When AI Produces Science Faster Than Humans Can Verify It

A deep dive into OpenAI’s three scientific claims of 26 October 2026

Status reviewed: 27 October 2026

Bottom line. OpenAI has announced three extraordinary AI-generated results in mathematics, theoretical computer science and mathematical physics. Two have public manuscripts and Lean formalizations; the third is being examined privately by a distinguished physics group. None has completed independent community validation. If the claims survive scrutiny, the results would be scientifically important. The larger story may be just as important: AI can now generate research claims quickly enough that verification, explanation, access and credit are becoming bottlenecks.

What OpenAI announced

The unusual part is not just the science

On 26 October 2026, OpenAI announced three results produced by its internal models. The company itself used unusually careful language about how they should be treated: it described the manuscripts as “raw research data” that still need to be converted into scientific understanding by human researchers. OpenAI is testing three different release models at once: immediate public release for the mathematics result, public release plus an organized external group for the computer-science result, and a temporarily withheld expert-led process for physics.

This distinction matters. A model can generate a long argument. A proof assistant can check a precisely formalized theorem. But science also asks whether the formal statement matches the intended claim, whether assumptions are appropriate, how the new ideas connect to existing literature, whether independent experts can reconstruct the reasoning, and what the result actually changes. Those are human and community-level tasks as well as formal ones.

1. Mathematics: waves, Kakeya and a very large claim

The first result concerns the local smoothing conjecture, proposed by Christopher Sogge in 1991. In simple terms, it describes how waves can focus. A wave may become extremely concentrated at a particular instant, but the conjecture says that such extreme focusing cannot persist in the same way when behavior is averaged over time.

OpenAI says its model found a proof in every spatial dimension at least two, based on a new auxiliary partial differential equation. That is much broader than the previous human state of the art. Guth, Hong Wang and Ruixiang Zhang proved the local smoothing conjecture in two spatial dimensions (the 2+1-dimensional wave equation). Separately, Hong Wang and Joshua Zahl resolved the three-dimensional Kakeya conjecture in 2025. Higher-dimensional Kakeya remained open.

Why Kakeya appears here. Imagine a very thin laser beam or pencil that must point in every possible direction while occupying as little space as possible. The Kakeya conjecture says that even though such a set can have surprisingly small volume, it must still have full mathematical dimension. OpenAI’s local-smoothing result claims consequences strong enough to settle the Kakeya problem in all dimensions, along with major restriction and harmonic-analysis estimates.

The formal evidence is substantial. OpenAI’s Lean scope document says the formal development covers the Euclidean half-wave local-smoothing estimate, Bochner-Riesz bounds, spherical restriction, a Kakeya maximal estimate, full Hausdorff dimension of Kakeya sets, and Stein square-function estimates. That makes this far more than a casual language-model answer. But the public mathematical community still has to understand, audit and contextualize the proof.

2. Computer science: the Exponential Time Hypothesis may be false

The second claim may be the easiest to explain and the hardest to absorb. The Exponential Time Hypothesis (ETH) is a long-standing assumption about the difficulty of Boolean satisfiability problems. Roughly, it says that 3-SAT cannot be solved in subexponential time in the worst case. For decades, computer scientists have used ETH as a foundation for conditional lower bounds: “if ETH is true, then this other problem also cannot be made dramatically faster.”

OpenAI’s paper claims a deterministic algorithm that does something even stronger than merely deciding whether a fixed-width Boolean formula is satisfiable: it exactly counts its satisfying assignments in subexponential time while using polynomial space. The 3-SAT case would directly contradict ETH.

This does not mean P = NP. A subexponential algorithm need not be polynomial-time. It also does not mean every NP-hard problem suddenly becomes easy, and it does not by itself produce practical speedups on ordinary computers. Asymptotic complexity describes how running time grows for very large inputs; constants and implementation details still matter.

If the result is correct, hundreds of ETH-based papers do not become “wrong.” Their conditional statements can remain logically correct, but ETH would no longer be a true premise from which to infer real lower bounds. The field would need to revisit which hardness conclusions survive under weaker or different assumptions.

