What ILO Working Paper 140 Shows About AI Exposure

What does Working Paper 140 show about income groups?

Working Paper 140 reports a difference in modeled generative-AI exposure across national income groups: 11% of employment in low-income countries and 34% in high-income countries fell in occupations assigned to one of its four exposure gradients. Globally, the corresponding share was about one quarter. The headline is an estimate about where employment sits across occupational categories, not a count of workers using generative AI or jobs already changed. The paper was published on 20 May 2025; these figures describe its 2025 framework and evidence. The global quarter includes all four gradients, from low exposure with substantial variation among tasks to high and more consistent exposure. It is distinct from the highest-exposure category: Working Paper 140 reports that 3.3% of global employment was in Gradient 4. That figure answers a narrower question than the “some exposure” total. A role in a lower gradient may contain particular tasks with potential for automation while other tasks have less modeled exposure; Gradient 4 indicates higher and more consistent potential across the occupation’s documented task bundle. Neither category says that every worker performs tasks in the same way. To produce the country-group percentages, the ILO assigned potential-automation scores to tasks within detailed occupations, summarized those scores into occupational gradients, then applied the results to labour-force survey data for more than 140 countries. “How might generative AI impact different occupations?” describes this as estimating potential exposure in employment. Thus 11% and 34% are shares of employment in occupations classified as having some modeled exposure. They are not measurements of generative-AI use, employer adoption rates, or the share of jobs that will disappear. The contrast is useful as aggregate context: employment is distributed differently across the model’s occupational categories in low- and high-income country groups. Group averages cannot tell an individual worker how much of their role is exposed, whether an employer will adopt a tool, or whether tasks will be redesigned. The ILO calls these potential-exposure estimates, not actual impact, and notes barriers to full implementation, including infrastructure, cost, digital skills and operational difficulty. For a personal decision, use the figures as a reason to examine recurring tasks and local workplace evidence, not as a probability of job loss or a verdict to leave a career.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; How might generative AI impact different occupations? (ILO Data)

How did the researchers turn tasks into employment shares?

The ILO’s *Generative AI and Jobs: A Refined Global Index of Occupational Exposure* (Working Paper 140) did not count workplaces where generative AI was already in use. It built an occupational exposure index in stages: researchers sampled task descriptions from Poland’s detailed occupational catalogue, collected workers’ judgments, compared a subset with expert assessments, used checked scores to predict scores for tasks in the international classification, grouped occupations by task-score patterns, and applied those categories to employment data. The final country figures estimate potential exposure from ratings, not automation observed in firms. The starting catalogue contained 29,753 tasks across 2,541 occupations in Poland’s six-digit classification. Researchers selected a representative sample of 2,861 tasks for human assessment, distributed across occupational groups. “Representative” describes the task sample, not a survey of workers worldwide. Poland’s classification is linked to the International Standard Classification of Occupations (ISCO-08), allowing detailed task descriptions to map to an international occupation framework. Workers employed in Poland rated tasks associated with their broad occupational group for perceived potential for generative AI to automate them. The survey initially included 1,640 participants. That is the initial count, not the number retained in the task-rating analysis: Working Paper 140 says 105 people who answered “don’t know” to most questions and completed fewer than half the required rankings were excluded. Researchers also removed 1,167 task responses marked “don’t know.” After these exclusions, 1,535 respondents contributed 52,558 valid ratings across the same 2,861-task sample. Thus 1,640 is not the final analytic sample. This design brings occupational familiarity into ratings, but has limits. But respondents were not necessarily rating tasks they personally performed: the paper reports that most task assignments did not match a similar task in each respondent’s own job. These are judgments about described tasks, not direct observation or employer automation data. “Don’t know” responses were about twice as common in manual or technical groups as elsewhere, averaging roughly 10% versus 5%. The authors suggest some participants may have felt less able to judge specialized technical or manual tasks. Missing judgments prove neither exposure nor protection. Researchers compared worker assessments with evaluations from national and international experts, then used iterative discussion and manual resolution where scores differed substantially. This was a check on task judgments, not a measurement of what employers had implemented. In the resulting process, survey scores and expert inputs informed adjusted scores; AI-assisted predictions extended scoring to tasks without the same direct assessment. Workers rated sampled Polish tasks, experts reviewed selected scores, and a model extended them to a broader catalogue. For the international step, researchers used the checked task-score repository to predict scores for tasks documented in ISCO-08. They summarized scores within four-digit occupations using both the average and variation among tasks. Similar averages can conceal different bundles: consistently scored tasks in one occupation, a mix of high and low scores in another. Working Paper 140’s four gradients reflect combinations of average exposure and task variability; they do not mean every task or worker in an occupation is equally affected. Finally, the researchers linked occupation classifications to labor-force data and calculated employment shares for countries and income groups. The calculation asks what share of employment sits in occupations assigned to each modeled exposure category. Worker ratings were not a count of global AI users or displaced employees. Employment weighting gives the comparison geographic scale, while the task and occupation model supplies the exposure classification. The result is a comparison based on documented tasks and employment composition, not a census of automation or forecast of job losses.

