By Dr. Gyan Pathak
AI-Driven automation could close off many pathways to decent work, including clerical and administrative positions, particularly for women and young workers, which historically represent relatively higher-quality jobs in lower-income countries. Not only that, the disruption by Generative Artificial Intelligence (GenAI) in developing economies may materialize faster than productivity gains due to existing digital gaps and differences in how work is performed.
These are among the key findings of a new joint working paper by the International Labour Organization (ILO) and the World Bank titled “Disruption Without Dividend? – How the digital divide and task differences split GenAI’s global impact. This background paper for the World Development Report 2026 examines labour market exposure to GenAI across 135 countries, covering around two-thirds of global employment. It says that GenAI is set to reshape labour markets worldwide, but with uneven impacts across countries.
It is the latest study among the rapidly growing body of literature documenting the potential labour market impact of GenAI even since the launch of ChatGPT in November 2022.While early experimental studies focused on specific occupational segments reveal substantial heterogeneity in impacts, the aggregate implications for labour demand, job quality, and earnings distribution remain highly uncertain.
Employment exposure to generative artificial intelligence positively correlates with economic development, the paper finds. Advanced economies exhibit the highest exposure rates, exemplified by the United States and France with 30 percent or more of employment exposed to GenAI. Conversely, lower income economies such as Ethiopia and Zimbabwe demonstrate substantially lower exposure levels, at 10 percent or less of employment.
The predominant category of exposed employment—about 17 percent globally—falls within moderate exposure levels (gradients 1 and 2) that favour job augmentation rather than substitution. In contrast, roughly 8 percent of employment lies within gradients 3 and 4, which correspond to jobs facing higher automation risks. Although substantial variation in automation exposure exists between poorer and wealthier nations, these disparities are much smaller for the augmentation-related gradients.
This asymmetry suggests that, if occupational structure is the main driver of GenAI exposure, potential productivity gains from generative AI are more evenly distributed across countries than automation risks, which remain concentrated in high-income economies. In other words, based solely on theoretical occupational exposure, developing countries appear relatively better positioned to benefit from the adoption of these technologies compared with the automation risks they face. Realizing this optimistic scenario requires assuming equal implementation opportunities across countries at different income levels.
On GenAI exposure by internet access, the paper says that correcting for the digital divide amplifies the income-related differences in countries’ total exposure. As a starting point, developing countries have a smaller total share of occupations exposed to GenAI. Such low level of exposure falls even further in several countries where a significant fraction of such workers–in some countries the majority of them–do not have internet connectivity. The implications of the digital gap become even more apparent when examining specific occupations. Take, for example, shop sales assistants—a common job type potentially exposed to GenAI augmentation through tools that support customer interaction, inventory management, or sales analytics.
Among all countries in the sample, there are 441.8 million jobs that fall within the Gradient 1 and2 exposure categories, meaning that they could leverage GenAI for job augmentation. However, 66.9 million (15.1 percent) of those jobs do not have internet access. The broad occupational categories within this category include managers, technicians, and service and sales workers. In low- and lower-middle-income countries, plant and machine operators are highly prominent in this group, with a large part of this results due to vehicle drivers. One has to keep in mind that we lack high quality labour force and internet data for many low-income countries(LICs), meaning that in reality this gap is likely significantly higher.
On the automation side, in the GenAI exposure index applied, clerical support workers and, to a lesser extent, professionals are among the occupational categories with the highest exposure (gradients 3 and 4). This, in turn, translates directly into a strong link between exposure to GenAI automation and GDP per capita, given that the employment shares of these occupations are higher in prosperous economies. These vulnerable positions are disproportionately held by women and younger workers.
On average, 17 and 10 percent of female workers in high and upper-middle-income countries face automation risk from GenAI, compared to approximately 11 and 6 percent of male workers, respectively. In LICs, exposure to GenAI automation is significantly smaller for both men and women, representing less than 3 percent of total employment.
There are age disparities in exposure, particularly in upper-middle-income countries, where about 10.4 percent of workers aged 16-35 face automation risk, while only 4.3percent of workers aged 56-65 are in that position. In contrast, exposure to GenAI job augmentation is more equally distributed across genders and age groups, indicating that GenAI’s productivity benefits may be more equitably shared than its displacement risks.
The intersection of automation exposure with internet access further exposes income-based differences among countries. In high- and upper-middle-income nations, workers most vulnerable to GenAI automation tend to have high internet access levels, suggesting that displacement effects could theoretically occur rapidly. In contrast, less developed nations, particularly low income countries, have minimal shares of employment exposed to potential GenAI automation as a starting point. However, where such jobs exist, they are mostly digitized, which means that the digital divide does not protect these jobs from potentially rapid negative effects.
Further distributional implications of GenAI emerge when examining patterns across the levels of education. A common feature across all countries is that GenAI exposure increases with educational attainment, which highlights the particular ability of this technology to interact with what is broadly understood as ”knowledge work”. Importantly, workers with comparable education levels exhibit remarkably similar exposure rates regardless of their country’s income level. For example, college graduates face exposure rates ranging from 32.5 percent in low-income countries to 37.7 percent in high-income countries, which represents a far narrower gap than the differences observed across the entire employed population.
In LICs and LMICs, internet connectivity is also disproportionally higher among more skilled workers and among those at risk of automation. For instance, among college graduates in LICs,13.2 percent face GenAI automation exposure, and 72 percent of these workers already have internet access. This implies that initial displacement effects could concentrate among the narrow segment of skilled and digitally-connected workers currently holding some of the better jobs in those economies.
The breakdown of exposure by income suggests that GenAI’s first-order effects are likely to concentrate on middle-class occupations, although the magnitude and distributional direction of these effects remain uncertain. (IPA Service)
