OpenAI's Navier-Stokes Proof Raises Trust Questions for Math Research

OpenAI's Historic Math Claim Meets a Storm of Controversy

On September 8, 2026, OpenAI announced a result that would be a landmark in both mathematics and artificial intelligence: an internal model, significantly more capable than GPT-6 Astra, had produced a proof addressing the Navier-Stokes existence and smoothness problem—one of the seven Millennium Prize Problems. The company released a 165-page proof, along with a formalization in the Lean proof assistant, and stated the effort involved up to 10,000 AI agents running for roughly 88 hours at a cost of millions of dollars.

But the celebration has been short-lived. A dispute over scientific credit and the use of unpublished research data has quickly overshadowed the technical achievement. Two prominent mathematicians—Tristan Buckmaster of Princeton and Levent Alpöge of Harvard—are raising serious questions about whether OpenAI's systems may have benefited from their confidential work, which they had shared with the company's tools during the research process.

The Core Trust Problem: Unpublished Math and AI Training Data

The controversy strikes at a fundamental issue for the future of AI-assisted science: can researchers safely use frontier labs' tools to work on unpublished discoveries? If AI companies can potentially absorb and leverage the private intellectual work of researchers through their products, the entire collaborative fabric of mathematics could be compromised.

Buckmaster and Alpöge had been using OpenAI's Codex and ChatGPT to explore related problems in fluid dynamics. On September 7, they published three results on finite-time blowup for incompressible porous media, the Boussinesq system, and 3D incompressible Euler equations—work that is closely related to the Navier-Stokes problem OpenAI claims to have solved. The pair's Euler work, in particular, is acknowledged by OpenAI as having priority, but the company insists its proof for Navier-Stokes uses a different approach.

Direct Allegations and OpenAI's Qualified Response

Buckmaster has stated he asked OpenAI directly whether its systems had accessed the pair's Codex sessions, which contained their unpublished research. The company's response was evasive at best. OpenAI said its researchers and agents did not inspect those prompts or proofs to direct the project. However, it acknowledged: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."

This statement has done little to assuage concerns. In fact, it has amplified them. The admission that de-identified data may have influenced model improvements—without any way to audit or verify that claim—raises a profound provenance question. As one commentator on the Mastodon thread put it, "It's very easy to tell whether or not something went into the training data. It has nothing to do with interpretability."

A Pattern of Opaque Data Practices

This is not the first time OpenAI's data practices have come under scrutiny in the mathematical community. Andreas Thom, a mathematician at the University of Leipzig, has recounted a similar exchange with OpenAI following its announcement of a non-sofic group result, which used methods developed by Thom and his collaborator Gabor Kun.

Thom had emailed OpenAI researchers Mark Sellke and Sébastien Bubeck, expressing concern that his and his colleague's active use of ChatGPT while working on the expander matching problem might have influenced the training data. He explicitly asked two questions: whether their conversations entered training data, and whether they were accessible to the solving process. Sellke's complete answer was: "Regarding your conversations with ChatGPT: that did not happen."

Thom now views that response as misleading. "The categorical answer now looks as though it addressed only direct access under (2). No such qualification, explanation, or evidence was given. I take this as dishonesty to say the least." This history suggests a pattern where OpenAI's assurances may not be as comprehensive as researchers might hope.

The Data Controls Illusion

OpenAI offers users a setting to opt out of having their data used for model improvement, but its effectiveness is questionable. One Mastodon user pointed out that even with the opt-out enabled, data can still be ingested if users click thumbs up or down on a reply, select a preferred output, or trigger any ambiguous safety filters. Furthermore, backend model activations can be mapped and used, which may not count as "your data" under GDPR but can still capture ideas and implementations.

Another user highlighted a subtle trap: if you opt out of training but continue chatting in a conversation created before the opt-out, OpenAI may still appropriate that data. This complexity makes it nearly impossible for researchers to know whether their work has been absorbed into OpenAI's models, regardless of their privacy settings.

Legal and Ethical Implications

The dispute also raises novel legal questions. Under U.S. copyright law, an unpublished mathematical paper can receive protection once it is fixed in a sufficiently permanent form—publication is not required. However, copyright protects the author's expression, not the underlying mathematical idea or principle. This means that even if OpenAI used Buckmaster and Alpöge's unpublished proofs to inform its model, the legal recourse may be limited.

There is also a question of credit. Buckmaster alleges that OpenAI researcher Sébastien Bubeck discussed an arrangement under which Buckmaster alone could write up OpenAI's Navier-Stokes result after publishing the Euler work. Buckmaster further claims Bubeck twice raised removing Alpöge from that proposed authorship because Alpöge works at Anthropic, a competing AI lab. This adds a layer of corporate rivalry to an already complex scientific dispute.

What's at Stake for the Scientific Community

The lasting importance of this episode may not be whether OpenAI's proof is correct, but the precedent it sets for AI-assisted research. Frontier laboratories increasingly possess both the tools researchers use to develop unpublished ideas and the computational power to compete with those researchers. This creates a basic trust problem for science.

If major AI-assisted discoveries are to command confidence, laboratories may need far stronger provenance standards. This could include auditable training-data boundaries, timestamps, prompt histories, records of human intervention, and transparent rules for scientific attribution.

As one analysis concluded: "If machines can now discover mathematics, the next difficult problem may no longer be just 'Is the proof true?' but 'Where did the proof come from?'"

The Unresolved Question

OpenAI's statement that it "cannot rule out" the use of de-identified data does not establish that Buckmaster and Alpöge's unpublished mathematics influenced the Navier-Stokes result. But it also does not establish that it did not. The provenance question remains unresolved on the publicly available evidence.

For researchers considering whether to trust OpenAI with their unpublished work, the answer is far from clear. The company has not provided the transparency needed to reassure the mathematical community. As Thom put it, there is "a certain (frankly unacceptable) lack of transparency here; and I fear it will damage the communal process of math more than the new AI-generated results will benefit the subject."

The Navier-Stokes proof may eventually be verified and celebrated. But the trust deficit it has exposed could have far more lasting consequences for the relationship between AI companies and the scientific community.