OpenAI’s 722 Papers Put AI Mathematics Under Verification

OpenAI has released 722 mathematical manuscripts produced by an unnamed internal AI model, giving researchers a large body of machine-generated work to examine. The scale of the release is striking, but its importance will depend on how many results survive formal checking, expert review, and attempts to reproduce the underlying arguments.

The papers are available through GitHub and are organized into 372 families covering different areas of mathematics. The release includes PDF files, source materials, citation information, and formal proofs written in Lean, a language that allows a computer to check whether the steps in a proof follow from its definitions and rules.

That package gives mathematicians something concrete to inspect. It also makes the limits of the release visible. OpenAI says some of the results have not been formalized and may contain errors. The company plans to correct those errors while preserving earlier versions, creating a public record of how the work changes over time. The collection therefore begins as a research archive and a test of an AI system’s mathematical reliability.

OpenAI says the model received approximately 4,000 open mathematical problems. Each result required, on average, computing resources equivalent to about three hours of ChatGPT Pro’s Thinking mode. The company has also published ten shortened descriptions of the model’s reasoning.

The figures suggest a system designed for sustained exploration rather than quick answers. It was asked to work across many fields, generate candidate solutions, and produce material in a form that could be examined by specialists. The 722 papers vary in status, and the release does not mean that every manuscript presents a new theorem or a correct proof.

The collection reportedly includes several kinds of work:

That range matters because mathematical value can take different forms. A proof may settle a question. A counterexample may show that a promising conjecture needs to be narrowed. A new construction may open a path toward a result that remains out of reach. A manuscript can also be useful because it exposes a failed approach that researchers can avoid or repair.

The release alone cannot determine which category applies to each paper. That work belongs to mathematicians who can check the definitions, test the claims, compare the arguments with existing literature, and judge whether a result is genuinely new.

OpenAI has released 722 mathematical manuscripts generated by an unnamed internal AI model, along with source files, citations, version information, and Lean formalizations for some results. The collection gives researchers a substantial body of machine-generated work to inspect, but its paper count does not establish that the manuscripts are correct or novel. Lean can mechanically check formal proof steps, while mathematicians must still assess definitions, assumptions, originality, significance, attribution, and the connection between a formal statement and the underlying claim. The archive is therefore best understood as a large public verification project whose value will be determined by the results that withstand expert review and independent reproduction.

Lean formalization is one of the most important parts of the publication. In an ordinary mathematical paper, a proof is presented in natural language and mathematical notation. Specialists examine it line by line, filling in omitted steps and looking for hidden assumptions. That process remains essential, but it depends heavily on human interpretation.

A Lean proof places the argument inside a formal system. When the code compiles, Lean has checked that the stated result follows from the definitions, lemmas, and rules used in the file. This can catch gaps that are easy to overlook in prose. It can also force a team to state every assumption with greater precision.

The check has a clear boundary. Lean verifies the formal proof that it receives. It does not decide whether the theorem is interesting, whether the definitions capture the intended mathematical idea, or whether a result has been described accurately in a paper. It also cannot by itself establish that the work is original or that its citations are complete.

For the OpenAI release, the distinction is especially important because only some results have been formalized. A formalized proof comes with a strong mechanical checkpoint. An informal manuscript still requires the conventional process of expert review, even when its argument appears detailed and plausible.

A practical reading of the archive may therefore involve several levels of confidence:

Those levels should remain separate. A paper can advance through them at different speeds, and some may never reach the final stages.

The most prominent example connected to OpenAI’s mathematical work is a proposed solution to the existence and smoothness problem for the Navier-Stokes equations. The problem asks whether solutions describing fluid motion remain smooth for all time under appropriate conditions, or whether singularities can develop.

It is one of the seven Millennium Prize Problems, a group of major mathematical questions selected by the Clay Mathematics Institute. A correct solution to one of those problems carries exceptional significance, both for mathematics and for the broader understanding of mathematical physics.

OpenAI announced that its internal model had produced a proposed solution and later shared a writeup and formal proof materials. The mathematical community still has to assess the argument. The presence of a Lean file can help with checking formal steps, but it does not remove the need to understand the definitions, translate the intended claim correctly, and verify that the formalized statement addresses the original problem.

