New computer science graduates face 7.0% unemployment, among the worst rates of any college major. Computer engineering grads sit at 7.8%. The story behind those numbers is not that AI took the senior engineer's role. It is that AI dissolved the staircase the senior engineer once climbed.
The headline most labor coverage chases is "AI replacing workers." The more accurate, less photogenic phenomenon is something else: AI is removing the entry-level slot that converts an unproven graduate into an experienced professional. The job remains. The on-ramp does not.
What the data actually shows
Entry-level job postings have fallen roughly 35% since January 2023, even as overall job openings remained relatively healthy through 2025. Employment among workers aged 22 to 25 in AI-exposed roles is down 6% from late 2022. Young software developers have seen employment drop by nearly 20% in the same window.
Strip the abstraction off these numbers and a pattern emerges. The roles disappearing first are exactly the ones organizations historically used to vet and grow new talent: junior analyst, paralegal, tier-one support, associate copywriter, first-year associate. These roles were never primarily about output. They were apprenticeships dressed up as paid work.
A junior associate's first deposition memo was not worth what the firm paid for it. The point was that, by year three, that same person could draft for a partner unsupervised. Take away the year-one slot and you have severed the pipeline producing year-three talent — and bet, implicitly, that AI will keep producing year-three output forever.
Why this is qualitatively worse than past polarization
Previous waves of automation hollowed out specific occupations. A factory worker displaced in 1985 had a discernible, if difficult, retraining path: by 2005, that displaced worker, or more often their children, could in principle become a data analyst.
The displaced data analyst of 2026 looks at a labor market in which "data analyst" itself is the role being eliminated, and the hiring funnel into anything cognitive starts with a sentence that did not exist five years ago: "We don't hire juniors anymore. The model handles that work." There is no symmetric retraining path because the entry-level rung of the next career is being optimized away in parallel with the current one.
This is the structural feature past polarization waves did not have. Earlier automation eliminated mature jobs but left intact the cognitive on-ramps — university, internship, junior role — that fed new ones. Generative AI's signature is to attack those on-ramps directly, because that is where the savings are easiest to capture. A senior engineer is hard to replace; a junior who does the senior's prep work is not.
So-so automation, applied to the staircase
Daron Acemoglu's framework explains why this looks like recent labor history rather than a clean break from it. Acemoglu's research finds that automation accounts for 50 to 70 percent of the rise in U.S. wage inequality between 1980 and 2016. The mechanism he identifies is what he calls "so-so automation" — technology just good enough to replace a worker but not productive enough to grow the pie.
Generative AI is, for many tasks, exactly this. It can generate the first draft, summarize the deposition, triage the support ticket, build the analyst's deck. It cannot yet make a senior partner ten times more productive. The result is a clean cost line on the CFO's deck — fewer juniors hired — and no offsetting productivity surge from those who remain.
Capital's share of those savings rises. Labor's share falls. The graduates who never get hired in the first place do not show up in headline unemployment cleanly, because they are still officially in the labor force "looking." But the economic cost of a class of workers who never get to log their first year of professional experience compounds for forty years afterward.
What stops being possible
A society stops being mobile not by a dramatic event but by the slow, statistically invisible removal of rungs. The probability that a child born to the bottom quintile reaches the top quintile by age forty has been falling for decades. The on-ramp collapse is the mechanism by which it falls again.
The forward question is not whether AI capability keeps improving. It is whether anyone in a position to hire feels political pressure to keep training the next cohort, when the model is cheaper, faster, and never asks for a promotion. If no one does, the ladder is gone, and the policy fight that follows is the one that defines the next labor era.