Mathematicians Release Recommendations for Responsible AI-Generated Math

A group of mathematicians has published a set of recommendations for AI labs generating significant mathematical results, emphasizing that human understanding must remain central to scholarly output. The guidelines argue that while AI can produce complex proofs, the current practice of testing advanced problems on proprietary models without community access is problematic and should cease.

Core Principles

The recommendations are built on three overarching principles:

• AI labs must responsibly release significant results as soon as possible.
• Labs releasing output without immediate human understanding must fund and support the development of that understanding.
• The process of understanding must remain organic and community-led, not directed by the AI labs.

Technical Release Standards

For results not yet understood by humans, labs are recommended to take specific technical steps before and during release:

• Literature Review & Citation: Labs must scour literature for related ideas and cite them, even if the AI discovered them independently.
• Standardized Formatting: Proofs should be rewritten in traditional mathematical style with clear introductions and theorem statements, avoiding non-standard terminology.
• Repository Deposition: Results must be deposited in neutral, scholarly repositories with persistent identifiers, not used as marketing vehicles.
• Full Transparency: Labs must publish the model name, prompts, summarized chain of thought, time taken, and estimated computation cost.
• Formalization: Proofs should be formalized where possible, meeting community standards for artifacts and metadata.
• Contextual Documentation: Each release must document how AI was used, including the number of comparable problems the models failed to solve and how problems were selected.

Supporting Human Understanding

The document asserts that AI labs have a responsibility to provide funding for the mathematical community to assimilate AI-generated results. This support should be distributed by independent nonprofit institutions and could fund:

• Conferences or summer schools for complex proofs.
• Workshops and long-term working groups.
• Postdocs or students to assist with verification and application.
• Experts writing books or expository articles.

Access and Equity

The guidelines warn against a two-tier system where proprietary models allow labs to outrun the broader field. They advise granting the global mathematical community broad, equitable access to publicly available models to prevent exacerbating existing inequalities and to ensure that collective verification and shared intuition can thrive.