A proposed solution to a Millennium Problem has put into dispute something no equation decides alone: who deserves credit when humans and AI discover together.
OpenAI announced, in 8 September, a proposed solution for the existence and smoothness problem of Navier–Stokes. The result could be historic. But the announcement sparked a dispute over authorship and trust between the company and researchers working on related issues. OpenAI released a written demonstration and a formalization in Lean; this is not, by itself, definitive recognition from the Clay Institute. OpenAI Announcement · Clay Rules.
Beyond its possible scientific impact, the case raises three different questions: is the proof correct? Has the discovery's history received proper credit? And what happens when the tool provider also contests the discovery? Mixing them makes for an explosive headline. Separating them helps understand why this story matters.
What Navier–Stokes asks — without needing to go back to college
The equations describe fluid motion. The Millennium Challenge is not to find a formula that solves any stream of water. It is to clarify, under precise mathematical conditions, whether an initially smooth motion can cease to be smooth over time. This is one of the problems highlighted by the Clay Institute.
In this mathematical context, OpenAI's manuscript presents a three-dimensional, incompressible construction with positive viscosity. There is a smooth external force. The velocity starts at zero and becomes unbounded in finite time, while kinetic energy remains bounded. The text claims to satisfy alternatives C and D of the official formulation. Manuscript, theorem 1.1.
"With external force" doesn't automatically mean the company chose an irrelevant problem: this possibility is part of the official statement. It also doesn't allow announcing that all fluid and turbulence questions are solved. This is a mathematical result with specific hypotheses, not an experimental demonstration of water reaching infinite velocity.
Why does this matter outside mathematics?
You don't need to know Navier–Stokes to live surrounded by its applications. The air passing over a wing and the blood flowing through an artery are fluids in motion. Turning that motion into calculations allows studying situations that would be expensive, difficult, or impossible to observe in full detail.
| Where it appears | What the mathematics already serves for |
|---|---|
| ✈️ Airplanes | Simulate the air around the aircraft and estimate forces and stability. NASA used simulations in the X-57 project to build aerodynamic data and feed its flight simulator. NASA Case. |
| 🌦️ Weather forecasting | Calculate the atmosphere's evolution from equations of motion and other physical processes. The Met Office's Unified Model is an example—with hypotheses different from the incompressible problem under discussion. Met Office. |
| 🫀 Blood circulation | Estimate how blood flows and compare simulations with measurements. One study combined fluid dynamics and MRI to reconstruct carotid flow. Original research. |
These applications existed before the announcement. They show why studying fluids matters; they are not benefits produced by the new proof. In practice, simulations work with approximations and specific conditions that must be confronted with observations. It is not necessary to solve the Millennium Problem to calculate useful flow for engineering. NASA Introduction.
And what exactly can this discovery change?
Think of the equations as a map. Using the map to take a trip is one thing; proving it works in all allowed situations is another. A counterexample can show the limits of a general guarantee without invalidating the routes we already know.
This is the immediate scientific value if the proof is validated: to show that, under the conditions constructed in the manuscript, starting with smooth motion is not enough to guarantee velocity remains bounded indefinitely. It is an answer about the model's mathematical limits, not a recipe to predict any fluid. Presented theorem.
Why value discoveries like this? Because mathematics does not just produce answers: it produces tools and identifies which guarantees are possible. A demonstration can open techniques for other research; a proven limit can prevent pursuing an impossible guarantee. This is the connection we propose here — a possibility for future advancement, not an industrial application already demonstrated in this case.
One possible path would be: validated proof → new questions and methods → tested software → practical benefit. No step is automatic. There is no evidence in the sources consulted that this result reduced airplane fuel consumption, improved rain forecasting, or brought a new medical treatment. It also does not mean a real fluid will reach infinite velocity, nor that existing simulations have stopped working.
In summary: the breakthrough may be huge for knowledge without changing a product tomorrow. Understanding the rules — and their limits — is also a way to prepare technology for the future, still without a guaranteed deadline or specific benefit.
It was not a magic question to ChatGPT
According to OpenAI, the responsible group gathered about 10 thousand agents and reached the result in 88 hours. The internal model used was more capable than GPT-6 Astra; Astra participated in the subsequent formalization and verification. These are company statements, not an independent performance audit. Experiment report.
The distinction matters: an impressive answer in a conversation and a research program with tools, coordination, and verification are not the same. For those following AI, the story is less about a perfect phrase and more about how to organize difficult intellectual work.
