On September 8, OpenAI published what it says is a solution to the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. An unreleased model, organized into roughly 10,000 concurrent agents, reached the result in 88 hours. OpenAI says the effort consumed about 130 billion output tokens and 2.7 million agent messages. The system produced a construction in which a smooth-forced fluid flow develops a singularity in finite time. (OpenAI)
The mathematics will take time to judge. Clay Mathematics Institute rules require a proposed solution to appear in a qualifying outlet, survive at least two years of scrutiny, and achieve general acceptance before the institute considers awarding its $1 million prize. (Clay Mathematics Institute)
The more immediate problem is provenance. Science has procedures for deciding whether a result is correct and customs for deciding who deserves credit. It has almost no machinery for dealing with a third case: a researcher's unpublished thinking enters an AI system, contributes in some unknowable way to later model capability, and then reappears as part of a machine-generated discovery.
The dispute between OpenAI and mathematicians Tristan Buckmaster and Levent Alpöge puts that problem in concrete form.
Buckmaster and Alpöge had spent much of the previous year working on closely related fluid-dynamics problems. Buckmaster says they used several AI systems, especially OpenAI's Codex, throughout the project, and that their Codex sessions contained drafts, partial arguments and other working material. By August 15 they had obtained finite-time blow-up results with smooth forcing for Boussinesq and three-dimensional incompressible Euler. Buckmaster says the Euler result was verified in Lean on August 22. (Buckmaster statement)
OpenAI says that on September 1 it heard rumours that two Millennium Prize Problems had been resolved and decided to test its new internal model against the remaining problems. Its agents first produced an unforced Euler result, then the company concentrated resources on Navier-Stokes. OpenAI says it had the result by September 5. (OpenAI)
Buckmaster's account raises a separate concern. He says information about his and Alpöge's progress had reached OpenAI before its effort began. After learning that OpenAI had pursued the smooth-forcing route, which he regarded as unusually close to the programme he and Alpöge had been developing from earlier work by Diego Córdoba and Luis Martínez-Zoroa, he asked whether their Codex sessions had been available to the model or used in training. (Buckmaster statement)
OpenAI says its researchers and agents did not access their specific work. It also says: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." (OpenAI)
That sentence leaves the central issue unresolved. OpenAI's account may be entirely correct, but an outside observer still cannot reconstruct a clean chain of intellectual provenance from the information available.
A provenance problem
Buckmaster is careful in his own statement. He says he does not know whether his data was used and is "not accusing anyone of anything." OpenAI says the proofs differ significantly and that its agents produced their work independently. Sam Altman has said the approaches appear different now that both sides' work is visible. (Buckmaster statement)
All of that may be true. The trouble is that none of it resolves the structural issue.
Priority disputes in science usually leave evidence behind. Researchers have notebooks, emails, seminar slides, preprints, referee reports and version histories. Those records do not prevent disputes, but they give the scientific community a chance to reconstruct how an idea moved from one person to another.
Machine learning creates a route that is much harder to inspect. A private idea can enter a training process after identifying information has been removed. It can become one small influence among millions of examples. Months later, a model may be better at recognizing a promising technique or choosing a line of attack. Thousands of agents can then explore that line at machine speed.
There may be no copied paragraph, no employee who opened a private file and no prompt that can be placed beside another prompt. The author can disappear from the record while some informational residue remains useful to the model.
OpenAI's current policy says that content submitted through individual services such as ChatGPT and Codex may be used to train models unless the user opts out, while business products are excluded from training by default. The company says it takes steps to reduce personal information before training. (OpenAI data controls)
Those protections are designed for privacy. Scientific priority asks a different question: can the origin of an unpublished idea still be traced if that idea influences a model after identifying information has been removed? For a scientist, de-identification can make attribution harder even when it improves privacy.
When the notebook can also enter the race
AI companies are taking on several roles that used to be separate. The same company can provide the research tool, store the researcher's working material, train the next model and operate its own research programme. It may also have enough compute to enter the same scientific race once it learns that a breakthrough is close.
Finance has spent decades building rules around comparable conflicts. An institution that handles market-moving information while also trading related assets is expected to maintain information barriers because intention alone does not remove the conflict. Frontier science is beginning to need an equivalent discipline.
OpenAI has been unusually explicit about what triggered its Navier-Stokes push. The company says rumours that others had solved Millennium Problems prompted the September 1 effort. Altman has said OpenAI heard that Anthropic's models might have achieved one and wanted to see whether its own models could do the same. (OpenAI)
Whatever happened to the underlying mathematics, information about another group's progress changed how OpenAI deployed a large research resource. With thousands of agents available, even a rumour about where human researchers are close to success can be strategically valuable.
