OpenAI's mass math release leaves mathematicians years behind
Mathematicians told The Verge that OpenAI's nearly 400 AI-generated results may take years to assess, with about 42% of manuscripts formalized in Lean.
Mathematicians are struggling to digest a large batch of AI-generated research that OpenAI released this week, and several told The Verge that making sense of it could take years. The release is large, the verification is uneven, and some of the work looks significant enough to matter well beyond the release itself. That combination makes it a useful case study for anyone who relies on AI-generated technical output.
What OpenAI released
According to The Verge, OpenAI published nearly 400 AI-generated results spread across more than 700 manuscripts. The topics range over combinatorics, several branches of geometry, number theory, theoretical computer science, algebra, topology, probability, statistical mechanics and mathematical physics. The material lives in a GitHub repository that is big enough that OpenAI published a guide on how to navigate it.
The Verge's reporter, Robert Hart, spoke with more than three dozen mathematicians. Their descriptions ranged from 'staggering' to 'pure insanity', and many said the first task was simply reading. Several reported that working through the roughly 40-page table of contents and abstracts alone took most of an hour. Some researchers expect that understanding the full scope could take years, and a number of them worry that OpenAI will not wait that long before moving on or releasing more.
Why verification is the bottleneck
Formal verification is the main thing that separates a claim a machine can check from one that depends on human reading. Some of the manuscripts come with formalizations written in Lean, a programming language and proof assistant that lets results be verified computationally. Those formalizations helped researchers assess some of OpenAI's earlier mathematical claims.
OpenAI itself said the results are 'at different stages of verification' and that many, but not all, manuscripts have been formalized. The Verge reports that OpenAI put the figure at 300 top-line results out of 719 manuscripts, around 42 percent, and said it will add more formalizations as it obtains them.
A Lean file does not settle the question quickly, though. Researchers must confirm that the formal statement actually proves what the accompanying paper claims. Several mathematicians who examined the papers told The Verge that even where computer-checkable proofs existed, their quality was inconsistent and the verified statements did not always map neatly onto the manuscript's claims.
Kevin Buzzard, a mathematics professor at Imperial College London, described identifying many theorems in algebraic number theory, only about six of which immediately stood out to him. Few of those appeared to be formalized in Lean, which leaves him either reading unverified work, waiting for others to do so, or waiting for someone to formalize the results before he can be sure they are correct.
The slop problem and attribution
Researchers use the word 'slop' for low-quality and often erroneous AI-generated material. The Verge says such content is increasingly common online and in the real world, including in academic papers. Mathematicians told the outlet they had seen a sharp rise in it produced with tools such as ChatGPT and Claude. Much of it is hard to read, shows little grasp of the subject, and is especially weak at crediting other researchers.
OpenAI's earlier mathematical write-ups were widely criticized by experts, particularly for poor or missing attribution. Before this release, several researchers had begun calling the expected flood of papers a 'slopocalypse' or something similar.
Whether that flood arrived is hard to judge given the volume. Early signs suggest OpenAI took more care with at least some of these papers. Several mathematicians said their first impressions were better than they had expected, while acknowledging that the bar set by the company's previous publications was low.
Better, though, is not the same as meeting the standards expected for academic work that makes claims of this size. Nalini Joshi, a mathematics professor at the University of Sydney, found little overlap with her own work in a quick search, but noted that some papers she examined have short bibliographies. Given the earlier criticism, she said she is wary that attribution may be missing part of the story. Brendan Hassett of Brown University said a write-up of a problem he knows well made little sense after one read. He made that remark before OpenAI retracted three papers, according to The Verge.
Why researchers still see real value
The Verge reports a broad consensus that the release also contains genuinely impressive work, even with the presentation problems. Numerous researchers said that in a world without AI, many of these results would clearly merit top-tier journals and could be enough to secure an academic career for their authors. A handful were described as the sort of work that could make a mathematician a serious candidate for a Fields Medal.
Stanford mathematician Jared Duker Lichtman said there were 'tens' of results in that category. He cited progress toward the Riemann hypothesis, a special case of the Hodge conjecture, and a solution to the four-dimensional Kakeya conjecture. Riemann and Hodge are both among the seven Millennium Prize problems, so these are not minor claims. That is exactly why the verification gap matters: strong results with weak documentation are hard to trust and hard to build on.
What it means in practice
For working researchers, the immediate job is triage. Abstracts are the first filter, but the papers are the main way to check, understand and place a result in context when formal proofs are missing. The Verge's reporting suggests that the quality of those papers therefore matters more than usual. Researchers also have to decide which claims to chase now and which to wait on until someone formalizes them.
The lesson extends beyond mathematics. Any team that receives a large volume of AI-generated technical material faces the same pattern: the volume outruns review capacity, some items are excellent, and the ones that need checking are mixed in with the rest. Confidence should follow evidence, such as a machine-checkable proof, a clear attribution trail or independent replication, rather than fluency or the scale of the release.
How to work through a large paper set in Grafy
Grafy Reader is built for reading, searching and questioning PDFs and papers with AI, so it fits the first stage of this problem: getting oriented across a large body of documents. You can use it to search a set of papers, ask questions about specific sections, and check what a paper says about the work it cites. Our guide on a reading room for papers, with the AI on tap covers the workflow, and chatting with a PDF and knowing what was read explains how to keep track of the source text behind an answer.
Reader does not verify proofs. It cannot confirm that a Lean formalization matches a paper's claim, and nothing in Grafy should be read as a substitute for expert review. Treat any AI answer as a pointer back to the original text, then check the argument and the cited prior work yourself. You can open Reader and try it on your own papers.
Frequently asked questions
What did OpenAI release, according to The Verge?
The Verge reports that OpenAI released nearly 400 AI-generated mathematical results across more than 700 manuscripts. The topics include combinatorics, several geometry branches, number theory and mathematical physics. The material is hosted in a GitHub repository, and OpenAI published guidance to help readers navigate it. OpenAI said the results are at different stages of verification and gave a figure of 300 formalized top-line results out of 719 manuscripts.
What does Lean formalization add to a mathematical result?
Lean is a programming language and proof assistant that allows results to be verified computationally. A formalization gives researchers confidence that a claim is logically correct, even if they do not fully follow the argument. The Verge notes that such formalizations helped assess some of OpenAI's earlier claims. Researchers still need to check that the formal statement actually proves what the accompanying paper claims, which takes additional time.
Can Grafy verify these kinds of proofs?
No. Grafy Reader lets you read, search and question PDFs and papers with AI, which can help you get oriented across a large set of manuscripts. It does not verify mathematical proofs and cannot confirm that a Lean formalization matches a paper's claims. Treat any AI answer as a starting point, then check the original text, the formal proof and the cited prior work yourself before relying on a result.
