OpenAI's math proof storm challenges academia
OpenAI's Math Proof Storm Challenges Academia, but Mathematicians Aren't Buying It
On October 6, US time, OpenAI published 722 mathematical manuscripts all at once in the openai/math repository on GitHub, grouping them by result relevance into 372 "result families," covering 17 fields including number theory, geometry, and mathematical physics. The Economist's comment was: this day may be
💡 What You Will Learn
OpenAI's Math Proof Storm Challenges Academia, but Mathematicians Aren't Buying It On October 6, US time, OpenAI published 722 mathematical manuscripts all at once in the openai/math repository on Gi
📜 Table of Contents
OpenAI's Math Proof Storm Challenges Academia, but Mathematicians Aren't Buying It
On October 6, US time, OpenAI published 722 mathematical manuscripts at once in its openai/math repository on GitHub, grouped by result relevance into 372 "result families," covering 17 fields including number theory, geometry, and mathematical physics. The Economist's verdict: this day may go down in the history of mathematics. Among them are a rational-number-field version of Hilbert's tenth problem, a counterexample overturning the Kaplansky zero-divisor conjecture, and the Navier–Stokes equation proof that already went viral in September. Less than 48 hours after the manuscripts went live, the math community's reaction shifted from shock to collective resistance.
By the official account, these results are shockingly cheap
According to OpenAI itself, these results came from an unreleased internal model, which was fed 4,000 math problems during evaluation, with each result consuming on average the equivalent of three hours of a ChatGPT Pro subscription in compute. Problems that generations of human mathematicians couldn't crack, at three hours of subscription fees apiece. About 42% of the 722 papers come with Lean formal proofs that can be verified line by line by computer; the rest are still in progress.
Two days later, the first erratum arrived
On October 7, the repository updated its first log: 3 papers retracted, 14 revised. The cause was absurdly small: in a paper related to the Hodge conjecture, the sign of a certain geometric operation was recorded as +1, when by the paper's own convention it should have been -1. One plus, one minus — a count that should have canceled to zero became nonzero, the entire argument collapsed, and two other papers depending on it were retracted along with it. The total manuscript count went from 722 to 719. OpenAI stressed that what was retracted were proofs, not the propositions themselves, and the original drafts remain on file for inspection.
Why the math community isn't grateful
On the day of release, the advisory group OpenAI convened this year at the Institute for Advanced Study (AGMAI) issued a statement. Among the nine members are three Fields Medalists, and the wording was restrained: this is a big deal, but the advisory group's recommendations are just recommendations — they have not endorsed these results or the release process. Harsher still was the Association for Human Mathematics (AHM), led by Terence Tao, whose statement said outright that "nobody asked you to solve these proofs" and urged mathematicians worldwide to stop collaborating with OpenAI. An earlier first report from the advisory group had drawn a red line: frontier AI companies should not unilaterally test deep mathematical problems on internal models. OpenAI ignored it.
Dana Moshkovitz of UT Austin saw the Unique Games conjecture — her life's work — "proven" in these documents, read through them overnight, and her assessment was: it reads like something written by someone on hallucinogens. Logical leaps, piled-up conclusions, not a single derivation a human can follow. This is also where mathematicians' anger lies: the model generates hundreds of papers in days, while a human takes weeks to read one — this pile of gibberish effectively conscripts scholars worldwide as free proofreaders.
How many of the 719 will hold up
There's no answer yet. The 42% with formal proofs are highly credible; the remaining 58% are, by the official account itself, at various stages of verification. As for the September fluid dynamics proof, experts still haven't finished reading it. The AHM characterized this release as "a display of power, not scholarship" — harsh words, but the pace of one erratum every two days suggests it's at least half right: the quality of these manuscripts will take the math community months, maybe years, to weigh.
| Metric | Number |
|---|---|
| Manuscripts released | 722 (now 719) |
| Result families | 372, across 17 fields |
| Lean formalization | ~42% |
| Evaluation problems | 4,000 |
| Compute per result | ~3 hours of ChatGPT Pro |
| Two days after launch | 3 retracted, 14 revised |
Of the 719, how many will ultimately hold up? The math community is still grading. Will AI upend academia as we know it? The big names of the old academic world certainly don't want to see this happen. What do you think? Let's talk in the comments.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.