OpenAI’s Math Breakthrough Ignites Debate Over Attribution in the Age of AI

OpenAI’s recent announcement that it used an artificial intelligence model to solve a significant challenge in fluid dynamics—the Navier-Stokes problem—has triggered a broader discussion regarding intellectual attribution as AI increasingly permeates scientific research. The claim, unveiled by the San Francisco-based company on September 8, is expected to fundamentally alter mathematical practices, yet it has also raised alarms about whether the system may have absorbed insights from human researchers who were simultaneously attempting to crack the same problem.

The incident highlights a growing concern among scholars: as AI models become ubiquitous across disciplines, the traditional notion of scholarly credit faces unprecedented risks due to the difficulty of tracing the origins of training data. Luke McDonagh, an intellectual property law expert at the London School of Economics and Political Science, noted that many academics may not fully understand the implications of uploading their knowledge into personal AI accounts.

The issue gained urgency when Tristan Buckmaster, a mathematician at New York University, and his collaborator Levent Alpöge from Harvard University, raised objections just before OpenAI confirmed its breakthrough. The two researchers had been utilizing OpenAI and Anthropic tools to investigate aspects of the Navier-Stokes equations for nearly a year. They alleged that OpenAI had accelerated its efforts after learning of their work and suggested the AI model might have learned from their interactions with ChatGPT.

In response, an OpenAI spokesperson told Nature that an internal investigation confirmed no user inputs entered after July 3 could have influenced the system. The company stated it began working on the problem on September 1 and had not accessed the researchers’ unpublished work through any channels.

Buckmaster clarified that he maintained three separate ChatGPT accounts, only two of which had the training data opt-out feature enabled. While the Clay Mathematics Institute in Oxford, UK, will evaluate the validity of the solution only after peer-reviewed publication and community vetting, many experts in the field argue that Buckmaster, Alpöge, and other Spanish researchers such as Diego Córdoba and Luis Martínez Zoroa deserve substantial credit for their contributions.

The controversy extends beyond this single case. Andreas Thom, a mathematician at Dresden University of Technology, pointed to a separate incident involving non-sofic groups. Last August, OpenAI published a preprint describing the first example of such a group, employing a strategy similar to one Thom had discussed with ChatGPT during brainstorming sessions. Although OpenAI correctly cited Thom’s published papers, Thom emphasized that informal academic exchanges are typically acknowledged in human research, raising questions about whether his unpublished insights contributed to the AI’s output.

Thom expressed frustration at the lack of clear attribution, stating that if a human colleague had used his office discussions without acknowledgment, he would have been upset. OpenAI maintained that users control whether their conversations aid model improvement and asserted that once opted out, data is not used for training purposes. As these cases accumulate, the mathematical community faces the complex task of determining how to assign credit when AI tools are deeply integrated into the creative process.