AI has already proven itself to be useful at solving complicated Math problems, but some experts now believe that it is sufficiently good to cause a rethink of how Math itself is taught at a high level.
Few people are better placed to make that argument than Terence Tao. The UCLA mathematician — widely considered one of the greatest mathematical minds alive, and a Fields Medal winner at just 31 — has issued a pointed warning: the way we train the next generation of mathematicians may already be outdated.
“I think one urgent portion of our mathematical system we need to evaluate right now is graduate education,” Tao said. “Because the type of problems that we give to graduate students to train them to become good mathematicians now has extremely large overlap with the sort of problems that AIs can plausibly solve.”
The concern is not abstract. The traditional PhD pipeline in mathematics works by handing students problems of increasing difficulty — problems designed to build rigor, intuition, and originality. The publication of early research papers is treated as a rite of passage, proof that a student has arrived. But Tao suggests that this model may now be measuring the wrong things.
“This is actually a real problem — or maybe it is not. I mean, it’s currently a problem because we are putting so much value on pushing out papers, on having these initial publications.”
The culture of early publication, in other words, creates a perverse incentive at exactly the moment AI is best equipped to help game it. If the goal is simply to produce a result that can go into a journal, and AI can plausibly produce that result, then the training exercise stops being a test of mathematical thinking and becomes something closer to a workflow.
Tao’s proposed response is deliberate and somewhat counterintuitive: keep assigning problems that AI could technically solve, but insist that students work through them unaided. The struggle, not the solution, is the point.
“We may have to change our culture a little bit at the graduate education level — where we are giving graduate students problems which technically could be solved with AI assistance. But the whole point is to do them without the assistance.”
He acknowledged that this is easier said than done. “It’s hard to even articulate how we would make that shift. But this is something which I think we have to address pretty soon.”
The urgency in Tao’s words reflects a broader moment in the relationship between AI and high-level intellectual work. Tao himself has been among the more thoughtful voices tracking this shift. As recently as September 2024, he described AI’s mathematical abilities as comparable to “a mediocre, but not completely incompetent, graduate student.” By early 2026, his assessment had changed substantially — at a conference at UCLA’s Institute for Pure and Applied Mathematics, he declared that AI is now “ready for primetime” in math and theoretical physics, because it “saves more time than it wastes.”
The implications extend well beyond academia. Citadel CEO Ken Griffin recently said that AI is doing the work a PhD does in months in just days — and that these are not mid-tier white-collar jobs being automated, but “extraordinarily high-skilled” ones. That shift is happening in finance. In mathematics, the equivalent disruption is arriving in the graduate school, where the very problems used to forge expert mathematicians are now within the reach of a capable AI model.
Fei-Fei Li, the Stanford AI pioneer, has similarly argued that degrees are becoming less important relative to the ability to learn and adapt with new tools — a sentiment that, applied to mathematics, raises the same uncomfortable question Tao is wrestling with: if the credential is the paper and the paper can be AI-assisted, what is the degree actually certifying?
Tao’s answer, implicitly, is that graduate education must pivot toward cultivating something AI cannot replicate — judgment, taste, and the capacity to choose the right problem in the first place. He has said elsewhere that as AI drives down the cost of routine problem-solving, “the scarce skill becomes choosing the right problem, designing the workflow, and checking the result.” That is a different kind of mathematician than the ones current PhD programmes are built to produce, and building the institutional infrastructure to train that kind of mathematician — without clear metrics, without easy publication benchmarks, and against the grain of decades of academic culture — is the challenge Tao is flagging. It is not a small one.