Who Takes the Hit? World Bank Finds GenAI Job Losses Are Uneven

Job postings for occupations most vulnerable to generative AI fell 5.8% in high-income countries after ChatGPT arrived. In middle- and low-income countries, the same measure barely moved. If you want to know who is actually absorbing the first wave of GenAI displacement, that gap is the answer — and a new World Bank study puts hard numbers on it across 84 countries.

The paper, "Who Takes the Hit? The Uneven Impacts of Generative AI on Labor Demand Across Countries," is a background study for the World Bank's World Development Report 2026, written by Jingyun Huang, Yan Liu, He Wang, and Shu Yu. [Fact] The team analyzed 555 million online job postings collected by Lightcast across 84 countries, covering the first quarter of 2021 through the second quarter of 2025. The authors treat ChatGPT's release in November 2022 as a technology shock and apply difference-in-differences and event-study designs — the same quasi-experimental toolkit economists use to separate cause from coincidence.

The 5.8% divide

[Fact] In high-income countries, postings for occupations with above-median vulnerability to generative AI declined by 5.8% on average relative to less substitutable occupations after ChatGPT's launch — and the effect intensifies over time rather than fading. [Fact] In middle- and low-income countries, the estimated impact is much smaller and statistically insignificant.

Read that second finding again, because it cuts against a decade of automation forecasting. The standard story held that developing countries, with their higher shares of routine work, faced the greater automation risk. Generative AI inverts the map: it targets cognitive, language-heavy tasks — drafting, summarizing, coding, customer correspondence — and those tasks cluster where offices, degrees, and English-language digital work cluster. In rich economies.

[Fact] The study finds that, holding GDP per capita constant, displacement is larger in countries with greater GenAI adoption, more years of schooling, stronger English proficiency, and deeper specialization in digital services trade. The very ingredients that made a workforce competitive in the digital economy are, for now, the ingredients that expose it to substitution.

In the US, the bottom rung of the ladder takes the hardest hit

The cross-country paper builds on a companion study of the American labor market: World Bank Policy Research Working Paper 11263, "Labor Demand in the Age of Generative AI: Early Evidence from the U.S. Job Posting Data," which examined 285 million US postings from 2018 through mid-2025.

The American numbers are blunt. [Fact] Postings for occupations with above-median AI substitution scores fell 12% relative to below-median occupations, and the gap widened from 6% in the first year after ChatGPT's launch to 18% by the third year. [Fact] Entry-level positions took the sharpest losses: postings requiring no advanced degree fell 18%, and postings requiring no extensive experience fell 20%. [Fact] By sector, administrative support postings dropped 40% and professional services postings dropped 30%.

Twenty percent fewer openings for people with no experience. That is the career ladder losing its bottom rung.

The pattern matters beyond the individuals affected. Entry-level clerical and support roles have historically been how people without elite credentials entered the formal workforce. If those openings shrink 20% while senior openings hold steadier, the labor market does not just get smaller — it gets harder to enter. Our own occupation-level tracking points the same direction: administrative assistants sit among the highest AI-exposure groups in our database. See the detailed data for administrative assistants, or compare customer service representatives and paralegals — three occupation groups standing closest to the blast radius the World Bank describes.

Developing countries are shielded — by the wrong things

[Estimate] The World Bank team attributes the muted impact in developing economies to limited GenAI adoption and missing complementary factors: less reliable connectivity, fewer firms integrating AI into workflows, less English-language digital work. Workers there are protected not by resilience but by distance from the technology.

That distance is closing, and the World Bank's own numbers sketch the trajectory. [Fact] The Bank's Global Economic Prospects report (June 2026) estimates that about 60% of jobs in advanced economies are exposed to AI, compared with roughly 40% in emerging market and developing economies and 26% in low-income countries.

A related ILO–World Bank background study published in March 2026, "Disruption without Dividend?," adds a sharper warning from 135 countries. [Fact] It finds that clerical and administrative workers in lower-income countries — jobs that have offered a pathway to decent work, particularly for women and young workers — often already have the internet access needed for their tasks to be automated, while the workers who might use GenAI to become more productive frequently lack reliable connectivity. Displacement can arrive quickly; the dividend cannot.

[Estimate] Put together, the shield protecting developing-country workers looks less like immunity and more like lag. As adoption spreads — and knowledge-intensive service exports are already showing early disruption signals — the entry-level and administrative exposure documented in the United States is likely to replay elsewhere, in labor markets with thinner safety nets.

The counterargument, taken seriously

Job postings are not jobs. A decline in postings can reflect hiring caution, interest-rate pressure, or reorganization rather than workers being replaced — US hiring cooled broadly over 2023–2025 for reasons that have nothing to do with ChatGPT. The difference-in-differences design addresses part of this by comparing more-exposed occupations against less-exposed ones within the same economy, but what it measures is demand at the margin, not net job destruction.

The authors are candid about the other side of the ledger, too. [Fact] The US paper notes that generative AI creates new occupations and enhances productivity, which may increase labor demand over time; the early evidence mainly suggests that some occupations are less likely to be complemented than others. And there is a coverage limit worth naming: Lightcast captures online, mostly formal-sector postings. In economies where most employment is informal, posting data sees only a slice of the labor market. This is three years of early evidence, not a verdict.

What to do with this

If you work in an administrative, clerical, or entry-level professional role in a high-income country, the World Bank's numbers describe your segment of the labor market, not someone else's. The practical response is repositioning, not panic: the postings that survived skew toward tasks that require judgment, relationships, and accountability — the parts of a role an employer cannot yet delegate to a model. Building a visible track record in those tasks, inside or outside your current job title, is the most direct hedge this data supports.

If you work in a developing economy, the study reads differently: as a window. The displacement wave documented in the United States has not fully arrived, which leaves time — for workers to build complementary skills, and for governments to extend the connectivity, training, and social protection that the ILO and the World Bank both flag as the deciding variables in whether GenAI becomes disruption with a dividend, or without one.

Sources

• World Bank, "Who Takes the Hit? The Uneven Impacts of Generative AI on Labor Demand Across Countries" — World Development Report 2026 background paper (Huang, Liu, Wang, and Yu)
• World Bank, Policy Research Working Paper 11263, "Labor Demand in the Age of Generative AI: Early Evidence from the U.S. Job Posting Data"
• ILO–World Bank, "Disruption without Dividend? How the digital divide and task differences split GenAI's global impact" — WDR 2026 background study, March 2026
• World Bank, Global Economic Prospects, June 2026 (AI exposure estimates)

AI-assisted analysis: this article was drafted with AI assistance from the cited World Bank and ILO primary sources and reviewed before publication. All quantitative claims are attributed to their sources; where we extrapolate beyond the data, the sentence is tagged as an estimate.