AI Generates More Hypotheses Than Labs Can Test, New Study Finds

Artificial intelligence is creating an unexpected problem for scientific research: too many ideas and not enough ways to test them.

A study released this week by Google, Google DeepMind and MIT FutureTech found that 41 percent of surveyed scientists report a growing backlog of untested hypotheses, while 44 percent say their main research bottleneck has shifted downstream toward physical experimentation and data collection over the past two years. The findings, first reported by Scientific American, suggest AI's impact on science remains uneven across disciplines.

The research combined a survey of 637 scientists with analysis of 15 million Gemini conversations and an inventory of more than 2,600 specialized AI models. While mathematicians report AI has solved open problems and integrated into their workflows, biologists and chemists describe AI as still struggling to deliver comparable acceleration in experimental work and drug discovery.

The verification bottleneck

One key finding explains why AI proves more useful in some fields than others: verification costs. According to lead author Mihai Codreanu, a research economist at Google, 89 percent of scientists who save time using AI spend more than a tenth of that saved time checking AI outputs for accuracy. Nearly half spend more than a quarter of their time on verification.

"A mathematical proof can be checked against formal rules," Codreanu told Scientific American, "but a predicted protein function has to be built in a lab and tested in a living system."

This verification gap creates fundamental limits on where AI can accelerate research. Fields with clear ground truth—like mathematics—see faster progress than those without, according to MIT economist and Nobel laureate Daron Acemoglu. "In most of medicine, there is no ground truth," he noted.

Physical and economic constraints

Stanford biomedical data scientist James Zou, who builds methods for automated research including virtual labs, pointed to practical hurdles. Automated labs remain limited to certain chemistry experiments, while work involving animals or complex biological systems proves much harder to automate. The existing automated labs can also be prohibitively expensive.

Robots handling real-world materials raise safety questions that don't exist in software environments. Clinical trials move at the pace of regulatory review, constrained by safety checks that AI cannot compress—nor should it, researchers note.

Why it matters

This research quantifies a shift in how science operates. As AI lowers the cost of generating hypotheses, the economics of research are changing: ideation becomes cheap while experimental validation remains expensive and slow. For business leaders investing in AI-driven R&D, the findings suggest returns will vary dramatically by field. Drug discovery and experimental biology face structural barriers that pure computational advances cannot overcome. Organizations should calibrate expectations accordingly and invest in physical infrastructure and automation alongside AI capabilities.

Efforts to close the gap

Some researchers are working to automate more of the experimental cycle. Julius B. Lucks at Northwestern University received $20 million from the National Science Foundation this summer to build programmable cloud laboratories that let researchers remotely design, build and test proteins with automated equipment. His DREAM Cloud Lab is part of a new national network taking incremental steps toward closing the hypothesis-testing gap.

The study was detailed in a report titled "AI in Science: Early Insights" by Mihai Codreanu and colleagues, released in September 2026, and reported by Scientific American.