Artificial intelligence now assists mathematical research along two very different roads, and the difference between them determines whether a headline about an AI breakthrough means anything. One road produces human-readable arguments that can contain subtle logical errors; the other produces machine-checkable code in a formal language such as Lean, where a compiler verifies every logical step. Recent high-profile claims by OpenAI—including a claimed proof of the local smoothing conjecture—illustrate both roads at once, and show why the mathematics community treats formal verification and human understanding as separate achievements.
AI's Dual Approach to Mathematical Proofs
Why can't mathematicians simply ask an AI for a proof and be done with it? The core problem is reliability. A research paper on AI-driven formal proof search notes that while large language models have shown remarkable promise on complex mathematics, unreliability remains the primary barrier to integrating them into research. Mistakes in unreviewed intermediate steps can cascade through a proof, limiting the complexity of tasks that can safely be delegated to AI.
The field's answer is to split the work. AI can generate proofs in natural language, the prose mathematicians traditionally write—or it can generate proofs in formal languages like Lean, in which a compiler automatically verifies every logical step. These two approaches carry sharply different levels of trust, and most serious recent efforts combine them.
The Role of Large Language Models (LLMs)
LLM-generated natural language proofs have an obvious appeal: they are intuitive, readable, and resemble how mathematicians actually think and write. As Carnegie Mellon University researcher Sean Welleck put it, neural models learn patterns from data, which allows them to generate conjectures, suggest proof strategies, and work in ways that resemble human intuition.
But that flexibility comes at a cost. The formal proof search paper warns that LLM-generated natural language proofs can contain subtle logical errors, or "hallucinations," and therefore require expensive expert review. A plausible-sounding argument is not a proof; a single unspotted gap in an intermediate step can invalidate everything built on top of it. This is why an unreviewed AI argument, however elegant, remains a speculative claim rather than established mathematics.
Formal Verification with Proof Assistants like Lean
Lean offers a different bargain. It is a theorem prover—software that checks a proof step by step and refuses to accept any step it cannot verify. If the code compiles, the proof is considered complete. This shifts the burden of checking from overworked human referees to a machine that cannot be persuaded by a fluent but flawed narrative.
Two qualifications matter, though. Lean verifies the syntactic compliance of reasoning steps; it does not automatically guarantee that a proof is insightful, or that its mathematical content is free of loopholes in the sense of addressing the question mathematicians actually care about. And as Carnegie Mellon's reporting emphasizes, the productive direction is combining neural models with proof assistants—pairing AI's pattern-matching flexibility with the certainty of formal verification rather than relying on either alone.
Case Study: OpenAI and the Local Smoothing Conjecture
OpenAI's October 2026 mathematics release shows the full pipeline in action. According to reporting on the announcement, OpenAI's model generated a natural language proof of the local smoothing conjecture—a problem proposed by Christopher Sogge in 1991—and the proof was formalized in Lean the following day, October 19. OpenAI published the formal proof on GitHub and also released a natural language explanation, so human mathematicians do not have to read raw Lean code to follow the argument.
The same pattern recurred with related claims. Reporting on OpenAI's subsequent results indicates the company obtained AI output on October 18, formalized it in Lean the next day, and then contacted an expert to discuss forming a collaboration around the result. For its computer science claim—a disproof of the Exponential Time Hypothesis—OpenAI invited the scientists who originally proposed that hypothesis, Russell Impagliazzo and Ramamohan Paturi, to participate in reviewing the AI's work. Crucially, these claims remain under review by the mathematics community; formalization in Lean does not replace traditional peer review.
Evaluating AI-Generated Mathematical Breakthroughs
How should you judge the next announcement of an AI-proved theorem? The evidence suggests four questions, and the matrix below turns them into a reusable assessment tool. The central distinction is between an unverified LLM output, a formally checked proof, and a formally checked proof that human experts have actually understood and explained.
| Claim type | Generation method | Verification status | What it means |
|---|---|---|---|
| Raw natural language output | LLM-generated prose proof | Unchecked; may contain hallucinations or subtle logical errors | A speculative claim requiring expensive expert review |
| Formally verified proof | AI-generated code in Lean | Compiler checks every step and refuses unverifiable ones | Formally correct, but not automatically insightful or understood |
| Verified and human-interpreted result | Formal proof plus human-authored explanation | Formally checked and reviewed by experts, with the argument made comprehensible | A genuine contribution to human mathematical understanding |
Applying this to the OpenAI case: the local smoothing claim sits between the second and third rows. The Lean formalization compiled, and a natural language explanation was published—but community review is ongoing, and understanding why the proof works is a separate, unfinished task. The framework's value is precisely that it prevents a compiled Lean file from being mistaken for a completed mathematical contribution.
The Evolving Role of the Human Mathematician
Does a verified AI proof make human mathematicians obsolete? The recent record suggests the opposite. NPR reported in September 2026 that when OpenAI beat humans to a solution of the Navier-Stokes problem, mathematicians said the AI's 166-page solution wasn't telling them much—even as few believed it was wrong, because the accompanying Lean formalization compiled and, as mathematician Javier Gómez-Serrano told NPR, "the community seems to have the consensus that it is correct."
Correctness and understanding are different outputs. As one analysis of OpenAI's claims argued, counting solved statements measures one output, while understanding why they are true, how their methods relate to earlier work, and where they can be extended measures a different output—and a scientific community needs both. A shorter conceptual explanation, that analysis noted, might itself become a major contribution because it tells other researchers which parts of the machinery can be adapted.
This is where the human role is being redefined rather than eliminated. Interpreting, refining, and explaining AI-generated proofs—turning a compiled artifact into transferable insight—is a major intellectual contribution in its own right. OpenAI itself signaled this shift in an October 15, 2026 update, stating it wants to support the mathematics community in turning AI progress into deeper human understanding and new discoveries.
Integrating AI: The Future of Mathematical Discovery
The decision for the field is clear: mathematicians should integrate AI tools as powerful assistants for proof generation and verification, while maintaining their essential role in human interpretation, refinement, and the pursuit of deeper understanding. The two-track model—LLMs for intuition and conjecture, Lean for certainty—works only when humans supply the meaning.
The measurable indicator to watch is the acceptance and citation of AI-assisted proofs in peer-reviewed mathematical literature, accompanied by human-authored explanations and insights. That, not a compiled Lean file or a corporate blog post, is the point at which an AI-generated argument becomes mathematics.
Sources
- From Proof to Program: CMU and the Rise of AI-Driven Mathematics — Carnegie Mellon News
- Advancing Mathematics Research with AI-Driven Formal Proof Search — ARXIV
- OpenAI Says Its Models Produced Three Proofs in Math, CS, and Physics — Unite.AI
- OpenAI says its models proved the local smoothing conjecture in every dimension — Crypto Briefing
- OpenAI’s Science Claims: Kakeya, ETH and Quantum Fields Explained — Kingy AI
- AI solved one of math's hardest problems. Humanity learned nothing (so far) — NPR
- OPENAI