The human work that doesn't fit in the footnote
Tristan Buckmaster, from NYU, reports a personal collaboration with Levent Alpöge, a researcher employed by Anthropic, without institutional agreement between employers. The pair published related results, including Euler with smooth force, and used tools from both companies. Buckmaster attributes the basic program idea to Diego Córdoba and Luis Martínez-Zoroa. Buckmaster's statement.
The OpenAI manuscript version consulted in 10 September also discusses Córdoba and Martínez-Zoroa's work in its historical section. This allows verifying the presence of these references today; it doesn't prove all previous versions contained the same content. Section 1.1 of the manuscript.
Recognizing antecedents does not diminish a later breakthrough. A discovery can depend on a good question, a prior strategy, and new execution. Crediting only the last participant impoverishes the story; crediting only the first also doesn't explain what was missing to solve.
Where the drama begins
In the statement, Buckmaster describes publication proposals and alleges pressure to exclude Alpöge from presenting OpenAI's result. He also reports a remark about harming his career. He questions whether his Codex sessions may have contributed to training but states he doesn't know if his data was used. Original report.
There is a rebuttal. Sébastien Bubeck's response, reproduced with link to the original by Growth Academy, denies requesting Alpöge's removal from the work's authorship; places the discussion in a rewrite of OpenAI's proof and records an apology for the career remark. Response reproduced and contextualized · Bubeck's post on X.
OpenAI denies access to unpublished work and specific researchers' data to solve the problem. At the same time, it says it cannot completely exclude indirect contributions from anonymized product use data to improve its models. Company position.
These statements do not demonstrate plagiarism. Nor do they close questions about credit and trust. The responsible point is to maintain attribution of each account and acknowledge what remains without public proof — without exchanging investigation for bias.
Lean helps verify; it doesn't decide who deserves credit
OpenAI made available a repository of formalizations in Lean 4. A formal verifier allows checking arguments expressed in a rigorous language. The evaluation also needs to consider the formalized statement, hypotheses, and correspondence with the mathematical claim. Our editorial reading is not equivalent to executing or auditing this proof.
Even a correct proof doesn't decide how a collaboration happened. Authorship and conduct require other types of evidence. Likewise, a controversy over behavior does not automatically make a demonstration false.
The Clay requires publication in a qualified venue, a minimum two-year interval, and general acceptance by the mathematical community before considering a proposal for the prize. This deadline is a prize rule, not a prohibition on experts evaluating the work earlier. Official rules.
The lesson for companies: intelligence requires responsibility
From XMACNA's perspective, the business question is not whether every company needs thousands of agents. It is whether it can explain how its AI reached an important decision.
Imagine a team using AI to develop its own solution. Before celebrating speed, it needs to be able to answer: what contributions came from people? What documents support the result? Who reviewed it? What usage conditions were chosen for that environment?
It is not necessary to turn every task into an academic investigation. Control should be scaled to the consequence. An internal draft allows for one level of review; a public statement about third parties requires another.
For AI agents in companies, three decisions help make this concrete:
- Define what counts as a completed task. Generated text does not replace verified results.
- Preserve the origin of the work. Sources, contributions, and revisions must remain identifiable.
- Name who is responsible. AI involvement does not remove who decides to publish or act.
This reasoning applies from a research study to a Digital Vendor: autonomy is only useful when the organization knows what it has delegated and can track the outcome. This is the criterion we advocate for Digital Employees, without promising that a commercial product will replicate this scientific experiment.
Moving faster is valuable. Being able to explain the path remains essential.
Want to identify where this combination makes sense in your operation? Start with the XMACNA AI assessment.
Frequently asked questions
Did OpenAI definitively solve Navier–Stokes?
They presented a proof proposal. Announcement, independent evaluation, and Clay recognition are distinct stages; this report does not certify the demonstration.
Does the problem with an external force count?
Alternatives C and D in the official statement consider external forces under specific conditions. The important thing is to verify these conditions, not dismiss the result simply due to the word “force.”
Does the dispute prove research theft?
No. Conflicting reports and an unresolved question alone do not constitute proof of plagiarism.
Does this already improve airplanes, weather forecasts, or treatments?
No benefit of this type has been demonstrated for this proof in the sources consulted. These areas already use fluid models. If validated, the new result clarifies a fundamental issue; later applications would depend on their own research, development, and testing.
What can a company learn from this?
That quality execution must be accompanied by proportional review, preserved sources, and clear responsibility — even when AI does most of the work.
Updated in 10 September 2026. Scientific assessment and controversy may have new developments.