Mathematicians Are Using AI Tools They Publicly Distrust
Powerful AI models pose an existential threat to mathematics research, yet mathematicians keep relying on them because they work.
Mathematicians have spent the past two years warning that large language models threaten the foundations of their field, and then using those same models to do their jobs. That contradiction is the subject of a new Wired report describing a discipline caught between professional anxiety and practical dependence. The tension is not abstract: researchers who argue AI undermines mathematical rigor still reach for it when the work gets hard.
The fear is not primarily about lost jobs, though that is part of it. Mathematics prizes proof — a claim is true only when a human can verify each step — and language models generate plausible-looking arguments without any guarantee that the reasoning holds. A model can produce a proof that reads correctly and collapses under inspection. That failure mode strikes at the discipline's core method, not just its workflow.
The reliance, meanwhile, is easier to explain. Models are useful for the unglamorous parts of research: drafting code, checking cases, translating between formal and informal notation, and scanning literature that no individual can keep up with. For a mathematician facing a 400-page paper or a stubborn lemma, a tool that offers a candidate approach in seconds is difficult to refuse. The result is a field that criticizes the technology in public while quietly building it into daily practice.
The stakes extend beyond mathematics. Proof assistants and formal verification have become central to how AI labs test whether their systems can reason at all, and mathematicians are the people who judge whether those systems succeed. If the discipline outsources its judgment to the models it is supposed to evaluate, the outside world loses an independent check on claims about machine reasoning — a check that labs, funders, and regulators currently have few substitutes for.
What comes next is a question of norms more than tools. Some researchers want disclosure requirements for AI assistance in published work; others argue that policing usage is pointless when the productivity gap is so wide. The unresolved issue is whether mathematics can adopt these systems without quietly loosening the standard of proof that makes the field worth trusting in the first place.