## wpiea2024199-print-pdf

## Source details

**Canonical URL:** [wpiea2024199-print-pdf](https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024199-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2024/english/wpiea2024199-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2024/english/wpiea2024199-print-pdf.pdf.json)

---

### Main empirical finding: AI exposure and employment-to-population
- Commuting zones with a higher share of AI adopting firms experienced a more significant decline in the overall employment-to-population ratio during 2010-2021.
- A one standard deviation increase in AI exposure leads to 0.976 percentage points lower employment-to-population.
- Employment-to-population in commuting zones at the 75th percentile of AI exposure declines by 1.25 percentage points more than commuting zones at 25th percentile of AI exposure.
- Heterogeneous impacts:
  - Primarily borne by the manufacturing and low-skill services sectors.
  - Disproportionately affects middle-skill workers and non-STEM occupations.
  - Larger adverse impact on individuals at the two ends of the age distribution (young and older workers).
  - More pronounced effect on men than women.
- Robustness and falsification:
  - Main findings robust to alternative definitions of US industry-level AI adoption, alternative constructions of AI exposure and its IV with local employment shares in alternative years, and using 2019 as the end year to address concerns about Covid-19 employment impacts.
  - Falsification tests: after controlling for a wide range of commuting zone covariates, AI adoption in 2010-2021 does not affect past changes in the employment-to-population ratio in 1980-2010 (insignificant coefficients reported: 2.217, -1.199, 0.716 in Table 1 columns (4)-(6)).

### Data and measurement
- Industry-level AI adoption:
  - United States: Annual Business Survey (ABS) technology module (2018, 2019, 2021); paper uses 2021.
    - Five AI technologies in the ABS: machine learning, machine vision, natural language processing, voice recognition software, and automated-guided vehicles (AGVs).
    - Baseline industry-level adoption rate: percentage of firms in an industry that adopt a given AI technology, averaged across AI technologies.
    - ABS coverage: 47 industries (2-digit NAICS; 3-digit NAICS for manufacturing; 4-digit NAICS for professional, scientific and technical services).
    - Highest industry adoption reported at 6% for data processing, hosting, and related services.
  - European Union: ICT Usage in Enterprises (European Commission) under NACE Rev. 2; baseline measure uses 2021 percentage of enterprises that use at least one of: text mining, speech recognition, natural language processing, machine learning, AI-based software robotic process automation, and autonomous robots/vehicles/drones.
    - EU data includes 27 industries in the ICT Usage in Enterprises data.
- Commuting zone aggregation and labour outcomes:
  - 722 commuting zones using Autor and Dorn (2013) crosswalks.
  - American Community Survey (ACS) 5% sample from IPUMS used to compute commuting zone characteristics and employment-to-population (employed working-age individuals aged 16-65 divided by total working-age population, using census weights).
  - County Business Patterns (CBP) provides county-level industry employment used to construct local employment shares and Bartik-style commuting zone exposure to AI, industrial robots, and Chinese import competition.
- Additional controls:
  - Initial commuting zone demographic characteristics and industrial structure (examples: share of foreign born, share of population with college degrees), initial share of routine occupations, initial share of high offshorability occupations, and Bartik exposures to robotization (IFR) and Chinese import competition (CEPII BACI).

### Commuting-zone exposure construction and IV strategy
- Baseline Bartik-style exposure (USExposure_i):
  - USExposure_i = Σ_j (L_{ij2010} / L_{i2010}) · Δ^{2021}_{2010} AIAdoption^{US}_j
    - Weights: local employment share of industry j in commuting zone i in 2010.
    - Shift: nationwide industry-specific change in AI adoption in 2010-2021 in the US from ABS (2021).
- Instrument (EU-based IV) to mitigate anticipation and local-demand bias:
  - EUExposure_i = Σ_j (L_{ij1990} / L_{i1990}) · Δ^{2021}_{2010} AIAdoption^{EU}_j
    - Uses 1990 local employment shares to mitigate anticipation/path-dependence concerns.
    - Shift uses industry-level AI adoption in the EU to capture global technological advances.
  - Diagnostics:
    - Strong positive relationship between industry-level AI adoption in the US versus the EU: linear regression fit coefficient 5.255 with standard error 0.874.
    - First-stage F-statistic reported as 58.2 (described as “well above 10”).
    - Robustness checks use 1995 shares and 1990-1995 average; robustness to excluding top 1% commuting zones by USExposure_i.

### Empirical specification and estimation
- Baseline long-difference specification:
  - Δ^{2010 2021} Y_i = α_{d(i)} + β · AIExposure_i + γ X_i + ε_i
    - i denotes commuting zones; α_{d(i)} is census division fixed effect.
    - Dependent variable: long difference of Yi between 2010 and 2021; 2010 set as start year under assumption of no AI adoption in 2010.
    - Xi baseline controls: log of population size; share of female population; share of population aged above 65; share of white/black/American Indian or Alaskan native/Asian population; share of foreign born; share of college-educated workers; manufacturing share; light manufacturing share; initial share of routine occupations; initial share of high offshorability occupations; Bartik exposures to robotization and Chinese import competition.
- 2SLS implementation:
  - First stage: USExposure_i = ˜α_{d(i)} + ˜β · EUExposure_i + ˜γ X_i + ˜ε_i
  - Second stage: Δ^{2021}_{2010} Y_i = α_{d(i)} + β · ˆUSExposure_i + γ X_i + ε_i
  - Regressions weighted by commuting zone population in 2010. Standard errors clustered at the state level.

