What is the Jacobian conjecture and why did it matter?

Stated in full generality in 1939, the Jacobian conjecture asked whether a polynomial map whose Jacobian determinant is a nonzero constant must have a polynomial inverse. It sat on Smale's 1998 list of mathematical problems for the century, survived at least five flawed published proofs, and resisted careers' worth of effort — until a three-line counterexample ended it.

How was the Jacobian conjecture disproved?

A mathematician working with an AI model produced an explicit polynomial map from three-dimensional space to itself with constant Jacobian determinant -2 in which three distinct points map to the same output. A non-injective map has no inverse, so the conjecture fails. The check requires only polynomial algebra; any computer algebra system can confirm it in seconds.

What does the result mean for AI and mathematics?

The asymmetry is the lesson: verifying the counterexample is undergraduate algebra, finding it defeated professionals for 87 years. When a machine can compress the finding, the scarce human contribution moves up a level — to knowing which question is worth asking, and recognizing an answer when it arrives.

Easy to Verify, Nearly Impossible to Find

July 21, 2026 · By Nbidea

On the night of the World Cup final, a number theorist posted a short message. It was written in lowercase, thanked two friends, and mentioned the football almost apologetically. Buried in the middle of it was a polynomial map in three variables — and the end of a conjecture that had been open since 1939.

The Jacobian conjecture asked something a calculus student can parse. Take a polynomial map from n-dimensional space to itself. Suppose its Jacobian determinant — the local measure of how the map stretches space — is a nonzero constant everywhere. Must the map then have a polynomial inverse? Locally, calculus says yes without hesitation. Globally, for polynomials, everyone believed yes. For eighty-seven years, nobody could prove it.

The problem earned its reputation the hard way. It sits at number sixteen on Smale’s 1998 list of mathematical problems for the century, in the company of the Riemann Hypothesis. At least five published proofs later turned out to contain errors. Careers were spent circling it; doctoral theses were written on it; a field developed institutional reflexes for greeting any new claim with folded arms. The safest sentence in algebraic geometry was “the Jacobian conjecture is still open.”

Three tenants, one address

What arrived on Sunday was not a proof. It was the other thing — the thing that takes three lines. An explicit map from complex three-dimensional space to itself, polynomial in every coordinate, whose Jacobian determinant works out to the constant −2. Nonzero everywhere. Every condition the conjecture demands, satisfied.

And then: three different points that all land on the same output. Three tenants registered at one address. A map that sends distinct inputs to the same place cannot be undone — there is no way to look at the output and know which input produced it. Not invertible, by inspection. The conjecture does not survive contact with this object.

The check is almost insultingly simple. Expand the polynomials, take nine partial derivatives, evaluate a determinant, substitute three points. Any computer algebra system confirms it in seconds; exact rational arithmetic, no floating point, no trust required. We ran the verification ourselves before writing this — thirty lines of code, constant determinant, three collisions, done. A first-year student can grade the answer. Finding it defeated the profession for the better part of a century.

The asymmetry is the story

That gap — trivial to verify, nearly impossible to find — is the oldest asymmetry in mathematics, and it is exactly the shape of what just changed. The counterexample was found by a mathematician working with an AI model, in roughly the time it takes to play a football match. He described the model as a friend who kept working during the final.

It is worth being precise about what happened, because two wrong readings are already circulating. The first wrong reading: the machine did it alone. It did not. A working number theorist chose the problem, framed the search, recognized what came back, and staked his name on it. Colleagues were quick to note that hunting a counterexample in this haystack requires knowing which part of the haystack has ever contained needles. The prompt was not “solve the Jacobian conjecture.” The prompt was the last move of a long game only a trained player could see.

The second wrong reading: the human did it and the machine merely typed. Also no. The search space here is not one a person walks through in an evening while a match plays in the background. Something real was compressed — months or years of trained exploration folded into hours. The mathematicians reacting in public understood this immediately. One called himself very bullish on near-term AI impact on mathematics, and predicted more such results soon. Another, a Fields Medalist, said it was the first time a language model had settled a problem outside his own field that he had definitely heard of. The reaction was not “impossible.” It was “so it begins.”

Where the scarce thing moved

Every technology that compresses a kind of work relocates the scarcity somewhere else. When finding becomes fast, the scarce thing is no longer the search. It is the taste that chooses the question, and the judgment that recognizes an answer wearing unfamiliar clothes.

Notice what the mathematician actually contributed on Sunday night: a question worth asking, asked of the right instrument, at a moment when nobody was demanding he ask it. His friend had mentioned the problem in conversation. He was watching football. The whole thing has the texture of play — and that is not a decoration on the story, it is the mechanism. The conjecture did not fall to a funded program with a roadmap. It fell to curiosity with good tools and an evening off.

There is a lesson here that has nothing to do with algebraic geometry. The instrument is now available to anyone. The question is not. For eighty-seven years the bottleneck was the haystack; now the bottleneck is knowing that this haystack, of all the haystacks, was worth an evening. That knowledge does not live in the model. It lived in one person’s trained sense of where mathematics was soft — and in his willingness to ask a question most of his field had filed under “open, probably forever.”

The result still awaits formal review, and the field’s reflexive caution around this particular conjecture is well earned. But a counterexample is the least deniable object in mathematics: it does not ask you to follow an argument, only to do the arithmetic. The arithmetic says three points, one address, determinant −2.

An eighty-seven-year-old question closed during a football match. The interesting question was never whether machines would learn to search. It is whether we will keep training the thing that cannot be delegated: knowing what to ask.