How to Compare AI Exposure Scores With Different Task Definitions

How to compare unlike AI exposure scores: the short answer

Compare definitions and task evidence before comparing numbers. If two AI exposure measures count different things, such as tasks, abilities, or work activities, or use different AI scopes and aggregation rules, their values are not directly comparable. A shared label or normalized scale does not establish that they measure the same quantity. Identify what each score includes and how its inputs are combined.

Next, align what can be aligned: the assessed unit, AI capability, time reference, and how evidence becomes a score. Check representative duties against your recurring work, including important tasks either measure omits. The Bureau of Labor Statistics’ “AI Exposure Categories” explains that its framework combines distinct theoretical and observed-use inputs, ranks them within sources, and groups occupations. Those choices describe that framework, not a universal personal measure. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us—and what they don’t” cautions that exposure indicators do not establish employment outcomes.

If definitions and units genuinely match, compare numbers only within that shared scope. If they differ, compare directionally where overlap is clear; otherwise read each score separately. Do not average unlike values or assume the higher one is truer. Treat disagreement as a reason to inspect tasks and consider one small, reversible observation. An exposure score does not forecast your personal job loss.

Sources: AI Exposure Categories; Workers’ Exposure to AI: What Indicators Tell Us—and What They Don’t

What question does each kind of score answer?

Before reading a score, ask what relationship it measures. A task automation or time-saving measure asks whether a defined AI capability could assist with, complete, or reduce time on specified tasks. An ability-based measure asks whether AI capabilities correspond to abilities required in an occupation. An observed-use measure starts with activity recorded in a particular product or service, then maps it to occupational tasks or work activities. They answer different questions; their values are not competing estimates of one quantity. The distinction appears in sources combined by the U.S. Bureau of Labor Statistics (BLS). Its methods page describes one theoretical approach that asks whether AI applications relate to occupational abilities. Other inputs assess whether large language model capabilities could reduce time spent on occupation-specific tasks. The observed inputs use product activity mapped to tasks or work activities. BLS says these observed measures do not directly show whether workers in a given occupation used AI on the job. Product interactions mapped to a duty are evidence about activity in that data source; they do not establish that every employer adopted the tool or that a whole job can be performed without human contribution. The measure’s definition and mapping matter as much as its label. The OECD’s 2026 paper, “The OECD AI exposure measure: Mapping the OECD AI Capability Indicators to occupations,” illustrates this distinction. It maps AI capability indicators across nine cognitive, social, and physical domains to occupational requirements and constructs an AI Capability Gap index. A lower gap means current AI systems are closer to the capability profile required by an occupation, indicating higher potential exposure. The paper gives the measure a forward-looking, five-to-ten-year analytical orientation. That horizon describes its stated purpose; it is not a timetable for a particular employer or worker. The gap score is not a count of tasks already automated or a record of observed workplace use. These definitions clarify why carefully constructed scores may answer different questions. One may estimate potential assistance or time reduction across listed tasks; another may compare capabilities with occupational requirements; a third may summarize where a product’s use data maps onto work. A task can be technically assistable without the tool being used in a workplace. A capability match does not show that the relevant task dominates someone’s week. Observed use does not establish broad adoption, changed staffing, or displacement. Translate each score into a sentence beginning, “This measure asks whether…” If the sentence names a different unit or relationship, the scores are not direct numerical rivals. They may still provide context, but compare their definitions with recurring duties before deciding what they mean for your work.

Sources: The OECD AI Exposure Measure: Mapping the OECD AI Capability Indicators to Occupations; AI Exposure Categories

