Experts vet OpenAI's AI-made Yang-Mills proof, with limits
A review group says OpenAI's AI-generated proof builds a quantum Yang-Mills theory, checked in Lean. It does not prove the mass gap, and reviewers say they do not yet fully understand it.
An AI-generated mathematical proof from OpenAI on a core problem in quantum physics has passed an independent review, according to a report in Scientific American and the review group's own public repository. The reviewers are careful about what has, and has not, been shown.
What was reviewed
OpenAI submitted its proof on 28 October 2026, according to the review group's repository on GitHub. The work concerns quantum Yang-Mills theory, a foundation of the Standard Model of particle physics and the subject of one of the Millennium Prize Problems. Scientific American reports that the proof addresses the "existence" part of that problem.
OpenAI appointed a committee led by theoretical physicist David Gross to vet and release the result, Scientific American said. Harvard's Michael R. Douglas is among its members, and the committee has four other members besides Gross. The group's assessment is dated 10 November 2026.
Verified points
- Both sources say the committee issued a statement on 9 November saying the result proves the existence of pure Yang-Mills theory, and that an assessment followed on 10 November.
- Both sources say the proof package includes Lean code, and Scientific American reports the committee affirmed that the proof was formally verified in Lean.
- Both sources say the mass gap part of the problem is not resolved.
What the reviewers say it shows
In the group's opinion, as stated in its repository, the submission shows that a relativistic quantum field theory satisfying the Wightman axioms can be built as a scaling limit of four-dimensional Wilson lattice gauge theory, for any simple compact gauge group. The group says the submission does not prove that the infinite-volume limit is unique and does not prove a mass gap. It describes the assessment as its opinion and limited in scope.
Reported claims and caveats
According to Scientific American, the committee used OpenAI's "Astra" model to help digest the Lean code and wrote: "We do not claim to fully understand the proof yet." The same report quotes Scott Armstrong of New York University welcoming the Lean release while arguing that AI labs should share results more openly.
The repository adds practical caveats: posting a submission does not endorse or certify its claims, the group does not promise to review every submission, and a second OpenAI submission dated 4 November is listed as work in progress with no results yet. It also notes that its recorded macOS verification mode lacks Linux isolation. Scientific American further reports that OpenAI announced its claim in late October, ahead of a collaboration between another mathematician and an Anthropic employee on the same result.
Analysis
This is analysis, not reporting: the episode shows how formal proof tools such as Lean may change how AI-generated mathematics is checked, since reviewers can verify code without fully following every step by hand. Still, the reviewers' own admission that they do not yet fully understand the proof suggests the wider physics community will need time before treating the result as settled.
Sources
Related stories
SpaceX catches Starship from orbit for the first time
SpaceX's launch tower arms caught a Starship upper stage returning from orbit on November 14, a first for the programme. The booster was lost after engine trouble, so the planned double catch did not happen.
Geologists link Earth's missing rock layer to deep mantle
A Nature study modelled mantle motion to explain the Great Unconformity, suggesting uplifted continents were eroded, possibly by Snowball Earth glaciers.
Astronomers weigh a 'Planet Nine' candidate in old data
A team combing 1980s and 2000s infrared surveys flagged one possible distant world, but its steep orbit does not match the long-proposed Planet Nine, writes astronomer Phil Plait.
How HyperLogLog counts the internet with coin flips
A Scientific American explainer describes how the HyperLogLog algorithm estimates unique visitors using a tiny amount of memory, based on the logic of coin-flip streaks.