This changes the economics of scientific secrecy. Researchers once worried mainly about another laboratory hearing an idea at a seminar and publishing first. Now the platform used to develop the idea may also belong to an organization capable of putting thousands of agents on the same problem within hours.
The authorship dispute
What happened after OpenAI says it obtained its result has produced a second controversy.
Buckmaster alleges that during discussions on September 6 he was offered two arrangements. In one, he and Alpöge would release their Euler work before OpenAI released its Navier-Stokes result. In the other, Buckmaster would become lead author of a presentation of OpenAI's Navier-Stokes result while Alpöge would be excluded. Buckmaster says Alpöge's employment at Anthropic was explicitly given as the reason. He also says that when he threatened disclosure, he was asked, "Why would you ruin your career?" (Buckmaster statement)
Sébastien Bubeck has described allegations against him as false and inflammatory. Altman's account says OpenAI was trying to collaborate, wanted Buckmaster's group to publish first, and found extending the same authorship arrangement to Alpöge difficult because he worked for Anthropic and was unwilling to coordinate with OpenAI. (Stanford Tech Review)
The competing accounts do not establish that OpenAI used Buckmaster's research. They do, however, expose an awkward collision between scientific authorship and corporate affiliation. Authorship is supposed to record intellectual contribution. If someone's employer becomes relevant to whether that person can appear on a paper, credit is no longer being allocated on scientific grounds alone.
That should concern researchers even if every claim about data usage is resolved in OpenAI's favour.
Trust becomes part of the product
A month before the dispute, Princeton mathematician Shou-Wu Zhang had voiced a related worry at the International Congress of Basic Science in Beijing. He said he was reluctant to discuss his ideas with AI because the machine might run faster once given the direction, and because he did not know where the idea might go afterward. He was especially worried about younger researchers, who do not have the protection of an established career. (The Paper)
Zhang was describing a practical limit on scientific AI. The most capable research assistant is useful only if researchers are willing to trust it with work that has not yet been published.
That creates a difficult incentive for AI companies. Better models become more useful for frontier research at exactly the point when confidentiality matters most. A weak model can be given routine work with little risk. A strong model invites users to share the tentative idea, failed proof or half-formed conjecture that might contain the next breakthrough.
The companies that win scientific users may therefore be the ones that can offer the strongest guarantees about what happens to unpublished work. Capability will matter. So will the ability to demonstrate that product data cannot leak, directly or indirectly, into a competing research programme.
OpenAI has already built part of this model elsewhere. In July it launched a programme offering frontier-model access to as many as 100,000 academic researchers. The company says those dedicated workspaces have business-grade protections and are not used for model training by default. (OpenAI academic research programme)
That programme weakens the claim that free academic access is simply a way to harvest research. It also shows that stricter defaults are possible.
For frontier research, they should become standard. Scientific workspaces should be non-training by default. Companies should maintain auditable barriers between product data and internal teams competing in the same fields. Provenance records for major machine-generated discoveries should identify the model version, relevant training cut-offs, external data access, human interventions and the flow of information between agents. Independent auditors need enough access to test those records without exposing confidential research.
Universities, meanwhile, should negotiate AI access as carefully as they negotiate intellectual-property terms in industry partnerships. A cheap or free research tool is not cheap if its governance makes priority impossible to establish later.
Science needs a chain of custody
The OpenAI result may turn out to be a major mathematical achievement. Its agents produced both an analytical argument and a Lean formalization at a scale that would have been difficult to imagine only recently. If the proof survives the process set out by Clay, it will deserve recognition on mathematical grounds. (OpenAI)
The institutional problem will remain either way.
Most arguments about AI and intellectual property concern finished work: books, images, music, software and other material that already exists in a recognizable form. Frontier science is different because the most valuable information often appears before publication, when an idea is incomplete and its ownership is easiest to lose.
Researchers cannot solve this by refusing to use AI. If the productivity gap keeps widening, abstention will become professionally costly. Nor is it reasonable to ask scientists to trust assurances that outsiders cannot audit. The sensible answer is to build systems that preserve a chain of custody for ideas while still allowing researchers to use powerful models.
Science has always needed to answer two questions about a breakthrough: whether the result is correct, and who should receive credit for finding it. Formal verification may make the first question easier in some areas of mathematics. AI makes the second harder because the path from private human thought to model behaviour can be opaque even when nobody intends to copy anything.
A machine-generated proof therefore needs more than mathematical verification. It needs a credible provenance record.
Until that exists, every major AI-assisted discovery will carry a question that the theorem itself cannot answer: where did the idea come from?