OpenAI has announced a proposed solution to the Navier–Stokes Millennium Prize Problem.

But behind the mathematical breakthrough is a much stranger question:

Could researchers using AI to make scientific discoveries also be indirectly helping train the AI that eventually competes with them?

Here is what we actually know.

Mathematicians Tristan Buckmaster and Levent Alpöge had been working for roughly a year on a line of research involving fluid equations.

Importantly, the underlying mathematical program was not invented by them or by an AI. Buckmaster credits Diego Córdoba and Luis Martínez-Zoroa with pioneering the forced-blowup approach they were extending.

What changed was how aggressively AI could push that research forward.

Buckmaster says he and Alpöge used several frontier models throughout the project, including Claude and OpenAI Codex. On August 15, they obtained a finite-time blow-up result for the Euler equations with smooth forcing, and by August 22 they had verified it in Lean.

Buckmaster also says drafts from the project were being put into Codex sessions and that he personally paid a substantial OpenAI bill from his research funds.

Then the timeline starts to become interesting...

OpenAI says it began training a new, unusually strong mathematical model on August 28.

On September 1, OpenAI heard rumors that two Millennium Prize problems may have been solved.

It responded by launching an enormous automated research effort across the remaining Millennium problems.

For Navier–Stokes, OpenAI says the successful effort eventually involved on the order of 10,000 concurrent AI agents, millions of agent messages and around 130 billion generated tokens.

By September 5, the agents had produced the claimed Navier–Stokes construction. Lean verification followed.

And so the controversy starts.

Buckmaster says the direction OpenAI reached, a singularity involving smooth forcing, immediately caught his attention because it was closely related to the research program he and Alpöge had been pursuing privately.

He asked whether OpenAI's model had access to or had been trained on their Codex sessions.

OpenAI denies that its researchers or agents accessed their private work to solve the problem.

But OpenAI also included an unusually important sentence in its announcement:

It cannot rule out that de-identified data derived from the researchers' use of OpenAI products contributed to improving its models.

That statement is super important!!.

It does not prove that scientists gave OpenAI their unpublished proof, OpenAI trained on it, and the model then reproduced it but neither does it prove the opposite.

OpenAI's current consumer data policy says that ChatGPT and Codex interactions can contribute to model improvement when the relevant training setting is enabled, and users can opt out. We do not publicly know the exact data-control configuration used for every session involved in this research, additionally begs the question, if you opt-out, are you truly out?

That creates a completely new scientific dynamic.

For most of modern science, a researcher could privately work on an idea using a calculator, Mathematica, Python or a notebook without worrying that the tool itself might become a better competing researcher because of those interactions.

AI changes that equation.

Human researchers push into unexplored territory.

They use frontier AI systems to help reason, prove, code and test ideas.

Those interactions may, depending on the product and data settings, contribute signals toward improving future AI systems.

The AI lab then deploys a stronger generation of the model with thousands of agents and enormous compute.

The AI begins attacking the same frontier at machine scale.

That is not what we have proven happened in the Navier–Stokes case.

But for the first time, it is a mechanism serious researchers have to think about.

OpenAI denies using the researchers' private work directly, while acknowledging that it cannot completely rule out indirect influence from de-identified product-usage data.

Because if AI really is crossing from being a tool scientists use into becoming a scientist capable of racing them, academia is going to need new rules very quickly.

Data confidentiality.

Research provenance.

Training opt-outs.

Priority and attribution.

And perhaps most importantly:

What does it mean to share an unpublished idea with an AI system when the next version of that system may be capable of finishing the idea before you do?

Hence the push for sovereign AI, local models, sacrificing frontier glamour for privacy, else your idea becomes everyone's idea each time they update the model.

For an employer-facing view of making control conditions part of an AI proposition, read how to position an enterprise AI pilot.