OpenAI AI Math Papers: Why 722 Proofs Sparked a Backlash

OpenAI's latest push into machine-generated mathematics has produced a striking volume of results and an equally striking backlash. On October 6, the company published 722 AI-written manuscripts. Within roughly a day it had withdrawn three of them because of errors. The episode has become a flashpoint in a growing argument over who checks AI-produced mathematics, who gets credit for it, and whether releasing work at this scale helps the field at all.

What OpenAI released

The manuscripts were posted to GitHub and claim progress on 372 families of open problems. They were generated by an unreleased internal frontier model. 2 The Verge describes the batch as grouping related papers into these "result families." According to AGMAI, a newly formed independent advisory group of prominent mathematicians assembled to help communicate the findings responsibly, the batch addresses "hundreds" of open questions. 4

The release had been anticipated. In September, OpenAI said its model had "resolved more than 100 long-standing open problems across most areas of mathematics," though it did not say which problems or when details would appear. 4 Alongside the papers, the company published summaries of the model's reasoning, compute estimates, and figures on how many problems it attempted. It claims the average result required the equivalent of about three hours of ChatGPT Pro thinking time. 4

One detail stands out. Roughly half the manuscripts went out without independent verification. 2

The retractions

The weakness of that approach showed up quickly. A sign error invalidated the argument in one paper. Because two other papers built on that paper's constructions, they failed too, and all three were pulled. Fourteen more were revised. 2 The dependency chain matters here. When AI-generated results feed into one another, a single slip can spread through a body of work faster than any human reviewer can follow it.

MIT's Andrew Sutherland said the quick withdrawals were the responsible move. 2 That is a fair point, but it points both ways. Fast correction is good practice. It also shows that errors got past whatever internal checks OpenAI applied before publishing.

"A demonstration of power"

The sharpest reaction came from organized groups of mathematicians. The Association for Human Mathematics called for an end to cooperation with OpenAI. It described the release as a show of technological power rather than a scholarly contribution, and argued that dumping hundreds of manuscripts at once places an unreasonable evaluation burden on researchers. 1 A group letter circulating online makes a similar case. It says mathematicians did not ask for this work, and that releasing more than 700 files at once demonstrates power, not scholarship. 2 Futurism quoted the critics directly: "We reject OpenAI's assertion that this release advances our subject." 3

The objections go beyond volume. Critics are uneasy about OpenAI's models potentially drawing on human mathematicians' work. They also resent the implied expectation that the community will do the hard labor of unpacking and checking the output. 3 Questions about attribution of prior human research are central to the dispute. 1 The release also arrived soon after a separate controversy over how OpenAI's agents reached a claimed solution to the Navier-Stokes equations, which had already strained trust. 13

The verification gap

The most constructive criticism came from Cornell's Alex Townsend. He argued that OpenAI should have led with the subset of results verified in Lean, the formal proof-checking language. 2 This is the crux of the matter. A Lean-checked proof can be trusted mechanically. An unverified natural-language manuscript needs expert hours that the field does not have in surplus. Publishing everything at once, mixed together, effectively hands the verification cost to unpaid outsiders.

The outlets broadly agree on the facts but frame them differently. The Verge places the release within a run of breakthroughs that have "impressed and unsettled" mathematicians. 4 Futurism and the advocacy coverage put the anger front and center. 13

What it means

My reading is that the model's mathematical ability is not really what is in dispute. The dispute is about process. OpenAI appears to have treated mathematics as a benchmark to clear at scale, while the field treats it as a social system of trust, credit, and careful review. Three retractions out of 722 is a low error rate. But the error rate across the unverified half is unknown, and that uncertainty is exactly what mathematicians are being asked to absorb.

If AI labs want their results accepted rather than resented, the path is fairly clear: lead with formally verified results, release in reviewable batches, credit the human work being built on, and pay for the checking. Until then, results this large risk being judged by their worst paper rather than their best.