What ILO’s 2025 AI Index Says About Task Variation

What does variability measure inside an occupation?

In the ILO’s 2025 index, task variability means how widely automation-potential scores differ across the tasks listed for a standardized occupation. The ILO’s “How might generative AI impact different occupations?” explains that each task receives a score from 0, where performing it with generative AI is considered not possible, to 1, where it is considered entirely possible. For each occupation, the index plots the average of those task scores, called the mean, alongside their standard deviation: a summary of how dispersed the scores are around that average. The mean describes the center of the task-score profile; standard deviation describes its spread. Neither statistic measures how much time workers spend on a task. Consider an illustrative pair of task lists, not actual ILO occupations. Each has four task scores averaging 0.5. One might contain 0.4, 0.5, 0.5, and 0.6: scores cluster near the mean. Another might contain 0.1, 0.2, 0.8, and 0.9: scores are more spread out despite the same average. This shows why a mean alone can hide different profiles. It does not reveal which real tasks have higher scores; that requires task-level entries. Nor does a larger spread say that a worker performs the high-scoring tasks or that they are frequent. The unit summarized is a set of scored tasks grouped under an occupational classification. It is not a sample of people with that title, one employer’s workflow, or a time-use diary. In “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” the ILO describes a task-level framework for assessing exposure across occupations. That classification supports comparison, but cannot show how duties are divided among workers or weighted by frequency, hours, importance, or accountability. A rarely performed task and a daily task can both appear in a list; the statistics do not encode their different shares of a working week. Read variability narrowly: it tells you whether the assessed task scores in an occupation are relatively clustered or mixed around the mean. It does not measure differences among workers who share a title. To understand a role, compare the listed tasks with work done, then check which duties recur and which carry responsibility. That distinction matters before using an occupational profile to judge personal exposure. The next step is to see how the ILO combines mean and spread when assigning exposure gradients.

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

How do the gradients combine the mean and spread?

The ILO’s four exposure gradients combine two summaries of task scores: the mean, or average potential-automation score across an occupation’s listed tasks, and the standard deviation, which describes how dispersed those scores are around the mean. A gradient therefore describes a standardized occupational task profile. It does not rate a particular worker, measure time spent on tasks, or show whether an employer uses a tool. The ILO explainer presents the gradients as combinations of occupational exposure and task variability, not levels of an individual’s job-loss risk.

An average can conceal different patterns. Imagine two illustrative occupations with the same mean score. In one, most listed tasks sit near that average; in the other, lower scores sit alongside a smaller group of much higher scores. Their means match, but the second profile is more dispersed. These are hypothetical patterns, not ILO occupation results. Standard deviation flags this difference, but does not reveal which tasks create it. To locate them, a reader needs task-level scores or descriptions.

The gradient labels are not a simple ladder where each step means the same kind of change. In the ILO chart explainer, Gradient 1 pairs low overall exposure with high variability: the mean is low, while some task scores are more elevated. Gradients 2 and 3 also have high variability, but describe moderate and above-moderate occupational exposure. Gradient 4 describes high exposure with low variability, meaning scores are more consistently high across listed tasks. The categories called minimal exposure and not exposed sit outside the four gradients: they pair low exposure with moderate variability, and low exposure with low variability, respectively.

Ask where the mean sits and how much task scores vary around it. Low mean plus high spread describes a low-average profile with some higher-scoring tasks; high mean plus low spread describes a more consistently high-scoring profile. Similar means can have different spreads. The gradient compresses these dimensions into a category, useful for comparison but unable to replace the underlying numbers or task list. The ILO Working Paper, Generative AI and Jobs: A Refined Global Index of Occupational Exposure, calls the 2025 framework four progressively increasing exposure gradients; the chart explainer provides the specific descriptions of mean and variability.

When reading a gradient, check which combination of average and dispersion it represents, then inspect the task evidence. Standard deviation alone cannot identify whether a higher score belongs to drafting, checking, exception handling, or another listed duty. Nor does the gradient show how often that duty appears in someone’s actual role. Variability has a narrower, practical use: it warns against treating an occupation’s average as if every task shared that score. Use it as a reason to look inside the occupational profile and compare the listed work with the real job, not as a complete summary of an individual role or a prediction of what happens next.

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

What evidence sits behind the 2025 task scores?

