OpenAI's 700-Paper Math Dump Ignites a Fields Medalist Backlash

OpenAI released more than 700 mathematics "preprints" — preliminary research papers — on GitHub this week, escalating a conflict with the academic math community that has already drawn an open letter signed by 28 Fields Medal winners.

The letter, hosted on a website called Math and AI, accuses AI companies of pursuing goals that are "severely misaligned" with those of mathematicians. The Fields Medal, awarded only every four years to three or four mathematicians, is widely considered the Nobel Prize of math. The scale of the signatory list signals how deep the rift now runs between frontier AI labs and the discipline whose accumulated knowledge trained their models.

Harvard professor Curtis McMullen, one of the Fields Medalists who signed the declaration, told CNET in an email that AI companies are not coordinating with math experts or deploying their models in ways that could practically benefit society.

"The rollout of mathematical work by frontier AI companies has been carried out with disregard for the profound disruption to our community and its academic ecosystem," McMullen said. "Its goal appears to be to increase the valuation of these companies at any cost."

OpenAI did not immediately return a request for comment.

Why are mathematicians angry now?

The GitHub drop landed roughly a month after OpenAI announced it had solved the Navier-Stokes equation — a problem so difficult it carried a $1 million bounty. According to Scientific American, that earlier announcement left the field "already in shock." The 700-preprint release has now piled on.

The frustration predates both events. Early large language models, trained largely on written language and lacking a grasp of ground truth, were notorious for mathematical errors — misplacing decimals, hallucinating incorrect answers. But the models improved at startling speed, going from self-described math dunces to problem-solvers capable of attacking longstanding open questions.

Mathematicians' objections, however, are not primarily about job security. They focus on two structural problems:

• A lack of transparency about how AI systems achieve their mathematical results.
• A lack of engagement with why specific problems matter to mathematics in the first place.

OpenAI says it has formed an advisory group hosted at the Institute for Advanced Study, but the company has not detailed how the group will operate or whether it will share its most advanced models. OpenAI has called the initiative only "a first step."

What did McMullen actually say?

McMullen acknowledged in his email that AI's potential as a tool for math research is "undeniable." His complaint is about method, not capability.

"In particular, the models used to achieve these results should be shared with the scientists who understand and formulated the conjectures or questions under investigation," McMullen said.

He also pushed back on the framing of mathematics as a benchmark to be cleared. "Mathematics is not a game of chess," he said. "Its primary aim is to develop human understanding of abstract structures. Indeed, it is the breadth and versatility of this kind of understanding that underpins technological progress."

That last point carries an irony the math community is quick to note: the AI models now solving these problems were themselves built by ingesting vast stores of human knowledge, including mathematics.

Can anyone verify the results?

The GitHub publication has left many researchers without enough information to evaluate how many of the problems in the new batch were actually solved. Gary Marcus, a cognitive scientist and professor emeritus at New York University, wrote on Substack that the publication "would never pass peer review."

"We don't know what the procedure was … From the initial report we can tell almost nothing," Marcus wrote.

That verification gap sits at the center of the dispute. OpenAI has said its internal large language models are solving an unprecedented number of longstanding mathematics problems with the goal to "push the frontier of human knowledge." But without shared methods, shared models, or peer review, the mathematical community has no standard mechanism to check the claims.

Is the entire math world opposed?

No. Levent Alpöge, a mathematician working at Anthropic, publicly congratulated the OpenAI team on X.

"There are some sad stories related to their users getting scooped … we should put that aside for today though. It's obviously the most significant moment in mathematical history," Alpöge wrote.

His framing — significant, but not without casualties — captures the tension. Some researchers report having work preempted by AI systems. Others see a genuine advance in mathematical capability that traditional institutions have not yet learned to absorb.

What happens next?

OpenAI's Institute for Advanced Study advisory group is the only bridge currently on the table, and the company's own description of it as "a first step" leaves open whether its most capable models will ever be shared with the mathematicians whose conjectures they target. McMullen's demand is specific: give the scientists who formulated these questions access to the models that answered them. Whether AI-generated proofs advance human understanding of abstract structures or simply produce unverified output at scale is the question now hanging over both the 700 preprints and every release that follows them.