## Section 2 — AI Exposure and Adjusting for Potential Complementarity (wpiea2023216-print-pdf)

## Source details

**Canonical URL:** [Section 2 — AI Exposure and Adjusting for Potential Complementarity (wpiea2023216-print-pdf)](https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023216-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2023/english/wpiea2023216-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2023/english/wpiea2023216-print-pdf.pdf.json)

---

### Motivation and conceptual framework
- Exposure: potential for AI integration into each occupation based on tasks and skills; agnostic to whether AI complements or substitutes labor.
- Complementarity: driven by social, legal, and technical factors independent of exposure; occupations mapped into four quadrants by Exposure (low/high) and Complementarity (low/high):
  - High Exposure & High Complementarity: more likely productivity gains conditional on access to infrastructure and appropriate skills.
  - High Exposure & Low Complementarity: greater risk of substitution and long-lasting falls in labor demand.
  - Low Exposure quadrants: complementarity matters less for labor demand but can still influence AI integration.
- Higher complementarity does not guarantee no risk: workers in highly complementary occupations lacking required skills may still experience lower employment opportunities and wages.

### Complementarity-adjusted exposure: definition, formula, and calibration
- Complementarity-adjusted exposure (C-AIOE) for occupation i:
  - C-AIOE_i = AIOE_i * (1 - (θ_i - θ_MIN))
- Definitions and calibration details:
  - θ_i: measure of potential complementarity of AI for occupation i.
  - θ is bounded between 0 and 1 by construction (mean of six components divided by 100).
  - θ_MIN (lowest θ across occupations) = 0.31 (Hand Cutters and Trimmers, US SOC 2010 code 51-9031).
  - Highest θ observed = 0.78 (Oral and Maxillofacial Surgeons, US SOC 2010 code 29-1022).
  - Median θ = 0.58.
- Interpretation:
  - The adjustment uses θ_MIN so AIOE and C-AIOE coincide for the occupation with the lowest θ.
  - A higher C-AIOE index implies a greater risk of replacement at the occupation level.
  - Both AIOE and C-AIOE are relative measures, not absolute counts of exposed workers.

### AIOE methodology (Felten et al. 2021): exact formula and scope
- AIOE for occupation i:
  - AIOE_i = (sum_{j=1 to 52} A_kj · L_ji · I_ji) / (sum_{j=1 to 52} L_ji · I_ji)
    - where k represents the AI application, j indexes occupational ability, i stands for the occupation
    - A_kj = exposure to AI of ability j (sum of relatedness scores of the ability with each of the 10 AI applications)
    - L_ji = ability j prevalence in occupation i
    - I_ji = ability j importance in occupation i
- Scope:
  - Connects 10 AI applications with 52 occupational abilities via a crowd-sourced relatedness matrix; occupations weighted by ability prevalence and importance from O*NET.
  - Focuses on “narrow” AI (software that finds patterns in data and makes predictions); encompasses Generative AI but not “general” AI.

### Construction of θ: components, data sources, and rationale
- Data source: O*NET repository (work contexts and job zones) mapped to US SOC 2010 to align with Felten et al. (2021).
- Inputs and transformations:
  - From 57 O*NET work contexts, 11 were selected and aggregated into 5 groups; each work context value ranges from 0 to 100.
  - Job zones: ordered categorical values 1 to 5 converted to 20–100 by multiplying by 20.
- Six components (θ is arithmetic mean of these six components, then divided by 100):
  1. Communication: Face-to-Face; Public Speaking.
  2. Responsibility: Responsibility for outcomes; Responsibility for others’ health.
  3. Physical Conditions: Exposure to Outdoors Environments; Physical Proximity to Others.
  4. Criticality: Consequence of Errors; Freedom of Decisions; Frequency of Decisions.
  5. Routine: Degree of Automation; Unstructured vs Structured Work (automation score inverted).
  6. Skills: Job Zones (captures education, on-the-job training, and professional experience).
- Rationale: components capture dimensions affecting likelihood of AI replacing human activities or being adopted in supervised/complementary manner; averaging components reflects uncertainty about relative importance.

### Main empirical patterns from AIOE vs θ and effects of adjustment
- Occupational-level illustrations:
  - Lawyers and judges: high AIOE and high θ.
  - Telemarketers: similar AIOE to lawyers but low θ.
  - Surgeons: low AIOE in baseline but highest θ.
- Aggregated 1-digit ISCO-08 patterns:
  - High-skill occupations (e.g., managers) display greater complementarity than low-skill counterparts despite similar AI exposure as clerical support workers.
  - Significant within-group variability of θ among craft and trade workers, services and sales, and plant and machine operators; smaller variance among clerical/support and skilled agricultural/fishery workers.
- Effects of complementarity adjustment:
  - Adjustment reduces average differences in AI exposure among broad occupation groups and increases within-occupation variance.
  - Managers and professionals (previously high under AIOE) experience lower C-AIOE after adjustment.
  - Clerical jobs (previously high AIOE) demonstrate the highest C-AIOE after adjustment.
  - Agricultural workers rank relatively low on both AIOE and C-AIOE.

