More than 150 mathematics professors from institutions across the world have signed a formal declaration warning that artificial intelligence companies are systematically overstating the mathematical capabilities of their models — and that governments, funding bodies, and universities are making consequential decisions on the basis of those inflated claims.

The document, known as the Leiden Declaration, was coordinated out of Leiden University and presented to bodies including the International Mathematical Union. Its signatories are not peripheral critics of technology. They are working research mathematicians — people whose professional lives consist of constructing and verifying formal proofs — and their core message is blunt: the AI systems being marketed as capable of doing mathematics are not doing mathematics in any meaningful sense of the word.

The central technical concern is what the field calls "hallucination" in a mathematical context. Large language models can produce output that is formatted like a proof, uses the correct symbols, and follows the superficial grammar of mathematical argumentation — and is nonetheless wrong. Not slightly wrong. Structurally, foundationally wrong, in ways that a competent graduate student would catch but that a non-specialist reviewer, a journalist, or a government policy advisor would not. The declaration states explicitly that AI systems have produced incorrect proofs that were initially mistaken for valid work, creating downstream risks for the integrity of the published literature.

There is a commercial logic to this that the declaration names directly. The companies building and selling these systems have financial incentives to claim frontier mathematical capability. "Solving mathematics" or "achieving superhuman reasoning" are phrases that move stock prices, attract venture capital, and — critically — influence the policy positions of governments deciding how to regulate the sector, or whether to regulate it at all. The signatories argue that what is being sold as a research breakthrough is in many cases a sophisticated pattern-matching system that has ingested a large corpus of mathematical text and learned to mimic its surface features.

The declaration makes a specific demand for transparency: that AI systems used in or around mathematical research be subject to rigorous, independent verification of their claimed capabilities, with results published in a form that the mathematical community — not the company's communications department — can scrutinize. This is not a fringe ask. It mirrors what any peer-reviewed journal requires of a human researcher submitting a claimed proof. The industry has so far operated under no equivalent standard.

There is also a structural concern about credit and the future of the discipline. Mathematics builds on itself. Attribution — knowing who proved what, under what constraints, using which methods — is not vanity. It is the epistemological scaffold of the field. When AI-generated content enters the pipeline without clear disclosure, it corrupts that scaffold. Junior researchers whose careers depend on original contributions find their work devalued or obscured. The declaration calls for explicit policies at journals, conferences, and funding agencies to require disclosure of AI involvement in any submitted work.

What makes this moment distinct from previous rounds of AI skepticism is the institutional weight behind it. These are not technophobes or Luddites. Several signatories have themselves published research on the legitimate uses of computational tools in mathematics. Their objection is not to AI as a category. It is to the specific practice of claiming capabilities that current systems do not possess, and to the political and economic pressure that those claims are being used to generate. The International Mathematical Union, one of the oldest and most authoritative bodies in the discipline, has been formally presented with the declaration — a signal that this is not a fringe protest but a coordinated intervention by the profession's mainstream.

The timing is not accidental. Multiple AI laboratories have made high-profile announcements in the past eighteen months claiming their models can solve olympiad-level problems, generate novel proofs, or assist in frontier research. Some of those claims have been walked back quietly after scrutiny. Others remain in circulation, cited in regulatory submissions, congressional briefings, and grant applications. The Leiden signatories are, in effect, saying that the peer review the industry has avoided needs to happen — and that if the industry won't impose it on itself, the mathematical community will begin imposing it from the outside.

The harder question — the one the declaration does not fully answer — is what enforcement looks like. Mathematics journals can set their own policies. Funding agencies can attach conditions to grants. But the companies making the most aggressive capability claims are primarily accountable to their investors, not to the International Mathematical Union. The declaration is a shot across the bow. Whether it lands depends on whether governments treat it as expert testimony or file it in the same drawer as every other letter from concerned academics.