OpenAI Announces Three AI Scientific Breakthroughs and Introduces 'Shepherd' Verification Mechanism
OpenAI reveals three major AI-generated scientific results in math, CS, and physics, debuting a new 'shepherd' release process where external experts verify and interpret raw research data.
On October 26, 2026, OpenAI announced that its internal models achieved three significant discoveries in mathematics, computer science, and physics: a proof related to the local smoothing conjecture in harmonic analysis (implicating the Kakeya conjecture), a refutation of the Exponential Time Hypothesis (ETH) in computational complexity, and an existence proof for four-dimensional Quantum Chromodynamics (QCD). OpenAI characterized these outputs as "raw research data" and introduced a novel release strategy inviting external experts to act as "shepherds," organizing groups to understand and explain the results.
For the computer science result concerning the ETH refutation, OpenAI formed a collaboration group led by Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams. Impagliazzo and Paturi are the original proposers of the ETH, while Williams is a leading expert in the field. OpenAI committed to making all raw data public to facilitate independent verification and pledged to fund a conference to explore the broader implications for algorithm design and complexity theory. For the mathematical result on local smoothing, OpenAI chose to directly release the AI output and Lean formalization proofs, leaving researchers free to organize their own efforts to explain them. In the physics domain, OpenAI deferred the release timing to the external collaboration group.
This initiative aims to address the trust deficit surrounding AI-generated scientific claims. By adhering to principles of transparency, no academic credit for OpenAI, recognition for contributors, and independence, OpenAI seeks to establish a new paradigm where human experts translate AI's raw outputs into accepted scientific knowledge. Although these results have not yet undergone full peer review and their correctness remains subject to community validation, this "human-AI collaborative verification" model may serve as a crucial reference for future AI involvement in frontier scientific research.