---
id: "lib-2023-08-00-generative-ai-and-jobs-a-global-analysis-of"
type: "library"
title: "Generative AI and jobs: A global analysis of potential effects on job quantity and quality (ILO Working Paper 96)"
lang: "EN"
org: "ILO"
source_type: "Paper"
date: "2023-08-01"
week: "2023-W31"
url: "https://nosov.ai/library/lib-2023-08-00-generative-ai-and-jobs-a-global-analysis-of/"
original_url: "https://www.ilo.org/sites/default/files/wcmsp5/groups/public/%40dgreports/%40inst/documents/publication/wcms_890761.pdf"
tier: "S"
tier_name: "Shapes the field"
tier_reason: "ILO Research Department working paper by Gmyrek, Berg and Bescond that set the augmentation-over-automation reading of generative AI exposure across 436 occupations and income groups; one of the most cited early estimates."
summary_basis: "full-text"
original_minutes: "60"
tags: ["occupation exposure", "clerical work", "job quality"]
---

# Generative AI and jobs: A global analysis of potential effects on job quantity and quality (ILO Working Paper 96)

Tier S (Shapes the field): ILO Research Department working paper by Gmyrek, Berg and Bescond that set the augmentation-over-automation reading of generative AI exposure across 436 occupations and income groups; one of the most cited early estimates.

Summary based on the full text.

## In short

ILO Working Paper 96 (August 2023) uses GPT-4 to score how exposed each task in the ISCO-08 occupation list is to generative AI, then maps the scores onto employment data by country income group. Only clerical work comes out highly exposed. For other occupation groups, highly exposed tasks are 1 to 4 per cent. The authors conclude the main effect is augmentation of jobs rather than their full automation, with a strong gender skew because clerical work is mostly held by women.

## Key numbers

- 24 per cent: share of clerical tasks scored highly exposed, with a further 58 percent at medium exposure
- 0.4 per cent vs 5.5 percent: share of employment potentially exposed to automation in low-income vs high-income countries
- 3.7 per cent vs 1.4 per cent: share of female vs male employment worldwide in jobs potentially automatable with generative AI
- 10.4 percent vs 13.4 percent: share of employment with augmentation potential in low-income vs high-income countries
- 8.6 per cent and 281 million: share and count of workers in the 'big unknown' group between automation and augmentation

## What it reports

- 24 per cent of clerical tasks are highly exposed and 58 percent medium; for other groups, medium-exposed tasks do not exceed 25 per cent.
- Automation exposure grows with income: 0.4 per cent of employment in low-income countries, 5.5 percent in high-income countries.
- Augmentation potential is larger: 10.4 percent of employment in low-income countries, 13.4 percent in high-income and 13.5 per cent in upper-middle-income countries (global estimate).
- 3.7 per cent of female employment worldwide is potentially automatable, against 1.4 per cent of male employment; in high-income countries 7.8 per cent vs 2.9 per cent.
- A "big unknown" group of occupations with high mean scores and high variance, 8.6 per cent of global employment and 281 million workers, could move either way.
- Infrastructure gaps, such as internet access, will limit use in lower-income countries and may widen the productivity gap.
- The authors argue job quality effects (work intensity, autonomy) may matter more than job counts.

## How they got it

GPT-4 generated 10 typical tasks for each of 436 ISCO-08 four-digit occupations and scored 3,123 tasks from 0 to 1 for automation potential, separately for high- and low-income settings. Occupations with a high mean score and low spread were classed as automation; low mean and high spread as augmentation (following Carbonero et al. 2023). A consistency test of 100 runs on 5 tasks gave SDs no larger than 0.05. Scores were mapped to labour force microdata for 59 countries and to ILO modelled global estimates.

## Limits

- Exposure is scored by a language model, not observed; the authors call it an upper-bound estimate and say the precise numbers matter less than the distribution.
- Technical exposure says nothing about adoption speed, cost or demand effects.
- Only 59 countries have four-digit occupation microdata; the global figure is modelled from coarser data.
- Written in 2023 on GPT-4 capabilities; later agentic tools may shift the task scores.

## Why it matters for companies

The paper gives a defensible starting frame for workforce planning: judge exposure task by task, and expect most roles to be reshaped rather than removed. Clerical and administrative pools are where automation pressure concentrates, and those pools skew female, so redesign and redeployment plans need a gender view. The "big unknown" group is where planning effort pays off most, because outcomes there depend on how tasks are redesigned.

Original: https://www.ilo.org/sites/default/files/wcmsp5/groups/public/%40dgreports/%40inst/documents/publication/wcms_890761.pdf