An AI-generated mathematical proof can pass one kind of check while leaving crucial questions unanswered. Does the argument establish the intended claim? Are its assumptions justified? Was the result already known? Who contributed the ideas, and can other mathematicians explain why it works?

Those questions sit behind Terence Tao’s discussion of “Maths 2.0”. In his 22 October Nature World View, Tao argues that a growing supply of AI-generated proofs could outpace mathematicians’ ability to absorb them. He raises concerns about attribution, presentation, training and the time needed to understand an argument. These are concerns about how mathematics develops knowledge, beyond the production of plausible manuscripts.

OpenAI’s October mathematics collection supplies a concrete reason to ask what a verification claim covers. Its own disclosures describe mixed stages of verification, and its correction history records withdrawals and repairs. A useful assessment starts with the exact statement, the checking method and the version that was checked.

OpenAI’s correction history makes the checking problem concrete

OpenAI dates the collection’s original release to 6 October 2026. In the repository README preserved at its 30 October revision, it describes the outputs as raw research data, acknowledges differing stages of verification and warns that results without formalizations in the Lean proof assistant may contain issues. It leaves questions of credit, precedence and concurrent work to the mathematical community.

That framing matters when a manuscript’s title makes a striking mathematical claim. The collection is an invitation to examination; each result needs its own evidence.

The repository’s dated history records three withdrawals on 7 October. OpenAI attributes them to a sign error that invalidated an argument and a construction used by two dependent papers. The same entry records proof repairs, corrected statements and clearer hypotheses and dependencies.

A shared flaw can therefore affect several arguments. A repaired proof might require a changed statement or an explicit condition. Tracking the dependency and revision is part of assessing what has been established. These disclosures do not provide a collection-wide error rate.

Review status also needs a date and a verb. The history’s 26 October entry says an independent working group “will examine and explain the proof” associated with the manuscript titled The Exponential Time Hypothesis is False. That is an announcement of review. The repository evidence inspected on 6 November does not establish that the review has been completed, and this article does not certify the manuscript’s conclusion.

Formal verification checks a precisely encoded claim

A conventional mathematical proof presents an argument in words, symbols and diagrams. Readers assess whether the reasoning is valid, the assumptions are appropriate and the detail is sufficient to fill in the steps.

Formal verification makes the statement and derivation precise enough for a proof checker to validate under specified rules and axioms. Lean is one system used for this work. It can reduce reliance on a reader noticing every local logical error.

The encoded statement still needs scrutiny. A checker can validate the derivation of a claim whose assumptions are stronger than readers expect, or whose definitions differ from the intended problem. Checking the correspondence between the mathematical question and its formal representation is a separate task.

Admitted gaps require attention too. The Lean reference introduction demonstrates the placeholder sorry, including an example involving a purported proof of falsehood, and shows a warning that the declaration uses it. A file being accepted with an admitted gap is insufficient evidence of a complete proof.

A useful verification report therefore identifies the statement, assumptions, checker and relevant version, and establishes that the proof contains no admitted gaps. Numerical spot checks can help find mistakes, but checking selected cases cannot by itself prove a claim about all cases.

AlphaGeometry illustrates another kind of checking

The January 2024 AlphaGeometry work shows how AI generation and symbolic deduction can cooperate. It is a historical example involving competition problems with known human solutions.

In the paper published on 17 January 2024, a language model proposes auxiliary constructions: extra points or relationships that can help unlock a geometry problem. A specialized symbolic engine then derives deductions within its geometry representation. This uses a different representation and checking process from Lean.

The paper’s illustration of International Mathematical Olympiad 2015 Problem 3 includes three auxiliary points. Its displayed proof is explicitly shortened and edited. The authors report submitting solutions to the 2000 and 2015 geometry problems to the USA IMO team coach, who recommended full scores. DeepMind’s accompanying account identifies the evaluator as Evan Chen.

That is author-reported external expert evaluation of particular outputs. This article has not audited the full supplementary proof or replicated the system, and the example provides no validation of OpenAI’s October 2026 collection.

AlphaGeometry also illustrates why representing the intended question matters. In its discussion of a generalized, translated 2004 problem, the paper distinguishes a condition that points lie on the same line from the original requirement that one point lie between two others. The generated generalized solution does not guarantee that between-ness condition. A valid deduction must be read alongside what the representation actually expresses.

Human contributions also extend well beyond writing demonstrations for training. Researchers designed AlphaGeometry’s representation, symbolic rules, synthetic-data process, training and benchmark translations, and arranged evaluation. Describing one stage as automated does not account for the whole process.

Correctness leaves novelty, credit and understanding to establish

A correct proof may offer a new route to an old result. Establishing novelty requires comparison with existing mathematics: what was already proved, under which assumptions, and whether the contribution is a theorem, a stronger statement, a different proof or a formalization.

Credit requires an account of contributions and sources. OpenAI’s README identifies an unreleased internal model and discloses human editing for readability of a specified zeta-region writeup. That is useful production information; it does not supply a complete contribution history for every result.

Understanding asks further questions. Which idea makes the proof work? Which assumptions are essential? Can the argument be adapted to a related problem? A checked derivation can support that investigation while still requiring substantial explanatory work.

In her 6 November Nature World View, Ursula Martin advocates precise AI disclosure, citation policies and independently verifiable arguments for which human authors can answer. She calls for institutional participation and explicitly values Tao’s leadership. The accessible portions of their essays offer complementary concerns about how the field should handle this work. Both Nature commentaries were only partially accessible for this article. Our guide to checking AI-generated references explains how to verify that a citation exists and supports the claim.

The Leiden Declaration offers recommendations for that responsibility: disclose tool use, pursue proper attribution, support reviewing and retain human responsibility for correctness. These are proposed standards of practice, rather than evidence that any particular proof is correct.

When a result is described as “verified”, ask for the exact artifact and revision, what was checked, by whom or by which system, and under which assumptions. Then ask separately for the evidence of novelty and the account of human and AI contributions. A useful answer should let another reader locate the checked claim and follow the work behind it.