Mathematicians Publish Leiden Declaration on AI Risks to Mathematics

On June 2, 2026, mathematicians published the Leiden Declaration on Artificial Intelligence and Mathematics, a statement of values and recommendations for the mathematical community developed after a 2025 workshop at the Lorentz Center in Leiden, per the declaration website and Leiden University. Leiden University says the declaration was issued by 16 researchers from 15 universities and is endorsed by the International Mathematical Union, and Gizmodo reported it had attracted more than 130 signatories by the time of publication. The London Mathematical Society and Scientific American summarized the declaration's core concerns as the reliability of automatically generated results, attribution of work that uses proprietary models, and effects on publication and peer review.

The text centers on the role of proof in mathematical practice and the norms of attribution and responsibility for correctness, per the declaration preamble. It discusses both symbolic and neural methods for generating and formalizing mathematics and highlights tensions between rapid public dissemination of AI-generated claims and established procedures for validation, according to the declaration and reporting by Scientific American and Gizmodo. Rather than calling for a ban, the declaration recommends that authors disclose AI assistance, that results undergo peer review, and that funders and professional bodies develop policies to preserve scrutiny and accessibility, as described in the London Mathematical Society summary and on the declaration site. The Isaac Newton Institute likewise frames the effort as seeking guardrails rather than prohibition.

Industry-pattern observation: the issues the declaration raises, reproducibility and attribution, are the same technical-governance fault lines that recur whenever powerful automation enters a research discipline. If journals and funders begin to require disclosure of AI use and verifiable artifacts, demand tends to grow for tooling that emits machine-checkable proofs, for example formalization in proof assistants, and provenance metadata that records how a result was produced. Collective, field-level standards frequently follow workshops and declarations of this kind, which can accelerate editorial and funding guidelines even before formal regulation exists. A practical tension is that proprietary models complicate attribution and reproducibility precisely because their training data and internal reasoning are not open to inspection, so disclosure norms only partially resolve the verification problem.

Useful signals include whether major mathematics journals and funding bodies adopt formal AI-disclosure requirements, whether the International Mathematical Union or national societies issue implementing guidance, and whether adoption of proof-assistant formalization and provenance tooling rises in response. Growth of the signatory list well beyond the initial 130-plus, and uptake among large professional bodies, would indicate how much consensus the declaration commands across the discipline.