Why a shared scale does not make scores equivalent

Putting two exposure results on a 0–1 scale or giving each a percentile makes them easier to display side by side. It does not make them measurements of the same thing. A percentile answers a relative question: where does an occupation’s value sit among those covered by this measure? It does not mean this share of tasks is exposed or equals the same percentile elsewhere. The Bureau of Labor Statistics’ method makes the distinction concrete. BLS combines five inputs that, it says, do not all measure equivalent concepts. One input relates AI to abilities; others assess possible task time reduction. Observed-use measures map product activity to work tasks. BLS calculates each occupation’s percentile within each source, putting those within-source ranks on a 0–1 scale. It then takes separate medians for theoretical and observed inputs before assigning four relative categories. The display scale does not erase the inputs’ different questions. Comparability has three layers. The construct is what a measure represents: capability matched to abilities, potential time reduction on tasks, or observed work activity. The bridge connects evidence to occupations through classifications, crosswalks, coverage, and missing-value treatment. The scale turns evidence into a number or category, such as a raw rating, rank, median, or threshold. Matching one layer cannot fix mismatch at another: shared occupation codes and percentile formats can still represent different constructs and source populations. Those choices shape the label. BLS says categories depend on its occupation universe, five inputs, crosswalks, imputation, normalization, and categorization. It imputed missing source ranks for 75 occupations, rather than directly observing every input for them. BLS also says categories are relative to other occupations, not absolute levels. A “high” label needs its comparison universe and construction beside it. The Budget Lab’s comparison of seven measures shows the effect. After crosswalking classifications to SOC 2018, researchers analyzed 867 occupations with at least one score; 710 had all seven. They found broad agreement about which occupations showed exposure, but more disagreement about magnitude among highly exposed occupations. Computer programmers ranked near the top across measures, yet their reported percentile varied: the GenAI Total measure placed them at the 99th percentile of exposed occupations, while the AI Applicability Score placed them at the 88th. This does not estimate an individual programmer’s duties or prospects. Crosswalking enabled comparison without making the measures identical. Shared percentiles do not establish a common denominator; coverage, inputs, and transformations may differ. Before comparing magnitudes, check what each score ranks and how it was constructed. If details align, a bounded comparison may help. Otherwise, treat the label as a within-source ranking and inspect the task definitions. A score alone cannot say how much of one worker’s job is exposed.

Sources: AI Exposure Categories; Labor Market AI Exposure: What Do We Know?

Which parts of two measurement contracts must match?

For a numerical comparison, check the measurement contract before looking at whether one score is higher. Check assessed unit, AI scope, occupational mapping and geography, time reference, and aggregation rule. If a difference could change what work is counted or how it is counted, write it down. A matching display does not repair a mismatch in meaning. Use this compact audit for each source: • Unit: Does it score task statements, required abilities, work activities, or an occupation-level bundle? BLS combines ability associations, task-level time-reduction judgments, and observed product interactions mapped to tasks or activities. The OECD AI exposure measure instead maps AI capability indicators across nine cognitive, social, and physical domains to occupational requirements and builds a capability-gap index. Do not silently translate between them. • AI scope: Does the measure concern generative language models, broader AI capabilities, or embodied systems such as robotics and machine vision? Does exposure mean assistance, completion, or time reduction under stated assumptions? Does it include tasks needing additional software? BLS documents these distinctions. The OECD measure explicitly includes multiple capability domains, so it may cover a wider range of technologies than a language-model-only estimate. If unclear, record “not stated.” • Bridge: Which occupation classification, code version, and geography map tasks or requirements to a job title? Are two scores based on the same occupation code, or has one been crosswalked from another system? A crosswalk aligns records; it cannot guarantee identical underlying work. National taxonomies do not directly describe an employer’s role. • Time and aggregation: Is it tied to a capability date, observed-use window, or forward-looking horizon? Are tasks weighted equally, by importance or time, or combined through a median, sum, or threshold? BLS explains that its percentile ranks are calculated within each source, then combined into separate theoretical and observed medians before occupations are grouped. The result is a relative category, not a shared absolute unit. The OECD describes a forward-looking measure with a five-to-ten-year analytical horizon; not a timetable for a particular worker. Finally, record whether the displayed number is absolute or relative to that source’s occupation universe. Compare numerically only when definitions, mapping, scope, time basis, and aggregation support the same interpretation. If one material element differs, compare direction cautiously and name it. If essential method details are unavailable, keep the scores separate. The ILO record, “Generative AI and jobs: a refined global index of occupational exposure,” verifies a 2025 working paper but provides too little accessible scoring detail to fill unknown fields. Unknown is a useful result: it prevents a neat-looking score from being mistaken for comparable evidence.

Sources: AI Exposure Categories; The OECD AI Exposure Measure: Mapping the OECD AI Capability Indicators to Occupations; Generative AI and Jobs: A Refined Global Index of Occupational Exposure

Why can two well-built scores disagree?