The 2025 scores are estimates of technical potential assembled from human judgments, expert review, and AI-assisted extension. They are not observations of routine tool use inside employers. The index asks how feasible it may be for generative AI to perform described tasks; workplace adoption asks whether an organization actually deploys and relies on a system. The ILO publication record for Generative AI and Jobs: A Refined Global Index of Occupational Exposure describes a multi-stage assessment intended to build task coverage across occupations, not a survey of employer deployment. The starting task universe was Poland’s occupational classification, which contains 29,753 tasks. From it, researchers used a representative sample and surveyed 1,640 people employed across the major one-digit groups of the International Standard Classification of Occupations, or ISCO-08. The record reports 52,558 data points concerning perceived automation potential for 2,861 tasks. These numbers describe survey input: they do not mean each respondent rated every task, or that every task in the full Polish classification was directly assessed by workers. The record does not describe responses as time diaries or measurements of task frequency in a particular job. Researchers compared this input with an international expert survey and several Delphi-style discussion rounds among a smaller expert group. The ILO says this process produced a task-automation knowledge repository intended to extend beyond national specificities. An AI assistant then predicted scores for tasks in ISCO-08 technical documentation. This broadens coverage toward an international classification, but final task coverage does not consist solely of direct worker judgments: it includes expert review and AI-assisted prediction. Each stage addresses a different limitation while introducing a different boundary. Respondents can judge perceived technical feasibility, but that is not evidence that a tool is available, reliable, approved, or used at work. Expert review can refine judgments, yet agreement still concerns task descriptions rather than observed workplace outcomes. AI-assisted extension can fill classification gaps, but a predicted score is an extrapolation, not a direct measurement of the task in every country. The public abstract does not give a task-by-task error rate that would justify treating every extended score as equally certain. The companion ILO chart explainer says the assessment covers 436 detailed ISCO occupations and assigns each listed task a potential-automation score from zero to one, then summarizes scores by occupation. This standardization enables comparison, but a classification cannot capture every local version of a job. The Polish task source and international mapping are reasons to check whether descriptions fit the work under discussion, especially when duties differ by employer or country. The score also does not show which duties dominate a person’s week or carry the most responsibility. Use the index as a structured exposure screen: it organizes judgments about the technical potential of described tasks and helps identify where closer inspection may be useful. Claims about current use require separate workplace evidence or an adoption study suited to that question. Before applying a score to a role, compare the occupation’s task descriptions with actual duties, keeping perceived feasibility, expert assessment, predicted coverage, and observed deployment distinct.

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

Why is the 2023-to-2025 spread not a before-and-after measure?

The ILO reports that the standard deviation of occupational task-automation scores was 0.30 in its 2023 assessment and 0.14 in its 2025 assessment. The corresponding means were 0.30 and 0.29. These figures describe two versions of an index; they do not show that work itself became more uniform. The 2025 brief says the newer estimates use a refined methodology combining task-level data, expert input, and AI predictions, and cover nearly 30,000 tasks. A lower spread in the resulting scores is a change in assessment output, not a repeated measurement of the same workers' duties or workplace task bundles. [Generative AI and jobs: A 2025 update](https://www.ilo.org/publications/generative-ai-and-jobs-2025-update) Standard deviation describes how dispersed a set of scores is around its mean. Here, the set is the assessed task profile associated with standardized occupations. The comparison does not follow employees, observe the same jobs before and after new tools arrived, or record changes in employers' task assignments. The figures alone cannot tell how much reflects task coverage, scoring inputs, classification, or revised method. Calling the lower number a decline in workplace variability would give the index a time-series meaning this comparison does not establish. The two papers also use different assessment designs. The ILO's 2023 Working Paper 96 assessed occupational exposure with particular attention to GPT-4 and organized its interpretation around potential augmentation, automation, and an ambiguous middle. Its abstract describes augmentation as the likely main effect. The 2025 Working Paper 140 presents a refined global index drawing on task-level data, expert input, and AI model predictions, and groups occupations into four exposure gradients. It cites recent advances in GenAI and greater user familiarity as part of the update. These changes in evidence and classification do not mean the underlying task lists or jobs were held constant. [Generative AI and Jobs: A global analysis of potential effects on job quantity and quality](https://www.ilo.org/publications/generative-ai-and-jobs-global-analysis-potential-effects-job-quantity-and); [Generative AI and Jobs: A Refined Global Index of Occupational Exposure](https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure) Read the comparison narrowly: scores produced in the 2025 exercise were less dispersed, relative to their mean, than scores produced in 2023. It does not establish that tasks converged in real workplaces, employers adopted more consistent systems, or workers now face a more uniform pattern of change. The mean shifted only slightly in the reported summaries while the spread shifted more; neither fact supplies a causal explanation. Without a common task set, scoring procedure, and repeated observations of actual work, the difference is not a before-and-after labor-market result. For the assigned period, use the 2025 spread with the 2025 mean and gradient framework. Keep the 2023 figures as methodological context, not a baseline proving a decline in occupational variability. To apply the index to a role, check whether its listed tasks match the work, which duties recur or carry responsibility, and whether a tool is actually used in the workflow. The comparison sharpens those questions; it cannot answer them.