### Robustness checks summary (leave-one-out, PCA, correlations)
- Leave-one-out correlations of θ with AIOE (significance notation: *** p<0.01, ** p<0.05, * p<0.1):
  - θ full set: 0.170*** with AIOE; 0.179*** with AIOE:LLM
  - W/o communication: 0.059* with AIOE; 0.057 with AIOE:LLM
  - W/o responsibility: 0.351*** with AIOE; 0.369*** with AIOE:LLM
  - W/o physical condition: 0.414*** with AIOE; 0.421*** with AIOE:LLM
  - W/o criticality: 0.200*** with AIOE; 0.218*** with AIOE:LLM
  - W/o routine: 0.132*** with AIOE; 0.134*** with AIOE:LLM
  - W/o skill: -0.209*** with AIOE; -0.199*** with AIOE:LLM
- Pairwise component correlations (examples, significance as reported):
  - Communication vs Skills: 0.598***
  - Responsibility vs Physical condition: 0.510***
  - Physical condition vs Skills: -0.352***
  - Routine vs Skills: 0.476***
  - Communication vs Responsibility: 0.083**
- PCA:
  - The first two principal components explain about 65 percent of total variation across the 6 components.
  - Loadings emphasize:
    - criticality, responsibility, and physical condition for pilots and surgeons;
    - communication, skills, and routine for economists.
- Interpretation:
  - Components capture distinct facets of complementarity; “Skills”, “Responsibility”, “Communication”, and “Physical condition” are particularly influential. Omitting “skills” notably changes the sign of correlation with AIOE.

### Comparison with alternative AI exposure measures
- Baseline AIOE positively correlates with other measures (e.g., Webb (2020); Felten et al. (2023) LLM-only).
- Briggs and Kodnani (2023) replication shows a negative relationship with baseline; replication performed using documented methodology without original code—interpret with caution.
- Excluding Briggs and Kodnani replication, θ shows a concave hump-shaped relationship with exposure across measures.
- Conclusion: baseline AIOE-based insights largely carry over to most alternative measures.

### Detailed tabulations (selected exact figures from Table A.6)
- Employment shares:
  - Empl. Share for AIOE ≥50th percentile: 54.60
  - Empl. Share for AIOE ≥75th percentile: 26.04
  - Empl. Share for AIOE ≥90th percentile: 6.89
  - Empl. Share for θ ≥50th percentile: 40.24
  - Empl. Share for θ ≥75th percentile: 20.59
  - Empl. Share for θ ≥90th percentile: 8.29
  - Empl. Share for C-AIOE ≥50th percentile: 57.62
  - Empl. Share for C-AIOE ≥75th percentile: 37.98
  - Empl. Share for C-AIOE ≥90th percentile: 17.24
- Task composition (percent cognitive / routine / manual for selected buckets):
  - Cognitive:
    - AIOE ≥50th: 60.78
    - AIOE ≥75th: 69.90
    - AIOE ≥90th: 85.71
    - C-AIOE ≥50th: 64.84
    - C-AIOE ≥90th: 86.93
  - Routine:
    - AIOE ≥50th: 36.51
    - AIOE ≥75th: 29.88
    - AIOE ≥90th: 14.29
    - C-AIOE ≥50th: 27.33
    - C-AIOE ≥90th: 11.66
  - Manual (examples, small magnitudes):
    - AIOE ≥50th: 2.70
    - θ ≥50th: 12.52
    - C-AIOE ≥50th: 7.62
- Avg. RTI (employment-weighted mean of ln(RTI)):
  - AIOE ≥50th: -1.36
  - AIOE ≥75th: -2.01
  - AIOE ≥90th: -2.10
  - θ ≥50th: -1.82
  - θ ≥90th: -3.63
  - C-AIOE ≥50th: -0.28
  - C-AIOE ≤50th: 1.25
- Interpretation:
  - High-AIOE jobs skew cognitive, especially at upper percentiles, but C-AIOE shifts top-exposure roles toward routine occupations (e.g., routine clerical).
  - RTI relationships invert when moving from AIOE/θ to C-AIOE.

### Country-specific data, mapping, and aggregate cross-country results
- Microdata and sample:
  - Worker-level microdata from 6 countries (labor force and household surveys); sample restricted to individuals aged 16 to 64 who are employed in the reference period.
  - Education grouped into four categories (US-equivalent labels); India lacks exact category for higher education or incomplete college.
  - Earnings: gross (pre-tax) from main job; monthly for most countries, gross weekly for the UK, gross annual for the US.
  - AIOE and C-AIOE converted from US SOC 2010 to ISCO-08 using BLS crosswalk; non one-to-one mappings use simple averages.
- Aggregate exposure and country rankings (selected exact statements and figures):
  - Cumulative employment share at AIOE levels (examples):
    - UK: about 25 percent of employment below the lowest 40th percentile of AIOE (value of 0.4 on the x-axis).
    - Brazil: 50 percent below that threshold.
    - India: 70 percent below that threshold.
  - Country rankings:
    - UK: highest aggregate exposure; “40 percent of employment concentrated above the 80th percentile of AIOE.”
    - US: “almost 30 percent of employment concentrated above the 80th percentile of AIOE.”
    - Brazil, Colombia, South Africa: less than 15 percent of employment above the 80th percentile of AIOE.
    - India: lowest exposure due to sizable agricultural employment.
- Effect of complementarity adjustment:
  - Adjustment substantially diminishes previously observed differences in AI exposure between countries.
  - UK remains highest in complementarity-adjusted exposure, but the gap across AEs and most EMs is almost fully closed.
  - Exception: India, where the adjustment has only a marginal impact because of predominance of agricultural workers.

