What the ILO–NASK Gradient Says About Your AI Task Exposure

What does the gradient actually summarize?

The ILO–NASK Global Index gradient summarizes modeled generative-AI automation potential across the standardized task list for an occupation. It combines the list’s average task score and how much scores vary. A high, consistent result means modeled potential is relatively high across much of the occupation’s task bundle. A low, uneven result means the average is low, while some tasks score higher than others. Neither result estimates whether you personally will lose work. The index does not measure your weekly duties, an employer’s AI adoption, or your likelihood of job loss. In “How might generative AI impact different occupations?”, the ILO says tasks were scored from 0 to 1 for potential automation, then occupations were plotted by the mean of their task scores and the standard deviation, a measure of spread. These are modeled assessments of task potential, not observations of what every worker or employer is doing. The gradient has four exposed categories. Gradient 4 is the high-and-consistent end: the occupational mean is high and task scores vary relatively little. The model therefore judges many tasks in that standardized bundle to have substantial automation potential. Gradient 3 also has above-moderate occupational exposure, but greater task variation: some tasks remain less exposed even as overall modeled potential is elevated. Gradient 2 describes moderate average exposure with high variation, a mixed task list. Gradient 1 sits at the low-average, high-variation end: most of the bundle has low modeled potential, while some tasks have elevated potential. The ILO also distinguishes minimal and no exposure; these are not personal labels of safety. The mean asks, roughly, “How high are modeled task scores across this occupation’s list on average?” Variation asks, “How similar are those scores, or how widely do they differ?” Together, they distinguish broad consistency from a bundle containing a mix. They do not say which task takes most of your day, which carries the greatest consequence if wrong, or whether your organization can use a tool reliably within its rules and workflow. It is not your personal time sheet. That distinction matters at both ends. A high, consistent category is a reason to inspect your actual tasks closely, since the model indicates broad potential across the occupation’s bundle. It is not a reason by itself to leave the profession: capability does not establish adoption, reassignment, reduced demand, or displacement. A low, uneven category is not a personal all-clear. One higher-potential task may deserve attention if it is frequent, consequential, or central to a handoff in your work, even when the occupation-wide mean is low. Treat the gradient as an attention map for a task list, not a verdict about your job. Check whether the standardized occupation resembles your responsibilities, then compare its tasks with the work you repeat and how results are reviewed and used. A higher or more consistent result can justify closer review; a low but uneven result can prompt a focused check of a particular duty. The appropriate next step depends on that fit and on workplace evidence, not on the category label alone.

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

How was a global occupation map built from task evidence?

The 2025 ILO–NASK map was built by joining detailed Polish descriptions of work with judgments from workers and experts, then transferring those judgments to occupations in the international ISCO-08 classification. In “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” the researchers explain that Poland’s six-digit occupational system listed 2,541 occupations and 29,753 tasks. They selected a representative sample of those tasks and asked 1,640 people employed in Poland, across the broad ISCO-08 occupational groups, to assess their potential for automation by generative AI. The survey produced 52,558 data points covering 2,861 tasks. The task ratings capture perceived automation potential under the study's assessment framework, not workplace-use logs, observed production performance or employer plans. This distinction is important because the index estimates what technology could do, not whether a firm has put it into a workflow. The method combines several kinds of input, The Polish classification supplied a detailed task vocabulary, more granular than ISCO-08 and rooted in a national system. Workers who were employed in the relevant broad occupational groups assessed sampled tasks, The researchers then compared those assessments with a survey and iterative, Delphi-style discussions among a smaller group of international experts. Expert review could validate or adjust scores where wording or initial assessments called for scrutiny. This is structured estimation, not observation of every task with a live system. The researchers used the human and expert input to build a repository of task knowledge, then used a model to predict scores for the much larger set of tasks in the Polish classification. They also generated scores for task descriptions in ISCO-08, the common international classification that allows occupations to be compared across countries. The ILO Working Paper reports that this scoring process covered nearly 30,000 Polish tasks and tasks in 2,500 Polish occupations at the six-digit level, while the public global index presents occupational results at the four-digit ISCO-08 level. Thus, a global occupation map is not a survey of workers in every country. It is a modeled transfer from detailed national task descriptions and human judgments into a framework designed for international comparison. That chain strengthens the map in a specific way: it combines a fine-grained task list, worker familiarity, international expert review and a shared occupational framework. It also defines its limits. The Polish workers’ assessments and the expert review inform estimated feasibility; the predicted ISCO-08 scores extend that knowledge to standardized tasks. Neither step establishes that a particular workplace has adopted a tool, that a system can complete a task reliably in production, or that an employer will reorganize staffing. The index concerns generative AI, not every form of artificial intelligence, robotics or industrial automation. And because the international step relies on standardized descriptions, it cannot capture every national rule, organization, tool or local version of a role. For a worker, then, ISCO-08 is a useful starting point for finding a comparable occupation, not a personal diagnosis. Check whether the occupation description resembles the responsibilities you actually hold; titles can conceal different mixes of duties, and similar work can appear under different titles. If the classification is a loose fit, inspect nearby occupational descriptions and compare their task lists before drawing a conclusion. The gradient can tell you how modeled task potential is summarized for a standardized occupation. Your own recurring duties, review responsibilities and workplace evidence are still needed to decide what that summary means for your week.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Generative AI and jobs: a refined global index of occupational exposure

