Mathematicians Sign Declaration Against AI Use in Their Field

Hundreds of mathematicians have put their names to a declaration calling for a halt to—or at least serious regulation of—generative AI use in mathematical research and publishing. This is not a vague complaint about the future of work: it is a concrete document, with identified signatories, directed at publishers, academic journals, and research communities. The story was reported by Scientific American in early June 2026 and quickly circulated on Hacker News.

The timing is no accident. The past eighteen months have seen a surge of preprints and papers where generative AI tools appear as writing assistants, conjecture generators, or proof verifiers. For a significant portion of the mathematical community, this raises concerns that go beyond plagiarism or authorship: they touch on the very nature of what it means to prove something.

What the Declaration Says

According to the Scientific American article, the signatories are not calling for an absolute ban. The text distinguishes between auxiliary uses—literature searches, formatting, code assistance—and uses that compromise the integrity of mathematical reasoning: generation of proof steps without rigorous human verification, writing arguments that the author cannot reproduce from memory, or presentation of AI-assisted results without explicit disclosure.

The specific demands target three areas:

• Publishers and journals: require mandatory disclosure of AI use in the proof process, not just in writing.
• Academic institutions: do not incentivize publication speed over result verifiability.
• Peer review communities: reviewers have the right to reject work whose core reasoning cannot be traced without relying on a black box.

Why This Matters More in Mathematics Than Other Disciplines

Mathematics occupies a peculiar position in this debate. Unlike experimental biology or social sciences, where AI can assist with data analysis using auditable methodologies, in mathematics the proof is the result. A proof that no one can follow step by step is not a proof: it is a claim.

Current language models—including Claude Opus 4.7 with its one-million-token window, capable of sustaining very lengthy reasoning—are good at generating text that looks like a proof. They are less reliable at guaranteeing that each step is logically valid. Formal verifiers like Lean or Coq can close that gap to some extent, but most mathematical papers do not go through them.

The practical risk flagged by the signatories is the silent accumulation of incorrect results in the literature. A faulty but plausible proof, if it passes peer review and gets cited enough times, can contaminate entire lines of research before someone spots the error.

Why This Matters Beyond Academia

That mathematicians are signing does not mean the problem is purely academic. Several companies in the financial, cryptographic, and formal verification sectors rely on mathematical results assumed to be correct. If the reliability of mathematical literature degrades, the consequences could emerge in unexpected places.

For those working with Claude or any other LLM in contexts requiring formal reasoning—engineering, applied cryptography, algorithm design—the declaration is a useful reminder: a model's fluency in writing mathematics is not a guarantee of correctness. The criterion remains human judgment, or failing that, formal verification.

Editorial Note

The declaration is not alarmism: it is a reasoned response from a community well acquainted with the limits of its own verification tools. That it arrived now, not two years ago, says something about the speed at which LLMs have entered workflows where they were not previously expected.