Sources: ILO Working Paper 140 full text; Generative AI and Jobs: A Refined Global Index of Occupational Exposure; NASK index

Why does estimated exposure differ across income groups?

The difference is best read as a difference in the occupations represented in employment across country income groups, after those occupations receive different modeled exposure scores. Working Paper 140 reports that 34% of employment in high-income countries and 11% in low-income countries falls in one of its four exposure gradients. The ILO’s “How might generative AI impact different occupations?” explains that occupational scores are applied to labor-force survey data from more than 140 countries. The group total therefore combines two ingredients: how employment is distributed across occupations and the score assigned to each occupation. Different distributions can produce different totals. This is a compositional explanation of the statistic, not evidence that national income itself causes exposure.

The scores matter because the index does not treat each occupation as one indivisible block. The ILO explainer describes scoring tasks within 436 detailed occupations, then classifying occupations by average task score and variation among tasks. Its four gradients distinguish more consistent exposure from mixed or limited exposure. A group total can be higher when more workers are employed in occupations whose documented task bundles receive higher scores, even though many tasks still require human input. The ILO notes that nearly all occupations include some tasks requiring human input. The percentages represent employment in occupations assigned to gradients, not a count of tasks that can simply be removed from jobs.

The comparison is not a direct measurement of each country’s day-to-day work. It shows employment weighted by occupation-level task scores. A higher group estimate can be consistent with a larger employment share in occupations containing more tasks the framework considers potentially performable with generative AI. It does not show that workers in those occupations do identical tasks across countries, or that employers have introduced systems to perform them. The ILO calls the result “potential exposure” and explicitly distinguishes it from actual impact.

The group shares also do not isolate why occupational distributions differ. Working Paper 140 applies occupation scores to employment data; it does not test whether income, infrastructure, education, industry structure, or employer decisions caused the result. The ILO explainer separately names electricity and broadband constraints, limited digital skills, technology costs, and operational difficulties as adoption barriers. These may affect whether potential exposure becomes tool use, but they are not what the 34% and 11% figures directly measure. Occupational composition and adoption conditions belong to separate steps in the explanation.

A national average conceals variation among countries, sectors, workplaces, and workers with the same job title. The comparison is useful for showing that modeled potential exposure is unevenly distributed across income groups. It cannot tell a worker how exposed their actual task bundle is, whether their employer will adopt a tool, or whether their job will be redesigned. The practical follow-up is to identify recurring tasks that resemble those scored by the index, then look for local evidence that employers are changing how those tasks are done. A country’s income category can frame that inquiry, but cannot replace it.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; How might generative AI impact different occupations? (ILO Data)

What does 'exposure' measure—and what does it leave out?

