AI Didn't Prove the Conjecture. It Broke It.

Let me be upfront about something. There is almost no community discussion to draw on here. I went looking for threads from the last 30 days and came back with nothing. So this isn’t a roundup of how HN or math Twitter reacted. It’s an argument about why this category of event matters, from someone who can’t independently verify the specific claim or its current verification status. Read it as analysis, not reportage.

What the Jacobian Conjecture Actually Says

Ott-Heinrich Keller posed it in 1939. Stripped to the bone: if you have a map built out of polynomials, and the determinant of its derivative matrix is a nonzero constant everywhere, then that map must have a polynomial inverse.

It feels obviously true. A transformation that never collapses anything anywhere should be reversible. That’s the whole intuition.

Nobody has proved it in over eighty years. Nobody has found a counterexample either. Fields medalists have taken runs at it. Papers claiming proofs have appeared and been withdrawn after someone found the crack — repeatedly, enough times that the problem has a reputation. Among number theorists and algebraic geometers, the Jacobian Conjecture is the one that eats careers.

Proofs and Counterexamples Are Not the Same Sport

This is the part people skip past, and it’s the part that matters.

We’ve had “AI solves math problem” headlines before. DeepMind’s AlphaProof, formalization wins in Lean, Terence Tao’s various experiments with LLMs as research assistants. Those are all proof-side stories, and proof-side stories come with a catch.

A proof establishes that something holds in every case. It’s long, structural, and brutally hard to check. When an AI hands you 100 pages of argument, the mathematical community owes you months of reading before anyone can say whether it’s real. That verification bottleneck is why “AI proved X” news tends to age into “AI submitted X, still under review.”

A counterexample inverts all of that. You need exactly one. And checking it is almost trivially cheap. The claim is: here is a polynomial map satisfying the hypothesis with no polynomial inverse. You plug it in. You compute. Within hours you have a verdict, and there is no room for taste or judgment or “well, the argument in section 7 feels shaky.” It’s true or it isn’t.

So counterexamples sit in a very particular sweet spot: the thing machines are best at generating, and the thing humans are fastest at confirming. That combination is what makes this worth paying attention to.

Why Eighty Years of Humans Came Up Empty

Not because mathematicians weren’t trying. Because counterexample hunting is a search problem with an absurd search space.

The candidates are multivariable polynomial maps. Bump the number of variables, the degree, the coefficient structure — and the space explodes combinatorially. Humans cope by pruning. You develop a feel for which regions are worth exploring and which are dead ends, and you skip the dead ends.

That feel is exactly the trap. Eighty years of failure might mean no counterexample exists. Or it might mean everyone pruned the same branches. Mathematical intuition is trained, shared, and passed down through advisors and papers. It converges. A field-wide blind spot is a real thing, and it’s invisible from the inside by definition.

A model doesn’t share that intuition. It has a different one — pattern sense assembled from training data rather than from a lineage of advisors. It will happily rummage through regions a human dismissed as inelegant, because it has no aesthetic to offend. An explorer without taste is a liability in most contexts. In search, it’s an asset.

The Uncomfortable Follow-Up Question

Here’s the twist. A counterexample means the conjecture was false. Eighty years of serious people trying to prove a statement that was never true.

Which raises the obvious question about everything else on the list. Riemann. Collatz. P ≠ NP. How many of those are load-bearing truths, and how many are collective aesthetic hunches that nobody has stress-tested hard enough?

Mathematics has run on a particular workflow for a long time. Someone states a beautiful conjecture. Everyone spends decades trying to prove it. Counterexample hunting stays a low-status activity — it rarely produces papers, it usually fails, and it doesn’t build a career.

Cheap machine search changes that cost structure. When falsification gets cheap, doubting the conjecture stops being a fringe hobby and becomes standard procedure. New conjecture drops, you run a few thousand GPU-hours at it before anyone commits five years to a proof. That’s a genuinely different discipline.

So What’s Left for Mathematicians

The lazy read is “AI replaces mathematicians.” It doesn’t, and the counterexample itself shows why.

A counterexample is not an insight. It’s a fact. The interesting work starts after: Why does the statement fail here specifically? What additional hypothesis would rescue it? Is this a pathological one-off, or the visible edge of a structure nobody has named yet?

Those are human questions. The Jacobian Conjecture has spawned an entire body of partial results and reformulations — a counterexample doesn’t delete that work, it redirects it. Somebody has to figure out what the right conjecture was all along.

Machines are good at searching broadly. People are good at asking why. The division of labor is getting unusually legible.

The Takeaway

The headline here isn’t that an AI did mathematics. It’s that an AI walked directly into the region human intuition had written off, and produced something checkable in an afternoon.

Expect more of this. Over the next few years I’d bet on at least a couple of long-standing conjectures falling — not to clever proofs, but to counterexamples that turn out to have been sitting in the ugly part of the search space the whole time. Things we believed because they were beautiful.

Which is worth sitting with for a second. What are you currently treating as settled, that nobody has seriously tried to break? It doesn’t have to be a math problem.