### Employment shares by exposure/complementarity quadrants and occupation composition
- Quadrant shares (above/below median AIOE and θ):
  - UK: 51.9 percent of workers in highly complementary occupations; 32 percent in high-exposure but low-complementarity occupations.
  - US: 49.8 percent in highly complementary occupations; 29.7 percent in high-exposure but low-complementarity occupations.
  - Brazil, Colombia, South Africa: nearly 40 percent of workers in high-exposure occupations; ~20 percent of total employment in high-exposure & high-complementarity occupations and ~20 percent in high-exposure & low-complementarity occupations.
  - India: 26 percent of workers in high-exposure occupations, divided into 14 percent in high-complementarity and 12 percent in low-complementarity.
- Occupation composition drivers:
  - UK: nearly 30 percent of workers in professional occupations (drives high aggregate exposure); US just above 15 percent in professional occupations.
  - Managers: US 14.4 percent; UK 10.7 percent (managers exhibit highest degree of complementarity).
  - EM patterns: lower AI exposure driven by large shares in elementary occupations; India: over 30 percent employment in agriculture.
  - Developing economies have smaller proportions of workers in occupations with potential for complementarity:
    - Professionals: less than 10 percent in India and South Africa.
    - Managers: less than 5 percent in Brazil.
    - Clerical workers (high exposure, low complementarity) share under 10 percent in Brazil, Colombia, South Africa and less than 5 percent in India.

### Demographic patterns: gender, education, age (selected exact figures)
- Gender:
  - In five out of six countries, women face higher AI exposure than men.
    - US: 68 percent of women in high exposure occupations vs 51 percent of men.
    - Brazil: 52 percent of women vs 32 percent of men.
    - India exception: 24 percent of women in high-exposure occupations vs 28 percent for men.
  - Women have larger shares in professional occupations (second in potential complementarity) but are also more represented in clerical jobs (high exposure, low complementarity) in the UK, US, Brazil, and South Africa.
- Education:
  - College-educated workers are more exposed:
    - Approximately 90 percent of college-educated workers across most countries are in occupations with high exposure.
    - Workers without a high school diploma: in most countries, less than 20 percent are in high-exposure occupations.
    - UK exception: 40 percent of workers with middle school education or less are in high-exposure occupations.
  - Complementarity concentrated among college-educated:
    - UK example: share in high exposure & low complementarity across education levels differs by less than 10 percentage points (26 percent for middle school or below vs 36 percent for college or higher).
    - UK example for potential benefit: 17 percent of middle school or below in high-exposure & high-complementarity vs over 50 percent of college-educated.
- Age:
  - No straightforward association between age and AI exposure; youngest workers tend to have lower AI exposure than prime-age workers.
  - Conditional on being in high-exposure occupations, younger workers are less likely to be in jobs with high complementarity and thus more susceptible to negative impacts.

### Earnings, inequality, and transition implications (Section 4.4)
- AI exposure across earnings distribution:
  - Positive association between earnings and share of employment in high-exposure occupations in all countries.
  - US and UK: higher proportion of highly exposed workers across entire earnings distribution.
  - UK: almost all workers in the top decile are in highly exposed occupations.
  - India: AI exposure exceptionally low at lower end and increases with income.
- Distribution of high-exposure workers by complementarity:
  - High-exposure & low-complementarity workers are more equally distributed across the income distribution.
  - Employment in high-complementarity occupations concentrated in top deciles in all countries; concentration more pronounced in EMs and more gradual in AEs.
- Implications for inequality and worker transitions:
  - C-AIOE implies potential non-trivial wage changes between jobs and within jobs due to changing demand.
  - Final effects on income inequality depend on workers’ ability to transition (job-to-job switches or generational turnover).
  - Analysis does not account for creation of new tasks; historical evidence suggests new technologies create new tasks and jobs, reducing country-level exposure over the long term when considering occupational mobility.
- Cross-cutting observations:
  - AEs have larger shares in high-skill, highly exposed jobs that can also benefit from AI; after complementarity adjustment, some cross-country disparities are reduced.
  - AEs show more polarization: more employment in exposed occupations at both ends of complementarity spectrum, implying both greater substitution risk and greater productivity potential.

### Robustness and sensitivity visual/appendix materials (Annex A and B highlights)
- Annex A:
  - Lists of 57 O*NET work contexts used to form θ; component-level inspection shows no single occupation group systematically scores highest on all components.
  - Pairwise correlations and component–AIOE correlations reported (see correlation values above).
  - Leave-one-out and PCA analyses demonstrate multifaceted nature of complementarity.
- Annex B:
  - Additional country-level distributional analyses by gender, age, education, earnings quintile.
  - Sensitivity of C-AIOE across specifications with weight 0, weight 0.5, and weight 1 (AIOE/C-AIOE normalized between 0 and 1) affects cumulative employment distributions by country.

