Mathematicians In Denial About AI Replacing Them

Artificial intelligence is already performing at “gold-medal” levels on high-stakes mathematics problems, and the shift is likely to accelerate—pushing much of mathematical work toward automated theorem proving while leaving humans to review, interpret, and understand. The most consequential detail isn’t that AI can solve math; it’s that general-purpose reasoning models—rather than narrowly trained systems—can do it, catching many mathematicians off guard.

Earlier this year, both Google DeepMind and OpenAI reported top-tier performance on mathematics Olympiad problems. The surprise came from the method: the systems used general-purpose reasoning models, not bespoke training for those specific contests. Yet prominent mathematicians and commentators weren’t impressed by the comparison. Emily Rhiel, a mathematician at Johns Hopkins University, argued that Olympiad-style questions don’t match the kinds of problems professional mathematicians pursue. Terence Tao, a Fields Medalist, added that the comparison is structurally unfair because AI has a speed advantage and can generate many candidate proofs—something closer to the output of a large group than a single researcher.

That debate echoes an older pattern: calculators displaced human arithmetic, and theorem-proving systems may now displace substantial portions of proof production. DeepMind’s recent work on finding singularities in classic fluid equations—linked in spirit to the Millennium Problem about whether the Navier–Stokes equation can develop singularities—was cited as another sign of ambition. While the cited approach did not use the Navier–Stokes equation directly and instead worked with related two-dimensional fluid equations, the implication was that the endgame is still the hardest, most famous targets.

Some mathematicians are leaning into automation rather than resisting it, including through automatic theorem proving. Funding signals reinforce that momentum: the NSF has launched grant programs to support AI-supported mathematical discovery, and private foundations have joined in. At the same time, there are warning signs that AI-generated mathematics may be contaminating research channels such as arXiv, with at least one conspicuous example where an “error rate” was turned into a “blunder rate,” attributed to an author described as an “expert in AI at Google.”

The core technical friction is that large language models can mimic proof-like structure without reliably verifying logical correctness. Community norms on forums such as Math Overflow and Stack Overflow discourage or forbid AI-generated answers, partly because LLMs don’t “know” whether an argument is logically valid. Even so, the transcript argues that proofs follow repeatable patterns, and LLMs can learn those patterns—just not the truth-checking step that some dedicated math software can perform.

Two additional concerns sharpen the picture. First, LLMs may answer even when a question is ill-posed, failing to flag that no meaningful answer exists. Second, AI proofs may be hard to explain in human-comprehensible terms. Daniel Litt is cited for emphasizing that mathematics is about understanding, not merely producing a correct result.

The likely outcome, then, is not total disappearance of mathematics but a transformation: humans will increasingly outsource proof generation to AI and then sift through results, shifting the discipline toward something more empirical—studying what AI can do. The transcript frames this as a new stage of cultural adjustment, ending with a call to learn how AI works through interactive educational resources.