OpenAI unveiled a massive collection of AI-generated mathematics this week, consisting of almost four hundred results distributed across more than seven hundred manuscripts. The material touches on fields ranging from combinatorics and geometry to number theory, theoretical computer science, algebra, topology, probability, statistical mechanics and mathematical physics. Researchers interviewed by The Verge said the sheer breadth of the drop has left the community scrambling to understand its implications for future work.
OpenAI also published a guide for navigating the sprawling GitHub repository, acknowledging that even a first pass through the roughly forty-page table of contents can consume an hour. Álvaro Lozano-Robledo, a mathematics professor at the University of Connecticut, told The Verge that reviewing the entire list of abstracts felt overwhelming. The volume, he said, makes any preliminary assessment of the content a daunting task for scholars across the discipline.
Among the submissions, some include formalizations written in Lean, a proof-assistant language that can mechanically verify statements. The Verge reported that these Lean files have been useful for checking earlier OpenAI claims, yet their quality varies and they do not always correspond cleanly to the surrounding manuscripts. Kevin Buzzard, a professor at Imperial College London, noted that only about six of the algebraic number theory theorems he examined stood out, and few were accompanied by reliable Lean verification.
The community also expressed alarm over what they label “slop,” low-quality AI-generated output that frequently contains errors and poor attribution. Researchers said the term “slopocalypse” has been used to describe the surge of such material in recent years, especially from tools like ChatGPT and Claude. OpenAI’s earlier mathematical papers were widely condemned for sloppy citations, prompting many to anticipate a repeat of that problem with the current flood of preprints.
Several mathematicians described the new papers as difficult to follow, with some appearing almost unintelligible. Brendan Hassett, a professor at Brown University, told The Verge that the write-up of a familiar problem made little sense after a brief read, and that he would not invest further time in it. He added that the release initially included 721 preprints, though OpenAI later retracted three of them after concerns were raised.
Despite the shortcomings, a number of scholars highlighted genuinely impressive contributions. Stanford mathematician Jared Duker Lichtman said he could identify tens of results that might qualify for top-tier journal publication, with some potentially competitive for a Fields Medal. He cited advances toward the Riemann hypothesis, a special case of the Hodge conjecture and a solution to the four-dimensional Kakeya conjecture as examples of the most striking claims in the batch.