The Proof Machine: OpenAI Claims Its AI Has Produced Formally Verified Proofs of Three Mathematical Breakthroughs — Including the Yang–Mills Millennium Prize Problem

1. The Claims, Precisely Stated

OpenAI announced its results on 27 October via a formal research release and concurrent media briefing. [Established — Al Jazeera, “AI makes discoveries in mathematics, computer science, physics, OpenAI says,” 27 October 2026; QZ, “OpenAI said its AI cracked major problems in math, computing, and physics,” 27 October 2026.] Three results were presented: a proof of the local smoothing conjecture in harmonic analysis, generalised to all spatial dimensions; a disproof of the Exponential Time Hypothesis for a specific class of Boolean formulas; and a proof of the existence of pure Yang–Mills theory, bearing on the Yang–Mills existence and mass gap Millennium Prize Problem. All three were verified in Lean.

Each result operates in a different mathematical domain. The local smoothing conjecture belongs to real analysis — the study of how functions behave over continuous space. The Exponential Time Hypothesis is a conjecture in computational complexity theory, bearing on how long it takes to solve certain classes of problems. Yang–Mills theory is a framework from mathematical physics, underpinning the standard model of particle physics at the level of quantum field theory. The three claims have in common only their origin — an AI system — and their verification method: Lean.

2. The Lean Standard: What Formal Verification Does and Does Not Guarantee

Lean is an interactive theorem prover: a software system that mechanically checks whether a sequence of logical steps constitutes a valid proof given a set of axioms. [Established — Lean documentation; MindStudio, “OpenAI’s Three AI Science Claims: What Was Actually Proven?,” 27 October 2026.] When a proof is verified in Lean, it means that every step of the argument follows logically from the prior step, and that the argument rests on explicit foundational axioms. It does not mean that the problem is solved in the full sense required by the mathematical community — it means that the formalisation of the argument is correct.

This distinction matters. A proof must be formalised before Lean can check it. The formalisation itself — the translation from mathematical intuition and conventional notation into the formal language Lean accepts — requires choices about how concepts are defined, and those choices can introduce errors or fail to capture the full generality of a claim. If the formalisation does not accurately represent the mathematical question being asked, Lean’s verification proves something slightly different from what was intended.

That said, Lean verification is considerably stronger evidence than a proof that merely “looks convincing.” Several famous mathematical errors — Alfred Kempe’s 1879 false proof of the four-colour theorem, Vopěnka’s retractions — would have been caught by Lean had the formalisation been done correctly. The computer-assisted proof of the four-colour theorem (Appel and Haken, 1976) required years of expert debate before the community accepted it. The formal verification of the Kepler conjecture took more than a decade. Lean verification accelerates but does not conclude that process. [Assessed with high confidence — standard history of formal methods in mathematics.]

3. The Harmonic Analysis Result: Local Smoothing in All Dimensions

The local smoothing conjecture asks how much regularity is gained when averaging a function over curved surfaces — in practice, over spheres or their generalisations. The conjecture, formulated by Thomas Wolff in 1995, predicts that integrating over spheres of varying radii improves the regularity of the function by a specific amount, depending on the spatial dimension. [Assessed — standard formulation of the local smoothing problem in harmonic analysis; claim attributed to standard mathematical literature.]

Previous work had established the conjecture in three spatial dimensions. The AI’s claimed contribution is a proof valid in every spatial dimension. [Established — Al Jazeera, 27 October 2026; Analytics Insight, “OpenAI Reveals New Discoveries in Mathematics, Computer Science, Physics,” 27 October 2026.] If correct, this is a genuine generalisation — not merely a verification of known results, but an extension of the problem’s resolution to the full class of cases. The applications are downstream in the study of wave equations and dispersive partial differential equations, where local smoothing estimates underpin regularity theory. For specialists in analysis, this would be a significant result. [Assessed with moderate confidence — based on established significance of local smoothing in the harmonic analysis literature; full verification pending expert review.]