Scores can disagree because they start with different units and ask different questions. One may list occupational tasks and ask whether a language model could reduce the time they take; another may map AI capabilities to abilities required for an occupation; a third may map observed product activity to work. The OECD AI Exposure Measure (2026) maps capabilities across cognitive, social, and physical domains to occupational requirements and builds a capability-gap index. The BLS methods page describes theoretical inputs based on ability associations or task time-reduction judgments, alongside observed-use inputs mapped to tasks or work activities. Differences may reflect scope, not calculation error. The distinction between assistance and task completion also matters. A tool might help draft a spreadsheet summary without checking data, coordinating exceptions, or owning final figures. BLS describes source measures that distinguish direct exposure from tasks requiring additional software, with different weights assigned to software-enabled exposure in one approach. A measure of possible assistance can rank work differently from one asking whether AI performs a meaningful part of an activity. “Could help” and “can complete” are separate tests. AI scope creates another divergence. A language-and-reasoning index notices different capabilities from one that also includes machine vision, robotics, or embodied AI; the OECD measure explicitly treats exposure across those domains as multidimensional. Weights and aggregation matter too. BLS cautions that its inputs represent occupations as bundles of independent characteristics and do not account for bottlenecks or complementarities. The ILO’s Workers’ exposure to AI: What indicators tell us – and what they don’t says measures differ conceptually and rely on static task lists, and miss changing workflows. Imagine an operations analyst who drafts spreadsheet summaries, coordinates handoffs, investigates exceptions, and owns the final figures. One index might emphasize drafting and spreadsheet work; another might represent coordination, exception handling, or accountability differently, or weight those activities differently. Results could diverge even when both apply their rules consistently. This is an illustration, not a reported case or a claim about any index’s score for that occupation. Ask which description of work would make each result fit. When scores diverge, check the descriptions against recurring duties before choosing a winner. Ask whether each measure tests assistance, time reduction, or performance; which AI capabilities it includes; and how it combines tasks. The OECD measure notes that actual impacts also depend on adoption and organizational change; the ILO account says exposure indicators do not establish labor-market outcomes. Investigate the tasks; do not average unlike values or infer personal job-loss odds.

Sources: The OECD AI Exposure Measure: Mapping the OECD AI Capability Indicators to Occupations; AI Exposure Categories; Workers’ Exposure to AI: What Indicators Tell Us—and What They Don’t

What cross-measure research says about disagreement

In its February 2026 comparison, [“Labor Market AI Exposure: What Do We Know?”](https://budgetlab.yale.edu/research/labor-market-ai-exposure-what-do-we-know) examines seven occupation-level measures and finds an asymmetry: they generally agree which occupations show some exposure, but diverge more over its magnitude among highly exposed occupations. The analysis converted different occupational classifications to 2018 Standard Occupational Classification codes, leaving 867 occupations with at least one score and 710 covered by all seven. The crosswalk enabled comparison, but did not make the measures ask identical questions. The researchers standardized scores to z-scores, then used a principal-component-weighted summary and cross-measure variance to examine disagreement. Variance rose moderately with the summary exposure score. This is a pattern in occupation data, not a reading of an individual worker or future. The computer-programmer example is concrete. In the study, programmers ranked near the top overall, yet their cross-measure score variance was about 0.48; two measures placed them at the 99th and 88th percentiles, respectively. Those percentiles show positions within each measure’s occupational distribution. They do not mean that programmers’ tasks are 99 or 88 percent automatable, nor do they establish a personal probability of job loss. A worker may encounter a high signal and disagreement without either being a clerical mistake. The result supports inspecting definitions and mappings before interpreting the gap. The International Labour Organization’s April 2026 brief, [“Workers’ Exposure to AI: What Indicators Tell Us—and What They Don’t”](https://www.ilo.org/publications/workers%E2%80%99-exposure-ai-what-indicators-tell-us-%E2%80%93-and-what-they-don%E2%80%99t), helps explain why gaps can persist. It says exposure measures differ conceptually, rely on static task lists, embed judgments, and leave out economic conditions and institutional barriers to adoption. It also stresses that technological susceptibility is not an employment outcome: a score does not show whether a firm will find adoption worthwhile, redesign a workflow, or change staffing. The brief summarizes limitations; it does not validate an individual forecast. A counterpoint: harmonizing occupation codes and, where possible, task units can reduce noise from mismatched classifications. Yale’s crosswalk is a practical example of that work. But a common code is only a bridge between occupational labels. It cannot erase differences in what counts as exposure, which tasks or capabilities enter, or how they are weighted. For a worker, disagreement calls for an audit: identify differently framed or omitted duties, then check whether they recur in the role. Do not average unlike scores or select the more alarming one. Agreement can help screen where to look; divergence among highly exposed occupations is a reason to seek more task-level detail, not more precise personal odds.

Sources: Labor Market AI Exposure: What Do We Know?; Workers’ Exposure to AI: What Indicators Tell Us—and What They Don’t

How should I map the scores onto my actual work?

Start with a short list of work you repeat, not with your job title alone. For each duty, note its output, frequency, inputs, where judgment enters, what an error could affect, the review required, and who remains accountable. A label such as “prepare a report” can cover assembling figures, choosing which exceptions matter, explaining results, or approving a recommendation. The score’s wording may resemble your work while leaving out the part that consumes time or carries responsibility. Compare the whole duty, including handoffs and checking, rather than matching on a verb alone.

