Advisory group sets norms for releasing AI-generated mathematical results

The Advisory Group on Mathematics and Artificial Intelligence, an independent body of nine mathematicians (agmai.org), published recommendations on 29 September 2026 on how AI laboratories should release mathematical results produced by their models. The document, “Responsible Release of AI-Generated Mathematics”, draws on more than 600 replies collected from the mathematical community.

The authors say they do not endorse some frontier labs testing advanced mathematical problems on proprietary models the broader scientific community cannot access, and ask them to stop.

Why the community is concerned

Mathematical scholarship rests on long-standing norms: authors understand a paper's argument, verify its correctness and take responsibility for it. AI can now output arguments that the person who prompted it cannot understand or check. The group argues that human understanding of mathematics remains of paramount importance.

Two paths for release

The document separates results a human understands from results nobody understands yet. For the first, established practice applies: a preprint, peer review at a journal and talks explaining the work. For output nobody understands, the group sets technical norms for the initial release.

An AI lab should search the literature and cite the papers where related ideas first appeared, even if its model found them independently. Proofs should be rewritten in the style of a conventional paper, with precise theorem statements rather than wordy, non-standard exposition. Results should go promptly into scholarly repositories that no AI lab controls.

For every result, the lab should publish the model's name, the prompts used, a summarized chain of thought, the time taken and the estimated cost, and formalize the proof where possible. If many results are released at once, a separate document should reference them all and explain how many comparable problems the model failed to solve.

Funding understanding, keeping access open

Labs releasing substantial output without immediate human understanding, the group says, must provide significant support, including funding, so that understanding follows. Which efforts receive that money should be decided by existing nonprofit institutions, not by the group or the labs. Conferences, workshops, working groups, postdocs and expository articles are among the activities it suggests.

The document warns that proprietary internal models risk a two-tier system in which a few labs outrun the rest of the field, while unequal access to public models could deepen existing inequalities. It advises labs to grant the global mathematical community broad, equitable access to their publicly available models, since mathematics advances through collective verification, conceptual synthesis and shared intuition.