When AI Makes Deep Theorems No Longer Scarce: Mathematics Needs to Recalibrate What It Values

AI lowers the cost of producing sophisticated proofs, making deep theorems no longer scarce and weakening the old mechanism of using theorem production to identify deep thought. The author frames this as a shock to mathematics' evaluation system rather than a purely technical advance, noting that 'the machines are producing solutions faster than the mathematical community can read them, much less digest them.' This is an observational account based on the author's experience: she received several purported proofs of the Nivat conjecture this week, some authors acknowledged using AI only for polishing English or checking the proof, and none accepted her Zoom invitation to explain their arguments.

The Nivat conjecture is a long-standing problem at the intersection of combinatorics and dynamical systems; the author and Van Cyr proved a partial result in 2012, and Kari and Szabados later developed an algebraic approach, leaving a missing conceptual bridge between the two methods. The author uses her own research history to convey the problem's difficulty and the coexistence of two technical routes, and notes that frontier models excel at synthesizing different approaches, which is why the conjecture has been proposed for Google DeepMind's Formal Conjectures project. This is a retrospective account of the author's research lineage, including the 2012 partial result and the Kari–Szabados method, but it presents no complete proof or new theorem.

A proof is more than a certificate that something is true; it is a story, a picture, an insight, an explanation that opens new directions and enters the community's toolkit. The author cites Bill Thurston's 2010 MathOverflow statement that 'The product of mathematics is clarity and understanding. Not theorems, by themselves,' and applies this stance to evaluating AI-generated proofs. This is a normative argument supported by quotation rather than experimental data; its force comes from the author's experience in the mathematical community and her definition of what proofs do.

Existing journal publication, hiring and prize structures, and peer review relying on a small group of overworked reviewers cannot handle the volume AI brings; the roles of discovery, proof, formalization, and explanation need to be distinguished and credited. The author argues that 'producing a paper is no longer enough': authors must be able to explain the proof's mechanism and how they arrived at this point, and exposition must become an important component of any major intellectual achievement. This is a diagnosis of institutional conditions and a reform proposal, based on the author's observations about reviewer burden, incentives for young researchers, and reduced time for deep reflection.

This article is aimed at the mathematical community and readers concerned with AI's effect on knowledge production; it applies to a setting where AI can generate polished-looking proof manuscripts while evaluation still centers on theorem output. The author welcomes AI-generated proofs, believes human-only proofs will become rare, and argues for incorporating AI responsibly into mathematical practice.

The author states plainly that 'a proposed proof is not a theorem,' so whether the purported proofs of the Nivat conjecture are correct remains unknown; readers should still watch the verifiability of AI-generated proofs, whether authors can explain a proof's mechanism, and how the community can distinguish discovery, proof, formalization, and explanation without drowning reviewers.