What the ILO–World Bank 2026 Study Changes About AI Exposure

What does the 2026 study change about country comparisons?

The 2026 ILO–World Bank study makes cross-country comparisons more conditional by adding two filters to occupational mix: differences in task content within the same occupation, and internet connectivity. Its 135-country analysis models where generative AI may apply and how access conditions shape potential automation and augmentation. It does not report that employers have adopted these tools, that productivity has risen, or that a particular worker will lose a job.

In “Disruption without dividend?,” the authors report lower aggregate automation exposure in developing economies but comparable potential for augmentation; they also find that conventional occupational measures can overstate exposure when they assume the same task bundle across countries. Connectivity adds another condition: internet access can constrain whether potential augmentation is usable, even where tasks appear applicable. The ILO announcement likewise emphasizes uneven impacts shaped by task composition and digital infrastructure. Read together, the findings suggest a three-layer comparison: occupational mix, within-occupation tasks, and access. Each layer changes what a national average can mean, while none establishes actual workplace use or outcomes. For an individual decision, use country figures as context, then examine recurring tasks, permitted tool access, and how review responsibility is changing.

Sources: Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact | International Labour Organization; New ILO–World Bank paper highlights uneven global impact of generative AI on jobs | International Labour Organization

What do the headline country figures measure?

The headline percentages are modeled employment shares exposed to generative AI, not observed layoffs. The ILO working paper Disruption without dividend? reports that roughly 30–32 percent of employment in high-income countries falls within its exposure measure, compared with about 10–15 percent in low-income countries. These are ranges for income groups, not exact scores for every country. The paper attributes much of the difference to occupations with higher automation exposure, including clerical and some professional roles. By contrast, the share of employment with augmentation potential is described as more evenly distributed across income levels. That distinction matters because “exposed” is not one outcome. In the paper’s framing, automation-oriented exposure identifies work where GenAI could take over tasks; augmentation-oriented exposure identifies work where it could assist people doing tasks. The two categories describe different possible relationships between technology and work. They do not say whether a particular employer has bought or approved a tool, whether a worker uses it, whether output is dependable, or whether staffing changes follow. Nor does an augmentation estimate establish a productivity gain: that would require evidence about use and results in actual workplaces. The percentage aggregates employment classified through occupations and their task profiles. It is a summary of potential applicability across a workforce, not a direct count of automatable tasks performed in every workplace. A country with more workers in clerical occupations can therefore have a higher modeled automation-exposure share even if the measure says nothing about what any one clerk does each day. The reverse also needs care: a lower group average does not mean that every occupation, employer, or worker in a lower-income country has little exposure. The ILO record says augmentation potential is more evenly distributed across income levels, while its automation-oriented share is the main source of the high-income versus low-income gap. Reading only the total percentage can hide that difference between channels. The World Development Report 2026 concept note makes the underlying distinction explicit: task exposure can lead toward automation or can augment a worker. This is a useful way to read the headline, but the concept note is framing for the report’s analysis, not separate confirmation of the working paper’s numerical estimates. The paper’s headline result is a modeled comparison, so it should be stated with its denominator and category: an employment share in a country-income group, under a specified exposure framework. It is not the share of people who will be replaced, and it is not a forecast of net employment change. For a worker, the practical implication is limited but useful. These figures describe broad labor-market context and help explain why country comparisons can differ when occupational composition differs. They do not provide an individual forecast, an employer adoption rate, a measure of demand for a particular occupation, or a finding that jobs have already been redesigned. Those questions need other evidence. Compare like with like: check whether two numbers refer to automation or augmentation, a country or income group, employment or tasks, and modeled potential or observed use. Without those labels, a headline percentage can look more decisive than the evidence allows. The distinction is essential when figures circulate without their original chart notes.

Sources: Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact | International Labour Organization; World Development Report 2026 Concept Note

Why can the same occupation label produce different exposure?

