Capabilities

OpenAI's AI-Generated Math Proofs: What a Lean Check Proves and What Still Needs Human Review

OpenAI says its models disproved the Exponential Time Hypothesis and settled other famous problems. The Lean files are public, but the independent verdict is still pending.

Claire Nakamura

Technical Writer, Frontier AI Coverage

Published 7 min read
mathematics, math, pythagoras, theorem
In this story 7 sections

OpenAI announced on Monday, October 26, that its models produced a proof that the Exponential Time Hypothesis is false, plus results on two other famous problems. A Lean check, a machine verification of the proof steps, backs the first claim. Independent experts have not yet certified it, so treat the result as promising but unconfirmed.

The announcement came with files, a formal proof and a promise of review, which is better than most AI claims. It still leaves a gap between "the software accepted it" and "mathematicians agree it is true." Emergent Wire covers what AI systems can actually do, tested against the claims, and this piece is for developers and technical readers who want to know how to weigh a headline like this.

What Did OpenAI Announce?

Vivid, blurred close-up of colorful code on a screen, representing web development and programming.

OpenAI announced three results across mathematics, computer science and physics. According to Fox News's live coverage, the headline claim is a proof that the Exponential Time Hypothesis is false, which OpenAI says Lean verified. It also reported proofs addressing the Kakeya conjecture and results on the foundations of quantum chromodynamics.

Reporting from Cryptopolitan adds the timeline. It says the disproof was found on October 16 and verified in Lean the next day, and that the write-up and the Lean proof are on GitHub. The same report says OpenAI claims a proof of Christopher Sogge's 1991 local smoothing conjecture in every dimension, which it says would also settle the Kakeya and Stein restriction conjectures.

The physics claim is the most cautious of the three. A group led by Nobel laureate David Gross has chosen to hold back OpenAI's materials while it prepares its own analysis. That choice alone signals that the specialists want to do their own work before putting their names near the claim.

Why Does the Exponential Time Hypothesis Matter?

teacher, geometry, mathematics, dice

The Exponential Time Hypothesis, or ETH, matters because hundreds of results in computer science build on it. It concerns how hard it is to solve Boolean satisfiability problems, which ask whether a logical formula can be made true. ETH says that for a standard version of the problem, no algorithm can do much better than trying exponentially many possibilities.

Researchers use ETH as a working assumption. If it holds, many other problems must be hard too, so papers prove "this problem is at least as hard as satisfiability" and move on. If it fails, a long list of those hardness results could lose their foundation, or at least need a fresh look.

That is why a disproof would be a big deal. It would not break everyday software on its own, but it would change what theorists believe about the limits of fast algorithms. It would also be one of the most striking cases of a machine producing a result that experts had failed to find for decades.

What Does a Lean Check Actually Prove?

Three men collaborating and learning technology together on laptops indoors.

A Lean check proves that a formal statement follows from its assumptions by valid steps. Lean is a proof assistant, a program that checks each step of a proof against strict rules. If the file compiles with no gaps, the logic inside it holds.

The catch is the word "inside." The proof is only as meaningful as the statement written into the file. If the formal statement quietly differs from the real conjecture, the check can pass and still miss the point. A researcher who reads the Lean statement has to confirm that it says what everyone thinks it says.

Cryptopolitan reports that critics have made exactly this argument. It cites a paper by Anders Hansen and co-authors arguing that a Lean-verified formalization can compile while failing to match the intended natural-language proof. Separate research points the same way. The LeanMarathon paper on arXiv describes how, in long formalization projects, statements drift and dependencies tangle, which is why its authors add adversarial review to check target fidelity.

For developers, the lesson is familiar. A passing test suite shows the code meets the tests, not that the tests describe the right behavior. Lean is a very strong test suite. It is still a test suite.

There is also a human cost to checking. Reading a long formal proof and the paper behind it can take expert reviewers weeks. If a lab produces results faster than the community can verify them, the bottleneck moves from discovery to review, and the people doing the reviewing are often unpaid volunteers.

Who Is Checking OpenAI's Claims?

Three outside computer scientists are leading the review of the ETH result: Ryan Williams of MIT, and Russell Impagliazzo and Ramamohan Paturi of UC San Diego. Impagliazzo and Paturi are among the originators of ETH, which makes them natural judges. They published an open letter saying their involvement should not be read as certifying the claim.

Williams said on X that the group signed no nondisclosure agreement and will not be paid. According to Cryptopolitan, OpenAI agreed to publish all raw data behind the result and to fund a conference on its implications. OpenAI also gave $5 million to the Simons Foundation for workshops on AI-produced results.

The table below summarizes where each claim stands based on the reporting so far.

