The First Step Up Is Getting Harder to Find
Aggregate employment figures in developed economies look stable, but that surface calm is masking something specific and serious: early-career workers in AI-exposed roles are losing ground faster than any other group, and the window to act is already narrowing.
What the Data Actually Shows
A November 2025 working paper from the Stanford Digital Economy Lab tracked workers aged 22 to 25 in occupations with significant AI exposure. After the spread of generative AI, that group experienced a 16% relative decline in employment – even after researchers controlled for other variables that typically influence hiring decisions. Workers in those same roles who had more experience did not see comparable declines. An Anthropic report published in March 2026 arrived at similar conclusions through a different analytical path.
The pattern is not uniformly spread across entry-level work. Employment is holding steady in early-career roles with low AI exposure. The erosion is concentrated specifically in junior positions within fields where generative AI is used extensively – software developers, customer service representatives, computer programmers, and information systems managers among them. That specificity matters. It points toward a deliberate substitution: companies using AI to absorb the tasks that once gave new workers their first foothold.
Those junior tasks were never just busywork. Junior analysts learn which data points can be trusted and which cannot. Young software developers absorb how production systems actually fail under pressure. New marketing hires discover how customers behave beyond the sanitized language of dashboards. Early-career legal and financial staff learn how rules, deadlines, judgment, and human relationships intersect in real conditions. If AI takes on the drafting, triage, coding, summarizing, and administrative groundwork that once trained those workers, companies may see short-term efficiency gains while the broader workforce loses a generation of practically experienced professionals.
At the same time, the broader hiring environment for recent graduates is also deteriorating. The Federal Reserve Bank of New York reported that in the fourth quarter of 2025, the unemployment rate for recent college graduates reached 5.6%. The underemployment rate – the share of graduates working in jobs that typically do not require a college degree – hit 42.5%, its highest level since the Covid pandemic. No single figure proves AI is solely responsible, and post-pandemic hiring overall remains sluggish. But dismissing AI as a contributing factor would ignore what the occupation-specific data is already showing.
The Human Cost Behind the Statistics
Behind these numbers is a specific kind of strain that does not always appear in labor market reports. Recent graduates today frequently submit hundreds of applications before receiving a single offer. Surveys conducted across young worker populations consistently find elevated rates of anxiety, financial precarity, and burnout among those in extended job searches. The effects are not abstract – delayed financial independence, postponed major life decisions, and the corrosive experience of having sustained professional effort go unrewarded have real consequences for individuals and for household formation patterns at scale.
The industries where AI displacement is most visible – software development, customer service, information management – are also ones where entry-level positions historically served as onramps to mid-career advancement. When those onramps close or narrow, the pipeline for experienced workers in those fields does not simply pause. It starts to hollow out from the bottom.
What makes the current moment distinct from earlier periods of technological disruption is the speed at which generative AI has been absorbed into white-collar workflows. Previous automation cycles tended to displace physical or routine clerical tasks over years or decades, giving labor markets and educational institutions time to adjust. Generative AI reached usable capability in programming assistance, document drafting, and customer communication within a compressed window – short enough that hiring managers could redirect workloads before the next cohort of graduates arrived.
The response required is not simple, and it does not belong to any single institution. Educational institutions need to reorient their programs around an AI-augmented workforce rather than treating AI fluency as an optional add-on. Governments have a role in creating incentives for businesses to hire and train early-career workers rather than defaulting to AI tools for entry-level output. Businesses themselves need to weigh the long-term cost of bypassing junior talent pipelines – a workforce with deep AI experience does not materialize without workers who were once beginners learning alongside the technology.
Students entering the workforce also carry part of the responsibility: developing AI fluency is increasingly a baseline expectation, but knowing how to apply AI tools across different professional contexts – how to verify AI-generated analysis, when to override it, how to communicate its limitations to non-technical colleagues – is the layer of skill that will actually distinguish candidates in a compressed job market.
A System Built to Train Workers Is Losing Its First Stage
Entry-level employment was never just an economic transaction. It was a distributed training system – decentralized, imperfect, but functional, because it put young workers inside real professional environments where they absorbed judgment, context, and consequence in ways that no coursework fully replicates. That system is degrading in the exact occupations where AI is deployed most aggressively, and the degradation is showing up in measurable employment data before it shows up anywhere else.
With the underemployment rate for recent college graduates already at a post-pandemic high of 42.5%, and Stanford’s data showing a 16% relative employment drop in AI-exposed roles for workers aged 22 to 25, the question is no longer whether there is a problem. The question is whether a 22-year-old graduating this spring, applying to software or data or customer-facing roles, will find that the job category she trained for has quietly been reassigned to a model that doesn’t need onboarding, doesn’t ask for feedback, and doesn’t need to be paid.
