OpenAI’s Math Push Is Stirring Alarm Among Mathematicians Over Credit, Process and Power

OpenAI is preparing to publish a large batch of AI-generated mathematics results, and the move is deepening an already tense relationship with academic mathematicians who say the company is racing ahead of accepted scientific norms. The planned release, expected on GitHub, follows claims that OpenAI’s internal models have solved more than 100 open problems and has raised fresh concerns about credit, verification and the growing influence of AI labs over a field that depends on rigorous peer review.

The backlash matters because the dispute is no longer just about whether AI can prove theorems faster than humans. It is now about who gets credit for discoveries, how results are validated, and whether the world’s leading AI companies are setting the rules for mathematics on their own terms.

In August, OpenAI gathered roughly 40 mathematicians to ask a difficult question: what happens if AI outperforms people in advanced math? Participants say the company suggested its systems had already cracked hundreds of longstanding problems, but they were also told the findings would not be released all at once. Instead, OpenAI sought advice on how to disclose the work without triggering a backlash from the mathematical community.

That conversation was supposed to open a more careful channel between the company and the academics who would be affected by its progress. Instead, several mathematicians now say they see a pattern of rushed announcements, blog posts, tweets and private negotiations that leave the field struggling to understand — or trust — what OpenAI and its competitors are actually doing.

What OpenAI is planning to release

OpenAI is preparing to publish hundreds of mathematical results tied to an internal model that it says has solved more than 100 open problems across different areas of mathematics, according to people familiar with the plan and a company statement. The results are expected to appear on GitHub, though the company has not announced a firm release time.

The company says it is trying to handle the disclosure responsibly and has been taking advice from the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. But many mathematicians say that even if the content is impressive, the way it is shared could undermine its usefulness.

OpenAI says it is working to release the next wave of results responsibly and in line with advice from outside experts, while also making clear that it has not locked in an exact publication time.

The tension has been building for months as frontier AI systems become more capable at generating proofs, checking steps, and exploring large mathematical spaces. What once sounded speculative is now becoming routine enough that labs are treating it as a race — and that is exactly what worries many researchers.

Why mathematicians are upset about the release strategy

Mathematicians argue that the issue is not simply whether the results are correct. It is whether they can be independently checked, contextualized and built upon. In their view, publishing breakthroughs through short-form online posts or software repositories may be efficient, but it does not meet the standards of a discipline built on detailed exposition and verification.

Bryna Kra, a mathematician at Northwestern University, said the August meeting offered a promising start. But she and others expected OpenAI to follow a more formal path than the company’s earlier announcement of 10 solved problems, which was shared via a blog post and social media rather than through research papers.

According to Kra, the community asked for papers that would explain the work, show the methods and allow other mathematicians to evaluate the claims. She says that advice does not appear to have shaped the upcoming release.

Kra says the math community wanted thorough papers, not a quick burst of online announcements, because the field needs time to absorb, verify and reuse new results.

Her criticism cuts to a broader fear: that AI companies are treating mathematics as a showcase for model performance rather than a scientific discipline with its own norms. In that framing, the main question becomes not whether a theorem is true, but whether the lab can claim it first.

How do mathematicians think the process should work?

Mathematicians say new results should be documented like traditional research: with clear definitions, careful proofs, peer review where possible, and enough background for others to reproduce or build on the work. They also want proper attribution, especially when human researchers contributed ideas, tested approaches or had previous work overlooked in a public announcement.

That is why the mode of release matters so much. A theorem posted in a blog entry can attract attention, but a theorem embedded in a paper helps the field judge whether it is genuinely new, where it fits in the literature, and whether it can be extended.

The credit fight that escalated concerns

The biggest flashpoint so far came in September, when OpenAI reportedly deployed thousands of agents to work on a famous Millennium Prize problem after hearing rumors that others were close to a solution. The move triggered accusations that the company had moved too aggressively into work already underway by human mathematicians.

One of the sharpest complaints came from Tristan Buckmaster, a mathematician at New York University, who said OpenAI had front-run his effort on part of the problem. Buckmaster had been working with Anthropic employee Levent Alpöge in a private collaboration, and the pair had not yet published their work. According to Buckmaster, OpenAI’s actions appeared to jump ahead of that research.

Negotiations over credit did not go smoothly. Buckmaster says OpenAI researcher Sébastien Bubeck appeared to suggest that Alpöge should not be listed as an author, which Buckmaster rejected. Bubeck has denied asking that Alpöge be excluded, pointing to an earlier public statement.

Meeting notes reviewed by WIRED also suggest the discussion grew heated. In those notes, Bubeck reportedly expressed concern about what Anthropic might be doing and questioned what would stop the rival company from dedicating all of its computing resources to win a Millennium Prize problem of its own. He also allegedly warned Buckmaster that going to the press could damage his career. Bubeck disputes the characterization of the exchange.