*Source: IMF working paper (wpiea2023216-print-pdf), Section 2, Section 2.5, Section 4.4, Annexes A and B.*

### Section  3  describes  the  country-specific  data  sources  used  for  the  analysis.   Sections  4-5

### Section 2 — AI Exposure and Adjusting for Potential Complementarity

### Motivation and conceptual framework
- Exposure reflects the potential for AI to be integrated into each occupation based on the tasks and skills that characterize each job; exposure is agnostic on the likelihood of AI complementing or replacing labor.
- Complementarity is conceived as driven by a set of factors – social, legal, technical – that are independent of exposure itself. The framework maps occupations into four quadrants by Exposure (low/high) and Complementarity (low/high):
  - High Exposure & High Complementarity: more likely productivity gains conditional on access to infrastructure and appropriate skills.
  - High Exposure & Low Complementarity: greater risk of substitution and long-lasting falls in labor demand (analogous to routine-biased automation).
  - Low Exposure quadrants: complementarity matters less for labor demand but can still influence how AI is integrated.
- Higher complementarity does not guarantee no risk for individual workers: workers in highly complementary occupations lacking required skills may face lower employment opportunities and wages.

### Complementarity-adjusted exposure (definition and formula)
- For a given occupation i, let θ_i be a measure of potential complementarity of AI.
- The baseline AIOE exposure is adjusted to form C-AIOE as follows:
  - C-AIOE_i = AIOE_i * (1 - (θ_i - θ_MIN))
- θ_MIN is the minimum value of θ_i across all occupations; the adjustment makes the complementarity measure relative so that AIOE and C-AIOE coincide for the occupation with the lowest θ.
- A higher value of the C-AIOE index implies a greater risk of replacement at the occupation level.
- Both AIOE and the proposed C-AIOE are relative measures and are not intended to provide absolute counts of exposed workers.

### AIOE methodology (Felten et al. 2021) — brief overview and exact formula
- AIOE connects 10 AI applications with 52 occupational abilities via a crowd-sourced relatedness matrix; occupations are weighted by ability prevalence and importance from O*NET.
- AIOE for occupation i is calculated as:
  - AIOE_i = (sum_{j=1 to 52} A_kj · L_ji · I_ji) / (sum_{j=1 to 52} L_ji · I_ji)
    - where k represents the AI application, j indexes occupational ability, i stands for the occupation
    - A_kj is exposure to AI of ability j (sum of relatedness scores of the ability with each of the 10 AI applications)
    - L_ji is ability j prevalence in occupation i; I_ji is ability j importance in occupation i
- The measure focuses on “narrow” AI (software that finds patterns in data and makes predictions); it encompasses Generative AI but not “general” AI.

### Construction of the complementarity measure θ (data sources and components)
- Data source: O*NET repository (work contexts and job zones) mapped to US SOC 2010 classification to align with Felten et al. (2021).
- Selected inputs:
  - Work contexts: out of 57 contexts, 11 were selected and aggregated into 5 groups following O*NET grouping; each work context has a value between 0 to 100.
  - Job zones: ordered categorical values 1 to 5; these are multiplied by 20 to convert into values from 20 to 100.
- The 11 contexts and job zone are grouped into six components; θ is the arithmetic mean of these six components and then divided by 100 to bound θ between 0 and 1.
- The six components:
  1. Communication: i) Face-to-Face, ii) Public Speaking.
  2. Responsibility: i) Responsibility for outcomes, ii) Responsibility for others’ health.
  3. Physical Conditions: i) Exposure to Outdoors Environments, ii) Physical Proximity to Others.
  4. Criticality: i) Consequence of Errors, ii) Freedom of Decisions, iii) Frequency of Decisions.
  5. Routine: i) Degree of Automation, ii) Unstructured vs Structured Work (automation score inverted so occupations with low degree of automation have higher values).
  6. Skills: Job Zones (captures education, on-the-job training, and professional experience; occupations with longer professional development may better integrate AI complementary skills).
- Rationale: Components capture dimensions affecting likelihood of AI replacing human activities or being adopted in supervised/complementary manner; averaging components is a cautious stance given uncertainty about the relative importance of each factor.

### Key numeric values and calibration details
- Each O*NET work context value ranges from 0 to 100.
- Job zones are converted from 1–5 into 20–100 by multiplying by 20.
- θ is bounded between 0 and 1 by dividing the component mean by 100.
- θ_MIN (lowest θ across occupations) = 0.31 (corresponds to Hand Cutters and Trimmers, US SOC 2010 code 51-9031).
- Highest θ observed = 0.78 (Oral and Maxillofacial Surgeons, US SOC 2010 code 29-1022).
- Median θ = 0.58.
- The θ_MIN value is used in the C-AIOE adjustment formula to provide a relative interpretation.

### Main empirical patterns and sensitivity of adjustment
- Plotting AIOE against θ (complementarity) and segmenting by medians illustrates occupational interactions:
  - Example contrasts: lawyers and judges (high AIOE, high θ) vs telemarketers (similar AIOE, low θ).
  - Surgeons: categorized as low-AI exposure in AIOE but have the highest potential complementarity (highest θ).
- Aggregated patterns by 1-digit ISCO-08 groups (distributional findings):
  - High-skill occupations (e.g., managers) display greater complementarity than low-skill counterparts despite similar exposure to AI as clerical support workers.
  - Significant within-group variability of θ exists among craft and trade workers, services and sales, and plant and machine operators; variance is smaller among clerical/support workers and skilled agricultural and fishery workers.
- Effects of the complementarity adjustment:
  - The complementarity adjustment reduces average differences in AI exposure among broad occupation groups.
  - It increases within-occupation variance of AI exposure.
  - Managers and professionals (previously high under AIOE) experience lower C-AIOE after adjustment.
  - Clerical jobs (previously high AIOE) demonstrate the highest C-AIOE after adjustment.
  - Agricultural workers rank relatively low on both AIOE and C-AIOE.

