Why AI Is Fast‑Tracking Theory but Struggling to Keep Up in the Lab

If you ask a mathematician about artificial intelligence, the answer is often a grin and a story about a proof that appeared overnight. Ask a biologist, and you’ll hear about a “parking lot” of ideas that never made it past the notebook. That mismatch is exactly what a fresh report from Google, DeepMind and MIT’s FutureTech team set out to capture.

The researchers surveyed 637 scientists across dozens of disciplines, mined 15 million Gemini chat logs, and catalogued more than 2,600 specialized AI models. Their headline finding? Roughly 44 % of respondents say the biggest hurdle in their work has shifted downstream – toward the physical experiments and data collection stage – while 41 % admit their backlog of untested hypotheses has swelled.

In plain English: AI is handing scientists a flood of new ideas, but the “real” work of testing those ideas in a lab hasn’t kept pace. It’s not that the machines are lazy; it’s that checking their output still takes a lot of human time. “Eight‑nine‑y percent of people who save time with AI spend more than a tenth of that saved time double‑checking the results,” notes Mihai Codreanu, a research economist at Google and co‑lead author of the study. “Forty‑six percent spend over a quarter of the saved time on verification.”

This verification burden matters because not every field can easily confirm an AI’s answer. A mathematical proof can be run through a formal verifier in minutes; a predicted protein function, however, has to be synthesized, folded and examined under a microscope – a process that can take weeks or months.

Stanford’s biomedical data scientist James Zou has been tinkering with ways to hand over more of the research pipeline to large language models and tools like DeepMind’s AlphaFold. He’s built prototypes of fully virtual labs, but he admits the leap to physical automation is anything but trivial. “We can automate some chemistry,” he says, “but as soon as you bring animals or complex cell cultures into the mix, the challenges explode.”

Part of the problem is practical: real‑world robots that move tubes, heat plates or live organisms raise safety and ethical questions that software alone can’t answer. The existing automated workstations are also pricey, making them out of reach for many labs that operate on shoestring budgets.

Economist and Nobel laureate Daron Acemoglu adds another layer, pointing out that AI thrives where there’s a clear “ground truth.” In many areas of medicine, the true answer is elusive or disputed, so AI predictions can’t be neatly labeled right or wrong. That uncertainty slows adoption.

And let’s not forget the deliberate pace of some experiments. Clinical trials, for instance, are bound by regulatory review and safety checks that no algorithm can shortcut without risking lives. So even a perfect AI can’t magically compress a year‑long trial into a few weeks.

Still, some teams are pushing forward. Northwestern’s Julius B. Lucks runs the DREAM Cloud Lab, a $20 million NSF‑funded platform that lets researchers remotely design, build and test proteins with a fleet of robots. “We’re taking it one step at a time,” Lucks says, “and figuring out what the next piece to automate should be.”

All of this paints a nuanced picture. AI is undeniably reshaping the front end of scientific discovery – generating hypotheses, suggesting pathways, even writing code – but the back end, the messy, tactile world of wet‑lab work, remains a lagging partner. Bridging that gap will require cheaper hardware, better safety frameworks, and perhaps new ways of defining “truth” in biology.

Until then, scientists will keep leaning on AI for ideas, while rolling up their sleeves to test them the old‑fashioned way.