OpenAI’s coming math release faces a test of trust

A release without a date

People familiar with OpenAI’s plans told WIRED the company intends to post the results on GitHub on Tuesday. The work was discussed at an August meeting, where mathematicians urged OpenAI to publish scientific papers rather than blog posts or social-media announcements, as it had done with ten problems earlier that month. Brina Kra, a mathematician at Northwestern University, says the group wanted work that other researchers could study and build on. She says that advice appears to have gone unheeded.

“On August 28, we began training a new internal model. In addition to solving the Navier–Stokes problem from the Millennium Prize Problems list, it has solved more than 100 longstanding open problems across most areas of mathematics,” says McCallum. OpenAI is preparing further mathematical results from the model for responsible release, she says, and is taking into account advice and public recommendations from the Institute for Advanced Study’s Advisory Group on Mathematics and AI. The company has not set a publication date.

The planned GitHub release is one more addition to the tens of thousands of AI-generated mathematical solutions produced this year. But several prominent mathematicians see it less as a change in practice than as another sign that OpenAI has not learned much from earlier disputes.

Researchers told WIRED that mathematics has become a venue for OpenAI and Anthropic to demonstrate their models’ capabilities as both companies prepare for major IPOs. In trying to outdo each other, mathematicians say, the labs are neglecting established norms around scientific publication and authorship.

The authorship dispute behind the unease

The most prominent conflict came in September. After hearing that other researchers were close to solving a million-dollar Millennium Prize problem, OpenAI deployed thousands of AI agents to work on it. Tristan Buckmaster, a mathematician and professor at New York University, accused the company of getting ahead of his work on part of the problem. He had been working on it with Levent Alpöge, an Anthropic employee, as a personal collaboration. They had not published their result, but had used OpenAI tools.

Buckmaster says OpenAI researcher Sebastian Bubeck appeared to suggest leaving Alpöge off a paper to avoid complicating the situation. Bubeck referred WIRED to an earlier public statement in which he denied asking for Alpöge to be excluded as a co-author. Buckmaster said he rejected the suggestion.

Notes from a meeting seen by WIRED say Bubeck raised concerns about Anthropic’s activity. They attribute to him the remark that Levent was probably contacting company leadership to report that “fucking OpenAI can solve a Millennium problem,” and the question of what would stop Anthropic from putting all its computing resources toward solving one.

The same notes say that when Buckmaster threatened to tell the press he believed OpenAI had taken his work, Bubeck appeared to warn that doing so could destroy the mathematician’s career. “If you don’t want me to be nice, I can not be nice,” the notes quote him as saying.

Nestor Guillen, a visiting professor of mathematics at New York University, told WIRED that mathematicians had formed an “impression of mafia-like behavior” by AI companies. “We disagree with that characterization,” McCallum says.

“I feel deeply anxious, and I see that anxiety increasingly among my fellow mathematicians. But it’s not about AI; it’s about the companies developing it,” Guillen says. “I think a lot of people are worried about power being concentrated in one place.”

OpenAI formed an advisory group in mid-September to help set standards for evaluating and presenting new results, follow academic and professional norms, and build tools for mathematical research and education. Mathematicians say they have yet to see meaningful progress.

Their objection is practical as well as procedural: blog posts are harder to verify than papers, and earlier contributions by mathematicians can disappear from the story. The problem is not confined to labs. Shortly after the World Championship final, Alpöge announced on social media that he had disproved an 87-year-old conjecture.

“Publishing mathematics on social media and in press releases does not help sustain the scientific environment on whose fertile ground these models were trained,” Kra says.

Tools are available; trust is not

This year, as machine-assisted proofs have multiplied, specialists have built tools to help the community use new results and decide which deserve further checking and interpretation. Hexagon is a repository mostly for AI-generated material; Palomar is a registry of machine-verified mathematics.

Kra says OpenAI was directly urged to use these tools, but its behavior has not changed. She also does not think the company’s actions align with the Leiden Declaration, a call signed by more than 4,000 mathematicians urging AI companies to follow the standards of the mathematical community. Kra helped prepare the declaration.

People who have spoken with OpenAI employees say some believe the company’s technology has already made mathematics unnecessary. “We don’t believe the future of mathematics is predetermined,” McCallum says. “We’re working with the mathematical community to figure out together what it will look like.”

At the August meeting, the prospect of AI surpassing human researchers was discussed as a hypothetical. But attendees who spoke to WIRED took it as a warning about what lies ahead. One compared the way the company told mathematicians about its capabilities and breakthroughs to police warning a family before telling them a relative had died in a car crash.

Someone familiar with Bubeck’s views says he believes his company’s technology will soon end the careers of most mathematicians. Bubeck told WIRED that more capable AI would help scientists tackle more ambitious questions, better connect research to practical problems and make mathematics more accessible. He sees it as a way to expand the range of problems mathematicians can solve and increase the impact of their work.

Kra is glad to have a powerful new tool for solving problems, but wants companies to disclose more about their results so researchers can trust them and build on them. “This will change how we work, but I think right now we can think bigger,” she says. “It’s an anxious time,” she adds. “But it’s also genuinely exciting.”

My read is that the technical headline is arriving faster than the field’s ability to assess it. OpenAI can release hundreds of results on GitHub, but without papers, clear attribution and visible uptake of the community’s guidance, the repository risks becoming a record of claims rather than a body of work other mathematicians can reliably extend.