How an AI disproved the Jacobian conjecture

On July 19, 2026, while the World Cup final was still pulling attention away from almost everything else, Levent Alpöge was using Anthropic’s public Claude Fable 5 model on a problem Akhil Mathew had suggested: the Jacobian conjecture. What came back was not a polished proof in the usual sense. It was stranger, and in some ways more useful: a concrete counterexample to an 87-year-old claim, announced on social media quickly enough that mathematicians were checking it almost in real time.[1]

That is the short answer to how AI disproved the Jacobian conjecture. The safer answer is more precise: as of July 21, 2026, the result has not gone through formal peer review, and the full prompt history has not been disclosed. But the alleged counterexample is not just a dramatic claim about an AI model being clever. It is a degree-7 polynomial map from C³ to C³ with Jacobian determinant −2, and it sends three distinct inputs to the same output, (−¼, 0, 0).[1][2]

That last sentence is why the story deserves more than the usual “AI solved math” headline. A theorem-sized claim can float around for years if it is vague. A counterexample with explicit polynomials can be attacked immediately. Reports describe multiple mathematicians informally verifying the result, and mathematical commentary around the Xena project says a Lean formalization check has also been part of the verification conversation.[3]

What the conjecture was claiming

The Jacobian conjecture starts with polynomial maps. In ordinary classroom language, a polynomial map takes a point as input and returns another point, with each output coordinate built from polynomials. Instead of one formula like x² + 3x, imagine several polynomial formulas working together on several input coordinates.

The Jacobian determinant measures how such a map locally stretches or flips space. In calculus, a nonzero derivative tells you a one-variable function is locally well behaved near a point. In multivariable calculus, the determinant of the Jacobian matrix plays a related role. The Jacobian conjecture, proposed in 1939, said that if a polynomial map has a nonzero constant Jacobian determinant, then the map should have a polynomial inverse.[4]

For this article, the key idea is even simpler: if a map really has an inverse, it cannot send two different inputs to the same output. Once two distinct inputs land in the same place, the original input cannot be recovered. Alpöge’s reported counterexample is stronger than that minimum failure: it sends three distinct points to the same output, (−¼, 0, 0).[1][2]

The counterexample is the center of the story

The reported object is a polynomial map F: C³ → C³. Its Jacobian determinant is the constant −2. Under the conjecture, that kind of constant nonzero determinant should have forced the map to be invertible. But the map reportedly sends three distinct points to exactly the same output, (−¼, 0, 0), which breaks invertibility and therefore breaks the conjecture in dimension 3.[1][2]

Once a counterexample exists in dimension 3, the usual extension argument makes the conjecture false for every higher dimension N ≥ 3. The dimension-2 case, often called the plane case, is not settled by this counterexample and remains open.[3][4]

This is where the workflow matters. Claude Fable 5 did not publish a paper, referee itself, and declare the subject closed. The model appears to have been used in an expert-guided search over candidate polynomial families. Alpöge supplied the mathematical taste: what to ask for, which candidate shapes were worth testing, what constraints mattered, and how to recognize a result that would actually kill the conjecture rather than merely look interesting.[4]

That distinction is not cosmetic. Plenty of AI-generated math text sounds plausible while being wrong in the third line. A candidate polynomial map, by contrast, can be checked. Compute the Jacobian matrix. Take its determinant. Substitute the three proposed input points. Confirm that they land at the same output. None of that requires trusting Claude’s prose style.

For a student, the important word in that table is “compute.” A symbolic algebra system such as Wolfram Alpha or SymPy can verify the determinant and substitutions if the explicit polynomials are supplied. That does not make the surrounding research easy. It does mean the central failure is not locked inside a private model run or a famous person’s authority.

Why mathematicians were surprised

The Jacobian conjecture had accumulated the kind of history that makes mathematicians cautious. It dates to 1939 and later appeared on Stephen Smale’s 1998 list of 18 major mathematical problems for the 21st century.[4] It was not an obscure puzzle waiting for the first sufficiently bold chatbot prompt.

There is also a useful warning in its publication history: at least five published proofs of the conjecture were later found to contain errors.[1] That does not mean the field was sloppy. It means the problem was deceptive. A proposed proof could survive long enough to enter print and still fail under sharper inspection.