### Main quantitative results and diagnostics (overall employment-to-population, 2010–2021)
- Table 1 (2SLS second-stage baseline IV using 1990 share) — USExposure coefficients (standard errors):
  - Column (1): -7.511 (3.067) — significance: ∗∗ (5 percent).
  - Column (2): -5.699 (2.979) — significance: ∗ (10 percent).
  - Column (3): -8.375 (3.129) — significance: ∗∗∗ (1 percent).
- First-stage reported:
  - First-stage coefficient (column (1)): 0.075 ∗∗∗ (0.010).
  - First-stage F-statistic reported as 58.2.
- Interpretation:
  - A one standard deviation increase in AI exposure implies 0.976 percentage points lower employment-to-population.
  - Commuting zones at the 75th percentile of AI exposure decline by 1.25 percentage points more than those at the 25th percentile.
- Falsification results (1980-2010 dependent variable; Table 1 columns (4)-(6)):
  - Coefficients: 2.217, -1.199, 0.716 (insignificant).
- Robustness to outliers (excluding top 1% USExposure_i; Table 2):
  - USExposure coefficients reported: -8.968 ∗∗ (4.156), -6.345 (3.912), -10.054 ∗∗∗ (4.225).
  - First-stage coefficients ~0.060–0.067; first-stage F-statistics 41.5–43.1.
- Employment level (change in log overall employment level 2010-2021 controlling for changes in log working-age population; Table 3):
  - USExposure coefficients: -10.970 ∗∗ (4.856), -8.759 ∗ (4.701), -12.351 ∗∗∗ (4.951).
  - First-stage F-statistic example: 59.5 reported in one specification.
  - Conclusion: negative effect in employment-to-population ratio is driven by negative effect on employment (numerator), not solely by population changes.

### Geographic and industry patterns
- Highest commuting-zone AI exposure observed in San Francisco, Los Angeles, San Antonio, Seattle, Pittsburgh, New York, Washington D.C., and Boston under both USExposure_i and EUExposure_i specifications.
- Industry-level adoption: data processing, hosting, and related services highest at 6%; other high-adoption industries include computer systems design, publishing, machinery, computer and electronic products, paper products, plastic and rubber products, transportation equipment, and scientific research and development.

### Heterogeneity: sectoral, occupational, education, age, gender
- Sectoral impacts (Table 4; baseline IV with 1990 shares):
  - Agriculture: 0.914 ∗∗ (0.460)
  - Manufacturing: -5.118 ∗ (2.782)
  - Construction: 1.047 (1.091)
  - Low-Skill Services: -5.292 ∗∗∗ (2.039)
  - High-Skill Services: 0.939 (1.635)
  - Note: manufacturing includes manufacturing and mining; low-skill services and high-skill services classified as described in source.
- Occupation impacts (Table 5; baseline IV with 1990 shares):
  - Non-STEM occupations: -6.997 ∗∗∗ (2.881)
  - STEM occupations: -0.514 (1.049)
  - Low-Skill occupations: -0.230 (0.980)
  - Middle-Skill occupations: -4.936 ∗ (2.559)
  - High-Skill occupations: -2.345 (1.701)
- Education heterogeneity (Table 6; 2SLS, 1990 share):
  - Below High School: -2.598 (5.850)
  - High School: -9.723 ∗∗∗ (3.935)
  - Some College: -6.216 ∗ (3.729)
  - College and Above: -0.550 (2.401)
  - Observations: 722 per column.
- Age and gender heterogeneity (Table 7; 2SLS, 1990 share):
  - Age bins:
    - 16-25: -11.519 ∗∗ (5.606)
    - 26-35: -2.962 (3.423)
    - 36-45: -5.576 (4.007)
    - 46-55: -7.746 ∗∗ (3.750)
    - 56-65: -7.969 ∗ (4.108)
  - Gender:
    - Male: -9.191 ∗∗ (4.214)
    - Female: -5.581 ∗ (3.089)
  - Observations: 722 per column.
- Interpretation:
  - Negative impacts concentrate on manufacturing and low-skill services, non-STEM and middle-skill occupations, workers with high school or some college education, young (16-25) and older (46+) age groups, and men more than women.

### Robustness exercises and alternative measures
- Alternative IVs: 1995 local employment share; 1990-1995 average local share — results broadly consistent.
- Alternative AIAdopt_US_j measure: maximum adoption rate across five AI technologies — results consistent in magnitude and pattern across overall, sectoral, occupational, education, age, and gender breakdowns (selected estimates reported across Tables A.11–A.15).
  - Example overall (2010-2021; 1990 Share using maximum adoption): USExposure = -3.785 ∗∗ (1.628) — Observations 722; R-squared 0.20; First-stage coefficient 0.149; First-stage F-statistic 33.2.
- Alternative end year: using 2019 as end year (2010-2019) to address Covid-19 concerns — core negative results persist.
  - Example overall (2010-2019; 1990 Share): USExposure = -7.060 ∗∗ (3.088) — Observations 722; R-squared 0.37; First-stage coefficient 0.075; First-stage F-statistic 58.2.
- Alternative local shares for weights (2005, 1995, 1990-1995 average) — qualitative findings robust; some first-stage F-statistics vary (examples reported: 58.2, 52.8, 57.3; lower F-statistics noted in some specifications).

### Interpretation relative to prior labor shocks and literature contribution
- Unequal effects of AI parallel previous labor market shocks:
  - Routine-biased technological change (Autor et al. (2006); Goos et al. (2014) for skill group).
  - Offshoring (Goos et al. (2014) for skill group).
  - Robotization (Acemoglu and Restrepo (2020) for skill group and gender).
  - Import competition (Traiberman (2019) for age).
- Measurement and methodological contribution:
  - Uses firm/industry-level AI adoption (ABS) rather than occupational-task exposure approaches, enabling direct estimation of employment impacts using historical data.
  - Instruments US industry-level adoption with EU adoption to capture global technological advances and mitigate local-demand confounding.
  - Comparison to other empirical studies: firm/establishment-level vacancy studies show mixed findings; studies of generative AI find short-run labor demand reductions for freelancers.

### Policy implications and research priorities
- Policy recommendations:
  - Strengthen social safety nets targeted to affected groups: middle-skill workers, non-STEM occupations, workers in manufacturing and low-skill services, young and older workers.
  - Expand job retraining programs to facilitate reskilling, especially for middle-skill and older workers facing skill obsolescence.
- Research and policy priorities:
  - Improve large-scale, up-to-date micro-level panel data on AI adoption to analyze generative-AI era effects.
  - Develop fully-specified general equilibrium models to account for cross-region spillovers and to gauge aggregate effects from regional estimates.
  - Investigate geographic specialization across the AI value chain (AI production vs AI usage).
  - Extend empirical analysis to other outcomes: wages, housing prices, and political views.

