Young-worker employment is 19% behind in high-exposure jobs

The Stanford Digital Economy Lab updated its analysis of ADP payroll records on August 12 and found a widening gap for workers ages 22 to 25. By June 2026, employment in occupations with high AI exposure stood 19 percent below the level it would have reached if it had kept pace with less-exposed occupations.¹ ²

That figure is a relative shortfall built from two observed changes. Between November 2022 and June 2026, employment for young workers in the two most-exposed occupation quintiles fell about 11 percent. It grew roughly 10 percent across the three less-exposed quintiles. Starting both at 100, the first group ended near 89 while matching the second would have taken it to about 110. The 21-point difference between 89 and 110 is about 19 percent of the 110 reference level, up from 15 percent in the version using data through July 2025.¹ ²

Authors Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen describe the result as a pattern rather than a causal estimate. Its timing and concentration are consistent with generative AI affecting employment, but the payroll data cannot identify how much of the divergence AI produced. Education, the composition of the company sample, and trends that predate ChatGPT limit that interpretation.²

The divergence formed at the entry point

Employment can contract because more workers leave or because fewer are hired. The paper's decomposition points to the second route. Hiring and separation rates have both declined across groups since 2022, part of a low-hire, low-fire U.S. labor market. Among young workers, however, hiring rates diverged according to occupational AI exposure.²

Separation rates declined in both more- and less-exposed occupations. For 22- to 25-year-olds, they fell at least as much in the most-exposed group. The researchers found no increase in exits that could account for the gap. Fewer people entered exposed occupations, accumulating into lower employment for young workers.²

Our report on Big Tech layoffs follows announced cuts at large companies. The ADP panel crosses industries and measures payroll relationships, locating the change mainly in hiring. It does not count positions that employers say they automated.

That distinction also changes what the slowdown looks like to someone seeking a first role. A layoff leaves an event and a displaced employee in the record. A hire that never occurs appears only as weaker growth. The study does not observe a vacancy that a company decided against opening or the tool involved in that decision; it compares the changing stock of payroll relationships.²

AI exposure is assigned to occupations

The main classification combines standardized job titles with a measure estimating how susceptible each occupation's tasks are to large language models. Researchers split occupations into five exposure quintiles and track employment by age within them. Software, customer service, and accounting appear among the large occupations in the highest quintile, though the result extends beyond technology companies.²

The 2026 revision adds a division between uses oriented toward automation and uses that complement workers. It draws on the Anthropic Economic Index, which classifies Claude queries associated with occupational tasks. Occupations with the largest share of automation-classified usage concentrated the employment declines among young workers. Where queries indicated complementary usage, employment was flat or rising, particularly for more experienced people.²

This measure reflects conversations associated with tasks, not Claude adoption inside the employers in ADP's sample. Payroll outcomes and usage type meet at the occupation level. The analysis therefore supports an association: jobs whose tasks appear more often in automating uses also had the weakest trajectory for young workers. It does not identify a system replacing a task at a particular company.²

The same evidence offers one way to interpret the age difference. The paper finds declining young-worker employment in occupations that depend on codified knowledge, the formal material recorded in texts and procedures. Jobs requiring tacit knowledge gained through practice and repeated situations fared better for experienced workers. The codified-knowledge relationship weakens after controlling for education, leaving the proposed mechanism intertwined with workforce composition.²

The payroll sample is large and selective

ADP processes payroll for more than 26 million U.S. workers. The research uses a subset: a balanced panelA sample that retains the same set of companies throughout the period being analyzed. of companies observed every month from January 2021 through June 2026, covering 3.5 million to 5 million employees per month. It includes full-time workers under age 70 with positive earnings.²

Keeping the same companies improves comparison over time, while conditioning the sample on businesses that survived and remained with ADP. Manufacturing, wholesale trade, and large firms have more weight than they do across the U.S. economy. Retail, accommodation, and food services have less. The sample also includes a larger share of AI-exposed occupations than the national surveys used as benchmarks.²

Occupational coding introduces another uncertainty. Job titles are missing for about 30 percent of records. Where possible, the team imputes the occupation code using the worker's most recent known title or, failing that, the next known title. The authors report that the raw pattern is largely insensitive to this procedure, but it limits the precision of the match between exposure and employment.²

Education produces the most consequential specification test. In a regression on the panel balanced since 2021, the estimated difference between the highest and lowest exposure quintiles moves from about −19 percent without the control to −6 percent after adding the occupation's college share; the latter estimate is no longer statistically significant. Education could represent an independent labor-market shock or part of the exposure channel, because jobs built on formal knowledge also contain more codified tasks. The authors present the specifications as bounds on a range instead of selecting a causal account.²

Earlier trends also constrain interpretation. Highly exposed occupations experienced a relative boom during the pandemic and began declining before ChatGPT appeared. By November 2022, their relative position had returned roughly to its 2018–2019 baseline; the decline then continued for more than three years below that line. Excluding technology companies, removing computer occupations, or controlling for interest-rate and remote-work exposure preserves the direction of the result without turning correlation into causation.²

A Census Bureau working paper using separate administrative data and a different design also found reduced hiring of workers ages 22 to 24 in highly exposed industries, alongside trends beginning around the pandemic.⁴ Agreement across data sources makes early-career entry worth monitoring, while differences in samples and methods rule out treating any one magnitude as a national count.

The Stanford and ADP dashboard tracks the employment divergence each month; the paper locates the difference in hiring through June 2026.² ³ The data cannot yet show whether that gap will widen, stabilize, or reverse.

Sources

1. No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19% · Stanford Digital Economy Lab · https://digitaleconomy.stanford.edu/news/canariesaug26/ · Aug. 12, 2026
2. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence · Stanford Digital Economy Lab · https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf · Aug. 2026

Show 2 more sourcesHide sources

3. Canaries Dashboard · Stanford Digital Economy Lab / ADP Research · https://digitaleconomy.stanford.edu/project/indicators/canaries-dashboard/ · accessed Aug. 26, 2026
4. You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators · U.S. Census Bureau · https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-27.html · Apr. 2026