An occupation label is a useful starting point for comparing labor markets, but it is not a complete description of the work people do. An accounting clerk role can include varying proportions of drafting, record checks, exceptions and decision responsibility, reflecting local technology, skills and organization. If an exposure measure assigns the same task bundle to everyone with the same occupational code, it can mistake a shared label for identical work. That is the measurement problem the 2026 paper addresses. In “Disruption without dividend? – How the digital divide and task differences split GenAI’s global impact,” the International Labour Organization reports using skills-survey evidence from PIAAC and STEP covering 46 countries to compare task content within occupations. Its summary says workers in developing economies perform fewer non-routine analytical tasks, including within occupations classified as highly exposed; the same ISCO occupation can involve more routine or manual tasks in lower-income settings. When exposure estimates are adjusted for this task content, many developing countries move lower in the rankings. Conventional measures assuming uniform task content can therefore overstate GenAI’s transformative potential in those settings. This corrects an assumption in the comparison; it does not find that a worker or country is protected. It identifies an aggregate survey pattern, not that every worker with a title performs fewer analytical tasks or that routine and manual work cannot change. Modeled exposure establishes neither reliable production performance, employer adoption nor job removal. This matters when a national comparison describes an occupation across economies: it changes an assumed input to the estimate, rather than documenting a new outcome for workers. The narrower result is that occupational codes alone may miss the task mix behind an exposure judgment. Consider an example, not a case from the study. Two people share a customer-service title. One locates policy details and drafts routine replies from a knowledge base. The other handles more conversations requiring context, team coordination and judgment about escalation, and reviews suggested replies before they reach customers. A title-based measure could assign both workers a similar profile. A task-level view asks how much recurring work involves searchable, repeatable text versus context, interaction, exception handling and accountable review. It still cannot establish whether a tool is available, permitted, accurate enough or used at either workplace. The World Bank’s “World Development Report 2026: Artificial Intelligence for Development” concept note distinguishes automation-oriented from augmentation-oriented exposure, clarifying how one label can hide different channels. It describes planned analysis, however, and is not independent confirmation; the completed ILO paper directly supports the reported survey finding. Treat a job title as a prompt for investigation, not a verdict. List activities that recur in a normal week, especially those requiring local knowledge, judgment, exception handling or sign-off. Ask which recurring tasks are routine information work and whether that mix is changing at work. This account guides a proportionate next step; the country estimate remains context about modeled exposure, not a personal forecast. The adjustment changes how a cross-country ranking should be read, not the outlook for a named worker.

Sources: Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact | International Labour Organization; World Development Report 2026 Concept Note

Why do the 135-country and 46-country numbers both matter?

The 135-country and 46-country figures describe different evidence layers, so they should travel together. The 135-country figure refers to broad employment microdata used for the paper’s country-level exposure estimates. The narrower 46-country sample supplies skills-survey evidence used to adjust for differences in task content within the same occupation across countries. The study estimates a broad employment pattern across many countries while relying on a smaller set to examine whether people sharing an occupation label report comparable work. The figures are complementary, not competing counts of the same survey. The World Bank’s 2026 Concept Note states this distinction: detailed microdata cover 135 countries, while PIAAC and STEP surveys inform task-content adjustment in a 46-country sample. It describes the research design; it is not a second empirical replication of the completed paper. That difference matters because the adjustment challenges comparisons based on occupation codes alone. A common occupational classification makes cross-country estimates possible, but the code does not guarantee that tasks inside a job title are distributed similarly. Survey data in the smaller sample let researchers examine within-occupation differences rather than assume uniform task content. The ILO–World Bank paper reports that its country comparison accounts for such differences, while the concept note explains which survey sources support the adjustment. Broad employment data provide reach; detailed task data check whether an occupation label carries equivalent work content. The second layer makes interpretation more locally grounded where evidence is available. It does not follow that every country in the broad estimate received the same direct task-level measurement. A reader should not turn 135-country coverage into a claim that workers’ tasks were directly surveyed in all 135 countries. Nor should the smaller sample be treated as proof that the larger estimates are invalid. It marks a boundary on how directly the study can compare within-occupation task content across countries. Where direct skills-survey evidence is unavailable, the strength of a country-specific task comparison differs from that of broad employment coverage. That distinction affects how confidently one should rank countries. The World Bank’s reproducibility record says the paper’s exhibits were computationally reproduced by the Development Impact Analytics team and gives instructions for rerunning analyses when users have access to all data. The same record notes that some datasets are restricted or limited-access and are not included in the package. This supports computational reproducibility, not independent validation that task measures capture every local difference or that modeled exposure predicts employment outcomes. Reproduction can check whether code yields reported exhibits under available inputs; it cannot remove coverage limits or turn modeled estimates into observations of adoption or displacement. A careful comparison should state both denominators and what each supports: 135 countries for broad employment coverage, 46 for survey-based task-content adjustment. Country ordering is less definitive when task evidence is indirect or unavailable. When interpreting a national figure, ask: “Which countries are represented in each layer, and what was measured in each?” This discipline preserves the study’s contribution without implying equal measurement depth everywhere.

Sources: World Development Report 2026 Concept Note; Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact | International Labour Organization; Reproducibility package for ILO Working Paper 166 | World Bank

What does connectivity add to the comparison?

