Mathematics and physics: Nature describes a change in the way questions are asked, not in who answers them

The analysis identifies the uses that have become common—exploration of conjectures, formal verification, generation of counterexamples—and emphasizes what automation does not replace.

Nature devotes an analysis to the entry of language models and evidence systems into the daily work of mathematicians and theoretic physicists. The central finding is that human intuition has not been replaced: what changes is the cost of exploring a track. Test a conjecture on thousands of cases, look for a counter-example, reformulate a statement to make it attackable—these steps take weeks and are now done continuously.

The article draws on a series of recent results to show that the audit followed the move. The formal evidence assistants, Lean in particular, serve as a guard: a demonstration produced by a machine can be certified line-to-line, which avoids having to trust the reasoning displayed. This combination — wide generation on the one hand, mechanical verification on the other — is what makes the 2026 announcements stronger than the demonstrations two years ago.

The limits that the analysis highlights remain: the choice of problems that deserve to be posed, the assessment of what constitutes a conceptual advance rather than a successful calculation, and the rereading by the community. On these three points, the bottleneck remains human.