*Content drawn from wpiea2024199-print-pdf (excerpt provided).*

### 2010.  Therefore, long-run common factors are unlikely to be the main drivers for both

### wpiea2024199-print-pdf - 2010. Therefore, long-run common factors are unlikely to be the main drivers for both

### Main empirical finding: AI exposure and employment-to-population
- Commuting zones with a higher share of AI adopting firms experienced a more significant decline in the overall employment-to-population ratio during 2010-2021.
- A one standard deviation increase in AI exposure leads to 0.976 percentage points lower employment-to-population.
- Employment-to-population in commuting zones at the 75th percentile of AI exposure declines by 1.25 percentage points more than commuting zones at 25th percentile of AI exposure.
- The negative effect is heterogeneous across sectors, skills, occupations, age, and gender:
  - Primarily borne by the manufacturing and low-skill services sectors.
  - Disproportionately affects middle-skill workers and non-STEM occupations.
  - Larger adverse impact on individuals at the two ends of the age distribution.
  - More pronounced effect on men than women.
- Robustness:
  - Main findings robust to alternative definitions of US industry-level AI adoption, alternative constructions of AI exposure and its IV with local employment shares in alternative years, and using 2019 as the end year to address concerns about Covid-19 employment impacts.
  - Falsification tests: after controlling for a wide range of commuting zone covariates, AI adoption in 2010-2021 does not affect past changes in the employment-to-population ratio in 1980-2010.

### Interpretation and relation to prior labor shocks
- Unequal effects of AI parallel previous labor market shocks:
  - Routine-biased technological change (Autor et al. (2006), Goos et al. (2014) for skill group).
  - Offshoring (Goos et al. (2014) for skill group).
  - Robotization (Acemoglu and Restrepo (2020) for skill group and gender).
  - Import competition (Traiberman (2019) for age).
- Policy implication emphasized: need to consider unequal distributional consequences of labor market shocks and the importance of social safety nets and job retraining programs.

### Contribution to literature and measurement choices
- Distinction from occupational-task exposure approaches:
  - Many studies map AI progress to tasks/occupational content (Frey and Osborne (2017), Webb (2020), Felten et al. (2021), Eloundou et al. (2023)) and remain agnostic on complement vs substitute effects.
  - Cazzaniga et al. (2024) augment occupational exposure with a potential complementarity index across six countries.
- This paper uses AI adoption by firms (industry-level adoption) rather than occupational task content, enabling direct estimation of employment impacts using historical data and allowing instrumenting US industry-level adoption with EU data to capture global technological advances.
- Comparison to other empirical studies:
  - Firm/establishment-level studies using vacancy data produce mixed findings (Acemoglu et al. (2022b), Copestake et al. (2023), Babina et al. (2024)).
  - Hui et al. (2023) find generative AI reduces labor demand for freelancers in the short run.
  - Bonfiglioni et al. (2024) also analyze commuting zones and find stronger negative impact in more exposed zones; key difference: Bonfiglioni et al. (2024) measure AI exposure using changes in commuting zone employment share of 19 AI professions, whereas this paper leverages firm-level AI adoption from a nationally representative survey.

### Data sources and measurement details
- Industry-level AI adoption:
  - United States: Annual Business Survey (ABS), technology module for years 2018, 2019, and 2021; this paper uses 2021.
    - ABS reports number of firms that use a given AI technology at industry level.
    - Five AI technologies in the ABS: machine learning, machine vision, natural language processing, voice recognition software, and automated-guided vehicles (AGVs).
    - Baseline industry-level adoption rate: percentage of firms in an industry that adopt a given AI technology, averaged across AI technologies.
    - ABS coverage: 47 industries in the ABS (with 2-digit NAICS, 3-digit NAICS for manufacturing, and 4-digit NAICS for professional, scientific and technical services).
  - European Union: ICT Usage in Enterprises (European Commission) under NACE Rev. 2; baseline measure uses 2021 percentage of enterprises that use at least one of: text mining, speech recognition, natural language processing, machine learning, AI-based software robotic process automation, and autonomous robots/vehicles/drones.
    - EU data includes 27 industries in the ICT Usage in Enterprises data.
  - Rationale for EU IV: isolate US-specific shocks and capture global technological advances; coarser EU industry granularity mitigated by IV relevance (first-stage F statistic) and robustness checks excluding top 1% commuting zone by AI exposure.
- Commuting zone level data:
  - Aggregation using Autor and Dorn (2013) crosswalks; 722 commuting zones in total.
  - American Community Survey (ACS) 5% sample from IPUMS used to compute commuting zone characteristics and employment-to-population.
    - Employment-to-population defined as number of employed working-age individuals (aged 16-65) divided by total working-age population, using census weights.
  - County Business Patterns (CBP) provides county-level industry employment used to construct local employment shares and Bartik-style commuting zone exposure to AI, industrial robots, and Chinese import competition.
- Additional controls and confounders:
  - Wide range of initial commuting zone demographic characteristics and industrial structure from ACS to address comparability concerns (examples: share of foreign born, share of population with college degrees).
  - Bartik exposures computed for robotization using International Federation of Robotics (IFR) data and for Chinese import competition using CEPII BACI (Gaulier and Zignago (2010)) for 2010-2021.

### Empirical strategy and specification
- Identification approach:
  - Shift-share (Bartik) design exploiting regional variation in AI adoption to infer causal relationships, instrumenting US industry-level adoption with EU data to capture global technologies.
  - Econometric foundation relates to Bartik literature (Bartik (1991); Adão et al. (2019); Goldsmith-Pinkham et al. (2020); Breuer (2022); Borusyak et al. (2022)); cross-regional inference methods related to Nakamura and Steinsson (2018) and Acemoglu and Restrepo (2020).
- Baseline empirical specification (long-difference):
  - ∆2010 2021 Yi = αd(i) + β AIExposurei + γ Xi + εi
    - i denotes commuting zones; d(i) refers to the census division of commuting zone i; αd(i) is census division fixed effect.
    - Yi refers to labor market outcomes in commuting zone i (e.g., overall employment-to-population ratio or subgroup-specific employment-to-population).
    - Dependent variable is the long difference of Yi between 2010 and 2021; 2010 is set as start year under assumption of no AI adoption in 2010.
    - Main coefficient of interest is β, capturing effect of commuting zone-level AI exposure on local labor market outcomes.
    - Baseline controls Xi include initial demographic characteristics (log of population size; share of female population; share of population aged above 65; share of white/black/American Indian or Alaskan native/Asian population; share of foreign born; share of college-educated workers), initial industrial structure (manufacturing share; light manufacturing share), initial share of routine occupations (proxy for exposure to routine-biased technological change), initial share of high offshorability occupations, and Bartik exposures to robotization and Chinese import competition.
- Timing and robustness to Covid-19:
  - 2010 chosen as start year; EFF AI progress measures indicate little progress in 2010/2011.
  - Alternative specification uses 2019 as end year to alleviate Covid-19 concerns (results presented in Appendix E).

