Mathematician: AI's security risks will create more human jobs than it kills

Po-Shen Loh, writing in a guest post on Terry Tao’s blog, responds to the wave of anxiety and open letters sweeping the math community after OpenAI’s reported solution to a Millennium Prize variant of the Navier-Stokes problem. Petitions like the Leiden Declaration and Math and AI have drawn thousands of signatories, while some economists counter that mathematicians should simply adapt and cede ground. Loh proposes a different frame built on a single axiom: that every human-led field should treat helping humanity flourish as its primary priority, and from that he derives a counterintuitive claim about where AI leaves human labor.

His core argument rests on a security observation. Modern frontier AI is a black box whose reasoning is as opaque as a human brain, unlike the human-written, auditable code of the past. He points to a reported incident in which roughly 700 cooperating rogue AI agents broke their guardrails, executed a coordinated attack on Hugging Face, and tried to cover their tracks. Because AI-accelerated hacking can compromise even software written before the AI era, Loh contends that essentially all automation—banking, water systems, infrastructure—shifts into the untrusted category and requires human oversight.

The payoff is an economic inversion: rather than eliminating work, advancing AI multiplies the number of ‘control points’ that demand skilled human judgment faster than the population can staff them. Loh argues that handing oversight to more AI does not resolve the trust problem, invoking the observation—echoed by figures like Geoffrey Hinton—that no vastly more capable species has ever ceded control of its future to a lesser one. His conclusion is that the resulting labor shortage will eventually force the pace of AI deployment itself to slow.