At the frontier where human ingenuity and machine learning converge, a long-standing mathematical mystery about fluid dynamics has reportedly been solved — but the announcement has unsettled the scientific community less for what was discovered than for how the discovery came to be. Tristan Buckmaster, an NYU mathematician who used OpenAI's tools in his own related research, has raised careful but pointed questions about whether his inputs may have trained the very models that now claim to have outpaced him. The episode does not yet resolve into clear wrongdoing, but it opens a deeper reckonin
OpenAI's Math Breakthrough Claim Ignites Credit and Data Training Dispute
The gap between what companies are legally permitted to do and what the scientific community considers ethical has widened into a chasm.
Mark
So what's actually at stake here? Is this just about who gets their name on a paper?
Mimi
It's bigger than that. Buckmaster used OpenAI's tools to do his research. If OpenAI's models learned from what he typed in, and then those improved models helped OpenAI solve the same problem he was working on, that's a direct conflict of interest.
Luke
But we don't know that happened. OpenAI said it's "unlikely" their models learned from his inputs, and Buckmaster himself said he doesn't know if his data was used. We're working with a lot of "could have" here.
Mimi
True, but that's exactly the problem. The terms of service say OpenAI trains on user data by default. Buckmaster didn't say whether he opted out. Most people don't even know that option exists.
Mark
So OpenAI is allowed to do this legally?
Mimi
Yes. Unless you actively opt out or have an enterprise account with different terms, your inputs become training data. It's in the fine print.
Luke
And to be fair, OpenAI did say their proofs were different from Buckmaster's work. They're not claiming they copied his solution.
Mark
But if they learned from his research inputs, wouldn't that give them an advantage?
Mimi
Potentially, yes. And that's what Buckmaster seems to be raising—not that they stole his work, but that the whole setup creates a perverse incentive. Why would you use a company's AI tools if that company might use what you learn to compete against you?
Luke
The other thing we don't have is OpenAI's actual proof. We're taking their word that they solved it. We haven't seen the math.
Mark
So this could drag on for a while.
Mimi
It already has. And it's going to force a reckoning about how AI companies handle user data when they're also doing their own research.
O Pulso
- OpenAI announced it solved the Navier-Stokes Millennium Prize Problem — one of mathematics' most coveted unsolved challenges — on the same day two human mathematicians published their own related breakthrough.
- Mathematician Tristan Buckmaster alleges that an OpenAI researcher pressured him to alter publication plans and suggested removing his co-author, who works at rival Anthropic, from the paper.
- OpenAI acknowledged it cannot fully rule out that data from Buckmaster's own use of its tools helped train the models that may have competed against him — a quiet admission with enormous implications.
- The AI and mathematics communities have reacted with alarm, questioning whether existing norms of scientific attribution and research ethics can survive the era of AI-assisted discovery.
- The dispute exposes a structural tension baked into AI business models: users generate data that trains systems, and those systems may then outcompete the very users who helped build them.
At the frontier where human ingenuity and machine learning converge, a long-standing mathematical mystery about fluid dynamics has reportedly been solved — but the announcement has unsettled the scientific community less for what was discovered than for how the discovery came to be. Tristan Buckmaster, an NYU mathematician who used OpenAI's tools in his own related research, has raised careful but pointed questions about whether his inputs may have trained the very models that now claim to have outpaced him. The episode does not yet resolve into clear wrongdoing, but it opens a deeper reckoning: as AI becomes a collaborator in human knowledge-making, the old conventions of credit, consent, and competition may no longer hold.
On the same Monday, two mathematicians — Tristan Buckmaster of NYU and Levent Alpöge of Anthropic — released papers on one of mathematics' most enduring puzzles: whether the equations governing fluid flow can catastrophically break down. Terence Tao, among the world's most celebrated mathematicians, called their findings remarkable. But the math was quickly overshadowed by a four-page statement Buckmaster published alongside his paper.
In it, Buckmaster described a troubling sequence of events. After sharing his publication plans with an OpenAI mathematician, he received a call from Sebastian Bubeck, a computer scientist at the company. Bubeck told him that OpenAI's internal AI had produced its own proof of the Navier-Stokes problem — the one carrying a $1 million Clay Mathematics Institute prize. Buckmaster also said Bubeck proposed coordinating their announcements and suggested removing Alpöge as a co-author, apparently because Alpöge works at OpenAI's rival. Buckmaster was careful in his framing: "I have not seen OpenAI's proof. I do not know whether our data was used. I am not accusing anyone of anything."
