Even if language models worked perfectly, they could make research worse, not better. A new theoretical study argues that because AI saves time, researchers will spend less effort on each project, not more.

Language models are supposed to speed up science at every stage, from developing hypotheses to analyzing data to writing papers. The obvious hope is that if AI handles the tedious routine tasks, 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. 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 can't go toward starting a new one. Economists call this "opportunity cost," and the paper's model builds on that idea.

For their analysis, the authors deliberately idealize LLMs. They treat them as tools that cut time costs without introducing errors and at negligible financial cost, a setup designed to isolate the pure effect of time savings from the technology's well-known weaknesses.

A foraging model for scientific effort

The researchers built a mathematical model based on optimal foraging theory from behavioral ecology, a framework that describes how organisms allocate effort across competing opportunities. Adapted to science, the model simulates how researchers distribute their labor across projects and what happens when LLMs shorten different phases of the project lifecycle.

In the model, a research project unfolds in two phases. The researcher first checks whether an idea is even viable, then decides whether to abandon it or push forward. Moving forward involves a mandatory part like creating figures, formatting text, and submitting, plus a voluntary part like running extra experiments, doing deeper analysis, or polishing the prose. That voluntary part is what gets sacrificed when time becomes scarce, they argue.

Two out of three scenarios lead to worse research

The paper lays out three scenarios depending on where AI gets applied in the research process. In the first, 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.

In the second, AI helps with publishing by speeding up writing, formatting, and analysis. Because getting a paper out the door 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.

Only in the third scenario does AI actually improve quality. Here, it speeds up the voluntary deep-dive phase, things like extra experiments or more careful analysis. Because AI targets the exact stage where researchers have always cut corners due to time pressure, the time savings translate into more thorough work.

In two out of three cases, then, thoroughness drops. When time becomes more valuable, polishing a paper that's already publishable no longer makes sense. That time is better spent on the next project.

The fallacy of saved time

"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," the authors write. The idea that saved time automatically flows into deeper analysis doesn't hold up.

What the model describes in theory is already showing up in practice. A field report from OpenAI covering eight scientific case studies found up to 60x speedups when rewriting research software, but the bottleneck just shifted from coding to validation and long-term maintenance. The perceived time savings don't even have to be real to change behavior. A METR study found that experienced open-source developers using AI tools actually took 19 percent longer to finish tasks, even though they felt 24 percent faster.

The friction is visible in the publication system, too. In fields where LLMs speed up writing, submissions are already climbing fast and straining the already overloaded peer review system. Sakana AI's "AI Scientist-v2" pushed a fully AI-generated paper through an ICLR workshop, citation errors and all. Arxiv responded with tougher penalties, threatening a one-year submission ban for hallucinated sources or AI meta-commentary left in the text.

Institutional responses need to be discipline-specific, the paper argues, because AI's effect on research isn't a uniform acceleration. It depends on which phase of the process gets sped up. A recent study on software development describes a similar dynamic as a tragedy of the commons, where individual productivity gains come at the expense of the people who have to review and maintain the output later.

Read on for the full picture.
Subscribe for hype-free coverage.

• Full access to every article on THE DECODER
• No ads
• Join the comments and community discussions
• A weekly AI news recap via mail
• 6x/year: "AI Radar" — deep dives on the AI topics that matter most
• Daily AI news, always up to date
• Our full ten-year archive
• Covered by a team with 10+ years in AI