Two findings about artificial intelligence and employment are often presented as contradictory: AI has not produced broad job losses across the U.S. economy, and young workers in the most AI-exposed occupations are falling behind their peers. The latest evidence from the Stanford Digital Economy Lab supports both, and how they fit together matters more than either headline.
In an August 12 revision of their study “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Erik Brynjolfsson, Bharat Chandar and Ruyu Chen analyzed ADP payroll records covering millions of U.S. workers through June 2026. Their first fact is direct: “We do not see widespread, economy-wide job displacement associated with AI.” Their second is the one that has drawn attention. Employment of workers ages 22 to 25 in highly AI-exposed occupations now stands about 19% below where it would be had it kept pace with similarly aged workers in less-exposed occupations. At the July 2025 data vintage, the same gap was 15%.
What the 19% gap is, and what it is not
The 19% figure is a relative gap, not a 19% fall in young-worker employment. In levels, the paper reports that employment of 22- to 25-year-olds in the two most-exposed fifths of occupations fell about 11% between November 2022 and June 2026, while employment of the same age group in the three least-exposed fifths grew about 10%. The gap is the distance between those two paths.
“Exposure” is also a research measure, not an observation of what employers did. The authors rely mainly on task ratings from Eloundou and co-authors (2024), which estimate how susceptible occupational tasks are to large language models, and on usage data from the Anthropic Economic Index. An exposed occupation is one where the technology could plausibly do a meaningful share of the work, not one where it has necessarily been deployed.
Several further findings shape the interpretation:
- Experienced workers show no comparable gap. Older workers in the same exposed occupations did not diverge from their less-exposed peers.
- Hiring, not firing. The adjustment “appears to operate primarily through reduced hiring of young workers rather than increased separations.”
- Automation versus complement. The declines are concentrated in occupations where AI usage tends to automate tasks. Where AI is used more to complement workers, employment is flat or rising.
- Jobs, not pay. So far, the authors write, adjustment is showing up “primarily in employment rather than base pay.”
Why aggregate stability and entry-level pressure can coexist
If adjustment runs through hiring rather than layoffs, it will be hard to see in headline indicators. A firm that fills fewer junior positions issues no layoff notices, and people already in those jobs keep them. The change shows up as positions not opened, or candidates not hired, among a narrow group of new entrants.
The national data are consistent with that picture but cannot confirm it. The Bureau of Labor Statistics reported that payrolls rose by 162,000 in August and that the unemployment rate held at 4.1%. In July, according to the JOLTS survey, employers made 5.1 million hires while layoffs and discharges totaled 1.7 million. Those totals do not suggest a broad wave of dismissals. But they are not designed to isolate 22- to 25-year-olds in particular occupations; the Stanford authors note that standard federal surveys either do not collect that detail or lack precision for such small groups.
A separate working paper, “How Does AI Change Labor Demand? Evidence from 41 Countries,” published by the lab on September 21 by Chandar and Bouke Klein Teeselink, approaches the question from the employer side. Using Revelio Labs data covering 1.25 billion job postings and 154 million employment records, it compares foreign affiliates of companies that adopted generative AI with matched affiliates that did not. By March 2026, the junior share of employment at adopting affiliates was 1.9 percentage points lower than at controls. That decline came mainly from a 6.7% rise in senior employment; junior employment was 2.5% lower, an estimate the authors say is not statistically significant.
The mechanism differs from the U.S. payroll study: the entry-level share shrinks largely because senior headcount grows, not because junior roles are cut. What the studies share is a signal about workforce composition, especially the bottom rung of the career ladder, rather than total headcount.
What the evidence does not establish
The Stanford authors are explicit that their U.S. findings are “descriptive patterns, not causal estimates of the effect of AI.” They report that the gaps shrink when education is taken into account, that some differential trends were visible before generative AI was widely used, and that the estimated gaps are larger in the ADP sample than in national survey benchmarks. The ADP analysis sample also over-represents larger firms and higher-exposure occupations relative to the national workforce.
The 41-country study has its own limits. Its estimates describe multinational affiliates visible on LinkedIn, which over-represents managers and professionals, and they compare adopters with non-adopters rather than measuring economy-wide employment. The authors show that firm-level and aggregate effects can even differ in sign.
Neither study shows that a particular task has disappeared, that all entry-level roles are shrinking, or what is happening at any single employer or in any single metro area. Neither makes a forecast. The accurate summary is narrower: in current data, broad displacement is not visible, while a relative shortfall in young-worker employment in automation-heavy exposed occupations has widened.
What employers and workers can take from it
For employers, the practical question is whether entry-level hiring in exposed roles is changing inside their own organizations, and why. That calls for tracking junior requisitions, offers and hires by occupation, not just total headcount, and a deliberate look at the work junior staff used to do: if AI now handles tasks that once trained new hires, the training need does not disappear. The 41-country paper notes that a shift away from junior workers “may have implications for the pipeline of experienced workers that firms will draw on in the future.”
For early-career workers, one distinction is worth watching: declines are concentrated where AI tends to automate tasks, while occupations where it complements workers show flat or rising employment. That is a research pattern, not a guarantee for any role.
What to watch next: whether the young-worker gap keeps widening in later data vintages, which the lab tracks through its public AI Economic Indicators, and whether hiring rates in the monthly JOLTS release weaken. A persistent hiring-led gap would matter even if headline unemployment stays stable.
Sources
- Stanford Digital Economy Lab, “No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19%,” 2026-08-12.
- Stanford Digital Economy Lab (Brynjolfsson, Chandar, Chen), “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (revised paper),” 2026-08.
- Stanford Digital Economy Lab (Chandar, Klein Teeselink), “How Does AI Change Labor Demand? Evidence from 41 Countries (publication page),” 2026-09-21.
- Stanford Digital Economy Lab (Chandar, Klein Teeselink), “How Does AI Change Labor Demand? Evidence from 41 Countries (working paper PDF),” 2026-09-20.
- U.S. Bureau of Labor Statistics, “The Employment Situation, August 2026,” 2026-09-04.
- U.S. Bureau of Labor Statistics, “Job Openings and Labor Turnover Summary, July 2026,” 2026-09-01.
Featured photo: Programmers at work. Representative image of highly AI-exposed occupations. Photo: Lbronn, via Wikimedia Commons (CC BY-SA 4.0). Resized from the original.
