AINEWS
A Harvard working paper says junior employment fell roughly 9% in the six quarters after firms adopted generative AI, and a shelved U.S. nuclear plant shows what designing for expertise looks like.
A control-systems engineer once designed a fully digital sequence for a U.S. nuclear plant and deliberately left manual steps inside the automation. The plant was never built, but the design choice stuck: operators kept doing certain steps by hand on purpose, so they would keep their mental model of the system alive. The same design pattern is back, this time hanging over every firm rolling out generative AI: efficiency-by-default is dismantling the expertise pipeline that built the workforce in the first place.
A Harvard University working paper covering roughly 65 million workers at more than 280,000 U.S. firms found junior employment fell about 9 percent within six quarters of a company adopting generative AI, relative to non-adopters, while senior employment kept growing. The paper is not peer-reviewed and the authors are careful about causal language. It is a signal, not a verdict, but it is the first large-employer look at the question, and the firms in the sample are real.
A separate Stanford analysis of ADP payroll records shows the youngest workers in the most AI-exposed occupations losing ground after late 2022, while more experienced colleagues held theirs. The two studies agree on the shape of the pattern: losses concentrate where AI automates the work; where it merely augments, junior employment holds steady or rises. The lesson is not that AI is uniformly bad for early-career engineers. It is that the design choice between automating and augmenting decides who gets to keep learning on the job.
That is where the New York Fed enters the picture, and the story gets harder to tell cleanly. Fed researchers attribute much of the rise in young-graduate unemployment to remote work, not AI: firms are reluctant to hire inexperienced people they cannot train at a distance. The two explanations are not mutually exclusive. Remote work may have thinned the apprenticeship rungs first, and AI may be widening the gap by automating the very tasks those rungs used to teach. Either way, the cohort entering the workforce in 2023 and 2024 is paying for it.
So what does "designing for hands that stay warm" actually look like in practice? In the nuclear-controls parable, the answer was procedural: certain steps were kept inside the sequence on purpose, even when the automation could run them faster, because the cost of operators forgetting how the plant worked was higher than the cost of slower responses. The same arithmetic is available inside a software team. A junior engineer asked to write the test cases an AI assistant drafts in seconds is doing real work that builds judgment. A junior engineer asked only to review what the model produced is not. The first keeps the apprenticeship rung in place; the second welds it shut.
The constructive case is not a return to inefficiency for its own sake. It is a recognition that the cheapest version of a task is not always the cheapest version of the workflow, once you price in the next five years of who-knows-what. The Harvard working paper will be revised. The Stanford numbers will sharpen. The remote-work story will keep being told. What will not change is the design question: when the automation can do the work, what do you leave for the person who needs to learn it? The control engineer who left manual steps inside a sequence that never ran had an answer. The firms rolling out generative AI in 2026 do not, yet.