Workers aged 22 to 25 in the occupations most exposed to generative AI have lost 16 percent of their employment since ChatGPT became widely available. Older workers in the exact same occupations have not. The damage is age-specific, occupation-specific, and currently invisible in headline economic data.
The aggregate unemployment rate looks fine. The labour market is not collapsing. What is collapsing, quietly and specifically, is the bottom rung of the ladder, and the standard reassurances about technology creating more jobs than it destroys happen to be technically correct and almost entirely beside the point.
That is the looming crisis. Not robots in the streets. A generation that cannot get hired into the jobs that historically taught them how to think.
The numbers that the unemployment rate is hiding
Look at the top-line figures and nothing seems unusual. Look at recent graduates specifically and the picture changes. The unemployment rate for recent college graduates climbed to 5.6 percent in the fourth quarter of 2025, per the Federal Reserve Bank of New York, while the underemployment rate hit 42.5 percent, the highest reading since the pandemic.
Underemployment is the more revealing number. It captures the philosophy graduate working retail, the computer science graduate driving for a rideshare app, the marketing graduate doing admin temp work. These are people who did everything the system told them to do. The system has stopped reciprocating.
Researchers went looking for the canary in the coal mine. They controlled for firm-level shocks, excluded technology companies, excluded remote-friendly occupations, and the pattern still held. Early research suggests the adjustment in AI-exposed fields may be happening through employment rather than wages, and may be concentrated in occupations where AI automates rather than augments human work.
An Anthropic labour market study from March 2026 reached a similar conclusion from a different direction, looking at usage patterns of its own models. Software development, customer service, programming. The entry-level versions of these jobs are thinning out.
The training-ground problem
Here is what makes this different from previous waves of automation. Entry-level work is not just a paycheque for a young person. It is the mechanism by which an economy reproduces its own skilled workforce.
The junior associate who spends two years reading contracts learns, slowly, what a bad contract looks like. The junior developer fixing tickets learns which kinds of bugs cluster around which kinds of architectural choices. The new analyst building decks learns what a senior partner actually wants to see on slide three. None of this is taught in school. It is absorbed through repetition of tasks that often look, from the outside, like grunt work.
AI is now very good at the grunt work. That is the part the optimistic framing never quite explains. The suggestion that AI frees humans to do higher-value work assumes the higher-value work can be done by people who skipped the lower-value work. It cannot, or at least not reliably. The higher-value work is built on top of years of doing the lower-value work. Remove the bottom of the staircase and the top floors do not stay suspended in mid-air forever. They just become inaccessible to anyone who is not already there. The 22-year-old shut out of a junior role in 2026 does not simply lose a year of income. They lose the only sequence of repetitions that would have made them a competent 28-year-old, and the firm that did not hire them loses the only candidate pool it would have had for the senior roles of 2034. The damage compounds quietly, on a delay, and shows up in the data as a thin layer of mid-career professionals a decade after the original decision was made.
MIT Technology Review put the stakes plainly in an essay published last week: entry-level hiring is not just an expense, it is an investment in the future stock of judgement inside the firm. Firms making the cost-cutting decision in 2026 are also making a workforce-quality decision in 2034, whether they realise it or not.
Why the standard advice to learn coding stopped working
For most of the last fifteen years, the standard policy answer to technological displacement was a coding bootcamp. The federal government funded them. State governments funded them. Universities pivoted curricula toward them. The implicit promise was that software was eating the world and the people who wrote the software would be safe.
That promise has aged badly. The coding tasks that bootcamps were optimised to teach, building a CRUD app, wiring up a REST API, writing test cases for a known specification, are exactly the tasks that current language models perform competently. The graduate who can do nothing except those things is competing directly with subscription-based AI tools.
Coding is not useless. Junior coding jobs are scarcer. Those are different statements, and conflating them has produced a lot of bad career advice over the last three years.
The career advice that holds up better is unglamorous. Domain expertise plus AI fluency. A mechanical engineer who understands manufacturing and can also direct a language model. A programmer who understands financial services regulation and can also direct a language model. The combination is what is actually scarce, because the model on its own does not know which questions matter inside any specific industry, and the industry expert on their own cannot operate the model at the speed the market now demands.
The competition that actually matters
Most coverage of AI and jobs frames the contest as human versus machine. That framing flatters the technology and confuses the labour market.
The real competition is colleague versus AI-augmented colleague. The graduate who can produce in a day what used to take a team a week is not displacing a robot. They are displacing the four other graduates who have not learned to use the tool. This is closer to how the spreadsheet remade accounting in the 1980s than to how the assembly line remade manufacturing in the 1920s. Productivity per worker rose. The number of workers needed per task fell. The accountants who survived were the ones who treated VisiCalc as a power tool rather than a threat.
The difference now is speed. Job openings data from the Bureau of Labor Statistics shows hiring slowing in white-collar categories that historically expanded reliably across business cycles. Companies are not announcing waves of AI-driven layoffs. They are simply not backfilling the junior roles that turn over naturally, and they are not creating the next layer of junior roles that growth would normally have produced.
The effect is statistically quiet and individually brutal. A graduate sends hundreds of applications before getting a single offer. That is not a recession story. Recessions show up in the aggregates. This is a structural reshaping of one specific stage of one specific kind of career, happening at a speed faster than universities, hiring managers or policymakers have figured out how to respond to.
What universities are still getting wrong
The university response so far has mostly been to add a course. Courses with titles like AI for Business. Seminars with titles like Generative AI in the Humanities. A coding requirement bolted onto a non-technical major.
