OpenAI’s Math Dispute Is Becoming a Crisis of Trust

Introduction

AI systems are beginning to tackle problems that have challenged mathematicians for years. That prospect could be a major scientific advance, but it is also creating friction between frontier labs and the mathematical community. Twenty-five Fields Medalists have signed an open letter warning that laboratories are rushing to announce results without leaving enough time for verification, exposition, or proper citation of earlier work.

The concerns became more concrete this week. Tristan Buckmaster, a professor at New York University, accused OpenAI of pressuring him not to credit a collaborator who works for Anthropic in the solution of an important mathematical problem. He also questioned whether OpenAI had used work produced with Codex to generate its own major proof during a marathon inference effort. The proof in question has not yet been verified by mathematicians.

The central disputes

• A proof is more than an announcement. Mathematical results must be checked, written up in detail, and made understandable to other researchers. A rushed release can make it difficult to assess reliability or incorporate a new technique into the discipline.
• Attribution is becoming harder. When AI tools help search for a solution, researchers need clearer standards for distinguishing the origin of an idea, the execution of a proof, and the contribution of the model. Without them, disputes over plagiarism and misappropriation are likely to grow.
• Researchers fear a feedback loop. Some mathematicians worry that work carried out with Codex could be absorbed into future model development, potentially giving a lab an advantage over the people who first explored the idea.
• Institutional relationships are under strain. On Thursday, OpenAI withdrew sponsorship of a mathematics event at Caltech after criticism from researchers there, showing that the dispute now extends beyond individual authorship arguments.

Why it matters

Mathematics is not defined only by a final proof or the person who claims priority. Its value also lies in the concepts that emerge from a result, the questions it inspires, the way it is taught, and the chain of communication connecting generations of researchers. The signatories argue that AI-generated ideas cannot become part of the mathematical canon unless mathematicians are willing to verify, explain, and develop them.

There is also a structural risk. Frontier labs can spend enormous sums on computation and inference to search for a proof and potentially announce it before the original researchers. If scholars become unsure whether using an AI tool exposes their research direction or gives a competitor an advantage, open collaboration may give way to secrecy.

The letter follows the Leiden Declaration, released by a mathematicians’ working group in June, which addressed similar questions and offered recommendations for researchers, institutions, and policymakers. The underlying issue extends well beyond mathematics. In software, science, and other creative fields, AI is forcing society to ask not only whether a system can perform a task, but how to preserve responsibility, meaning, and the human transmission of knowledge.

TechCrunch AI