Math community frustration with OpenAI has a new focus: the advisory group that was supposed to calm things down. In a report published on 28 September 2026, The Verge spoke to Martin Hairer, a Fields Medallist and one of the group’s nine members, and to mathematicians outside it. They describe a rushed, confusing rollout, and a profession bracing for “more than 100” results that OpenAI says its unreleased model has produced.
Robert Hart’s article sums up the pattern in two lines: OpenAI can “make impressive breakthroughs in mathematics, then colossally screw up announcing them,” and its attempt to do better has been “botched” too. The criticism is not that AI is doing mathematics. It is about how the results reach the math community, who gets credit, and who pays the cost of checking them.
This article sets out what The Verge reported, how a new artificial intelligence advisory body for mathematics came about, who sits on it, what each side has promised in writing, and why so many researchers are anxious. It ends with what a responsible release process could look like, and what the episode means for any organisation publishing AI-generated work.
Table of contents
- What The Verge Reported About OpenAI and the Math Community
- Who Sits on AGMAI, and How It Began
- Why the Math Community Feels Bulldozed
- How Fast the Story Moved
- The 100 Results Waiting in the Wings
- What the Math Community Stands to Lose
- What a Responsible Release Could Look Like
- Why This Matters Beyond Mathematics
- What Comes Next for AGMAI and the Math Community
- Math Community FAQs
- References
What The Verge Reported About OpenAI and the Math Community
The Verge’s piece rests on interviews with three mathematicians and a close reading of the announcements. Its headline, “OpenAI keeps bulldozing mathematicians”, captures a view that has hardened across the math community since early September.
An advisory group few people understood
On 21 September a group of eminent mathematicians announced the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) in a guest post on Terence Tao’s blog. It would be independent of OpenAI and would advise it and other frontier labs “on the review and communication of emerging results.” OpenAI’s own announcement, published the same day, reached far more people. The Verge found that many in the math community assumed the group was OpenAI-appointed, an impression it calls “not entirely unreasonable”. AGMAI’s website says the group formed after OpenAI approached some of its members about an advisory board.
“They obviously are very good PR people”
Hairer told The Verge the group takes “no financial, technical, or other support” from OpenAI. “We don’t work for OpenAI, are not paid by them, and it’s totally independent,” he said. He was still exasperated by the announcement. “The way they phrased their announcement didn’t really help,” he said, adding that “if you read it exactly in detail, there’s nothing wrong in what they say”. He laughed as he added: “They obviously are very good PR people.”
Who Sits on AGMAI, and How It Began
AGMAI lists nine members and is hosted by the Institute for Advanced Study (IAS) in Princeton. Three of the nine are also on record backing “A Severe Misalignment of AI in Mathematics”, the 11 September declaration that set out the math community’s case against the AI labs.
| Member | Affiliation listed by AGMAI | Link to the declaration |
|---|---|---|
| François Charles | ENS-PSL | Not found in the endorser search |
| Camillo De Lellis | IAS, GSSI | Endorser |
| Timothy Gowers | Collège de France, Cambridge | Not found |
| Martin Hairer | EPFL, Imperial College London | Fields Medallist signatory |
| Nikhil Srivastava | Berkeley, Simons Institute | Not found |
| Ulrike Tillmann | Oxford, Isaac Newton Institute | Not found |
| Ravi Vakil | Stanford | Endorser |
| Edward Witten | IAS | Not found |
| Melanie Matchett Wood | Harvard | Not found |
We checked both of the declaration’s lists: the Fields Medallists on its front page and the searchable endorser register at mathandai.org. “Not found” means the name did not appear on 29 September, not that the person opposes the declaration.
Independent, but born of an OpenAI approach
AGMAI’s own account is direct: the group “came together after OpenAI approached some of its members about establishing an external advisory board. In agreement with OpenAI, they decided to instead create an independent group and invite the others to join.” Hairer’s blog adds that not every member was contacted by OpenAI, that nobody else takes part in the group’s conversations, and that its IT, legal and communications support comes from the IAS. The Verge notes it is still unclear which members OpenAI first approached.
What each side has put in writing
The documents agree on independence but differ in emphasis. OpenAI stresses what the group may do. AGMAI stresses what it cannot do. The math community has read both closely.
| Question | OpenAI’s announcement | AGMAI and Hairer |
|---|---|---|
| Is it independent? | “The group will operate independently from OpenAI” | “Operates independently of any AI company” |
| Is it paid? | Members “will not be paid by OpenAI” | Members “do not accept payment” |
| Can it go public? | Free to “make its advice public” | “We will publish our recommendations” |
| Does it have power? | Not on “how to pace our internal progress on mathematics” | “We do not have decision making power at any AI company” |
| Current task | Help assess significance and coordinate dissemination | Advise on releasing “a large number of significant results” |
Why the Math Community Feels Bulldozed
The Verge’s reporting, Hairer’s blog and the September declaration describe the same grievances. None of them is about whether the AI results are correct. They are about how results are released and how that affects people.
