- Leiden Warning: Mathematicians behind the Leiden Declaration are seeking rules for AI use in proof review and verification.
- Proof Risks: Automated systems can produce plausible but unreliable arguments, while proprietary models can complicate credit and auditing.
- Commercial Pressure: Named mathematicians argue commercial incentives and opaque evaluations can push research systems toward speed over verifiable proof.
- Field Response: The declaration already has 1,854 signatories and sets 23 recommendations before a July mathematics congress discussion.
Mathematicians behind the Leiden Declaration have put a public warning around AI’s rapid move into proof work. They argue, automated systems can produce plausible but unreliable mathematical arguments, creating pressure on peer review and independent verification if institutions treat speed as a substitute for verifiable proof. Its signatory page showed 1,854 signatories at time of writing, giving the warning early support beyond the original sixteen mathematics specialists who released the declaration.
AI systems are moving from benchmark math tasks toward work that touches proof itself. Most recently, a general-purpose reasoning model by OpenAI produced a new construction for the planar unit distance problem, pushing proof-related capability debates beyond specialist circles.
For mathematicians, the risk is institutional as much as technical. Faster automated reasoning can create more results to verify, more credit disputes to resolve and more pressure on systems that still depend on human judgment.
How AI Tests Proof, Credit and Review
The Leiden Declaration authors set out principles for using AI in mathematics without treating the technology as a replacement for proof culture. Their framework for proof and attribution identifies independent verification, shared standards and human judgment as values that make mathematics more than answer production.
Mathematical papers contain arguments that other researchers must be able to inspect, reproduce and challenge. Disclosure and responsibility become safeguards for the publication record when AI systems help generate or shape a result.
AI models can produce plausible but unreliable arguments that are difficult to distinguish from correct mathematical proofs. Fast-moving model outputs put review systems under pressure and jeopardize standards for correctness, transparency and independent verification.
A 2025 research-publishing precedent shows the operational version of that concern: arXiv tightened computer science submission rules amid pressure on peer review from low-quality AI-generated material.
Leslie Ann Goldberg, head of computer science at the University of Oxford, put the review problem in practical terms.
“Inaccurate AI-generated drafts are cheap to produce, and there is a risk of cluttering the literature with claimed results that are simply wrong.”
Leslie Ann Goldberg, head of computer science at the University of Oxford (via Ars Technica)
Proof checking remains labor-intensive even when software helps. Human proof checking keeps AI-generated conjectures and strategies dependent on people who can find errors.
Attribution creates a second pressure point. Results produced with trained proprietary models can make proper credit harder, especially when institutions reward speed without accounting for how results were produced.
Early-career researchers are especially exposed because reputations can depend on priority disputes, careful citation and the ability to explain how a result was obtained. Speed-based incentives may push researchers toward opaque tools even when the proof trail is harder to audit.
Why Commercial AI Has Mathematicians Uneasy
Commercial incentives sharpen the concern. Michael Harris, a Columbia University mathematician and declaration co-author, warns of “commercial logic” pushing against mathematical values. Kevin Buzzard, a mathematician at Imperial College London, called it striking that “tech companies are suddenly interested in their work”.
OpenAI’s recent geometry achievement involved a model-produced construction for the planar unit distance problem, a geometry problem first posed by Paul Erdős in 1946. Proprietary systems can make independent evaluation harder, turning the achievement into a disclosure problem as well as a technical milestone.
AI researcher and commentator Gary Marcus urged to view OpenAI’s work with skepticism, pointing to fine-print limits in how the system was tuned and evaluated.
Ethical control extends beyond publication mechanics. Mathematics also has applications in technology for warfare, oppression, mass surveillance and the undermining of democracy. Possible downstream use is the concern, not proof that current mathematical research has already produced those outcomes.
What the Declaration Asks For Next
Rather than banning AI, the declaration calls for rules. Its 23 recommendations are also circulating as a package for individual mathematicians, mathematical organizations, research funders and policymakers.
Their recommendations include disclosing the use of AI in research, ensuring papers are peer-reviewed, preparing guidelines for AI use in publication and review, and investing in public AI research laboratories.
Institutional support from the International Mathematical Union adds weight to the argument that mathematical research should remain guided by human judgment, transparency and shared disciplinary values. A July International Congress of Mathematicians discussion in Philadelphia could guide the field to decide which disclosure rules and public research support should govern AI proof systems before they become routine research tools.
Leiden_Declaration_on_Artificial_Intelligence_and_Mathematics
