All the drama around AI’s takeover of mathematics

Over a few months, OpenAI went from announcing solutions to 10 long-standing problems in August to claiming, in September, a solution to the Navier-Stokes problem, one of the seven Millennium Prize Problems, using an internal model more powerful than GPT-6 Astra running alongside 10,000 concurrent agents. Anthropic and other labs have announced their own results too.

By any normal standard, that’s historic. Instead, the story has become one of backlash: allegations of scooping rival researchers, questions about whether models drew on mathematicians’ unpublished work, and a hastily formed advisory panel that mathematicians describe as confusing. And OpenAI says scores more results from its unreleased model are on the way.

We want to hold two things at once, because both are true. First, this is some of the most impressive evidence yet that AI can do genuinely new intellectual work. Combining known methods in new ways, linking distant fields, resurfacing buried ideas from the literature: that is a large part of how mathematics advances, and machines are now doing it on problems that stumped people for decades. We think that’s worth celebrating, and we’d push back on anyone who dismisses it as autocomplete.

Second, the way OpenAI has gone about it is doing real damage. The Verge’s reporting describes a company that, after hearing other researchers were making progress on Navier-Stokes, appears to have thrown its resources at a last-minute effort to get there first. Mathematicians have raised allegations of scooping and spying and say long-standing academic norms were violated. One professor at Queen Mary University of London put it as the “kind of things that mathematicians will generally not do.” Separately, a mathematician whose area underpinned one of the August results has publicly questioned whether his and colleagues’ prior chatbot interactions contributed, and OpenAI acknowledged that result built heavily on his earlier work.

The framing in The Verge’s coverage lands for us: mathematicians want to advance the field, and OpenAI wants to win. Those goals overlap until credit, data provenance and timing are at stake, and then they collide. When a lab with near-unlimited compute races individual researchers to the finish on problems they’ve spent careers on, the chilling effect is predictable. Why share partial progress, or even talk through ideas with a chatbot, if it might feed a competitor that can outrun you?

The advisory panel announced in September is the right instinct executed badly, by mathematicians’ own accounts. Researchers called it a reasonable first step but were left asking what it will actually do, how much influence it will have, and whether such a small group can speak for the wider community. Its first job, helping coordinate the release of many more results, is exactly the scenario researchers dread.

Where this leads if it scales: proofs arriving faster than humans can check them, careers built on long problems disrupted overnight, and trust between AI labs and researchers eroding just when verification by people matters most. The fix isn’t to slow the math. It’s for labs to disclose what data and prior work their systems used, coordinate releases with the people doing verification, and give credit like members of the field rather than conquerors of it.

The Millennium Prize Problems are seven famous open problems, each carrying a $1 million reward. OpenAI says it began training the internal model behind the Navier-Stokes claim on August 28. The broader disquiet predates that result; an Oxford Fields Medalist told The Verge in August he had spent the year soul-searching about the future of the field.