Researchers are inviting the mathematics community to revise and scrutinize OpenAI’s AI-generated mathematical manuscripts, creating a public place to improve explanations, restore references and track corrections.

The October 24 announcement on Terence Tao’s blog is a guest post by Abhinav Nellore, Johannes Schmitt, Soledad Villar and Rachel Ward. The authors say OpenAI’s write-ups can be hard to understand and incompletely cited. Their invitation concerns the new community effort, not a new release of an OpenAI model or confirmation that every claimed result is correct.

A public record of review

The oai-math-revised repository is a community fork of OpenAI’s collection. Contributors can propose changes to the manuscripts, file issues or add notes collecting references, comments and follow-up work. The maintainers prefer edits to the underlying TeX sources and say submissions will be reviewed.

The contribution rules also request disclosure of substantial use of language models, proof assistants or other mathematical software. Notes record accepted changes and identify contributors, making the revision process visible rather than silently replacing a document.

A companion browsing site organizes the material by discipline and links papers to their notes. It explicitly warns that the manuscripts are not peer reviewed. Its October 24 activity includes a contributor adding a review reference for an integer-multiplication paper, alongside people registering interest in working on other manuscripts. Those entries show participation, not blanket validation of the collection.

What OpenAI has actually released

OpenAI’s own repository describes 719 manuscripts grouped into 372 families, produced by an internal model. Related papers within a family can include consequences, companion arguments or alternative proofs, so the manuscript count should not be read as a count of independently established breakthroughs.

OpenAI says the material is at different stages of verification and acknowledges that unformalized results may contain issues. It reports that roughly 42% of top-line results have formalizations, with supporting Lean files and verification information available. These are the company’s descriptions of its collection, not a PromptsZone audit of those proofs.

Why the review process matters

The project addresses a practical bottleneck in AI-assisted research: a manuscript must be understandable enough for specialists to inspect, relate to earlier work and reuse. Producing more text does not by itself complete those tasks. A public trail of revisions can help readers distinguish the original output from later human contributions.

It is too early to judge how much of that work this effort will accomplish. An invitation, a repository and early contributions do not establish comprehensive coverage or correctness. The useful test will be the substance of documented reviews and corrections over time, rather than the volume of manuscripts or online attention.

PromptsZone has checked the announcement and public repositories but has not independently verified the mathematical results.

Archival image: Terence Tao at IPAM’s Fireside Chat, March 4, 2026, Institute for Pure & Applied Mathematics, CC BY 4.0; frame from IPAM’s video. Display resized. Tao’s blog hosted the later guest announcement; this image does not depict that project or a PromptsZone interview.