Responsible Release of AI-Generated Mathematics

The document sets out community-informed recommendations for how AI labs should responsibly release significant mathematical results produced by their models, and it begins by urging labs to stop testing advanced problems on proprietary systems that exclude mathematicians. Based on over 600 responses, the core principles are that human understanding must remain central, labs that produce results they do not themselves understand must take responsibility for ensuring human comprehension follows, and the development of that understanding must be community-led rather than directed by the labs. Two tracks are specified: conventional scholarly treatment for papers a human fully understands, and a detailed protocol for results not yet understood.

For AI-generated results lacking human verification, the protocol requires labs to clean up outputs (including proper citation of prior literature and exposition in standard mathematical style), deposit materials in independent scholarly repositories with persistent identifiers and comment capability, and publish model identifiers, prompts, summarized chain-of-thought, runtime and cost estimates, and original LLM outputs where feasible. Proofs should be formalized when possible and the formalization status clearly stated. Labs must also fund community-led activities - workshops, courses, postdocs, expositions - to build understanding, with funding routed through independent nonprofit mechanisms. Finally, equitable access to powerful public models is urged to avoid a two-tier mathematical ecosystem and to prevent release being used as marketing.