How close is the occupation label to your actual week?

The occupation label is a standardized description of a bundle of tasks, not a personal time sheet. Compare the occupation’s tasks with your recurring duties, asking how often you do each, the consequence of error, who checks it, and what depends on it. It summarizes modeled task potential across the occupation’s bundle, not your personal frequency, consequence, checking, or dependencies. The method behind the index explains why the classification is useful and why it needs this second step. In “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” the ILO describes starting with Poland’s detailed occupational classification, which lists nearly 30,000 tasks across more than 2,500 occupations. A survey of 1,640 employed people in Poland collected judgments on a representative sample of tasks; international experts then reviewed and adjusted a large sample, and predicted scores were mapped to tasks in ISCO-08, the international classification used for comparison. This supports an internationally comparable model of task potential, not direct observation of every worker’s duties or task frequency at a particular workplace. That distinction matters even when your job title matches an index entry. Imagine two analysts whose roles both fit the same standardized occupation. One spends much of the week preparing recurring summaries and reconciling routine records, with a separate specialist approving the final recommendations. The other prepares fewer summaries but handles unusual cases, explains evidence to decision-makers, and personally signs off on recommendations. The occupation-level list may flag relevant task types, but it cannot say these analysts have the same personal exposure: their task mix, review arrangements, and responsibility differ. This is an illustration, not a finding about analysts from the ILO study. A useful mapping begins with a representative week rather than an unusually busy day or a formal job description written years ago. List recurring duties in plain verbs: draft, reconcile, explain, approve, coordinate, or resolve. Match each to the closest task descriptions in the relevant occupation. If the fit is loose, check adjacent occupation descriptions as well; many roles combine duties that a single title does not capture. This is an editorial mapping exercise, not an additional ILO measure. For each matched duty, note frequency, consequence, checking, and handoffs. Frequency shows whether a duty is occasional or repeated. Consequence asks what an error could affect; checking identifies who catches it and the review required; handoffs show whether others can proceed without your contribution. The ILO’s task-based account explains that automating a task does not by itself automate an occupation: the task’s importance and its relationship to the rest of the work matter. These questions apply the principle; they are not variables in the gradient. A rarely performed approval can still be a workflow bottleneck if later work cannot move until it is completed. Conversely, a frequently produced draft may be easier to assist or partially automate when a person checks it and remains responsible for the result. Neither pattern establishes adoption or role change. Do not equate hours spent with importance to the workflow. Mark dependencies beside frequency to see where capability could affect work sequence, while adoption, staffing, and outcomes remain open questions. ISCO-08 offers a shared vocabulary for comparison, while task allocation still varies across employers and teams. The title is a starting point, not the unit of personal analysis. If the occupational label fits poorly, do not discard the index or treat it as a diagnosis. Compare the nearest standard descriptions, identify the tasks that genuinely recur, and note which responsibilities are distinctive to your setting. Use the gradient to prompt inspection; the task map clarifies what applies to your week. A mismatch describes the limits of the match, not personal safety or exposure. The next question is what has actually changed in your workplace, which the modeled occupation summary cannot answer.

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

What can the gradient not tell you about capability, adoption or job loss?