The verification process is unusually well designed. Russell Impagliazzo and Ramamohan Paturi, who originated ETH with Francis Zane, and MIT’s Ryan Williams are organizing an open examination. Their message to the community explicitly says that their involvement must not be treated as certification. They asked for the raw materials to be public and for a conference to examine the claim; OpenAI agreed.

A small but important Lean nuance

OpenAI’s public Lean scope is enough to establish a subexponential algorithm, which is the key fact needed to refute ETH if the formal definitions correspond correctly to the intended problem. However, the scope note and the paper present slightly different detailed running-time bounds. The Lean note explicitly records a weaker general bound than the manuscript’s sharpest bound, while still remaining subexponential. This is a good example of why “Lean verified” should be read precisely: Lean verifies the formal theorem that was encoded, not every sentence or every stronger formulation in a human paper.

3. Physics: Yang-Mills, QCD and the result we cannot inspect yet

The physics claim may ultimately be the deepest of the three, but it is currently the least transparent. OpenAI says its models produced mathematically rigorous existence proofs for certain interacting quantum field theories in four dimensions, including quantum chromodynamics (QCD), the theory of the strong interaction between quarks and gluons, and a unified completion containing the gauge fields and fermions of the Standard Model.

The work also includes existence of pure Yang-Mills theory. That is part of the Clay Mathematics Institute’s Yang-Mills and Mass Gap Millennium Prize problem. The full Millennium problem asks for both the existence of the quantum Yang-Mills theory and a positive mass gap. OpenAI has announced the existence part; it has not announced a proof of the mass gap. So this should not be reported as “AI solved the Yang-Mills Millennium Prize problem.”

OpenAI says the model produced a natural-language Yang-Mills existence proof on 9 September, a QCD proof on 18 September, and completed the Yang-Mills Lean formalization on 16 October. A group led by Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, is shepherding the work. The group chose to hold back the AI-produced artifacts while preparing its own account.

That decision immediately became controversial. On 26 October, the independent Advisory Group on Mathematics and Artificial Intelligence (AGMAI) called on OpenAI to release the natural-language and Lean outputs without delay. AGMAI argued that the completed formalization removes the scientific reason for withholding the material and that understanding should develop openly with equal access. As of this review on 27 October, I found no public OpenAI reversal of the withholding decision.

What Lean verification really gives us

Lean is an interactive theorem prover. Its small trusted kernel checks whether a formal proof follows from the definitions, axioms and earlier theorems that have been encoded. For AI-generated mathematics this is extremely valuable: a model cannot simply wave past a logical step if the final formal object is genuinely accepted by the kernel.

But Lean does not automatically answer every scientific question. A formally valid theorem can still be the wrong theorem to formalize, use definitions that fail to match the intended real-world concept, omit a stronger claim made elsewhere in the manuscript, or be difficult for humans to understand. Lean’s own validation guidance explicitly distinguishes “does the theorem have a valid proof?” from “what does the theorem statement mean?”

Natural-language proof: helps humans understand the argument, but can contain gaps or mistakes.

Lean proof: gives very strong evidence that a precisely encoded formal statement follows from the encoded assumptions.

Independent scientific review: checks the match between the formal theorem, the intended claim, prior literature, assumptions, significance and broader consequences.

Why caution is justified without dismissing the work

OpenAI’s October 6 mathematics release provides a useful precedent. The repository initially contained hundreds of AI-generated manuscripts. On October 7, OpenAI withdrew three manuscripts after a sign error invalidated a key argument and two dependent papers, and revised fourteen other manuscripts with proof repairs, corrected statements or clearer hypotheses. The repository currently records 719 manuscripts and says about 42 percent of top-line results in that large release had formalizations at that point.

This should not be used as evidence that the new formalized results are false. It shows something more general: large-scale AI research production creates a new verification problem. Generating manuscripts can become cheap while expert attention remains scarce. The scientific bottleneck can move from “finding a possible proof” to “confirming exactly what was proved and integrating it into human knowledge.”

The new argument is about scientific governance too

One day before OpenAI’s announcement, AGMAI warned against AI companies designating particular mathematicians to receive or lead the interpretation of AI outputs. Its concerns included unequal access, loss of community agency, biased selection and what it called “borrowed reputation”: a famous researcher’s name can look like an endorsement before the work has actually been checked.