Sources: Generative AI and jobs: A 2025 update; Generative AI and Jobs: A global analysis of potential effects on job quantity and quality; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

What can an occupation average hide about actual jobs?

An occupational average summarizes the classified task bundle, not the task mix of a particular employee, employer, or country. In the ILO's 2025 chart, each standardized occupation is represented by scores assigned to its listed tasks; the mean summarizes those scores and the standard deviation describes their spread. Neither statistic records which duties a specific worker performs, how often they occur, or how much responsibility sits with them. A wide spread across tasks is therefore different from variation between workers who share an occupation code. It describes the list used in the index, not a survey of role-by-role differences inside an organization. This is a translation problem between a standard occupational description and a particular job. Employers can organize work differently; specialists may handle only part of a broader bundle, while senior staff may spend more time on review, exceptions, or sign-off. Local industries, regulations, software, and customer needs can also change which tasks are present. A mean and standard deviation do not encode time share, frequency, sequence, consequence, or who is accountable when an output is wrong. A rare task might carry substantial responsibility, while a routine task might occupy much of a workday. The two summary statistics cannot tell those cases apart. Consider an example, purely illustrative. Two workers have the same administrative occupation title. One spends much of the week preparing recurring reports from structured records, then routes unusual cases to a colleague. The other spends more time resolving exceptions, checking source documents, and signing off on final records. An index task list might include reporting, checking, and decision-related activities in the same occupational bundle, but its mean and spread would not reveal how those duties are divided between these two jobs. Nor would the statistics show whether a tool is available, approved, or used in either workplace. This is illustration, not an ILO finding. Geography adds another boundary. The ILO's 2026 Working Paper 166, “Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact,” examines 135 countries and reports that conventional occupational exposure measures can overstate effects in developing economies when they assume task content is uniform. Its summary says analysis using PIAAC and STEP skills-survey data for 46 countries found fewer non-routine analytical tasks within some highly exposed occupations, with the same ISCO-level occupation sometimes involving more routine or manual work in lower-income settings. This later analysis does not recalculate the 2025 index's variability statistic, and it does not show that every occupation differs across countries. It does support a narrower caution: the task bundle behind an occupation label may not transfer unchanged across national contexts. Before using an occupation's spread to think about your own work, compare the index's mapped tasks with the duties you actually perform. Note which recur, which consume meaningful time, which are consequential even when uncommon, and where judgment or sign-off remains yours. Then identify whether a suitable system is actually used in your workplace and whether it changes the workflow. A task inventory makes that boundary visible without pretending to resolve every uncertainty.

Sources: How might generative AI impact different occupations? (ILO occupational exposure chart explainer); Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact

What the variability score cannot establish about adoption or jobs

A high or mixed task-exposure profile does not establish that a suitable system is deployed, that an employer will redesign work, that labor demand will change, or that a worker will lose a job. The ILO’s 2025 chart assigns potential automation scores to tasks in occupations and summarizes them with means and standard deviations. It addresses potential capability across task lists. It does not record which tools an employer uses or what happened to staffing after workflow changes. The index stops before workplace adoption and employment outcomes.

