AI Came for Math. Now It’s Chasing the Rest of Science

For the past few years, mathematics has served as artificial intelligence’s proving ground. It was an obvious target: math is self-contained, its claims can be checked line by line, and a proof is either valid or it isn’t. That clarity made it the ideal arena for testing whether machines could do more than summarize text — whether they could reason.

The results have been striking enough to unsettle a discipline that long considered itself immune to automation. Systems that once stumbled over arithmetic now work through competition-level problems, propose intermediate steps, and collaborate with human mathematicians on research questions. Formal verification tools, which translate proofs into machine-checkable code, have given AI a built-in referee. If a model hallucinates a step, the checker catches it. That feedback loop — generate, verify, repeat — turned out to be enormously powerful.

Now the obvious question is whether the same approach can be pointed at the rest of science.

The harder problem

The trouble is that most of science doesn’t come with a proof checker. Biology, chemistry, materials science, and climate modeling all deal in messy, noisy, incomplete data. A hypothesis in cell biology can’t be verified by symbolic manipulation; it has to be tested in a lab, often over months, sometimes with results that don’t replicate. Where mathematics offers a clean signal, experimental science offers ambiguity.

That hasn’t stopped researchers from trying. The strategy taking shape across academic labs and corporate research groups is to find or build the equivalent of a verifier for each domain — a simulation, a physical experiment, an automated instrument — that can tell a model whether its proposal was any good. Protein structure prediction already showed what becomes possible when a field has large, standardized datasets and a well-defined scoring system. The push now is to manufacture that kind of structure where it doesn’t naturally exist.

Robotic labs are part of the answer. So-called self-driving laboratories, where software designs an experiment, machines execute it, and the results feed directly back into the model, are being built for chemistry and materials discovery. The appeal is speed: a system that can run hundreds of variations without waiting for a graduate student to pipette them by hand.

What could go wrong

The enthusiasm comes with warnings. Scientific literature is already strained by a flood of low-quality submissions, and tools that make it trivial to generate plausible-sounding papers threaten to make that worse. Peer review, which depends on scarce human attention, was not designed for machine-scale output.

There are also deeper concerns about what gets lost. Science is not only a search for correct answers but a search for understanding — for the mechanism behind a result, the intuition that lets a researcher ask the next question. A model that proposes a working catalyst without explaining why it works has delivered something useful but incomplete. Several researchers have argued that AI’s real contribution may be in narrowing enormous search spaces rather than producing finished insights.

And then there is the matter of trust. Mathematics can verify its own output. Most sciences cannot, which means claims produced by AI will have to be validated the old-fashioned way, by people with instruments and time.

What seems clear is that the pattern established in math — reasoning paired with rigorous checking — is now the template. The field is racing to figure out which parts of science can be made legible enough for machines to work on, and which will stubbornly resist. The answer will shape how research gets done for the next decade. Read More