In mathematics a result becomes real when strangers can check it. That has not happened to the claim now absorbing the field. OpenAI says one of its models proved a problem that had resisted attack for years. The write-up, by the company’s own count, runs 1,147 pages. It has not been released.
The silence is doing most of the arguing. Formal verification groups that have examined the claim disagree about whether the conclusion holds, according to people familiar with those reviews. Nobody in the dispute questions that a language model can produce a very long argument. The question is whether the argument is correct at every step, and whether anyone outside the company can establish that independently.
Verification is a mechanical business, which is what makes the standoff odd. A proof of this scale is meant to be transliterated into a proof assistant, a program that checks each inference against formal rules and refuses a step it cannot derive. Run that process and the answer is yes or no. What the checkers report instead is subtler: some sections go through, some do not, and the ones that fail may reflect gaps in the formalization as easily as gaps in the mathematics.
The ambiguity is expensive, because a proof is an all-or-nothing object. Ten thousand correct steps and one invalid step is a failed proof, and locating the invalid one in a document of 1,147 pages is a search that grows with every page. Human referees manage long arguments by trusting structure, checking the load-bearing lemmas and skimming the routine ones. That habit assumes an author with a reputation to protect and a career that depends on not being caught. Neither assumption transfers cleanly to a model.
There is a commercial reason the claim was made loudly, and a different reason the document may be hard to publish. Labs have spent years hunting for tasks that survive benchmark saturation. Research mathematics is one of the few left where a strong result cannot be dismissed as contamination or memorization. A settled conjecture is the cleanest demonstration available. Publishing the artifact invites the one outcome a demonstration cannot survive, which is a checker that says no.
Consider what release would actually require. The text would have to appear in a form a stranger could validate, with enough detail about how it was produced that other teams could attempt the same run. If the proof emerged across many sessions with human steering between them, the boundary between what the model found and what its operators supplied becomes the interesting question, and it is also the least likely to be documented. Independent groups would need months to work through the pages.
The ordinary channel exists and is unglamorous. Mathematicians post to a preprint server, wait for comments, submit to a journal, and accept a referee report that may take two years. The process is slow because it is adversarial, and it works because the artifact is public long before anyone certifies it. A claim that skips the first step and asks for credit at the last is asking the field to invert its own procedure.
The charitable reading deserves a hearing. The proof may be sound and the field’s verification pipeline may simply be too slow for output at this scale. Formalization is specialist work, and translating a 1,147-page informal argument into machine-checkable form could occupy a team for a year. On that account the delay reflects capacity, and impatience reflects a misunderstanding of how mathematics absorbs results.
The historical comparison cuts both ways. When the classification of finite simple groups was completed, the manuscript was published, many volumes of it, and sat in libraries where anyone could read it, argue with it, and find the errors that were in fact found over the following decades. The standard was never that a proof arrive verified. The standard was that it arrive.
What this episode disturbs is where trust sits. For a century the referee was the gate. The gate is now an artifact, a machine-checkable file that a single company controls, and the institutions that certify mathematical knowledge have no procedure for a claimant who declines to hand it over. The result is a claim that circulates as reputation rather than knowledge, cited in fundraising conversations and argued about on mailing lists by people who cannot read the text.
The incentives run in one direction. Announcing a theorem costs nothing and buys months of favorable coverage. Retracting one costs a great deal. Between those prices sits a document nobody has seen, and every week it stays private, the claim’s value as marketing holds while its value as mathematics decays. That asymmetry is not evidence of bad faith. It is evidence of a gap between how labs report and how the field verifies.
The technical question underneath is narrower than the publicity suggests. A model that can generate a long, structured argument is doing something real, and the capability is worth measuring whether or not this particular proof survives. What remains untested is whether the same system can find its own error, or whether the discovery of a flaw still requires a human who understands the subfield well enough to suspect the lemma on page 800.
The near-term test is unglamorous. Release the document. Let the checkers work. Accept that the answer may take a year and may be no. Until then the accurate description is that a company has asserted a theorem and asked the field to take it on faith, which is the one currency mathematics has never accepted.
