The numbers, read carefully
OpenAI framed the release as 372 result families. Read it the other way: the model was pointed at roughly 4,000 problems and produced publishable claims for fewer than one in ten. Of those, fewer than one in four arrived with machine-checkable proofs. The retraction cascade is the telling detail: one sign error in a single manuscript invalidated the results of two more, which is what happens when outputs are generated faster than they are verified, by a system whose error does not know where it ends.
This is not a criticism of the underlying capability. Even the critics grant the model did something remarkable. The objection is to the release strategy: dropping the unverified remainder on a community whose checking bandwidth is fixed, then calling the result a contribution. Verification is the scarce resource in mathematics, and the October 6 dump spent a large amount of someone else's.
The boycott, in brief
The Association for Human Mathematics is new, formed in response to the release. Its statement argues that publishing at this volume is a way of claiming territory: hundreds of researchers whose projects overlapped the 372 result families now face preemption questions through no choice of their own, and only labs with frontier-scale compute get to play the game at all. The group letter circulating on social media put the sharp version: "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power."
The community is not unanimous. Levent Alpöge, a mathematician at Anthropic, called the release potentially "the most significant moment in mathematical history" for what it showed about AI's usefulness to working mathematicians, while conceding there would be "sad stories" for preempted researchers. OpenAI said its goal is to "provide tools for mathematicians and scientists" and that it consulted an independent advisory committee before publishing. The critics' reply is that the advisory committee's recommendations were not followed, and that 722 simultaneous filings are not how tools are delivered.
The September rehearsal
The Mathathon fight was the preview. In September, Caltech undergraduates planned a 40-hour contest with $2 million in AI compute credits (OpenAI and Anthropic at $1 million each) for 100 teams to attack open problems with language models. Seven hundred seventy-one mathematicians signed a letter calling it slop mathematics, arguing it would flood the field with hastily produced results that others would have to verify without credit. OpenAI withdrew its sponsorship within hours of the letter's publication. The October dump is the same playbook at larger scale, minus the student intermediary.
What it means in practice
If you read AI research claims for a living, the 162/722 ratio is a useful new sniff test. When a lab announces a mass release of AI-generated science, ask what fraction arrived machine-checked before you update your priors. Prose proofs from a system that cannot show its working transfer the verification bill to the reader; Lean-formalized proofs keep the bill with the author. The field's message this week is that it has noticed the difference, and that it is done paying.
Sources
- [1] AI Weekly, “OpenAI posts 722 AI math preprints; mathematicians balk” (Oct 9, 2026; reporting Nature)Read source
- [2] TechSignal, “OpenAI's AI Math Release Sparks a Call for Boycott”Read source
- [3] OfficeChai, “Mathematicians Did Not Ask For So Many Proofs, OpenAI Has Total Disregard For Norms Of Scientific Research: Association Of Human Mathematics”Read source
- [4] Cointime, “OpenAI Withdraws Three Math Papers Due to Symbol Error; Terence Tao Shares Call to End Collaboration”Read source
- [5] HypeFresh, “OpenAI Withdraws $1 Million Caltech Mathathon Sponsorship Following Mathematician Boycott”Read source