Leiden Declaration Warns AI Puts Mathematical Integrity at Risk

When more than 2,654 mathematicians put their names to a document warning about AI, the argument is worth reading on its own terms. The Leiden Declaration on Artificial Intelligence and Mathematics, released June 2, 2026 and endorsed by the International Mathematical Union, is not a blanket rejection of AI tools. It is a careful enumeration of what properties make mathematics trustworthy and which AI-driven pressures are currently putting those properties at risk.

Mathematics is unusual among intellectual disciplines in that its outputs, proofs, are supposed to settle questions rather than just advance them. The declaration identifies five properties worth protecting: proof as a foundation for certainty and understanding, clear attribution and accountability, transparency that allows independent verification, shared community standards for evaluating significance, and researcher autonomy to choose problems based on intellectual merit rather than commercial viability. AI systems that generate plausible but incorrect mathematical arguments, the declaration argues, undermine all five simultaneously by producing results that look right and are difficult to check.

Beyond bad proofs, the declaration names a cluster of structural risks: attribution collapse when AI tools trained on published works fail to properly cite sources or use data obtained without consent, incentive distortion when research directions bend toward AI capabilities rather than significant open problems, and peer review bypass when results get announced through press releases rather than vetted publication. These are claims that current AI adoption is creating pressures no individual mathematician can counter alone, which is why the declaration makes separate recommendations to individual researchers, publishing organizations, policymakers, and commercial AI developers.

The honest caveat is that the declaration does not address enforcement. The recommendations are clear in direction but silent on mechanism: who decides whether a tool disclosure is adequate, or whether a journal's AI guidelines actually meet community standards. The source does not indicate whether any major journals have committed to the proposed guidelines, so for now this remains a statement of values with institutional backing rather than a set of binding rules.

The 2,654 signatories, which include Fields Medal winner Terence Tao and algebraic geometer Peter Scholze, along with the International Mathematical Union's endorsement, give the declaration enough weight that funders and publishers will have difficulty treating it as fringe opinion. For AI developers building tools for mathematical research, the declaration is also a specification: the mathematical community is telling you plainly what trustworthy assistance would have to look like.