Potential scores are several steps from employment effects: whether a system handles the task reliably reliably in real workflows; employer access, governance and adoption; changes in responsibility and review; effects on output, service or cost; and resulting labor demand. Each needs its own evidence. A task can be technically addressable yet unsuitable with inconsistent inputs, costly errors, or required human approval. A tool may assist one step while leaving verification or exceptions to a worker. Neither possibility follows from the occupation’s spread statistic.

The ILO explainer, “How might generative AI impact different occupations?”, calls its employment estimates potential exposure and an upper threshold conditional on full implementation. It names constraints including electricity and broadband infrastructure, digital skills, technology cost, and operational difficulties. These can separate potential from implementation but do not identify barriers at a particular workplace or establish employer adoption. The explainer’s view that job transformation is more likely interprets its assessment broadly, not a measured outcome for every occupation or worker.

A separate ILO brief, “The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence,” published in June 2026, reviews experiments, firm data, platform studies, and surveys in seven countries. Its summary reports productivity gains that are real but often unverified and uneven. It says large-scale displacement remains limited in the evidence reviewed, and worker-reported time savings of a few per cent of working hours have not yet translated into higher measured output, earnings, or employment. These early observations do not validate the 2025 variability statistic. It also flags inequality, younger workers’ opportunities, and work organization. It neither proves displacement cannot grow nor predicts an outcome for one occupation or employer.

For a career decision, use a high or mixed profile to choose a task to inspect, not as a replacement signal or proof of safety. Check whether the occupation’s listed tasks match your duties, then check workplace evidence: which tool is used, what work step it changes, who checks outputs, how exceptions are handled, and whether demand or staffing has changed. Without that evidence, the index describes potential exposure, not the result for your job. Base a larger move on evidence about your work and constraints, rather than variability alone.

Sources: How might generative AI impact different occupations? (ILO occupational exposure chart explainer); The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence

How should a worker translate the index into a task-level check?

Use the ILO index to identify a task worth examining, then check whether that task actually appears in your work and how it is done. The occupation-level score is a starting point for a question, not a finding about your particular job. The ILO’s “How might generative AI impact different occupations?” explains that its chart scores tasks in standardized occupations and summarizes their mean and dispersion. Its accompanying Working Paper 140 describes task-level inputs and a revised exposure framework. Neither source supplies a time-use diary for your role or evidence that your employer has adopted a tool. Translation therefore requires three separate checks: what potential the index assigns to listed tasks, how closely those tasks match your own duties, and whether a tool is actually being used in a way that changes the workflow. Start with one concrete task rather than the job title. For a recurring report, for example, write down what information arrives, what the report must contain, how often you prepare it, and who relies on it. Compare those steps with the relevant occupation’s listed tasks. This is an illustration of how to inspect a task bundle, not a documented case from the ILO research. Then check fit and frequency. Does the proposed system have access to the necessary inputs? Is the output format predictable? Which parts are repeated, and which depend on local context or an exception? Record how often the task occurs, but do not let frequency stand in for importance. A task performed rarely may still carry substantial responsibility if an error affects a client, a payment, or a compliance decision. The ILO index’s task-score spread does not encode how often a particular worker performs each listed task, how much time it takes, or what is at stake when it is wrong. Your own task inventory adds information that the occupation-level summary cannot provide. If a tool is approved for a small trial, keep a simple record before and during it. Note the usual time to complete the task, the time spent reviewing and correcting an assisted output, the kinds of errors or omissions found, and the points that still need escalation to a person. Record who checks the result and who remains accountable for the final decision. These are proposed observations for a local workflow check; they are not measures reported by the ILO index, and a small trial cannot establish what an entire occupation or labor market will experience. This record can make a manager conversation more specific: the task may be technically suited to assistance, yet still require source verification, exception handling, or accountable sign-off. The practical outcome is a better-defined work question and a clearer basis for discussion, not a personal job-loss estimate.

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

What is the index good for—and what is the fair verdict?

The ILO’s 2025 index is useful as a comparative map of potential GenAI exposure across standardized occupational task profiles. Its mean summarizes the average task score within an occupation; its standard deviation shows how dispersed those scores are around that mean. The chart explainer groups occupations using both dimensions, so spread adds information when task scores are mixed. Its verdict is narrow: spread across a classified task list. It does not describe how duties differ among people sharing a title, and it is not an individual assessment, adoption tracker, or labor-market forecast.

