Claude Fable Cracks the Jacobian Conjecture: Math Enters the 'Machine Proposes, Humans Verify' Era

A Counterexample You Can Compute

On July 21, 2026, mathematician Alpo Gerard publicly presented a set of three-dimensional polynomial maps found by Claude Fable. The construction is striking in its simplicity: the Jacobian determinant is identically equal to -2, yet the map sends three distinct inputs to the same output. That violates the injective behavior the Jacobian conjecture would require, and it does so in a way anyone can check with direct computation.

The Jacobian conjecture, a long-standing problem about polynomial maps with constant nonzero Jacobian determinant, has resisted proof and disproof for decades. What makes this event unusual is not just the outcome but the form of the evidence. The model did not deliver a hundred-page proof requiring years of expert refereeing. It delivered a short object: a concrete polynomial map whose relevant properties can be verified by running the numbers.

Not a Flashy Proof, but a Checkable Object

Much of the public conversation about AI and mathematics has fixated on proofs, on whether a model can reason its way through an argument the way a human mathematician would. This case suggests a different and arguably more disruptive contribution. The scarce insight a counterexample provides is compressed into an artifact that is:

• Short. The map itself is a compact object, not a sprawling manuscript.
• Executable. Its key claims, the constant determinant and the colliding inputs, can be checked by direct calculation rather than by interpretive judgment.
• Falsifiable. If the arithmetic were wrong, anyone could show it. The claim carries its own test.

This is what the framing 'machine proposes, humans verify' actually means in practice. The model's output is not an assertion to be believed. It is a candidate to be run. That distinction matters enormously, because it changes what the human side of the collaboration is responsible for.

The Bottleneck Moves to Verification

The Xena Project discussion, which drew substantial attention on Hacker News with 694 points and 447 comments, zeroed in on exactly this shift. Once a model can cheaply generate candidate counterexamples at scale, the question 'should we trust the model' becomes less useful than a set of operational questions: Can we run the computation? Can we reproduce the colliding inputs independently? Can we formalize the check so it is machine-verifiable and citable?

If generation becomes abundant, verification becomes scarce. The mathematics community would then need not just more generative capability, but public, fast, citable verification pipelines: infrastructure for independent recomputation, for minimizing constructions, and for formalizing checks in proof assistants. This is a tooling problem, not a trust problem. And it is a problem the community can actually engineer its way out of, which is precisely why the Xena Project conversation framed the event as a turning point rather than a curiosity.

Who Is the Discoverer?

The tempting answer is that the model is merely a tool, that the discovery still belongs to the humans who posed the question and verified the result, just as a calculator does not become a mathematician by finding a prime. That answer is not wrong, but it misses how the production structure of discovery is changing.

The scarce step in mathematical work has often been described as inspiration: finding a construction worth trying at all. In this event, the model searched an enormous space and returned a short, checkable object. The human contribution migrated toward designing the search environment, recognizing that the output mattered, and building a credible chain of verification. Credit, in other words, is becoming a question about a pipeline with distinct roles, and the community's task now includes separating the contributions of model generation, question design, and formal verification rather than collapsing them into a single heroic act of discovery.

Falsifiability Before Understanding

Perhaps the most counterintuitive implication is the order in which AI may automate parts of mathematics. A model does not need to understand everything a field means in order to force a theory to update. It only needs to produce a counterexample that runs independently. Falsifiability, it turns out, may be automatable before explanation is.

That does not make mathematicians obsolete. It makes their work look more like security research: reproduce the finding, minimize the construction, delineate exactly where the original conjecture's scope breaks, and harden the verification chain so the result can be cited with confidence. The immediate agenda following Gerard's announcement reflects this: independent recomputation, minimization of the construction, clarification of the conjecture's valid range, and a careful accounting of who did what.

The headline will read that an AI overturned a famous conjecture. The reality is more interesting. A model handed mathematics a small, executable piece of evidence, and the field's center of gravity shifted overnight from finding answers to building the machinery that certifies them. The era of 'machine proposes, humans verify' is not coming. As of this week, it is here.