“Exposure” in Working Paper 140 means modeled potential for generative AI to automate tasks described within an occupation. It does not mean that a worker has used a tool, that an employer has adopted one, or that a job has changed. The ILO assigns each task a potential automation score from 0 to 1. It then summarizes an occupation’s task scores using their average and their variability. It estimates task potential, not workplace automation. The four exposure gradients help preserve that distinction. Gradient 4 describes high, relatively consistent exposure across an occupation’s tasks: the average is high and task-level variation is low. Gradient 3 describes significant exposure with greater variation; some tasks remain less exposed. Gradient 2 indicates moderate average exposure and a mixed task bundle. Gradient 1 has low overall exposure but substantial variation, meaning a few tasks may have higher potential even while most of the occupation’s work remains less exposed. The categories represent both level and distribution of exposure. The broader “some exposure” total includes all four gradients, not only Gradient 4. That structure also explains why the headline employment share must be read conditionally. The ILO applies occupation scores to labor-force survey data for more than 140 countries. Its explainer calls the resulting estimates an upper threshold for the share of employment that could be affected if the technology were fully implemented. The same explainer names limits on implementation: electricity and broadband access, digital skills, technology costs, and operational difficulties. The index does not measure employer adoption, reliability in a particular organization, or organizational choice. A useful way to keep the terms separate is to follow the steps. Technical capability concerns whether a system can perform or assist with a task under some conditions. Use means someone actually applies it. Employer adoption means the organization makes that use part of a workflow, with the necessary access, rules, and support. Job redesign concerns how responsibilities and handoffs change after adoption. Labor demand concerns what employers need or hire as those workflows change. Displacement is a further possible outcome, not a score embedded in the exposure category. The index does not observe each step directly. For example, a task may receive a high potential score because its output is largely digital and its steps can be represented in a task description. That does not show that a specific employer has connected a tool to its records, permits its use, trusts its output for consequential decisions, or has decided to remove a position. These are separate operational facts. The ILO’s release likewise says the figures reflect potential exposure rather than actual job losses, and notes that many tasks still require human involvement. The report’s transformation-over-replacement conclusion is an interpretation, not an observed outcome for each occupation. So the 34% reported for high-income-country employment is not the percentage of jobs already changed, nor the share likely to disappear. It is the employment share assigned to occupations in the four modeled exposure gradients under the paper’s framework. Use it for aggregate context. To judge a role, ask which recurring tasks are changing, what the employer has adopted, and who checks consequential work. The group estimate cannot answer those questions by itself.

Sources: How might generative AI impact different occupations? (ILO Data); One in four jobs at risk of being transformed by GenAI, new ILO–NASK global index shows

Does the 2023 comparison show that exposure has risen?

No direct rise follows from comparing the 2023 and 2025 headline percentages. ILO Working Paper 96, published in August 2023, divided potential employment effects into automation and augmentation. Working Paper 140, published on 20 May 2025, uses a refined index with four gradients of exposure. The categories answer different questions about potential task change, so the 2023 shares cannot be subtracted from the 2025 “some exposure” shares to calculate how much exposure grew. Working Paper 96 estimated that potential automation effects covered 0.4% of employment in low-income countries and 5.5% in high-income countries. It estimated potential augmentation effects at 10.4% and 13.4%, respectively. In that paper, automation meant tasks with potential for replacement, while augmentation meant tasks that could be assisted or changed while other duties remained. The authors described augmentation as the larger potential effect. These figures are employment shares assigned through a task-based model, not observed counts of jobs already automated or workers already using generative AI. The paper cautioned that its estimates did not account for infrastructure constraints that could limit use in lower-income countries. Working Paper 140 instead groups occupations into four progressively higher exposure gradients, using both an occupation’s mean task exposure score and the variability of scores across its tasks. Its headline estimate that 11% of employment in low-income countries and 34% in high-income countries falls in occupations with some exposure therefore has a broader threshold than Working Paper 96’s automation category. “Some exposure” includes several gradients; it does not mean every task in those occupations is automatable, or that those jobs will disappear. The comparison can show that both studies found a higher modeled potential share in high-income than low-income employment. It cannot show that the difference between the papers’ figures is a change that occurred between 2023 and 2025. The ILO’s research brief, “Generative AI and jobs: A 2025 update,” makes the method change explicit. It says the revised assessment combines task-level data, expert input and AI predictions, covers nearly 30,000 tasks at a more detailed occupational level, and defines four gradients by mean score and task variability. The brief reports a mean automation score of 0.30 in 2023 and 0.29 in 2025, alongside a standard deviation of 0.30 and 0.14. Those are outputs of the two index assessments, not measurements showing that AI use or automation in workplaces fell. The update also says capabilities in voice, image and video generation raised scores for some media- and web-related tasks. Neither detail isolates a cause for every score difference. Improved task evidence and expert input can alter how the index classifies tasks even when no workplace has changed. The 2025 brief says the update draws on a more refined methodology; its reported score movement should be read within that revised measurement framework. The papers do not provide a repeated observation of the same employers, workers or job tasks over time, with adoption and employment outcomes tracked between dates. They therefore do not establish a before-and-after labor-market trend. For a reader assessing a role, Working Paper 96 remains useful as the earlier comparison that separates potential augmentation from potential automation. Working Paper 140 offers a later, more granular map of modeled exposure. Read together, they support a cautious directional observation about occupational structure across income groups, not a claim that exposure rose by a calculable amount. The group estimates cannot tell whether a particular employer has adopted a tool, whether a worker’s tasks have been redesigned, or whether labor demand has changed. Use the 2023 paper as a predecessor with different categories, not as a personal forecast.