Connectivity adds an access condition to the study’s estimates. A task may be technically applicable to generative AI, yet a worker cannot use a tool in day-to-day work without a workable internet connection. The ILO working paper “Disruption without dividend?” describes an uneven pattern: people in roles exposed to automation are often already connected, while people whose tasks might be augmented frequently face infrastructure gaps. The paper therefore suggests that some developing economies may face disruption pressures before realizing productivity benefits. That is a conditional interpretation of modeled exposure and connectivity, not a report that job losses or productivity gains have occurred.

The two sides matter because one national exposure total can hide different routes to change. The study reports a larger aggregate share of employment exposed in high-income countries, mainly because clerical and some professional occupations contribute more automation-oriented exposure. It also says augmentation potential is more evenly distributed across income groups. But when internet access is considered, potential benefits may be less available in practice: some workers whose tasks could be assisted lack reliable connectivity, while some workers in exposed office roles are online. The ILO’s 27 March 2026 announcement summarizes this contrast. Together, the findings complicate a simple country ranking. Lower aggregate automation exposure does not mean there is no concentrated pressure, and comparable augmentation potential does not mean workers have comparable access to possible benefits.

The paper gives a sense of scale for the countries with detailed data: it reports 441.8 million jobs in augmentation-oriented exposure gradients, of which about 66.9 million lack internet access. These are modeled counts tied to the study’s exposure categories and connectivity data. They do not count workers who would certainly gain from a tool or measure realized productivity loss. A connection is only one condition. A worker may still lack an employer-approved service, stable access, training, permission to use workplace data, a suitable task, or time to check output. Access also does not show that an employer has adopted GenAI or changed staffing. The estimates do not establish those additional steps.

Connectivity is therefore a filter on potential, not a proxy for implementation. “Disruption before benefits” describes a concern about sequencing: connected workers may encounter automation pressure while infrastructure limits augmentation elsewhere. The paper frames this as a possible pattern for developing economies, not a universal timeline for countries or workplaces. Reliable internet might remove one barrier, but it cannot establish that a workflow fits, outputs are dependable, productivity rises, or employment changes. Task organization and skills also shape what happens.

For an individual worker, the country-level adjustment is context, not a personal forecast. It helps explain why technical exposure may differ from practical access to potential benefits. It cannot show whether an employer permits a tool, which recurring tasks are changing, or who checks consequential work. Examine local conditions directly: list repeated tasks, confirm whether an approved tool is available, and observe any changes in time, review, workload, or accountability. If policy permits, a small, reviewable workflow trial can reveal more about one’s work than an internet-access average. The comparison adds an access layer while leaving adoption and outcomes open questions.

Sources: Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact | International Labour Organization; New ILO–World Bank paper highlights uneven global impact of generative AI on jobs | International Labour Organization

How is this different from the 2025 ILO exposure index?

The 2025 and 2026 ILO papers answer related questions, but they refine different parts of the comparison. “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” asks how to score tasks for potential exposure more carefully, then aggregate those task scores into occupational and employment estimates. “Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact” asks a further cross-country question: does the task bundle behind the same occupation code vary by country, and does internet access change whether potential exposure can be used? The later study adds local task context and connectivity to international interpretation; it does not establish that the earlier index was invalid. The 2025 index is more detailed than assigning one risk label to every job title. Its method draws on a representative sample from a repository of 29,753 tasks in Poland’s occupational classification, a survey of 1,640 employed people across broad ISCO-08 groups, and 52,558 reported judgments about automation potential for 2,861 tasks. The ILO record says those worker responses were compared with surveys and several rounds of Delphi-style discussion among international experts. The resulting task repository informed model-assisted predictions for tasks in ISCO-08 technical documentation. The paper then arranged scores into four exposure gradients. These steps improve how potential task exposure is assessed; they do not show that employers adopted tools or that the scored tasks were actually automated. That distinction matters because occupations combine tasks. A higher exposure score can identify work where some activities may be affected while leaving other activities dependent on human input. The 2025 paper’s conclusion that transformation is more likely than complete replacement follows from this mixed-task structure. It is a framing of potential occupational change, not an observed displacement rate. The 2026 paper changes the comparison unit in a different way. It reports that conventional measures can overstate exposure in developing economies when they apply a common task profile to an occupation across countries. Drawing on PIAAC and STEP skills-survey data for 46 countries, it finds fewer non-routine analytical tasks among workers in developing economies, including within occupations classified as highly exposed. The paper’s wider headline employment coverage is 135 countries, so the 46-country task evidence should not be mistaken for direct task surveys in every country covered. The adjustment asks whether the occupation label carries comparable work content; it does not prove that every worker with that title performs the same local pattern. Connectivity adds yet another filter. The 2026 paper treats internet access as a condition for realizing potential augmentation, and reports that access gaps can widen income-group differences when exposure is interpreted through connectivity. That is useful context about practical access, but connectivity alone does not establish workplace permission, tool quality, training, adoption, productivity, or displacement. The two papers therefore should not be ranked as if one produces the definitive country score: the 2025 index refines task scoring, while the 2026 analysis adjusts for local task composition and access conditions. Before comparing figures, ask whether the measure varies task scores, occupation mix, within-occupation content, connectivity, or observed use. Those are distinct layers, and none is an individual job-loss probability.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure | International Labour Organization; Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact | International Labour Organization

What can a country comparison not tell a worker?

