OpenAI Dumps 700 AI Math Manuscripts on Preprint Servers

OpenAI has dumped more than 700 AI-generated mathematics manuscripts onto public preprint servers, claiming solutions to open problems — and the response from working mathematicians has ranged from fascination to paralysis.

A math blog collecting reactions drew more than 100 entries. The most-circulated came from Hugo Duminil-Copin, a 2022 Fields Medalist, who wrote: "I am paralysed." The episode, first reported by The Decoder, has escalated a long-simmering debate inside pure mathematics over whether a discipline built on hand-crafted proofs can absorb corporate-scale AI output.

What did OpenAI actually publish?

The submissions are manuscripts posted to preprint servers, including arXiv. They claim solutions to open math problems — questions that have defeated the world's working mathematicians for years. Seven hundred documents from a single lab in one discipline is a volume the field has never absorbed.

OpenAI's broader research mission emphasizes AI that accelerates scientific discovery. The Decoder's report framed the mathematicians' response as a mix of "shock and disgust," and the reaction blog's headline — "How much beauty have we lost?" — captured the affective register several respondents used.

How broad is the reaction?

The blog that collected the responses drew contributions from researchers at institutions around the world. The editors grouped them into three broad registers:

• fascination with novel moves the AI made inside otherwise-flawed proofs
• institutional panic about how to triage and review 700 incoming documents
• a third, rarer thread: open grief over a perceived loss of what drew people to the field

The medium matters here. A preprint server is not a journal. Most papers posted to arXiv receive no formal peer review before publication. They go up so other specialists can read, cite, and contest them. A flood of 700 documents from one source strains that voluntary review system in ways smaller AI contributions never have.

Who is Hugo Duminil-Copin?

Duminil-Copin won the Fields Medal — math's highest honor — in 2022 for work on the geometry of random systems. He holds positions at the Institut des Hautes Études Scientifiques and the University of Geneva.

A one-line reaction from a figure of his standing carries unusual weight inside a discipline where reputation moves slowly and public statements land at discipline-defining moments. Publishing "I am paralysed" is the kind of statement a Fields Medalist reserves for events that reframe the field.

Why does this matter beyond math?

Pure mathematics is one of the cleanest tests of AI reasoning. A model that proves an open theorem is cited — by OpenAI, its competitors, and the wider AI research community — as evidence of general cognitive ability. Solving open problems is the most flattering benchmark the lab can publish.

The 700 manuscripts therefore test three things at once. They test the capability of OpenAI's models. They test the resilience of the math community's voluntary review infrastructure. And they test whether preprint servers can keep functioning when a single source dominates incoming submissions.

The second of those tests is now public. The response blog exists, in part, because mathematicians cannot tell at a glance which of the 700 documents deserve serious attention and which are machine-generated noise that wastes review cycles.

What is the "lost beauty" framing about?

The reaction blog's headline — "How much beauty have we lost?" — borrows from a long-running argument inside mathematics over the relationship between correctness and aesthetic form. The most cited articulation of that argument is G.H. Hardy's 1940 essay "A Mathematician's Apology," which held that the strongest proofs are also the most beautiful. Several respondents echoed that framing, suggesting they read AI proofs as technically possible but formally and aesthetically impoverished.

This is a critique AI cannot easily refute. Correctness is a checklist. Beauty is the thing mathematicians say they trade in.

What happens next?

OpenAI has not, in the reporting to date, responded to the math community's reaction. The blog that collected the 100-plus answers continues to accept new submissions. The 700 manuscripts will be triaged one at a time. The slower they are triaged, the more they burden a volunteer review system that already runs near capacity during normal preprint traffic.

OpenAI's stated mission centers on AI that accelerates research. The math community's reaction is a working case study in what acceleration costs when the discipline has spent centuries producing work by hand. The company can solve 700 open math problems, or solve them in a way mathematicians will accept as their own. By current behavior, it is doing neither.

The phrase "I am paralysed" will outlast the manuscripts that triggered it. It captures the moment when AI did not just answer a math question. It changed what it felt like to spend a career asking them.