Who Is Checking OpenAI's Claims?
ClaimEvidence releasedIndependent status
Exponential Time Hypothesis is falseWrite-up and Lean proof on GitHubUnder review by Williams, Impagliazzo and Paturi
Local smoothing, Kakeya and Stein restrictionWrite-up and Lean proofNo independent verification reported
Existence of certain quantum field theoriesMaterials withheld by Gross groupGross group preparing its own analysis

Why Is Caution Warranted?

Detailed view of a microchip on a printed circuit board, showcasing electronic components.

Caution is warranted because earlier AI math announcements needed corrections. Cryptopolitan recalls that in October 2025 an OpenAI vice president claimed a model had solved 10 open Erdős problems, and that Thomas Bloom, who maintains the Erdős problems site, called this a dramatic misrepresentation because the model had mainly surfaced existing papers.

The same report says OpenAI released 722 preprints on 372 math problems on October 6 and withdrew three within a day because of a sign error that invalidated one argument and two papers that depended on it. It revised 14 others. These are the report's figures, and OpenAI has not necessarily described them the same way.

None of this means the new claim is wrong. It means the cost of checking is real, and the community has been asked to pay it before. Our earlier piece on Meta Muse Wonder benchmark claims showed the same pattern: results arrive with a press release first and scrutiny later.

What Should Developers and Readers Take From This?

Read AI discovery claims the way you would read a vendor benchmark. Ask who ran it, who can reproduce it and what exactly was tested. Emergent Wire's own approach is to separate what a lab says from what outsiders have confirmed, and to say plainly when the second column is empty.

Here is a short checklist you can apply to any claim like this one.

  1. Is the artifact public, such as code, data or a formal proof?
  2. Does an independent party with real expertise have access?
  3. Did they agree to be named, and were they paid or bound by a nondisclosure agreement?
  4. Does the formal statement match the plain-language claim?
  5. Has the lab walked back earlier results, and how did it handle that?

For builders, the practical effect is limited today. Nothing in the announcement changes how you ship software this week. The longer-term signal is about tooling, since a model that can write checkable proofs is also a model that can write checkable code. Our look at Coding Agent Index v2 scores shows how hard it is to say what a single score tells you about real work.

It also connects to the question of machine-found flaws. Our report on AI-found PHP vulnerabilities described the same dynamic in security: a machine finds something real, and humans still have to confirm it.

One more habit is worth keeping. Separate the tool from the claim. Models that help mathematicians find and check arguments can be useful even if a particular headline result fails. A single overstated announcement does not cancel the real progress in proof assistants, and a single success does not guarantee the next one will hold.

The Takeaway on OpenAI's Math Claims

OpenAI has put real artifacts behind a bold claim, and respected outsiders have agreed to examine it. That is a better process than most announcements get. It is not yet a result, because the Lean check proves the proof is consistent with its own statement, not that the statement is the right one.

Your next step is simple. Watch for the reviewers' verdict from Williams, Impagliazzo and Paturi, and until then, do not cite the disproof as settled. If you work with formal verification yourself, read the Lean statement for any claim you depend on and confirm it says what the headline says.

What is the Exponential Time Hypothesis?
The Exponential Time Hypothesis, or ETH, is a conjecture that solving a standard version of the Boolean satisfiability problem requires time that grows exponentially with the problem size. Hundreds of computer science results assume it, so a confirmed disproof would force researchers to revisit many hardness arguments.
Did OpenAI really disprove the Exponential Time Hypothesis?
OpenAI says so and has released a write-up and a Lean proof. However, three outside experts, Ryan Williams, Russell Impagliazzo and Ramamohan Paturi, are still reviewing it and say their involvement is not a certification. Until their verdict arrives, the claim should be treated as unconfirmed.
What does a Lean proof check actually verify?
Lean verifies that each step of a formal proof follows validly from its assumptions. It does not check whether the formal statement matches the plain-language claim. A file can compile and still prove something slightly different, so experts must read the formal statement itself before trusting the result.
Should developers change anything because of this announcement?
Not yet. The announcement does not change how software is built or shipped this week. The broader signal is that models are getting better at producing machine-checkable work, which could matter for formal verification tools later. For now, wait for independent review before relying on the claim.

Written by

Claire Nakamura

Technical Writer, Frontier AI Coverage

8+ years experience

Claire Nakamura started as a developer relations engineer at two developer-tooling companies before moving into full-time technical journalism. She spends most of her working week with API documentation, model cards, and changelogs that most outlets skip past, and she is the person on staff most likely to have actually run the eval a press release is citing. At Emergent Wire she covers developer-facing AI tooling, agent frameworks, and the practical gap between a model's stated capabilities and what shows up in a production integration.

Credentials

  • Former developer relations engineer, developer-tooling sector (6 years)
  • Hands-on tester of API-based AI tools and agent frameworks
  • Regular contributor on model card and changelog analysis