On August 12, 2026, the Stanford Digital Economy Lab released a revised working paper examining the relationship between generative AI and youth employment. Analysis of U.S. payroll records found that among 22-25 year-olds, employment in occupations with high AI exposure fell by about 11% since November 2022, while employment in low-exposure occupations rose by about 10%. Compared to a scenario where the high-exposure group grew at the same pace as the low-exposure group, the former lagged by 19%. However, what the study captured is a correlation between an occupation's AI exposure and employment levels—not a causal estimate that AI eliminated 19% of jobs.
This distinction significantly changes the strength of the conclusion. In the data, the gap showed up more in reduced hiring at the entry point than in increased departures among young workers. At the same time, AI-exposed occupations had already shown different employment trends since the COVID-19 period, and the gap narrows when controlling for education levels. Readers need to weigh both the sample size of 3.5 to 5 million people and a research design that cannot cleanly separate causation.
An Observational Study Tracking 3.5 to 5 Million People Monthly
The study by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," is a Stanford Digital Economy Lab working paper. A first version appeared in 2025, and the August 2026 version extends the observation period through June of that year. It has not been published in an academic journal and carries no DOI. At this stage, it is a pre-peer-review study, not the result of an experiment or simulation.
The underlying data comes from anonymized administrative payroll records provided by ADP, a payroll processing company. While ADP provides payroll services to more than 26 million people in the United States, the study primarily used a balanced panel of companies observable every month from January 2021 through June 2026. The monthly sample size ranged from 3.5 to 5 million people. The analysis was restricted to full-time workers under 70 with positive wages. For analysis extending further back to confirm trends before ChatGPT's public release, the researchers also used a smaller panel of companies that had existed continuously since January 2018.
Each individual's occupation was mapped from standardized job titles to the Standard Occupational Classification system. However, about 30% of job titles were missing. The research team filled in codes where they could be inferred from a worker's own prior and subsequent job history, and excluded remaining unmapped records. The team reports that descriptive patterns for young workers largely held up even when the imputation method was varied. Even so, this does not prove the missing data was random.
AI exposure was measured using two distinct yardsticks. One is a theoretical measure of automatability, based on evaluating O*NET job tasks using GPT-4. The other is the Anthropic Economic Index, which maps Claude usage logs to job tasks. The latter distinguishes between "automation" usage, where AI completes work on a person's behalf, and "augmentation" usage, where AI collaborates with a person. Neither measure directly observes whether companies using ADP had actually adopted AI.
The 19% Figure Is a Relative Gap Between an 11% Decline and a 10% Increase
The 19% figure mentioned above does not mean employment in high-exposure occupations fell by that amount. From November 2022 through June 2026, employment among 22-25 year-olds fell by about 11% in the top two quintiles of AI exposure, while it rose by about 10% in the bottom three quintiles. The gap between the two groups was 21 percentage points, and the shortfall—assuming the high-exposure group should have grown at the same pace as the low-exposure group—comes to 19%.
The top two exposure groups accounted for about 57% of employment among 22-25 year-olds at baseline. The decline in this group pulled down overall employment for this age bracket by about 6 percentage points, but growth in the low-exposure group offset most of that. As a result, overall employment for 22-25 year-olds fell by only 1.9%. Across the full ADP sample spanning all ages, employment rose by about 6%, and even the most exposed fifth quintile saw about 4% growth. The study did not observe an economy-wide employment collapse.
The 2026 version also changed its headline metric from the previous year's version. The 13% figure that the 2025 version highlighted was a regression estimate adjusted for company-level shocks. The revised 19% is a descriptive comparison that requires no model selection. When aligned on the same comparative scale over time, the shortfall stood at 15% as of July 2025 and widened to 19% by June 2026. It would be inaccurate to directly frame this as a worsening from 13% to 19%.
The revision also included improvements to the crosswalk table linking occupation codes to external indicators, as well as imputation of missing occupations. While the raw employment patterns did not change qualitatively, the Poisson event-study measuring within-company changes proved sensitive to specification. Under the current minimal filter, the coefficient for 22-25 year-olds was minus 6.8 log points (p=0.03) for the most-exposed fifth quintile and minus 11.3 log points (p=0.003) for the fourth quintile. When the same strict filter used in the prior year's version was applied to the revised data, the fifth-quintile coefficient became minus 5.3 log points (p=0.26), no longer statistically significant at the conventional 5% level. This is why the research team shifted its emphasis toward the simpler descriptive figures.
Reduced Hiring, Not Increased Departures, Drives the Gap—Correlated with Automation
Breaking down the pathways behind the decline in headcount, the study found no evidence that departures increased in high-exposure occupations among young workers. The study defines "hiring" as a new combination of worker and company, and "departure" as the disappearance of a combination that existed the previous month, calculating a 12-month cumulative rate divided by the headcount 12 months prior. Because this departure metric does not distinguish between company-initiated layoffs and voluntary resignations, it cannot determine whether layoffs specifically increased or decreased. Since 2022, both hiring and departure rates have broadly declined across age groups and exposure levels, reflecting a "low-hiring, low-departure" environment. In high-exposure occupations among 22-25 year-olds, departure rates fell by as much or more than in low-exposure occupations. What created the gap was the decline in hiring rates.