Buckmaster says the episode looked like a fight over scientific priority; OpenAI says the company’s account of the conversation is different.

For mathematicians already uneasy about how AI labs operate, the episode reinforced a sense that the traditional safeguards around scholarly work can be brushed aside when companies are under pressure to win a public race.

What is driving the race between OpenAI and Anthropic?

The competition between OpenAI and Anthropic is intensifying because both firms are preparing for large-scale corporate milestones, including potential blockbuster initial public offerings. That creates a strong incentive to publicize visible breakthroughs, especially in a field as prestigious as mathematics.

Researchers who spoke with WIRED say that competitive pressure is now shaping how results are communicated. Instead of slow, paper-based disclosure, the companies are increasingly using blog posts, tweets and software repositories to announce progress in near real time.

To many academics, that is the problem. The labs may see themselves as accelerating science. Mathematicians worry they are turning scholarly work into product marketing.

How are mathematicians responding?

Mathematicians are not rejecting AI outright. Many say they welcome powerful new tools, especially if those systems can help explore difficult areas faster or reveal connections that humans might miss. What they reject is the sense that the companies are setting the terms without respecting the norms of the field.

Some researchers have responded by building infrastructure designed to handle machine-assisted mathematics more responsibly. Two examples are Hexagon, a repository focused on AI-generated material, and Palomar, a registry for machine-verified mathematics. Both are meant to help scholars track what has been produced, what has been checked, and what still needs human attention.

Those tools reflect a practical recognition: if AI is going to generate more proofs, the field will need new systems for cataloging them. But the very need for those systems is also a sign of how quickly the landscape is changing.

What is the Leiden declaration and why does it matter?

The Leiden declaration is an effort backed by more than 4,000 mathematicians to persuade AI companies to follow standards that better align with academic practice. Kra helped write it, and she says OpenAI’s behavior has not matched the declaration’s spirit.

At the core of the complaint is a simple idea: if AI companies want to use mathematics as a proving ground for their systems, they should also accept the responsibilities that come with the field’s long-standing culture of openness, attribution and verification.

How OpenAI describes its own role

OpenAI says it does not share the darker interpretation that some mathematicians have attached to its work. The company rejects the suggestion that it is trying to displace mathematicians or ignore their concerns, and it says its goal is to collaborate with the community while developing more powerful tools.

In its statement, OpenAI said it does not believe the future of mathematics is predetermined. Instead, it says it wants to work alongside mathematicians to navigate the changes ahead.

OpenAI says it sees the technology as a partner to mathematics, not a replacement for the field, and argues that better models could help researchers tackle bigger questions and make the discipline more accessible.

Sébastien Bubeck has also pushed back on the notion that the technology means the end of human math research. He says more capable AI could help mathematicians ask more ambitious questions, connect theoretical work to real-world problems and broaden access to the subject.

That optimism is not universally shared. Some mathematicians say that even if AI does expand what can be explored, the labor of mathematics — especially the judgment involved in deciding what matters, what is true and what should be trusted — will remain human for the foreseeable future.

What this means for the future of mathematical research

The bigger issue is not whether AI can solve a handful of famous problems. It is whether its growing role will alter the culture of mathematics itself. If labs can generate large numbers of proofs, then the field may need to rethink how it records authorship, validates claims and teaches younger researchers.

That change may be unavoidable. But many in the discipline want it to happen through consultation, not disruption. They fear that if companies rush ahead, trust in the results will erode, and the very people best equipped to understand the findings will be left out of the process.

There is also a generational issue. Younger mathematicians may grow up working with AI tools as naturally as earlier generations used computer algebra systems. But if the first wave of AI-generated results is handled poorly, the field could begin that transition under a cloud of mistrust.

For now, mathematicians are left in an uneasy middle ground: excited by what these systems can do, frustrated by how they are being presented, and unsure whether the companies leading the charge understand the difference between speed and scientific legitimacy.

What comes next

If OpenAI follows through on the planned GitHub release, the reaction will likely hinge on two questions. First, are the results genuinely important and well documented? Second, does the company present them in a way that allows mathematicians to verify, cite and extend them?

That could determine whether the release is seen as a milestone or another misstep. For a field that relies on trust as much as talent, the method of disclosure may matter almost as much as the mathematics itself.

Kra summed up the mood as both anxious and hopeful. She wants the tools to exist. She wants them to help. But she also wants the companies behind them to respect the system that made those breakthroughs possible in the first place.

As she put it, the moment is unsettling, but it is also full of possibility — if AI labs and mathematicians can find a way to work together without turning discovery into a branding exercise.