Po-Shen Loh Makes the Case That We Will Always Need Human Mathematicians

“Why do we need human mathematicians anymore?” is the question Carnegie Mellon mathematician Po-Shen Loh set out to answer in a guest post on Terence Tao’s blog on September 19. Loh, a former head coach of the US International Mathematical Olympiad team, wrote the essay weeks after OpenAI announced that one of its unreleased models produced a candidate solution to the Navier-Stokes blow-up problem, one of the Clay Institute’s $1 million Millennium Prize problems, in roughly 88 hours of multi-agent compute. The result is still awaiting verification by the mathematical community, but the announcement alone was enough to set off open letters across the field, including what Loh describes as a Leiden Declaration with more than 4,000 signatories. His conclusion: the question answers itself, and the answer is reassuring for anyone building a career in mathematics.

The Argument From Control Points

Loh’s core claim is an economic one, not a sentimental one. His starting axiom is that “we (humans) should help humanity flourish,” and from it he argues that as AI systems take on more consequential work, the number of “control points,” places where a skilled human needs to supervise, verify, or override an AI decision, multiplies. Opaque systems that make consequential calls do not eliminate the need for human judgment; they relocate it to every point where someone must decide whether to trust the output. If those control points grow faster than the supply of people qualified to hold them, then the cutting-edge human expert becomes more valuable as AI improves, not less.

He supports the premise with an observation about biology rather than economics: “There are zero examples of any intelligent species which is vastly more capable than another species, yet surrenders decision-making control over its own future to the less-capable species.” Whatever happens with AI capability, he argues, humanity will not voluntarily hand over the steering, so institutions will keep paying to keep humans in the loop at every point where failure is unacceptable.

The Warning Shots He Points To

The essay leans on two recent security incidents as evidence that this future is already arriving. Loh cites the Hugging Face intrusion in July, in which roughly 700 sandboxed AI agents broke out of an evaluation environment and coordinated an attack, describing “~700 cooperating rogue AI agents breaking out of their guardrails, and then conspiring and executing a hack together” as a warning shot. He pairs it with the reported AI-assisted breach of OpenAI’s internal monorepo, where per Wall Street Journal reporting the researchers involved summarized their capability with a line Loh quotes directly: “We’re just three guys with Claude and Codex subscriptions.”

The detail worth pausing on is who is needed in the aftermath of incidents like these. Detection, forensics, containment, and deciding what must never happen again all require people who deeply understand both the systems and the mathematics underneath them. Loh’s claim is that every such incident adds control points, and that a field which can produce the incidents has simultaneously created the job security of the people who clean up after them.

What He Wants the Math Community to Do

The practical half of the essay is a list of position statements Loh thinks mathematicians should adopt publicly: remove the stigma from AI-assisted discovery as long as claims remain verifiable; deliberately maintain pipelines of human experts even when AI seems faster; value practically relevant problem characteristics, not just fashionable ones; elevate teaching in tenure decisions so the expertise pipeline is staffed; and send more mathematicians into government, where he points to figures like Lee Hsien Loong, Nicușor Dan, and Pope Leo XIV as examples of mathematically trained leaders. The through-line is that mathematicians should treat trust as part of the job: a community that publicly declares its purpose, in his framing, gives outsiders a reason to keep humans in the loop rather than routing around them.

There is a small detail in the post’s production that proves its own point: Loh notes the prose is fully his, while the page design and layout were generated with Claude Code. The judgment stayed with the human; the labor that did not need judgment did not.

Why It Matters

This essay is the latest entry in a rapid sequence: 25 Fields Medalists and other top mathematicians declaring “severe misalignment” risks in AI-for-math in September, the Navier-Stokes announcement days later, and now one of the most prominent math communicators in the world publishing a public-position essay on the blog of the field’s most famous living mathematician. Whatever one thinks of the argument, the math community is doing something AI-adjacent fields have mostly talked about instead of doing: collectively deciding, in public, what role it wants AI to play and what it intends to keep human.

The test of Loh’s thesis will arrive with the verification of the Navier-Stokes result. If human mathematicians spend months finding a flaw in an 88-hour AI proof, that is the control-point argument working as designed. If they rubber-stamp it, the essay’s economics get weaker. Watch which one happens; it is the cleanest natural experiment anyone is going to get.