OpenAI vs Mathematicians: Who Really Solved Navier-Stokes?

A breakthrough claim enters its second week

As of September 19, 2026, the safest answer to “who solved Navier-Stokes?” is still conditional: OpenAI says an internal artificial-intelligence system produced a solution to the Navier-Stokes Millennium Prize Problem, while mathematicians are asking whether the proof, the priority and the credit have been settled in the same sense [1]. The dispute is not merely about whether a computer-generated manuscript is correct. It is about whether researchers who used commercial AI tools may have exposed unpublished ideas to a company powerful enough to race them to the finish line [1].

The stakes are unusually high because Navier-Stokes is one of the Clay Mathematics Institute’s Millennium Prize Problems, a list selected in 2000 and associated with a $1 million prize for each accepted solution [3]. The equations describe the motion of fluids, including water, air and blood, and are central both to pure mathematics and to fields such as aerodynamics, climate modeling and medicine [2]. OpenAI’s September 8 claim therefore landed at the intersection of scientific prestige, industrial AI competition and academic trust [1].

Nature reported on September 17 that OpenAI’s announcement has sparked controversy over whether AI tools might have learned from human mathematicians who were pursuing related work, and whether current systems can preserve the scholarly norm of crediting conversations and informal insights [1]. That is the central issue: not only “did the proof work?” but “whose ideas made it possible?”

The contested timeline

The controversy centers on Tristan Buckmaster of New York University and Levent Alpöge, described in fresh reporting as a mathematician at Harvard and also associated with Anthropic-related work in the wider dispute [1][3]. According to Nature, Buckmaster and Alpöge had been working on an aspect of the Navier-Stokes problem using AI tools from OpenAI and Anthropic before OpenAI publicly announced its own result [1]. Buckmaster told Nature that he had used OpenAI tools for about a year and had three ChatGPT accounts, only two of which were opted out of the setting that permits conversations to improve models [1].

OpenAI has rejected the implication that those interactions affected its claimed proof. A company spokesperson told Nature that “no user inputs past July 3rd could have influenced this system in any way,” and that OpenAI started working on the problem on September 1 [1]. OpenAI also said it had not seen Buckmaster and Alpöge’s work “through any means” until they released it publicly [1]. Those statements directly address one version of the concern, but they do not fully dissolve the wider anxiety: researchers now realize that commercial AI tools can sit inside the earliest stages of mathematical discovery, before preprints, seminars or acknowledgments exist [1].

The unresolved question is not only legal. Luke McDonagh of the London School of Economics told Nature that academics might not have fully understood the consequences of uploading data and knowledge to personal AI accounts [1]. In traditional mathematics, a colleague who learns a key technique through discussion is expected to acknowledge that intellectual debt. With AI systems, the path from prompt to model behavior can be opaque, and the “colleague” may be a proprietary product owned by a potential competitor [1].

Proof, prize and community verification

Even if OpenAI’s proof is ultimately judged correct, formal acceptance is not instantaneous. Nature reported that OpenAI says it verified the proof using Lean, a programming language used for machine-checkable mathematics [1]. But the Clay Mathematics Institute told Nature that it will consider whether a solution is valid only after publication in a peer-reviewed venue and further community vetting [1].

That distinction matters. A Lean formalization can strengthen confidence in a proof’s internal logic, but it does not automatically settle questions of scope, interpretation, priority, exposition or prize eligibility. The mathematical community still has to determine what exactly has been proved, whether the assumptions match the official problem and how the result fits into earlier work [1][4].

A September 17 arXiv paper by Peter Constantin, Mihaela Ignatova and Vlad Vicol illustrates how fast expert scrutiny has begun [4]. The authors analyze regularity for asymptotically axisymmetric solutions to the 3D Navier-Stokes equations with analytic forcing, explicitly framing their work in response to OpenAI’s recently announced construction [4]. Their result does not read as a simple refutation of OpenAI. Rather, it narrows the mathematical landscape by showing that, under certain structural properties attributed to the OpenAI construction, analytic or locally vanishing forcing would impose regularity instead of blow-up [4]. In plainer terms: the follow-up literature is already probing which features of the claimed construction are essential, fragile or physically interpretable.

