Researchers who share draft work with AI companies now face a new risk: being outpaced by their own findings. The public dispute over OpenAI’s claimed AI-generated result on the Navier-Stokes equations, one of the Clay Millennium Problems with a $1 million prize, has put pressure on the norms that have governed open mathematics for centuries.
Mathematicians Tristan Buckmaster and Levent Alpöge say they uploaded drafts of their work to OpenAI’s Codex systems, then watched OpenAI announce a related result soon after. OpenAI’s CEO and a lead researcher deny using those drafts, but the company’s own blog post acknowledges it cannot fully rule it out.
What happened with the Navier-Stokes problem?
Buckmaster and Alpöge had been working on the Navier-Stokes problem, one of seven Clay Millennium Problems each carrying a $1 million prize. At some point, they entered portions of their approach into OpenAI’s products. The two researchers later accused OpenAI of using that input as training data, pressuring Buckmaster, and trying to remove Alpöge from any resulting paper because Alpöge is employed by Anthropic, a direct competitor to OpenAI.
Buckmaster described the situation as “absolute academic malpractice.”
What does OpenAI say?
OpenAI employee Sébastien Bubeck, who led the internal effort, denied the plagiarism allegation. Another OpenAI employee, Boaz Barak, argued the model did not need outside help. “It’s just cope to think that the model would have needed this. It actually started off by proving a stronger claim than they did,” Barak wrote. “Anyone who has seen this model at work would not think it needs ‘hints.'”
OpenAI’s official blog post conceded the gap in certainty, writing: “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.”
OpenAI CEO Sam Altman backed Bubeck publicly on X, saying the team had “acted with integrity and generosity throughout.” Altman also confirmed that the effort began after “rumors on the internet last week that Anthropic’s models had solved a millennium problem and we were curious if ours could do it too.” OpenAI has suggested that Buckmaster and Alpöge should receive credit for the prize.
Where do the two sides disagree?
Buckmaster and Alpöge dispute OpenAI’s claim that its solution differs significantly from theirs. They say they had entered a similar approach into OpenAI’s systems, and as far as Buckmaster understands, that input ended up in the training data.
OpenAI employees counter that the chance of training on the submitted solutions is low, especially if the two researchers had disabled the option to allow their inputs to be used for training. Whether Buckmaster and Alpöge disabled that option is not publicly known.
On the authorship question, Alpöge directly contradicted Altman. Altman wrote that it was difficult to extend the same authorship offer to Alpöge because he “was not willing to talk or coordinate with us anyway.” Alpöge responded that he would have liked to work with OpenAI and that the authorship question did not matter to him. “I also like the idea of the labs cooperating, and even better on scientific progress. It’s a shame!” Alpöge wrote.
What does this mean for open science?
Mathematician Terence Tao, one of the most influential living mathematicians, warned on Mastodon that “even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential.”
Tao added: “The incentives may now be pointing in the direction of no longer sharing any promising research directions with the broader community, which would reverse centuries of traditions of open science and do serious long-term damage to the future of the field.”
For academics and companies, the practical lesson is straightforward. Anyone who feeds research data into AI lab systems risks being beaten by their own findings. Opting out of data training through product settings offers thin protection at best, especially since AI labs have not earned the benefit of the doubt on training and data practices.
Why does this case matter beyond mathematics?
The incident shows how a single rumor of progress can now trigger an industrial-scale response from a well-funded AI lab. OpenAI heard that Anthropic’s models had potentially solved a Millennium Problem, then pointed its own massive resources at the same problem and produced a competing result within days. The question of whether inputs submitted to AI products can be used to train future models has moved from a policy footnote to a frontline concern for every researcher who experiments with AI tools.
FAQ
What is the OpenAI Millennium Proof dispute about?
Researchers Tristan Buckmaster and Levent Alpöge accused OpenAI of using drafts they uploaded to OpenAI’s systems as training data to produce a competing result on the Navier-Stokes problem, one of the Clay Millennium Problems. OpenAI denies plagiarism but says it cannot rule out that de-identified data from those uploads improved its models.
Did OpenAI train its models on the researchers’ drafts?
OpenAI’s official statement says it is unlikely but cannot be ruled out. The researchers believe their input ended up in training data. OpenAI employees have argued the model did not need outside help to produce its result.
Why does Terence Tao think this is dangerous for open science?
Tao warned that AI labs can now mobilize large resources based on rumors of progress, potentially overtaking original research before it is published. He said this could push researchers to stop sharing promising directions openly, reversing centuries of scientific tradition.
This article summarizes reporting from the-decoder.com. See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.