Who checks the math if only OpenAI can run it

OpenAI has released solutions to 372 families of long-standing mathematics problems, generated by an unreleased frontier model and distributed across 722 manuscripts. The work has impressed mathematicians and unsettled them in equal measure. The unsettling part is not the results themselves — it is the structure through which they arrive.

Thirty years ago, arXiv shattered academic publishing by letting physicists post papers directly online, skipping journal gatekeepers entirely. Anyone could download the preprint, check the proofs, replicate the work. Verification became a collective act. Truth was public by default. That revolution built the modern internet's relationship with knowledge — open, verifiable, distributed.

OpenAI's strategy inverts this. A mathematician can see that a proof exists. She cannot run the system that generated it. She cannot probe the reasoning, test edge cases, or ask why the model took one path instead of another. To verify OpenAI's mathematics requires access to OpenAI's proprietary frontier model, which means commercial terms, rate limits. The company's permission to exist. The AGMAI advisory group. Formed to examine research ethics. Apparently raised concerns the excerpt does not quite name. The concern is structural: verification has moved behind an API.

This matters because mathematics was supposed to be the one domain where access didn't matter. A proof is either true or false regardless of who checks it. But a proof you cannot independently examine is not quite a proof in the old sense. It is a claim made by a powerful system, from inside that system, that can only be verified by running that same system again. The asymmetry is different from journal gatekeeping. Journals at least published the full work. But it points in the same direction: certain people can verify. Certain people cannot.

For researchers, funding committees, and anyone building on these results, this creates a dependency that looks like progress. For the structure of how mathematical knowledge gets validated, it is a quiet reversal. When arXiv opened the papers, it democratized verification. This locks it away behind a commercial moat, just wearing the language of openness. The mechanism, once named, makes the problem obvious: you cannot build knowledge collectively on a foundation you cannot inspect.