*Italic: Content drawn from wpiea2024199-print-pdf (excerpt provided).*

### 3.2    Commuting Zone Level Exposure to AI

### 3.2    Commuting Zone Level Exposure to AI

### Methodology: Commuting-zone AI exposure (Bartik-style)
- USExposure_i is computed as:
  - USExpsoure_i = Σ_j (L_{ij2010} / L_{i2010}) · Δ^{2021}_{2010} AIAdoption^{US}_j
  - Weights: local employment share of industry j in commuting zone i in 2010, L_{ij2010}/L_{i2010} (“share”).
  - Shift: nationwide industry-specific change in AI adoption in 2010-2021 in the US (Δ^{2021}_{2010} AIAdoption^{US}_j) from ABS (2021).
- Baseline industry-level AI adoption measure: average percentage of adopting firms across five AI technologies (AGV, machine learning, voice recognition, speech recognition, text mining). Highest industry adoption reported at 6% for data processing, hosting, and related services.
- Key methodological concerns noted:
  - Local employment shares in 2010 may reflect anticipation of AI arrival and path dependence of technological change → simultaneity bias.
  - Shift component (US industry adoption) may reflect US-specific demand shocks correlated with local labor demand → positive bias in OLS.

### Instrumental variable (EU-based IV) and relevance
- Instrument constructed as:
  - EUExposure_i = Σ_j (L_{ij1990} / L_{i1990}) · Δ^{2021}_{2010} AIAdoption^{EU}_j
  - Uses 1990 local employment shares to mitigate anticipation/path-dependence concerns.
  - Shift uses industry-level AI adoption in the EU to capture global technological advances.
- Justification and diagnostics:
  - Figure 2: strong positive relationship between industry-level AI adoption in the US versus the EU; linear regression fit coefficient 5.255 with standard error 0.874.
  - Reported first-stage strength: F-statistic reported in text as 58.2, described as “well above 10,” indicating EUExposure_i is a strong instrument.
  - Robustness checks: local employment shares in 1995 and average 1990-1995 used; robustness to excluding top 1% of commuting zones by USExposure_i.

### Empirical strategy (2SLS)
- First stage:
  - USExposure_i = ˜α_{d(i)} + ˜β EUExposure_i + ˜γ X_i + ˜ε_i
- Second stage:
  - Δ^{2021}_{2010} Y_i = α_{d(i)} + β · ˆUSExposure_i + γ X_i + ε_i
- Regressions weighted by commuting zone population in 2010. Standard errors clustered at the state level.

### Main results: Overall employment-to-population ratio (2010–2021)
- Core finding: commuting zones with higher AI exposure experienced stronger declines in employment-to-population ratio during 2010-2021.
- Key point estimates and diagnostics reported:
  - Table 1 (2SLS second-stage estimates, baseline IV using 1990 share):
    - USExposure coefficient (column (1)): -7.511 (standard error 3.067), significance: ∗∗ (5 percent).
    - Column (2): -5.699 (2.979), significance: ∗ (10 percent).
    - Column (3): -8.375 (3.129), significance: ∗∗∗ (1 percent).
    - First-stage coefficient reported as 0.075 ∗∗∗ (0.010) in column (1); first-stage F-statistic reported as 58.2.
  - Interpretation in text: a one standard deviation increase in AI exposure implies 0.976 percentage points lower employment-to-population ratio. The estimate also implies that commuting zones at the 75th percentile of AI exposure decline by 1.25 percentage points more than those at the 25th percentile.
- Falsification:
  - Regressing past changes in employment-to-population (1980-2010) on future AI exposure (2010-2021) yields insignificant coefficients (columns (4)-(6) of Table 1: 2.217, -1.199, 0.716), suggesting AI exposure affects outcomes only for 2010-2021 and not for 1980-2010.
- Robustness to outliers:
  - Excluding commuting zones in the top 1% of USExposure_i (Table 2) leaves the negative effect on employment-to-population ratio robust.
  - Table 2 reported USExposure coefficients (columns (1)-(3)): -8.968 ∗∗ (4.156), -6.345 (3.912), -10.054 ∗∗∗ (4.225). First-stage coefficient ~0.060–0.067 with first-stage F-statistics in the 41.5–43.1 range.
- Employment level vs population:
  - Using change in log overall employment level (2010-2021) controlling for changes in log working-age population (Table 3) shows negative effects on employment:
    - Table 3 reported USExposure coefficients (columns (1)-(3)): -10.970 ∗∗ (4.856), -8.759 ∗ (4.701), -12.351 ∗∗∗ (4.951).
    - First-stage coefficient and F-statistics reported (e.g., first-stage F-statistic 59.5 in one specification).
  - Conclusion: negative effect in employment-to-population ratio is driven by negative effect on employment (numerator), not solely by population changes.

### Geographic distribution and descriptive patterns
- Figure 3 (maps): Highest commuting-zone AI exposure observed in San Francisco, Los Angeles, San Antonio, Seattle, Pittsburgh, New York, Washington D.C., and Boston under both USExposure_i and EUExposure_i specifications.
- Industry-level adoption (Figure 1): data processing, hosting, and related services highest at 6%; other high-adoption industries include computer systems design, publishing, machinery, computer and electronic products, paper products, plastic and rubber products, transportation equipment, and scientific research and development.

### Heterogeneity of effects
- Four main heterogeneous-results findings summarized in text:
  1. Manufacturing and low-skill services sectors are negatively affected.
  2. The negative impact falls mainly on middle-skill workers (consistent with routine-biased technological change and job polarization patterns).
  3. AI exposure reduces employment-to-population ratio for age groups 16-25 and above 46.
  4. The adverse impact is more pronounced on men than women.
- Sectoral impacts (Table 4; baseline IV with 1990 shares):
  - Agriculture: 0.914 ∗∗ (0.460)
  - Manufacturing: -5.118 ∗ (2.782)
  - Construction: 1.047 (1.091)
  - Low-Skill Services: -5.292 ∗∗∗ (2.039)
  - High-Skill Services: 0.939 (1.635)
  - Note: manufacturing includes manufacturing and mining; low-skill services and high-skill services classified as described in the source.
- Occupational impacts (Table 5; baseline IV with 1990 shares):
  - Non-STEM occupations: -6.997 ∗∗∗ (2.881)
  - STEM occupations: -0.514 (1.049)
  - Low-Skill occupations: -0.230 (0.980)
  - Middle-Skill occupations: -4.936 ∗ (2.559)
  - High-Skill occupations: -2.345 (1.701)
  - Interpretation: negative employment impact concentrated in non-STEM and middle-skill occupations.

