AI can now disprove decades-old conjectures and offer a claimed solution for a $1 million math problem. But in mathematician Terence Tao’s terms, it is still a knowledgeable and slightly drunk assistant. The danger lies not in what AI can do but treating its results like an oracle.
“It’s like having someone who knows a lot and is … slightly drunk and is sort of throwing out ideas,” mathematician Terence Tao remarked about artificial intelligence. “With enough guidance,” he says, “you can actually extract useful output.”
The artificial intelligence model is the machine that makes mistakes, that relies on the mathematician to sort the ten bad ideas from the one good one. Even if it can try to work through many different directions quickly—and it may be very good at that—the mathematician is often still the one judging what works and understanding why the result holds. And when that division of labor is hidden and AI is treated as if it is this all-knowing oracle, the cost is not wrong answers but a failure in understanding—results that no one has digested—and people are told that how we think no longer matters.
After the pocket calculator had become widely available, math teachers at the National Council of Teachers of Mathematics (NCTM) met in 1986 to discuss when to introduce it in education. Six thousand teachers showed up to the conference, and around twenty held signs outside the conference, protesting against the NCTM’s policy recommending the integration of calculators “at all grade levels in classwork, homework and evaluation.”
The protestors argued that calculators should only be used to apply skills that students have already learned, and warned against introducing them too early. Judging and understanding were to come before the tool. They did not argue for a ban: their signs read “turn off until upper grades.”
But an AI model, of course, is not a calculator, and calculators cannot act “drunk.” They are deterministic machines that output a fixed answer given a problem. In many ways, the drunkenness of AI can benefit mathematics instead. By May 2026, OpenAI stated that its models had disproved Erdős’s eighty-year-old planar unit distance conjecture, a problem for which Erdős himself had offered $500 and which a standard reference calls “possibly the best known (and simplest to explain) problem in combinatorial geometry.” In comments from nine mathematicians, Thomas Bloom called the result “surprising and impressive,” and W. T. Gowers, a Fields medalist, said that if a human submitted the paper to the prestigious Annals of Mathematics, he would have recommended acceptance “without any hesitation.” Gowers suggests that the gap between human mathematicians and AI is probably one of efficiency and not of kind:
…we are unlikely to be able to show that there is something we can do that current AI models cannot in principle do at all, but we might be able to show that there are things we can still do much more efficiently than those models.
Bloom partly attributes AI’s sudden breakthrough to the fact that not many mathematicians would be “seriously trying to disprove [the conjecture], despite the oft-repeated belief of Erdős that it is true.” In a way, that drunk assistant was able to go where sober humans did not. If Gowers is right about the gap between AI and humans, then the real question is not what AI can do, but how its results are used.
On the same day, the nine mathematicians—to whom OpenAI had sent the result before announcing it—posted a short, human-verified version of what OpenAI’s models wrote, and a later paper reported that the mathematicians had traced the key ideas to earlier work. Bloom noted that the disproof did not introduce powerful new geometric tools of the kind a proof of the conjecture would have needed, though human understanding of the disproof led four mathematicians, including Bloom, to disprove another long-standing Erdős conjecture. That is the kind of progress that comes from digesting a result instead of just accepting it.
But the AI is also drunk in the sense that it is not reliably correct. On 6 October 2026, OpenAI posted 722 manuscripts covering 372 problems, but within a day it withdrew three and revised 14 others. Within a day 17 errors were caught, and most of those manuscripts had not yet been checked by anyone. In its announcement, OpenAI stated that it had consulted an independent advisory group on how to release the results, and that it will fund workshops on understanding results produced by AI.
So is the objection to hallucination and the fact that AI gets some of the math wrong? Not quite: While some of the proofs were flawed, others were formalized in Lean, a programming language that checks whether a proof is correct. Only 162 of the 722 manuscripts had their main results formalized when they were released; the other 560 still depended on human review.
The gap is in understanding. When an AI company releases proofs faster than mathematicians can digest them, as OpenAI did with its 722 manuscripts, those proofs cannot enter the process that human-written ones enter—review from colleagues who are interested in understanding and building on the methods used. Checking these proofs takes specialists weeks or months. Workshops can help get people to read AI-generated proofs, but they cannot create the time or specialists needed to review hundreds of them.
Perhaps the most dangerous use of AI comes from treating it as something it is not: an oracle with convincing answers that can be accepted right away. “In many cases,” Melanie Matchett Wood writes in her comments on the disproof of Erdős’s conjecture, “it will be easier for AI to convince humans it has a proof than to come up with a correct mathematical argument.”
On 8 September 2026, OpenAI announced that its models—10,000 agents using around 130 billion tokens—had proved that the Navier–Stokes equations can break down in finite time. The Navier–Stokes problem is one of the seven Millennium Prize Problems, each of which carries a $1 million award, though OpenAI has stated that it would not claim the prize. The Clay Mathematics Institute, which awards the prizes, said that the problem has “apparently been settled” but that its review would be deliberately unhurried.
OpenAI’s own researcher Noam Brown said that nobody knows exactly what the 10,000 agents contributed, though he estimated that they accounted for less than 10 percent of the success. A paper posted on 6 October argued that the natural-language proof did not match what was formalized in Lean. There are still attempts to translate part of OpenAI’s manuscript into a “readable and accessible version.” Evidently, humans are still the ones doing the digesting.
Just five days before OpenAI claimed to have solved the Navier–Stokes problem, Terence Tao argued that problems in pure mathematics are primarily posed not for their solutions but because human-directed efforts to solve them and digest the solutions develop the field. He states that “prematurely” solving these problems with AI, especially without transparency in the process, can be a net negative for mathematics. In a later post, he warns that “indiscriminate use of powerful solution-extraction tools” solves the problems at hand at the cost of sustaining the field and, paradoxically, of “understanding the progress already obtained.”
When IBM’s Deep Blue beat then-World Chess Champion Garry Kasparov in 1997, that looked like the end of human chess. But it wasn’t. Chess surged in popularity in the early 2020s even after the prospect of a human ever beating an engine again was destroyed. The World Chess Championship still runs today, and chess players still value human judgment—understanding the moves, executing them, finding tactics your opponent couldn’t. Chess is played for its own sake, and mathematics claims to produce knowledge, which only makes understanding more important. For mathematics to mean anything, the understanding needs to happen in human minds.
It is still humans’ job to understand the move the computer finds, not just use it. “Drunk” assistants are important, but they are certainly not the only ones playing the game. If students still need to learn how to calculate after the advent of calculators, if calculating by hand still has use, then the art of the proof will remain as well. If all we can say about the result is the name of someone who could afford 10,000 agents that produced it, and not how the result was reached, then we have the answer without the mathematics. That is the modern calculator: It is what makes students believe that math is no longer worth learning, and that the work of humans will be replaced. That is not true.
