Entry-level jobs in AI-exposed roles fell 16% since 2022

By Claire Bennett · 2026-08-04

What the payroll records show

This research is not a survey, not a model, not a projection. Stanford Digital Economy Lab researchers fed individual-level monthly payroll records from ADP covering more than 3.5 million workers into their analysis, and the numbers that came back were blunt.^[1] Employment for workers aged 22 to 25 in the most AI-exposed occupations fell significantly relative to their peers since late 2022, after controlling for firm-level shocks, according to the Stanford research.^[1] The granularity of the data makes it hard to dismiss.

Monthly individual-level payroll data matters because it can isolate specific workers within specific firms across time. Aggregate employment statistics flatten variation across industries, ages and occupations into a single figure; this approach does not. When researchers applied firm-time fixed effects, a method that strips out the influence of any single company hiring or shedding workers en masse, the relative employment decline for 22 to 25-year-olds in the highest AI-exposure quintiles still stood.^[1] No other age group showed a statistically significant effect.

The age split in numbers

Between late 2022 and September 2025, employment for workers aged 22 to 25 in the highest AI-exposed quintile declined in absolute terms.^[1] Across the same period and the same occupational bucket, workers aged 35 to 49 grew employment by more than 8 per cent, a sharp divergence between two cohorts sitting inside the same job categories at the same firms.

Erik Brynjolfsson, the Jerry Yang and Akiko Yamazaki Professor at Stanford's Digital Economy Lab, said the findings were consistent with AI having a disproportionate impact on entry-level workers. Brynjolfsson said the research provides early, large-scale evidence consistent with the hypothesis that the AI revolution is beginning to have a significant and disproportionate impact on entry-level workers in the American labor market.^[3] Entry-level roles in software engineering, customer service and document-heavy professional services are disproportionately where generative AI has been deployed since late 2022, and they are also disproportionately where 22 to 25-year-olds tend to start.

Substitution versus augmentation

The study's most practically useful finding is the distinction between substitution and augmentation. Not all AI-exposed occupations behaved the same way.^[1] Roles where AI automates the core tasks, such as drafting, coding, data entry and basic analysis, shed junior headcount, while roles where AI assists rather than replaces showed no comparable employment decline.

Employment declines for young workers were concentrated in occupations where AI automates tasks, while occupations with high augmentation exposure showed no similar employment decline, according to the Stanford research.^[1] Job classification now requires a second dimension: it is not enough to ask whether a role is AI-exposed. You need to ask whether the AI is there to replace the worker or to make the worker faster. For young workers entering the labour market, the answer to that question may determine whether there is a role to enter at all.

The timing of the divergence is not incidental. The sharpest acceleration in employment declines for young workers coincided with the broad release of ChatGPT in late 2022 and the subsequent wave of enterprise AI tool adoption through 2023 and 2024, which is consistent with the substitution hypothesis, though the researchers are careful not to claim it as proof of causation alone.

Wages held while headcount fell

One finding that runs against intuition is what did not change: pay. Compensation trends showed little divergence by age or AI-exposure level across the study period, indicating the adjustment occurred through employment rather than wages, according to the Stanford research.^[1] Workers in AI-exposed occupations were not paid less; firms simply hired fewer of them.

That distinction matters for how policymakers and employers read the headline numbers. A wage divergence would signal that AI is degrading the value of certain skills; a headcount divergence signals something different, that firms are performing the same work with fewer junior bodies. For a 23-year-old trying to get a first job in software or professional services, the output staying stable is cold comfort.

Where the disagreement between studies lies

Not all research reaches the same conclusion. A separate National Bureau of Economic Research working paper, published in February 2025 and using aggregated occupation-level data, found that the net employment effects of AI adoption have been modest.^[2] The difference is not that the researchers disagree on the facts of individual firms; it is that aggregated data can hide what is happening inside occupational categories.

Menaka Hampole, an author on the NBER paper, said the offsetting mechanism works as follows: despite strong substitution at the task level, overall employment effects are modest, as reduced demand in exposed occupations is offset by productivity-driven increases in labor demand at AI-adopting firms.^[2] When AI makes a firm more productive, that firm sometimes expands and hires more people overall, just not necessarily the entry-level workers it used to. Aggregate figures look benign because total employment at AI-adopting firms is not collapsing, but individual-level data reveals that the distribution of who gets those jobs has shifted sharply toward more experienced workers.

The Stanford approach, using monthly individual-level ADP payroll records covering more than 3.5 million workers, can observe that distributional shift directly.^[1] Aggregate studies cannot. If a government or employer reads only the aggregate literature, they will see modest net effects and may not register that the workers absorbing those effects are clustered in a single narrow age band trying to enter the workforce for the first time.

What the findings mean in practice

The Stanford researchers titled their paper "Canaries in the Coal Mine." Young workers in AI-exposed roles are not simply adjusting to a slower hiring cycle; the evidence from nearly three years of individual-level payroll data suggests something more structural, that firms deploying AI for task automation are rebuilding their staffing pyramids from the top down, keeping experienced workers and skipping the junior intake.^[1]

For anyone managing graduate hiring programmes or workforce planning in technology, professional services or any sector with high AI-tool penetration, the implication is direct. The roles that used to absorb the first two or three years of a worker's career, the roles that built the pipeline of experienced staff, are shrinking. The Stanford data covers the period through September 2025.

Claire Bennett

Claire Bennett writes about work and workplace culture. She is interested in the gap between how organisations describe themselves and what it feels like to work inside them.