4. The Complexity Theory Result: The ETH and a Crucial Caveat

The Exponential Time Hypothesis — formulated by Impagliazzo and Paturi in 1999 and extended by Impagliazzo, Paturi, and Zane — states that 3-SAT cannot be solved in sub-exponential time. [Established — Impagliazzo, Paturi, and Zane, “Which Problems Have Strongly Exponential Complexity?,” Journal of Computer and System Sciences, 2001. Tier 1 primary source.] The ETH is not a proven theorem; it is a widely believed conjecture that underpins a large body of complexity-theoretic lower bounds. If it were false in its full generality, many established results about the hardness of NP-complete problems would require revision.

OpenAI’s claim is more specific: it disproved the ETH for fixed-width CNF formulas, via a deterministic algorithm that counts satisfying assignments in subexponential time and polynomial space. [Established — Al Jazeera, 27 October 2026; QZ, 27 October 2026.] This is an important caveat. Fixed-width CNF is a restricted class of Boolean formulas — formulas where every clause has a fixed number of literals. The full ETH applies to k-SAT formulas without width restriction. A result limited to fixed-width formulas does not disprove the ETH as it is standardly stated. It demonstrates that within this restricted class, subexponential computation is possible — a meaningful contribution, but not the demolition of the ETH headline formulation.

The distinction matters structurally. The full ETH, if it holds, protects cryptographic constructions and lower bounds that practitioners in complexity theory and theoretical cryptography rely on. A disproof of the full ETH would be among the most consequential results in theoretical computer science. The fixed-width result is narrower — significant for specialists but not existentially destabilising for complexity theory as currently practised. [Assessed with high confidence — standard complexity-theoretic analysis of the ETH and its scope.]

5. The Yang–Mills Claim: The Most Consequential Statement

The Yang–Mills existence and mass gap problem is one of seven Millennium Prize Problems identified by the Clay Mathematics Institute in 2000, each carrying a prize of one million US dollars. [Established — Clay Mathematics Institute, “Millennium Problems,” 2000. Tier 1 primary source, claymath.org.] Only one Millennium Prize has ever been awarded: the Poincaré conjecture, resolved by Grigori Perelman in 2003 and fully verified by the mathematical community by 2006. Perelman declined the prize.

The Yang–Mills problem has two parts. The first asks whether a four-dimensional Yang–Mills theory — the mathematical framework that underpins the quantum field theories describing the strong and electroweak nuclear forces — can be constructed in a mathematically rigorous way. The difficulty is that physicists have used Yang–Mills theories successfully for fifty years in calculations, but the theories have never been given a rigorous mathematical foundation. The second part asks whether such a theory necessarily has a “mass gap”: a minimum positive energy for any particle state, which explains why the strong nuclear force is short-range despite being mediated by massless gluons in classical field theory.

OpenAI claims to have produced “mathematically rigorous” proof of the existence of pure Yang–Mills theory — the first part of the problem. [Established — Al Jazeera, 27 October 2026; QZ, 27 October 2026.] The mass gap component is not mentioned in available reporting. The result, if it holds, would be the first progress on a Millennium Prize Problem since Perelman. The phrase “existence of pure Yang–Mills theory” refers specifically to the construction of a quantum field theory satisfying the Wightman axioms or their constructive equivalents — a problem that has resisted decades of effort by some of the most capable mathematical physicists working, including Fields Medalists. [Assessed with high confidence — standard characterisation of the Yang–Mills problem’s mathematical difficulty.]

The Clay Institute’s prize requires that a solution be “published in a refereed mathematics journal of worldwide repute” and that “a general consensus in the mathematical community” emerge within two years. [Established — Clay Mathematics Institute prize rules. Tier 1 primary source.] Lean verification is not publication. Expert community consensus has not formed. The prize has not been awarded. [Established — Clay Mathematics Institute website, as of 28 October 2026.]