The ILO–NASK gradient cannot tell you whether a task will work reliably from beginning to end, whether your employer uses generative AI, or whether your job is likely to disappear. It estimates modeled technical potential across occupational tasks. Those are separate questions. A task may look automatable in a description yet require context, access, judgment or correction that prevents a tool from completing it in practice. Even demonstrated capability does not show that a workplace has adopted the tool, assigned work to it, or reduced demand for human labor. The 2025 ILO Working Paper 140 frames generative AI's effects through the interaction between technology's capacity to automate tasks and the task composition of occupations. It describes possible transformation as well as substitution; it does not turn the gradient into an individual outcome estimate. The gradient is an attention map for examining work, not a forecast of your employer's next decision.

Between a model's ability and a changed job sit practical steps that an occupational score cannot observe. A draft may be easy to produce but costly to verify. If an error could mislead a client or trigger a consequential decision, a qualified person may still need to check the result, explain exceptions and take responsibility. Sensitive data may be barred from an unapproved tool. Connecting a tool to existing records can require integration work, and a workflow may depend on customer expectations or human sign-off. These factors can make assistance useful while limiting autonomous completion. They can also shift effort: less time composing a routine first version might mean more time checking, resolving unusual cases or communicating decisions. That is a possible workflow mechanism, not a claim that every organization will redesign work this way. The index does not measure local review burden, data rules, integration costs, customer preferences or accountability. A high task score cannot answer whether transfer would save total effort or make human review more important.

The evidence about workplace experience should be read on its own terms. The OECD project page reports surveys conducted in early 2022 with 5,334 workers and 2,053 firms in manufacturing and finance across Austria, Canada, France, Germany, Ireland, the United Kingdom and the United States. The related OECD policy paper, Using AI in the Workplace, reports workers' self-assessed performance and enjoyment responses. This is evidence about surveyed workers' experiences with AI in those sectors and at that time. It is not a measurement of current generative-AI adoption across all occupations, a test of the ILO–NASK task scores, or a counter-forecast showing that exposed work is safe. Positive reported experience can coexist with task change, and neither positive nor negative experience establishes what happened to staffing or labor demand. Keep observed use, employer adoption, reassignment and employment outcomes as different rungs of evidence.

The strongest reason to take exposure seriously is that repeated adoption accompanied by actual transfer of recurring tasks is more relevant to a worker's decisions than modeled potential alone. If procedures change, work is routinely routed through a tool, review expectations shift, or responsibilities move, those are local signals worth investigating. But reassignment does not by itself prove that total labor demand fell or that a particular person will lose a job. Work may expand elsewhere, review may remain, or the organization may change service volume; the gradient cannot resolve those possibilities. Before making a large career decision, look for approved tools and written procedures, ask which recurring duties have changed, who verifies outputs, and whether workload or staffing expectations have shifted. Separate a pilot from routine use and a manager's interest from a formal change. If the only signal is the occupation's gradient, monitor and map your own duties. If actual transfer is visible, a focused conversation about responsibilities and review is proportionate. Stronger local evidence improves the question you can ask; it still does not convert an exposure category into a personal job-loss probability.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Using AI in the Workplace; AI surveys of employers and workers

How is this different from other AI exposure measures?

