Mathematicians who say powerful AI systems threaten the norms of their profession are continuing to use those same systems in their daily work. WIRED’s interviews reveal a field caught between concerns about credit and verification and the practical advantage of tools that can accelerate research and writing.
New York University professor Tristan Buckmaster remains at the center of the conflict. He has accused OpenAI of using his work to overtake him on the Navier-Stokes existence and smoothness problem. Yet he told WIRED that he still uses OpenAI’s Codex to edit papers and to reconstruct how the company’s agents may have moved from his earlier work toward their claimed proof.
New Answers Do Not Settle the Attribution Dispute
OpenAI investigated Buckmaster’s allegation and amended its announcement. The company said his Codex prompts during the two months before its September 8 paper could not have influenced the system through training or any other route. That is OpenAI’s conclusion; Buckmaster continues to question how the company learned that a solution was close and then deployed tens of thousands of agents.
A second mathematician described a similar uncertainty. Andreas Thom said OpenAI’s Astra model used techniques from geometric group theory to prove a problem he had worked on. After he challenged the company’s claim that no progress had occurred for a decade, OpenAI updated its release to acknowledge earlier work, including his 2019 paper.
Thom said OpenAI researchers told him that his ChatGPT interactions were not used for training. He nevertheless doubts he will ever know whether his work contributed indirectly. His concern is not only ownership of a particular result. He argues that AI makes it much harder to trace the chain of ideas and people behind a proof.
The source does not establish that either man’s private interactions trained an OpenAI model. It documents their unresolved doubts alongside the company’s denials and corrections.
Useful Tools Create Pressure to Keep Participating
Despite those doubts, Buckmaster and Thom both continue using AI. Thom has adjusted privacy settings and uses ChatGPT to speed up paper writing. Buckmaster says that opting out is difficult because the systems are useful and the leading providers hold unusually strong positions.
Cornell mathematician Alex Townsend sees the same tension among colleagues. Since August, he has watched researchers investigate subscriptions to more capable models. He described feeling excited by work that AI makes possible while also questioning the role left for human mathematicians.
That pressure may be especially strong for early-career researchers. Thom warned that refusing tools that deliver major efficiency gains could leave them isolated. At the same time, Buckmaster fears that years of unpublished work could enter a provider’s data through an apparently routine editing request and later surface elsewhere.
WIRED notes that university and enterprise versions of OpenAI products default to not using customer data for model training. That policy addresses one route of concern, but it does not solve the broader problem of documenting how agent-generated proofs build on published research or model-learned patterns.
Researchers Want Ground Rules Without Giving Up AI
The dispute has produced organized resistance. Twenty-five Fields Medal recipients signed an open letter saying AI companies and mathematicians are severely misaligned. More than 4,000 people have signed the Leiden Declaration, which proposes recommendations for researchers, funders and policymakers. More than 2,000 people connected to Caltech also sought to suspend an AI mathematics hackathon; OpenAI later withdrew as a sponsor.
Buckmaster is now calling for a temporary detente while laboratories and mathematicians establish rules for announcing results and citing prior work. He also plans to revise papers he posted quickly during the competition with OpenAI, acknowledging that one was not complete.
The emerging position is not a rejection of machine assistance. It is a demand that speed does not erase attribution, comprehensible reasoning or the time required for verification. Whether laboratories accept common procedures remains open. For now, researchers are trying to negotiate standards while continuing to use the technology whose influence they want to constrain.
