This site is a work in progress and has not been widely shared. Content may contain errors. Feedback is welcome.
This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.
Back to papers

Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality

Gmyrek, Berg, Bescond

2023ILO Working Paper Series52 citations
Exposure / measurementInterdisciplinary
LLM / Generative AIGenderDeveloping economiesAugmentation vs. substitutionGeneral automation
DOI: 10.2139/ssrn.4584219
Summary

Gmyrek, Berg, and Bescond use GPT-4 API calls to score automation potential of 3,123 ISCO-08 tasks, link these scores to ILO employment microdata and modelled estimates across 189 countries, and estimate global exposure to generative AI by country income group and gender.

Main Finding

Clerical work is the most exposed occupational group, with 24% of tasks highly exposed and 58% with medium exposure to GPT-like technology; augmentation effects (10.4-13.4% of employment across income groups) exceed automation potential (0.4% in low-income to 5.5% in high-income countries); women face more than double the automation exposure of men in high-income countries (7.8% vs 2.9% of employment).

Primary Datasets

O*NET

Other

O*NET; ISCO classification

Secondary Datasets

ILO labor force surveys

Key Methods
Sequential GPT-4 API calls to score 3,123 ISCO-08 tasks and 4,360 GPT-generated tasks on automation potential (0-1 scale); task-level embeddings and K-means clustering; aggregation to occupational means; linking to ILO labor force survey microdata and modelled employment estimates; weighted calculations by country income group
Sample Period
Cross-sectional
Geographic Coverage
International
Sample Size
436 ISCO-08 4-digit occupations, 3,123 official tasks, 4,360 GPT-generated tasks; 59 countries with microdata; global employment estimates for 189 countries covering ~3.2 billion workers
Level of Analysis
Task, Occupation, Country
Occupation Classification
ISCO-08
Industry Classification
None
Notes
ILO Working Paper 96; Non-US paper; included for international comparison [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding. [Claude classification]: Upper-bound estimate acknowledged by authors due to: (1) tech-optimism bias in GPT-4 training data, (2) high-income country context used for all scores, (3) focus on technical feasibility ignoring adoption constraints. Cascading estimation procedure from ISCO 4-digit (59 countries) to 3-digit, 2-digit, 1-digit to achieve global coverage. ~25,000 GPT-4 API calls conducted. Authors emphasize value is in understanding direction of change rather than precise estimates. Gender effects highly salient finding.