A country comparison cannot tell an individual worker whether an employer has adopted a tool, which parts of that worker’s job will change, or whether a job will disappear. The 2026 ILO–World Bank paper combines modeled task exposure with country-level task-content and connectivity evidence. Those layers improve the interpretation of labor-market context; they are not observations of a particular workplace or a probability of personal job loss. The paper’s central contrast is useful precisely because it resists a single “more exposed” story. It reports lower aggregate automation exposure in developing economies than in advanced economies, alongside comparable potential for task augmentation. It also says internet gaps may leave some workers who could use GenAI for augmentation unable to realize possible productivity benefits, while workers in connected, automation-exposed roles may face pressure sooner. The authors therefore suggest that disruption could arrive before productivity gains in some developing economies. That is a conditional interpretation of modeled exposure and access, not a record that displacement has already occurred or that gains are guaranteed elsewhere. Even potential and actual use do not line up perfectly. The World Bank Atlas of Global Development’s “Inequalities in Use of and Exposure to Artificial Intelligence” gives managers as an example: their theoretical exposure can be high while observed use is relatively low. This supports a general caution about interpreting exposure as adoption; it is not a direct test of the 2026 paper’s country estimates. A task may appear technically applicable without a worker having an approved tool, suitable data, reliable service, training, or a workflow that makes its use worthwhile. A country score cannot settle either question. The 2025 ILO paper, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” helps explain why exposure still matters without turning it into a replacement forecast. Its method combines worker perceptions about 2,861 tasks, expert input and model-assisted predictions; it reports four rising exposure gradients. The paper’s abstract says most occupations include tasks requiring human input and frames job transformation as the most likely impact. The 2026 study adds local task-content and connectivity conditions to a country comparison; it does not show that the 2025 index measured realized use or that its task-level method was invalid. For an individual, neither result can establish whether a model performs a task reliably enough for a real workflow, whether a manager will reorganize work around it, or whether saved time will reduce staffing, increase output, or create more review work. Those are separate questions about capability, adoption, job design, demand and displacement. The country comparison supplies context about where potential exposure and access constraints may differ; it cannot substitute for evidence from the worker’s own tasks and workplace. So do not use a country rank to conclude that your role is safe, doomed, or due for an immediate exit. Start with recurring tasks and look for observable changes: which work is drafted, searched, summarized or formatted; where a person still checks accuracy, context and consequences; whether a permitted tool is actually available; and whether workload or review responsibility has shifted. If policy allows, a small, bounded workflow test can show whether a tool helps with one task and what verification it requires. If access is absent, that is evidence about the present work setting, not a guarantee about future conditions. The practical question is not whether a country’s average predicts your fate, but what is changing in your task bundle and what evidence would justify your next step.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure | International Labour Organization; Inequalities in Use of and Exposure to Artificial Intelligence | World Bank Atlas of Global Development; Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact | International Labour Organization

What is a proportionate next step after reading the estimates?

Use the country estimate as context, then make a decision from work you can observe. The ILO working paper, “Disruption without dividend?,” compares modeled exposure and access across countries; the 2025 ILO index, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” measures occupational exposure, not your employer’s adoption or your chance of displacement. Neither says whether a workflow has changed. List recurring tasks from a normal week. Separate drafting, searching, summarizing, and formatting from fact checking, applying local context, handling sensitive information, earning trust, and taking responsibility for decisions. This inventory helps inspect the work; no category is always automatable or protected. Exposure matters alongside the quality threshold, review burden, and consequences of error. Check whether a tool is available and permitted. National connectivity does not establish employer approval, safe data access, training, or reliable output. If policy allows, choose one low-stakes, repeatable task for a bounded trial. Record changes in completion time, corrections, handoffs, and accountability. Stop if the result cannot be reviewed safely or requires more checking than it saves. Then choose a proportionate response: practice verification or task framing, ask about approved training, or explore an adjacent role if the work is shifting. Keep the next step within your time and income constraints. One trial informs one workflow; it does not forecast labor demand. The free task-level change-pressure checker can organize questions and first actions, but its signals are not a validated probability of job loss.

Sources: Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact | International Labour Organization; Generative AI and Jobs: A Refined Global Index of Occupational Exposure | International Labour Organization

Questions readers ask