OpenAI picked the people who get to check its physics
29 Oct 2026 · News
Three weeks ago OpenAI dumped its AI math on GitHub and let anyone check it. This week it chose who looks first.
A 1147-page physics proof written by an OpenAI model exists, and almost nobody is allowed to read it.
On Monday OpenAI announced three results from its internal models, in math, computer science and physics. Each one took a different route to the public, and each time OpenAI chose the people who decided it, Science reported on Wednesday.
The math result went out as is, because the expert OpenAI consulted said to release it. The computer science result refutes the exponential time hypothesis, a 27-year-old assumption that some puzzles always take exponentially long to solve. It went to three researchers OpenAI asked to "shepherd" it. They spent 10 days rewriting a first version one of them called "completely unreadable," unpaid and with no NDA, then posted everything. The physics result, an existence proof for quantum Yang-Mills theory, went to a five-person committee chaired by Nobel laureate David Gross. That committee is holding the proof until it is convinced it is correct.
To be fair to the committee, OpenAI picked only Gross. He chose the other four, and their statement says "We have no responsibility to OpenAI." These are serious people, and 1147 pages is a lot to check.
The process is still the problem. OpenAI's own unpaid advisory group of mathematicians said the day before the release that companies should not handpick shepherds. The day after, it asked for the Yang-Mills proof to be released, noting that OpenAI had said it was waiting on a Lean check, and that check is done. Lean is a programming language that verifies every step of a proof by machine. OpenAI's answer, per Science, was to take it up with Gross.
Ryan Williams, who leads the computer science shepherds, put it plainly. "The main difficulty is that a company was selecting or attempting to select who is taking charge of it." He would rather academics pick the committee.
Then there is Anthropic. Its model found a near-identical computer science proof on the same day, 16 October, and Anthropic just posted the paper. Nobody outside got it first. A spokesperson told Science the field "should lead the discussion" on what it means. Both labs say Lean checked their proofs. No human has fully confirmed either yet.
On 8 October we called OpenAI's math dump good news because three papers were retracted within two days, which meant someone outside was reading. Three weeks later the open dump has given way to a curated reading list. We have watched this pattern on safety audits, where the auditor moved into the lab. Now it has reached the science the labs produce.
Our call is that the open release wins. By 30 June 2027 a frontier lab will let an outside body it did not choose, such as a scientific society, journal or academic committee, pick the reviewers for an AI-produced result. We are wrong if every major AI proof by then still reaches the public through people the lab selected.
A result counts as checked when people the lab didn't pick have read it. Lean already qualifies, since nobody chose it. The humans should qualify too.