This treats AI as a topic. It is closer to a medium. Literacy in AI is not something you learn for three credits in junior year and check off. It is something that has to be embedded in how every subject is taught, from the first semester onward, because every subject is now being practised professionally with these tools.
A law school that teaches contract drafting without teaching students how language models draft contracts is preparing students to lose to classmates who learned both. A journalism school that teaches reporting without teaching how to use models for research, transcription and source verification is doing the same. Domain expertise still matters, arguably matters more, but it now has to ship with fluency in the tools that the field actually uses.
The harder institutional shift is admitting that the bachelor’s degree alone no longer signals readiness for entry-level work the way it did even five years ago. Employers used to outsource the final layer of training to the first job. The first job is now doing less training, because the tasks that constituted the training are being absorbed by software.
The policy gap nobody wants to name
Governments have spent the past three years funding AI compute, AI safety institutes, AI sovereignty efforts and AI national strategies. The amount spent on incentivising employers to actually hire and train early-career workers in this environment is, by comparison, negligible.
This is upside down. The compute will get built either way, because the economics are too compelling. What does not happen automatically is the maintenance of an apprenticeship system in an economy where entry-level tasks are vanishing. That requires deliberate policy: wage subsidies for genuine training positions, tax treatment that recognises junior hires as capital investment rather than pure operating expense, accreditation regimes that distinguish actual workforce development from credential mills.
The countries that figure this out first will compound the advantage for a generation. The ones that do not will discover, around 2032 or 2034, that they have an unusually thin layer of mid-career professionals, the cohort that should have been hired and trained in 2025 and 2026 but never quite was. By then it will be too late to fix retroactively.
The longer historical view
This is not the first time a technology has rewired the bottom of the labour market. Silicon Canals has previously looked at Keynes’s 1930 prediction that his grandchildren would be working fifteen-hour weeks by now. Technologically, he was roughly right about productivity. He was wrong about what societies would do with the surplus, which turned out to be: redirect it into consumption, status competition and longer working hours for the people who still had jobs.
The current moment rhymes. The productivity gains are real. The distribution of those gains is, again, going to be decided by institutions rather than by the technology itself. If entry-level hiring continues to be treated as a discretionary cost line that AI conveniently eliminates, the productivity gains will accrue to capital and to the already-credentialed. If it is treated as the load-bearing structure of the long-run workforce, the gains can be spread more broadly.
Neither outcome is technologically determined. Both are policy choices being made right now, mostly by default.
What young workers can actually do
Practical advice, given the structural problem, is harder than it sounds. “Learn AI” is necessary and insufficient. “Get domain expertise” is necessary and insufficient. The combination is the thing, and the combination is hard to acquire when the entry-level jobs that historically taught the domain piece are precisely the ones being thinned out.
Three strategies are visible in the graduates who are landing roles. The first is to build a specific, narrow application of AI to a real problem in a real field, not vague claims about ChatGPT proficiency but a portfolio of concrete artefacts. A tool that automates a step in legal discovery. A model fine-tuned on a particular regulatory regime. Something that demonstrates the candidate has already done a junior version of the job, before being hired into it. This substitutes for the apprenticeship the firm is no longer offering.
The second is to chase smaller employers earlier. Large companies are the ones most aggressively shrinking junior headcount, because they have the systems to absorb AI capacity at scale. Mid-sized firms and regional employers, the kind that do not make headlines, still hire juniors, still train them, and increasingly cannot compete with large-firm salaries anyway. The graduate who takes the less prestigious job in 2026 and emerges in 2030 with four years of actual experience will out-compete the graduate who spent those years sending applications to the firms everyone has heard of.
The third is to treat AI fluency as a baseline, not a differentiator, and to be honest about the fact that the model is faster than they are at the mechanical work. The value a junior offers now is judgement about which questions to ask the model, which outputs to trust, and which parts of the answer to bring to a human decision-maker. That is a senior skill being demanded of juniors. It is unfair. It is also the actual job.
What firms should stop pretending
The cost-cutting framing inside companies is seductive. Replace the junior analyst with a model. Save the salary. Ship the same deliverables.
It works for one quarter. It works for one year. Five years in, the firm has no mid-level analysts, because there were no juniors to promote. Ten years in, the firm has no senior analysts, because there were no mid-levels to promote. The talent pipeline is not a metaphor. It is a literal sequence of jobs, and if any link in the sequence is removed, everything downstream eventually breaks.
The firms that will be in the strongest position by 2030 are the ones currently treating junior hiring as a strategic investment, even when the spreadsheet says it is not the most efficient use of capital this quarter. The ones treating it as a line item to optimise away are quietly hollowing themselves out.
This is the part that is hardest to communicate to a CFO. The savings show up immediately and the costs show up in five years, by which point the person who made the decision has typically been promoted, transferred or hired elsewhere. The incentive structure produces the wrong answer reliably.
So here is the question worth sitting with. What happens to a profession once its firms discover they can function, quarter to quarter, without hiring anyone under 26? And once that discovery is made at scale, once the junior tier has been gone long enough that nobody inside the building remembers what it was for, can the ladder actually be rebuilt? Or do the missing rungs harden into a permanent shape, with the 22-to-25-year-olds whose employment quietly dropped 16 percent since ChatGPT becoming, a decade on, the thin cohort that the firms above and below them learned to live without?
The headline unemployment rate will keep looking fine. It always does. The number that matters is already telling a different story, and the only honest question is whether anyone with the authority to act on it will, before the default answer becomes the permanent one.
