AI Labs Now Run Experiments, But Real-World Proof Lags Behind

AI Is Speeding Up Science, and the Power Bill Is Growing

A robot-run lab called A-Lab was handed a list of 58 materials to make and delivered 41 of them in 17 days, working out recipes from published research and adjusting course when a route failed. A new review in the journal The Innovation holds that up as the real change in AI for science: machines now run the whole cycle that turns data into knowledge, from reading papers to testing ideas. The same review adds a catch, though. Automated checks of A-Lab’s output may mislabel what was actually made, and some materials billed as new turned out to be messier versions of known ones.

More than 100 researchers from universities and institutes in China, the United States, Europe, Australia, and Singapore wrote the review. Their central claim is that AI has gone from a narrow tool to something closer to a research partner, one that can lend a hand at every stage of a project. It is already posting eye-catching results. AlphaProof solved three of five non-geometry problems at the 2024 International Mathematical Olympiad, a result described as silver medal performance. AlphaEvolve found a way to multiply two 4-by-4 grids of numbers using 48 multiplications, improving on a method mathematician Volker Strassen published in 1969. The authors say the most credible path is stronger teamwork between humans and AI under clear validation rules.

AI for Science Now Follows a Six-Step Research Loop

No volunteers or patients were tested, so there is no sample size in the usual sense. This is a survey. The authors sorted published studies from more than 10 fields, including physics, chemistry, medicine, and social science, into six connected stages.

Those stages mirror how a human researcher works. AI reads the literature, hunts for patterns in data, proposes educated guesses, and designs tests, either as computer simulations or on real lab equipment. Then it writes up the results, charts included, and feeds what it learned into the next round. Some systems add built-in self-checks to catch inconsistencies and made-up results before errors snowball. A few studies are even testing whether AI can sit in on peer review, the step where experts vet each other’s work before it is published.

AI Math and Medicine Results Come From Specific Studies

Math is the cleanest scoreboard, since a proof either works or it does not. A model called Intern-S1 scored 102 points to win gold at the 2025 Chinese Mathematical Olympiad, well past the 78-point cutoff. AlphaEvolve, by comparison, often rediscovered known best solutions across dozens of unsolved problems and only occasionally produced real improvements. Still, the review notes that current AI provers need large amounts of computing power, struggle with long or truly new proofs, and need humans to verify an idea before it counts as an accepted result.

Medicine shows gains too, but in individual studies rather than everyday hospital care. A cancer-decision tool that analyzed medical images while pulling up treatment guidelines lifted accuracy from 30.3 percent, with a standard chatbot-style AI, to 87.2 percent. A 3D planning tool for lung surgery improved the spotting of unusual anatomy by 8 percent and cut errors by 41 percent in a study at several medical centers.

Three Levels of Evidence Separate Real Discoveries From High Scores

A high score is not the same as a discovery, the authors argue, and they lay out three levels of proof. The first is doing well on standard tests. The second is passing a sanity check against established science or expert judgment. The third is holding up in a real experiment, a step the review says many fields still lack.

Other problems linger. Good data is scarce. Many models cannot explain why they reached an answer, which makes their reasoning hard to inspect. Some AI-generated results may look plausible while breaking the laws of physics. Messeri and Crockett, cited in the review, warned that AI can create “illusions of understanding,” potentially producing more output but less insight. In medicine, patient privacy and biased training data that could widen health gaps add to the worry.

Data Centers Behind AI for Science May Double Their Power Use by 2030

Scale carries a price, too. Data centers worldwide used about 415 terawatt-hours of electricity in 2024, and the review cites a projection that demand will roughly double by 2030, reaching about what all of Japan uses in a year today. US data centers get about 56 percent of their power from fossil fuels. Training one big model can cost up to hundreds of millions of dollars, which the authors say could confine frontier research to the best-funded institutions.

A-Lab’s 41 of 58 shows what a self-running lab can do, and the review’s own warning shows why a success count is not a discovery count. Until AI-driven results hold up in real experiments, AI for science has earned credit for speed. Proof is still to come.

---