Think of a map that smoothly stretches and warps as you zoom in, but never folds over itself or tears. In geometry, as long as your transformation doesn’t flat-line anywhere — meaning its local scale factor, or Jacobian determinant, never hits zero — you would expect it to cover the space cleanly, without any two points landing on top of each other.
For 87 years, mathematicians believed this intuitive rule was ironclad for polynomial transformations of complex space. It seemed almost self-evident: if a map nowhere squeezes space down to zero, the global picture ought to be a perfect one-to-one transformation.
Then, on 19 July 2026, that assumption fell to a single polynomial map short enough to fit in one social media post — and, remarkably, it was found with the help of an AI model.
What Was the Jacobian Conjecture?
Formulated by mathematician Ott-Heinrich Keller in 1939, the Jacobian Conjecture is one of the most famous open problems in algebraic geometry. It concerns polynomial maps of complex space: if a map F: Cⁿ → Cⁿ has a Jacobian determinant that is a non-zero constant everywhere, must F have a polynomial inverse — that is, must it be perfectly reversible?
(Over the complex numbers, “non-zero everywhere” and “non-zero constant” turn out to be the same condition: a non-constant polynomial always has a zero somewhere in complex space, so if the determinant never vanishes, it has to be constant.)
Keller originally posed the question in general dimension, though he flagged even the two-dimensional case as difficult. Mathematician Stephen Smale later placed it on his influential 1998 list of mathematical problems for the 21st century.
- In 1D (n=1): The rule is trivially true. A polynomial with a nowhere-zero derivative is just a strictly increasing or decreasing curve — it never turns back on itself.
- In 2D (n=2): The conjecture stood as one of algebraic geometry’s most stubborn open problems, resisting a long parade of attempted proofs (including a flawed one from 1884) and remaining unresolved even today.
Because non-zero local scaling prevents local collapse, generation after generation of mathematicians assumed the rule would simply carry over into higher dimensions too.
How an AI Helped Break the Rule
The break came from Levent Alpöge, a Harvard-trained mathematician working at Anthropic. Together with colleague Akhil Mathew, he put the problem to Anthropic’s Claude Fable 5 AI model — and posted the result on X on 19 July 2026: an explicit degree-seven polynomial map from C³ to C³, just 216 characters long.
Its Jacobian determinant is identically −2 — a non-zero constant, satisfying the conjecture’s hypothesis perfectly. And yet three distinct input points all map to the exact same output. The map is not invertible. The conjecture, at least for n = 3, is false.
The counterexample was short enough that other mathematicians could verify the arithmetic by hand or with tools like Wolfram Alpha within hours; it was later machine-checked in the formal proof language Lean. Because any counterexample in three dimensions can be padded out with extra “dummy” variables, the result immediately implies the conjecture is false in every dimension greater than two — leaving only Keller’s original two-dimensional case unresolved.
[ 2D Space ] [ 3D+ Space ]
Local non-zero Jacobian Local non-zero Jacobian
│ │
▼ ▼
Conjecture Remains Open Conjecture Is False
(Global 1-to-1 status unresolved) (Hidden collisions exist)It was the mathematical equivalent of checking every square inch of a road for obstructions, finding none, and yet discovering three different vehicles parked in exactly the same spot.
Why Did Intuition Fail?
Fields medallist Terence Tao published a detailed “digestion” of the construction just two days later, and mathematicians David Speyer and Patrick Gallagher rapidly generalised it further. At its heart, the trick exploits a simple fact about complex polynomials: a generic cubic curve in two variables can always be factored into three linear pieces. That means there are three different ways to split it into a “linear polynomial times a quadratic polynomial” pair — and multiplying each pair back together lands on the very same cubic. The map behaves perfectly well close up (it’s locally invertible everywhere), but three genuinely different starting points converge on one output.
More generally, higher-dimensional polynomial maps have enough algebraic freedom to route different points around each other and let them collide far from where any local measurement would notice. Unlike closed shapes such as spheres, the infinite, unbounded space of Cⁿ gives polynomial curves room to loop back from an unexpected direction — invisible to any test that only looks at a single point at a time.
Why This Shattered the Rule
This wasn’t a minor technical curiosity — it sent shockwaves through pure mathematics.
- A cautionary tale for higher dimensions: it proved that dimension 3 (and above) possesses topological freedoms that dimension 2 may not — you cannot simply assume 2D geometric intuition carries over.
- A wave of generalisation: within days, other mathematicians had built an infinite family of such counterexamples and produced explicit examples in four and five dimensions too, each with more points colliding at once.
- A live question for algebraic geometry: it forces a re-examination of related conjectures — including the Dixmier and image conjectures — that lean on Jacobian-style conditions for global invertibility.
It’s also worth noting some pushback in the mathematical community: a few commentators pointed out that the underlying technique draws on ideas already present in earlier work on the real-variable case from the 1990s, and that the excitement partly reflects the novelty of an AI model contributing to the discovery, rather than a wholly new proof technique. Formal peer review of the result was still pending at the time of the announcement, though independent verification and rapid generalisation by several mathematicians followed within days.
The Bigger Lesson
The collapse of this 87-year-old rule is a masterclass in the nature of mathematical truth. Intuition is a powerful guide, but higher dimensions are notoriously counterintuitive — and 2026 has been a striking year for that lesson, with AI systems also contributing to a disproof of the Erdős unit distance conjecture and solutions to several other long-standing Erdős problems.
A rule can hold for nearly a century, tested and trusted by thousands, only to be dismantled by a single polynomial — 216 characters long — that finds the one gap in the architecture of space.
Frequently Asked Questions
What is a Jacobian determinant?
A calculation derived from the matrix of all first-order partial derivatives of a vector-valued function. It represents the factor by which a map locally expands or shrinks space.
Does this mean the entire Jacobian Conjecture is false?
It’s now false for every dimension n ≥ 3, following the July 2026 counterexample and its generalisations. The original two-dimensional case posed by Keller in 1939 remains open. (A separate, older result — Pinchuk’s 1994 counterexample — had already shown the real-variable version of the conjecture false in two dimensions, but that concerned a different, weaker condition than Keller’s complex-variable conjecture.)
Who actually found the counterexample, and how?
Mathematician Levent Alpöge, working at Anthropic with colleague Akhil Mathew, used Anthropic’s Claude Fable 5 AI model to construct the explicit polynomial map and posted it publicly on 19 July 2026. It was independently verified by other mathematicians within hours and later formally checked in the Lean proof assistant.
Why is 3D so different from 2D in this context?
Three-dimensional (and higher) space introduces extra algebraic and topological degrees of freedom. Distinct points can be routed to the same destination in ways that are simply not available on a two-dimensional plane.