OpenAI Picks "Shepherds" to Vet Its Latest Math, Physics, and CS Claims, and Researchers Aren't Thrilled

OpenAI announced three new research results this week in mathematics, computer science, and physics, each paired with a different disclosure strategy decided by outside expert "shepherds." It's a notable shift from the company's previous approach of mass-releasing AI-generated preprints with little curation. But if OpenAI expected this to smooth things over with the research community, it hasn't quite worked out that way.

The backdrop here matters. At the start of October, OpenAI released hundreds of preprints on mathematicians with essentially no triage, a move one researcher at the time called "pure insanity" given the sheer volume of work needed to even begin checking it. This week's announcement is OpenAI's attempt at a more deliberate process, but the deliberation itself is now the problem for a lot of scientists.

Here's what OpenAI actually claims to have found:

• Mathematics: A purported resolution of the Kakeya conjecture in all dimensions. The problem asks how little space you need to rotate a needle through every possible direction. A proof of the three-dimensional case was part of why NYU mathematician Hong Wang won a Fields Medal earlier this year, so a full all-dimensions resolution would be a big deal. Brown University's Brendan Hassett told The Verge the claim "looks well supported," though he hadn't yet checked the Lean formalization, the machine-checkable proof format used to verify this kind of work computationally.
• Physics: Work touching the Standard Model, specifically quantum chromodynamics (the theory describing the strong nuclear force) and its relationship to Yang-Mills theory, a foundational framework in particle physics. Yang-Mills also underpins one of the seven unsolved Millennium Prize problems, though OpenAI's result doesn't touch the famous "mass gap" piece of that puzzle.
• Computer science: A claimed disproof of the exponential time hypothesis, the long-standing assumption that certain computational problems can't be solved meaningfully faster than brute force. This one is closely tied to P vs NP, arguably the biggest open question in theoretical computer science.

For the math result, OpenAI released the AI-generated output and is letting the community organize its own verification effort. For the computer science result, it's assembling an external group to help explain and vet the work. The physics result hasn't been released at all, OpenAI says it's deferring entirely to an outside group of experts on how and when to share it.

The Access Problem

That tiered, expert-gated rollout is exactly what's generating friction. Researchers The Verge spoke to broadly took the underlying science seriously, calling the results "huge," "major breakthroughs," and "significant." The frustration isn't with the math. It's with who gets to see it first.

Constantin Kogler, a researcher at the Institute for Advanced Study, welcomed the smaller batch of three results compared to the hundreds dumped in October. But he pushed back on OpenAI choosing who gets early access. "It doesn't make any sense that first only a few distinguished scientists can look at them," he said.

Andreas Karch, a physics professor at UT Austin, was blunter: "The fact that they handpick the group is a disaster. The large labs have enough power as it is. They should not be picking winners and losers." He floated a competitive selection process instead, though he admitted it would probably land on a similarly qualified group anyway. His bigger ask was simpler: just release the proof now, alongside the expert panel. "This would democratize access while still showing that they care about making the results meaningful and understandable," he said.

This isn't a fringe complaint. The Advisory Group on Mathematics and Artificial Intelligence, a panel OpenAI had previously committed to consulting, published a warning against exactly this kind of handpicked early access just a day before OpenAI's announcement. AGMAI argued the practice is unnecessary, improperly borrows the credibility of the selected researchers, and risks unequal access, biased selection, and unnecessary delay. Its preferred model: dump outputs on scholarly repositories with full Lean formalizations and transparent selection criteria, then fund the community effort to actually digest it.

After OpenAI's announcement, AGMAI specifically called for the physics result's natural-language and Lean outputs to be released immediately, noting that OpenAI's own stated reason for delay, an incomplete Lean formalization, no longer applies by the company's own account. The group's statement didn't mince words: "understanding should develop in the open, with everyone given the same access."

There's also a practical science argument for openness beyond fairness. Hassett pointed out that broader access accelerates adoption of whatever new techniques underlie these results. "Having many groups around the world wrestle with these results supports rapid adoption of new tools by those that would like to use them," he said. Lock the outputs behind a small reviewed circle, and you slow down exactly the kind of distributed scrutiny that makes a result trustworthy in the first place.

Not everyone thinks there's a clean solution here. Stanford mathematician Jared Duker Lichtman framed it as damage control: "My sense is that for any approach, there is going to be someone who is upset. So it's just a question of finding the least offensive option." That's a pretty low bar for a company claiming to have cracked some of the hardest open problems in several fields simultaneously.

Georgia Tech's Xiaoyu He went further, arguing the release strategy is a side issue. The real problem, he said, is structural: results that would normally take a research community years to produce and verify are now arriving faster than humans can meaningfully engage with them. "The results are very impressive," He said, "but I think they are making a show of trying to be responsible and thoughtful about how they release results while not being able to change the main issues: that fundamental results are being solved so quickly without human contribution, and that the papers are difficult to read and verify."

Key Takeaways

OpenAI's shepherd model is a step away from the chaotic mass-dump approach that defined its October release, but it trades one problem for another. Researchers broadly trust the underlying science, several independently described the results as major, yet they're increasingly skeptical of a process where a frontier lab decides who gets to check its homework first.

The tension boils down to a pretty simple question: is curated, expert-mediated disclosure actually safer science, or is it just a more polished form of gatekeeping? AGMAI's position is unambiguous, release everything, let the whole field work on it. OpenAI's framing, that it's "experimenting" with disclosure approaches, suggests the company doesn't have a firm answer either. Given how fast these claimed breakthroughs keep arriving, that's a question the field is going to have to resolve soon, whether OpenAI leads that conversation or not.