💻 computer science

The Jagged Global Economy: Frontier AI Unevenly Exposes National Economies

This paper introduces a national AI exposure metric for 141 countries, revealing that high-income nations and women face significantly higher direct exposure to frontier AI than low-income nations and men, while also uncovering a critical indirect exposure mechanism through cross-country income dependencies like remittances that necessitates tailored, non-generalizable policy responses.

Original authors: Arul Murugan, Tomás Aguirre, Abhishek Nagaraj, Rishi Bommasani

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arul Murugan, Tomás Aguirre, Abhishek Nagaraj, Rishi Bommasani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A "Jagged" Wave

Imagine a massive, uneven wave of new technology (Frontier AI) rolling across the globe. The authors argue that this wave doesn't hit every country with the same force. Instead, the wave is "jagged."

Think of the capabilities of AI like a jagged mountain range. Some peaks are very high (AI is great at writing code or analyzing legal contracts), while some valleys are very low (AI is bad at fixing a leaking pipe or harvesting crops).

Because different countries have different "landscapes" of jobs, this jagged wave hits them differently:

Rich countries (like the US or Luxembourg) have economies built on "white-collar" jobs (offices, finance, tech). These jobs sit right on the high peaks of the AI mountain. The wave crashes hard here.
Poorer countries (like Burundi or Mozambique) have economies built on farming and manual labor. These jobs sit in the low valleys where the AI wave barely reaches.

The Main Tool: Measuring "Exposure"

The researchers created a new ruler called "National AI Exposure."

Imagine you are a weather forecaster, but instead of predicting rain, you are predicting how much a country's workforce will be touched by AI. You don't need to know if the AI will actually be used tomorrow; you just measure how much of the country's work could be done by AI right now.

How they did it: They looked at the "job menu" of 141 countries. They asked: "What percentage of people in this country do jobs that AI is currently good at?"
The Result: The difference is huge. The most exposed country (Luxembourg) is 2.6 times more exposed than the least exposed country (Burundi).

Key Findings in Plain English

1. The Rich Get More "Exposed" (For Now)

There is a strong link between how rich a country is and how much AI can touch its workers.

The Analogy: Think of AI as a high-speed train. Countries with lots of office workers (white-collar jobs) are standing right next to the tracks. Countries with mostly farmers are standing in a field miles away.
The Data: High-income countries are much more exposed. Europe and Central Asia are 50% more exposed than Sub-Saharan Africa.

2. The Gender Gap: Women Are Closer to the Tracks

In 91% of the countries studied, women are more exposed to AI than men.

Why? In many places, women are concentrated in office jobs, sales, and service roles—the very jobs AI is good at. Men are often more concentrated in agriculture or manual labor, where AI is less useful.
The Exception: In countries like Pakistan or India, where women's work is still heavily focused on farming or household enterprises, men are actually more exposed because their jobs are more likely to be office-based.
The Concern: Since women often already face barriers to finding work and earn less, this "exposure" adds another layer of risk. If AI disrupts the office jobs women hold, they might have fewer safety nets.

3. The "Remittance Ripple" (The Indirect Hit)

This is the paper's most unique discovery. Sometimes, a country isn't exposed directly, but it gets hit anyway because of its neighbors.

The Analogy: Imagine a small village (Tajikistan) where many young people have moved to a big city (Russia) to work. The village relies on the money these workers send home (remittances) to survive.
The Scenario: The big city (Russia) has a high-tech office economy. AI comes in and changes how those office workers do their jobs. Even if the small village has no AI, the village is now in trouble because the money flowing from the city might dry up or change.
The Data: Tajikistan has low direct AI exposure, but because 37% of its GDP comes from money sent by workers in highly exposed countries (like Russia), its "remittance-adjusted" exposure becomes very high. The paper calls this a second-order effect.

Does This Prediction Work?

The authors tested their "Exposure Ruler" to see if it actually predicts what's happening in the real world.

They checked data from three giant AI companies: Anthropic, Microsoft, and OpenAI.
The Result: Their ruler worked perfectly. Countries with high "Exposure" scores were the ones actually using the most AI tools.
The Math: A small increase in a country's exposure score meant a massive jump in how many people were using AI chatbots there.

The Bottom Line

The paper concludes that one size does not fit all.
Policymakers in the US or Europe cannot just copy-paste their AI rules and expect them to work in Africa or Asia. The "jagged" nature of AI means that:

Rich nations face immediate disruption to their office workers.
Poor nations might face indirect economic shocks if their rich trading partners change how they work.
Women in many nations are standing closer to the AI wave than men are.

The world isn't getting hit by a uniform AI tsunami; it's getting hit by a jagged, uneven wave that requires different survival strategies for every country.

