OpenAI puts three AI-generated science claims on separate review tracks

OpenAI disclosed three claimed AI-generated scientific results on October 26, putting each on a different release and external-examination track. Artifacts for the mathematics and computer-science claims are public, while the physics work remains withheld as an outside group prepares its own account.

The company called the manuscripts “raw research data” and said it is testing three stewardship models. It released the mathematics work immediately, paired public computer-science artifacts with a verification group and deferred the physics release to an external group. None of the opened sources included a completed independent assessment confirming any of the three scientific conclusions.

In mathematics, OpenAI claims its systems proved the local smoothing conjecture in every dimension, which it says would also resolve the Kakeya and Stein restriction conjectures in all dimensions. The natural-language manuscript and a description of the Lean formalization’s scope were publicly accessible at the research cutoff. OpenAI said it consulted an expert before following that person’s recommendation to release the output, but its announcement did not identify the expert or publish an assessment.

For computer science, OpenAI claims it disproved the Exponential Time Hypothesis. Its public paper is accompanied by a Lean scope document. The scope document says the formalized result proves, for each fixed integer k of at least 2, a deterministic exact counting algorithm for canonically encoded k-CNF formulas with subexponential dependence on the number of variables. The opened evidence did not contain an independent conclusion that this formal theorem and the natural-language argument settle the broader hypothesis as intended.

OpenAI said Russell Impagliazzo, Ramamohan Paturi and Ryan Williams would lead a group to verify and shepherd the computer-science work through an open community examination. The opened sources did not independently confirm each member’s participation or give a date for the group’s assessment.

Physics is following the most restrictive process. OpenAI claims a rigorous existence proof for certain four-dimensional interacting quantum field theories, including quantum chromodynamics and a unified completion of the Standard Model. The company said a group led by David Gross and including Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov chose not to release the AI-produced artifacts immediately and would prepare its own write-up. Because the proof and formalization were unavailable, their contents and scope could not be inspected.

OpenAI characterized its Yang–Mills result as an existence subproblem, not a solution to the full Millennium Prize problem. The Clay Mathematics Institute’s problem statement requires both construction of a non-trivial four-dimensional quantum Yang–Mills theory and proof of a positive mass gap. OpenAI’s announcement did not claim to have completed both parts.

The company said the external collaborations would operate transparently and rely on voluntary, unpaid participation. OpenAI said it would not seek academic credit, that inviting experts to shepherd the work would not itself grant them discovery credit, and that collaborators should be recognized for their explanatory contributions. The company did not disclose the exact models, prompts, compute budget, generation procedure or evaluation record behind the results. OpenAI dated the AI outputs from September 9 through October 18 and said Lean work followed on later dates.