OpenAI claims its AI has disproved the Exponential Time Hypothesis (ETH), a major assumption computer scientists use to establish limits on how efficiently certain problems can be solved. If independent researchers confirm the result, some existing conclusions about computational complexity will need to be reconsidered.
The company’s October 26 announcement also reports discoveries in wave mathematics and quantum field theory. OpenAI is using different approaches to release the findings, including a physics collaboration whose decision to withhold the underlying research has drawn criticism from independent mathematicians.
The disagreement raises a challenge for scientific institutions: AI could produce complex discoveries faster than researchers can independently verify and explain them. How OpenAI makes its findings available may therefore be as important to scientific progress as the results themselves.
What OpenAI says its models discovered
OpenAI’s computer science claim challenges ETH, which states that a standard form of Boolean satisfiability cannot be solved in subexponential time. Boolean satisfiability asks whether a collection of true-or-false conditions can all be satisfied simultaneously.
Researchers use ETH to establish conditional limits on the performance of other algorithms. If the reported disproof holds, those ETH-based arguments would need to be reassessed, although the result would not resolve the separate P versus NP problem or automatically make everyday software faster. Some computational limits could also remain valid for reasons independent of ETH.
In mathematics, OpenAI reports a proof involving local smoothing, which examines how the concentration of waves described by mathematical equations changes when their behavior is measured over time. The company says the result covers all relevant dimensions and has implications for the Kakeya and Stein restriction conjectures, longstanding problems in geometry and mathematical analysis.
The physics claim concerns the mathematical foundations of quantum field theory, which describes how particles and fields interact. OpenAI says its AI established existence results in four-dimensional spacetime, including one involving quantum chromodynamics, the theory governing interactions between quarks and gluons. The announcement also describes a unified completion of the Standard Model.
Establishing mathematical existence would not demonstrate that a proposed unified theory accurately describes the physical universe. OpenAI also reports a pure Yang–Mills existence result, but the Clay Mathematics Institute’s Millennium Prize problem also requires proof of a positive mass gap, a minimum amount of energy needed to excite the system above its vacuum state. The reported existence result alone would not resolve the full prize problem.
How OpenAI is releasing the research
OpenAI has made the mathematical proof and its Lean formalization publicly available, so researchers outside the collaboration can examine the argument. The computer science proof materials are also public. Russell Impagliazzo and Ramamohan Paturi lead an external examination alongside Ryan Williams.
The physics collaboration, led by David Gross, has taken a different approach. Its members plan to prepare their own explanation before the underlying AI-generated materials are released, temporarily restricting examination to researchers with access to the work.
According to OpenAI, the collaborators participate voluntarily and independently. They receive no payment for their involvement and can publish their own conclusions. Their participation should not be interpreted as confirmation that the reported discoveries are correct.
OpenAI previously used a public repository and Lean formalizations in its October 6 mathematics release. The latest announcement introduces a more prominent role for external researchers in examining and explaining the findings, although the three collaborations differ in how quickly the wider scientific community receives access.
Why machine-checked proofs still need expert review
Lean is software that checks whether mathematical arguments follow formal logical rules. An accepted proof establishes a specific statement under its defined assumptions, but the software cannot determine whether that statement accurately represents the research problem scientists intended to solve.
Researchers must examine the definitions and confirm that the published conclusions reflect what the proof actually establishes. A mathematically valid argument could still leave an important part of the original problem unresolved if its formal statement is incomplete or incorrectly formulated.
A separate study of Lean benchmarks identified cases where formal statements did not accurately represent their intended problems. Although the study does not assess OpenAI’s new discoveries, it illustrates why successful machine checking cannot replace independent examination of the underlying mathematical claims.
Researchers also need explanations they can teach and techniques they can apply to other problems. A machine-checked proof may establish formal validity, but its wider scientific contribution depends partly on whether other scientists can understand the reasoning and build on the result.
Who should receive early access to AI discoveries?
Giving selected experts early access may help organize the review of AI-generated research. However, those researchers can also examine the results and pursue follow-up work before others can inspect the same materials.
On October 26, the Advisory Group on Mathematics and Artificial Intelligence called for the immediate release of OpenAI’s physics outputs. It argued that completed formalization removed the stated justification for withholding them.
The group’s October 25 guidance also warned that letting companies select initial reviewers could discourage researchers outside those collaborations. OpenAI acknowledges that providing early access to selected experts creates differences in who can examine the evidence.
The disagreement highlights the difference between independent collaboration and open scientific scrutiny. Experts can assess OpenAI’s claims independently, but withholding the underlying materials prevents other researchers from checking the same evidence or developing alternative interpretations.
Public release also gives scientists the opportunity to challenge assumptions that a smaller review group might overlook. For discoveries with potentially major implications, examining competing explanations is part of establishing confidence in the results.
Scientific verification will need funding, too
The Simons Foundation’s $15 million initiative combines funding and resources to support mathematicians as they examine AI-generated research and adopt new tools. The program includes contributions from Anthropic and OpenAI, with OpenAI having announced a $5 million commitmen in early October.
Funding could help researchers examine proofs and independently reproduce findings, rather than leaving those tasks dependent on unpaid academic work. The need becomes more significant when a result requires specialized mathematical knowledge or substantial computing resources to assess.
Universities may need dedicated support for researchers who verify AI-generated results, particularly when the examination requires expertise across multiple fields. Journals could also distinguish checking a proof’s formal validity from evaluating its scientific significance, while recognizing researchers who produce independent explanations.
As The Neuron previously reported, outside mathematicians examined an earlier AI-generated geometry result and developed improvements. Their work illustrates how independent review can extend a discovery rather than merely confirm its correctness.
OpenAI’s three collaborations show that evaluating AI-generated scientific discoveries involves more than checking whether a proof is logically valid. Researchers also need access to the original work and time to establish what the findings mean for existing research.
If AI systems continue producing complex results at this pace, universities and scientific organizations may need to devote more resources to independent verification. The credibility of future discoveries will depend partly on whether researchers beyond the companies developing these models can examine the evidence and build on it.