*Source: IMF working paper content (3.2 Commuting Zone Level Exposure to AI and related tables and figures).*

### 2021.  The list of STEM occupations are from O*NET. High-skill occupations are management, business

### wpiea2024199-print-pdf - 2021.  The list of STEM occupations are from O*NET. High-skill occupations are management, business

### Main empirical approach and identification
- Uses a shift-share measure of commuting zone AI exposure constructed from 2010 local employment shares and nationwide industry-level AI adoption.
- Instruments US commuting zone exposure using 1990 local employment shares and industry-level AI adoption in the EU to mitigate bias from local demand shocks, anticipation, and path dependence.
- Regressions weighted by 2010 commuting zone population. Robust standard errors in parentheses and clustered at the state level.

### Aggregate employment effect
- Estimated effect: commuting zones at the 75th percentile of AI exposure experience a decline in the employment-to-population ratio of 1.25 percentage points more than commuting zones at the 25th percentile of AI exposure (2010-2021).

### Heterogeneous effects by education (Table 6: 2SLS estimates; dependent variable = change in employment-to-population ratio by education, 2010-2021)
- Columns (1)-(4) correspond to: Below High School, High School, Some College, College and Above.
- USExposure coefficients and standard errors:
  - Below High School (1): -2.598 (5.850)
  - High School (2): -9.723 ∗∗∗ (3.935)
  - Some College (3): -6.216 ∗ (3.729)
  - College and Above (4): -0.550 (2.401)
- Observations: 722 for each column.
- Finding: Employment of individuals with middle levels of education—particularly high school graduates and those with some college—is negatively affected by AI exposure; the effect on college and above is small and not statistically significant.
- Interpretation: Consistent with routine-biased technological change, negative impacts concentrate on middle-skill/middle-education groups, contributing to job polarization.

### Heterogeneous effects by age and gender (Table 7: 2SLS estimates; dependent variable = change in employment-to-population ratio by 10-year age bin or gender, 2010-2021)
- Columns (1)-(5): age bins 16-25, 26-35, 36-45, 46-55, 56-65. Columns (6)-(7): Male, Female.
- USExposure coefficients and standard errors:
  - 16-25 (1): -11.519 ∗∗ (5.606)
  - 26-35 (2): -2.962 (3.423)
  - 36-45 (3): -5.576 (4.007)
  - 46-55 (4): -7.746 ∗∗ (3.750)
  - 56-65 (5): -7.969 ∗ (4.108)
  - Male (6): -9.191 ∗∗ (4.214)
  - Female (7): -5.581 ∗ (3.089)
- Observations: 722 for each column.
- Findings:
  - Negative employment impact of AI falls primarily on the very young (16-25) and older workers (46 and above).
  - Male employment experiences a more pronounced decline than female employment.
- Interpretation:
  - Young workers: lower employment-to-population ratios due to staying in school longer and a higher share of lower-skill jobs vulnerable to displacement.
  - Older workers: skills obsolescence, lower adaptability to new technologies, higher specific human capital and switching costs.
  - Gender: women may be relatively less adversely affected because occupations held by women can be more complementary to AI.

### Sectoral and occupational heterogeneity (robustness and appendices)
- Robustness checks using alternative IVs (1995 local employment share; 1990-1995 average local share), alternative AI adoption measure (maximum adoption rate across five AI technologies), and different end years (2019) produce consistent findings.
- Appendix B (1995 share in IV) — selected results:
  - Broad sector (Table A.3): USExposure coefficients (observations = 722):
    - Agriculture: 0.866 ∗ (0.489)
    - Manufacturing: -4.069 ∗ (2.321)
    - Construction: 1.671 (1.162)
    - Low-Skill Services: -6.058 ∗∗∗ (1.712)
    - High-Skill Services: 1.890 (1.496)
  - Occupation (Table A.4): USExposure coefficients (observations = 722):
    - Non-STEM: -6.045 ∗∗ (2.821)
    - STEM: -0.346 (1.019)
    - Low-Skill: -0.394 (1.239)
    - Middle-Skill: -4.020 (2.491)
    - High-Skill: -1.285 (1.750)
- Findings:
  - Negative employment effects are primarily borne by manufacturing and low-skill services.
  - Occupation-level: non-STEM and middle-skill occupations drive much of the negative employment impact; STEM occupations show smaller or insignificant effects.

### Robustness exercises summarized
- Alternative industry-level AI adoption measure: maximum adoption rate across five AI technologies — results consistent.
- Using 2019 as end year to address Covid-19 concerns — results consistent.
- Using 2005 local employment shares for USExposure to mitigate anticipation/mean reversion — results consistent.
- Overall robustness: negative employment effect concentrated in manufacturing, low-skill services, middle-skill workers, non-STEM occupations, and at the two ends of the age distribution; larger adverse impact on men.

### Policy implications and recommendations
- Unequal distributional consequences of AI adoption underscore the need for:
  - Social safety nets targeted to affected groups (middle-skill workers, non-STEM occupations, workers in manufacturing and low-skill services, young and older workers).
  - Job retraining programs to facilitate reskilling, especially for middle-skill and older workers facing skill obsolescence.
- Research and policy priorities:
  - Improve large-scale, up-to-date micro-level panel data on AI adoption to analyze generative-AI era effects.
  - Develop fully-specified general equilibrium models to account for cross-region spillovers and to gauge aggregate effects from regional estimates.
  - Investigate geographic specialization across the AI value chain (AI production vs AI usage).
  - Extend empirical analysis to other outcomes: wages, housing prices, and political views.

