Ethical Guidelines for AI-Generated Mathematical Research
Mathematicians propose norms to maintain integrity in AI-assisted proofs
1Challenging proprietary AI research
Leading AI labs currently test advanced mathematical problems on proprietary models, creating an accessibility gap in the scientific community. Over 600 mathematicians responded to a survey about responsible AI use in mathematics, revealing strong opposition to closed systems. The mathematical community emphasizes that authors must understand and verify their work, a principle threatened by opaque AI systems. This tension highlights the need for new norms in AI-assisted mathematical research.
2Core scholarly principles at risk
Traditional mathematics values human verification, peer review, and author responsibility - norms developed over centuries. Mathematicians typically present work through seminars and conferences, fostering collective understanding. AI-generated proofs bypass these mechanisms when developers can't explain the output. The community warns this could degrade mathematical reliability while creating a two-tier research system favoring well-funded labs.
3Technical norms for AI proofs
Recommendations demand AI labs properly attribute prior work, even if independently rediscovered. Proofs must follow standard mathematical writing conventions with clear theorem statements. Labs should publish model details, prompts used, computation time and costs for each result. Outputs require deposition in independent repositories with version control, not company-controlled platforms that could serve marketing purposes.
4Formalization requirements
AI-generated proofs should be formalized using community standards including machine-readable metadata. When full formalization causes delays, papers must clearly state completion status. For multiple results, labs must document problem selection criteria and failure rates on comparable problems. These measures aim to make AI-assisted research reproducible and evaluable by human mathematicians.
5Supporting human understanding
Labs must fund efforts to help mathematicians comprehend AI outputs, with funding decisions made by existing nonprofit institutions. Suggested activities include workshops, explanatory papers, and verification projects. The community stresses that accepting such support doesn't legitimize questionable AI practices, but holds labs accountable for their intellectual impact.
6Access inequality concerns
Proprietary models risk creating research hierarchies where labs outpace academic mathematicians. Public models face access barriers from economic and institutional disparities. Recommendations urge broad, equitable access to public models to prevent knowledge hoarding. Mathematics thrives on collective verification and shared intuition - processes undermined when key tools remain restricted.
7Two-path publication system
For understood results, traditional peer review and preprint posting remains standard. For opaque AI outputs, additional safeguards apply: literature reviews for prior art, style standardization, and formalization. This dual approach balances innovation with accountability, recognizing different levels of human comprehension in AI-assisted research.
8Preserving mathematical culture
These guidelines aim to protect mathematics' collaborative nature against commercialization and secrecy. By maintaining attribution, accessibility and verification standards, the community hopes to integrate AI as a tool rather than replacement for human insight. The recommendations represent a proactive stance to shape AI's role in mathematics before problematic practices become entrenched.