When an AI Lab Chases a Millennium Problem
The conflict began with an extraordinary claim: an AI system had produced a proof of the Navier-Stokes equations, one of seven Clay Mathematics Institute Millennium Problems and a challenge that has resisted mathematicians for decades. The episode should have been a landmark for AI-assisted research. Instead it has become a cautionary tale about the way ideas move between independent researchers and the powerful labs that control both the algorithms and the platforms used to develop them.
All major parties have now made public statements, and the accounts differ sharply. At the center is mathematician Tristan Buckmaster, who says OpenAI engaged in misconduct after learning of a research project he was pursuing with Levent Alpöge. OpenAI denies the most serious accusations, but admits its actions were shaped by rumors about what rival lab Anthropic’s models had achieved.
Allegations of Pressure and Plagiarism
Buckmaster has been unusually blunt in his criticism. He says OpenAI took material from drafts he and Alpöge uploaded to OpenAI’s Codex environment, pressured him during the research process, attempted to remove Alpöge as co-author simply because Alpöge is employed by Anthropic, and threatened him with consequences for his career. In Buckmaster’s view, the behavior amounts to absolute academic malpractice.
The accusation of pressure is impossible to verify from the outside. What can be documented is that OpenAI heard rumors that Anthropic’s models had solved a Millennium Problem and, by its own admission, pointed its resources at the same target. OpenAI CEO Sam Altman and researcher Sébastien Bubeck deny plagiarism. They acknowledge their team developed a model specifically for the problem after the rumor circulated, but they reject the suggestion that the company’s result was lifted from Buckmaster and Alpöge.
Two Very Different Versions of the Same Work
The technical core of the disagreement is originality. OpenAI maintains that its solution differs significantly from the mathematicians’ work. Buckmaster and Alpöge disagree. They note that they had entered a similar approach into OpenAI’s systems, and they suspect that the input ended up in training data. If true, the company may have accelerated past its competitors using the very ideas those competitors had entrusted to its development tools.
That possibility matters for a simple reason: researchers do not normally submit their unpublished work to a competitor for training. But when AI laboratories provide coding assistants and cloud platforms, users may be feeding their most promising intellectual property directly into a system that can improve the lab’s own models. The boundary between using a tool and surrendering research to it becomes difficult to police.
The Training Data Question Remains Unanswered
OpenAI employees point out that the chance the company trained on the submitted solutions is low, especially if the two researchers disabled the option that would allow their inputs to be used for training. Whether they did so is not publicly known. From a scientific standpoint, that uncertainty is exactly what makes the dispute dangerous. When an allegation of this kind cannot be investigated independently, researchers are left relying on the word of a corporation that has no legal obligation to open its training pipeline to outside inspection.
OpenAI researcher Boaz Barak has pushed back on the idea that the model needed outside help at all. In his telling, the model did not merely reproduce a known argument; it began by proving a stronger claim than the mathematicians had made. To suggest otherwise, in his view, is coping rather than realistic assessment. Yet this defense does little to resolve the deeper question of provenance. Even a model that ultimately exceeded the researchers’ work could have been influenced by their drafts. Scientific integrity is not only about whether output was copied; it is equally about whether an idea was taken without permission.
A Warning From Terence Tao
One of the most sobering responses has come from mathematician Terence Tao. He has warned that future events of this kind may leave independent mathematicians in an impossible position. When a rumor reaches a large technology company, that company can mobilize substantial compute and talent within days. A small team of academics cannot match that pace, particularly if the rumor is based on their own unpublished work. The result may be that AI labs routinely overtake original research projects simply because they happen to have better ears and bigger machines.
Tao’s warning extends beyond the immediate dispute. It points to an emerging asymmetry in science: the groups creating the most powerful models are also in the best position to hear about early results, integrate them into training runs, and publish first. Academic incentives are built around priority and openness, but those incentives become meaningless if a rival can absorb undisclosed data without consent.
What This Means for Open Science
The Navier-Stokes controversy is therefore not just a dispute between two individuals and one lab. It raises fundamental questions about the relationship between AI development and open research.
- Can researchers safely use AI coding platforms while keeping their unpublished ideas out of a company’s training data? Most users have no way to verify that such protections are effective.
- Should AI labs be allowed to redirect their research efforts based on confidential or rumored work from another organization, especially when the rumor originates inside their own tools?
- When a co-author is employed by a competing lab, should that fact influence authorship? Buckmaster says OpenAI attempted to remove Alpöge from the paper for that reason, while the company denies misconduct but has not clarified its authorship policy.
- Who should have access to training data and model weights when plagiarism is alleged? Current dispute resolution mechanisms provide no clear path for outside auditors.
- What happens to the field’s willingness to share early results if junior researchers fear their work will be mined by a more powerful actor?
These are not abstract concerns. The dispute has already damaged trust in the idea that AI companies can serve as neutral scientific partners. Even researchers who have no interest in Navier-Stokes might reasonably hesitate before uploading a promising lemma or an unpublished method to a platform operated by an AI lab.
Trust Cannot Be Assumed
OpenAI and its defenders may be right that no misuse occurred. The available evidence is incomplete, and the possibility of an innocent coincidence is real. But the episode has raised the bar for what responsible AI research looks like. A lab that develops tools for mathematicians cannot rely on a vague promise not to use their work. It must demonstrate, through transparent data-handling policies and independent audit mechanisms, that user contributions will not appear in future models without consent.
The broader lesson may be that scientists and AI labs operate under different logics. Academics share to build knowledge; companies share to build advantage. When the two systems collide, the weaker side is usually the researcher who is outside the corporate walls. The Navier-Stokes dispute will be remembered not only for the mathematical achievement that triggered it, but for what it revealed about the vulnerability of open science in the age of generative AI. Unless clear norms are established soon, the next millennium problem may be solved not by a lone mathematician on a chalkboard, but by a corporate team that saw the answer coming before the question was ever published.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
