Mathematicians call for scrutiny of AI proof claims

A group of 16 mathematicians working with colleagues and mathematical organizations worldwide has published the Leiden Declaration on Artificial Intelligence and Mathematics, urging governments, funders and researchers not to overstate AI systems’ mathematical capabilities. The declaration says mathematical work should remain open, transparent and independently verifiable as AI models begin attracting attention for research-level results.

The warning follows OpenAI’s announcement in late May that one of its models had disproved a notable 80-year-old conjecture in combinatorial geometry, one of about 1,200 open questions proposed by Hungarian mathematician Paul Erdos. OpenAI also released an assessment by independent mathematicians, including Timothy Gowers, winner of the 1998 Fields Medal, who said no previous AI-generated proof had reached a comparable standard.

The declaration’s authors say commercial developers have strong incentives to promote breakthroughs while disclosing little about methods, training data or unsuccessful attempts. They warn that AI-generated proofs can look persuasive while containing subtle errors, and they call for disclosure of AI use, proper credit for prior research, rigorous peer review and stronger public support for academic mathematics. The International Mathematical Union has endorsed the declaration.