OpenAI’s Navier–Stokes Breakthrough Sparks Debate Over AI and Scientific Credit
14 Sept 2026
OpenAI says a system of roughly 10,000 AI agents produced a machine-verifiable solution to the decades-old Navier–Stokes problem in 88 hours. The achievement has also triggered a dispute over research priority, private AI-tool usage and how scientific credit should work when machines accelerate discovery.
Artificial intelligence is beginning to move beyond assisting scientific research and into something much more consequential: generating results at a scale and speed that can compete directly with human researchers.
On 8 September 2026, OpenAI published a proposed solution to the Navier–Stokes existence and smoothness problem, one of mathematics’ Millennium Prize Problems. The company released both a written proof and a formalisation in the Lean theorem prover, allowing the mathematical argument to be checked computationally.
OpenAI says its internal system used about 10,000 coordinating AI agents, with the successful effort taking approximately 88 hours. The agents exchanged around 2.7 million messages and generated roughly 130 billion output tokens, followed by another 17 hours of Lean formalisation and verification.
The achievement is extraordinary in its own right. But the scientific debate that followed may ultimately prove just as important.
A 90-Year Mathematical Problem Meets 10,000 AI Agents
The Navier–Stokes equations describe the movement of fluids such as water and air. Mathematicians have understood the equations for centuries, but a central question about whether smooth three-dimensional solutions can always remain well behaved has resisted proof for roughly 90 years.
OpenAI says its system found a finite-time singularity, demonstrating that smooth fluid dynamics can break down under the conditions considered in its proof. Unlike ordinary generative-AI output, the result was also translated into Lean, a formal theorem-proving system that checks mathematical arguments step by step.
That distinction matters.
AI systems can produce highly convincing but incorrect mathematical reasoning. Formal verification creates a separate layer of scrutiny by requiring the proof to satisfy machine-checkable logical rules.
OpenAI itself has described the work as a proposed solution rather than treating the result as the final word. The company has also said it does not intend to claim the $1 million Millennium Prize, leaving broader mathematical review and institutional assessment to continue.
Then Came the Scientific Credit Dispute
The announcement quickly became more complicated because Tristan Buckmaster, a mathematics professor at New York University, and mathematician Levent Alpöge, who works at Anthropic, had been pursuing closely related work.
Buckmaster said he and Alpöge had been making significant progress and that drafts of their research had been entered into OpenAI’s Codex during the project. He subsequently questioned whether those interactions could have influenced OpenAI’s system. Contemporary reporting documented his concerns and the disagreement between the researchers and OpenAI.
This is where an important distinction is necessary.
An article circulated on Medium claimed OpenAI had effectively admitted that Buckmaster’s data might have contributed to model training. However, OpenAI later updated its official account on 10 September after conducting an investigation, stating that Buckmaster’s Codex prompts from the previous two months “could not have influenced the system in any way, including through training.”
OpenAI says its researchers and agents did not see Buckmaster and Alpöge’s work before it became public and that the two approaches also produced materially different results.
The allegation of direct appropriation therefore remains disputed, and the stronger claim that OpenAI “stole” the proof is not established by the available evidence.
The Controversy Is Still Bigger Than One Proof
Even if OpenAI’s investigation closes the narrow question of whether Buckmaster’s recent Codex prompts trained the system that produced the result, the episode exposes a much larger problem for science.
Researchers increasingly use AI systems while their work is still unfinished.
They upload code, equations, experimental ideas, failed approaches and early drafts. In traditional science, those materials might remain inside notebooks, university servers or private correspondence until publication.
With cloud-based AI tools, they can instead pass through systems operated by some of the most powerful technology companies in the world.
That changes the meaning of scientific confidentiality.
The question is no longer simply whether an AI model can solve a problem. It is whether researchers clearly understand what happens to the knowledge they contribute while trying to use those systems.
When AI Accelerates Science, What Counts as Priority?
Scientific culture has traditionally relied on concepts such as authorship, priority, peer review and citation.
AI complicates every one of them.
If thousands of agents can explore millions of mathematical pathways in several days, the period between a promising idea and a completed proof can collapse dramatically. Researchers who once expected months or years to develop a result may find themselves competing with machine systems capable of testing huge numbers of approaches simultaneously.
Buckmaster’s experience illustrates this pressure even without assuming wrongdoing.
According to OpenAI, it began its large-scale effort after hearing a rumour that another team might be making progress. It then deployed enormous computational resources and produced a result within days.