*Source: IMF working paper (wpiea2023216-print-pdf), Section 2 and related methodological exposition.*

### 2.5    Robustness Checks on the Complementarity Adjustment

### 2.5    Robustness Checks on the Complementarity Adjustment

### Robustness and sensitivity analyses
- Performed correlation checks between individual components of complementarity (Table A.3).
- Compared each component with Felten et al. (2021)’s AIOE (Figure A.2; Table A.4).
- Assessed alignment of complementarity with other AI exposure measures from the literature (Figure A.7).
- Graphically inspected component influence (Figure A.1) and implemented a “leaving-one-out” analysis by excluding one component at a time and recomputing complementarity based on the remaining five components (Figures A.3 and A.4; Table A.5).
- Conclusion from checks: the selected components “effectively encapsulate diverse and significant factors crucial to the interaction between AI and workers,” providing a “comprehensive and multifaceted measure of complementarity.”

### Principal component analysis (PCA) on complementarity components
- PCA indicates the components are not all systematically interrelated.
- The first two principal components explain about 65 percent of the total variation in the 6 individual components of complementarity.
- PCA emphasizes the importance of:
  - criticality, responsibility, and physical condition for professions such as pilots and surgeons;
  - communication, skills, and routine for roles like economists.
- Interpretation: the work contexts and skills chosen provide a multifaceted view of potential complementarity (Figure A.5).

### Comparison with other AI exposure measures
- Baseline AIOE exhibits positive correlation with other frequently cited AI exposures, such as Webb (2020) (Figure A.6).
- Positive correlation also holds with Felten et al. (2023)’s measure that includes only exposure to Large Language Models (LLMs).
- Exception: Briggs and Kodnani (2023) shows a negative relationship with the baseline measure (caveat: their AI exposure measure was replicated using only their text-described methodology without original data or exact coding procedure).
- Overall inference: findings derived from the baseline AI exposure measure should be robust and applicable to most alternative measures.

### C-AIOE and routine-biased technical change
- Table A.6 reports employment shares in occupations categorized as cognitive, routine, and manual above thresholds of AIOE, θ, and C-AIOE.
- Main takeaways:
  - High-AIOE occupations are tilted towards cognitive jobs, but “it is only at uppermost quantiles of exposure that almost all jobs are cognitive.”
  - High-complementarity occupations have a greater association with cognitive jobs than with routine and manual ones.
  - The C-AIOE measure shows much lower overlap with cognitive occupations and larger overlap with routine and manual jobs.
- Alignment with main results: professions and managerial positions mostly fall into cognitive categories; Table A.6’s tabulations are consistent with the paper’s main findings on the effect of the complementarity adjustment.

### Country-specific data and mapping
- Worker-level microdata used for 6 countries from labor force and household surveys (Table 1 lists surveys, years, and ISCO-08 granularity).
- Sample restricted to individuals aged 16 to 64 who are employed in the reference period.
- Educational attainment grouped into four categories (US-equivalent labels): middle school and below; high school; higher education or incomplete college; college degree or higher. (Note: India lacks an exact survey category corresponding to higher education or incomplete college.)
- Earnings: gross (pre-tax) income from main job; monthly for most countries, gross weekly for the UK, gross annual for the US.
- AIOE and C-AIOE converted from US SOC 2010 to ISCO-08 using a Bureau of Labor Statistics crosswalk; non one-to-one mappings use simple averages.

### Aggregate cross-country results (AIOE and C-AIOE)
- Cumulative employment share at different levels of AIOE:
  - Example: the UK has about 25 percent of employment below the lowest 40th percentile of AIOE (value of 0.4 on the x-axis); Brazil has 50 percent; India 70 percent.
- Country rankings and aggregate exposure:
  - The UK: highest aggregate exposure; “40 percent of employment concentrated above the 80th percentile of AIOE.”
  - The US: “almost 30 percent of employment concentrated above the 80th percentile of AIOE.”
  - Brazil, Colombia, South Africa: less than 15 percent of employment above the 80th percentile of AIOE.
  - India: lowest exposure due to sizable agricultural employment.
- Effect of complementarity adjustment (C-AIOE):
  - Adjustment substantially diminishes previously observed differences in AI exposure between countries (Figure 4b).
  - The UK remains highest in complementarity-adjusted exposure, but the gap across AEs and most EMs is almost fully closed.
  - Exception: India, where the adjustment has only a marginal impact because of the predominance of workers in the agricultural sector.

### Employment shares by exposure and complementarity quadrants
- Occupations categorized by above/below median AIOE and complementarity into four quadrants; focus on:
  - High Exposure and Low Complementarity (most vulnerable); and
  - High Exposure and High Complementarity (most likely to benefit).
- Country-level shares (Figure 5):
  - UK: 51.9 percent of workers engaged in highly complementary occupations; 32 percent in high-exposure but low-complementarity occupations.
  - US: 49.8 percent in highly complementary occupations; 29.7 percent in high-exposure but low-complementarity occupations.
  - Brazil, Colombia, South Africa: nearly 40 percent of workers in high-exposure occupations; ~20 percent of total employment in high-exposure & high-complementarity occupations and ~20 percent in high-exposure & low-complementarity occupations.
  - India: 26 percent of workers in high-exposure occupations, divided into 14 percent in high-complementarity and 12 percent in low-complementarity.

