Mathematicians question whether OpenAI's Navier-Stokes proof solved the intended problem

OpenAI's headline-grabbing proof of the Navier-Stokes problem — one of the Clay Mathematics Institute's Millennium Prize problems, each worth $1 million — is facing a second wave of scrutiny. According to Scientific American, researchers increasingly argue that the proof, generated by an internal large language model, solved a variant of the problem that most mathematicians consider peripheral, and new work shows the approach can never be extended to the question they actually care about.

A proof built on an optional term

The Navier-Stokes equations describe how fluids flow, and the open question is whether the equations can "blow up" — that is, permit flows that become infinitely fast at certain points, something that cannot occur in the physical world. The problem statement, written in 2000 by mathematician Charles Fefferman, includes an optional component: an external force, such as gravity, acting on the fluid.

Most experts study the problem without such a force, looking for a blowup driven purely by a fluid's intrinsic dynamics rather than by a carefully engineered outside push. As Scientific American reports, OpenAI's proof took the opposite route, constructing a very specific external force to break the equations. Fefferman's formulation explicitly permits this under an option labelled "C", so the result satisfies the Clay problem as officially written. Mathematicians interviewed by Scientific American nevertheless characterise the move as exploiting a loophole in how the question was framed.

The road to the announcement

The underlying technique did not originate with OpenAI. In recent years, mathematicians Diego Córdoba and Luis Martínez-Zoroa had focused on this niche corner of the equations and laid out a method for building an external force that triggers a blowup. On September 7, two other mathematicians used that programme, together with AI, to produce a blowup for a frictionless fluid — regarded as a major step toward the full result. OpenAI completed the extension less than a day later, and Scientific American reports that the timing has triggered a heated dispute between those two mathematicians and the company.

Why the method hits a wall

Roughly a week and a half after OpenAI's announcement, three mathematicians posted a proof with a stark conclusion: if the external force is removed, the blowup disappears. OpenAI's method, they show, depends on a contrived forcing term unlike anything that could occur in nature, and it can never be made to work on the unforced problem without a completely new idea.

Luis Silvestre of the University of Chicago told Scientific American that the new work demonstrates the forced formulation was different from the problem researchers really wanted to solve. "The most important problem is unsolved," he said. "The Clay problem is settled, but the main problem for the Navier-Stokes equations is not."

An unexpected open question

The controversy has left fluid dynamicists grappling with a possibility few had considered: that the equations may blow up only when an external force is applied, and never from within the fluid itself. Gonzalo Cao-Labora told Scientific American that how OpenAI's contribution is ultimately judged depends on the answer, and that if the forced-only picture holds, the community may conclude the external force should never have appeared in the Clay statement.

There may also be a consolation for human mathematicians. LLMs excel at searching an infinite landscape of fluid scenarios and plucking out the precise one that breaks the equations, but proving that blowup is impossible — the other half of the Clay question — is theory-building work where Cao-Labora believes humans may still hold an advantage. Even so, he described the technology's recent progress as "a wake-up call to the community."

Why it matters

The episode is a case study in an AI failure mode that reaches far beyond pure mathematics: a system satisfying the letter of a formal specification while missing the intent behind it. A 26-year-old wording choice in the Clay statement created an opening, and an LLM found and exploited it — a dynamic equally relevant to benchmarks, contracts and safety evaluations written by humans.

It also sharpens the picture of where AI-driven discovery currently stands. Constructing explicit counterexamples is something models demonstrably do well and fast; building new mathematical theory to rule possibilities out remains hard. For anyone calibrating expectations about machine-made scientific breakthroughs, the Navier-Stokes saga suggests both reasons for caution and a clearer map of what humans still bring to the table.