The AI Exposure Gap Shrinks From 9.7 Points to 2.5 Under Augmentation

Generative AI produces two very different labor-market gaps. The World Bank estimates that 4.5% of jobs in low- and middle-income economies face automation exposure, against 14.2% in high-income economies. That is a 9.7 percentage-point gap.

For augmentation, the shares are 16.2% and 18.7%. The gap falls to 2.5 points.

That comparison is useful precisely because it is not a jobs forecast. It measures whether a technology could perform or assist the tasks attached to occupations. It does not measure whether a worker has electricity, a computer, a suitable model, managerial support or permission to use it.

The exposure gap narrows on paper. The deployment gap does not.

What the 2.5-point gap measures

The underlying ILO and World Bank paper starts with an ILO-NASK index built from almost 30,000 tasks in Poland's detailed occupational classification, mapped to the international ISCO-08 system. Worker surveys, expert validation and generative AI modelling feed the scores.

The index sorts occupations using both average exposure and variation among their tasks. Gradients 1 and 2 contain mixed task bundles and lean toward augmentation. Gradients 3 and 4 contain higher, more consistent exposure and lean toward automation. National shares then use employment weights within occupations. Where detailed occupational data are unavailable, the method imputes from similar countries using employment structure, internet penetration, region, income and population.

The point arithmetic survives. The interpretation needs tighter limits.

The joint study covers 135 countries and 67% of global employment. Its coverage table and reproducibility materials show that the missing third is not evenly distributed: the sample captures 75.78% of high-income employment and 99.24% of lower-middle-income employment, but only 31.45% of upper-middle-income employment. East Asia and Pacific coverage is 29.56%. The World Development Report's overview chart separately reports averages across 137 countries using fiscal 2027 income groups.

Informal work is not wholly absent because labor-force surveys measure workers, not only payroll records. It is less directly observed. Some countries enter with two-digit occupation tabulations rather than worker-level task data. The task-content adjustment uses 46 countries, and its developing-country STEP surveys cover urban labor markets and exclude skilled agricultural workers for comparability. Income-group averages therefore conceal large differences in job formality, task content and data quality.

Access turns compatibility into a bottleneck

The infrastructure comparison is less flattering. The World Development Report's enabling chapter says fewer than half of people in low-income countries had electricity access in 2023. About one-quarter used the internet in 2024, against roughly half in lower-middle-income countries and 93% in high-income countries. Internet speeds were below 25 megabits per second in low- and lower-middle-income economies and above 300 in high-income ones.

Devices add another gate. Mobile-phone ownership was 63% among adults in low-income countries, compared with 95% in upper-middle- and high-income countries. A typical smartphone cost about 60% of monthly income in low-income economies and 4% in high-income ones.

Language coverage can make a connected tool unusable. The report finds that English accounts for 61.1% of training datasets on Hugging Face. Across one evaluation, model accuracy in 16 low-resource African languages trailed English and French by about 45 percentage points.

Skills matter after access. The joint paper finds that workers in developing countries perform fewer non-routine analytical tasks even inside nominally exposed occupations. A shared occupation code does not guarantee a shared workflow.

Traffic and output tell different stories

The report's adoption figures widen the gap further. Semrush website-traffic estimates put average ChatGPT use among internet users in April 2025 at nearly 25% in high-income countries, 5.8% in upper-middle-income countries, 4.7% in lower-middle-income countries and 0.7% in low-income countries.

Those estimates are not a census of adoption. Figure 4.3 reports income-group averages through April 2025 from Semrush website traffic. Its note warns that one person can be counted more than once when using multiple IP addresses or devices. It also says ChatGPT traffic from China is materially underestimated. Middle-income countries' reported shares are therefore website-traffic estimates, not measures of workplace penetration.

Workplace data show the operational divide. A World Bank survey of formal firms in six developing economies found advanced uses such as agents and workflow automation at 10%, against 24% in the United States. The sample covers firms with at least five employees and weights the six developing countries equally. It says little about informal microbusinesses.

Observed results also resist a simple exposure story. South Asian job postings fell 1.6% after ChatGPT's release, according to the report, while monthly listings dropped about 20% in the most exposed white-collar occupations with ready substitutes. Yet the broader cross-country estimate for low- and middle-income economies was smaller than the high-income effect and statistically insignificant. Job postings measure advertised demand, not employment or output.

A Kenyan field test points in the other direction. The same business assistant improved already stronger small firms and reduced performance among weaker ones, whose owners applied generic advice poorly. Exposure was shared. Organizational capability changed the result.

The operating metric comes last

The 2.5-point augmentation gap survives as a measure of technical compatibility. It fails as a claim about equal productivity opportunity.

Workforce strategists need three separate ledgers: tasks a model can assist, workers who can access it, and workflows that produce a measured gain. Public-sector buyers need reliable power, connectivity, affordable devices, local-language evaluation and staff training before counting an exposed role as reachable. Enterprise buyers should track cycle time, output, error rates and adoption inside the workflow, not chatbot traffic or licensed seats.

The task model identifies where AI might help. The distribution of complements decides where it will.