### Occupation composition insights
- Professional and managerial occupations drive the UK’s high aggregate exposure:
  - Nearly 30 percent of workers in the UK employed in professional occupations; US just above 15 percent.
  - Managers: US 14.4 percent; UK 10.7 percent (managers exhibit highest degree of complementarity).
- In EMs, lower AI exposure driven by large shares in elementary occupations (low exposure).
- India: over 30 percent employment in agriculture, contributing to low AI exposure.
- Developing economies have smaller proportions of workers in occupations with potential for complementarity:
  - Professionals: less than 10 percent in India and South Africa.
  - Managers: less than 5 percent in Brazil.
  - Clerical workers (high exposure, low complementarity) share under 10 percent in Brazil, Colombia, South Africa and less than 5 percent in India.

### Demographic patterns: gender, education, age

- Gender
  - In five out of six countries, women face higher AI exposure than men.
    - US: 68 percent of women in high exposure occupations vs 51 percent of men.
    - Brazil: 52 percent of women vs 32 percent of men.
    - India exception: 24 percent of women in high-exposure occupations vs 28 percent for men (driven by higher female share in elementary and agricultural jobs).
  - Women have larger shares in professional occupations (second in potential complementarity), suggesting higher likelihood to benefit from AI proliferation.
  - However, women are more represented in clerical jobs (high exposure, low complementarity) in the UK, US, Brazil, and South Africa, increasing vulnerability to adverse impacts.

- Education
  - College-educated workers are more exposed to AI than lower-educated workers:
    - Approximately 90 percent of college-educated workers across most countries are in occupations with high exposure (primarily professional roles).
    - Workers without a high school diploma: in most countries, less than 20 percent are in high-exposure occupations.
    - UK exception: 40 percent of workers with middle school education or less are in high-exposure occupations.
  - Complementarity potential concentrated among college-educated workers:
    - Example (UK): difference in share in high exposure & low complementarity across education levels is less than 10 percentage points (26 percent for middle school or below vs 36 percent for college or higher).
    - Potential to benefit shows larger disparities: 17 percent of middle school or below are in high-exposure & high-complementarity occupations vs over 50 percent of college-educated workers.

- Age
  - No straightforward association between age and AI exposure; patterns intertwined with country-specific trends in educational attainment and female labor force participation.
  - General observation: youngest workers tend to have lower AI exposure than prime-age workers.
  - Conditional on being in high-exposure occupations, younger workers are less likely to be in jobs with high complementarity and thus more susceptible to potential negative impacts from widespread AI adoption.

*Source: 2.5    Robustness Checks on the Complementarity Adjustment, wpiea2023216-print-pdf*

### 4.4    Earnings

### 4.4    Earnings

### AI exposure across the earnings distribution
- A positive association between earnings and share of employment in high-exposure occupations (occupations above the AIOE median) emerges in all countries.
- The US and UK possess a higher proportion of highly exposed workers across the entire earnings distribution.
- In the UK, almost all workers in the top decile are in highly exposed occupations.
- In Brazil, Colombia, and South Africa, AI exposure patterns across the earnings distribution are relatively similar, with high-income workers exhibiting greater exposure, especially noticeable in Colombia.
- In India, AI exposure is exceptionally low at the lower end of the distribution and progressively increases with income.

### Distribution of high-exposure workers by complementarity
- There is a more equal distribution of workers with high AI exposure and low complementarity across the income distribution, indicating risks from widespread AI adoption may be broadly evenly distributed across the earnings distribution.
- Employment in high-complementarity occupations is concentrated in the top deciles of the earnings distribution in all countries in the sample.
  - This concentration is more pronounced in Emerging Markets (EMs) and more gradual in Advanced Economies (AEs).
  - In the UK, the top four deciles are fairly levelled in terms of employment in high-complementarity occupations.

### Implications for inequality and worker transitions
- The C-AIOE measure suggests the possibility of non-trivial changes in wages between jobs (due to shifts in demand for different occupations) and within jobs (based on workers’ skills).
- The ultimate effect on income inequality will depend on workers’ ability to transition from jobs with shrinking demand to jobs experiencing growing demand.
  - Reallocation could occur through job-to-job switches or through generational turnover with new cohorts entering growing jobs in greater proportions.
- Even with complementarity adjustment, the analysis does not account for the creation of new tasks; historical evidence shows new technologies have also created new tasks and new jobs, which would reduce country-level AI exposure over the long term when considering occupational mobility.

### Cross-cutting observations relevant to earnings outcomes
- AEs display a larger share of employment in high-skill jobs that are highly exposed to AI; these occupations can also greatly benefit from AI, reducing some cross-country disparities when complementarity is considered.
- AEs have more employment than EMs in exposed occupations at both ends of the complementarity spectrum, suggesting a more polarized impact of AI on labor markets in AEs: greater risk of labor substitution but also greater potential productivity benefits.
- The potential for both negative and positive outcomes associated with AI is distributed across different demographic and income groups within and across countries in complex patterns, challenging simplified narratives of AI as solely a threat to employment.