The ILO–NASK gradient and U.S. Bureau of Labor Statistics (BLS) categories both summarize occupational exposure, but they measure different things. The 2025 ILO–NASK Global Index estimates generative AI’s modeled potential across standardized occupational tasks, summarizing each occupation with the mean task score and variation among scores. BLS’s categories for its 2025–35 employment projections combine theoretical exposure measures with mapped evidence of observed AI use, then group occupations by relative exposure. Neither result gives the probability that you personally will lose your job. Their scales and labels are not interchangeable: read each as an answer to its own measurement question, then compare either with your actual duties and workplace evidence. The difference begins with the inputs. “How might generative AI impact different occupations?” describes ILO task scores on a 0–1 scale and a chart organized by occupational mean and standard deviation. Its gradients distinguish high, consistent modeled potential from lower overall potential with substantial task variation, alongside intermediate patterns. These are categories in a GenAI task-potential framework. BLS combined five external sources for U.S. occupations: three theoretical measures and two sources mapping observed model interactions to occupational tasks or work activities. It calculates separate median percentile ranks for theoretical and observed dimensions, then uses clustering to assign low, moderate, high and very high relative exposure. One framework summarizes modeled task-score level and unevenness; the other combines two kinds of relative evidence. A BLS “high” label cannot be converted into an ILO gradient number. Geography and timing matter. The ILO index is designed to compare occupations internationally through standardized classifications, and its 2025 estimates concern potential exposure to generative AI. BLS categories accompany U.S. 2025–35 employment projections, but their exposure inputs have distinct dates and limits. BLS says theoretical sources describe capabilities available no later than mid-2023; observed sources may reflect early adopters of specific models. Its current-evidence dimension includes Claude conversations and API traffic, and Microsoft Copilot interactions, mapped to occupational tasks or activities. “Observed” means platform usage data were mapped to tasks associated with occupations. It does not mean BLS watched a named worker at work or verified that a particular person uses AI. Reading both can help because different inputs reveal different parts of the picture. A theoretical task match asks whether technology could assist or complete some work; mapped usage adds evidence that model interactions occurred for associated tasks. That evidence does not establish employer-wide adoption, changed workflows, reduced staffing, wage effects or displacement. BLS also notes that its method treats occupational components as independent and does not capture bottlenecks or complementarities. The ILO–NASK gradient has another limit: it is an occupational task summary, not a record of your task frequency, consequences, review burden or local rules. Neither measure establishes whether a duty can be handled reliably in your workflow. When you encounter an exposure measure, note its subject, inputs, geography, date and scale. Ask whether it reports modeled capability, mapped use or both; whether categories are relative to other occupations; and which outcomes the authors say it cannot establish. Then return to your work: which recurring duties match the standardized description, how often do you perform them, what must be checked, and has your workplace changed its tools, procedures or responsibilities? That comparison supplies context the index lacks. Use the ILO–NASK gradient to inspect your task bundle, and BLS as a distinct U.S. comparison where relevant. Local workflow evidence remains necessary before choosing a test, raising a redesign conversation or monitoring change.

Sources: Artificial Intelligence (AI) exposure categories; How might generative AI impact different occupations?; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

What can change inside a job without changing the job title?

A job title can stay fixed while its work sequence shifts. Generative AI may help draft, retrieve or organize routine material, but that does not show whether a task disappears or staffing changes. The ILO’s Generative AI and Jobs: A Refined Global Index of Occupational Exposure models task potential across occupational bundles; it does not observe an employer redesigning a role. The question is who checks output, handles exceptions and remains accountable. Consider an illustrative analyst workflow. A tool might draft a recurring report summary, leaving the analyst to trace unusual figures, compare them with local definitions and explain their meaning. That is assistance if the analyst still reviews evidence and owns the judgment. A bounded step might instead be automated, such as extracting standard fields before a report enters a queue; work then moves to checking exceptions. If output is unreliable, the team may retain its process or use it only as a rough start. These are possible mechanisms; outcomes depend on error costs and verification time. A faster draft may shift effort rather than remove it: less composition can mean more validation. If an exception requires context missing from the input, the shortcut can add investigation. The gradient does not score who sets definitions, approves decisions or communicates consequences, so it cannot tell how much work remains. Evidence on skill demand complicates the picture, but is no verdict on one worker. In its November 2024 policy brief How Is AI Changing the Way Workers Perform Their Jobs and the Skills They Require?, the OECD examined online vacancy data and identified programmers, budget analysts and administrative assistants among occupations more exposed to AI. Its abstract says management, business and digital skills remained among the most demanded, while demand for these groups fell in the workplaces most exposed; the decline was relatively small and should be monitored. This measured labor-market pattern does not show that every worker should acquire a specific skill or that domain knowledge is dispensable. Exposure does not mean every related skill is vanishing; continued demand cannot guarantee an individual’s duties will stay unchanged. The OECD’s Using AI in the Workplace surveyed workers and firms in manufacturing and finance across seven countries in early 2022. This cannot represent every occupation or establish current adoption from an ILO–NASK gradient. A tool may help with part of a job while management separately changes targets, staffing or review. Positive user experience is not occupational demand. For an individual, learning should connect to a real workflow change. Domain knowledge may help spot conflicts with established definitions; verification skill may help check outputs against source records; explanation may make exceptions understandable. These can extend contribution, but do not establish immunity. A short course fits a defined gap, such as evaluating outputs against known cases. A project may fit a practical gap, such as documenting a review step. A larger credential or pivot needs its own comparison of prerequisites, time, cost, location and personal constraints; the gradient alone does not warrant it. Pick one task and check whether its workflow has changed. If an approved tool is in use, ask what remains under human review, how exceptions are handled and whether responsibility moved. If nothing changed, keep watching rather than reorganizing your career. If routine production is shifting and review or explanation is growing, record what that work requires and discuss how it should be supported and checked. This evidence can guide modest learning or a redesign conversation, but cannot turn exposure into a prediction of job loss.

