On Monday OpenAI announced three high-profile discoveries that touch on long-standing open problems in mathematics, computer science and physics. The company said it would adopt a new model in which external experts, termed “shepherds,” help decide the most appropriate way to disseminate each result, marking a shift from earlier mass releases of AI-produced papers.
Each of the three findings follows a distinct rollout plan. The mathematical output was posted publicly, allowing researchers to organise their own analyses. For the computer-science breakthrough, OpenAI will convene an outside team to craft an explanatory framework. The physics result remains unpublished, with OpenAI deferring release decisions to a separate group of specialists.
Scientists familiar with the recent surge of AI-generated preprints expressed weariness after hundreds of papers appeared at the start of October. Many said the rapid influx already strained community resources, and that adding a layer of curated access could further delay critical peer review and validation of the claims.
Constantin Kogler, a researcher at the Institute for Advanced Study, welcomed the more manageable batch of three papers but warned that limiting early access to a select circle of scientists “doesn’t make any sense.” He noted that broader availability would let the community engage with the results without unnecessary bottlenecks.
Andreas Karch, a physics professor at the University of Texas at Austin, described the hand-picking of reviewers as “a disaster,” arguing that large AI labs already wield considerable influence. He suggested a competitive selection process might be fairer, but ultimately urged OpenAI to release the proof immediately so that the wider research community could begin independent verification.
The Advisory Group on Mathematics and Artificial Intelligence, which OpenAI had previously pledged to consult, issued a warning a day before the announcement. The group recommended that AI firms avoid selective distribution, citing risks of biased access, reputational borrowing, and delayed progress, and instead publish outputs on open scholarly repositories with full proofs and transparent selection criteria.
Hassett, speaking to The Verge, argued that open dissemination would not only improve scientific rigor but also accelerate the adoption of AI tools worldwide. By allowing multiple research teams to grapple with the findings, the community could more quickly develop robust methods for interpreting and extending the AI-generated results.