A Fields Medalist has told Le Monde that OpenAI’s week of AI-generated mathematics left his speciality wiped out overnight. “They have annihilated my field of research,” Hugo Duminil-Copin, who won the 2022 Fields Medal for statistical physics, said in an interview republished on Tuesday 13 October, adding that he “cried so much I woke up my daughter” after the release of the 722-paper collection, according to the paper’s account.
The collection itself is now a fortnight old — OpenAI published 722 AI-written papers in a public GitHub repository on 7 October, generated by an internal frontier model the company has not released, spanning number theory and algebraic geometry in 372 groups. But the reaction wave is live and escalating right now: new interviews, boycott calls, and a fresh technical challenge to the company’s flagship Navier-Stokes proof all reached print inside 48 hours.
The open letters have moved to grief
The 722-paper release followed OpenAI’s announcement that a model had claimed the Navier-Stokes existence and smoothness problem, deploying up to 10,000 agents in parallel over roughly 88 hours, which more than two dozen Fields Medalists answered with an open letter warning the field was being undermined. The release wave has now produced a second, more personal complaint. OfficeChai’s account of the Duminil-Copin interview says the mathematician saw his working area swept to done within days, with the emotional toll — a father crying loud enough to wake his child — as striking as the claim. Le Monde’s framing is harsher still: mathematicians, it writes, have been left feeling their careers were consumed as training fuel for a corporate model, which is exactly the sentiment the earlier open letter had already flagged in more clinical terms.
Duminil-Copin is not alone. According to the Korea Daily’s summary, US-based scholars have separately criticised the speed of the release: proofs that rushed past community review carry a social cost, not just a technical one. And a widely shared open letter now circulating calls for a boycott of AI-generated mathematics results until labs accept standards from bodies like the Institute for Advanced Study’s Advisory Group on Mathematics and Artificial Intelligence.
What is actually verified in the release
Strip away the emotion and the release has hard technical edges. According to Quartz’s report, the model attempted roughly 4,000 problems, and the average result used the equivalent of about three hours of ChatGPT Pro compute; WSJ adds that the batch touches three still-unsolved Millennium Prize Problems including the Riemann hypothesis. Many of the proofs ship with Lean formalizations, machine-checkable transcripts that are a genuine step up toward auditability, alongside ten reasoning summaries and compute estimates. OpenAI’s own posting commits to updating the repository with more formalizations over time and to funding community workshops around AI-generated results.
The verification layer is also where trust frays. New Scientist reports that mathematicians claim there is a mismatch between the two versions of OpenAI’s Navier-Stokes proof: the human-readable paper and the machine-checked Lean transcript diverge, meaning what was written for humans is not what was verified by code. Cambridge professor Anders Hansen points out the consequence: an enormous added burden lands on human readers because LLM proofs must be manually checked, not just machine-executed. The mismatch does not mean the problem is unsolved in OpenAI’s favour or against it — it means the certificate of correctness humans were promised and the artifact that exists are not the same object.
The incentive system nobody designed
Two things are happening at once, and conflating them is the trap. The technical layer — 722 papers, Lean transcripts, a model that produces them in hours — is real, and by some assessments of one release-adjacent result, “if a human did this, it would be an instant Fields Medal, no questions asked,” as Rutgers mathematician Alex Kontorovich wrote, quoted by Quartz. The cultural layer is the part being actively destroyed. Mathematics runs on priority, reputation, and the decades-long gamble that an obscure open problem is worth a career. When a frontier lab releases a fortnight of machine output that collapses whole subfields at once, the gamble structure breaks: nobody can win a race against a compute budget that spends millions of dollars per problem, and mid-career researchers whose entire capital is a speciality are the ones left holding the risk. That is the specific grief in Duminil-Copin’s “annihilated” — not that the papers are wrong, but that the reward economy his entire generation structured their lives around no longer applies to them, while the verification burden that would make the new work trustworthy has actually increased.
For New Zealand’s small research community the pattern is a warning about sequencing. The earlier Claude Riemann zeta bound and the Erdős disproof read as breakthroughs because they landed against problems mathematicians still owned. A 722-paper dump reads as dispossession because it arrives all at once, sourced from a model nobody outside OpenAI can inspect, with the Institute for Advanced Study advisory group’s guideline against testing frontier problems on proprietary models openly declined — the one deviation OpenAI acknowledged from the community guidelines it otherwise followed. University heads who are this month negotiating AI tooling policies would do well to include research-priority protections, not just academic-integrity boilerplate.
❓ FAQ
What did OpenAI actually publish? 722 AI-generated papers organised into 372 groups, released on GitHub on 7 October and produced by an internal frontier model, spanning number theory and algebraic geometry — including results bearing on the Riemann hypothesis and other Millennium Prize Problems.
Who is Hugo Duminil-Copin, and what did he say? He won the 2022 Fields Medal for his work in statistical physics. Speaking to French newspaper Le Monde, he called his field “annihilated” by the AI-paper wave and described the emotional weight of watching an entire research area get consumed in one release cycle.
What was the New Scientist mismatch claim? A team of mathematicians reviewed OpenAI’s Navier-Stokes proof pair and reported the human-written proof and the formally verified Lean transcript do not fully match. That does not settle the theorem — it means the machine-verified object and the human-readable object are not the same thing, and it revives the earlier review complaints about gaps a reader could not independently fill.
Is OpenAI doing anything for the math community? It says it will fund workshops, conferences, and programs around AI-generated mathematics and is working to responsibly release the model that produced the results. It also published Lean formalizations, reasoning summaries, and compute estimates alongside its papers.
Sources
- Le Monde — OpenAI’s latest blow to mathematicians
- Quartz — OpenAI’s AI solved one famous math problem. Then it cracked hundreds more
- OpenAI — Sharing AI progress in mathematics
- New Scientist — OpenAI mistranslated mathematics
- OfficeChai — Duminil-Copin interview account
- SiliconANGLE — OpenAI publishes 722 AI-generated math discoveries
