After Math's Upheaval, Physicists Ask Whether Science Is AI's Next Conquest

Krishnan cites other limits. Good scientists depend on subtleties that rarely reach peer-reviewed literature: negative results and tacit, hard-won intuition. LLMs also homogenize their responses, blunting their usefulness for genuinely novel ideas. And simulations cannot replace reality: designing a material in silico is no substitute for synthesizing it in a lab.

Materials science illustrates the gap. AI-designed materials have historically produced unphysical results or trivial novelties, and the most promising evidence of AI accelerating the field turned out to be completely fabricated. Still, Princeton physical chemist Leslie Schoop, who also consults for Anthropic, says those failures may not predict the future given AI's dramatic recent gains. "It can now do mostly what I can do, except it doesn't have a lab," she quips. She recently asked AI models to predict a novel structure her lab had found through serendipity, intuition and repetition; none succeeded, because "there's no analytical way to do it."

Takeaway: math verifies itself; the messy physical world does not — yet.

Where might AI still fail?

The most ironic candidate is the study of intelligence itself. "My sense is that the hard questions in the field are really not going to be solved by adding intelligence," says Mark Histed, a neuroscientist at the National Institute of Mental Health. Neuroscience needs better data instead: better neural recordings, more thorough studies of biological networks, and new measurements using frontier tools such as optogenetics.

Even with better data, complicated fields face a hard wall. "At some point the amount of data you need to estimate your downstream nonlinear function is greater than can be acquired before the heat death of the universe," Histed says. The problem is chaos: in systems like Earth's weather — and perhaps the brain — infinitesimal initial changes magnify until predictions diverge.

Takeaway: more intelligence cannot fix fields starved of data or ruled by chaos.

What does the 'AI grad student' look like?

In December 2025, Schwartz began feeding graduate-student-level physics problems to Anthropic's Claude Opus 4.5. The model had plenty of limitations, but "it would come back in two minutes with an answer rather than two weeks." He documented the experiment in what he calls "the most important paper I've ever written."

The trick, Schwartz says, is finding "AI-shaped" problems. Last month he and collaborators released 36 preprints across 22 areas — from outlining the geological history of Earth's atmospheric oxygen to extending phylogenetic trees and determining sunspot demographics. The shared features: checkability, robust data, and often the import of a mathematical tool across disciplinary boundaries.

Human experts remain essential. In one case, the model claimed it had improved earthquake forecasting by 15 percent over current methods; Schwartz's independent human expert easily refuted the claim. "Claude convinced me that this was interesting, but then the expert said, 'Not really,'" Schwartz recalls.

Takeaway: AI compresses weeks of work into minutes, but human judgment still audits the output.

How far does this go?

Schwartz predicts the gap will keep closing: "The content of any paper we could write today, Claude could just generate by thinking." Schoop is less sure the curve holds. "You know, every exponential rise eventually flattens out," she says.

Researchers who spoke about the story showed little consensus on what AI will do or how it should be used — but all agreed science stands on the cusp of massive change. "Right now everyone is trying to figure out where AI is going to go, and scientists are not alone in that," Histed says. "There's a lot of uncertainty."

"My wife hates AI. And a lot of people do. It's scary, and it's impersonal and inhuman, and it's encroaching on who we are," Schwartz says. "But as a scientist, I just can't not use the best available tools to solve problems."