Consider two illustrative occupations with the same mean. In one, listed task scores might sit close to that average; in another, some might be much higher and others lower. The mean alone makes the profiles look alike, while the spread flags a difference worth investigating. This explains the statistic, not actual ILO occupation values. Standard deviation also cannot identify which tasks score highest; that requires task-level results. A gradient compresses mean and variability into a category. It helps compare profiles, but cannot explain the work. The ILO’s 2025 update says its four exposure gradients depend on both mean exposure and task variability. The chart explainer distinguishes low-exposure, mixed-task profiles from high, more consistent exposure. A gradient remains a summary of an occupational model, not a verdict about a worker.

Compare the standardized list with a worker’s actual duties. The chart covers 436 detailed occupations in ISCO, the International Standard Classification of Occupations. Its unit is neither employee nor workplace. A listed task may be absent, rare, central, or performed with different accountability in a particular job. Mean and spread do not encode time spent, frequency, sequence, consequence, or who signs off. The 2026 ILO paper “Disruption without dividend?” analyzes skills-survey data for 46 countries and reports that workers in developing countries perform fewer non-routine analytical tasks even within occupations classified as highly exposed. This supports the caution that occupational titles need not mean identical task content across countries; it does not revise every job or country’s score. Compare the list with assigned work and local differences.

Potential exposure is also separate from adoption and employment outcomes. The ILO chart explainer calls its estimates potential exposure and an upper threshold conditional on full implementation. It identifies infrastructure, skills, cost, and operational difficulties as barriers. Feasibility does not show a suitable system is available or used. Adoption alone does not establish redesign, changed output, or labor demand. The separate 2026 brief “The impact of GenAI on jobs, productivity and work organization” synthesizes experiments, firm data, platform studies, and worker surveys across seven countries. Its summary describes gains as uneven and often unverified, and large-scale displacement as limited in the reviewed evidence. That concerns workplaces, not validation of the statistic.

The 2023-to-2025 comparison needs similar restraint. The ILO’s 2025 update reports mean automation scores of 0.30 in 2023 and 0.29 in 2025, with standard deviations of 0.30 and 0.14. These are separate exercises, not repeated measures of the same workers or workplaces. The brief describes the 2025 method as refined, drawing on human and AI input and broader task coverage; 2023 belongs to an earlier framework. The lower 2025 spread cannot establish that work became more uniform over time. Changed coverage and scoring make them non-equivalent snapshots; summaries do not isolate each change’s contribution.

The fair verdict: variability usefully supplements an occupation-level mean by showing whether assessed task-potential scores are concentrated or mixed within the standardized bundle. Use it to frame a question, not answer a personal career decision. More specific interpretation would require task-level frequency and importance, local mapping to actual duties, representative workplace adoption measures, and longitudinal employment evidence. Until then, cite the index as a description of variation in assessed task potential within a classification. Check real duties and workplace evidence before making a larger career decision.

Sources: How might generative AI impact different occupations? (ILO occupational exposure chart explainer); Generative AI and jobs: A 2025 update; Disruption without dividend? - How the digital divide and task differences split GenAI’s global impact; The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence

What is the next useful step?

The next useful step is small: choose one recurring or consequential task from the occupation description and compare it with the work you actually do. The ILO’s “How might generative AI impact different occupations?” explains that its scores describe potential automation across tasks in standardized occupations. They do not tell you whether your employer uses a tool, how often you perform a task, or how much of your week it occupies. Treat the index as a prompt to investigate, not as a personal diagnosis. Write down the task’s inputs and expected output. Then note how often it occurs, what exceptions require judgment, who checks the result, and who remains accountable if it is wrong. Separate three observations: whether the index lists potential for the task; whether a tool available to you can handle the real inputs and standards; and whether your workplace has adopted it in a way that changes the workflow. If a limited trial is permitted, record time saved alongside corrections, errors, review effort, and escalations. Frequency and consequence are different: a rare task can still carry important responsibility. The free task-level change-pressure checker can help organize these questions, but its result is not a validated probability of displacement. If you later need to compare staying and redesigning your role with adjacent or larger changes, the career roadmap is a separate option. For now, verify one task against your own duties and local workflow before making a larger career decision.

Sources: How might generative AI impact different occupations? (ILO occupational exposure chart explainer)

Questions readers ask