*Source: https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024199-print-pdf.pdf*

### 2021.  The list of STEM occupations are from O*NET. High-skill occupations are management, business

### wpiea2024199-print-pdf - 2021.  The list of STEM occupations are from O*NET. High-skill occupations are management, business

### Occupation (1995 Share in IV)
- Non-STEM: USExposure = -5.527-8.665 ∗∗ -3.5860.272 (standard errors: (6.209)(3.502)(3.815)(2.494)) — Observations 722.
- Note: Table A.5 reports second stage estimates β from equation (5); dependent variable is change in employment-to-population ratio by education levels in 2010-2021. Regressions weighted by 2010 commuting zone population; robust standard errors in parentheses and clustered at the state level.

### Age and Gender (1995 Share in IV)
- 16-25: USExposure = -11.499 ∗∗ (5.260) — Observations 722.
- 26-35: USExposure = -1.458 (3.860) — Observations 722.
- 36-45: USExposure = -6.657 ∗ (3.789) — Observations 722.
- 46-55: USExposure = -6.620 ∗ (3.853) — Observations 722.
- 56-65: USExposure = -0.826 (3.980) — Observations 722.
- Male: USExposure = -6.613 ∗ (3.758) — Observations 722.
- Female: USExposure = -4.596 (3.273) — Observations 722.
- Note: Table A.6 reports second stage estimates β from equation (5); dependent variable is change in employment-to-population ratio by 10-year age bins or gender in 2010-2021.

### Broad Sector (1990-1995 Average Share in IV)
- Agriculture: USExposure = 0.878 ∗ (0.453) — Observations 722.
- Manufacturing: USExposure = -4.976 ∗ (2.605) — Observations 722.
- Construction: USExposure = 1.260 (1.126) — Observations 722.
- Low-Skill Services: USExposure = -6.432 ∗∗∗ (1.743) — Observations 722.
- High-Skill Services: USExposure = 0.895 (1.484) — Observations 722.
- Note: Table A.7 dependent variable is change in sectoral employment-to-population ratio in 2010-2021. Definitions: Manufacturing includes manufacturing and mining. Low-skill services are wholesale trade, retail trade, utilities, transportation, information, real estate, administrative support and waste management, arts and entertainment, accommodation and food services, and other services. High-skill services are finance and insurance, professional scientific and technical services, management of companies and enterprises, education, health, and social assistance.

### Occupation (1990-1995 Average Share in IV)
- Non-STEM: USExposure = -7.781 ∗∗∗ (2.838) — Observations 722.
- STEM: USExposure = -0.594 (1.096) — Observations 722.
- Low-Skill: USExposure = -0.559 (1.124) — Observations 722.
- Middle-Skill: USExposure = -5.090 ∗∗ (2.504) — Observations 722.
- High-Skill: USExposure = -2.726 (1.814) — Observations 722.
- Note: Table A.8 dependent variable is change in occupational employment-to-population ratio in 2010-2021. The list of STEM occupations are from O*NET. High-skill occupations are management, business and financial occupations, professionals, and technicians. Middle-skill occupations are office and administration, sales, construction and extraction, mechanics and repairers, production, transportation and material moving. Low-skill occupations are personal services and agriculture occupations.

### Education (1990-1995 Average Share in IV)
- Below High School: USExposure = -7.109 (5.782) — Observations 722.
- High School: USExposure = -10.546 ∗∗∗ (3.703) — Observations 722.
- Some College: USExposure = -5.820 (3.873) — Observations 722.
- College and Above: USExposure = -1.555 (2.553) — Observations 722.
- Note: Table A.9 dependent variable is change in employment-to-population ratio by education levels in 2010-2021.

### Age and Gender (1990-1995 Average Share in IV)
- 16-25: USExposure = -13.546 ∗∗∗ (5.441) — Observations 722.
- 26-35: USExposure = -3.810 (3.609) — Observations 722.
- 36-45: USExposure = -7.546 ∗ (3.967) — Observations 722.
- 46-55: USExposure = -9.109 ∗∗ (3.972) — Observations 722.
- 56-65: USExposure = -4.921 (3.898) — Observations 722.
- Male: USExposure = -9.945 ∗∗∗ (4.099) — Observations 722.
- Female: USExposure = -6.534 ∗∗ (3.309) — Observations 722.
- Note: Table A.10 dependent variable is change in employment-to-population ratio by 10-year age bins or gender in 2010-2021.

### Alternative Measure: Use Maximum for AIAdopt_US_j — Overall (1990/1995/1990-1995 shares)
- 2010-2021 (1990 Share): USExposure = -3.785 ∗∗ (1.628) — Observations 722; R-squared 0.20; First-stage coefficient 0.149; First-stage F-statistic 33.2.
- 2010-2021 (1995 Share): USExposure = -2.897 ∗ (1.584) — Observations 722; R-squared 0.25; First-stage coefficient 0.165; First-stage F-statistic 29.2.
- 2010-2021 (1990-1995 Average): USExposure = -4.325 ∗∗∗ (1.740) — Observations 722; R-squared 0.16; First-stage coefficient 0.166; First-stage F-statistic 30.7.
- 1980-2010 (1990 Share): USExposure = 1.117 (2.403) — Observations 722; R-squared 0.55; First-stage coefficient 0.149; First-stage F-statistic 33.2.
- 1980-2010 (1995 Share): USExposure = -0.610 (2.746) — Observations 722; R-squared 0.55; First-stage coefficient 0.165; First-stage F-statistic 29.2.
- 1980-2010 (1990-1995 Average): USExposure = 0.370 (2.622) — Observations 722; R-squared 0.55; First-stage coefficient 0.166; First-stage F-statistic 30.7.
- Note: Table A.11 reports second stage estimates β from equation (5). USExposure_i is computed using the maximum over AI technologies for AIAdopt_US_j. Dependent variable is change in employment-to-population ratio in 1980-2010 and 2010-2021 as indicated.

### Alternative Measure: Use Maximum for AIAdopt_US_j — Broad Sector (1990 Share)
- Agriculture: USExposure = 0.461 ∗ (0.243) — Observations 722.
- Manufacturing: USExposure = -2.579 ∗ (1.472) — Observations 722.
- Construction: USExposure = 0.527 (0.551) — Observations 722.
- Low-Skill Services: USExposure = -2.667 ∗∗ (1.048) — Observations 722.
- High-Skill Services: USExposure = 0.473 (0.827) — Observations 722.
- Note: Table A.12 dependent variable is change in sectoral employment-to-population ratio in 2010-2021. USExposure_i computed using maximum over AI technologies for AIAdopt_US_j.

