On Tuesday, October 6, OpenAI published a repository of more than 700 machine-written papers claiming solutions to at least 372 long-standing mathematical problems — produced, it says, by an unreleased frontier model, in weeks, at a cost the company has put in the “several million dollars” range. The Economist promptly called the date possibly one of the most consequential in the history of mathematics. By the next morning, the Association for Human Mathematics had posted a rebuttal on Terence Tao’s blog, and 25 Fields Medalists — laureates spanning 1978 to 2026 — had signed on as initial endorsers of a declaration titled “A Severe Misalignment of AI in Mathematics.”

That is a great deal of credential, and credential is precisely what the reflexive right-of-center reader is trained to distrust. When an association puts “Human” in its name and issues a joint statement, the sound you are supposed to hear is a cartel creaking. This time I’d resist the reflex. The AHM is asking for something much smaller, and much more consequential, than job protection. It is asking who gets to own the word “proved.”

The Guild Wants Less Than You Think

Tao’s framing is the one to understand. Mathematics has three stages, he argues: generation, verification, and digestion — understanding, contextualizing, and explaining a result well enough that the community can actually use it. In an August talk at the Simons Foundation, he said models have gotten “pretty good” at the first two. The third is where the trouble lives. “We can have these 100,000-line proofs that we have to verify, but no one understands them,” he said. “Even the humans who entered in the prompts to generate the proofs might not understand them.”

Read that as a plea for tenure lines if you like. It is also, and more importantly, a demand for a legible audit trail. Verification that only the producer’s own machinery can perform is not a proof. It is a receipt.

”The Automated Verifier Says So” Is Not a Truth Condition

Here is the line that should be getting more attention. According to Nature’s coverage of the claim, OpenAI says the results have been validated using an automated verification method “becoming the standard for rigour in mathematics.” An automated method. Applied to output from the same company’s unreleased model. Disclosed to the public at the level of a marketing bullet.

This is the credit-rating problem in a lab coat. After 2008 we rewrote the rules for private firms whose model output had quietly acquired the force of public fact — because “the model says it’s fine” turned out to be a private opinion with systemic consequences. Nobody on the right objected to that in principle. Audit trails, reproducibility, adversarial checking: these were conservative asks before they became unfashionable. A verification standard that lives entirely inside one vendor is not a standard. It is a chokepoint with a press release.

A postdoc working on formalization put it to me at a departmental tea last Thursday, holding a cookie and a stack of grading: “The verification is probably real. But no hiring committee is going to read seven hundred papers, so the verification is going to end up being a line in a slide deck.” She isn’t wrong, and that is the problem.

The Field Has Been Faking It for Longer Than You Think

Now the part the AHM will not enjoy. The declaration concedes the capability question outright — its own premise states that the mathematical capabilities of LLMs “have improved dramatically, to the point that they can solve major outstanding problems in many fields of mathematics.” If the signatories will grant that, then the fight was never about whether the machine did it. It is about who certifies that it did.

And “nobody fully understands it” is not a new condition in mathematics. There are results sitting in the literature whose complete verification is a multi-decade community project, with errata, repair programs, and graduate seminars that exist precisely to reverse-engineer what the original authors meant. The field has never answered that by voiding the theorem. It answered by building better machinery — formalization, expository journals, refereeing norms that reward explanation alongside results.

So the honest objection cannot be scale. Seven hundred papers is not disqualifying; it is Tuesday in a field already drowning in output. The objection is ownership. Who holds the checker, and on what terms does anyone else get to run it?

What a Right-of-Center Position Would Actually Look Like

If you believe the 372 results, do not accept them on faith — that is the side of this argument you are supposed to be against. Demand the artifacts. Publish the verifier. Publish the prompt trail. Publish enough of the model that an outside group with no relationship to OpenAI can reproduce a nontrivial subset. Then call it mathematics. Until then, call it a demo: an extremely impressive one, with the volume knob at eleven, released by a company that, as the AHM statement notes, is simultaneously defending lawsuits alleging plagiarism, copyright infringement, and trademark dilution. That is a curious week to be asking the public for epistemic trust.

The uncomfortable bit for the labs: volume is not rigor. A dump of 700 papers is a publicity strategy wearing a corpus’s clothes. The uncomfortable bit for the guild: digestion is the cheapest of the three stages, and if the AHM wants to own it, it should be funding it — with the several million dollars, not with a statement.

Tao is right that proofs have to be explained. What he did not quite say is that “explained by us, at our pace, on our terms” is a claim about jurisdiction. Jurisdictions get renegotiated.

The 372 proofs may all be correct. That was never the interesting question.

Sources