OpenAI Research Transparency: Open Access Now! 🔍
OpenAI research transparency: explore AI-generated research, open scientific research, and AI science verification—learn how to share results ethically and openly.
OpenAI's Science Breakthroughs Raise a Hard Question: Who Gets to Check Them?
AI systems can now help with work on problems that have stumped researchers for years. But a result only matters if other people can look at it closely, test it, and understand how it was produced. That's the uncomfortable tension behind OpenAI's latest announcements in mathematics, computer science, and physics: the claims are big, yet what's available to the broader community doesn't always line up across fields.
OpenAI says it's using expert "shepherds" to decide how each result is shared. The idea is to make dense research easier to explain. But researchers quoted by The Verge push back on the premise: if a company controls who gets early access, should that really happen when public scrutiny is what determines whether the work holds up?
This isn't just academic wrangling. It's a practical problem for any AI lab releasing consequential work: how do you balance accuracy, safety, clear communication, and equal access? Below, I break down OpenAI's three announcements, why selective review can backfire, and what a more accountable disclosure process could look like.
What did OpenAI announce?
OpenAI announced three results in mathematics, theoretical computer science, and physics. The company says the sharing plan varies depending on the result, based on advice from expert shepherds. That variation matters. An announcement, a publication of evidence, and independent verification aren't the same step - even if they get described like one story.
Mathematics: a claim about the Kakeya conjecture
In mathematics, OpenAI's announcement claims to resolve the Kakeya conjecture in all dimensions. The basic intuition behind the problem is about how little space you need to rotate a needle through every direction. It's a deep question in geometric measure theory - not something with an obvious "try this and you're done" answer.
OpenAI released AI-generated output for the result and asked researchers to help explain it. Brendan Hassett, a Brown University mathematician, told The Verge that the claim looked well supported - though he also said he hadn't checked the Lean formalization yet. Lean is software used to represent mathematical proofs in a way that can be verified by machines.
And that difference is real. A plausible proof and a fully verified proof are different stages. A formalization can catch some errors, but the work still has to be examined and understood by people who know the domain well. So the result should be treated as a serious claim that still needs scrutiny, not as settled science just because it was announced.
Computer science: a challenge to the exponential time hypothesis
OpenAI's computer science result claims to disprove the exponential time hypothesis (ETH). ETH is a widely held assumption that certain computational problems can't be solved much faster than brute force. It also connects to the famous P versus NP question, which asks whether problems with solutions that are easy to check are also easy to solve.
OpenAI said it would bring in an external group to explain the finding. That could make the work more legible to people outside the immediate niche. But "explain" and "verify" aren't interchangeable. A group can produce a helpful write-up and still not be the only group capable of inspecting the underlying argument.
Physics: work on quantum chromodynamics and Yang-Mills theory
The physics announcement concerns quantum chromodynamics (the theory of the strong nuclear force) and its relationship to Yang-Mills theory, a core framework in particle physics. Yang-Mills theory is also tied to one of the seven Millennium Prize problems in mathematics.
At the time of the announcement, OpenAI hadn't released the physics materials. The company said it would defer to outside experts on how to share them. The announcement also doesn't address the unresolved mass gap problem - an important detail to clarify when people are assessing how significant the work really is.
Across all three areas, the evidence available to the broader research community isn't the same. That makes it harder to evaluate the claims together, and on equal terms. It also matters that some words get treated like synonyms when they aren't: announced, published, independently checked, and accepted describe different realities.
Why is expert-led access controversial?
The strongest case for expert shepherds is pretty straightforward: frontier research is hard to communicate, and a result without a clear explanation can be misunderstood. Specialists can point out what's genuinely new, prepare documentation that others can actually use, and help plan releases that involve real technical or safety considerations. In that sense, involving experts can improve the process.
The worry starts when "guidance" turns into a gate. If a company selects a small group to see or interpret results before everyone else, it's effectively deciding who gets time to dig in, who helps shape the first public explanation, and whose expertise gets treated as the reference point. Even if the chosen researchers are legit, the process can still create unequal access - and give the impression of endorsement.
Selection can affect the science, not just the optics
Independent review works best when researchers can judge the claim based on the same underlying evidence. A limited first look can make it harder for other people to reproduce an argument, spot a missing assumption, or use a different approach to test it. It can also concentrate effort in a small group, which can slow down scrutiny when the result needs a wider range of expertise.
OpenAI previously pledged to consult the Advisory Group on Mathematics and Artificial Intelligence (AGMAI). That group warned against giving selected researchers first access. Its objections included unnecessary delays, biased selection, and the risk that the company might borrow the credibility of whoever it chose. AGMAI recommended sharing outputs through scholarly repositories, alongside formal proofs and information about how the problem selection process worked.
A public announcement creates a reasonable expectation
Researchers quoted by The Verge were more positive about OpenAI's smaller batch of results than an earlier release of hundreds of preprints. Still, their objection wasn't simply that OpenAI involved experts. The issue was that major claims were public while access rules left most researchers waiting.
This came through most sharply when an announced result came with a full formalization. After OpenAI's physics announcement, AGMAI urged the company to release the natural-language and Lean outputs without delay. Their argument was that the stated reason for holding things back - the idea that releasing those materials depended on completing the formalization - no longer applied.
