Harker: AI cut the entry-level training subsidy

Harker brings a decade at the Federal Reserve Bank of Philadelphia to the question of whether AI is destroying entry-level white-collar work. He notes that job postings for the occupations most exposed to AI peaked and began to fall in March 2022, when the FOMC started raising rates, eight months before ChatGPT existed. Zanna Iscenko and Fabien Curto Millet, analyzing 238 million job postings, tied that posting decline to the tightening cycle, and Harker adds that the Economic Policy Institute found young workers without college degrees, whose occupations score negative on AI exposure, saw unemployment rise at a similar pace. The Stanford researchers themselves, he points out, describe their findings as descriptive patterns, not causal estimates, and say they do not see widespread, economy-wide job displacement associated with AI.

Harker acknowledges it is too early to know what will happen as AI is deployed, because overlapping shocks cannot be separated in real time. But he says one fact is clear in the data: firms are not firing junior employees, they are hiring fewer of them, and the decline concentrates where AI automates work rather than complements it. He argues this may explain why the Stanford employment gap persists even as rates have come down. The junior role, he writes, was never primarily about output productivity; a first-year associate's document had to be redone and a resident's 2 AM patient history retaken, but firms bought that work because it was how a junior became a senior partner.

AI can now do that junior work, so firms invest less and the pipeline thins. Harker cites Matt Beane's observation that surgical robots cost residents case time years before ChatGPT. He argues employers should treat junior hiring as a capital investment mislabeled as an operating expense, and universities face the same problem, with Nature warning this spring that early-career researchers risk having training tasks done by a machine. The stakes, he suggests, hinge on whether employers keep cutting formation budgets and universities keep certifying machine-done work, since the postings data will recover when the hiring cycle turns, but the old bargain in which production paid for formation will not rebuild itself.