Top mathematicians are outraged by OpenAI’s methods
By [Author Name] Last updated: September 2026 | 13 min read
Why This Matters
Twenty-five of the most decorated mathematicians alive — every one of them a Fields Medal winner, the field's equivalent of a Nobel Prize — just signed a public letter accusing AI companies of damaging the culture that makes mathematical research work. This isn't an abstract academic spat. It's a fight over whether AI-generated proofs can be trusted, who gets credit when a machine and a human reach the same answer around the same time, and whether the race to claim "AI solved a famous problem" as a marketing win is quietly eroding decades of scientific norms. If you've been reading headlines about AI "solving" unsolved math problems and assumed that was straightforwardly good news, this dispute is the reason it's more complicated than it looks.
What Actually Happened
The dispute has been building for months, but it came to a head this week. Twenty-five Fields Medalists signed an open letter warning that AI labs are threatening the intellectual culture of mathematics by racing to solve famous problems as demonstrations of raw model capability, rather than as genuine scientific contributions. The letter follows an earlier document called the Leiden Declaration, published in June by a working group of mathematicians and formally endorsed by the International Mathematical Union, which had already raised concerns about how large language model proofs are changing the field.
The immediate trigger was messier and more personal than an abstract policy dispute. NYU professor Tristan Buckmaster publicly accused OpenAI of pressuring him not to credit a collaborator — a mathematician affiliated with Anthropic — after the pair made progress on a hard problem related to the Euler equations. Buckmaster also raised the possibility that OpenAI's own subsequent proof of a related, larger result (the Navier-Stokes equations, a longstanding open problem in fluid dynamics) may have drawn on the same approach he and his collaborator had been developing, after OpenAI became aware of their work. OpenAI has denied that its internal model's solution depended on the researchers' approach, with an OpenAI mathematician stating in a press briefing that the model reached its result through an independent method.
Whatever the outcome of that specific dispute, it's not an isolated incident. Fields Medalist Terence Tao has separately argued that AI could bring mathematics to its most serious identity crisis since Gödel's incompleteness theorems reshaped the field last century. And in a widely mocked episode from a related controversy, an OpenAI executive publicly claimed the company's model had solved ten previously "unsolved" Erdős problems — only for the mathematician who actually maintains the reference website listing those problems to clarify that "unsolved" meant only that he personally hadn't seen a published solution, not that none existed. The model had found existing literature, not broken new mathematical ground, and the claim was walked back.
The Core Complaint, in Plain Terms
Strip away the specific incidents, and the mathematicians' objection comes down to three linked concerns:
1. Speed versus verification. When a result is announced the same weekend it's produced, there's no time for the normal process of independent scholars checking the work, understanding the method, and confirming it holds up. Several of OpenAI's announced results have gone out before independent verification was complete, leaving claims of "solved" resting on the company's own internal review.
2. Attribution. Mathematics has always run on a credit system — who gets named for which idea shapes careers, funding, and institutional trust. When an AI company's model produces a result using methods that closely track unpublished work already underway by human researchers, and that overlap isn't transparently disclosed, it raises exactly the kind of priority dispute the field has spent over a century building norms to avoid.
3. Incentives toward secrecy. This is the part with the longest-term consequences. If mathematicians start to suspect that sharing early-stage work — including using AI coding tools themselves — risks that work being absorbed into a company's next model and then "solved" by that company first, the rational response is to stop sharing early. Mathematics has historically progressed through open collaboration; a shift toward guarded, secretive research would slow the field down for everyone, including the AI companies hoping to learn from it.
A Comparison: Two Different Uses of AI in Mathematics
AI as a Research Tool | AI as a Capability Benchmark | |
Who benefits | The mathematician, whose work speeds up | The company, whose model looks more impressive |
Verification | Built into the normal collaborative process | Often announced before full independent review |
Attribution | Standard academic credit norms apply | Frequently unclear or contested |
Effect on the field | Accelerates genuine progress | Risks distorting incentives toward speed over rigor |
Example from this dispute | Alpöge reportedly using Claude as part of his own research process | OpenAI announcing a marathon-weekend proof as a standalone achievement |
The mathematicians signing this letter aren't rejecting AI's use in their field — several have said plainly that they see real potential for AI to accelerate research. Their objection is specifically to solving problems as a demonstration of model capability, disconnected from the normal scientific process that makes a result trustworthy and properly credited.
Why a Company Would Want to "Win" a Math Problem
It's worth being direct about the commercial logic here, since it explains why this keeps happening. A public claim like "our model solved a decades-old unsolved problem" is enormously valuable marketing — it's concrete, dramatic, and easy to report on, unlike more nuanced claims about incremental capability improvements. That incentive exists independent of whether the underlying research process was sound, which is exactly the gap the mathematicians' letter is trying to name.
This dynamic isn't unique to mathematics — it echoes tensions in medical AI, where a splashy diagnostic accuracy claim can outrun the peer review needed to confirm it, and in software security, where a company might announce a vulnerability "discovered by AI" without giving the researchers who actually reported it proper credit. Wherever a research claim doubles as a product marketing moment, the incentive to move fast and skip steps grows.
Where This Might Go From Here
A few plausible paths, based on how similar disputes have played out in other fields:
Verification standards tighten. Journals and conferences may start requiring a waiting period or third-party review before AI-assisted results can be publicly announced as "solved," similar to embargo norms in medical research.
A parallel credit system emerges. Some mathematicians have floated the idea of a formal disclosure standard — requiring companies to publish exactly what prior published and unpublished work a model's output drew on, reducing ambiguity in attribution disputes.
The chilling effect proves real. If researchers do start withholding early-stage work from public view or from AI-assisted tools out of fear of being scooped, the field could see a genuine slowdown in the open collaboration that's historically driven it forward — the opposite of what AI boosters promise.
Nothing formal changes, and this becomes a recurring pattern. Given how much commercial value companies get from these announcements, the incentive to keep making them likely outweighs reputational pressure from open letters alone, unless institutions like journals or funding bodies attach real consequences.
The Broader Pattern This Fits
This dispute is a specific, high-profile instance of a much broader question every technical field is starting to face: when an AI system produces a result, who verifies it, who gets credit, and who's accountable if it's wrong or improperly attributed? Mathematics happens to have unusually clear, long-established norms for exactly these questions — which is precisely why violations of those norms are so visible and provoke such a strong reaction from a normally understated academic community. Fields with looser existing norms may not even notice the same violations happening.
What to Watch Next
The Leiden Declaration and this new letter both call for institutional responses — funding bodies, universities, and journals adopting clearer standards for how AI-assisted results get verified and credited before they're publicly announced. Whether any of that translates into enforceable practice, or whether commercial pressure simply outpaces it, will likely become clearer within the next year as more of these disputes surface.
https://www.gammateksolutions.com/post/2026-price-comparison-of-hci-hyper-converged-infrastructure-solutions https://www.gammateksolutions.com/post/fortinet-cyber-security-pricing-2026-firewall-cost-guide https://www.gammateksolutions.com/post/amazon-prime-members-just-got-a-huge-new-perk-here-s-how-to-use-it
Comments