AI May Push Scientists to Publish More, But Less Thoroughly, Study Argues

The study argues that language models do not automatically turn saved time into better research. Instead, when AI removes work from a scientist’s schedule, the remaining hours become more valuable, creating a stronger incentive to start new projects rather than improve existing ones.

The Decoder frames this as an opportunity-cost problem: every hour spent polishing a current paper is an hour not spent launching the next one. In the model, that shift can reduce the thoroughness of individual publications.

The researchers use a mathematical model based on optimal foraging theory, adapted from behavioral ecology, to describe how scientists allocate effort across competing research opportunities. A project is modeled in phases: first checking whether an idea is viable, then deciding whether to abandon it or continue.

Once a project moves forward, the model separates mandatory work, such as figures, formatting, and submission, from voluntary work, such as extra experiments, deeper analysis, and stronger writing. The voluntary layer is the part most at risk when AI makes starting the next project more attractive.

The study outlines three scenarios based on where AI speeds up the research process. If AI helps evaluate early ideas, researchers may become more selective, but even promising projects may receive less follow-through because moving on is cheaper.

If AI speeds up publishing tasks like writing, formatting, and analysis, weaker projects may become worth pursuing, increasing output while making individual papers shallower. The only modeled case where quality improves is when AI directly accelerates the voluntary deep-dive phase, such as extra experiments or more careful analysis.

The Decoder notes that AI’s effects are not uniform across research fields because the outcome depends on which part of the project lifecycle becomes faster. The article connects the model to broader pressure on publication systems, including rising submissions and strain on peer review where writing becomes easier.

The practical takeaway is that AI adoption in science should be evaluated by where it saves time, not just how much time it saves. Tools that support validation, deeper analysis, and careful follow-up may be more likely to improve research quality than tools that mainly accelerate output.