722 Math Papers in a Day Broke the Review Queue

722 Math Papers in a Day Broke the Review Queue

On 6 October 2026, OpenAI published 722 manuscripts in 372 result families to a GitHub repository, all produced by an internal frontier model the company has not released. A day later, it withdrew three of them over a sign error and revised fourteen more.

That one-day withdrawal is the whole story in miniature. Producing mathematics at industrial scale turns out to be easy. Reviewing it, understanding it, and deciding what it means is the hard part — and that work still belongs to humans.

This is a follow-up to my two September posts: the technical breakdown of the Navier–Stokes claim and the credit dispute around it. What happened next moved the argument from one proof to the entire production line.

What OpenAI actually released on 6 October

The release lives in the public openai/math GitHub repository. OpenAI says it consulted the independent Advisory Group on Mathematics and AI at the Institute for Advanced Study and drew on that group's public recommendations to shape the release, including protocols for paper revisions and citations.

Concretely, the release includes Lean formalizations for many proofs but not all, with more promised later, plus 10 reasoning summaries, compute estimates framed as ChatGPT Pro usage — the average result cost roughly three hours of Pro thinking — and statistics on attempted problems. The model itself stays internal, though OpenAI says it is working to responsibly release it, alongside funding for workshops on understanding AI-produced results.

Every item on that list concedes that dumping files is not doing mathematics. The governance layer is being built after the factory already shipped.

Three withdrawn, fourteen revised

Retraction Watch reported on 8 October that OpenAI announced the withdrawal of three manuscripts on 7 October because of a sign error. The error invalidated the argument in one manuscript and the construction used by two dependent papers. Each withdrawn paper carries a notice explaining the gap. In the same announcement, OpenAI said it had revised fourteen other manuscripts with proof repairs, corrected statements, clearer hypotheses and dependencies, and one fixed obsolete citation.

That is revision control working as designed: notices on the papers, stated reasons, dependent papers flagged rather than silently patched.

But count the reviewers. Three withdrawals and fourteen revisions within 24 hours of a 722-file drop means seventeen papers changed before anyone outside the company could have read even a fraction of the release. A Lean formalization can catch a wrong step, but only if the formalization exists — and OpenAI concedes many results do not have one yet. Everything else waits in the same queue it has always waited in: unpaid human referees, one careful reading at a time.

A script makes the scale concrete. Given a local clone of a mass-release repository, this walks every manuscript folder and reports which ones carry withdrawal or revision notices:

import os
import sys

NOTICE_FILES = ("WITHDRAWN.md", "withdrawn.md", "REVISION.md", "revision.md")
NOTICE_MARKERS = ("withdrawn", "gap in the proof", "revised", "proof repair")


def audit_manuscripts(root):
    total = 0
    flagged = []
    for entry in sorted(os.listdir(root)):
        path = os.path.join(root, entry)
        if not os.path.isdir(path):
            continue
        total += 1
        hits = []
        for name in sorted(os.listdir(path)):
            full = os.path.join(path, name)
            if os.path.isdir(full):
                continue
            if name in NOTICE_FILES:
                hits.append(name)
                continue
            if name.endswith(".md"):
                try:
                    with open(full, encoding="utf-8") as fh:
                        head = fh.read(2000).lower()
                except OSError:
                    continue
                if any(m in head for m in NOTICE_MARKERS):
                    hits.append(name)
        if hits:
            flagged.append((entry, hits))
    return total, flagged


if __name__ == "__main__":
    root = sys.argv[1] if len(sys.argv) > 1 else "manuscripts"
    try:
        total, flagged = audit_manuscripts(root)
    except FileNotFoundError:
        print(f"missing folder: {root}")
        raise SystemExit(1)
    print(f"{total} manuscript folders, {len(flagged)} carry notices")
    for slug, hits in flagged:
        print(f"- {slug}: {', '.join(hits)}")

Seventeen moving files out of 722 is a 2.4% day-one churn rate. For a normal journal that would be a scandal. For a repository drop it is presented as process. Both descriptions are true, and the gap between them is the governance problem.

The mathematicians who saw this coming

The 6 October drop did not land on an unprepared community. On 11 September, three days after the Navier–Stokes announcement, 25 Fields Medalists signed a declaration titled A Severe Misalignment of AI in Mathematics, published on Terence Tao's blog. Its core charge: "the push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics," and "the goals of the AI companies and the goals of the mathematical community are severely misaligned."

