The Next Bottleneck in AI for Science

Google, Google DeepMind, and MIT FutureTech surveyed 637 US and UK researchers and analyzed about 15 million Gemini interactions and 2,690 specialized models to see what AI is doing to science. The headline result: researchers report saving about seven hours a week, but nearly half say their backlog of untested hypotheses is growing. AI is speeding up individual tasks, but turning predictions into validated discoveries is still bottlenecked. Mihai Codreanu, a Google senior research economist who co-led the study, told the author that the micro-level gains are clear, but macro-level discovery depends on everything else in the workflow, including steps outside the scientist's direct control. The bottleneck has moved downstream. Nearly half of participants said their main constraint had shifted to later stages over the past two years, with about a quarter pointing to physical experimentation or data collection, rising to roughly a third in physical sciences and engineering. That suggests AI will speed discovery fastest in fields where hypotheses can be tested computationally, since those can close the loop in silico. Even there, compute access, simulation costs, and high-quality data remain limits. The study also found a large "verification tax": 88% of scientists spent more than 10% of their saved time checking AI outputs, and nearly half spent over a quarter. Verification looks different by field, from reviewing proofs and code to running wet-lab tests, and Codreanu doesn't expect it to disappear as models improve, since clinical testing and expert judgment remain central to the scientific method. The study also raises a selection problem. With a growing list of plausible hypotheses, labs must decide what to test, weighing feasibility against expected value, and a cheap incremental test can crowd out a costly, uncertain one with bigger payoff. About half of respondents said AI pushed them toward safer projects even as it broadened their access to ideas from other fields. Codreanu worries about a "streetlight effect," where AI makes data-rich, benchmarked problems easier to pursue than harder questions. To judge the impact, he says he'd need longitudinal data tracking citations and whether a result opens new research directions, because being correct is not the same as being useful.