OpenAI has set off a fresh wave of intense discussion and skepticism across the global mathematical community by publishing 722 math manuscripts on GitHub. The papers, released on Tuesday, were entirely generated by an internal, unreleased artificial intelligence model. According to an OpenAI spokesperson, nearly all of the results stemmed from a single prompt handed to a single AI agent, though some individual outputs may have required multiple attempts to complete successfully.
The release represents a bold claim and a potentially significant breakthrough in the field of mathematical research. However, it has also met with immediate pushback from academic experts who are questioning the verifiability, transparency, and practical utility of the AI-generated findings.
The publication follows a pattern of high-profile AI math announcements. Last month, a separate OpenAI claim involving the Navier-Stokes equations utilized 10,000 coordinating agents working continuously for 88 hours. In contrast, the newly released papers required an average compute investment roughly equivalent to three hours of ChatGPT Pro thinking compute per result, according to figures shared by the company.
Rather than representing 722 distinct solved problems, the manuscripts are grouped into 372 "families" of related results. Within these families, a main theorem is often bundled with companion arguments, secondary consequences, or alternative proofs. OpenAI reports that it initially posed roughly 4,000 mathematical problems to its unreleased model, ultimately selecting and publishing the outputs it deemed significant enough to share with the public.
Despite the scale of the release, academic skepticism remains high. Prominent mathematicians have pointed out that independent verification is currently limited, making it difficult to assess the true validity of the work.
"Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified," Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, told Scientific American. "We should ask for receipts."
The issue of verification is complicated further by how the papers have been processed. Out of the 722 published manuscripts, only 162 come with a computer-checked main result, according to the formalization catalog included in the GitHub repository. This translates to roughly 22% of the collection being translated into Lean, a specialized software program designed to check every logical step of a proof mechanically.
OpenAI itself has acknowledged that not all manuscripts feature Lean formalizations, admitting that "some of the unformalized results could have issues." In practical terms, this means a substantial portion of what the company has published could ultimately prove to be incorrect. Furthermore, researchers note that a passing Lean check only confirms that a proof follows logically from the statement as written in the Lean language. It does not verify whether that statement accurately reflects the original problem, nor does it guarantee that the resulting mathematical conclusion is genuinely new or important—leaving those crucial judgments entirely up to human scholars.
This dynamic has already led to puzzling encounters for researchers attempting to read the repository. Dmitry Rybin, a mathematician sharing his experiences online, noted that he attempted to read an OpenAI proof regarding the chromatic number of a plane being greater than or equal to six, describing the output as "totally unbelievable alien math" that suddenly introduced concepts like weakly measurable K-coloring seemingly out of nowhere.
Other researchers have pointed out administrative and logistical hurdles in the repository itself. Keith Adler, a mathematician working to formalize OpenAI’s proof of Saxl’s Conjecture in Lean 4, observed that the GitHub repository had issues turned off and had never accepted a pull request, creating a barrier for independent scientists attempting to contribute formalizations back to the project.
The broader implications of autonomous mathematical discovery have also alarmed prestigious research institutions. In a public statement, the Institute for Advanced Study in Princeton, New Jersey, highlighted a profound philosophical and practical shift taking place within the discipline.
"It is now the case that AI can output mathematical arguments in situations without the human who prompted it being able to understand the arguments, verify them, or take responsibility for them," the Institute said. "We believe that human understanding of mathematics remains of paramount importance. How, in this new era, can we work towards a new paradigm that includes human understanding of mathematics as part of responsible scholarly output?"
Not all reactions from the academic community have been critical, however. Some researchers view the release as a watershed moment for the application of artificial intelligence to theoretical sciences. Professor Abhishek Saha described the drop as "a very big day for mathematics," though he categorized most of the generated problems as "exceptional advances within an existing program" rather than entirely unprecedented breakthroughs.
Saha suggested that the vast majority of the problems in the set fall into intermediate categories rather than game-changing territory, noting that the status of solving ultimate challenges—such as the millennium problems—remains largely untouched, with only one problem in the entire collection approaching that echelon: the Quasi-Riemann Hypothesis.
The release also falls short of specific recommendations put forward by an advisory group at the Institute for Advanced Study, which suggested on September 29 that AI companies should proactively disclose the model name, exact prompts, summarized chains of thought, execution time, and total compute cost for every published result. OpenAI has thus far published average compute figures and 10 reasoning summaries, but has omitted the prompts, indicating that it is still working on a responsible path toward releasing the underlying model itself.
Other mathematicians have adopted contrasting perspectives on OpenAI’s strategy of holding back the model. Daniel Litt of the University of Toronto argued that there is no inherent reason for the company to keep the answers to these mathematical questions secret. The approach contrasts with recent moves by competitors like Anthropic, which published all 13 million lines of a Lean-checked proof of Fermat’s Last Theorem on GitHub last month—though that proof formalized an existing theorem published by Andrew Wiles in 1995 rather than generating entirely new mathematical findings.
OpenAI has stated that it intends to add further Lean formalizations to the repository as they become available. For the time being, however, the mathematical community is left to sift through hundreds of unverified manuscripts, balancing excitement over the raw potential of AI-driven discovery against the sobering reality of verifying alien proofs without a clear map.
Web3 & Decentralized Web
Web3 & Decentralized Web
Web3 & Decentralized Web