### Alternative Measure: Use Maximum for AIAdopt_US_j — Occupation (1990 Share)
- Non-STEM: USExposure = -3.526 ∗∗ (1.509) — Observations 722.
- STEM: USExposure = -0.259 (0.534) — Observations 722.
- Low-Skill: USExposure = -0.116 (0.495) — Observations 722.
- Middle-Skill: USExposure = -2.487 ∗ (1.327) — Observations 722.
- High-Skill: USExposure = -1.182 (0.871) — Observations 722.
- Note: Table A.13 dependent variable is change in occupational employment-to-population ratio in 2010-2021. The list of STEM occupations are from O*NET. High-skill occupations are management, business and financial occupations, professionals, and technicians. Middle-skill occupations are office and administration, sales, construction and extraction, mechanics and repairers, production, transportation and material moving. Low-skill occupations are personal services and agriculture occupations.

### Alternative Measure: Use Maximum for AIAdopt_US_j — Education (1990 Share)
- Below High School: USExposure = -1.309 (2.971) — Observations 722.
- High School: USExposure = -4.900 ∗∗ (2.050) — Observations 722.
- Some College: USExposure = -3.132 (1.903) — Observations 722.
- College and Above: USExposure = -0.277 (1.209) — Observations 722.
- R-squared by column: 0.25, 0.26, 0.25, 0.24.
- Note: Table A.14 dependent variable is change in employment-to-population ratio by education levels in 2010-2021. USExposure_i computed using maximum over AI technologies for AIAdopt_US_j.

### Alternative Measure: Use Maximum for AIAdopt_US_j — Age and Gender (1990 Share)
- 16-25: USExposure = -5.804 ∗∗ (2.871) — Observations 722.
- 26-35: USExposure = -1.493 (1.744) — Observations 722.
- 36-45: USExposure = -2.810 (2.079) — Observations 722.
- 46-55: USExposure = -3.904 ∗ (2.030) — Observations 722.
- 56-65: USExposure = -4.016 ∗ (2.145) — Observations 722.
- Male: USExposure = -4.631 ∗∗ (2.180) — Observations 722.
- Female: USExposure = -2.812 ∗ (1.628) — Observations 722.
- Note: Table A.15 dependent variable is change in employment-to-population ratio by 10-year age bins or gender in 2010-2021. USExposure_i computed using maximum over AI technologies for AIAdopt_US_j.

### Alternative End Year: 2019 as End Year — Overall (2010-2019)
- 2010-2019 (1990 Share): USExposure = -7.060 ∗∗ (3.088) — Observations 722; R-squared 0.37; First-stage coefficient 0.075; First-stage F-statistic 58.2.
- 2010-2019 (1995 Share): USExposure = -6.450 ∗∗ (2.918) — Observations 722; R-squared 0.38; First-stage coefficient 0.084; First-stage F-statistic 52.8.
- 2010-2019 (1990-1995 Average): USExposure = -8.240 ∗∗∗ (3.129) — Observations 722; R-squared 0.35; First-stage coefficient 0.086; First-stage F-statistic 57.3.
- 1980-2010 (1990 Share): USExposure = 2.217 (4.739) — Observations 722; R-squared 0.56; First-stage coefficient 0.075; First-stage F-statistic 58.2.
- 1980-2010 (1995 Share): USExposure = -1.199 (5.402) — Observations 722; R-squared 0.55; First-stage coefficient 0.084; First-stage F-statistic 52.8.
- 1980-2010 (1990-1995 Average): USExposure = 0.716 (5.075) — Observations 722; R-squared 0.55; First-stage coefficient 0.086; First-stage F-statistic 57.3.
- Note: Table A.16 reports second stage estimates β from equation (5). Dependent variable is change in employment-to-population ratio in 1980-2010 and 2010-2019 as indicated.

### Alternative End Year: 2019 as End Year — Broad Sector (1990 Share)
- Agriculture: USExposure = 1.160 ∗∗ (0.474) — Observations 722.
- Manufacturing: USExposure = -5.771 ∗∗ (2.504) — Observations 722.
- Construction: USExposure = 1.482 (1.107) — Observations 722.
- Low-Skill Services: USExposure = -4.773 ∗∗∗ (1.561) — Observations 722.
- High-Skill Services: USExposure = 0.842 (1.590) — Observations 722.
- Note: Table A.17 dependent variable is change in sectoral employment-to-population ratio in 2010-2019.

### Alternative End Year: 2019 as End Year — Occupation (1990 Share)
- Non-STEM: USExposure = -6.259 ∗∗ (2.909) — Observations 722.
- STEM: USExposure = -0.801 (0.808) — Observations 722.
- Low-Skill: USExposure = 0.550 (1.098) — Observations 722.
- Middle-Skill: USExposure = -5.114 ∗∗ (2.347) — Observations 722.
- High-Skill: USExposure = -2.496 (1.529) — Observations 722.
- Note: Table A.18 dependent variable is change in occupational employment-to-population ratio in 2010-2019. The list of STEM occupations are from O*NET. High-skill occupations are management, business and financial occupations, professionals, and technicians. Middle-skill occupations are office and administration, sales, construction and extraction, mechanics and repairers, production, transportation and material moving. Low-skill occupations are personal services and agriculture occupations.

*Source: wpiea2024199-print-pdf - 2021. The list of STEM occupations are from O*NET. High-skill occupations are management, business and financial occupations, professionals, and technicians.*

### 2019.  The list of STEM occupations are from O*NET. High-skill occupations are management, business

### wpiea2024199-print-pdf - 2019.  The list of STEM occupations are from O*NET. High-skill occupations are management, business

### Occupation (2010–2019; Table A.19 / 2010–2021; Table A.23)
- 2010–2019: Effect of AI on employment-to-population ratio by occupation (second-stage β, 1990 local employment share to compute IVEUExposure_i)
  - Non-STEM: -1.662 (1.976) Observations 722
  - STEM: 0.146 (1.166) Observations 722
  - Low-Skill: -3.313 (2.059) Observations 722
  - Middle-Skill: -2.652 (2.613) Observations 722
  - High-Skill: -0.241 (1.598) Observations 722
- 2010–2021 (USExposure uses 2005 local employment share; Table A.23):
  - Non-STEM: -6.623 ∗∗ (2.725) Observations 722
  - STEM: -0.486 (0.992) Observations 722
  - Low-Skill: -0.217 (0.926) Observations 722
  - Middle-Skill: -4.672 ∗ (2.516) Observations 722
  - High-Skill: -2.220 (1.545) Observations 722
- Notes: The list of STEM occupations are from O*NET. High-skill occupations are management, business and financial occupations, professionals, and technicians. Middle-skill occupations are office and administration, sales, construction and extraction, mechanics and repairers, production, transportation and material moving. Low-skill occupations are personal services and agriculture occupations. All regressions weighted by 2010 commuting zone population. Robust standard errors in parentheses and clustered at the state level.

