OpenAI’s latest scientific claims have opened a dispute over who gets to examine the evidence. On October 26, the company announced AI-generated results in mathematics, computer science and physics. An independent mathematics advisory group called for immediate release of the withheld physics output.
Three claims, different release paths
According to OpenAI’s announcement, the results concern wave behaviour and the Kakeya conjecture, a claimed refutation of the Exponential Time Hypothesis, and existence proofs involving quantum field theories.
The company released mathematical and computer-science materials, including Lean formalizations. External groups will examine the computer-science and physics work; publication of the physics artifacts is deferred to that collaboration. OpenAI describes the manuscripts as raw research data.
The advisory group wants equal access
The Advisory Group on Mathematics and Artificial Intelligence (AGMAI) challenged that approach in statements dated October 25 and 26. It asked OpenAI to publish the natural-language and Lean outputs behind the announced physics results without delay.
The group respects the selected scientists but argues that understanding should develop openly, with the same materials available to everyone. Its wider objections to company-selected reviewers include unequal access, influence over research priorities and the risk that respected names are mistaken for an endorsement.
That distinction matters when reading this announcement: being invited to examine a result is different from having established its correctness. AGMAI also states that its advice does not represent the entire mathematical community.
What a machine-checkable proof does—and does not—tell readers
In its October 6 research release, OpenAI explained that Lean allows mathematical proofs to be checked by a computer and published formalizations alongside research materials. That earlier release also included information about attempts and computing effort.
For a non-specialist reader, the useful questions remain specific: which statement was formalized, which assumptions were used, and how does it connect to the headline claim? This article does not independently verify the proofs. It reports the release and the disagreement about how researchers should assess it.
Funding the work of understanding
There is a separate effort to support that assessment. On October 15, the Simons Foundation announced a combined $15 million in funding and resources with Anthropic and OpenAI to help mathematics researchers navigate AI. The foundation will manage awards, including support for workshops, evaluation of AI-assisted results and community standards.
The initiative addresses a practical challenge: generating a long technical argument and helping a research community understand it are separate pieces of work. Both require resources.
What to watch next
For readers following AI research, this is also a useful test of transparency. A strong release should make it possible to trace an ambitious claim to its evidence and follow subsequent corrections or clarification. Readers should distinguish those developments from the company’s initial claims.
For related context, read our coverage of AI language models and research integrity.
