AI saves scientists time, so they cut corners: Princeton model finds research quality drops

The pitch for AI in science has always been seductive: let language models handle the tedious stuff, and researchers will finally have bandwidth for the deep thinking that actually produces breakthroughs. A new theoretical paper from Princeton and the University of Washington takes that assumption apart with a simple economic mechanism: opportunity cost. Make a scientist’s time cheaper, and every hour spent polishing a nearly-finished project is an hour they could spend starting the next one. So they don’t polish.

The researchers built a mathematical model borrowed from behavioral ecology — optimal foraging theory, which describes how animals decide when to abandon one food patch for another. Applied to research, it simulates a two-phase project: first, a scientist checks whether an idea is viable, then either drops it or pushes forward. The push-forward phase includes mandatory busywork like formatting and submission, plus voluntary deep work like extra experiments and careful analysis. That voluntary part, the authors argue, is exactly what gets sacrificed when time becomes more valuable.

They modeled three scenarios depending on where AI gets applied. If it speeds up idea evaluation, researchers become pickier about which projects to start but spend less effort on each one that survives. If it speeds up writing and formatting, weaker projects suddenly clear the bar, flooding journals with shallower papers. Only when AI accelerates the deep-dive phase itself — the extra experiments and careful analysis that time pressure normally kills — does research quality actually improve. That’s one scenario out of three.

The pattern is already visible outside the model. OpenAI’s case studies found up to 60x speedups in rewriting research software, but the bottleneck just moved downstream to validation and maintenance. A METR study found experienced developers using AI tools took 19% longer to finish tasks while feeling 24% faster. And the publication system is straining: Sakana AI’s fully AI-generated paper made it through an ICLR workshop with citation errors intact, prompting arXiv to threaten one-year submission bans for hallucinated sources. The paper’s core warning — that saved time doesn’t automatically become better thinking — reads less like a prediction than a postmortem of a transition already underway.