On a quiet Sunday afternoon while much of the world watched the World Cup final, an AI model produced a result that has occupied mathematicians since 1939: a counterexample to the Jacobian conjecture. The announcement—shared by Anthropic employee Levant Alpöge and widely circulated online—was rapidly verified and has reignited debates about what it means for machines to solve deep, conceptual problems in pure mathematics.
What the model produced
The Jacobian conjecture, originating with Ott-Heinrich Keller in 1939, concerns when a polynomial map can be inverted from its outputs. Alpöge’s result reportedly constructs a mapping whose Jacobian determinant is constant at −2 everywhere yet sends three distinct inputs to the same output, a direct counterexample to the conjecture’s conditions. The post announcing the result drew extraordinary attention on X, reportedly registering more than 20 million views.
The development follows a string of AI-driven advances in mathematics: in mid-2025 language models solved five of six problems at the International Mathematical Olympiad; an OpenAI model later challenged an 80-year-old conjecture in combinatorial geometry in May; and in June a group of researchers published the Leiden Declaration urging new norms around transparency, attribution, and peer review as AI reshapes mathematical practice.
Verification vs. understanding
Colleagues at Imperial College London and elsewhere said the result had been checked quickly. Kevin Buzzard noted the paper had already been formalized in Lean, a proof assistant that checks each logical step mechanically. That machine-checked verification removes ambiguity about whether the construction meets formal criteria.
Yet several mathematicians expressed unease about what getting a mechanically verified answer does and does not deliver. A recurrent concern is that current language models provide the “how” of a construction without the human-style “why”—the narrative, intuition and conceptual story that makes a result intelligible and teachable. Akhil Mathew, whom Alpöge credited with proposing the problem, told reporters that while a result can be checked, it remains valuable to be able to tell the story behind it.
Generating long, delicate proofs still presents challenges. Formal proofs that span hundreds of steps require a level of sustained, gap-free reasoning where language models often resort to plausible-sounding interpolations. Unlike human researchers, models do not face reputational costs if they produce errors, and that difference affects how mathematics is produced and trusted.
Proof assistants and the changing toolkit
Buzzard has long promoted Lean as a way to make mathematical reasoning machine-checkable. He observed that once models produce proofs that are then checked in systems like Lean, a significant barrier between human and machine mathematical labor disappears. In this case, the machine-checked result existed by the time many learned of it, illustrating how proof assistants and large models can be combined to produce formally verified outcomes.
But verification via a checker is not the same as producing a human-comprehensible proof. Mathematicians prize proofs not only for correctness but for explanatory power: a proof that makes the result feel inevitable and gives insight into related problems. That kind of explanation remains predominantly a human achievement.
Implications for careers, culture and research
The broader reaction among mathematicians mixes amazement with anxiety. Some see AI as a new tool that will enlarge what is possible; others worry about the profession’s social and economic future. Observers have noted falling public research support and shifting institutional appetites. The article points to steep cuts to federal funding for mathematics research—reported as roughly a 72% decline under recent NSF budget changes—and to declines in PhD admissions at top programs. One university was reported to be admitting no funded math doctoral students this year.
Those structural pressures compound concerns about the role of mathematicians. Michael Harris argued in a recent essay that industry often treats reasoning as commercially expendable, which could reshape what kinds of mathematical labor are valued. Some have suggested a rediscovery of amateur or independently funded mathematics, but Alpöge’s background—as a Harvard valedictorian who has spent a decade on algorithmic approaches—underscores that experts remain central to posing and interpreting the field’s most consequential questions.
At the heart of the discussion is a question of taste and creativity: machines can verify and sometimes discover constructions, but humans still appear to hold an advantage in asking the right questions. Buzzard emphasized that many landmark mathematical problems are remembered for who posed them as much as for who solved them; formulating deep, fertile questions is a creative act that has not been automated.
Where the field goes next
The episode is likely to accelerate conversations already underway about standards for transparency, peer review, and attribution when AI contributes to mathematical results. The Leiden Declaration and recent commentary from figures across academia and industry recommend new guardrails before AI fully reshapes how mathematical knowledge is produced and certified.
For now, the community is left balancing a validated, machine-checked milestone against longstanding norms about exposition, pedagogy and professional practice. Whether the arrival of AI as a routine creator of formal results will shift the center of gravity in mathematics—or simply enlarge the toolkit for human inquiry—remains an open and contested question.
Source: Fortune