The declaration makes three arguments worth stating plainly. First, famous problems are landmarks for measuring understanding, not leaderboard entries; treating them as benchmarks inverts the purpose. Second, rushed announcements skip the slow work — proper writeups, isolated methods, citations of prior work — which raises attribution and plagiarism questions. Third, the scarcest resources are students and ideas, and mass-producing "true/false" statements without the human transmission chain of talks, simplifications, and teaching leaves AI-conceived ideas inert: "the mass production at faster and faster pace of 'true/false' statements could destroy fertile ground instead of breathing life into new ideas."

A month later, the Association for Human Mathematics answered the 722-file drop in the same register. As quoted by Retraction Watch, the group said: "Mathematicians did not ask for this work to be done," adding that "releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power," and urging mathematicians to discontinue work with OpenAI in favour of "a vision of science that centers human understanding."

You do not have to endorse the boycott call to see the point. Scholarship is rate-limited by design. Power is the ability to ignore the rate limit.

Even OpenAI's own mathematicians cannot keep up

Scientific American reported that an OpenAI spokesperson told the magazine many of the newly released results are not yet understood by the company's own mathematicians — and that the company does not plan to slow down, because these problems are the test it uses to show its models are getting smarter.

The producer cannot yet explain its own output. The reviewers have 722 files to triage. The model stays internal, so nobody else can probe how the results were reached. The only verification layer that scales — Lean checking — covers many proofs but not all, and checking a formal statement is not understanding it. My September breakdown of the Navier–Stokes formalization made exactly this distinction: the checker confirms the theorem follows from its assumptions, while humans must still decide what the theorem means.

Each AI-produced result arrives with a verification debt, and the debt compounds with volume — as the earlier Erdős conjecture episode already showed.

Clay's clock is deliberately slow

Against the speed of releases, the Clay Mathematics Institute's 11 September statement reads like a different century — deliberately. Clay said it shares the excitement as it contemplates the announcement that the Navier–Stokes problem has "apparently been settled," and hopes for "waves of new human understanding" as the work is analysed. Then the brake: the prize rules describe the evaluation and credit process, "the process is deliberately unhurried," and updates will follow.

The rules behind that sentence are specific. Before Clay considers a proposed solution, three conditions must all hold: publication in a qualifying outlet, at least two years since publication, and general acceptance by the global mathematics community. Clay accepts no direct submissions. OpenAI, for its part, says it does not intend to claim the prize.

So two clocks are running. The repository clock measures files per day. The Clay clock measures years and consensus. The first serves model evaluation; the second serves mathematics. The medalists' declaration is, at bottom, a demand that the first clock stop pretending to be the second.

What to watch instead of the file count

File counts are the wrong metric from here. Four signals matter more.

First, Lean coverage. Every result that gains a machine-checkable formalization moves from trust-me to check-this. Watch the fraction, not the total.

Second, the withdrawal-and-revision log. A public, prompt correction record is the difference between a repository and a dump. The 3-and-14 episode is the baseline; the question is whether the next correction is equally visible.

Third, exposition. Workshops, readable writeups, isolated methods, proper citations — the slow goods the medalists asked for. The promised conference funding is a start, not a result.

Fourth, the model. An unreleased model producing hundreds of unreviewed results concentrates both capability and accountability inside one company. OpenAI's promise of a responsible release is the load-bearing commitment in the whole arrangement.

The Navier–Stokes episode began as a priority dispute between researchers. A month later it is a governance dispute between a production line and a profession. Proofs are now cheap. Understanding is still handmade. Until the review queue catches up with the release pipeline, every new file count measures the backlog, not the breakthrough.

Sources

• OpenAI — Sharing AI progress in mathematics
• Clay Mathematics Institute — Navier-Stokes Announcement
• Clay Mathematics Institute — Rules for the Millennium Prize Problems
• Terence Tao — A Severe Misalignment of AI in Mathematics
• Retraction Watch — OpenAI withdraws three preprints a day after releasing 722 manuscripts
• Scientific American — OpenAI unleashes hundreds more math results upon a field already in shock
• OpenAI — On the Navier–Stokes Millennium Prize Problem

---

Editorial note

Counts (722 manuscripts, 372 families, 3 withdrawn, 14 revised) and the advisory-group, Lean-coverage, and compute details are attributed to OpenAI's release page and Retraction Watch's reporting as linked. Quotations from the Fields Medalists' declaration, the Clay statement, and the Association for Human Mathematics are presented as those parties' statements, not as established fact. This post is a governance follow-up to two earlier hejes.my articles and does not re-litigate the September authorship dispute.