OpenAI sets three review paths for AI-generated research claims
OpenAI has announced three AI-generated research claims in mathematics, computer science and physics, alongside different plans for expert scrutiny and release. The company says its models produced a proof of the local smoothing conjecture in every dimension, a disproof of the Exponential Time Hypothesis (ETH), and rigorous existence proofs involving quantum chromodynamics (QCD) and a unified completion of the Standard Model.
The results are being handled through three distinct strategies. For mathematics, OpenAI has released the model output and left researchers free to organise their own examination. For computer science, it has formed an external collaboration while publishing the underlying natural-language proof and Lean formalisation. For physics, an external group will lead the write-up and has chosen to defer release of the AI-produced artifacts.
Different pathways from model output to scientific understanding
OpenAI characterises the manuscripts as raw research data rather than finished scientific understanding. It says external experts were asked to act as shepherds for groups that can test, explain and contextualise the claims. The company acknowledged a trade-off: selecting initial shepherds can shape early participation and create information asymmetry.
The collaborations are intended to operate transparently, without nondisclosure agreements, and participants may discuss the work, share intermediate materials and publish conclusions. OpenAI says participation is voluntary, unpaid and independent; it does not seek academic credit for discoveries made by its internal models. It also says researchers who develop explanations and write-ups should receive recognition for that work.
Claims cover harmonic analysis, ETH and quantum field theory
In mathematics, the claimed result concerns Sogge’s 1991 local smoothing conjecture, which addresses the behaviour of waves as they focus. OpenAI says a proof in every dimension also resolves the Kakeya conjecture and Stein’s restriction conjecture in all dimensions. It attributes the approach to a novel auxiliary partial differential equation. The model output was obtained on October 18, 2026, formalised in Lean the following day, and both the natural-language write-up and Lean proof were released through the company’s public GitHub repository.
For computer science, OpenAI says its system found a natural-language disproof of ETH on October 16, 2026, followed by a Lean formalisation on October 17. ETH is a longstanding complexity-theory conjecture about Boolean formula satisfiability and has supported conditional hardness results across multiple fields. The review group is led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, with Ryan Williams of MIT. The group requested public release of raw materials to support independent verification and an open examination of the claim.
The physics collaboration will examine claims of mathematically rigorous existence proofs for interacting four-dimensional quantum field theories with gauge fields and fermions, including QCD and a unified completion of the Standard Model. The work also includes an existence proof for pure Yang–Mills theory, which is part of the Quantum Yang–Mills Theory Clay Mathematics Institute Millennium Prize problem. David Gross will lead a group that includes Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov.
Verification remains the operating requirement
OpenAI says it completed a Lean formalisation of Yang–Mills existence on October 16 after model outputs for Yang–Mills and QCD dated September 9 and September 18, respectively. It has also committed funding for a conference on the implications of the ETH work and noted a recently announced $5 million contribution to the Simons Foundation for workshops, conferences and programmes related to AI-generated research.
For businesses following AI research, the immediate implication is not that these claims are settled, but that model-generated technical output needs a deliberate validation process. Formalisation, public access to evidence, independent domain review and clear attribution are separate parts of turning a result into something that can responsibly inform decisions.