The AI employment shock remains mild, but occupation-level pressure is becoming hard to ignore

Overview

Morgan Stanley uses data updated through the first half of 2026 to assess AI's impact on the U.S. labor market. The conclusion is that the overall shock remains relatively small, showing a pattern of “micro stronger than macro”: evidence at the occupation and worker levels continues to strengthen, while aggregate industry employment and job-posting data remain mixed. The unemployment rate for high-AI-exposure occupations continues to rise, while other exposure groups are broadly stable; this change is more pronounced after controlling for occupational cyclical sensitivity.

Core views

First, the cyclically adjusted unemployment rate for high-AI-exposure occupations is now about 0.5 percentage points above the level implied by historical relationships, and statistical significance has risen to the 90% confidence level. Second, the worsening in unemployment comes from both a decline in reemployment rates and an increase in layoffs since 2025, while changes in labor force participation across AI exposure groups show no obvious differences. Third, workers aged 22 to 27 are at the forefront of the adjustment, with especially clear unemployment-rate divergence in high-exposure occupations. Fourth, AI is accelerating the reorganization of work tasks, but the share of high-exposure workers reporting task changes declined in the first half of 2026, which may reflect short-term noise or a slowdown in AI implementation. Fifth, industry-level employment data and job postings are not yet sufficient to support a broad substitution narrative, and medium-exposure industries have recently performed even better than low- and high-exposure industries.

Analysis framework

The report builds an AI labor disruption tracking framework, integrating CPS household survey microdata, worker-flow statistics, task-reallocation indicators, industry payroll employment, job postings, and company earnings-call text. The study groups occupations and industries by AI exposure and uses regressions to control for the historical cyclical sensitivity of different groups; it mainly uses the AI exposure measure proposed by Felten et al. in 2021, while also conducting robustness checks with alternative measures from Morgan Stanley's thematic research team.

Methodology notes

• Integrated monitoring frameworkAI Labor Disruption Tracker

  Cross-identify AI-related employment changes using multiple types of labor-market indicators

  The framework covers unemployment rates, worker flows, task changes, industry employment, job postings, and corporate text to compare signals at the micro and macro levels.

• Exposure measurementFelten et al. (2021) AI Exposure Measure

  Classify occupations into low-, medium-, and high-exposure groups based on the extent to which they are affected by AI capabilities

  This measure is used for the report's main results, supplemented by alternative exposure measures from Morgan Stanley's thematic research team for robustness checks.

• Econometric analysisCyclically Adjusted Residual Analysis

  Control for the historical cyclical relationship between each exposure group and the overall labor market

  The study regresses the unemployment rate of each exposure group on the overall unemployment rate and analyzes the residuals to measure the degree to which actual unemployment deviates from the normal cyclical relationship.

• Time-series processing12-Month Moving Average

  Smooth seasonality and high-frequency noise in non-seasonally adjusted CPS data

  Because the relevant CPS data are not seasonally adjusted, the report uses 12-month moving averages to observe sustained changes across different AI exposure groups.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• U.S. labor market

  Core macro observation object for AI disruption

  Strengths

  The overall employment impact remains small, industry employment has not shown a broad decline, and labor force participation has not displayed systematic divergence across exposure groups.

  Weaknesses

  The unemployment rate in high-AI-exposure occupations continues to rise, reemployment has become more difficult, and the layoff rate remains elevated relative to pre-2025 levels.

  Comparison

  Disruption signals at the occupation and worker levels are clearly stronger than the macro signals reflected in industry payroll employment and job postings.

  Risks

  If weakness at the occupation level continues to spread, it could push up the overall unemployment rate and increase the risk of long-term unemployment.

• Artificial intelligence theme

  A structural theme in which productivity gains and job substitution coexist

  Strengths

  Corporate discussion still focuses mainly on productivity and process-efficiency improvement, and work-task restructuring indicates that AI is gradually entering production processes.

  Weaknesses

  The social costs of job substitution are beginning to appear in high-exposure occupations and youth groups, which may raise regulatory and policy attention.

  Comparison

  Evidence of task changes is relatively strong, but total industry employment and job-posting data have not yet validated large-scale substitution.

  Risks

  If job reductions from AI adoption occur faster than new job creation, thematic valuations may face social, regulatory, and demand-side risks.

Key data

• Upper-bound impact on the overall unemployment rateAbout 15 basis pointsA rough estimate as of June 2026, higher than about 10 basis points in December 2025; part of the difference may still be caused by non-AI factors.
• Abnormal unemployment increase in high-AI-exposure occupationsAbout 0.5 percentage pointsRelative to the normal level predicted by overall labor-market conditions and historical cyclical relationships.
• Employment share of high-AI-exposure occupationsAbout 30%Used to convert abnormal unemployment within the occupational group into an impact on the overall unemployment rate.
• Statistical significance90% confidence levelAs of the first half of 2026, unemployment residuals for high-exposure occupations had exceeded the two-standard-error range; previously they were significant only at the 68% confidence level.
• Key affected age group22 to 27 years oldUnemployment-rate divergence among young workers in high-AI-exposure occupations is significantly greater than for the overall workforce.
• Data observation periodThrough the first half of 2026The report was published on August 3, 2026.

Impact & implications

The current evidence does not yet support the view that AI has triggered a broad unemployment shock in the U.S. labor market, but it suggests that structural adjustment may first appear in high-exposure occupations, among labor-market entrants, and in specific work tasks. If occupation-level divergence continues to widen and gradually transmits to industry employment, the overall unemployment rate and wage growth may face more visible pressure; conversely, if companies continue to focus mainly on process optimization and productivity improvement, AI's positive impact on output efficiency may emerge before large-scale job substitution. Current investment and policy judgments should not rely solely on aggregate industry data, but should also track more granular occupation-level and worker-flow indicators.

Risks

• There is an omitted-variable issue: monetary policy, tariffs, immigration policy, and post-pandemic hiring normalization may all cause divergence across occupations.
• The report's estimated 15 basis points is an upper-bound impact, not a causally identified pure AI effect.
• Industry-level data are noisy; medium-exposure industries have performed relatively strongly, which does not support a monotonic relationship between AI exposure and employment deterioration.
• The sample of non-college youth in high-exposure occupations is relatively small, and the related estimates have a lower signal-to-noise ratio.
• Task-change indicators cannot distinguish between AI's augmentation effect and substitution effect on workers.
• The decline in job postings predates the release of ChatGPT and cannot be simply attributed to generative AI.

What to watch

• Whether cyclically adjusted unemployment residuals for high-AI-exposure occupations continue to widen and remain statistically significant.
• Changes in the reemployment rate, unemployment duration, and layoff rate for unemployed workers in high-exposure occupations.
• Unemployment-rate divergence among workers aged 22 to 27 across different education groups.
• Whether employment growth in high-AI-exposure industries weakens further relative to historical trends.
• Whether job postings shift from post-pandemic normalization to a systematic decline related to the degree of AI exposure.
• Changes in company earnings-call language related to productivity improvement, job substitution, and job creation.
• Whether the pace of task changes in high-exposure occupations reaccelerates or continues to slow.