Tristan Buckmaster, a mathematician at the University of Maryland, has leveled serious allegations against OpenAI, claiming researchers pressured him to drop his co-author, Levent Alpöge, from a breakthrough paper on the Navier-Stokes equations. The tension arose after Buckmaster and Alpöge used AI models, including OpenAI’s Codex, to make progress on one of mathematics’ most famous unsolved problems. According to Buckmaster, OpenAI researcher Sébastien Bubeck insisted Alpöge be removed because he works at rival Anthropic. When Buckmaster refused, Bubeck allegedly asked, “Why would you ruin your career?” OpenAI denies seeing any of their private work before its public release.
The dispute centers on who solved what, and when. OpenAI published its own solution to the Navier-Stokes problem, stating it was generated by 10,000 coordinated AI agents in 88 hours using an internal model described as “significantly more capable than GPT-6 Astra.” Mark Chen, OpenAI’s research lead, noted the compute cost ran into the millions. The company claims it only began focusing on the problem in early September after hearing rumors of progress, ultimately producing a Lean-formalized proof. OpenAI asserts its approach differs fundamentally from Buckmaster and Alpöge’s, which targeted the forced Euler problem rather than the full Navier-Stokes equations.
Buckmaster remains skeptical of OpenAI’s timeline and ethics. He points out that OpenAI admits to training models on customer data from the period when he and Alpöge were actively working on their proofs. “Is it ethical to use customer’s data to try to scoop their customer?” Buckmaster asked on Mastodon. While OpenAI states no specific user data was accessed to solve the problem, it acknowledges that de-identified data may have improved its models. This raises uncomfortable questions about intellectual property in the age of large language models, especially when users haven’t explicitly opted out of data usage policies.
The broader implication is a shifting landscape for mathematical discovery. Noam Brown, an OpenAI developer, noted that tasks costing $500,000 in compute today could cost $20 within a year. If AI agents can indeed solve Millennium Prize-level problems in days, the barrier to entry for high-level mathematics collapses. But if proprietary models are trained on unpublished academic work without clear consent, trust between researchers and tech giants may erode. The math community now watches closely to see if formal verification can settle the priority dispute, or if this becomes a legal battle over data rights.
