Annelies Gamble, writing in the Mixture of Experts newsletter, argues that the slow step in AI-assisted science is no longer coming up with ideas. It is finding out which ones are right. Her piece builds on a conversation with Mihai Codreanu, who works as an economist on Google’s research staff and co-led AI in Science: Early Insights, a study by MIT FutureTech together with Google and Google DeepMind.
The study analyzed 15 million Gemini interactions and 2,690 specialized models, and it surveyed 637 researchers in the US and UK. Respondents said AI saves them about seven hours a week. Nearly half also reported a growing backlog of hypotheses nobody has had time to test. Gamble reads that as a gulf between the results scientists can design or forecast and the ones they can confirm.
On where the bottleneck sits today, she reports that 44 percent of surveyed scientists say their main constraint has moved downstream over the past two years. Nearly a quarter put it in physical experiments or data collection, and the share climbs to about a third in the physical sciences and engineering. Computational fields have an edge, because in Codreanu’s words they “can close the entire loop in silico.” Even there, access to compute, expensive large-scale simulations, and thin data or benchmarks still hold work back.
The cost that surprised Codreanu most was what he called “the sheer size of the verification tax.” Gamble relays that 88 percent of scientists spend more than 10 percent of their saved time checking AI outputs, and close to half give over a quarter of it to that job. Checking means different things by field. In math and computer science it is reviewing a proof or code. Researchers in health and life sciences devote the biggest share of saved time to validation, and for them it means wet-lab testing. In physical domains, checking the output and running the experiment are the same act. Codreanu does not expect the tax to vanish as models improve, since clinical testing, replication, and expert judgment sit at the core of the scientific method.
How the bottleneck is moving matters as much as where it sits. A growing pile of credible hypotheses forces a choice for any lab with limited experimental capacity. Codreanu told Gamble the study did not assess the quality of the backlog, though he suspects the ideas are strong enough to be worth queuing. He is “bullish” on AI helping to rank them, for instance with scoring frameworks that weigh feasibility, novelty, and expected value. Gamble points out that the hard part is choosing what to optimize: an easy test of an incremental idea against a costly, uncertain one with far more upside.
She also flags a second tension. Most respondents said AI gives them access to other disciplines and encourages broader, more ambitious agendas. About half also said it nudges their own projects toward safer ground. Codreanu calls the risk a possible “streetlight effect,” where researchers pick problems with plenty of data and established benchmarks because those are the ones AI makes cheap. He wants to track citation patterns over time to see whether diversity across subfields grows or shrinks.
Most of the industry’s pitch for AI in science is counted in hours saved. This survey suggests the hours did not disappear; they moved to the bench, the wet lab, and the review pile. A tool that generates hypotheses faster is selling into the stage that is already ahead.
For a lab weighing an AI purchase, the sharper question is how many extra experiments per month it can run, because that number, not the size of the idea list, sets the pace of discovery.
Reported by Annelies Gamble in the Mixture of Experts newsletter, published 20 October 2026.
