Researchers balk as OpenAI gates access to math, CS and physics breakthroughs

Three results, three release strategies

On Monday, OpenAI announced advances in mathematics, computer science and physics, along with what it describes as an evolving approach to communicating its findings. According to The Verge, the company consulted expert "shepherds" on how each result should be shared, and each received different treatment.

The mathematics output was published in full, leaving researchers to organize their own efforts to explain it. For the computer science finding, OpenAI plans to assemble an external group of researchers to interpret the result. The physics output has not been released at all; OpenAI said it will defer to outside experts on how to handle the material.

The announcement follows a turbulent stretch between OpenAI and mathematicians, who are still working through the hundreds of preprints the company released at the start of October, The Verge reports.

Substantial claims, uneven visibility

The mathematics result purports to settle the Kakeya conjecture in every dimension — a problem about how little space is needed to rotate a needle through all possible directions. A proof of the three-dimensional case was part of the work that earned NYU mathematician Hong Wang a Fields Medal earlier this year. Brown University professor Brendan Hassett told The Verge the claim "looks well supported," though he had not yet examined its Lean formalization, the machine-checkable version of the proof.

The physics finding concerns the Standard Model, specifically the relationship between quantum chromodynamics — the theory of the strong nuclear force — and Yang-Mills theory, a cornerstone of particle physics tied to one of the seven Millennium Prize problems. The Verge notes the work does not address the unresolved "mass gap" at the center of that problem.

The computer science result claims to disprove the exponential time hypothesis, the widely held assumption that certain problems cannot be solved much faster than brute force, and is closely related to P versus NP, another Millennium Prize problem.

Experts who spoke to The Verge described the results as "huge" and significant, and said there are good reasons to take them seriously — but the uneven disclosure makes independent judgment difficult.

Scientists welcome the work, not the process

Much of the frustration centers on OpenAI deciding who gets first access. Constantin Kogler of the Institute for Advanced Study told The Verge he welcomed the manageable batch of three results but objected to limiting early scrutiny to chosen figures: "It doesn't make any sense that first only a few distinguished scientists can look at them."

University of Texas at Austin physicist Andreas Karch was blunter, calling the handpicking "a disaster" and arguing that large labs already hold enough power and "should not be picking winners and losers." He suggested OpenAI could simply release the proof alongside its expert team, which would widen access while still showing it cares about making the results understandable.

Hassett added that open availability would likely produce better science and speed up adoption of AI tools by groups around the world who want to use them.

Advisory group had warned the day before

The release strategy is exactly what the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) — a body OpenAI had previously pledged to consult — warned against one day before the announcement. AGMAI said selectively granting early access is "unnecessary," borrows the reputation of the chosen researchers, and risks unequal access, biased selection and delays. It recommended publishing outputs on scholarly repositories with formalized proofs and details on how problems were selected, along with funding and support for researchers doing the work of interpretation.

After Monday's announcement, AGMAI pressed OpenAI to make the physics outputs public "without delay," noting that OpenAI's stated reason for waiting — an incomplete Lean formalization — was, by the company's own account, finished.

Others see no clean solution. Stanford's Jared Duker Lichtman told The Verge that any approach will upset someone, making it "a question of finding the least offensive option." Georgia Tech's Xiaoyu He said the results are "very impressive" but argued the problem runs deeper than announcement mechanics, suggesting the company is performing thoughtfulness about release practices without changing the underlying situation. A recurring theme among researchers was that AI-generated proofs arriving rapidly and largely without human contribution are difficult to read and verify, straining the community's normal review processes.

Why it matters

If AI labs are going to produce candidate solutions to some of the deepest open problems in mathematics, computer science and physics, the process for checking those claims becomes critical scientific infrastructure. The dispute is less about whether the results are real and more about who controls verification: when a single company decides which scientists see work first, it shapes credit, careers and the direction of fields it is only beginning to enter. The standoff also sets a precedent for open science norms — open repositories and community-wide scrutiny versus curated access — at precisely the moment AI output threatens to outpace the human capacity to review it.