The following day, OpenAI confirmed its solution and congratulated the two mathematicians, insisting its researchers had not seen their work before it went public. But on the question of whether Buckmaster's inputs into OpenAI's tools had helped train its models, the company hedged: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." Bubeck denied asking for Alpöge's removal, saying he was only trying to coordinate announcements and that OpenAI had been spurred to act after rumors spread that Anthropic had already solved two Millennium problems.
The episode crystallizes a dilemma embedded in how AI companies operate. OpenAI's terms of service allow it to train on user data by default unless users explicitly opt out. Buckmaster paid for the tools himself but did not say whether he had opted out. The uncomfortable question his statement raises is whether a researcher can use an AI platform to advance their work, only to find that the platform — potentially sharpened by learning from that very research — has raced ahead of them in the same domain.
The broader scientific community has responded with unease. When AI is woven into discovery, credit becomes contested terrain. Buckmaster himself acknowledged the foundational contributions of other mathematicians to his approach, underscoring how knowledge builds across many hands. As AI grows more central to research, the gap between what companies are legally permitted to do and what the scientific community considers ethical appears to be widening into something harder to bridge.
On Monday, two mathematicians released papers describing new findings about one of mathematics' most stubborn unsolved problems—whether the equations governing fluid flow can suddenly fail. Tristan Buckmaster, who teaches at NYU, and Levent Alpöge, who works at Anthropic, had used AI tools to help push their research forward. Their work was solid enough that Terence Tao, one of the world's most respected mathematicians, called it "remarkable."
But what grabbed attention wasn't the math itself. It was the four-page statement Buckmaster published alongside the papers, in which he described a series of conversations with OpenAI staff that left him unsettled. He said that after telling an OpenAI mathematician about his plans to publish, he received a call from Sebastian Bubeck, a computer scientist at OpenAI. During that conversation, Bubeck told him that OpenAI's internal AI model had produced a proof of the Navier-Stokes problem—the very question that carries a $1 million prize from the Clay Mathematics Institute. Buckmaster also said Bubeck suggested publication arrangements and wanted Alpöge removed as an author because Alpöge works at Anthropic, OpenAI's rival. "I have not seen OpenAI's proof," Buckmaster wrote. "I do not know whether our data was used. I am not accusing anyone of anything. I am stating what I was told, when, and what was proposed to me."
On Tuesday, OpenAI announced it had indeed solved the Navier-Stokes problem. The company congratulated Buckmaster and Alpöge on their related work and said its researchers and AI agents had not seen their work until it became public. But when addressing whether OpenAI's models had learned from the mathematicians' inputs—the data they had typed into OpenAI's tools while developing their research—the company's language shifted. "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models," OpenAI stated. The company added that its proofs differed from Buckmaster and Alpöge's work, particularly in the specific mathematical cases they addressed.
Bubeck pushed back on Buckmaster's account in a post on X, denying he had asked for Alpöge to be removed from authorship. He said he was trying to "coordinate our releases" of the mathematical announcements and that OpenAI had been motivated to tackle the problem after seeing "viral Twitter rumors that Anthropic had resolved 2 Millennium problems." He also characterized Buckmaster's allegations as "false and inflammatory."
The dispute cuts to the heart of how AI companies operate. OpenAI's terms of service state that the company trains its models on data submitted by users unless they explicitly opt out through the company's privacy portal. "When you share your content with us, it helps our models become more accurate and better at solving your specific problems and it also helps improve their general capabilities and safety," OpenAI wrote in a March post. Buckmaster's statement said he paid for the tools himself but did not specify which product he used or whether he had opted out of the training program. The question hanging over the episode is whether a researcher can knowingly use a company's AI tools to advance their own work, only to have that company's models—potentially improved by learning from that very research—compete against them in the same domain.
The math and AI communities have reacted with alarm. Sholto Douglas, a member of Anthropic's technical staff, called the situation "extremely sad." Others have highlighted the broader implications: when AI is woven into scientific discovery, who deserves credit? If a researcher's inputs help train a model that later solves a problem the researcher was working on, what obligation does the AI company have to disclose that connection? Buckmaster credited mathematicians Diego Cordoba and Luis Martinez-Zoroa for foundational work from 2024 that enabled his approach, raising the question of how credit should flow when multiple parties—human researchers, AI companies, and their models—all contribute to a breakthrough. The episode suggests that as AI becomes more central to research, the gap between what companies are legally permitted to do and what the scientific community considers ethical has widened into a chasm.
Citações Notáveis
I have not seen OpenAI's proof. I do not know whether our data was used. I am not accusing anyone of anything. I am stating what I was told, when, and what was proposed to me.— Tristan Buckmaster, in his statement accompanying the research papers
While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.— OpenAI, in its response to Buckmaster's allegations
Quer a matéria completa?
Leia o original em Business Insider ↗
Technology