Results first, scholarship later
The declaration says AI solutions “are announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others.” Hairer’s blog is blunter, describing recent releases “on privately controlled websites” combined with “abject levels of scholarship and only minimal presentation effort”. Researchers told The Verge of poorly written manuscripts and thin engagement with the literature. Both make it harder for the math community to judge a result’s significance, and they leave earlier work without credit.
Quiet edits and “sloppy scholarship”
The Verge says OpenAI sometimes changed documents after release without announcing the changes or keeping a clear record. It noticed this with OpenAI’s “Ten advances in mathematics and theoretical computer science” in early August. Hairer described times when manuscripts seemed to be altered “sneakily” after criticism. “That also makes people paranoid, right?” he said, calling it “shoddy” and “really bad and sloppy scholarship”. In mathematics, a paper’s version history is part of the record of who knew what, and when.
“That sentence makes no sense in mathematics”
Álvaro Lozano-Robledo of the University of Connecticut objected to how OpenAI talks about its results. “This is not how any academic behaves,” he said. “I don’t go around saying, like, ‘Oh, I’ve proved all these things, but I don’t know what to do with them.'” When OpenAI’s Laurance Fauconnet told The Verge the company had “made substantial progress” on another Millennium Prize problem, he replied: “No mathematician would go out and say that. You either have solved it, or you’re still trying.”
How Fast the Story Moved
The dispute has moved at the speed of product launches, not of academic publishing. The chart counts days from 28 August, when OpenAI says it began training the model behind these results, to each milestone.
A group two days old
Asked about AGMAI’s plans, Hairer laughed: “So we, well, you know, started two days ago. So it’s not like we have a big master plan.” That is the heart of the math community’s unease. A body designed to slow things down was assembled in days, because OpenAI is holding results it clearly wants to release.
The 100 Results Waiting in the Wings
OpenAI’s announcement says the model it began training on 28 August has, besides the Navier-Stokes problem, “resolved more than 100 long-standing open problems across most areas of mathematics”, and that its pace “has surprised the mathematicians within OpenAI.” AGMAI’s website names its current task as advising OpenAI “on how to coordinate the release of a large number of significant results”.
What OpenAI has, and has not, said
OpenAI has not listed the problems. Its announcement says the group “will not be responsible for advising us on how to pace our internal progress on mathematics”, so AGMAI can shape how results are released, not whether the work continues. For the math community, that leaves a known quantity (more than 100 results) attached to unknown subjects, which is precisely the condition that feeds anxiety.
Why the labs chase mathematics
Mathematics suits the way frontier models are now trained. Labs increasingly use reinforcement learning on tasks whose answers can be checked automatically, and a proof written in the Lean language can be verified by a computer line by line. That makes open problems an attractive benchmark: a solved problem is an unambiguous, headline-friendly result. The September declaration argues that this is exactly the misalignment, because the math community values the understanding a solution brings, while a benchmark only counts the answer.
Lean checks correctness, not significance
OpenAI has published a Lean formalisation alongside its Navier-Stokes work, and formal verification genuinely helps. It removes the worry that a machine-written proof contains an error. It does not tell anyone whether a result matters, which earlier ideas it builds on, or what new method it contains. Those judgements still fall to people, and they are the part of the job the math community says OpenAI keeps skipping.
A flood is different from a breakthrough
One breakthrough is news. A hundred, released without context, is a workload. Lozano-Robledo called the stream of AI-generated solutions a “burden”, because “solutions are often less important than the understanding that comes with them.” “They need our expertise,” he said of the labs. “They need us to celebrate that solution.” Someone in the math community still has to read, check, contextualise and teach each result, and that time is unpaid.
What the Math Community Stands to Lose
The Verge’s most striking passages are about people, not proofs. Researchers described fear that problems they have worked on for years “could be next on the chopping block”, dispatched by a company some suspect of chasing publicity before an IPO.
Researchers deciding whether to rush
Colva Roney-Dougal of the University of St Andrews said it is “somewhat agonizing to know that a ‘large number’ of results are likely to be announced soon,” adding: “The more there are, the more likely it becomes that my ongoing work, or that of my students, suddenly becomes irrelevant.” She is unsure whether to “rush out as many papers as possible”, wait, or carry on as normal. When a whole math community faces that choice at once, rushed and weaker papers become more likely.
Students and the problems chosen for them
The declaration notes that supervisors “often suggest problems with the core intention of developing skills”. A PhD problem is chosen to be hard enough to teach and open enough to publish. If an AI system resolves it midway through a doctorate, the student loses a thesis result even though the training still happened. Hairer said this potential disruption was one reason he added his name to the declaration.
Credit, citation and trust
Priority matters in mathematics because careers are built on it. Unclear version histories, thin citations and surprise releases all erode the shared record the math community relies on to assign credit. That is why the Navier-Stokes dispute, with its accusations of scooping, still shadows every OpenAI announcement.
