Language models are supposed to speed up science at every stage, from developing hypotheses to analyzing data to writing papers. The hope is obvious: if AI handles the tedious routine work, researchers will have more time to actually think.

A theoretical economics paper by researchers from Princeton, the University of Washington and other institutions pushes back on that assumption. The thesis: when AI takes work off a scientist's plate, that scientist's time becomes more valuable. Every hour spent on an existing project is an hour that cannot go toward starting a new one. Economists call this "opportunity cost," and the paper's model builds on it.

The perfect-AI assumption

The methodological choice is notable. The authors deliberately idealize language models: they treat them as tools that cut time costs without introducing errors and at negligible financial cost.

The point of that setup is to isolate the pure effect of time savings from the technology's well-known weaknesses. So "AI produces wrong information" is not the subject here. The question is sharper: what happens even if the model never errs?

A model adapted from foraging theory

The researchers built a mathematical model based on optimal foraging theory from behavioral ecology, a framework describing how organisms allocate effort across competing opportunities.

Adapted to science, the model simulates how researchers distribute their labor across projects and what happens when language models shorten different phases of the project lifecycle. In the model, a project unfolds in two phases:

  • Screening — the researcher first checks whether an idea is even viable, then decides whether to abandon it or push forward.
  • Executing — moving forward involves a mandatory part (creating figures, formatting text, submitting) plus a voluntary part (running extra experiments, deeper analysis, polishing the prose).

The authors' emphasis falls on the second: when time becomes scarce, what gets sacrificed is the voluntary part.

Two of three scenarios go badly

The paper lays out three scenarios depending on where AI gets applied.

First: early idea evaluation. AI helps evaluate early ideas. Researchers become pickier because starting over is cheaper, so only the most promising projects move forward. But even those get less thorough treatment, since the time saved is better spent launching something new. The authors say this pattern is typical of technical fields.

Second: the publishing stage. AI speeds up writing, formatting and analysis. Because getting a paper out takes less effort, weaker projects become worth pursuing. More papers enter circulation, but each one ends up shallower. This pattern is typical of fieldwork-based disciplines.

Third: the deep-dive phase. This is the only case where quality actually improves. When AI speeds up the voluntary deep-dive work — extra experiments, more careful analysis — it targets exactly the stage where researchers have always cut corners, so the result gets better.

Why it matters

The value of this study is that it introduces the missing variable in the AI productivity debate: where the saved time goes.

The common assumption about productivity tools is that saved time goes into doing the same work better. The model shows the opposite: in a system that rewards competition and output count, saved time goes into starting new work.

The finding is not specific to science. The same logic holds in any job measured by output count — software, journalism, consulting. Where the tool gets applied matters more than how much it speeds things up.