A new Stanford Digital Economy Lab update says AI has not triggered a broad employment breakdown across the economy, but it is making the path into work harder for younger entrants.
The latest figures show that workers aged 22 to 25 in occupations with high AI exposure have 19% fewer employment opportunities than same-age peers in low-exposure occupations. A year earlier, that gap stood at 15%.
The paper, titled Canaries in the Coal Mine? Six Facts About Recent AI Employment Effects, was updated in August 2026 by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen. It uses U.S. ADP payroll data spanning November 2022 to June 2026 and tracks actual hiring records for workers aged 22 to 70. Its headline conclusion is straightforward: overall employment has not clearly worsened, but barriers to entry are rising.
Hiring slowed more than firing rose
The researchers said the widening gap is not mainly the result of layoffs. It comes from slower hiring.
Since 2022, employment for young workers in high-exposure occupations has fallen by about 11%, while employment for same-age workers in low-exposure occupations has grown by 10%. The two lines have continued to drift apart.
The underlying data showed little difference in quitting rates or voluntary job switching between young workers in high-exposure and low-exposure occupations. What changed was employers’ willingness to open new positions.
In practical terms, existing workers are largely not being replaced. Fewer openings are being left for recent graduates.
The impact shows up in hiring, not broad wage erosion
The paper defines AI exposure as the share of tasks in a job that current AI models can perform or speed up. By that measure, the gap is reflected almost entirely in hiring counts. Wage levels themselves did not show a system-wide decline, suggesting that more experienced workers already in those jobs remain relatively stable.
Automation-heavy occupations look weaker for young workers
The study also matches a framework used by Anthropic in its Economic Index, which separates AI use into two categories: automation and augmentation. Automation refers to AI replacing human tasks. Augmentation refers to AI helping workers do their jobs faster or better while remaining necessary.
According to the article, accountants, auditors, and reception or advisory roles fall more heavily into the automation category, while chief executives and registered nurses lean more toward augmentation. Stanford’s data points in the same direction: the higher the automation share, the worse the employment picture for young workers; the higher the augmentation share, the more likely employment is to stay flat or keep growing.
Codified knowledge and tacit knowledge split the outcomes
The researchers also looked at the type of knowledge a job depends on. They distinguish between codified knowledge and tacit knowledge. Codified knowledge can be written into textbooks and taught through formal education. Tacit knowledge is built through practice, mentoring, and repeated trial and error.
Using formal education requirements as a proxy for dependence on codified knowledge, the team found that occupations with a higher share of codified knowledge showed slower employment growth for new entrants. Occupations with more tacit knowledge showed faster growth for mid-career and senior workers.
Not an economy-wide shock
The paper does not present these findings as a forecast of mass unemployment across the economy. When the team compared occupations with the highest and lowest AI exposure, overall employment growth was almost the same. The damage was concentrated at the intersection of specific age groups and specific job categories rather than spread across the full labor market.
Education also appeared to cushion the effect. In occupations with a higher share of college graduates, the gap between high-exposure and low-exposure jobs narrowed. In occupations with a lower graduate share, the divide was sharper: low-exposure jobs kept growing, while high-exposure jobs continued to shrink.
An early warning, with limits in the data
The researchers noted that the ADP database covers formal employment records only. Gig work and other nontraditional forms of employment are outside the dataset, so the true gap could be either understated or overstated.
Stanford Digital Economy Lab described the study on its announcement page as an early warning signal. In comments to The Washington Post, Brynjolfsson said, “The entry effect we are measuring is real, persistent, and growing.” He added that he is more concerned than before about a labor market in which the top-line employment numbers stay steady while younger workers are quietly kept out.