OpenAI Posts 372 AI-Generated Math Proofs on GitHub, Challenging Academia to Catch Up

Meta description: OpenAI releases 372 AI-generated math results on GitHub, using Lean formal verification to ease review as the math community debates scale and rigor.

Tags: OpenAI, AI mathematics, formal verification, Lean, research infrastructure


OpenAI has released 372 AI-generated mathematical results produced by an internal frontier model, publishing them on GitHub instead of in academic journals. The release is important because it tests whether machine-checked formal verification, rather than traditional peer review, can keep pace with a large batch of potentially new mathematical claims.

Background: The Review Bottleneck Before the Release

Before this release, AI-generated mathematics faced a practical obstacle: human review. A proof can be logically sound but still difficult for another mathematician to read quickly, especially when many claims arrive at once. Formal verification addresses that by encoding a proof in a machine-checkable language, allowing software to confirm the logical steps.

OpenAI chose GitHub, a code hosting platform, rather than peer-reviewed journals. That choice signals that the traditional publication pipeline may be too slow for a volume of AI-generated results that could quickly exceed the math community's capacity for manual checking.

Development: From Single Prompts to Lean Formalization

OpenAI says the collection contains 372 new mathematical results, each intended to solve an open problem or make substantial progress toward one. The GitHub repository includes revision logs and citations, and the results cover improvements to major computer algorithms and advances related to the Riemann hypothesis.

A key detail is how the results were produced. Nearly every result came from a single prompt to a single AI agent, although some took multiple attempts. On average, each result consumed roughly three hours' worth of ChatGPT Pro Thinking compute.

That is very different from the earlier Navier-Stokes problem solution attributed to the same internal frontier model. The Navier-Stokes solution required a swarm of 10,000 agents and millions of dollars in compute, and it has been under formal review for weeks.

Many of the newly published proofs come with formalizations in Lean, a programming language built for machine-checkable mathematical proofs. More formalizations are planned. OpenAI says the sheer volume of AI-generated results could overwhelm manual review, so machine checking is meant to make the review process more practical.

The company also published methodology details, including summaries of the reasoning process, statistics on how many problems the model attempted, and estimates of compute costs. It did not publish prompts and shared average compute costs rather than per-problem figures.

Latest Progress: Advisory Groups, Workshops and Community Pushback

OpenAI consulted with the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study and loosely followed its public recommendations. The group includes Fields Medal winner Timothy Gowers and other renowned mathematicians. The company set a boundary before that consultation: mathematicians may advise on how results are communicated, but not on whether or how fast they are produced.

OpenAI also announced plans to fund workshops and conferences focused on understanding AI-produced results. It acknowledged the need to improve citation quality and presentation, and said it is working on a responsible release of the model to "directly empower scientists with state-of-the-art capabilities."

The reaction shows why this release is contested. Lean can verify logical correctness, but it cannot judge whether a result is mathematically relevant or original. Reactions range from excitement to frustration.

In an open letter titled "A Severe Misalignment of AI in Mathematics," 25 Fields Medal winners warned of a deep disconnect between the AI industry's goals and those of mathematics. They wrote that problem-solving is merely a tool and proxy for the real goal of conceptual understanding and insight. They argued that mass-producing true statements could destroy fertile ground rather than bring new ideas to life, and said the effect would spill over into other fields.

Gowers warned that within one to two decades, mathematical literature could grow enormously while no human community remains that truly understands it. Terence Tao, another Fields Medal winner, said training young mathematicians needs to emphasize the human side and tightly limit AI tool use so that genuine learning and understanding survive.

The release is therefore best understood as an early stage of AI-assisted mathematics: a large batch of claims, partial formal checking, limited public prompts, and an ongoing debate over whether machine-generated results create new knowledge or mainly create review pressure.