*Source: 4.4 Earnings, wpiea2023216-print-pdf*

### References

### References and Annexes (wpiea2023216-print-pdf - References)

### Major themes and scope
- Bibliographic scope
  - Citations include foundational and recent works on AI, automation, and labor markets by Acemoglu and Restrepo; Agrawal, Gans, and Goldfarb; Autor; Brynjolfsson et al.; Felten et al. (2021); Webb (2020); Eloundou et al. (2023); Briggs and Kodnani (2023); and multiple IMF, OECD, ILO, and NBER reports and working papers.
- Annex focus
  - Annex A: Construction, inspection, and robustness analysis of a complementarity index (θ) combining six components derived from 57 O*NET work contexts.
  - Annex B: Additional labor market exposure analysis by country, gender, age, education, and earnings quintile, and sensitivity of complementarity-adjusted exposure (C-AIOE).

### A. Complementarity index: construction and robustness checks
- Data inputs and scope
  - Tables A.1 and A.2 list all the 57 work contexts from O*NET used to form the θ index.
  - θ is constituted from six distinct components (referred to throughout as: Communication, Responsibility, Physical condition, Criticality, Routine, Skills).
- Component-level inspection
  - Average contributions by occupation group (Figure A.1)
    - No single occupation group systematically scores highest on all components.
    - Example: Professionals score very low in "Responsibility" and "Physical Conditions" but have a high average θ driven by a high score for "Skills".
    - Removing the "Skills" component would mainly change the ranking for "Professionals"; otherwise, top/bottom half groupings are largely preserved.
  - Pairwise correlations (Table A.3)
    - Communication vs Responsibility: 0.083** 
    - Communication vs Physical condition: -0.071* 
    - Communication vs Criticality: 0.121*** 
    - Communication vs Routine: 0.379*** 
    - Communication vs Skills: 0.598*** 
    - Responsibility vs Physical condition: 0.510*** 
    - Responsibility vs Criticality: 0.455*** 
    - Responsibility vs Routine: -0.066* 
    - Responsibility vs Skills: -0.200*** 
    - Physical condition vs Criticality: 0.363*** 
    - Physical condition vs Routine: -0.002 
    - Physical condition vs Skills: -0.352*** 
    - Criticality vs Routine: 0.203*** 
    - Criticality vs Skills: 0.127*** 
    - Routine vs Skills: 0.476*** 
    - Significance notation: *** p<0.01, ** p<0.05, * p<0.1
  - Interpretation: components capture distinct facets of complementarity rather than redundant information.
- Association with AI exposure (AIOE)
  - Correlations of each component with AIOE and AIOE:LLM (Table A.4)
    - Communication: 0.502*** with AIOE; 0.548*** with AIOE:LLM
    - Responsibility: -0.394*** with AIOE; -0.412*** with AIOE:LLM
    - Physical Condition: -0.555*** with AIOE; -0.547*** with AIOE:LLM
    - Criticality: -0.026 with AIOE; -0.058 with AIOE:LLM
    - Routine: 0.245*** with AIOE; 0.286*** with AIOE:LLM
    - Skills: 0.723*** with AIOE; 0.724*** with AIOE:LLM
    - Note: "AIOE" uses the 10 AI applications in Felten et al. (2021); "AIOE: LLM" focuses on large language models.
  - Observed patterns (Figure A.2 and discussion)
    - Skills, Communication, and Routine positively correlated with AIOE.
    - Responsibility and Physical condition negatively correlated with AIOE.
    - Criticality shows no discernible correlation with AIOE.
    - Occupations can shift quadrants depending on which θ component is emphasized (example: lawyers).
- Leave-one-out sensitivity (Figures A.3, A.4; Table A.5)
  - Correlations of θ with AIOE and AIOE:LLM under component exclusions (Table A.5)
    - θ full set: 0.170*** with AIOE; 0.179*** with AIOE:LLM
    - W/o communication: 0.059* with AIOE; 0.057 with AIOE:LLM
    - W/o responsibility: 0.351*** with AIOE; 0.369*** with AIOE:LLM
    - W/o physical condition: 0.414*** with AIOE; 0.421*** with AIOE:LLM
    - W/o criticality: 0.200*** with AIOE; 0.218*** with AIOE:LLM
    - W/o routine: 0.132*** with AIOE; 0.134*** with AIOE:LLM
    - W/o skill: -0.209*** with AIOE; -0.199*** with AIOE:LLM
    - Significance notation: *** p<0.01, ** p<0.05, * p<0.1
  - Interpretation
    - "Skills", "Responsibility", "Communication", and "Physical condition" are particularly influential; omitting "skills" notably changes the sign of the correlation.
    - Some occupations (e.g., economists) can have substantially different complementarity estimates if "responsibility" is omitted.
- Principal Component Analysis (Figure A.5)
  - The first two principal components account for about 65% of total variation.
  - Implication: complementarity variability is distributed across multiple dimensions; more than two components are needed to preserve information.
  - Variable loadings highlight:
    - "Criticality", "Responsibility", and "Physical condition" as important for occupations like pilots and surgeons.
    - "Communication", "Skills", and "Routine" as important for occupations like economists.

