Skip to Content

Stanford Finds a 19% Employment Gap for Young Workers in AI-Exposed Jobs. Exposure Scores Describe Jobs, Not Your Teams.

November 11, 2026 • Levos Marketing • 7 min read
Stanford Finds a 19% Employment Gap for Young Workers in AI-Exposed Jobs. Exposure Scores Describe Jobs, Not Your Teams.

Stanford's Digital Economy Lab reports that employment among workers aged 22 to 25 in highly AI-exposed occupations is about 19% below where it would be had it kept pace with similarly aged workers in less-exposed occupations, using ADP payroll data through June 2026. The authors find no widespread, economy-wide job displacement, and they describe the pattern as descriptive. An AI exposure score is a rating of an occupation, not of an employer, so the finding cannot tell any one company what AI is doing to its own teams.

Two things are true at once. The Stanford gap is a carefully built pattern in early-career employment, and it is a population pattern that says little about what is happening inside your organization.

What did Stanford's revised Canaries paper find?

In August 2026, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen published a revised version of "Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence," extending the ADP payroll data to mid-2026. The lab posted its summary on August 12 2026. We read it on the lab's site; the full paper PDF would not render for us, so every figure below is cited as that summary states it.

The headline gap

Employment among workers aged 22 to 25 in highly AI-exposed occupations stands about 19% below where it would be if it had kept pace with similarly aged workers in less-exposed occupations. By the same measure the shortfall was 15% at the July 2025 data vintage. Experienced workers show no comparable gap.

The lab's footnote gives the levels behind the ratio. Employment of 22- to 25-year-olds in the two most-exposed quintiles fell about 11% between November 2022 and June 2026, while the same age group in the three least-exposed quintiles grew about 10%. Our arithmetic, 0.89 divided by 1.10, is about 0.81, which lands near the 19% headline; the lab does not say that is how the 19% is built.

Measure Figure Period Source
Gap, ages 22 to 25, highly exposed vs less-exposed occupations About 19% As of June 2026 Stanford DEL
Same gap at the earlier data vintage 15% July 2025 vintage Stanford DEL
Change in employment, ages 22 to 25, two most-exposed quintiles About -11% November 2022 to June 2026 Stanford DEL
Change in employment, ages 22 to 25, three least-exposed quintiles About +10% November 2022 to June 2026 Stanford DEL

Where the adjustment shows up

The authors report that the adjustment appears to operate mainly through reduced hiring of young workers, not increased separations, and so far in employment rather than base pay. The declines are concentrated in occupations where observed AI usage tends to automate human tasks. Where AI is used more to complement workers, employment is flat or rising, particularly among experienced workers.

What the revision adds on knowledge type

The revision adds a distinction between two kinds of knowledge. Employment fell for young workers in occupations that rely heavily on codified knowledge, meaning formal, standardized, documented knowledge that can be taught through education, textbooks, or written procedures. Employment rose among experienced workers in occupations that rely more on tacit knowledge, acquired through practice and mentorship.

Does the Stanford study show that AI caused the gap?

The authors say they cannot yet answer that question definitively, and we would not read it as stronger.

The divergence remains when they exclude technology firms and computer occupations, when they control for exposure to interest-rate increases and remote work, when they include firms that enter or leave the sample, and when they use alternative exposure measures. They also give reasons to think AI plays a role: the gap kept widening after interest rates peaked, it sits where AI use is more automating than complementary, it shows a clear age gradient, and Census Bureau administrative data show broadly consistent raw patterns by age and industry.

What the authors flag as limits

The gaps between more- and less-exposed young workers shrink when education is accounted for. Some differential trends are visible before generative AI was widely used. The estimated gaps are larger in the ADP sample than in national survey benchmarks, with the discrepancy concentrated in education, health care, and public administration, and part of it reflecting Census Bureau weight revisions. And once estimates account for overall changes in firm hiring, they stay directionally consistent but become more sensitive to specification choices. The authors write that they do not view this paper, or any single study, as definitive evidence. The Canaries Dashboard, built with ADP Research, updates monthly (Stanford HAI, October 21 2026).

What does an AI exposure score measure, and what does it leave out?

The Stanford analysis groups workers by occupation, using exposure measures from Eloundou and coauthors (2024), as the lab's summary cites them. An exposure score attaches to a job category. It is not a record of anyone using AI, and it does not show whether use at a given employer automates a task or assists the person doing it.

The question a CHRO or CFO is actually asked is local: did AI change hiring in our teams, and by how much? A population gap cannot answer it in either direction. Turning a pattern across occupations and age groups in a payroll sample into a claim about one organization's teams is the most common error in this area.

