Gen Z Unemployment Is Double the National Rate. "AI Took the Job" Isn't the Real Story.
US unemployment for Gen Z workers — the cohort born 1997 to 2012 — now sits at 8.3%, roughly double the national rate of 4.2% Fortune. Goldman Sachs estimates AI is displacing about 16,000 US jobs a month, concentrated on exactly this age group Fortune. One in three employers surveyed say they are already replacing entry-level roles with AI, tech and manufacturing most exposed, and 89% of the class of 2026 say they're worried AI will take their entry-level job before they get one — up from 64% just a year earlier Fortune.
The headline writes itself: AI is destroying jobs, and the young are first against the wall. It's not wrong, exactly. But it's a blunt instrument applied to a precise wound, and the precision is the actual story.
The number that doesn't fit the headline
Here's the puzzle. A Danish administrative-data study tracking eleven AI-exposed professions over two years found that chatbots saved workers only about 2.8% of their working hours on average — and produced no statistically significant change in earnings or overall employment. That's it. A small, almost boring macro effect, the kind of result that would normally support the "AI is overhyped" camp, not the "AI is destroying jobs" one.
And yet, inside those same eleven professions, employment for 22-to-25-year-olds in the most AI-exposed roles fell by roughly 13% relative to their less-exposed peers — while experienced workers in the identical professions kept their jobs at normal rates. Same technology. Same occupation. Opposite outcome, depending entirely on where you stood on the ladder when the tool arrived.
That mismatch — a macro effect too small to move national statistics, sitting on top of a micro effect brutal enough to cut youth employment by double digits — is worth sitting with before reaching for a broader theory. It rules out the simplest story ("AI does the work now, so fewer workers are needed") because if that were true, senior employees in AI-exposed roles should have thinned out too. They didn't. Something more specific than "AI replaces labor" is happening, and it's happening specifically at the entry rung.
What actually gets automated
A new paper by Anton Pankratov, "The Human as a Generator of Distinctions", part of the Observer-Dependent Theory of Everything (ODTOE) project, offers a reframe that fits this asymmetry more tightly than "AI takes jobs" does. ODTOE proposes that what automation actually consumes isn't skill in general — it's the crystallized product of past distinctions: a configuration that was genuinely hard to work out the first time, and has since hardened into something repeatable, executable, low in ambiguity. Learning to format a legal brief, triage a support ticket by category, write boilerplate code against a known pattern, summarize a standard document — these were once acts of judgment. By the time a new graduate is taught them, they've already crystallized into procedure. That is precisely the kind of task a large language model is good at compressing.
Entry-level jobs are built disproportionately out of exactly this material. That's not an accident of history; it's close to the definition of "entry-level" — the work new hires are handed is, almost by design, the work that has already been figured out by someone else and reduced to a teachable procedure. Seniority, by contrast, is built on the opposite activity: continually generating new distinctions in response to problems that haven't shown up in a training set yet — judgment calls under ambiguity, the version of the client relationship or the technical decision that doesn't reduce to precedent. In ODTOE's terms, this is the human functioning as a self-observing observer, an operator that can fold back on itself — Ô(Ô), in the paper's notation — and keep generating structure the world hasn't crystallized into a manual yet.
That's the piece that explains the Danish numbers better than a blanket "automation hits everyone" story: the tool is roughly equally capable across the whole profession, which is why the aggregate hours-saved number is small and flat. But the exposure is wildly uneven, because entry-level work is stacked with crystallized configurations and senior work isn't. The paper's claim is not that juniors are less skilled than seniors — it's that what juniors are asked to do, structurally, overlaps far more with what a model can already imitate.
Why "it'll come for seniors eventually" undersells the point
It's tempting to read this as a delay rather than a distinction — AI eats the bottom of the ladder now, works its way up later. That may partly be true, and ODTOE doesn't claim otherwise; the paper is explicit that this is a framework-relative argument, built on top of real labor-market data rather than a new empirical finding of its own. It's worth being honest about that limit before going further: this is a way of reading numbers that already exist, not a prediction verified independently of them.
But the ladder metaphor undersells what's structurally different about the top. A senior engineer's coherence — what the paper formalizes as B(O,C), a multiplicative measure of how well an observer's judgment holds together against a given configuration — depends on qualities that don't crystallize the same way skill does: sustained focus across ambiguous problems, accumulated cross-domain experience (the paper's term is Λ, for the empirical residue of past distinctions actually paying off), and low internal contradiction under pressure. Those don't get "used up" and handed to a model the way a formatting convention does. They get exercised. This is also the structural worry MIT's Andrew McAfee has raised separately: if companies automate away the entry-level roles where junior employees would otherwise build toward that senior capacity, they risk breaking the talent pipeline they'll need in ten years — cutting the rung that trains the very judgment that stays scarce Fortune.
The profession isn't the unit anymore
If the paper's frame holds, the practical implication for a 23-year-old isn't "pick a more AI-proof job title" — Handshake data already shows entry-level postings running 12% below pre-pandemic levels across categories, so the safe-category strategy is getting harder to execute on anyway Fortune. It's that a profession is, in this frame, a temporary configuration with a finite shelf life, not an identity to plant a flag in. The more durable unit is the observer's own trajectory — the accumulated, transferable pattern of how you generate good distinctions, carried across whichever configuration currently holds your paycheck.
That reframes what school is actually for. The ODTOE paper's argument is that education's real product was never mainly the crystallized skills it hands out — those were always going to harden and, eventually, automate. Its more durable product is the capacity behind them: the ability to keep producing new, useful distinctions when the old configuration stops working, what the paper labels Λ, the accumulation that lets an observer's coherence survive a change of scenery. That's a stronger claim than "education is still useful in the AI era" — the paper argues it becomes closer to infrastructure for it, precisely because crystallized skill is now cheap to imitate and the capacity to generate new distinctions is not.
None of this makes the Gen Z numbers less real or less urgent — an 8.3% unemployment rate and a talent pipeline quietly narrowing are not abstractions. What ODTOE offers is not a fix for that, but a sharper description of the mechanism underneath it: not "AI is destroying jobs" in general, but AI consuming crystallized configurations first, disproportionately at the rung where those configurations are concentrated. The full argument, including the paper's own stated limits and its four testable predictions, is at odtoe.org.