Mathematicians open a GitHub hub to fix citations in OpenAI's 719 AI-written math papers

What was launched

On 24 October 2026 a guest post on Terence Tao's blog, What's new, announced oai-math-revised, a community project for improving the mathematical write-ups OpenAI released earlier this month. The post is signed by four mathematicians, Abhinav Nellore, Johannes Schmitt, Soledad Villar and Rachel Ward, and is short. It says OpenAI's papers "can be difficult to understand and can have incomplete citations", and invites students, postdocs, professors and enthusiasts to help fix that.

The work lives in a public GitHub repository under the mathcommunity account, which GitHub's records show was created on 10 October and carries an Apache 2.0 licence. Its tagline reads: "Math is of the people, by the people, and for the people, and we will cite the people." There is also a browsable website, labelled beta, that lists every manuscript by discipline with OpenAI's own summaries, community notes, and buttons for reporting problems or adding references.

The organisers are also recruiting. The post asks people who want to serve as admins or reviewers to email the project and say which mathematical categories, taken from OpenAI's own overview document, they could cover. The project is independent of OpenAI; nothing on the site or in the post says the company is involved or paying for it.

Why OpenAI's papers need the work

The papers come from OpenAI's math repository. As we reported when it appeared, OpenAI released 722 AI-written maths manuscripts with Lean proof artifacts on 6 October. Its README now describes 719 manuscripts in 372 result families, produced almost entirely by an unreleased internal model. OpenAI says the model was posed about 4,000 problems and each result used, on average, three hours of ChatGPT Pro thinking compute.

Not everything held up. OpenAI's own history file records that on 7 October it withdrew three manuscripts after a sign error invalidated an argument in one paper and a construction two others depended on, and revised 14 more with proof repairs and corrected statements. About 42 percent of the top-line results, 300 of 719, have Lean formalisations, which a computer can check. The rest rest on prose proofs that people have to read.

Readability and credit are the problems the new project targets. The write-ups were generated at scale, and mathematicians have complained that they often skip the history of a problem and the people who worked on it. OpenAI's README says it is "exploring community-hosted repositories for these materials". oai-math-revised is the community deciding not to wait.

How contributions work

The README describes three ways to help. The first is a pull request: fork the repository, edit a paper's TeX source, recompile the PDF under its existing file name, and submit the change. Edits can add a literature review, cite related results, improve the exposition, or flag and fix errors. Because the full repository is large, the organisers document a sparse checkout that downloads only the papers you want.

The second is simply opening an issue that proposes a change. The third is a NOTES.md file in each paper's folder with four parts: a dated log of accepted changes, links to follow-up work, links to constructive comments and reviews, and a list of people who want to meet and work through, or in the project's word "digest", that paper together.

Two rules stand out. Contributors are asked to disclose substantial use of automation, which the README spells out as language models, proof assistants and other mathematical software. And people do not have to use their real names; the README says an alternate GitHub handle is fine and that every submission will be reviewed. The website also carries automatically collected notes for all 719 papers, and it labels them as unreviewed AI notes so readers can tell them apart from human review.

What has happened so far

The repository's pull request list shows work began before the public announcement. The first pull request, on 18 October, came from Nellore and added references to a write-up on circulant Hadamard matrices. On 18 and 20 October the account teorth, which Tao uses, added remarks and his own proof sketch to the paper on Landau-Siegel zeros. Schmitt then built the website and issue forms on 20 October and added the collected notes on 21 October.

After the blog post, outside contributions started arriving. By the evening of 24 October the notes pointed readers to Kevin Shu's blog exposition of a paper on a nonspectrahedral hyperbolicity cone, to Mario Bonk's talk on the AI proof related to the Cannon conjecture, and to David Harvey's note on a paper about integer multiplication below n log n. One volunteer signed up to work on two papers. The same day the maintainers cut 344 low-value items from the automatic notes and tightened their inclusion rules.

The numbers are still small. When AI Lookout checked on 25 October, the project had 26 stars, six forks, about 20 pull requests and a handful of open issues, against more than 15,000 stars on OpenAI's own repository. The maintainers also record OpenAI's upstream commits as merged without taking their content, so the community copy keeps its own edits apart from OpenAI's revisions.

The pushback in the comments

The comments under Tao's post were mostly sceptical. Several readers asked why unpaid academics should polish a for-profit company's output. "Why is it our job to revise OpenAI's manuscripts?" wrote Lior Silberman. "If they want these services they should pay for them." Another commenter worried the improved write-ups would become training data for future models, and one said the organisers, several of whom work in machine learning, are not a neutral group.

Others defended the effort. One reply asked why the work should be seen as done for OpenAI rather than for the community, and a commenter named Nestor compared it to volunteers cleaning up a mess in a national park because they care about the park, with accountability for whoever made the mess left for later. A third suggested independent papers on the same results would be a better model than revising OpenAI's text.

The split matches what has happened since 6 October. Mathematicians formed the Association for Human Mathematics, which now has more than 1,000 members pledging not to work for AI firms. At the same time the Simons Foundation, OpenAI and Anthropic set up a $15 million fund for checking AI mathematics. oai-math-revised sits between those positions: it treats the results as worth understanding while insisting on human credit.

What it means for AI-made mathematics

Proof checkers can tell you whether a formal statement follows from axioms. They cannot tell you whether a paper explains itself, whether it cites the people whose ideas it uses, or whether a result matters. Those are the parts of a mathematical paper that turn it into something other mathematicians can build on, and they are exactly what the four organisers say OpenAI's batch is missing.

The project is also a test of whether the mathematical community can absorb AI output at this volume. Hundreds of manuscripts arrived at once. Labs keep releasing new training material for maths models, such as the FineEnvs environments built from OpenAI's math repository, and the stream is unlikely to slow. A shared, versioned place to log corrections, references and reviews is a sensible response, but it depends on volunteers, and the comments show many mathematicians are not keen to volunteer.

For readers outside mathematics, the useful lesson is about provenance. An AI system can produce a plausible paper quickly. Turning it into reliable, properly credited knowledge still takes people, time and a record of who checked what. The hub makes that record public.

What to watch next

Three things will show whether the project works. The first is participation: whether reviewers sign up for the categories the organisers listed, and whether pull requests start changing the papers themselves rather than only the notes. The second is OpenAI's response, since its README already mentions community hosting. If OpenAI points readers to the hub or folds corrections back into its own repository, the volunteer work gains reach.

The third is credit. If the revised versions start citing prior work properly and listing human contributors by name, the project will have answered its own tagline. AI Lookout has not checked any of the mathematics in OpenAI's papers or in the community edits; we report what the project and its contributors have published.