Everything About AI
NextFin News - The clearest signal that artificial intelligence is reshaping the labour market is not a spike in unemployment. It is a missing rung on the career ladder. Three years after generative AI entered the workplace, the aggregate jobless rate for highly exposed workers has not budged - but employment for junior software developers has fallen nearly 20% from its 2024 peak while employment for developers aged 30 and over kept growing, and hiring of workers aged 22 to 25 in the occupations most exposed to AI has slowed. The disruption is arriving not as a wave of layoffs but as a door quietly closing behind the least experienced.
The question now facing economists and executives is whether this is the first clear evidence of AI's labour-market impact - or merely the first visible edge of a much larger structural shift. The answer matters because it determines whether the right response is patience or preparation.
The Evidence: What the Data Actually Shows
On March 5, 2026, economists at Anthropic published the most comprehensive industry study of AI's actual labour-market effects to date. Maxim Massenkoff and Peter McCrory introduced a new measure they call "observed exposure," which combines theoretical AI capability with real-world usage data rather than relying on hypothetical task lists. Their central finding runs against the catastrophe narrative: "We find no systematic increase in unemployment for highly exposed workers since late 2022, though we find suggestive evidence that hiring of younger workers has slowed in exposed occupations."
That conclusion carries weight because it is not alone. An IMF analysis of Denmark, covering 2023-2024 tool adoption across roughly 25,000 employees and 7,000 firms, found no significant impact on wages or hours worked. The Stanford AI Index 2026, published April 13, 2026, reports that large-scale aggregate job losses have not materialized in the overall employment data. Three methods, three datasets, one convergent conclusion on the aggregate question: so far, AI has not produced mass unemployment.
But the aggregate number is the wrong place to look for the first signal. Stanford's Digital Economy Lab, in a November 2025 study led by Erik Brynjolfsson, documented a 13% relative decline in early-career employment in the most AI-exposed occupations, even after controlling for firm-level factors. The same research found employment for software developers aged 22 to 25 down nearly 20% from its 2024 peak, while developers aged 30 and over continued to gain ground. The ladder is being pulled up at the bottom, not collapsed from the top.
Goldman Sachs, in research published in April 2026, put a monthly figure on the net effect: economist Elsie Peng estimated AI has reduced US payroll growth by about 16,000 positions a month - roughly 25,000 positions eliminated through substitution, partially offset by about 9,000 added through augmentation. Annualized, that is approximately 192,000 net positions, or about 0.1% of the 160-million-person US workforce. The bank's framework combines AI exposure scores with a complementarity index from IMF economists to separate roles AI substitutes for from roles where it augments human work.
The counterintuitive detail is who is exposed. The public story has always been that automation hits low-wage, low-skill work first. The measured data says the opposite. Workers in the top quartile of AI exposure earn 47% more, on average, than workers with zero observed AI coverage, and are 16 percentage points more likely to be female, 11 points more likely to be white, and almost twice as likely to be Asian, according to the Anthropic study. Computer programmers, customer-service representatives, and financial analysts sit at the top of the exposure rankings. AI is landing first on high-skill, high-value desk work - the inverse of the usual automation story.
At the other end of the distribution, about 30% of workers have effectively zero AI coverage because their tasks appeared too infrequently in usage data to meet the minimum threshold - cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers. These are jobs requiring physical presence that no large language model can currently replicate. The theoretical capability frontier is far ahead of actual coverage: in computer and mathematics occupations, 94% of tasks are theoretically automatable but only about a third are actually being done by AI today. AI is diffusing unevenly, and the unevenness is what makes the first signal local rather than economy-wide.
"The track record of past approaches gives reason for humility," Massenkoff and McCrory write, noting that a prominent study of job offshorability once flagged a quarter of US jobs as vulnerable, yet a decade later most of those jobs had maintained healthy employment growth.
Why the Aggregate Data Misses the First Signal
The reason aggregate unemployment stays flat while entry-level hiring deteriorates is not a statistical accident. It is the mechanism itself. Massenkoff and McCrory argue that AI's effects "might be less like COVID and more like the internet or trade with China" - diffuse, gradual, and easily clouded by the business cycle. A pandemic shuts the economy down in weeks; trade competition and technological diffusion work through hiring margins over years.
Firms rarely fire experienced workers the moment a tool improves. They stop posting junior roles. They let attrition do the work. A senior engineer with ten years of context is expensive to replace and hard to automate wholesale; a trainee whose job is to write boilerplate, triage tickets, and summarize meetings is performing tasks that current models handle at near-zero marginal cost. The displacement shows up in job openings and hiring rates long before it shows up in layoffs.
This is why the 22-to-25 signal is more informative than the unemployment rate. The Anthropic authors explicitly test the young-worker channel, matching Brynjolfsson's cohort definition, and find suggestive slowing in exposed occupations. The Stanford work, built on ADP payroll data covering millions of workers monthly, finds the same pattern concentrated in the first rung. Both are consistent with a displacement model that operates through reduced entry-level demand rather than through dismissals.
Cyclical Noise or Structural Change?
