Human Mathematics Association Urges Boycott of OpenAI After Release of 700-Plus Math Manuscripts

The statement also raised questions about the basic rules of mathematical research: whether authors understand a proof, whether they have checked it themselves, who bears responsibility for it, and how papers are cited and peer reviewed. Critics quoted by ifanr said OpenAI does not understand what it publishes and cannot read the papers and reports its systems generate, releasing conclusions without verification. The association further argued that a company using internal models the public cannot access to mass-produce results, and then deciding when and what to publish, could gain the power to set the research agenda in mathematics.

Fields Medalist Terence Tao reposted the statement on his personal blog. The repost quickly circulated online as though Tao had led the statement, but he does not appear on the association's membership list. According to ifanr, the association was founded in August 2026 and has 809 members, most of them graduate students rather than mathematicians with significant published results. About a third of the members are French, about a third American, and 11 are from China.

Membership requires three pledges: not to supply technical labor, mathematical knowledge, consulting or promotional services to commercial artificial intelligence companies; not to publish mathematical text generated by AI models, including papers, reviews, referee reports, lecture notes, teaching materials and books; and, for members of its non-AI subgroup, not to use AI models in their own research. Some online commenters noted that when the statement itself was submitted to the AI detection tool Pangram, the result came back as 100 percent AI-generated, while the association's line that mathematicians had not asked for the work was called by one commenter "the most comical, absurd statement I've read recently."

On the substance of the release, ifanr reported that of the 722 papers, only 162 had their main results verified by computer. The day after publication, OpenAI withdrew three manuscripts after a notational error invalidated a core argument and affected two papers that depended on it; 14 more were revised and 13 had citation relationships updated. OpenAI has acknowledged that not all of the papers used the Lean formalization method. A README in its public GitHub repository states that "some unformalized results may have problems. We will work to fix any such issues as soon as possible."

Supporters of the release argue that the manuscripts push mathematics into a stage where its working methods must be redefined, and some researchers say they want to read the proofs with students and postdocs to look for new connections. Within OpenAI's developer community, users have carried out preliminary checks: one reran the Lean verification of a quasi-Riemann result and reported that the code passed, and another extended a matrix multiplication result to a more general field, which also passed Lean verification.

In earlier statements, Tao and others did not deny that AI can help mathematics. Their argument, as ifanr described it, is that treating the solving of open problems as a capability benchmark damages what mathematics values: understanding, insight, the posing of questions and the transmission of knowledge between generations. Commenters cited by ifanr said that OpenAI published the manuscripts for its own commercial benefit without regard for the field's existing structure of named authors, journal submission and peer review, and that the unresolved problems accumulated over decades or centuries give mathematics its structure and its measure of progress.

The Human Mathematics Association is one of several groups formed around the issue. A declaration at mathandai.org, signed by 28 Fields Medalists and supported by more than 8,000 people, describes what it calls a serious imbalance in AI's role in mathematics. A separate body, the Mathematics and AI Advisory Group at agmai.org, is hosted by the Institute for Advanced Study and consists of nine senior mathematicians who work independently and without pay, issuing comments and recommendations on how laboratories should release AI-generated results.