Technical Summary: The Jagged Global Economy: Frontier AI Unevenly Exposes National Economies

Problem Statement
While the labor-market effects of frontier AI are a critical concern for workers, firms, and policymakers, existing evidence is heavily skewed toward a handful of high-income economies. Frontier AI capabilities are "jagged," meaning they vary significantly across different work tasks. Concurrently, national economies diverge substantially in how they allocate human labor across occupations. Consequently, a uniform understanding of AI's impact is insufficient; there is a need to quantify how the specific composition of a nation's labor market aligns with the capabilities of frontier AI. Current policy responses calibrated to U.S. or European markets may not generalize to the diverse global landscape.

Methodology
The authors introduce a National AI Exposure Metric designed to capture heterogeneity across 141 countries. The methodology involves two primary data sources:

Occupational Exposure Estimates: The study utilizes exposure scores for the International Standard Classification of Occupations (ISCO-08) developed by Gmyrek et al. [2025a]. These scores are derived by mapping occupations to tasks and estimating the potential for AI to save time on those tasks while preserving quality. The authors aggregate these scores to the 2-digit ISCO-08 level (40 occupational categories).
International Employment Data: The study combines these exposure scores with employment statistics from the International Labour Organization (ILO). These data provide the share of the workforce employed in each 2-digit ISCO-08 category for 141 countries.

The National AI Exposure ( $n_i$ ) for a country $i$ is calculated as the employment-share-weighted average of the exposure scores ( $e_j$ ) for each occupation $j$ :
$n_i = \frac{\sum_j c_{i,j} e_j}{\sum_j c_{i,j}}$
where $c_{i,j}$ represents the employment count in occupation $j$ for nation $i$ .

The study further validates these estimates by predicting observed national AI adoption statistics using data from Anthropic (Claude usage), OpenAI (ChatGPT usage rankings), and Microsoft (Generative AI adoption rates). Additionally, the authors model a novel mechanism for indirect exposure via remittances, using World Bank KNOMAD bilateral remittance matrices to calculate how exposure in a source country (e.g., Russia or the U.S.) propagates to a recipient country (e.g., Tajikistan or Honduras) through income dependencies.

Key Results

Substantial Cross-National Variation: National AI exposure varies widely. The most exposed nation, Luxembourg ( $n=0.37$ ), is 2.6 times more exposed than the least exposed, Burundi ( $n=0.14$ ).
Regional and Income Disparities: High-income countries are substantially more exposed than low-income countries. The regions of North America and Europe & Central Asia are approximately 50% more exposed than Sub-Saharan Africa.
Predictive Drivers: The "white-collar share" of a country's labor force is the strongest predictor of national AI exposure ( $R^2 = 0.91$ ). Gross National Income (GNI) is also predictive but less so than occupational composition ( $R^2 = 0.77$ ).
Gender Gap: In 91% of the studied countries (126 out of 138), women are more exposed to frontier AI than men. This is driven by the concentration of women in white-collar and sales occupations. In countries where women's employment remains concentrated in agriculture and household enterprises (e.g., Pakistan, India), the gap reverses or narrows.
Validation: The exposure metric strongly predicts observed AI adoption. A 0.10 increase in national AI exposure correlates with a 12-fold increase in per-capita Claude usage and a 19 percentage-point increase in the national generative AI adoption rate. However, once white-collar share and GNI are controlled for, the exposure metric adds little additional explanatory power, suggesting exposure serves as a structural summary of these broader economic factors.
Indirect Exposure via Remittances: The study identifies a second-order exposure channel. For countries heavily dependent on remittances, the AI exposure of the source country significantly impacts the recipient country. For example, Tajikistan has below-average direct exposure, but because 37% of its GDP comes from remittances (primarily from highly exposed Russia), its "remittance-accounted exposure" rises to above-average levels.

Significance and Claims
The paper claims to provide the first comprehensive, comparable measure of frontier AI exposure at the national level for a broad set of 141 countries. Its primary contribution is demonstrating that the economic impact of frontier AI is not uniform; rather, it is "jagged" and highly dependent on local labor market structures.

The authors argue that policy responses must account for this heterogeneity. Policies designed for advanced economies with high white-collar shares may be irrelevant or harmful for economies where labor is concentrated in agriculture or where income is heavily dependent on remittances from AI-adopting nations. By identifying that women in most nations are disproportionately exposed due to occupational segregation, the paper highlights a specific distributional risk that warrants attention given existing labor market barriers for women. Finally, the identification of indirect exposure channels suggests that national economic planning must consider how AI adoption in foreign labor markets will ripple through global income dependencies.