The strongest alternative explanation is simple: this is the tech downturn, not AI. Young workers in software and professional services were hit by hiring freezes and layoffs through 2024 and 2025 as interest rates rose and venture funding tightened. A 20% drop in junior developer employment could be the business cycle doing what it always does to the least senior cohort.
That argument deserves weight, but three pieces of evidence point the other way. First, the decline is age-specific within the same occupation: developers aged 30 and over kept gaining employment while the 22-to-25 cohort fell. A cyclical downturn does not usually split a single occupation so cleanly along an experience line. Second, the studies control for firm-level factors - Brynjolfsson's team compares similar workers inside the same firms across different occupations, and Anthropic compares exposed and insulated workers facing the same macro conditions. Third, the timing aligns with the diffusion of generative AI, not just with rate hikes: the slowdown concentrates in the tasks and roles where AI capability has actually been deployed.
The verdict: this is a structural change in how young workers enter the labour market, not a cyclical dip. A cyclical claim would require mean reversion - junior hiring rebounding as the cycle turns. But the mechanism here is durable: the routine tasks that traditionally trained entry-level workers are the same tasks AI now performs cheapest and most reliably. Even in a strong growth environment, a firm that can cover junior output with AI has less reason to hire the junior. The missing rung does not grow back on its own.
That said, the structural call comes with a boundary. The measured effects reflect tools workers actually used through 2024 and early 2026, not the frontier models arriving now. "No aggregate effect yet" is well-supported for the period measured. It is not a forecast that the effect stays zero as capability climbs. The honest reading is that we are early, and that the first clear evidence is the quietest kind: a narrowing of entry, not a widening of exits.
The Second-Order Consequence: A Broken Apprenticeship Model
The first-order effect is fewer junior jobs. The second-order effect is what that does to the pipeline. Junior roles are not just cheap labour; they are the apprenticeship through which senior workers are made. If the first rung disappears, the second rung eventually has no one to step onto it.
Industries that depend on accumulating tacit knowledge - software engineering, law, financial analysis, consulting - train people by giving them low-stakes versions of the real work. AI is best at exactly those low-stakes, routine versions. Remove them, and you remove the practice field. The short-term productivity gain is real; the long-term cost is a thinner bench of experienced workers five to ten years out. Companies are harvesting the productivity dividend today and pushing the training bill into the future.
This is the gap between what the market has priced and what the data implies. Equity markets have rewarded AI adoption as a margin-expansion story - fewer junior hours, same output. The labour-market data suggests the margin expansion is genuine but partially financed by deferred human-capital formation. Whether that trade is wise depends on how much of senior expertise is transferable without the apprenticeship that produced it.
The Strongest Counter-Argument, and What Would Prove It Wrong
The counter-thesis is straightforward: the studies are measuring a narrow, early window, and the absence of aggregate pain is the real story. Actual AI coverage is a fraction of theoretical capability, and about 30% of workers have effectively zero exposure. On this view, the junior-developer decline is a tech-specific adjustment, the Goldman Sachs estimate of 16,000 net positions a month is a regression inference rather than a direct count, and the economy-wide impact remains too small to act on.
There is also the offset the optimists emphasise. The Goldman analysis does not fully capture the hiring surge tied to AI infrastructure - data centres, power systems, construction - nor the incremental labour demand created when AI-driven productivity lowers costs and expands markets. And the company driving the disruption is funding the research into it: on June 11, 2026, Anthropic announced a $350 million commitment split between a $200 million Economic Futures Research Fund and a $150 million "Claude Corps" fellowship placing 1,000 early-career professionals in US nonprofits. The Associated Press reported the announcement; the detail is worth noting because the vendor has incentives on both sides of the question.
So what would change the call? The falsifying signal is specific and observable: if employment of 22-to-25-year-olds in high-exposure occupations re-accelerates to match the 30-and-over cohort for two consecutive quarters while AI usage continues to climb, the "missing rung" thesis fails. A second test: if the Bureau of Labor Statistics revises its 2034 occupational-growth projections upward for high-exposure occupations, the structural-displacement reading weakens. Until one of those prints, the weight of evidence favours a structural shift at the entry level, masked by aggregate stability.
What to Watch Next
Short term (next 6-12 months): monthly hiring rates for 22-to-25-year-olds in software, customer service, and financial analysis; any inflection in the net displacement estimate; whether the AI infrastructure hiring surge begins to offset white-collar substitution.
Medium term (1-3 years): BLS 2034 occupational projections and their revisions; wage growth differentials between exposed and insulated occupations; whether the 30-and-over employment premium in exposed fields persists or narrows as the junior pipeline thins.
Long term (3-10 years): the apprenticeship question - whether industries rebuild training pathways outside the traditional first-rung model, and whether the productivity dividend survives the thinning of the experienced cohort that produced it.
The bottom line: yes, this is likely the first clear evidence of AI's labour-market impact - but it looks nothing like the headline scenario. Mass unemployment was never the most probable first signal. The first signal was always going to be quieter: fewer openings, slower hiring, a generation finding the first door closed. The aggregate data will eventually catch up, or it will not. What is already clear is that the market for experience is holding while the market for potential is contracting - and that is a structural change no business cycle will reverse on its own.
Explore more exclusive insights at nextfin.ai.
