Forty-four percent of surveyed scientists report their bottleneck has moved downstream to the bench

The cost of checking is where the report's authors put the asymmetry. Mihai Codreanu, a research economist at Google and a lead author, said "89% of those who save time using AI spend more than a 10th of the time that they're saving actually checking AI outputs," and added that "46% spend more than a quarter." ^[8][6][7] For that second group, ten hours of saved work returns at least two and a half hours of verification, leaving seven and a half at best. ^[27] Codreanu said the checking burden shapes which tasks AI can take and how useful it is field by field: "There is a lot of heterogeneity." ^[10][9] A mathematical proof can be checked against formal rules. A predicted protein function has to be built in a lab and tested in a living system. ^[11]

The survey findings are self-reports, and they are modest in absolute terms: 44 percent of 637 is about 280 scientists, and 41 percent is about 261. ^[25][26] Respondents came from online panels, so frequent AI users may have been overrepresented. ^[3] The Scientific American account does not break the 637 down by discipline, and field-level splits are the substance of the heterogeneity claim. ^[28] A shift in where a scientist reports their bottleneck is also not a count of experiments run. ^[4]

The researchers quoted point to different causes for the jam. ^[29] James Zou, a Stanford biomedical data scientist who has built fully virtual automated labs and handed research to tools including DeepMind's AlphaFold, described a hardware limit. ^[12] "They're still fairly limited in the kinds of experiments that can be automated, mostly some chemistry experiments," Zou said. ^[13] "But things that involve, let's say animals? It becomes much harder." ^[14] He also said existing automated labs can be prohibitively expensive. ^[15] Daron Acemoglu, the Nobel laureate MIT economist who studies AI's labor impacts, located the problem in the questions themselves. "When you look at the tasks where AI has made much progress, they are the ones where you have ground truth," he said. "In most of medicine, there is no ground truth." ^[17][18] The first diagnosis is a capital problem. The second says some questions have no checkable answer to automate against.

Some of the slowness is deliberate. Clinical trials move at the pace of regulatory review, and any experiment involving people is constrained by safety checks before, during and after. ^[19]

Money is moving toward the bench anyway. Julius B. Lucks, principal investigator of the DREAM Cloud Lab at Northwestern University, is trying to automate more of the protein-engineering cycle, and his lab received $20 million from the National Science Foundation over the summer as part of a new national network of programmable cloud laboratories. ^[20][21] The same agency told researchers in a memo last week that AI will "transform" the questions they can answer and how science is done. ^[22]