The mathematical community is pushing back against frontier AI labs that announce advanced mathematical results without providing the underlying human comprehension or full technical details. A new report published on September 29, 2026, by the Association for the Generalization of Mathematics and AI (AGM-AI), outlines strict recommendations for how these labs should handle AI-generated discoveries. The core argument is simple: if an AI solves a problem that no human yet understands, the lab that produced the result bears the financial and logistical responsibility to facilitate that understanding.

The Two-Tier Release Standard

The report divides AI mathematical output into two distinct categories, each requiring different handling. For papers where a human mathematician fully understands and verifies the content, labs must adhere to traditional academic norms: post preprints, submit for peer review, and give seminars. However, for 'black box' resultsβ€”those generated by AI that no human can currently verify or explainβ€”the bar is raised significantly. Labs are required to release the model name, the exact prompts used, a summarized chain of thought, computation time, and estimated costs. They must also scour existing literature to ensure proper citation of related ideas, even if the AI discovered them independently.

Funding Human Interpretation

Perhaps the most controversial recommendation is the demand for financial support. The report argues that AI labs cannot simply dump complex, unverified proofs into the wild and walk away. Instead, they must fund the 'organic' development of human understanding. This includes paying for summer schools, dedicated workshops, and long-term working groups where mathematicians can dissect and verify AI outputs. The authors explicitly state that accepting this funding does not legitimize the labs' opaque practices but rather holds them accountable for the externalities of their marketing-driven releases.

Breaking the Proprietary Barrier

Underpinning these technical recommendations is a broader critique of proprietary model access. The report warns that testing advanced math on closed-source models creates a 'two-tier system' that alienates the global mathematical community. It urges labs to grant equitable access to their publicly available models to prevent exacerbating existing inequalities in institutional wealth and technological privilege. The document emphasizes that mathematics thrives on collective verification and shared intuition, neither of which can survive if the tools generating frontier results remain locked behind corporate paywalls.

Key Takeaways

  • AI labs must release model names, prompts, and cost estimates for all AI-generated mathematical results.
  • Labs are required to fund workshops and postdocs to help humans understand AI proofs they cannot yet verify.
  • Proprietary testing of advanced math is discouraged in favor of broad, equitable access to models.
  • 'Marketing vehicles' for AI models should not come at the expense of mathematical rigor or community trust.

The Bottom Line

This report is a necessary reality check for AI labs treating mathematics as a leaderboard rather than a discipline. If you can't explain the proof, you can't claim the creditβ€”and you definitely need to pay for the cleanup.