Researchers Criticize OpenAI for Gatekeeping Its Breakthrough Results
OpenAI’s latest attempt to share major AI research results has drawn both cautious respect for the scientific breakthroughs and deepening frustration over how the company controls access to them. On Monday, OpenAI announced new results in mathematics, computer science and physics, describing them as extending earlier work and opening new ground. The company said it is experimenting with communication approaches and is working with outside experts, which it calls shepherds, to determine how each finding should be made available. The three disclosures reflect different strategies. For the mathematics result, OpenAI released the AI-generated output and left researchers to organize their own efforts to interpret it. For the computer science result, it plans to convene an external group to explain the finding. The physics result has not been released, and OpenAI said it will defer to an external group on how to share the materials. Several researchers characterized the underlying results as significant and potentially major breakthroughs, saying they address consequential unsolved problems. But many also argued that OpenAI’s release method risks undermining trust and slowing scrutiny. The concern is not merely the volume of prior unfiltered preprints dumped on communities in October, but the perception that a private lab is choosing which external scholars may first inspect or explain findings, even as the work is publicly announced. Researchers including Constantin Kogler of the Institute for Advanced Study welcomed the smaller batch of three results but questioned why only a handpicked group of scientists would initially examine them. Andreas Karch of the University of Texas at Austin said allowing large labs to select privileged reviewers creates an unhealthy concentration of power, and suggested the underlying proofs or outputs should be made public immediately even if an advisory group is also formed. The Advisory Group on Mathematics and Artificial Intelligence, which OpenAI had pledged to consult, warned against the practice the day before the announcement. It said selective early access is unnecessary, risks biased selection and delays, borrows the reputations of chosen experts, and should be avoided in favor of open scholarly repositories, formal proofs, clear problem selection details, and support for broader verification. After the latest release, the group urged OpenAI to publish the natural-language and Lean outputs behind its physics results without delay. Others argued open access could improve science, broaden tool adoption and reduce reliance on curated channels. Yet OpenAI’s situation is difficult: researchers say any approach will leave someone dissatisfied, while some contend the deeper issue is that fundamental results are being produced so rapidly, often in difficult-to-verify formats, without substantial human contribution. The company has acknowledged tradeoffs and says it is experimenting, but the episode has reinforced growing exhaustion among academic communities with being asked to validate AI-generated breakthroughs under conditions they see as uneven and overly controlled.
