Essay: AI Can Now Do Real Math—Just as the US Guts Its Math Pipeline

This arXiv essay frames two simultaneous trends as a self-inflicted strategic wound. AI systems have crossed into producing genuine research-level mathematics—the author points to a May 2026 machine-generated disproof of a long-standing Erdős conjecture on the planar unit distance problem—while cuts to US federal support for the mathematical sciences are eroding the pipeline that trains people who can actually vet what those systems output. The argument is that the ability to verify, interpret, and challenge mathematical reasoning is human infrastructure built up over generations, not a free side effect of machines cranking out theorems.

The core claim is that mathematical capacity deserves to be treated as a strategic national asset on the level of semiconductor manufacturing: it can’t be spun up on demand once the institutions that sustain it are hollowed out. Automating theorem production while shrinking the population capable of understanding those proofs leaves a society dependent on outputs it can no longer audit—automation without understanding.

As a partial remedy, the essay proposes that AI systems performing consequential reasoning be required to expose their decision-critical claims in formal, machine-checkable form. That would convert part of AI reasoning from opaque persuasion into verifiable structure, preserving a check on the results even as human expertise thins.