What did the alleged solution mean physically?

The September 18 Nature physics analysis adds another layer: even if the mathematical claim is confirmed, its direct physical consequences may be limited [2]. The report describes OpenAI’s posted solution as involving a vortex that stretches until it becomes extremely long and thin [2]. Applied mathematician George Karniadakis estimated for Nature that, in air, the relevant singularity would appear when the vortex becomes roughly 70 nanometres wide, around the scale where the continuum approximation behind Navier-Stokes becomes suspect [2].

That point is important for readers outside mathematics. Navier-Stokes equations treat fluids as continuous media, not as discrete molecules [2]. At tiny scales, especially when only a small number of molecules span the width of the flow structure, that assumption begins to fail [2]. So a singularity in the mathematical model is not the same thing as a practical recipe for aircraft design, weather prediction or blood-flow simulation.

Nature also emphasized that the Millennium Problem concerns incompressible fluids, a very good approximation for liquids such as water, whereas other fluid models already have known limitations in rarefied gases or compressible settings [2]. This helps explain why the breakthrough can be mathematically historic while still having no immediate engineering payoff. It may change how mathematicians understand the limits of the equations, without changing how engineers run everyday simulations.

A split reaction from mathematicians

The social reaction has been as dramatic as the mathematics. Live Science reported from the Heidelberg Laureate Forum that a September 15 panel of prominent mathematicians confronted OpenAI’s announcement as part of a broader crisis over AI in the discipline [3]. Columbia mathematician Michael Harris described the episode in severe terms and called for a regulatory framework to protect mathematics and mathematicians from similar behavior [3].

Not everyone on the panel read the moment the same way. Jacob Tsimerman, a 2026 Fields Medalist who had joined OpenAI as an AI safety researcher, argued that the top-line achievement was being buried and that the result should inform the public about AI capabilities [3]. Peter Scholze, a 2018 Fields Medalist, reportedly dismissed the announcement as “some kind of PR” and stressed that the purpose of Millennium Problems is not merely to obtain an answer, but to generate new perspectives and human understanding across related mathematics [3].

That split captures the larger dilemma. If the proof is correct, OpenAI may have demonstrated a startling new mode of discovery. But if the proof is correct and the credit system fails, the episode could still damage mathematics by making researchers afraid to use the tools that accelerate their work.

So who really solved Navier-Stokes?

There are three defensible answers, depending on what “solved” means.

If “solved” means “first produced the proof now under review,” then OpenAI claims that role for its internal system, backed by a Lean-verified proof and its account of an independent September effort [1]. If “solved” means “created the mathematical path that made the result possible,” fresh reporting says many researchers believe significant credit should also go to Buckmaster and Alpöge, as well as Diego Córdoba and Luis Martínez Zoroa for earlier forcing techniques [1]. If “solved” means “has been officially accepted as a Millennium Prize solution,” the answer remains no: the Clay process, peer review and community vetting have not been completed [1].

For now, the OpenAI versus mathematicians dispute is not a clean story of machine replacing human genius. It is a story about a new research stack in which humans, AI agents, formal proof systems, private prompts and corporate incentives are entangled. The proof may eventually stand. The credit system, however, is already under strain.

Sources from the last 72 hours

1. [1]Who gets credit in the AI era? OpenAI maths bombshell sparks debateSep 17, 2026, 12:00 AM UTC
2. [2]AI cracked the Navier–Stokes challenge. What does that mean for physics?Sep 18, 2026, 12:00 AM UTC
3. [3]'Predatory behavior': Elite mathematicians clash over OpenAI's 'solution' to million-dollar math problemSep 17, 2026, 12:00 AM UTC
4. [4]Regularity of asymptotically axisymmetric solutions to the 3D Navier-Stokes equations with analytic forcingSep 17, 2026, 5:57 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.