Association for Human Mathematics urges OpenAI boycott


The appeal follows a release of hundreds of claimed solutions and a dispute over how to check, credit and explain SI-generated proofs.
The Association for Human Mathematics is calling on mathematicians to stop collaborating with OpenAI over its publication of SI-generated proofs, The Decoder reported on 8 October. The association accuses the company of violating scientific research norms. Its statement is a boycott appeal, not a rule imposing restrictions on researchers.
The release contains 722 manuscripts covering 372 “result families,” or groups of related papers, generated by an unreleased model, according to The Verge’s report on 6 October. It follows OpenAI’s September claim to have resolved more than 100 long-standing open problems across most areas of mathematics, before it disclosed the problems or a precise publication date. The papers are on GitHub, a site for hosting code and project files, with revision records and citations. The independent Advisory Group on Mathematics and Artificial Intelligence (AGMAI) says the collection includes solutions to “hundreds” of open questions. How many claimed solutions independent mathematicians will confirm remains unclear.
The dispute concerns whether a problem can count as solved before independent mathematicians have worked through its proof and made it understandable to humans. The Association for Human Mathematics describes the release of more than 700 files at once as a demonstration of power rather than scholarship. It also alleges that OpenAI ignored what the association considers a central premise of AGMAI’s advice: that advanced mathematical problems should not be tested on internal models.

Terence Tao, who chairs the association, shared its statement on his blog. In an earlier statement, Tao and 24 other winners of the Fields Medal, a mathematics award, argued that solving problems serves a broader goal: understanding the mathematical ideas behind them. They warned that rapidly producing solutions could reduce opportunities to develop new ideas.
Those signatories also criticized rushed announcements without proper explanations or citations of relevant earlier work. They said these practices raise “severe attribution and plagiarism questions.” These are the mathematicians’ concerns, not established findings that OpenAI plagiarized the work.
The Decoder illustrates the readability problem through researcher Dana Moshkovitz’s response to a claimed proof of the Unique Games Conjecture, an open problem in theoretical computer science. In messages published by complexity theorist Scott Aaronson, Moshkovitz said much of the paper was unclear and cited earlier work without explaining why it applied. She said the paper was impossible to read without superintelligence (SI) help. Her criticism concerns the presentation of that claimed proof; it does not establish that every manuscript is unreadable or incorrect.
Tao argues that traditional solutions to long-standing problems generate talks, workshops and collaborations through which researchers understand and build on the results. He says autonomous SI solutions can produce less of that exchange when the people presenting them cannot explain the proofs or answer questions. He advocates judging progress more by explanation, community-building and new research directions, with corresponding changes to training, publication and career advancement.
Other mathematicians assess the work differently. In Fortune’s report on 7 October, University of Toronto mathematics professor Dan Litt welcomed solutions affecting problems he wanted to understand and praised their accessibility on GitHub. Litt, who was not an AGMAI member, said extracting understanding would require substantial human work, but compared it with work mathematicians have long done.
Litt approved of most aspects of the publication. In his assessment, OpenAI could not produce rigorous academic papers for every result because its models struggle to explain mathematics and cite prior work, and the company lacks enough mathematicians with expertise across the relevant fields. His assessment does not confirm every claimed solution.
AGMAI takes a different position from the association making the boycott appeal. According to The Decoder, it sees the release as a first step, with mathematical understanding of the results only beginning. The advisory group has no say over the pace of OpenAI’s internal research.
AGMAI nevertheless warns that mathematics cannot become solely the work of interpreting results produced by SI laboratories. It says researchers need fair access to powerful tools and sufficient computing resources to pursue questions of their own.
Mathematicians must be able to formulate their own questions, develop their own approaches, and explore directions that have not been selected as examples of an AI system's capabilities.
Many proofs have formalizations in Lean: versions written in a programming language that allows software to verify mathematical logic. The Decoder reported on 7 October that these checks could help with a volume of results that might overwhelm manual review. They cannot establish mathematical relevance or originality, and having formalizations for many proofs does not make the entire collection machine-checkable.
OpenAI says it will share more formalizations as it obtains them.
According to OpenAI, nearly every result came from a single prompt—an input given to the model—to a single SI agent, although some required multiple attempts. The company estimates average computing use equivalent to roughly three hours of ChatGPT Pro Thinking. It supplies some reasoning summaries and attempt statistics, and says the repository includes protocols for revising and citing the papers.
In late September, AGMAI recommended following traditional mathematical research-paper conventions so researchers could scrutinize and learn from SI-generated proofs. It also urged prompt publication through established academic channels where possible.
For every solution, the group requested the model’s name, prompts, reasoning steps, time taken, approximate computing cost and an explanation of why the problem was selected. For batch releases, it also wanted an account of why those problems were targeted and how many comparably difficult problems the model tried unsuccessfully.
Fortune reported on 7 October that OpenAI followed some, but not all, of those recommendations. OpenAI said it supplied reasoning summaries, computing-resource estimates and attempt statistics for ten problems, rather than every solution. It has not released the prompts. The Decoder’s report on 7 October gave average computing figures; Fortune described estimates for ten problems, not computing-cost disclosures for every solution.
In its statement on 6 October, AGMAI reaffirmed its recommendations, described discussions with OpenAI as constructive and left assessment of compliance to the mathematical community. Its earlier recommendations also urged companies not to use mathematical results to market their models, arguing that this significantly harms the mathematical community.
OpenAI says it is exploring community-hosted alternatives for the release that meet AGMAI’s guidelines. It has committed to improving citations, mathematical explanations and presentation in future releases, and says it will fund workshops, conferences and programs to help mathematicians understand the results.
Sierra, Walmart and Stripe are working with Meta on rules the partners say would help businesses recognize authorized agents and see their activity.
2 min

A patch is reported to fix the flaw; CISA also recommends limiting control-system network exposure.
2 min

The vendor says SOI versions 2.0.0–2.2.0 are affected and recommends its cumulative EP2 patch. CISA has republished the advisory verbatim.
2 min