Recommendations for the Responsible Release of AI-Generated Mathematics

Summary

A mathematics-community document argues that AI labs should stop testing advanced mathematical problems solely on proprietary models, which can leave the wider field unable to inspect the systems or understand their results. The recommendations respond to more than 600 survey replies about an earlier announcement of mathematical results whose details were not provided. They distinguish between papers fully understood by a responsible mathematician and papers whose AI-generated arguments are not yet understood by anyone. The first category should follow conventional practice: preprints, peer review, and talks. For the second, labs should search the literature and provide proper attribution, use models to produce clear theorem-and-proof write-ups, deposit results in independent scholarly repositories, and disclose the model, prompts, a summarized chain of thought, elapsed time, and estimated computational cost. The document also recommends formalizing proofs where possible, stating the formalization status when delays prevent it, and explaining how AI was used, including comparable problems attempted and failed. Because technical release alone may not create human understanding, labs should fund conferences, workshops, working groups, students, postdocs, and expository writing through established nonprofit institutions, with priorities set by the mathematical community rather than the labs. The authors warn that proprietary systems could create a two-tier mathematical world and say publicly available models should be accessible to the global community on an equitable basis. They emphasize that accepting support would not legitimize the labs’ practices, but would help the community understand and assess the released work.