### A.2 Alternative AI exposure measures: comparisons and robustness
- Scatter and binned scatter comparisons (Figures A.6 and A.7)
  - Felten et al. (2021) metrics ("AIOE" and "AIOE: LLM") generally show positive association with other commonly referenced AI exposure measures (e.g., Webb (2020), Eloundou et al. (2023)).
  - Briggs and Kodnani (2023) shows a notable negative correlation in the authors' replication; the replication follows documented methodology but may lack exact coding.
  - Excluding Briggs and Kodnani replication, θ exhibits a concave hump-shaped relationship with exposure across measures.
  - Conclusion: baseline AIOE-based insights largely carry over to most alternative measures.
- Occupational employment share and task intensity by exposure buckets (Table A.6)
  - Employment share and task composition across percentiles (Table A.6):
    - Empl. Share for AIOE ≥50th percentile: 54.60
    - Empl. Share for AIOE ≥75th percentile: 26.04
    - Empl. Share for AIOE ≥90th percentile: 6.89
    - Empl. Share for θ ≥50th percentile: 40.24
    - Empl. Share for θ ≥75th percentile: 20.59
    - Empl. Share for θ ≥90th percentile: 8.29
    - Empl. Share for θ for AIOE ≥50th percentile: 24.08
    - Empl. Share for θ for AIOE ≥75th percentile: 30.53
    - Empl. Share for C-AIOE ≥50th percentile: 57.62
    - Empl. Share for C-AIOE ≥75th percentile: 37.98
    - Empl. Share for C-AIOE ≥90th percentile: 17.24
  - Task composition (percent of employment within bucket classified as Cognitive, Routine, Manual)
    - Cognitive shares (examples):
      - AIOE ≥50th: 60.78
      - AIOE ≥75th: 69.90
      - AIOE ≥90th: 85.71
      - θ ≥50th: 57.78
      - C-AIOE ≥50th: 64.84
      - C-AIOE ≥90th: 86.93
      - θ for AIOE ≥50th: 40.14
      - θ for AIOE ≥75th: 33.03
      - θ for AIOE ≥90th: 30.91
    - Routine shares (examples):
      - AIOE ≥50th: 36.51
      - AIOE ≥75th: 29.88
      - AIOE ≥90th: 14.29
      - θ ≥50th: 29.59
      - C-AIOE ≥50th: 27.33
      - C-AIOE ≥90th: 11.66
      - θ for AIOE ≥50th: 56.10
      - θ for AIOE ≥75th: 47.66
      - θ for AIOE ≥90th: 49.41
    - Manual shares (examples and small magnitudes):
      - AIOE ≥50th: 2.70
      - θ ≥50th: 12.52
      - C-AIOE ≥50th: 7.62
      - θ for AIOE ≥75th: 1.41
      - C-AIOE ≥90th: 3.72
      - θ for AIOE ≥90th: 0.28
    - Avg. RTI (employment-weighted mean of ln(RTI)):
      - AIOE ≥50th: -1.36
      - AIOE ≥75th: -2.01
      - AIOE ≥90th: -2.10
      - θ ≥50th: -1.82
      - θ ≥75th: -1.70
      - θ ≥90th: -3.63
      - θ for AIOE ≥50th: -3.18
      - θ for AIOE ≥75th: -0.23
      - θ for AIOE ≥90th: -0.31
      - C-AIOE ≥50th: -0.28
      - C-AIOE ≤50th: 1.25
  - Interpretation:
    - High-AIOE jobs (above median) are largely cognitive but not exclusively; jobs just above median include about 40 percent non-cognitive tasks.
    - Jobs surpassing the 90th percentile in AI exposure skew heavily toward cognitive tasks.
    - After adjusting for complementarity (C-AIOE), high-exposure jobs distribute more uniformly across cognitive, routine, and manual tasks; the highest-exposure roles shift toward routine occupations (e.g., routine clerical).
    - RTI analysis: higher AIOE and θ correspond to lower RTI (less routine intensity), but this relationship inverts for C-AIOE.
    - Note: job classification into cognitive/routine/manual is from Cortes et al. (2020); RTI from Autor and Dorn (2013). Employment shares reported as percentages.
- Reproducibility and caution
  - The Briggs and Kodnani (2023) measure was reproduced using their documented methodology without exact code; results with this measure should be interpreted with caution.

### B. Labor market exposure to AI: distributional analysis (additional)
- Figures and dimensions analyzed (Figures B.1–B.7)
  - Employment share across major occupation groups presented:
    - By country (Figure B.1)
    - By gender (Figure B.2)
    - By age (Figure B.3)
    - By education level (Figure B.4) — note: for India no category for "Some College or Higher Education"
    - By earnings quintile (Figure B.5)
  - Aggregate shares of AI exposure and complementarity by demographic groups (Figure B.6)
    - Plots share of employment in high-exposure occupations distinguishing high vs low complementarity (high/low defined as above/below median).
  - Sensitivity of complementarity-adjusted exposure across specifications and countries (Figure B.7)
    - Cumulative employment distributions for C-AIOE specifications with weight 0, weight 0.5, and weight 1, where AIOE or C-AIOE normalized between 0 and 1.
- Key distributional observations (textual summary)
  - Employment shares in high-exposure, high-complementarity occupations vary markedly across countries and demographic groups.
  - Gender-, age-, education-, and earnings-conditional employment share plots indicate heterogeneous exposure and complementarity patterns across groups.
  - C-AIOE specification weight materially affects cumulative employment distribution by country.

*This overlay summarizes the References section and Annexes A and B from the supplied content unit.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023216-print-pdf.pdf_