists to “do more, less well — rather than the same amount, better.”
The prediction demands that a comfortable assumption be abandoned. Many expected that AI would let scientists take the same amount of work and make it better. The model says the opposite. Time saved by handing writing, figures and editing to a machine will not flow back into the same paper. It will flow toward the next project, and the one after that.
The authors are careful to add that the tools do not bear all of the blame. The model describes how scientists will reallocate their effort once the machines are in place. The direction of that shift is the problem. It points away from care and toward volume.
A Foraging Lens on Research Habits
To reach that prediction, the researchers divided scientific work into discrete phases. First comes a discovery phase, when scientists brainstorm hypotheses and run initial experiments to determine whether a project is worth pursuing. Then comes a development phase with two parts.
The two parts are not equal. Required work includes creating figures and drafting papers - tasks that must be done. Discretionary development includes follow-up experiments and polishing the writing - tasks that can be shortened or dropped when time runs short. Once an LLM handles part of the required work, the model asks, where does the saved time go?
The framework comes from optimal-foraging theory. In biology, this theory analyses how an animal maximizes energy gain while conserving resources. The authors applied the same logic to scientific effort.
The model was generous to the technology. It assumed the LLMs are functioning at their best - cheap, fast and accurate. Even under those ideal conditions, the predicted result is more output and less care. The study was posted on the arXiv preprint repository — an open online archive where researchers post papers before formal peer review — on 19 July and has not yet been peer reviewed (that is, checked and approved by independent experts before publication).
The Mirror Science Needs to Face
The study’s co-author, Carl Bergstrom, a biologist at the University of Washington in Seattle, points to a deeper cause. [1] The results reflect a flawed incentive system in science, one that prioritizes quantity over quality.
Scientists are incentivized to churn out papers, Bergstrom says. Large language models help them reach this ever-growing quota. The faster a paper is finished, the sooner the next one can begin. Quantity wins, and the quota keeps moving.
The questions the model raises are already visible outside the model. What the model describes in theory, the published record appears to show in practice.
That leaves a contradiction at the centre of the debate. “LLMs are rarely the problem themselves,” Bergstrom says. “LLMs hold up a mirror to problems that we already have.” If the tools are only a mirror, then making the tools better will not make the science better. The contradiction has not yet been resolved.