Sources: ILO Working Paper 96; Generative AI and jobs: A 2025 update; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

Who is concentrated in the highest gradients?

The highest exposure gradient is concentrated in clerical work, but the ILO–NASK finding is about modeled occupational task bundles and employment distributions, not a verdict on every person with a clerical title. The ILO’s 20 May 2025 release for “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” says clerical occupations have the highest exposure overall. It also reports that Gradient 4, the highest gradient, contains 9.6% of women’s employment and 3.5% of men’s employment in high-income countries. The percentages describe employment distribution, not who will lose work or use GenAI. The distinction between the top gradient and “some exposure” matters. The paper’s global figure of about one quarter of employment includes occupations across the gradients; its highest-gradient share is much smaller, at 3.3%. Gradient 4 indicates a concentration of tasks the index rates highly exposed, rather than a finding that an entire occupation can be performed by a system. The ILO says many tasks still require human involvement, making transformation more likely than full automation. The gender difference is an occupational-composition result, not an individual vulnerability estimate. Why clerical work ranks high is connected to the work the index can describe: tasks that are often digital and involve handling, producing, or organizing information. But “clerical” is still a broad occupational grouping. Actual jobs may also involve checking exceptions, resolving context, coordinating, and taking responsibility for records or responses. These activities can shape how a workflow changes. The index classifies potential exposure in task descriptions; it does not observe how much time a particular worker spends on each task or whether an employer has integrated a tool into a live process. The ILO–NASK release also identifies some highly digitized cognitive work in media, software, and finance-related occupations as exposed. This is a useful corrective to the assumption that exposure belongs only to administrative jobs. It does not mean every professional role is equally exposed. A software role may include generating or revising code alongside requirements clarification, testing, maintenance, and accountability for what runs. Media work can combine text production with sourcing, editorial judgment, audience knowledge, and rights checks. Finance-related tasks can involve repeatable document analysis as well as interpretation under organizational rules. These illustrations show task variation; they are not measured task shares from the paper. A category is therefore a starting question: which recurring tasks in this role resemble the exposed work, and which depend on context, verification, coordination, or accountable judgment? For a worker, the practical evidence is the actual workflow: what is being trialed or adopted, what outputs need correction, and whether responsibilities or job requirements are changing. The ILO’s group comparison can direct attention toward occupational patterns, but it cannot answer those workplace questions. Its release explicitly says the figures describe potential exposure, not actual job losses, and notes that implementation varies with infrastructure and skills. Use the label to examine task mix, not to make an individual career decision from the occupation name alone.

Sources: One in four jobs at risk of being transformed by GenAI, new ILO–NASK global index shows; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

Do lower-income countries face less practical change?

Working Paper 140 reports a lower share of employment in occupations with modeled generative-AI exposure in low-income countries than in high-income countries. That does not establish that workers in poorer countries will face less practical change, or that they are protected. The ILO’s 2025 explainer says the estimates describe potential exposure under full implementation; electricity and broadband gaps, limited digital skills, technology costs and operational difficulties can slow adoption. Those conditions affect whether a technical capability becomes an available workplace tool. Even where a tool is available, an employer still has to choose whether and how to use it, reorganize tasks, change hiring, or alter staffing. The paper’s group-level employment shares do not measure those later steps.

This distinction matters because potential exposure and practical effect can move in different directions. Poor connectivity may limit access to tools that could assist with analysis or drafting, while some workers with reliable access may encounter pressure around tasks that software can produce. Neither possibility is a conclusion about every workplace in a country. The 2025 result averages occupations and employment distributions; it cannot reveal whether a particular office has adopted a tool, whether its outputs meet local requirements, or whether management uses any resulting time savings to reduce headcount, serve more clients, or improve review. The ILO’s list of adoption barriers therefore qualifies how the index might translate into use; it does not convert the exposure gradient into a forecast of who benefits or loses.

