AI labs should responsibly release AI-generated mathematics, survey of 600 respondents says

The recommendations propose two pathways: when a human fully understands a result, it should follow traditional academic norms—preprint, peer‑review, and talks. For results not yet understood, labs must publish the model name, prompts, chain‑of‑thought, compute time and cost, and deposit papers in independent repositories with persistent identifiers. They should also provide funding for workshops, summer schools, or post‑doctoral work to help the community assimilate the findings. Finally, labs are asked to grant broad, equitable access to publicly available models to avoid a two‑tier system that privileges well‑funded institutions.

The authors stress that these steps aim to preserve the core scholarly values of verification and collective insight while ensuring that powerful AI tools do not create new inequities in the mathematical community.