OpenAI's 722 Math Papers: AI Proof or Marketing?

The Paper Drop That Isn’t Just About Paper

OpenAI dumped 722 math papers onto GitHub on October 6, most bundled with formal proofs written in Lean, the theorem-proving language that has become the gold standard for machine-verifiable mathematics. The headline claim—that the company’s frontier model made progress on problems related to the Landau-Siegel zero conjecture—is precisely the kind of result that would reshape analytic number theory if it held up.

The catch, as usual with OpenAI’s math claims, is the word “if.”

Why Landau-Siegel Matters

The Landau-Siegel zero conjecture sits at the intersection of two of the deepest questions in number theory: the distribution of prime numbers and the behavior of L-functions. Roughly, it asks whether certain L-functions can have a real zero unusually close to s = 1—a so-called exceptional zero. If such a zero exists, it would have cascading consequences for class number formulas, primality testing, and the broader architecture of analytic number theory that underpins modern cryptography.

No one has proved the conjecture. No one has even ruled out the exceptional zero conclusively. The fact that OpenAI claims its model produced results “related to” this problem is notable precisely because it signals the model touched a region of mathematics where a single wrong step can derail an entire argument—and where the barrier to entry is genuinely high.

The Process Behind the Papers

According to OpenAI, the 722 papers came from evaluating approximately 4,000 unsolved problems against an unpublished model. The filtering process narrowed those outputs down to results deemed sufficiently important. Each paper consumed, on average, computational effort equivalent to three hours of ChatGPT Pro’s thinking mode. Two results—the one on zeta function zeros and another on the Hodge conjecture for CM abelian varieties—are explicitly called out as exceptions to the standard procedure.

That transparency is actually unusual. Most AI-generated research outputs don’t come with this level of procedural disclosure. OpenAI also published reasoning traces for 10 of the papers, covering topics from the irrationality degree of pi tokaplansky’s direct finiteness conjecture in characteristic 2 to self-magnetization in quantum Heisenberg ferromagnets.

But here’s the thing about reasoning traces: they show you the path the model took, not whether the path was correct. And in pure mathematics, a plausible-looking derivation can hide a subtle error that only a human expert would catch.

The Lean Question

OpenAI says many but not all of the 722 papers include Lean formalizations. It acknowledges that unformalized results may contain errors and promises to fix them if found. This is both honest and strategically important. Lean proofs are machine-checkable; without them, a result is only as credible as the peer review it can attract—and right now, the peer review pipeline for AI-generated math is basically non-existent.

The distinction matters because the mathematics community has spent the last decade building Lean as a verification tool precisely to eliminate ambiguity. An AI-generated proof that hasn’t been formalized is closer to a very confident conjecture than to a theorem.

AGMAI: Advisory, Not Endorsement

The Advisory Group on Mathematics and Artificial Intelligence (AGMAI), formed in September, provided guidance on OpenAI’s disclosure process. Their September 29 recommendations were thorough: register in externally managed repositories, disclose model names and prompts, publish reasoning summaries and compute costs, formalize with Lean, explain selection methodology and failed attempts.

Then AGMAI immediately walked most of that back. In its response to OpenAI’s latest release, the group stressed that its advice was not an endorsement of the results, that it does not represent the mathematical community, and that evaluation belongs to mathematicians—not to AI companies or their advisory boards.

That caveat is significant. It means the very framework OpenAI built to lend credibility to this release simultaneously distances itself from the substance. It’s the mathematical equivalent of “we facilitated the conversation, we didn’t write the lines.”

The Pattern Is the Story

This isn’t OpenAI’s first rodeo with bold math claims. In September, the company announced its internal model had solved the Navier-Stokes existence and smoothness problem—one of the seven Millennium Prize Problems. Working mathematicians who followed the claim found it problematic, and the controversy forced a reckoning. Around the same time, OpenAI said its internal model had solved over 100 unsolved problems. In August, it reported new results on 10 problems using an internal version of its next flagship model, codenamed Astra.

And in October 2025, a company executive’s post claiming GPT-5 had solved an Erdős problem was deleted after backlash.

The pattern is clear: OpenAI repeatedly announces breakthroughs that the mathematical community treats with deep skepticism. Each time, the company adds more transparency machinery—advisory groups, Lean formalization, GitHub repositories—but the core tension remains unresolved. AI can generate mathematical output at scale. Human mathematicians can verify it at speed. Those two rates are not currently aligned.

Who Wins, Who Loses

If even a fraction of these 722 papers survives peer review, the winners are analytic number theorists who suddenly have a new tool for exploring conjectural territory. The Landau-Siegel conjecture, the Hodge conjecture for CM abelian varieties, and several others on the list are legitimate open problems where any progress is valuable.

The losers are the mathematicians who now have to sort signal from noise in a flood of AI-generated results. The field didn’t ask for this deluge, and the peer review infrastructure doesn’t exist to handle it at scale.

OpenAI wins regardless. Whether the papers hold up or not, the company has established a new template for how AI companies announce scientific成果—transparency theater that looks impressive from the outside and generates exactly the kind of headlines this release is already getting.

What Happens Next

OpenAI says it will fund workshops and special programs to help mathematicians understand AI-generated results, and that it’s working toward responsible public release of the model behind these papers. Both are reasonable steps. But the real test is simpler: do independent mathematicians verify the Lean proofs? Do the unformalized results survive scrutiny?

The Landau-Siegel conjecture won’t be solved by a GitHub dump. It will be solved by a chain of reasoning that other mathematicians can check, question, and build on. AI can help with that process. But right now, OpenAI’s 722 papers are more like a very large set of research proposals than a set of results.

The mathematics community will decide which. That’s the whole point of AGMAI’s disclaimer.