### Education (2010–2019; Table A.19 / 2010–2021; Table A.24)
- 2010–2019 (second-stage β, 1990 local employment share to compute IVEUExposure_i; dependent variable: change in employment-to-population ratio by education in 2010–2019; Table A.19)
  - Below High School: USExposure -0.320 (5.776) Observations 722
  - High School: USExposure -7.005 ∗∗ (3.416) Observations 722
  - Some College: USExposure -7.569 ∗∗ (3.726) Observations 722
  - College and Above: USExposure -2.139 (2.413) Observations 722
- 2010–2021 (USExposure uses 2005 local employment share; Table A.24)
  - Below High School: USExposure -2.459 (5.484) Observations 722
  - High School: USExposure -9.203 ∗∗ (3.920) Observations 722
  - Some College: USExposure -5.884 (3.693) Observations 722
  - College and Above: USExposure -0.520 (2.260) Observations 722
- Notes: All regressions weighted by 2010 commuting zone population. Robust standard errors in parentheses and clustered at the state level.

### Age and Gender (2010–2019; Table A.20 / 2010–2021; Table A.25)
- 2010–2019 (second-stage β, 1990 local employment share to compute IVEUExposure_i; dependent variable: change in employment-to-population ratio by 10-year age bins or gender in 2010–2019; Table A.20)
  - Age 16-25: USExposure -11.104 ∗∗ (5.448) Observations 722
  - Age 26-35: USExposure -4.588 (3.017) Observations 722
  - Age 36-45: USExposure -3.180 (3.977) Observations 722
  - Age 46-55: USExposure -7.393 ∗ (4.131) Observations 722
  - Age 56-65: USExposure -7.938 ∗∗ (3.914) Observations 722
  - Male: USExposure -7.685 ∗∗ (3.859) Observations 722
  - Female: USExposure -6.242 ∗∗ (3.171) Observations 722
- 2010–2021 (USExposure uses 2005 local employment share; Table A.25)
  - Age 16-25: USExposure -10.903 ∗ (5.653) Observations 722
  - Age 26-35: USExposure -2.804 (3.220) Observations 722
  - Age 36-45: USExposure -5.278 (3.599) Observations 722
  - Age 46-55: USExposure -7.332 ∗∗ (3.336) Observations 722
  - Age 56-65: USExposure -7.543 ∗∗ (3.843) Observations 722
  - Male: USExposure -8.699 ∗∗ (4.134) Observations 722
  - Female: USExposure -5.283 ∗ (2.795) Observations 722
- Notes: All regressions weighted by 2010 commuting zone population. Robust standard errors in parentheses and clustered at the state level.

### Alternative Share for USExposure (USExposure uses 2005 local employment share; Section F)
- F.1 Overall employment-to-population ratio (Table A.21; dependent variable: change in employment-to-population ratio in 1980-2010 for columns (4)-(6) and 2010-2021 for columns (1)-(3))
  - 2010-2021, 1990 Share (column 1): USExposure -7.109 ∗∗ (2.899) Observations 722 R-squared 0.27 First-stage coefficient 0.080 First-stage F-statistic 57.5
  - 2010-2021, 1995 Share (column 2): USExposure -5.160 ∗ (2.712) Observations 722 R-squared 0.30 First-stage coefficient 0.092 First-stage F-statistic 9.4
  - 2010-2021, 1990-1995 Average (column 3): USExposure -7.840 ∗∗∗ (2.946) Observations 722 R-squared 0.25 First-stage coefficient 0.092 First-stage F-statistic 6.6
  - 1980-2010, 1990 Share (column 4): USExposure 2.098 (4.479) Observations 722 R-squared 0.55 First-stage coefficient 0.080 First-stage F-statistic 57.5
  - 1980-2010, 1995 Share (column 5): USExposure -1.086 (4.892) Observations 722 R-squared 0.55 First-stage coefficient 0.092 First-stage F-statistic 9.4
  - 1980-2010, 1990-1995 Average (column 6): USExposure 0.671 (4.751) Observations 722 R-squared 0.55 First-stage coefficient 0.092 First-stage F-statistic 6.6
- F.2 Broad sector (Table A.22; dependent variable: change in sectoral employment-to-population ratio in 2010-2021)
  - Agriculture: USExposure 0.865 ∗∗ (0.429) Observations 722
  - Manufacturing: USExposure -4.845 ∗ (2.629) Observations 722
  - Construction: USExposure 0.991 (1.030) Observations 722
  - Low-Skill Services: USExposure -5.009 ∗∗ (2.028) Observations 722
  - High-Skill Services: USExposure 0.889 (1.573) Observations 722
  - Notes: Manufacturing includes manufacturing and mining. Low-skill services are wholesale trade, retail trade, utilities, transportation, information, real estate, administrative support and waste management, arts and entertainment, accommodation and food services, and other services. High-skill services are finance and insurance, professional scientific and technical services, management of companies and enterprises, education, health, and social assistance.
- F.3 Occupation (Table A.23; dependent variable: change in occupational employment-to-population ratio in 2010-2021)
  - Non-STEM: USExposure -6.623 ∗∗ (2.725) Observations 722
  - STEM: USExposure -0.486 (0.992) Observations 722
  - Low-Skill: USExposure -0.217 (0.926) Observations 722
  - Middle-Skill: USExposure -4.672 ∗ (2.516) Observations 722
  - High-Skill: USExposure -2.220 (1.545) Observations 722

*The Labor Market Impact of Artificial Intelligence: Evidence from US Regions Working Paper No. WP/2024/199*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2024/english/wpiea2024199-print-pdf.pdf_