AI-generated conceptual cover illustration. Research checked October 30, 2026.
AI companies are claiming solutions to problems that mathematicians have spent decades trying to crack. You might expect a celebration. Instead, one group of researchers has called on colleagues to stop working with OpenAI. What happened?
The immediate flashpoint was OpenAI’s October 6 release of hundreds of AI-written mathematical manuscripts. But the anger reaches beyond machines getting answers first. It concerns whether those answers are correct, whose ideas they build on, and whether producing a proof is the same thing as advancing human understanding.
There is no single mathematicians’ position on AI. Some researchers want to avoid it; others already use it. The disagreement is about how a powerful new research tool is entering a discipline built around careful explanation and trust.
What did AI actually do?
On October 6, OpenAI released a collection initially containing 722 manuscripts grouped into 372 related families. These were research claims across mathematical fields, rather than school exercises or competition puzzles. A family can contain several related papers, so 722 manuscripts does not mean 722 separate landmark problems were solved.
The collection is changing. When checked on October 30, OpenAI’s repository listed 721 manuscripts in 373 families following withdrawals and additions. OpenAI describes the material as research data at different stages of verification, rather than finished academic papers. Some results have computer-checkable proofs; others do not. The repository reports that roughly 42% of its top-line results have Lean formalizations.
That distinction matters. A company publishing a claimed solution, a computer checking a formal proof, and a research community understanding the result are different milestones. Treating them as one event makes the achievement sound more settled than the evidence warrants.
Why can’t everyone just check the answers?
A research proof is an argument connecting assumptions to a conclusion. Checking one can require years of specialist knowledge. Checking hundreds of unfamiliar arguments creates a substantial task for the people who can judge them.
This release already has a concrete example. OpenAI’s October 7 change log records three withdrawn manuscripts after a sign error invalidated an argument and a construction used by two dependent papers. It also records revisions to 14 other manuscripts, including proof repairs and corrected statements. That does not show the entire collection is wrong. It shows why a large initial release cannot be treated as a collection of settled discoveries.
Lean, a proof assistant, can help. It checks a precisely encoded mathematical argument. But the statement encoded in software still has to match the problem people think has been solved, with appropriate definitions and assumptions. Lean’s own validation guidance explains the checks needed to assess a proof. A formalization is valuable evidence; it does not remove the need to inspect what was formalized.
The sharper objection: who gets to set the rules?
The Association for Human Mathematics responded on October 7 by describing the mass release as a “demonstration of power” and urging mathematicians to discontinue work with OpenAI. That is the position of this association, rather than a verdict from the entire profession.
Its criticism reflects a wider concern: if companies select famous problems to test private models, they can influence what receives attention while leaving other researchers to check and explain the output. The researchers who do that work may have had different questions they wanted to pursue.
The independent Advisory Group on Mathematics and Artificial Intelligence has proposed responsible-release guidelines. It asks labs to stop testing advanced problems on proprietary models inaccessible to the wider community. Where results are produced, it recommends proper citations, readable arguments, responsible publication and support for developing human understanding. Its advice is more demanding than simply posting files online.
In its October 6 statement, the group also says that advising OpenAI should not be read as endorsing its process. Its October 25 follow-up raises concerns about companies choosing particular mathematicians to lead the interpretation of AI output, including unequal access and influence over the field’s priorities.
There is also a real dispute over credit
The October backlash followed September’s dispute around the Navier–Stokes equations, which describe fluid motion. OpenAI announced that an internal AI system had produced a proof of finite-time breakdown under smooth external forcing, claiming it resolves a version of the Millennium Prize problem.
Mathematician Tristan Buckmaster and collaborator Levent Alpöge had been pursuing related fluid-equation results. In Buckmaster’s public account, he raised concerns about how OpenAI learned of their progress and about proposed publication arrangements. He also credits earlier work by Diego Córdoba and Luis Martínez-Zoroa as the starting point for their approach. His statement makes clear that Buckmaster and Alpöge used AI tools themselves.
OpenAI disputes the implication that it used their private work. It says no specific user data was accessed for its solution, that Buckmaster’s recent prompts could not have influenced the system, and that the proofs differ significantly. Those allegations and denials should remain distinct. A public dispute is not proof that unpublished research was stolen.
What are mathematicians worried about losing?
The Leiden Declaration on Artificial Intelligence and Mathematics, dated June 2 and endorsed by the International Mathematical Union, identifies concerns about unreliable arguments, missing attribution, pressure on review and changes to research incentives. It also describes enthusiasm for AI’s potential alongside anxiety about its effects.
One concern deserves more attention than the idea that researchers dislike competition: training the next generation. Learning to investigate a problem develops judgment about which approaches fail, which assumptions matter and which questions are worth asking. The declaration warns that threats to mathematical practice can fall especially heavily on students and early-career researchers.
Our view is that a verified machine result can be a major advance while leaving a substantial educational and scientific task unfinished. Knowing that a statement is true is useful. Understanding why it is true helps people connect it to other ideas, explain it and decide what to investigate next.
So is AI good or bad for mathematics?
The evidence supports a more useful question: under what conditions does AI help people do better mathematics?
There are efforts to make that collaboration work. OpenAI announced on October 15 that it would contribute US$5 million to a Simons Foundation initiative supporting examination and understanding of AI-produced results. In Australia, MATRIX has announced a March 2027 workshop for human–AI collaboration on open mathematical problems, with interpretable insights and responsible use among its stated questions.
For the next spectacular AI-math headline, ask four things: What exact result is claimed? What has been independently checked? Are earlier contributions credited? Can people understand and build on the argument?
Those questions leave room for extraordinary AI achievements. They also explain the anger: researchers are being asked to absorb a rapid change in how discoveries are produced, checked and rewarded. The lasting value will depend on what the mathematical community can make of the results.
Editorial note: This is a research-based explainer, prepared with AI assistance and editorial source checking. Techotick has not independently verified the mathematical proofs. The cover is an AI-generated conceptual illustration, not a photograph of researchers or an actual event.
Sources
- OpenAI mathematics repository — OpenAI
- Mathematics release history and withdrawals — OpenAI
- Validating a Lean Proof — Lean
- Statement on OpenAI’s October 6 release — Association for Human Mathematics
- Responsible Release of AI-Generated Mathematics — AGMAI
- Statements on mathematical results and their release — AGMAI
- Public statement on fluid-equation research and publication — Tristan Buckmaster
- On the Navier–Stokes Millennium Prize Problem — OpenAI
- Leiden Declaration on Artificial Intelligence and Mathematics — Leiden Declaration
- Sharing AI progress in mathematics — OpenAI
- Human–AI Collaboration on Open SMMC Problems — MATRIX