Mathematicians Urge Caution After AI Disproves Conjecture

According to reporting by The New York Times, OpenAI announced in late May that one of its proprietary models had disproved an 80-year-old conjecture in combinatorial geometry originally posed by Paul Erdős. The New York Times reports that OpenAI released a research paper describing the result and a separate paper by several independent mathematicians reviewing the work. The New York Times quotes Jacob Tsimerman of the University of Toronto saying, "This is a really impressive piece of work, and I would accept it for any journal without hesitation." The New York Times also reports that Melanie Matchett Wood of Harvard noted the OpenAI-authored paper omitted appropriate references to a "history of closely related ideas in the literature." Reporting by The New York Times and commstrader says a group of 16 mathematicians, in consultation with international math organizations, published the Leiden Declaration on Artificial Intelligence and Mathematics to "frame the conversation about future directions," according to Dame Ursula Martin, one of the authors, as reported by The New York Times.

Industry observers note that large, opaque models can produce novel, publishable-seeming mathematical arguments while lacking traceable citations or human-style exposition. This pattern elevates practical verification challenges: formalization in proof assistants, reproducibility of computational steps, and independent peer review become central technical requirements. For practitioners, integrating formal-verification workflows, standardized benchmark tasks for symbolic reasoning, and tooling that exposes intermediate proof structure are immediate areas of interest.

Industry context: The combination of a high-profile model result plus a collective declaration from established mathematicians crystallizes tensions between accelerating automated discovery and preserving mathematical norms of citation, explanation, and reproducibility. Journals, referees, and research groups may need clearer policies about attribution, independent verification, and the evidentiary standards required for algorithm-generated proofs. Observers following the sector will watch whether leading journals adopt mandatory formal-checking or require authors to supply machine-verifiable artifacts.

Indicators include whether subsequent papers reproduce the OpenAI result via independent code and formal proofs, whether major journals update submission policies for AI-origin proofs, whether the Leiden Declaration gains endorsements from mathematical societies, and whether model developers release more transparent training or chain-of-thought artifacts for high-level mathematics. For practitioners: tracking uptake of proof assistants, benchmarks for automated theorem proving, and community-led standards for citation and provenance will matter for reproducible research.