6. The Pattern: AI as Theorem-Prover, Not Just Search Engine

The conventional criticism of AI in mathematics is that AI systems pattern-match: they identify structures that resemble known proofs and recombine them in ways that look plausible but fail under scrutiny. The reason this criticism has been persistent is that it has been repeatedly correct. AI-generated mathematical “proofs” have failed expert review at high rates, and the failures have often been subtle — correct in structure, wrong in detail, in ways that take specialists considerable effort to identify.

The Lean verification changes the baseline. If a proof has been formalised and Lean-verified, the failures that plagued earlier AI proofs — hidden circularity, informal steps that don’t bear weight, quantifier errors — should have been caught. The question moves from “is the proof internally consistent?” (Lean answers yes) to “does the formalisation accurately represent the intended mathematical statement?” (which requires human expert review). [Assessed with high confidence — standard analysis of Lean’s capabilities and limitations in mathematical verification.]

This represents a genuine shift in what AI can offer mathematics. Prior AI systems could suggest proof strategies; this system produced arguments complete enough to be formalised. The analogy in physics would be the difference between a simulation that suggests a result and an experiment that measures one. Neither is the final word, but they are not equivalent kinds of evidence.

7. The Institutional Stakes: Prize Committees, Peer Review, and the Expert Response

The mathematical physics community will respond to the Yang–Mills claim with careful scepticism — as it should. The Poincaré conjecture took three years of intensive expert scrutiny before consensus formed, despite Perelman being a known mathematician with an established research programme. OpenAI’s claims originate from an AI system in a commercial organisation, with no named mathematician as primary author. The peer review question is therefore both mathematical and sociological: how does the community evaluate a result with no human correspondent to interrogate?

The Clay Institute will almost certainly convene an informal expert panel before any prize consideration. The Institute is not obligated to respond on any particular timeline, and the historical record — the Poincaré case, the Kepler conjecture, Andrew Wiles’s Fermat’s Last Theorem — suggests that resolving major mathematical claims takes years, not weeks. [Assessed with high confidence — standard institutional practice of prize committees and mathematical community verification; historical precedents Established.]

One early signal: Terence Tao, the Fields Medalist who has worked extensively on harmonic analysis and contributed to progress on the Kakeya conjecture (an adjacent problem OpenAI also claimed progress on), has not publicly commented as of publication. His response, when it comes, will be among the most authoritative available. [Speculative — no statement confirmed at time of publication; Tao’s relevance to these problems is Established.]

8. What Changes — and What Remains the Same

If any of the three claims survive full expert review — and the Yang–Mills claim in particular — the structural implication for science is significant but not unlimited. Mathematics is a domain uniquely suited to AI-assisted proof search, because mathematical proof is a formal activity with clear criteria of correctness. Physics, biology, and economics are not. The applicability of this mode of AI-driven discovery to empirical science is a separate question that these results do not answer.

What the results do suggest, if they hold, is that AI has crossed a threshold in mathematics: from tool that assists human proof-finding to agent that initiates proof attempts capable of surviving formal verification. The four-colour theorem in 1976 and the Kepler conjecture in 1998 established that computer-assisted proof is legitimate mathematics. If AI-generated and Lean-verified proofs produce results of the quality OpenAI is claiming, the community will need to decide whether AI-originated proofs occupy the same legitimacy as computer-assisted proofs, or a distinct category.

The scientific method does not change. What may change is its pace — and the character of the partnership between human mathematical intuition and AI capacity for exhaustive formal search.

Bottom line: OpenAI’s 27 October announcement is the most consequential scientific claim made by an AI system, measured by the formal weight of the problems addressed. Three results, each significant in its own domain; all formally verified in Lean, which is a genuine advance over prior AI proof attempts. The Yang–Mills claim, if it holds, would be the first resolution of a Millennium Prize Problem in 23 years. But the Clay Institute has not awarded the prize, peer review has not begun, and the history of major mathematical claims — AI-generated or human-generated — counsels patience before celebration. What is certain: an AI system has produced a formally consistent argument for the resolution of a problem that has defeated human mathematicians for decades. That is not the end of the story. It is the beginning of a very careful reading.