The low degree of the reported counterexample added to the shock. Degree 7 is not tiny in a hand-computation sense, but it is surprisingly low for a counterexample to a conjecture that had resisted decades of work.[1][4] When a long-standing claim fails at a relatively accessible degree, the emotional reaction is partly mathematical embarrassment: the object was not impossibly far away.

That is also why the “AI did it” framing should be handled carefully. The model seems to have helped search a space of possibilities that humans had not efficiently exhausted. The human expert still mattered because most generated algebra is not a discovery. It becomes a discovery only when someone can state the object cleanly, test it, and understand which theorem it refutes.

What is verified, and what is still missing

By July 21, 2026, the public status is best described as informally but seriously checked, not formally peer reviewed. Reports say multiple mathematicians have verified the counterexample, and the Xena Blog commentary discusses Lean-based verification activity.[1][3] That is meaningful. It is not the same thing as a refereed journal publication with a complete discovery record.

• The concrete algebraic claim appears checkable: determinant −2 and three inputs mapping to (−¼, 0, 0).
• The broader mathematical consequence appears limited and clear: the conjecture is false for dimensions N ≥ 3.
• The plane case remains open, so “the Jacobian conjecture is completely finished” is too broad.
• The full prompt history and exact division of labor between Alpöge and Claude Fable 5 have not been fully disclosed.
• Peer review has not yet caught up with the speed of the public announcement.

Those limits do not make the result uninteresting; they mark what can be trusted now. The concrete map can be tested before the surrounding discovery story is complete.

The student lesson is not “let AI answer for you”

Students are going to see this story used in two lazy ways. One version says AI has made mathematical understanding obsolete. The other says the whole thing is probably a hallucination because language models often make mistakes. Neither version helps someone learn calculus, algebra, proof, or responsible AI use.

The better lesson is procedural. Alpöge’s result, if it continues to hold up, shows one way an AI model can be useful: propose candidates, search awkward spaces, generate objects worth testing, and let humans and formal systems apply the pressure. That is much closer to a research collaborator than an answer vending machine.

A student using AI well can borrow that shape without pretending to do frontier research. Ask the model for examples and counterexamples. Ask it to generate test cases. Ask it to explain what would falsify a claim. Then move the work into a system, textbook method, or proof assistant where the answer has to survive contact with calculation. That habit belongs beside any serious AI study workflow, not outside it.

In a homework setting, that might mean asking ChatGPT to produce several functions with a desired property, then checking derivatives yourself. In a research-reading setting, it might mean using a long-context tool to compare definitions across a paper and a set of notes. In a proof-heavy course, it might mean learning why a formal verification system is less forgiving than a fluent explanation. The common thread is not the brand name of the model. It is the discipline of turning generated text into checkable work.

That is why this story sits naturally beside guides to tools such as ChatGPT for studying, NotebookLM, and long-context AI systems. The most useful study tools are shifting from “explain this paragraph to me” toward “help me build, test, and revise a line of reasoning.” That shift is powerful only if the student keeps ownership of verification.

What Claude Fable 5 seems to have changed

Claude Fable 5 appears to have mattered because it helped find a specific object in a space where human search had not been enough. That is a different claim from saying the model “understood algebraic geometry” in the way a specialist does. It is also different from saying it merely copied a known answer. The current public evidence points to an AI-assisted discovery process whose final artifact can be inspected independently.[2][4]

If the result survives the next rounds of scrutiny, the Jacobian conjecture is false for all dimensions N ≥ 3. The dimension-2 case remains open. The degree-7 counterexample deserves attention because it is concrete enough to compute with, and the discovery process deserves more disclosure because prompt histories and failed attempts are part of the scientific record when AI systems participate.[3][5]

For students, the durable lesson is not that AI replaces understanding. It is that future study may increasingly require a new combination of skills: asking a model for useful candidates, checking the outputs with ordinary mathematics, and knowing when a formal system or a human expert needs to enter the loop.

References

1. AI’s solution to 87-year-old riddle takes mathematicians by surprise, New Scientist
2. Anthropic's Claude Fable 5 disproves Jacobian conjecture, Mashable
3. Human mathematicians are being outcounterexampled, Xena Blog, July 20, 2026
4. Fable 5 Jacobian Conjecture Counterexample Alpoge July 2026, explainx.ai
5. Anthropic's Claude Fable 5 Finds Counterexample to 1939 Jacobian Conjecture, ForkLog