Language models are often framed as tools that will give researchers more time to think. A theoretical economics paper by researchers from Princeton, the University of Washington, and other institutions argues that the opposite can happen: AI can make scientists more productive in volume while making individual projects less careful.
The central point is simple. If AI reduces the time needed for research work, a scientist’s remaining time becomes more valuable. That changes incentives across the whole research process, including which projects get started, which get abandoned, and how much extra effort goes into work that is already publishable.
The Time-Saving Assumption
The optimistic case for AI in science is straightforward. Language models can help with hypothesis development, data analysis, writing, formatting, and other routine tasks. If those steps take less time, researchers should have more room for deeper analysis and better judgment.
The paper challenges that assumption by using an idealized version of LLMs. In its model, the tools reduce time costs, introduce no errors, and cost almost nothing financially. That setup is important because it separates the effect of saved time from known problems such as hallucinations or inaccurate outputs.
Even under those favorable conditions, the model does not show a simple improvement in research quality. Instead, it shows that time savings can push scientists to reallocate effort away from polishing existing work and toward starting something new.
A Foraging Model For Research
The researchers use optimal foraging theory from behavioral ecology. In that field, the framework explains how organisms distribute effort across different opportunities. Applied to science, it becomes a way to model how researchers divide labor among possible projects.
In the model, a project has two broad phases. First, the researcher checks whether an idea is viable. Then the researcher decides whether to drop it or continue.
If the project continues, it includes mandatory work such as creating figures, formatting text, and submitting. It can also include voluntary effort, such as running extra experiments, doing deeper analysis, or polishing the prose. The paper argues that this voluntary layer is the part most exposed when time pressure rises.
The economic idea behind the model is “opportunity cost.” Every hour spent improving one project is an hour that cannot be spent testing or launching another. When AI makes some tasks faster, that tradeoff becomes sharper rather than weaker.
Where AI Helps Matters
The paper describes three scenarios, each based on a different point in the research process where AI saves time. The outcome depends heavily on where the acceleration happens.
In the first scenario, AI helps researchers evaluate early ideas. Because it becomes cheaper to restart, researchers become more selective. Only stronger projects move ahead. But even those projects may receive less thorough treatment, because the time saved can be used to begin another project instead. The authors describe this pattern as typical of technical fields.
In the second scenario, AI speeds up publishing-related work, including writing, formatting, and analysis. That makes it easier to move papers out the door. As a result, weaker projects become worth pursuing, more papers enter circulation, and each one can end up shallower. The source describes this pattern as typical of fieldwork-based disciplines.
The third scenario is the exception. If AI speeds up the voluntary deep-dive phase, such as extra experiments or more careful analysis, research quality can improve. In that case, the tool is aimed at the part of the process where scientists already tend to reduce effort because of time limits.
The result is not a general claim that AI always harms research. It is a narrower argument: AI’s effect depends on which part of scientific work it makes cheaper.
More Output Can Create New Bottlenecks
The source connects the model to signs already visible in research and software work. A field report from OpenAI covering eight scientific case studies found up to 60x speedups when rewriting research software. But the bottleneck moved from coding to validation and long-term maintenance.
A METR study found another mismatch between perception and measured time. Experienced open-source developers using AI tools took 19 percent longer to finish tasks, while feeling 24 percent faster.
The publication system shows related pressure. In fields where LLMs make writing faster, submissions are rising and adding strain to peer review. Sakana AI's "AI Scientist-v2" pushed a fully AI-generated paper through an ICLR workshop, with citation errors included. Arxiv responded with tougher penalties, including the threat of a one-year submission ban for hallucinated sources or AI meta-commentary left in the text.
"As a labor-augmenting technology, LLMs increase the opportunity cost of our time, impelling us to do more, less well—rather than the same amount, better,"
That quote captures the paper’s warning. Saved time does not automatically become better science. It can become more submissions, more projects, and more review or maintenance work for others.
Why Institutions Need Different Rules
The paper argues that institutional responses should be discipline-specific. AI is not a uniform accelerator. It changes the research process differently depending on whether it helps with early screening, publication tasks, or deeper voluntary work.
That distinction matters for universities, journals, labs, and peer reviewers. A policy that treats all AI assistance the same may miss the central risk. The issue is not only whether a tool is accurate, but whether it encourages researchers to trade depth for volume.
The source also notes a similar dynamic in software development, described as a tragedy of the commons. Individual productivity gains can create downstream costs for the people who later review, validate, and maintain the output.
The practical implication is clear: AI in science should be judged by where it saves time, not only by how much time it saves. Tools that make deeper work easier may improve quality. Tools that mainly make it easier to start or publish more work may increase output while reducing thoroughness.