OpenAI has given mathematicians hundreds of new results to examine, but not a shortcut to knowing which ones hold up. The company’s release spans 719 manuscripts, includes claims touching famous open problems, and varies widely in how far its results have been checked. Researchers say understanding the collection could take years, while its effects on academic careers are already being felt.
A release too large to assess quickly
The manuscripts cover fields including number theory, geometry, combinatorics, algebra, topology, probability, mathematical physics and theoretical computer science. OpenAI said its model tried more than 4,000 problems. A typical result used about three hours of ChatGPT Pro thinking compute.
The scale has made even an initial review difficult. Researchers said reading the collection’s roughly 40-page contents and abstracts took substantial time, before they could begin checking individual arguments or placing them in context. OpenAI also published guidance for finding material in the large GitHub repository, a practical admission that a research release can become its own filing problem.
Some mathematicians said their first impressions were better than they had expected, particularly given criticism of earlier OpenAI mathematical papers. That is encouraging, though the comparison point was not especially demanding. Researchers still described papers that were hard to follow, unusually brief, or thin on citations.
Formal proofs are useful, not automatic
Some manuscripts include Lean formalizations. Lean is a proof assistant that can check whether a formal argument follows from its stated premises. Such code can give mathematicians confidence in a result, even when they have not independently followed every step of the reasoning.
But the collection is far from uniformly formalized. OpenAI said 300 top-line results among the 719 manuscripts had been formalized, about 42 percent, and said it would add more. The company also acknowledged that the manuscripts were at different stages of verification.
A formalization is not a magic stamp. Researchers still need to check that the code proves the same claim made in the paper. Some mathematicians who examined the release said that the relationship between formal proofs and the accompanying manuscript claims was not always clear, and that the quality varied.
The repository’s own record showed corrections on more than a dozen manuscripts by October 8. It also listed three papers withdrawn after a sign error was found to invalidate an argument. That record is more transparent than silently changing a paper, but it also shows why readers cannot treat a large batch of results as a finished catalogue of established mathematics.
Major claims, still awaiting scrutiny
Researchers pointed to several results they considered potentially significant, including progress related to the Riemann hypothesis, a special case of the Hodge conjecture, and a solution to the four-dimensional Kakeya conjecture. The Riemann and Hodge conjectures are among the seven Millennium Prize problems. These are not obscure exercises, but their presence in a release does not by itself settle their status. Experts stressed that the work needs careful verification and explanation.
Some mathematicians saw dozens of results that could prove exceptionally important if they withstand scrutiny. Others said that the papers they had looked at were difficult to evaluate, or that work in their own areas appeared to revisit existing research. Those assessments remain preliminary, given the number of manuscripts and the limited time available to review them.
Researchers also noted that many of the results appear to use established methods in ingenious ways. That can still be significant mathematics. It may fill gaps in known territory without creating an entirely new mathematical framework, and it leaves room for people to draw connections, clarify arguments and develop new questions from the results.
The impact is not evenly distributed
The release has hit some research areas harder than others. Mathematicians described particular disruption in probability, combinatorics and theoretical computer science, as well as in parts of group theory. Researchers whose projects overlap with the published results may need to reconsider work that once formed the basis of a dissertation, grant application or research plan.
That uncertainty falls especially heavily on PhD students, junior researchers and others without permanent positions. A long-standing open problem can anchor years of work and help support a case for funding or a job. When a lab publishes a result that appears to address it, researchers must decide whether to keep pursuing the problem, shift direction, or wait for the argument to be checked. None of those choices comes with a ready-made timetable.
Some researchers said entire projects or research programmes had been overtaken by the release. Others found little overlap with their own work, at least on an initial review. The uneven effect matters: a large-scale AI result can be a profound disruption in one specialty and a distant development in another.
OpenAI’s response, and what it has not disclosed
OpenAI worked with the Advisory Group on Mathematics and Artificial Intelligence, a newly formed group of mathematicians, on how to handle the release. The group had called for papers that humans can understand, or support that gives researchers time and resources to understand AI-generated work. It also urged labs to formalize proofs where possible, disclose the models and prompts used, and avoid treating mathematical research as a vehicle for promoting proprietary systems.
OpenAI said it would fund workshops and conferences to help the mathematical community examine the work, although it did not provide details about their timing or format. It also described new protocols for paper revisions and citations, and said it would preserve the release history in public. The correction record gives researchers a way to see changes as they happen, rather than relying on a paper to remain quietly identical to the version first published.
Other requests remain unanswered. OpenAI did not identify the model behind the results, disclose the prompts, or publish the full set of problems it tried. It has said it plans to add more formalizations and improve the exposition and citations in future releases. The company has also said it intends to keep evaluating internal frontier models on mathematics and other sciences, and is working toward a responsible release of the model that produced the results.
That leaves researchers with a sizeable part of the work of assessment, without some information that could help them do it. Knowing the prompts, for example, would give mathematicians more context about how the model approached a problem. Knowing the full set of attempted problems could help clarify how the published results were selected.
The work after the answer
The dispute is not simply about whether AI should produce mathematics. Many of the researchers discussing the release use AI tools themselves, and some welcomed the possibility of new results. Their concern is the way those results arrive: in a large batch, with mixed levels of verification and explanation, followed by an expectation that the mathematical community will sort them out.
Mathematical research involves more than proving a proposition. Researchers develop methods, connect ideas, establish context and identify questions that others can pursue. When a result appears without that surrounding work, the burden shifts to people who may already be working on the same problems, often with little warning.
The community will need time to check the proofs and understand what the results mean. OpenAI’s release may also give mathematicians new material to build on, including ideas and connections that were not obvious before. But a result is not the same thing as an explanation, and a repository is not the same thing as a research programme. The proofs still need readers.