OpenAI’s attempt to transition from a disruptor to a collaborator in the scientific community has hit a new point of friction. On October 26, 2026, the company announced three significant AI-generated breakthroughs across mathematics, computer science, and physics, while simultaneously introducing a “shepherding” model designed to manage how this data enters the academic record.
The technical claims are substantial. OpenAI reported that its models have disproved the Exponential Time Hypothesis (ETH), a foundational conjecture in computational complexity that suggests certain problems cannot be solved faster than exponential time. In mathematics, the company claims to have proved Christopher Sogge’s 1991 local smoothing conjecture in all dimensions, a result that effectively resolves the Kakeya conjecture. In physics, the AI reportedly produced existence proofs related to Yang-Mills theory.
Rather than publishing finished peer-reviewed papers, OpenAI is framing these results as raw research data and handing them to “shepherds”—external academic leads who will guide the community’s verification process. For the computer science results, these shepherds include experts like Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams. The physics data is being directed toward a group led by Nobel laureate David Gross.
While OpenAI asserts it seeks no academic credit for these discoveries, the new model is fueling concerns about information asymmetry and scientific gatekeeping. By choosing which academics receive early or guided access to the data, critics argue the company is still exerting significant control over the scientific narrative and determining who gets to participate in the resulting breakthroughs.
This controlled release is a direct response to the “scoopolypse” of October 6, 2026, when OpenAI unexpectedly dumped 722 mathematical manuscripts onto GitHub. That event caused widespread panic in the math community, as researchers realized their long-term projects might have been rendered obsolete overnight by an automated system. NYU professor Tristan Buckmaster noted that such bulk releases have the potential to “wipe out” the research programs of early-career mathematicians who lack the resources to compete with AI-generated output.
The current “shepherding” approach for physics results specifically involves deferring the release of certain materials to the expert group rather than making them immediately public. While intended to prevent chaos, this selective disclosure creates a hierarchy where a small group of hand-picked experts holds the keys to foundational data.
To mitigate some of the administrative burden on the community, OpenAI has pledged $5 million to the Simons Foundation. This funding is intended to facilitate independent review and help the academic community digest the sheer volume of AI-generated proofs. However, for many in the field, the financial contribution does little to solve the core problem of a private corporation effectively “owning” the pace of discovery in fundamental science.