The history surrounding the claim has also affected the discussion. OpenAI began working on the problem after rumors that other researchers had made progress without publishing their results. Some mathematicians criticized the decision as a possible breach of accepted norms around scientific collaboration. Those norms are informal, but they matter in fields where ideas can circulate privately for months or years before a paper appears.

The episode gives the 722-paper release a broader context. Mathematical publication depends on more than producing an argument. It also involves attribution, priority, communication with researchers working on related questions, and a shared understanding of when a result is ready for public presentation.

OpenAI has not disclosed the name of the system that generated the manuscripts. The company describes it as an internal state-of-the-art model and says it is working toward a safe release.

That limited disclosure makes independent evaluation harder. Researchers can inspect the outputs, but they cannot fully study the model’s training data, architecture, search process, failure patterns, or internal selection criteria. They may be able to determine whether individual arguments work without knowing how often the model produces false leads or how much human intervention shaped the final papers.

OpenAI says the publication format was developed with an independent advisory group from the Institute for Advanced Study. The group helped establish guidelines for citations, corrections, and the presentation of results for the mathematical community. Those rules are useful because a large automated release can create several practical problems at once:

The decision to preserve earlier versions is particularly important. A living archive can improve as errors are found, but quiet revisions make it difficult to determine what was originally claimed or how a result changed after review. Version history gives researchers a way to track that process.

The release can happen in a day. Verification cannot.

Mathematical review is slow because a result may depend on a chain of definitions and technical lemmas that are individually familiar but difficult to combine. An argument can fail because of a missing hypothesis, an incorrect boundary condition, an unjustified limit, or a misuse of a result from another field. A formal checker may identify some of those problems, while others require a mathematician to examine what the formal statement actually means.

The 722 manuscripts will also compete for limited expert attention. Specialists usually focus on a narrow area. A researcher in algebra may not be equipped to assess a claim in partial differential equations, and a paper in theoretical computer science may require a different set of standards and tools. Organizing the archive into families helps, but it does not create enough reviewers by itself.

OpenAI’s own warning about possible errors sets an appropriate baseline. Researchers should treat the manuscripts as claims awaiting review. Headlines describing hundreds of solved problems move faster than the evidence needed to support them, especially when a result has not yet been independently checked.

The process will likely produce several outcomes:

That distribution would be normal for a large research program. It would also provide a more meaningful measure of the model than the raw paper count.

The ten summarized accounts of the model’s reasoning may help researchers study how it searches for mathematical ideas. Short summaries cannot expose every internal step, but they can show whether the system relies on familiar techniques, combines results from different areas, proposes useful intermediate lemmas, or repeatedly follows patterns that appear convincing but fail under inspection.

The source files and citation records add another layer. They may show how the model handles existing literature and whether it can identify the difference between a genuinely new result and a rediscovery of known work. That distinction is central to mathematical research. A correct proof of a known theorem can still have value if it introduces a useful method, but its contribution must be described accurately.

The archive also creates an opportunity to compare mathematical fluency with mathematical judgment. A model can produce polished notation, plausible terminology, and long chains of deductions. Researchers need to know whether it can select productive problems, recognize when an approach has failed, identify hidden assumptions, and explain why a result matters.

Those abilities are harder to measure than fluency. They require tracking errors across many attempts and examining the rate at which useful ideas emerge from the system’s search. The 4,000-problem starting point may eventually allow that kind of analysis, provided the release records enough information about unsuccessful work and human editing.

OpenAI’s release changes the conversation by placing a substantial body of AI-generated mathematics in public view. Researchers now have papers, code, formal proof artifacts, and version histories to examine. That is more useful than broad claims about an unnamed system’s capabilities.

The central question remains empirical: how many of the 722 manuscripts withstand serious checking, and what kind of mathematical contribution do they make? A high number of verified results would show that AI systems can support advanced mathematical research at a meaningful scale. A lower number would still reveal where automated mathematical reasoning breaks down and which tools help researchers find those failures.

The answer will emerge through ordinary scholarly work: reading, checking, formalizing, citing, correcting, and sometimes rejecting. Lean can accelerate part of that process. Public version histories can make it more accountable. Independent mathematicians will decide whether the claims hold.

For now, the archive represents a demanding verification project. Its importance will be measured by the quality of the results that survive that project, rather than by the size of the initial upload.