A later ILO analysis provides a further reason to be cautious about treating occupation labels as interchangeable across countries. Working Paper 166, published in March 2026, analyzes labor markets across 135 countries and reports that conventional occupation-based measures can overstate potential impact in developing economies when they assume identical task content within an occupation. Using skills-survey data for 46 countries, the paper finds that workers in developing countries perform fewer non-routine analytical tasks, even within occupations classified as highly exposed. It also reports infrastructure differences that may constrain the use of tools for augmentation. These findings point to two distinct mechanisms: actual tasks can differ under the same occupational title, and connectivity can shape who can use a tool. The paper’s abstract and findings describe modeled exposure and potential effects; they are not observations of job losses across all 135 countries.

Working Paper 166 should be read as a later, separate analysis, not as a correction that rewrites Working Paper 140’s 2025 result. It uses different evidence and examines task-content and connectivity differences that the earlier standardized comparison cannot settle. Together, the papers support a careful interpretation: Working Paper 140 shows that employment composition produces a lower aggregate modeled exposure share in low-income countries, while later evidence questions how well a common occupational task description represents work in every setting. Neither source supports a simple ranking in which one income group is safe and another is exposed. For a reader considering a change, country income is too coarse to guide an individual choice. The relevant evidence is the recurring task bundle in the role, local access to tools, and observable changes to workflows or job requirements. Those details can differ within a country and within a single occupation.

Sources: How might generative AI impact different occupations? (ILO Data); Disruption without dividend? How digital divide and task differences split

How should a worker compare the group estimate with evidence at work?

Use the group estimate as the widest frame in a four-step check: country or income-group context, occupation category, the recurring tasks in your own role, and direct evidence of what your workplace is changing. The ILO’s “How might generative AI impact different occupations?” applies occupation scores to labor-force data for more than 140 countries; it estimates potential exposure in employment distributions. It does not tell you whether your employer has adopted a tool, whether your tasks have changed, or whether a role will be cut. An occupation title is a useful starting index, but it is a coarse description of work. The ILO classified occupations using the average and spread of task-level potential automation scores. Two people with the same title may spend different amounts of time producing standard digital material, checking outputs against local rules, or making accountable decisions. List repeatable outputs, work needing context or verification, and decisions tied to personal responsibility. Then look for workplace signals that sit closer to your decision. Has a process actually changed? Are colleagues using a tool in a defined workflow, and who checks its output? Have responsibilities, handoffs, or role requirements changed? Capability does not establish that a tool is approved, integrated, or reliable enough for a consequential task. Keep use separate from redesign and labor demand too. Workflow change alone does not establish lower hiring or personal risk. If the output is low-stakes and reviewable, a bounded trial can answer a smaller question than the index can. Choose one recurring task, record the usual time and rework, then compare a limited tool-assisted attempt with the same quality checks. Count correction time and handoff quality, not just first-draft speed. Stop or revise the trial if verification consumes the apparent gain or errors create costs downstream. For a larger move, add local evidence: current role requirements, credible openings in your geography, and what your employer is asking people to do. They signal current demand or expectations, not future hiring or displacement. Weigh them against your experience, salary floor, location, available learning time, health, and family responsibilities. The group estimate can prompt review, not by itself justify costly retraining or exit. ILO Working Paper 166 later examines how task content and connectivity differ across countries, reinforcing why standardized occupation labels need local context.

Sources: How might generative AI impact different occupations? (ILO Data); Disruption without dividend? How digital divide and task differences split

What is a proportionate next step?

Start with the work you repeat. List the tasks then mark which have predictable digital inputs and outputs, which need contextual judgment, and which require accountable human review. Choose one relevant task with an output you can check, such as preparing a first draft for review. Before testing a tool, note the usual time, common errors, and rework. Afterward, compare those measures and check whether verification erased any time saved. This is a bounded workflow test, not evidence that an occupation will shrink or that a tool will perform reliably in all settings.

Next, look for evidence closer to your decision: whether your employer is changing the workflow and what local job descriptions require. The ILO’s *Generative AI and Jobs: A Refined Global Index of Occupational Exposure* applies occupational task scores to employment data; its national income-group estimates cannot select an individual career move. Use them as context, then let your tasks, workplace signals, salary floor, location, and available learning time guide action. The free task-level checker can organize that review, but its change-pressure signal is not a validated probability of displacement. The optional career roadmap compares scenarios against personal constraints; it does not promise income or employment. Ask your manager which part of a recurring workflow they expect to change and how quality will be checked.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure

Questions readers ask