What a Responsible Release Could Look Like
AGMAI says it has not settled its advice and has opened a feedback form. The existing norms of mathematics, and the rules OpenAI’s own headline problem is judged by, suggest what the math community will ask for.
The Clay rules are a useful benchmark
The Clay Mathematics Institute will not consider a Millennium Prize solution until it has been published in a qualifying outlet, “at least two years” have passed since publication, and it has “received general acceptance in the global mathematics community.” OpenAI’s Navier-Stokes result became public on 8 September. The chart compares how long it had been public by 29 September with the minimum wait Clay requires.
Five practices the math community already expects
None of these is new. They are how human mathematicians release work, and each answers a complaint in The Verge’s report.
| Practice | Complaint it answers | What it would mean for OpenAI |
|---|---|---|
| Full write-ups before announcements | Results “announced in a rush” | A readable preprint on a public server first |
| Proper citation of prior work | Missing credit for earlier ideas | Specialists check the literature review |
| Versioned documents | “Sneaky” edits after criticism | Numbered versions with change notes |
| Advance notice to active researchers | Fear of being scooped | Contact people working on a problem before release |
| Plain claims | “Substantial progress” language | Say solved or not solved, nothing in between |
Staggered, not simultaneous
A hundred results released together would overwhelm refereeing capacity in any field. Releasing them in batches by area, each with a named human reviewer and a written account of what is new, would let the math community absorb them. It would also give OpenAI’s results the credibility that only independent checking brings.
Why This Matters Beyond Mathematics
Mathematics is the first field where AI systems produce results fast enough to swamp the people who judge them, but it will not be the last. The same questions apply to AI-generated code, research summaries and engineering designs inside businesses.
Provenance, versions and review
If your organisation publishes AI-assisted work, keep the record the math community is asking OpenAI for: what the system produced, what humans checked, and every revision. Any AI strategy that involves publishing AI output should start from that principle, because trust is lost faster than it is built. Change control for documents is ordinary change management, applied to a new source of output.
Lessons for research and development teams
The same pressures appear wherever AI output meets expert review. A legal team handed a hundred AI-drafted contracts, or an engineering team handed a hundred AI-proposed designs, faces the math community’s problem in miniature: the output arrives faster than anyone can check it, and the checking is where the value lies. Set review capacity before you set output targets, record who approved what, and never announce a result your own specialists have not read.
Earlier coverage
This is the fourth time we have covered OpenAI’s clash with mathematicians. Our first report asked whether mathematicians’ unpublished work helped produce the OpenAI math results. The second covered the formation of the math advisory group and its overlap with the declaration’s signatories. The third looked at WIRED’s report on how AI math is changing a field that once treated great problems as an art form.
What Comes Next for AGMAI and the Math Community
The immediate test is the first batch of OpenAI’s “more than 100” results. If they arrive with full write-ups, proper citations and advance notice to the people working in each area, AGMAI will have earned the trust Hairer says it needs. If they arrive as a press release, the math community’s verdict on the group will be swift.
Hairer seems ready for either outcome. OpenAI may spin his involvement, he told The Verge, and he will let it. “They’re not going to sort of damage me within the math community,” he said. “In some sense, I don’t really care about what they say. But what I do worry about is the math community as a whole.”
Math Community FAQs
What is AGMAI?
The Advisory Group on Mathematics and Artificial Intelligence is a nine-member body hosted by the Institute for Advanced Study. It was announced on 21 September 2026 to advise AI companies on how they present and release mathematical results.
Is AGMAI part of OpenAI?
No. Both OpenAI and AGMAI say the group is independent and unpaid. It formed after OpenAI approached some members about an advisory board, which is why many in the math community first assumed it was OpenAI-appointed.
How many results is OpenAI holding?
OpenAI says its internal model has resolved “more than 100 long-standing open problems across most areas of mathematics” in addition to Navier-Stokes. It has not said which problems.
Why does the math community object if the proofs are correct?
The objections are about process: rushed announcements, thin write-ups, missing citations, quiet edits and the risk of wiping out years of human work, including students’ thesis problems, without warning.
Will AGMAI advise other AI companies?
Yes, according to both the group and Hairer. AGMAI says it will offer recommendations to any AI company whose models are likely to have a significant impact on mathematics, and Hairer told The Verge that conversations with other labs had already begun. He declined to say which ones.
How many mathematicians back the declaration?
On 29 September 2026 the “Severe Misalignment of AI in Mathematics” declaration listed 28 Fields Medallists and 8,063 endorsers.
References
OpenAI keeps bulldozing mathematicians (The Verge)
Advisory Group on Mathematics and Artificial Intelligence (AGMAI)
Advisory Group on Mathematics and Artificial Intelligence (OpenAI)
Advisory Group on Mathematics and Artificial Intelligence (Terence Tao’s blog)
Why I agreed to join AGMAI, by Martin Hairer (Proofs and Prompts)
A Severe Misalignment of AI in Mathematics (Math and AI)
Endorsers of the declaration (Math and AI)
Rules for the Millennium Prize Problems (Clay Mathematics Institute)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
