AI-exposed occupations show fewer young workers, not a clear fall in total jobs

A job count can stay roughly steady while the people holding those jobs change. That distinction is central to a November 2026 preprint by Jacob Dominski, Tyler Boes and Yong Suk Lee, examining US labor-market outcomes as AI progressed from chatbots to reasoning models and agents.

Their main finding is not evidence of widespread job disappearance. Instead, occupations with larger increases in estimated AI exposure had fewer young workers and fewer part-time workers relative to comparable occupations. These are statistical associations, not a demonstration that employers replaced particular workers with AI.

Measuring potential, not actual replacement

The researchers began with 19,281 occupational tasks from O*NET, a database describing work. One language model divided these into about 99,000 subtasks. Three models, from Anthropic, OpenAI and Google, then independently assessed the same subtasks against four descriptions of AI capability: pre-generative machine learning, chatbots, reasoning models and agents capable of executing workflows with tools.

The resulting exposure score combines three judgments: how much time AI could save, whether it could complete work independently, and whether people retain an advantage through physical presence, relationships, authority or expert judgment. The first two components incorporate estimated economic, regulatory and organizational constraints.

This is a model-generated index, not a measurement of tasks actually automated. Its construction assumes capabilities accumulate across stages. The models broadly agreed on occupational rankings but disagreed substantially about exposure levels. Changes in the index nevertheless correlated strongly with workers’ reported AI use in a separate survey.

The authors linked these scores to the Current Population Survey, focusing on January 2023 through June 2026 across 451 harmonized occupations. Their headline comparison asks how outcomes changed between the first half of 2023 and the first half of 2026, controlling for occupations’ initial workforce characteristics. Additional models compare changes within occupations over time while removing economy-wide monthly or quarterly movements.

Those individual-level regressions control for observed worker characteristics; they do not follow the same people through time.

A different workforce beneath the total

In the headline comparison, greater exposure growth was associated with lower total employment and higher unemployment, but neither estimate was statistically distinguishable from zero. That means the analysis found no clear relative change—not that it proved the absence of any effect.

The age pattern was sharper. Each additional standard deviation of exposure growth was associated with 8.2 log points lower employment among workers aged 18–29, approximately an 8 percent relative difference. A standard deviation here is about 12 points on the exposure scale, roughly the difference between average office-support and computer occupations.

The decline was concentrated among young workers without a bachelor’s degree, whose estimate was 9.8 log points. The estimate for young graduates was also negative but too uncertain to distinguish from zero—or conclusively from the non-graduate estimate.

Office and administrative support contributed heavily to the young-worker association, especially customer service. Computer and mathematical occupations, despite their larger estimated capability gains, contributed relatively little.

College graduates aged 30–49, meanwhile, gained full-time employment in the headline comparison. The full-time workforce in more exposed occupations became older and more educated.

Fewer entrants is plausible, not proven

Young-worker unemployment did not rise in the headline comparison. That is consistent with reduced entry rather than increased layoffs, because unemployed respondents are assigned the occupation of their last job. Someone never hired into an occupation will not appear among its unemployed workers.

But the survey cannot separate the relevant flows. Young people moving to other occupations or leaving the labor force could also produce this pattern.

Part-time employment fell by about 7 log points per standard deviation of exposure growth, while the part-time share declined by 0.74 percentage points. Average weekly main-job hours rose by about 12 minutes, potentially reflecting fewer part-time workers rather than longer schedules for everyone. Wage findings were much less consistent across specifications.

Why the causal question remains open

Remote work complicates the interpretation. Adding occupations’ initial telework shares and subsequent changes reduced the young-worker estimate from 8.2 to 4.7 log points, significant only at the weaker 10 percent threshold. An alternative remote-hours adjustment made it statistically insignificant.

Historical comparison tests generally supported the headline employment findings, but pandemic-recovery windows exposed misleading unemployment associations. Other forces, including post-pandemic hiring adjustments, remain possible explanations. The study also assumes early-2023 outcomes had not yet substantially responded to chatbots.

Finally, exposure is not randomly assigned, occupation-level survey samples are small, and the authors’ agent era covers only two quarters. Their relative comparisons cannot detect AI effects shared by all occupations. This preprint offers evidence of workforce reshuffling, not a settled causal verdict on AI and jobs.