Mathematicians want OpenAI to more clearly prove that it did not use unpublished research in its solution to the Navier-Stokes equations. At the center of the debate is a separate study conducted by New York University professor Tristan Buckmaster and Anthropic researcher Levent Alpöge. The knowledge that the two researchers were using Codex raised new questions about model training and user data.
According to The Verge, mathematicians demand not only the similarity of the final proof, but also that the OpenAI team explain what information it obtained on what date. Buckmaster says information about its work was provided to the company on September 3. OpenAI, on the other hand, states that it launched the Navier-Stokes initiative on September 1, following a rumor, and that it did not use the researchers’ prompts or proofs to direct agents.
The technical whitepaper published by OpenAI reports that the company worked with 10 thousand agents for approximately 88 hours, and Lean verification took another 17 hours. The system used approximately 300 billion output tokens while generating 4.9 million messages. The company says it did not access researchers’ private data; However, it does not completely exclude the possibility that de-identified usage data contributed to model development.
This distinction forms the crux of the debate. Without directly opening a particular conversation, a model can be influenced by patterns derived from previous training processes. OpenAI explains that when using individual ChatGPT and Codex, it can use content to develop models depending on the settings. Training is not provided by default in business and API products. Researchers want to know in advance and clearly which category unpublished results fall into.
Controversy magnifies trust issue in AI-supported research
The Navier-Stokes problem is one of the seven Millennium Problems that ask whether three-dimensional fluid motion always remains smooth. OpenAI states that its internal system provides proof that a singularity can occur in finite time. The Clay Mathematics Institute’s award process carries additional conditions, including independent review and acceptance of the published result. Therefore, the company’s announcement does not yet mean that the problem has been officially resolved.
Buckmaster and Alpöge’s work is based on a separate result on the forced Euler problem. OpenAI also argues that the first rumor was about this study and that the two teams proved different results. But since the two lines of research touch similar fluid dynamics methods, the issue of timing and attribution remains important. The company then made the priority issue visible by adding relevant studies to its narrative.
While AI tools accelerate the pace of experimentation in mathematics, they require new rules for research privacy. When a mathematician gives unpublished ideas to an assistant, he or she should be able to easily see the extent of storage, analysis, and training of that data. Laboratories must also disclose data provenance in an auditable manner when they are ahead of independent researchers with massive computing power. The lasting impact of this event can be seen in the standard of transparency that will be established for scientific trust rather than in the accuracy of a single proof.
Join Channel