1. 722 went up. 719 remain.
On 6 October, OpenAI published 722 manuscripts reporting work on 372 mathematical problems. On 7 October, it withdrew three. OpenAI's revision history says a sign error invalidated one construction and two manuscripts that depended on it. The same audit produced revisions to 14 other manuscripts and reference-version updates in 13 more.[1] Retraction Watch independently reported the withdrawals and OpenAI's explanation.[2]
The repository now lists 719 manuscripts, not 722, across the same 372 problem families. It also says 300 top-line results — about 42 percent — have been formalized, and warns that some unformalized results may contain issues.[3] Those are OpenAI's present numbers. They are not evidence that the other 419 manuscripts are wrong, and three withdrawals out of the original 722 are not a measured “failure rate.”
2. One sign error became three withdrawals
The withdrawals were not three unrelated typos. OpenAI says one sign error broke a stabilization-trace cancellation argument and the construction used by two dependent manuscripts. That is what scale changes: an error in one node can travel through a result family before anyone outside the lab has read the first paper.
The same October 7 audit revised 14 other manuscripts for proof repairs, corrected statements, clearer hypotheses or dependencies, and one obsolete citation. Thirteen more were updated to point at revised companion papers. Corrections are normal scholarship. The unusual part is the review surface arriving all at once.
3. Formalized is not the same as independently understood
The current repository reports 300 formalized top-line results out of 719 manuscripts, about 42 percent, and explicitly warns that some unformalized results could have issues.[3] Formal artifacts are substantial evidence: they let a computer check that a stated theorem follows inside a specified formal environment. They do not by themselves establish novelty, attribution, the quality of the exposition, or whether the formal statement captures every claim a reader takes from the prose.
The Advisory Group on Mathematics and Artificial Intelligence therefore recommends that formalization status be stated clearly, alongside the model, prompts, summarized reasoning, time, compute estimate, problem-selection process, and comparable failed attempts.[4] Those are not decorative disclosures. They are the information a reviewer needs to decide what has actually been checked.
4. The fight is over who absorbs the review cost
Nine days before the release, AGMAI published recommendations informed by more than 600 responses. Its central rule was direct: when a lab releases substantial mathematical output without immediate human understanding, the lab must take responsibility for ensuring that understanding follows, including significant financial support.[4]
The Association for Human Mathematics went further. In an October 7 statement reposted by Terence Tao, it described releasing more than 700 files at once as a display of power rather than scholarship and urged mathematicians to stop working with OpenAI.[5] That is one organization's position, not a vote by all mathematicians. Its statement also did not claim that every proof was false. The complaint was about process, scale, and who had been handed the work of evaluation.
The cost is not abstract. Fields Medalist Hugo Duminil-Copin told Le Monde that 16 of the released questions touched his own work and that more than 3,500 relevant pages now needed to be explored. He also worried that researchers would become less willing to discuss open ideas if a proprietary model could race ahead on questions gathered from that open culture.[6]
5. The strongest counterargument: this is what openness looks like
The repository is public. It preserves earlier versions, records withdrawals, publishes corrections, and includes machine-checkable artifacts for many results. OpenAI did not quietly delete the sign error. It documented the dependency failure and kept the withdrawn manuscripts accessible. That is materially better than an unverifiable announcement.
AGMAI itself called publication a first step, while stressing that its advice was not an endorsement of OpenAI's process and that public release begins rather than completes human understanding.[7] On October 15, OpenAI added a concrete response: a $5 million contribution to a new Simons Foundation initiative for mathematicians to examine, understand, and build on AI-produced results, plus model and compute access for supported researchers.[8]
Whether that is adequate is a legitimate argument. But it changes the facts. The story is not simply “OpenAI dumped 722 papers and left.” The sharper question is whether funding after publication can compensate for letting one company choose the questions, generate the volume, and set the community's verification agenda.
6. Every AI-generated result needs a status receipt
A folder full of PDFs is not enough. For a release at this scale, every manuscript should carry a machine-readable status record that answers the questions a reviewer otherwise has to reconstruct:
formalization status and exact scope
independent human review: none / partial / complete
upstream and downstream manuscript dependencies
model, prompt, reasoning summary, time, and compute
comparable problems attempted and failed
revision history with permanent prior versions
This would not decide whether a theorem matters. It would make the state of verification legible, separate a computer-checked claim from a human-understood one, and expose when a correction should propagate to dependent papers.
7. Where StrictCite fits — and where it stops
We audited the repository's 521 public BibTeX files at commit 1581d4c: 13,802 parsed entries, including 9,824 with a DOI field. Fifteen cross-file DOI conflicts survived automated normalization; manual review explained all 15. The strongest-looking error was actually a correct citation to an older arXiv version.[9]
That result does not clear every reference, and it says nothing about the proofs. It shows why verification needs receipts and why suspected errors must survive manual review. StrictCite checks the bibliographic foundation — whether identifiers, titles, authors, years, and versions agree with their sources. Lean and mathematical peer review address a different layer. Neither should pretend to be the other.
To inspect the evidence behind your own bibliography rather than only a pass/fail label: strictcite.com. The free tier requires no card.