Unemployment among 22-to-27-year-olds in AI-exposed occupations, analysts, accountants, judicial clerks, has climbed since 2023. Not collapsed. Climbed. That distinction matters if you are trying to argue that AI is just a productivity multiplier with no victims. The victims exist. They are just young and easy to ignore.
I have watched this pattern before, specifically in software. When GitHub Copilot shipped, senior engineers got faster. Junior engineers got fewer offers. The tool did not replace the senior developer who could review its output, catch the hallucinated API call, and understand why the suggested refactor would cause a race condition at scale. It replaced the ticket-grinding work that used to justify hiring someone at year one. White-collar knowledge work is now running the same script.
The Hollowing, Not the Flattening
Taylor St. Germain at ITR put it plainly: AI is hitting white-collar industries harder than blue-collar ones. White-collar layoffs are above their 10-year average. Manufacturing and construction are below it. The irony is sharp. The workers everyone assumed automation would displace first, the ones on factory floors, are currently more insulated than the ones with graduate degrees and Slack accounts.
Morgan Stanley's Diego Anzoategui argues the same technology that automates tasks can augment workers and boost demand in AI-exposed sectors. He is not wrong. The senior analyst who can prompt well, validate outputs, and synthesize across sources is genuinely more productive than she was in 2022. Her salary should reflect that. The problem is the pipeline feeding her role just got cut.
Janet Vertesi at Princeton called it "effacing expertise instead of enabling expertise." That framing is more precise than the standard augmentation pitch. You cannot augment your way to senior judgment without the years of grinding through the junior work that builds it. Giorgio Ascoli made the same point about scientific training: skip the hands-on learning and you cut your own roots. The white-collar version of that is a 28-year-old who never had to build a financial model from scratch because Copilot did it, and now cannot tell when the model is wrong.
What Actually Breaks in Production
Jade, an insurance tech worker in Raleigh, described her job as "AIs talking to each other, with barely a human involved." She built the automation herself. She knows it could replace her. That psychological weight is real, and 71% of Americans share some version of it. I will grant the optimists one point: historically, automation has expanded total employment over long time horizons. The industrial revolution did not produce permanent mass unemployment. That is a fair read of the data.
But the transition costs are not evenly distributed, and "long time horizons" is cold comfort if you graduated in 2024 and cannot get an entry-level role because the work that role used to do now costs $20 a month in API calls. The historical pattern of net job expansion does not guarantee that the people displaced by this wave are the ones who benefit from the next one.
The practical move for companies is to stop treating junior roles as pure cost centers and start treating them as the training infrastructure for the senior talent they will need in 5 years. That means deliberately preserving some of the work that AI could do, not because AI cannot do it, but because humans need to learn by doing it. Joanna Popper is right that AI lets creators move faster and cheaper. The question is whether faster and cheaper is the only variable that matters when you are building a workforce, not just shipping a sprint.
The white-collar job is not going away. The path into it might be.