This introduces an unfamiliar form of scientific competition: compute asymmetry.
A university mathematician may have an insight. An AI laboratory may have thousands of agents capable of pursuing that insight at industrial scale.
Scientific priority could therefore increasingly depend not only on who has the idea first, but on who has the computational resources to finish first.
Machine Verification Changes the Meaning of AI-Generated Knowledge
There is another side to the story.
The Lean formalisation demonstrates one of the strongest possible models for using AI in advanced mathematics: let AI propose reasoning, then subject that reasoning to a system that does not simply trust its language.
That approach could become extremely important for scientific AI.
Large language models are probabilistic systems. Formal proof assistants operate under explicit logical constraints. Bringing the two together creates a structure in which AI can explore aggressively while verification remains independent of the model’s confidence.
This does not eliminate scientific review. It does, however, offer a way to distinguish mathematically valid results from outputs that merely sound convincing.
OpenAI’s result therefore points to a future in which AI could help search enormous spaces of possible proofs while theorem provers provide an automated layer of mathematical discipline.
Science Now Needs Governance for AI-Assisted Discovery
The controversy also demonstrates that technical verification alone is not enough.
A proof can be mathematically valid while questions about authorship, data provenance, research ethics and institutional power remain unresolved.
Universities and laboratories may increasingly need explicit rules covering unpublished work entered into AI systems. Researchers need to know whether prompts are retained, how data is used, whether it can contribute to model improvement and what contractual protections apply.
The Medium article that triggered much of the debate framed this as a data-provenance problem rather than purely a credit dispute. That broader point remains relevant even after OpenAI’s subsequent clarification about Buckmaster’s specific prompts.
The responsible response is therefore not to conclude that researchers should stop using AI.
It is to build scientific norms that recognise AI systems as part of the research infrastructure.
The Beginning of a Different Kind of Mathematics
The lasting importance of the Navier–Stokes episode may not be whether an AI system reached one famous result before a human research team.
It is what happens next.
Ten thousand agents working simultaneously, exchanging millions of messages and producing a formally verified mathematical argument in days represent a fundamentally different research process from the traditional image of a mathematician working alone at a blackboard.
The human role does not disappear. Researchers still choose questions, evaluate meaning, develop theories and decide which results matter. But AI can dramatically alter the scale at which possible solutions are explored.
That leaves science with a difficult new challenge.
If machines can accelerate discovery faster than our institutions can adapt authorship, privacy and research norms, can scientific governance evolve quickly enough to preserve trust in the knowledge they produce?
About the Source and the Dispute
This article was prompted by Mohamed Abdelmenem’s 14 September analysis of the OpenAI–Buckmaster controversy, which raises concerns about scientific data provenance and the use of private research in AI systems.
Some claims in that article predate or conflict with OpenAI’s later public clarification. This adaptation therefore distinguishes between Buckmaster’s allegations, third-party reporting and OpenAI’s updated September 10 position rather than presenting the disputed data-use claim as established fact.
Sources
- OpenAI — “On the Navier–Stokes Millennium Prize Problem,” 8 September 2026, updated 10 September 2026. Includes the proof methodology, agent numbers, Lean verification and OpenAI’s subsequent investigation into Buckmaster’s Codex prompts. OpenAI research publication
- The Guardian — “OpenAI claims to have solved maths problem that stumped humans for decades,” 8 September 2026. Covers the mathematical result and early controversy. The Guardian report
- ABC News — “Controversy erupts as OpenAI claims solution to Navier Stokes maths problem,” 10 September 2026. Includes Buckmaster’s account of his discussions with OpenAI. ABC News report
- The Verge — OpenAI Navier–Stokes controversy, 12 September 2026. Examines the dispute over competition, research credit and AI laboratories’ growing role in mathematics. The Verge report
- Mohamed Abdelmenem — “OpenAI Solved a 90-Year-Old Math Problem. The Mathematician Says They Stole His Work,” 14 September 2026. Original commentary supplied for this adaptation.
Share
Sara Srifi
Sara is a Software Engineering and Business student with a passion for astronomy, cultural studies, and human-centered storytelling. She explores the quiet intersections between science, identity, and imagination, reflecting on how space, art, and society shape the way we understand ourselves and the world around us. Her writing draws on curiosity and lived experience to bridge disciplines and spark dialogue across cultures.
