OpenAI's internal LLM produced a proof that satisfies a Clay Institute variant of the Navier–Stokes Millennium Problem by using a specially designed external force. Many mathematicians argue that the force-dependent result does not settle the core, force-free question they consider physically meaningful. A recent critique shows the LLM's method cannot be adapted to the unforced case, leaving the central problem open and prompting broader debate about AI's role in mathematical discovery.
Did OpenAI Solve the Wrong Navier–Stokes Problem? The Debate Over a Force-Dependent Proof
OpenAI’s solution exploits an external force to break the Navier-Stokes equations, which are supposed to describe how fluids, such as water, flow.
Two weeks after OpenAI announced that an internal large language model had produced a proof resolving a Millenium Prize instance of the Navier–Stokes problem, the mathematical community is wrestling with a new controversy: did the company address the version of the problem that mathematicians consider meaningful?
The LLM's proof follows an approach many experts call unconventional: it constructs a very specific external forcing term and uses that force to produce a finite-time blowup. While this satisfies one option in the Clay Mathematics Institute's official statement, many researchers say the result is disconnected from the core, physically motivated question — namely, whether the Navier–Stokes equations can blow up without any contrived external force.
What OpenAI Claimed
OpenAI's submission reportedly meets the Clay Institute's option "C," which permits a prescribed external force in the formulation. Under that allowance, the LLM produced an explicit construction of initial data and a forcing term that leads to a singularity (a point where velocity becomes unbounded), thereby fulfilling the prize conditions for that variant.
Why Many Mathematicians Are Unsatisfied
Most researchers focus on the force-free Navier–Stokes problem as the most physically relevant and mathematically deep question. They seek a blowup that emerges solely from the equations' intrinsic nonlinear dynamics, without relying on a custom-designed external agent. As University of Chicago mathematician Luis Silvestre summarized: "The Clay problem is settled, but the main problem for the Navier–Stokes equations is not."
Adding to the debate, a recent paper by three mathematicians demonstrates that OpenAI's method cannot be extended to the force-free setting. According to that critique, the LLM's approach inherently depends on the special structure of the forcing term and so cannot, even in principle, produce a blowup when the force is removed.
Background And Recent Sequence Of Results
In recent years, Diego Córdoba and Luis Martínez-Zoroa concentrated on constructing very particular forcing terms that could trigger singularities. On September 7, two mathematicians applied that program (with AI assistance) to create a blowup for a frictionless fluid — seen by some as a step toward the broader problem — and OpenAI published a related construction less than a day later. The proximity of the results intensified disputes about credit, method and scientific norms.
Implications
The controversy has two main implications. Technically, it highlights that the Navier–Stokes question depends sensitively on how the problem is framed: allowing an external force makes a mathematically distinct problem that may be easier to break. Conceptually, the episode underscores strengths and limits of current LLMs: they excel at searching vast spaces to find explicit counterexamples or constructions, but they remain less persuasive at generating the new theory or impossibility proofs that often characterize deep mathematical advances.
"They essentially prove that the formulation with an external force was different from the problem we really wanted to solve," — Luis Silvestre.
What Comes Next
Fluid dynamicists must now weigh whether the Clay problem's option permitting external forcing was a misstep or a legitimate variant deserving the prize. If it turns out blowups can occur only with contrived forcing, mathematicians may debate whether the original statement should be revised. Regardless, the episode has prompted renewed scrutiny of AI-driven proofs and a wider conversation about how the mathematical community validates and credits results produced with machine assistance.
"We are really amazed with how [the technology] has evolved in the last year — so we don't know how it will look in one year," said Gonzalo Cao-Labora. "It's really a wake-up call to the community."
Help us improve.
