TL;DR
A new declaration signed by leading mathematicians and AI researchers warns that the unregulated use of large language models in mathematical research could erode the rigor and trust that underpin the entire field. The warning comes as AI systems have begun co-authoring published papers and generating novel proofs, raising urgent questions about verification and standards.
What Happened
More than 150 mathematicians and computer scientists from institutions including the Fields Institute, the Clay Mathematics Institute, and leading AI labs signed the "Montreal Declaration on Mathematical Integrity in the Age of AI" on Monday, June 1, 2026. The document, released jointly by the International Mathematical Union (IMU) and the Partnership on AI, calls for explicit disclosure requirements when AI tools are used in mathematical research, and warns that unchecked automation could "undermine the foundational trust that mathematics requires to function as a discipline."
Key Facts
- The Montreal Declaration was signed by 157 researchers from 23 countries, including three Fields Medalists and two Turing Award winners.
- The declaration identifies four core risks: AI-generated "hallucinated" proofs that appear valid but contain logical errors; erosion of peer review as AI-written papers flood journals; loss of human intuition in proof discovery; and concentration of verification power in proprietary AI systems.
- A 2025 survey by the European Mathematical Society found that 41% of mathematicians under age 35 had used AI tools in their research, but only 12% disclosed that use in their publications.
- The declaration calls for mandatory disclosure of AI assistance in all mathematical submissions to journals and conferences, modeled on existing disclosure rules in computer science and medicine.
- Google DeepMind and OpenAI both issued statements supporting the declaration's goals, though neither committed to specific enforcement mechanisms.
- The IMU plans to present the declaration to the International Congress of Mathematicians in July 2026 in Paris, where it could become official IMU policy.
- A pre-print study from MIT released alongside the declaration found that current large language models produce plausible-looking mathematical proofs with a 32% error rate that expert reviewers failed to catch in a blind test.
Breaking It Down
The Montreal Declaration is not a Luddite manifesto—it explicitly acknowledges that AI can accelerate discovery and help mathematicians explore complex spaces. Instead, it targets a specific, growing problem: the gap between how quickly AI is being adopted in mathematical practice and how slowly the field is adapting its verification and publication standards. The signatories argue that mathematics, unlike some empirical sciences, has no "replication crisis" to fall back on—a proof is either logically valid or it is not, and AI introduces a new class of error that existing review processes were not designed to catch.
In the MIT blind test, expert reviewers failed to detect errors in 32% of AI-generated proofs—a failure rate that, if extrapolated to the broader literature, would mean thousands of flawed results have already been published.
That statistic is the declaration's most damning data point. The MIT study used GPT-5 and Claude 4 to generate proofs in algebraic topology and number theory, then asked 48 professional mathematicians to review them alongside human-written proofs. The 32% miss rate was consistent across both AI systems and all subfields tested. The signatories argue this is not a temporary problem: as AI models improve, their outputs become more convincing to human reviewers, even as the models remain capable of subtle, logical errors that are harder to spot than the obvious mistakes of earlier systems.
The declaration also addresses a structural concern: the concentration of mathematical verification power in proprietary systems. If the most capable proof-checking and proof-generation tools are controlled by a handful of private companies—OpenAI, Google DeepMind, Anthropic—then mathematics risks becoming dependent on black-box systems whose internal logic is not transparent. The signatories call for open-source verification tools and public benchmarks for AI mathematical reasoning, modeled on existing efforts in computer vision and natural language processing.
What Comes Next
-
July 2026 – International Congress of Mathematicians in Paris: The IMU will put the Montreal Declaration to a formal vote. If adopted, it would become the first global policy on AI in mathematics, binding on IMU-member societies in over 80 countries. The vote is expected to be close, with opposition from researchers who argue the declaration is too restrictive.
-
September 2026 – Journal policy changes: Four major mathematical journals—Annals of Mathematics, Inventiones Mathematicae, Journal of the American Mathematical Society, and Compositio Mathematica—have already announced they will require AI disclosure statements starting with submissions received after September 1, 2026. The journals are coordinating on a standard disclosure format.
-
Late 2026 – Open-source verification benchmark: A consortium led by MIT, the Max Planck Institute for Mathematics, and DeepMind plans to release a public benchmark dataset of 10,000 AI-generated proofs with known error types, designed to train and test automated verification systems. The dataset is expected by December 2026.
-
2027 – Potential NSF and ERC funding requirements: The U.S. National Science Foundation and the European Research Council are both considering adding AI disclosure requirements to grant applications in mathematics, following the declaration's recommendations. Formal rulemaking could begin in early 2027.
The Bigger Picture
The Montreal Declaration is the latest example of a broader professional standards movement sweeping through academic disciplines as generative AI becomes embedded in research workflows. In 2025, the Association for Computing Machinery adopted disclosure rules for AI in computer science papers; the American Physical Society followed suit for physics in early 2026. Mathematics is now catching up, but with a unique twist: because mathematical truth is not empirical, the consequences of AI-generated error are different. A flawed AI proof that gets published could misdirect years of subsequent research, with no experimental results to correct it.
The declaration also reflects a growing tension between open science and proprietary AI development. The signatories explicitly call for open-source verification tools, positioning mathematics as a test case for whether academic disciplines can maintain independence from the AI industry. If mathematicians succeed in setting their own standards for AI use, it could provide a template for other fields—from law to history—that are grappling with similar questions about trust, verification, and the role of human judgment.
Key Takeaways
- [Montreal Declaration signed]: Over 150 leading mathematicians and AI researchers have called for mandatory disclosure of AI use in mathematical research, warning that undetected AI errors could undermine the field's foundations.
- [32% error detection failure]: A new MIT study found that expert mathematicians failed to catch errors in nearly one-third of AI-generated proofs, a rate the declaration's signatories call unacceptable.
- [July 2026 vote]: The International Mathematical Union will vote on adopting the declaration as official policy at the International Congress of Mathematicians in Paris, with major journals already planning to require AI disclosure.
- [Open-source verification push]: The declaration explicitly calls for public, transparent AI verification tools, framing mathematics as a test case for academic independence from proprietary AI systems.