No disclosure model will satisfy everyone. Jared Duker Lichtman, a Stanford mathematician, noted that every approach is likely to draw criticism. But approaches aren't all the same. A process should be judged by whether it gives people enough information to evaluate the claim, clearly explains why any restriction exists, and avoids turning access into something decided by a company's personal choice of reviewers.
What would responsible disclosure look like?
Responsible disclosure isn't the same thing as uploading a technical file and calling it "open." A result can be technically available but still practically unusable if key assumptions, methods, or evidence are missing. And a careful explanation can still be insufficient if people can't actually check it. A useful release needs to address both access and usability.
Separate the evidence from the explanation
The underlying research materials should be available to qualified reviewers - and, when there's no concrete safety or legal reason to limit them, to the wider research community. Explanations, summaries, and talks should support that access, not replace it.
For AI-generated work, the release package should spell out what the system produced and what people did afterward. When relevant, it should include the prompts or inputs, the tools and formalization used, what checks were performed, what's known to be missing, and what's still incomplete. Those details help reviewers separate what the AI contributed from what human interpretation and verification added.
Make restrictions specific, limited, and reviewable
Sometimes a staged release makes sense. A concrete risk might require withholding a particular capability or operational detail. But "the work is complicated" isn't enough on its own. If something is hard, the solution is better documentation and support - not automatic exclusion.
If a restriction is necessary, the lab should say exactly what's being withheld, why it's being withheld, who made the call, and when it will be reconsidered. It should also give a clear schedule for wider access. Without those details, a temporary review process can quietly turn into an indefinite bottleneck.
Fund scrutiny, not just announcements
Even when research is public, it doesn't mean everyone can assess it right away. People may need time, technical support, and sometimes funding to reproduce or formalize difficult results. That's why OpenAI's critics have argued for funding that supports the broader effort to understand the findings - not just a small group tasked with producing an official explanation.
A practical release checklist might include:
- The full claim and the precise limits of what it establishes.
- Natural-language papers, formal proofs, and relevant Lean outputs.
- A description of how the problem was selected and how the result was produced.
- Known uncertainties, failed approaches, and verification work that's still outstanding.
- A public channel for corrections and independent analyses.
- A stated timeline for any materials that are temporarily restricted.
This won't guarantee correctness. What it does do is make errors easier to find and clarifies what level of confidence the evidence supports. That's more meaningful than a polished announcement or a list of distinguished names.
What should AI labs and research institutions do?
The debate has a straightforward lesson for AI companies, universities, and funders: design the review process before a major result becomes a communications event. If a lab decides who sees a finding, how it will be explained, and when the evidence becomes public only after an announcement, it's likely to create needless friction.
For AI labs: treat release as part of the research
Labs should budget for verification and comprehension as part of the research work. That includes identifying external expertise, preparing materials others can reproduce, documenting the system's role, and setting a release schedule. If a result is hard to read, that isn't automatically a reason to restrict it - it's a reason to invest in better explanation and better supporting materials.
When a selective review group is useful, define what it's supposed to do. Is it checking technical correctness? Advising on communication? Assessing a safety risk? Coordinating follow-up? Those are different jobs. Naming the job helps prevent a group assembled to explain a result from being mistaken for a body that independently certified it.
For researchers: separate access from endorsement
Researchers can keep their questions grounded instead of treating the debate as a binary choice between openness and caution. What evidence is available right now? What's still private? What checks have actually been completed? Who picked the reviewers, and what authority do they have?
Those questions help separate a legitimate safety measure from a process that mostly controls the narrative.
Researchers should also be precise in their own assessments. If a proof seems plausible but hasn't been checked, say so. If a formalization exists but hasn't been inspected independently, say that too. Clear status labels help both the research community and the public avoid turning an exciting claim into an unsupported "this is settled."
For institutions: set shared norms before the next announcement
Universities, scholarly societies, and funders can help define minimum expectations for AI-assisted research disclosure. Those expectations should cover attribution, access to evidence, independent review, and support for replication. They should also allow exceptions when a specific risk justifies limited release - but still require a reason and a review date.
The trust and business consequences aren't theoretical. If labs keep announcing breakthroughs without making them meaningfully checkable, researchers may become less willing to spend their time on claims that never quite let them verify them. Public confidence can erode, too. If labs release complex work without context or support, people end up with confusion rather than useful adoption.
The goal isn't "maximum openness" as an abstract principle. The point is to build a process that lets the right people test the work - without turning access into a private privilege.
Key takeaways
- OpenAI announced three significant claims, but the evidence was shared differently across mathematics, computer science, and physics.
- Expert guidance can improve explanation; company-controlled access can still lead to unequal scrutiny and the appearance of endorsement.
- An announced result is not the same as an independently verified or accepted result.
- Strong disclosure pairs accessible evidence with clear documentation, known limitations, review timelines, and support for independent assessment.
- Access should only be restricted for a specific reason, and the scope and endpoint of that restriction should be clear.
Conclusion
AI-generated research is becoming consequential enough that how it's released matters almost as much as the result itself. OpenAI's latest announcements show why expert guidance can be helpful - and why it doesn't fully resolve the question of who gets to inspect the evidence and when.
For labs, the practical standard is clear: publish the supporting materials, explain what's still uncertain, separate review from endorsement, and justify any access limits. For everyone evaluating a breakthrough, hold the claim at arm's length until independent scrutiny catches up.