Question an employer asks What the Stanford data can speak to What the employer needs to see
Is AI changing early-career hiring in general? Yes, as a pattern across the ADP sample, with the limits above Nothing further for the general question
Is AI changing hiring in our own teams? No, the data is not about our organization Hiring and backfills by team, over time
Who in our organization actually uses AI? No, exposure is a property of the occupation Observed AI use by team, from work-tool signals
Is the use automating work or assisting people? Only at the level of occupations The same classification, within our roles
Does adoption explain the difference between teams? No Adopting teams against comparable teams that have not adopted yet

How can an employer check its own organization?

The method borrows the logic of a comparison group rather than the headline. Applied inside one organization:

  1. Separate hiring from separations. Track requisitions, offers, and backfills by team, not only headcount.
  2. Measure observed AI use by team. Use signals from the tools where work happens, not an exposure rating and not a survey of intentions. Our piece on 48% of surveyed US employees admitting they exaggerate AI use covers why self-report fails.
  3. Fix the comparison before looking. Compare teams that adopted with comparable teams that have not adopted yet, controlling for tenure, role, and tool stack. This is a stronger design than a company-wide before-and-after. See how to measure AI ROI across the workforce.
  4. Keep stated reasons apart from measured ones. A hiring decision justified by AI is a statement, not a measurement. We made that point in AI Is the Top Stated Reason for Job Cuts. Stated Is Not Measured.
  5. Report confidence and the limit. State what the result cannot show, such as early adopters differing in ways the controls miss.

Where does Levos fit?

Levos is a Human Capital Operating System: an intelligence layer above the existing stack. Existing tools keep doing their jobs. Levos reads behavioral signals from the tools where work happens, M365, GitHub, Salesforce, and AI tools among them, and shows, with a confidence score, how adopting teams compare with non-adopting ones. AI Impact is one of its six signal families. The measurement methodology uses controlled cohort analysis with confidence scoring, compares adopting teams with non-adopting teams, and discloses its limits. We do not claim full attribution.

Individual-level data is visible to the person and their direct manager. Above that level, people see team aggregates of at least 5 people, apart from logged escalations; see the privacy model.

Frequently asked questions

What did Stanford's revised Canaries paper find about young workers and AI?

Using ADP payroll data through June 2026, Stanford's Digital Economy Lab reports that employment among workers aged 22 to 25 in highly AI-exposed occupations is about 19% below where it would be had it kept pace with similarly aged workers in less-exposed occupations. The gap was 15% at the July 2025 data vintage. Experienced workers show no comparable gap, and the authors find no widespread, economy-wide displacement.

Does the Stanford study show that AI caused the employment gap?

No. The authors describe the patterns as descriptive and say they cannot yet answer that question definitively. They tested alternative explanations such as interest rates and remote work and found none that appeared sufficient to account for the pattern, but they also report that the gaps shrink when education is accounted for, that some differential trends predate generative AI, and that estimates are sensitive to specification choices.

What is an AI exposure score?

An AI exposure score is a rating attached to an occupation, based on how much of its work AI could affect. It describes the job category, not any employer and not any individual's actual use of AI tools.

Can a company use the 19% figure to predict its own hiring?

No. The 19% is a gap between groups of workers across the ADP payroll sample. It does not describe any single employer, and the authors note that gaps are larger in the ADP sample than in national survey benchmarks.

How can an employer test AI's effect on its own hiring?

Track hiring and backfills separately from separations, measure actual AI use by team from work-tool signals, and compare teams that adopted with comparable teams that have not adopted yet over the same window. Report a confidence level and a stated limit with the result.

See your own teams, not the population average

If your hiring story rests on an exposure score or a leader's rating of AI impact, nothing underneath it can be checked. Request a Demo and we will walk through what a comparison-based readout contains.

Levos is accepting design partner applications from US organizations of 100 employees or more with an active AI rollout. Large organizations typically start with one function or division, measured against comparable teams that have not adopted yet. That is a stronger attribution design than a company-wide before-and-after.

Stanford Digital Economy Lab, "No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19%," summary of the revised "Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence" by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, August 12 2026; ADP payroll data through June 2026. https://digitaleconomy.stanford.edu/news/canariesaug26/

Stanford Digital Economy Lab, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," publication page, revised August 2026. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/

Adam Hadhazy, "Tracking AI's Real-Time Footprint On The Labor Market," Stanford HAI, October 21 2026 (Canaries Dashboard built with ADP Research; updated monthly). https://hai.stanford.edu/news/tracking-ais-real-time-footprint-on-the-labor-market

Cited within the lab's summary and not opened by us: Eloundou et al. (2024), exposure measures; Tucker (2026), U.S. Census Bureau.

Share this article

Help others discover workforce intelligence insights

Levos.AI Editorial

Levos.AI Editorial publishes operator-grade research on workforce intelligence, AI deployment measurement, and human capital optimization. Reach the team at marketing@levos.ai