On October 26, 2026, OpenAI announced that its artificial intelligence models achieved breakthroughs in three major scientific domains: resolving the local smoothing conjecture for wave equations and the Kakeya conjecture in mathematics, disproving the exponential time hypothesis in computer science, and completing proofs concerning four-dimensional quantum chromodynamics and the Standard Model in physics. To release the findings, OpenAI appointed external academic experts as shepherds and established domain-specific disclosure paths. The company published raw mathematical outputs on GitHub, released the computer science findings alongside expert commentary, and withheld raw physics materials until an expert group including David Gross prepares a verification paper. The tailored rollout follows intense criticism from institutions such as the Institute for Advanced Study, which noted that even internal OpenAI teams entering the prompts could not fully grasp or reproduce the detailed proofs.
The revised protocol marks a significant quantitative pivot from OpenAI earlier publishing practices. In early October 2026, the company unilaterally released more than 700 pages of raw preprints touching on 372 mathematical problems without academic coordination. By narrowing the latest publication to three curated problems and involving external advisors, OpenAI sought to maintain tighter control over verification. However, academic researchers and mathematical advisory groups remain skeptical. Critics argue that granting privileged early access to a select tier of prominent scholars deepens information asymmetry across the research community while exploiting academic prestige for corporate promotion. Furthermore, scholars emphasize that outsourcing error-checking to external researchers without compensation violates basic scientific standards of peer review and reproducibility, especially when the underlying commercial models remain entirely closed.
This dispute highlights a structural bottleneck in foundational artificial intelligence research. While previous competitive benchmarks focused on inference computing power and the rapid generation of complex deductive formulas, the critical operational challenge has shifted to human intellectual verification. Evaluating unverified outputs from proprietary systems imposes heavy verification costs on research institutions, slowing institutional review and decision-making. Consequently, the ultimate value of applying advanced artificial intelligence to basic science will depend less on unilateral problem-solving claims and more on establishing transparent, trustworthy pipelines that the broader scientific community can independently inspect and validate.