AI and Science

Fields Medalists Ask AI Labs to Preserve What a Proof Is For

By Kaleido Field Staff · September 13, 2026

Citation-ready: The Fields medalists' declaration concerns understanding, attribution and transmission of mathematics, rather than a refutation of one named proof.

Evidence boundary: A public professional declaration, not an independent audit of a particular theorem or a research-paper result. We did not endorse the declaration or submit a signature.

What happened and why it matters

The declaration supplies a concrete professional response to rapid AI proof production, expanding the coverage question from whether an artifact checks to whether its ideas become usable knowledge.

Original source

Primary reference: Terence Tao's dated statement and Math and AI declaration. Kaleido Field checked the event date and the article's attributed facts against this source.

The objection is not to every use of AI

The signatories recognize the potential for AI to aid mathematical study. Their concern is that rapid announcements can outrun careful exposition, credit to earlier work and the human process through which an idea is understood and taught.

That distinction prevents a false binary. A reader can value automated checking while still asking for an explanation of the key construction. Conversely, a readable explanation does not excuse a gap in a formal argument. The two forms of scrutiny address different properties of the work.

Report the result and the route separately

For coverage of a new theorem claim, a useful record would identify the exact statement, the checkable artifact, the earlier work it builds on and the part of the method that is new. A separate account can explain who has examined it and what remains difficult to communicate.

Our Fermat formalization report distinguishes a checked formalization from a new mathematical discovery. The declaration adds a further question: how will the resulting ideas enter the literature and the classroom? It should not be used as evidence that the earlier formalization is invalid.

A score cannot carry the whole research story

A solved-problem count compresses work into an outcome. It cannot show, on its own, whether a technique generalizes, whether another researcher can adapt it or whether the attribution is complete.

Those omissions suggest additional reporting, not an invented replacement ranking. Follow-up coverage can examine a specific exposition, a reusable method or an attribution correction. This article records the September 11 statement; it does not infer that all signatories agree on every proposed AI policy.

Evidence boundary

A public professional declaration, not an independent audit of a particular theorem or a research-paper result. We did not endorse the declaration or submit a signature.