Analysis layered with Claude's actual usage classifications is also consistent with this hiring-decline interpretation. When automation, augmentation, and total usage from the March 2025 Anthropic Index are entered into a regression simultaneously, the coefficient for automation and the rate of employment change among 22-25 year-olds was minus 0.098. The standard error was 0.018, with a p-value below 0.01. Pooling indices from September 2025, January 2026, and April 2026 yielded a coefficient of minus 0.084 (standard error 0.014, p<0.01). Meanwhile, the coefficient for augmentation was not statistically significant for this age group.
This does not mean one can say "AI-driven automation reduced hiring." Claude's logs do not represent generative AI usage as a whole, and model-based estimation is also involved in the process of classifying conversations as automation. It remains possible that demand shifts specific to high-exposure occupations were related to both AI usage and hiring simultaneously. As the paper itself states, the findings are only strong enough to say that "substitutive usage is consistent with the decline in youth employment."
The research team also examined a hypothesis that younger workers tend to hold jobs relying more on codified knowledge, while more experienced workers draw on tacit knowledge. In an exploratory analysis, they had GPT-5.4-mini evaluate ADP's 7,514 job titles and descriptions twice each, scoring on a 1-to-10 scale both "codified knowledge" that can be formally taught and "tacit knowledge" acquired through practice and mentorship. Occupations with higher codified knowledge saw slower growth in youth employment, while occupations with higher tacit knowledge saw faster growth among mid-career and senior workers. However, the gradient for codified knowledge loses significance once college graduation rates are controlled for. This is a correlation consistent with the hypothesis, not evidence that demonstrates the mechanism.
The scope for measuring wages is also narrow. Differences in base pay trends by age or AI exposure were smaller than the roughly 20-percentage-point employment gap. However, ADP's base pay figures do not include bonuses, overtime pay, or commissions. Stock compensation and tips are also excluded. One can write that "adjustment is proceeding through employment, and is not conspicuous in base pay," but one cannot conclude there is no effect on total compensation.
National Statistics Show a Smaller Gap, and Causation Remains Undetermined
Three factors most strongly constrain causal interpretation: education, prior trends, and sample composition. In an occupation-level regression for 22-25 year-olds using the panel dating back to 2018, the coefficient comparing the most- and least-exposed groups fell by half, from an uncontrolled minus 0.18 to minus 0.09 once college graduation rates were included. The latter is significant only at the 10% level. Using the panel starting in 2021, the coefficient shrank from minus 0.19 to minus 0.06, losing statistical significance. Education may represent a pathway through codified knowledge that AI can more easily substitute, or it may represent a confounding factor reflecting a separate employment shock. Current data cannot distinguish between the two.
Employment in high- and low-exposure occupations did not move in tandem even before ChatGPT. The relative gap for high-exposure occupations rose by about 13 percentage points from the 2018-2019 baseline through mid-2020, then fell by about 30 points afterward. Of that, roughly 13 points represented a reversal of the pandemic-era spike, while the remaining roughly 17 points represented a decline below the earlier baseline. The fact that the decline continued after late 2022 is hard to explain solely as a temporary rebound, but this alone cannot pin AI as the cause either.
Comparison with the nationally representative American Community Survey (ACS) calls for even more caution in generalizing. For 22-25 year-olds from 2022 to 2024, the gap in employment growth rates between the most- and least-exposed groups was minus 2.2 percentage points in the ACS, with a 95% confidence interval spanning minus 5.5 to plus 1.1 points. Over the same period, the ADP sample showed a gap of minus 13.2 points. The direction is consistent, but the ACS confidence interval includes zero, and the magnitude of the gap is considerably smaller.
The ADP sample skews toward manufacturing and wholesale trade. It also includes many large companies and occupations with high AI exposure, while retail, hospitality and food service, and small businesses are underrepresented. Because the panel is limited to surviving companies, it cannot track cases where work shifted to companies outside ADP. The underlying microdata is not public, and while the occupation-level data behind the figures is available through a public dashboard, the analysis code must be requested from the authors. Furthermore, ADP is both the data provider and a corporate member of the Stanford Digital Economy Lab. The paper discloses that ADP retains the right to review the paper to prevent disclosure of confidential information.
Other studies using different data have reported findings pointing in the same direction. A U.S. Census Bureau working paper (CES 26-27) by Lee C. Tucker compared industry-by-state cells for 22-24 year-olds and reported that, ten quarters after ChatGPT's public release, regression-adjusted employment in the most-exposed quintile fell by 12%. However, both the age bracket and the exposure metric differ from the Stanford study, so it is not a direct replication. This paper is also pre-peer-review, has not undergone the standard review process for official Census publications, and does not represent the Census Bureau's official position.
The Stanford study alone has not settled whether "AI took away young workers' jobs." Related research includes findings pointing in the same direction as well as findings that detect no effect on aggregate employment. To move the assessment forward, further work is needed: tracking whether the hiring-rate gap persists using the public dashboard, breaking down national administrative records more finely by age and occupation, and comparing companies before and after AI adoption. When all three point in the same direction, the 19% early-warning figure will move one step closer to establishing causation.