Sources: How Is AI Changing the Way Workers Perform Their Jobs and the Skills They Require?; Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Using AI in the Workplace

Which next move fits the evidence you have?

The ILO explainer “How might generative AI impact different occupations?” describes the gradient; it is a reason to inspect a task, not a decision about your career. A proportionate move depends on task fit, frequency, and evidence of workplace change. The 2025 ILO–NASK paper models generative-AI task potential across occupational task bundles; it does not prescribe a response for an individual. Use the gradient to direct attention, not to choose a costly course, resignation, or occupational pivot by itself. Compare listed tasks with a representative week, not just your title. Mark recurring duties, duties requiring judgment or approval, and tasks where errors cause meaningful rework or harm. A potentially exposed task may take little time but carry significant accountability; another may recur daily yet be easy to check. If the occupation description fits poorly, check nearby standardized descriptions and write down your actual duties before interpreting the result further. One path is to test or upgrade part of your current work. Consider it when a recurring task matches the description, your organization permits the tool, and the work can be bounded and reversed. Choose a low-consequence step with a clear human check, such as drafting a routine internal summary for review, if policy allows. Before starting, choose what to compare: correction time, errors caught, verification effort, or handoff quality. If permission or access is absent, ask about approved tools instead of improvising with a public service. A second path is a redesign conversation or discussion of adjacent duties. Consider it when work has visibly moved: a procedure changes, a recurring task is reassigned, review expectations shift, or error responsibility becomes unclear. Ask what is changing, who checks results, and how quality and workload will be assessed. The ILO paper, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” frames occupational outcomes through task composition and technological capacity, without predicting an employer's choice. Keep the conversation on observable duties and accountability, including where domain knowledge may help specify requirements, resolve exceptions, explain results, or own decisions. These are contributions to discuss, not guarantees of protection or advancement. A third path is to monitor and gather evidence. This fits when the occupation label is a partial match, local procedures have not changed, or a test would violate policy or exceed your time and budget. Periodically check for approved tools, changed instructions, transferred work, altered review duties, or workload signals. An announcement or demonstration alone does not prove routine work has shifted. A gradient without local change can justify attention, but not a major retraining commitment by itself. If evidence accumulates, reconsider a bounded upgrade, adjacent responsibilities, or a larger change. Keep learning or career moves within your constraints. A focused course or project may address a demonstrated workflow gap; a longer program may fit a destination requiring foundations or credentials you lack. Neither follows automatically from an exposure category. Compare prerequisites, feedback, cost, time away from paid work, location, salary floor, health, and family responsibilities before committing. The occupational summary cannot replace a decision based on your actual work and circumstances. The free task checker can organize task-level change-pressure signals and possible first actions, but it is optional. Neither the index nor a checker result is a validated probability of job loss. A personalized roadmap can compare stay-and-redesign, adjacent, and larger-change scenarios against stated constraints; it cannot guarantee employment or income. Use either only if it helps structure a decision: establish task fit, look for workplace evidence, then choose an action proportionate to both.

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

What should you do this week?

Write down the recurring tasks from one representative week, not just your job title. Beside each, note whether it resembles the standardized duties behind your ILO–NASK result, how often you do it, and what would happen if the work were wrong. The ILO’s “How might generative AI impact different occupations?” explains that the gradient summarizes modeled scores across an occupation’s tasks; it does not supply your personal time sheet or employer’s adoption record.

Then identify one workplace signal to check: an approved tool entering the workflow, a task being reassigned, a new review step, or a change in accountability. If a matching task is low consequence and workplace rules permit it, propose a supervised test. If rules are unclear, ask which tools and data practices are approved. If responsibilities have shifted, ask how workload and accountability are changing. If unchanged, monitor rather than treating the gradient as a reason for an expensive move. The free task checker can organize change-pressure signals, but neither gives a probability of displacement. In a team conversation, ask: “Which task is changing, who verifies the output, and what evidence would show that the workflow has actually changed?”

Sources: How might generative AI impact different occupations?

Questions readers ask