OpenAI is trying to address some of those concerns. It says the external collaborators are unpaid, sign no NDAs, remain independent, and are free to publish their own conclusions. OpenAI also says it will not seek academic credit for discoveries generated by its internal models and has committed funding to help researchers examine AI-produced results. The Simons Foundation separately announced a $15 million combined initiative with OpenAI and Anthropic, including an open call for workshops and projects that help mathematicians evaluate and adapt to AI-generated research.

Still, there is a genuine unresolved tension. Expert shepherding can make an opaque machine-generated result understandable faster. But choosing the first experts can also shape who gets early access, who becomes associated with the discovery, and whose interpretation becomes dominant. The physics result makes that tension concrete because the wider community currently cannot inspect the underlying artifacts at all.

Timeline

If the claims survive scrutiny, what changes?

Mathematics: a large network of harmonic-analysis conjectures would move from open to solved, including higher-dimensional Kakeya results that had resisted human proof.

Theoretical computer science: ETH would fall. Researchers would need to revisit many conditional lower bounds and understand the new algorithmic ideas behind the subexponential SAT/counting result.

Mathematical physics: a rigorous construction of four-dimensional interacting theories such as QCD would be a major foundational advance. The announced Yang-Mills result would address existence, but not by itself the mass-gap half of the Millennium problem.

Scientific practice: AI could become not merely a tool that assists researchers, but a generator of results whose output rate exceeds the human community’s verification and explanation capacity.

What we should not say yet

“OpenAI has definitively solved all three problems.” Independent scientific validation is not complete.

“ETH false means P = NP.” It does not. The claimed algorithm is subexponential, not polynomial.

“This makes NP-hard problems easy in practice.” The paper establishes asymptotic theory, not a practical SAT breakthrough.

“OpenAI solved the Yang-Mills Millennium Prize problem.” The announced result concerns existence; the Millennium problem also requires a mass gap.

“Lean verification makes human review unnecessary.” Lean checks a formal statement; humans still must check meaning, correspondence, assumptions, novelty and implications.

BunCosmos assessment

The correct position today is neither disbelief nor celebration as settled fact. These claims are serious enough to demand attention because they come with unusually concrete artifacts, named expert review processes and formal proof work. They are also new enough that the scientific community has not had time to independently digest them.

The most important development may be the workflow itself. The traditional scientific sequence is often imagined as a researcher developing an idea, writing a paper and then asking peers to review it. Here the sequence is changing: an AI system generates a result, formal tools check parts of it, and entire expert communities must then reconstruct meaning, context and consequences. If this becomes common, the scarce resource in science may increasingly be human understanding rather than raw proof generation.

For BunCosmos, the story should remain open. The ETH working group’s public findings, independent responses to the local-smoothing proof, and the eventual release or expert write-up of the Yang-Mills/QCD work are the three developments worth tracking next.

Sources and further reading

1. Working with academic communities to build scientific understanding. OpenAI, 26 Oct 2026. Source

2. openai/math repository history. OpenAI / GitHub, updated 26 Oct 2026. Source

3. Lean scope: Local smoothing for the Euclidean wave equation. OpenAI / GitHub, 26 Oct 2026. Source

4. Lean scope: The Exponential Time Hypothesis is False. OpenAI / GitHub, 26 Oct 2026. Source

5. Message / public statement from Ryan Williams on the ETH examination. Bluesky, 26 Oct 2026. Source

6. Yang-Mills & the Mass Gap. Clay Mathematics Institute, current reference. Source

7. Advisory Group on Mathematics and Artificial Intelligence statements. AGMAI / Institute for Advanced Study, 25-26 Oct 2026. Source

8. Simons Foundation Launches $15M Effort To Help Mathematicians Navigate AI Era. Simons Foundation, 15 Oct 2026. Source

9. A Three-Dimensional Breakthrough (Kakeya). Institute for Advanced Study, 2025. Source

10. A sharp square function estimate for the cone in R3. Annals of Mathematics, 2020. Source

11. Validating a Lean Proof. Lean Language Reference, current reference. Source

12. On the Complexity of k-SAT. Impagliazzo & Paturi, Journal of Computer and System Sciences, 2001. Source

Editorial note: This is a status report, not a declaration that the claims are established. Because the review process is active, conclusions may change as experts publish checks, corrections or explanatory papers.