For a U.S. occupation, O*NET provides a useful reference vocabulary. Its database separates occupation-linked task statements and ratings from work activities, and includes mappings between detailed work activities and task statements. The O*NET 31.0 page also lists emerging tasks proposed for future data collection. These resources help you inspect what a measure means by a task or activity. They are a U.S. occupational taxonomy, not a live record of your employer’s process or a complete inventory of your duties. If you work elsewhere, check whether the occupation classification fits your country and role before using it.

Add duties a standard description may miss, such as a new workflow, coordination across teams, or an unusual exception you handle. Mark these as your observations, not official O*NET entries or measured score components. Then compare each measure’s unit with each duty. The U.S. Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” describes inputs that map different theoretical measures and observed-use measures to tasks or work activities. It cautions that observed activity in product data does not directly establish that workers in a particular occupation used AI on the job. A matching label therefore does not show that your process has adopted a tool or that it can complete the full duty.

As a practical interpretation, mark whether AI might assist a bounded step, perform a step with human review, or fit poorly where physical context, trust, exceptions, or accountability dominate. These are prompts for inspection, not universal ratings. Drafting a routine summary, for instance, may be assistable while deciding whether an unusual case warrants escalation still depends on context and responsibility. Then consider which duties recur and matter: a match on a rare activity may carry little practical weight, while an omitted duty may matter greatly if it involves consequential judgment. This remains a personal comparison, not a calibrated exposure estimate. Agreement between scores matters only if they cover the same recurring work; disagreement is useful when it points to specific tasks, omissions, or mappings to investigate. Neither converts an occupational score into a claim about your individual job security.

Sources: Database; AI Exposure Categories

What does an exposure score leave unresolved?

An exposure score describes a possible relationship between AI capability and defined work. It does not tell you whether an employer will adopt a tool, whether it will work there, or how management will use saved time. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” says measures rely on static task lists, differ in definitions and judgments, and omit constraints such as economic conditions and institutional barriers. They signal task change; they do not forecast employment, wages, demand, productivity, or displacement. Keep capability and career outcome separate. Capability means a system can assist with or complete a task under stated conditions. Observed use means a dataset records tool interactions, not that workers in an occupation use it on the job. Adoption means an organization has selected and enabled a tool. Redesign concerns changes to handoffs, review, exceptions, and accountability. Labor demand concerns how much work employers or customers need; displacement is an employment outcome. Reliability, costs, rules, management choices, and demand can interrupt this chain. The U.S. Bureau of Labor Statistics explains this boundary in “Artificial Intelligence (AI) exposure categories.” Its theoretical inputs concern capabilities available no later than mid-2023; observed-use inputs may lean toward early adopters. BLS says its categories do not measure adoption probability, employment growth or decline, wages, productivity gains, or worker replacement; they also do not distinguish automation from augmentation. These limits do not show workplace change is absent. A task may become faster while workers still handle verification or responsibility. Staffing depends on factors the category does not measure. Outcome evidence does not turn exposure into a forecast. In “What We Do and Don’t Know About How AI is Affecting the Labor Market,” Yale’s Budget Lab compares employment and wages in AI-exposed and unexposed occupations using a synthetic differences-in-differences design. It reports no clear statistically or economically significant effect so far, and says this may change with new data. The analysis notes that these occupation groups differ in ways that complicate comparison; broad exposure measures can hide varied outcomes within a group. This is what the method currently detects, not proof of no future or worker-specific effects. Use a score to identify tasks worth examining. Look for evidence: approved use, remaining review and exception work, changes in duties or staffing, and what local employers request. A high score alone does not justify abandoning experience or paying for a credential. If a task is changing, a bounded learning experiment may be reasonable; a larger move should fit income needs, location, training time, health, and family responsibilities. Exposure can focus attention, but cannot settle personal and labor-market questions.

Sources: Workers’ Exposure to AI: What Indicators Tell Us—and What They Don’t; AI Exposure Categories; What We Do and Don’t Know About How AI is Affecting the Labor Market

What evidence should change the next career move?

A mismatch between exposure scores is a reason to investigate, not by itself a reason to retrain or leave a job. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains that measures vary with their methods and definitions, often rely on static task lists, and do not establish whether firms will adopt AI, redesign workflows, or change employment, wages, or demand. An index may identify tasks to watch, but cannot show whether your employer will adopt AI or your role will shrink. Look for closer evidence: a recurring duty is changing in your workflow; an approved tool or implementation plan exists; and you can see how much review, correction, and exception handling it requires. If policy permits and the work is low stakes, test one step . Record total time including verification, errors or exceptions, and whether responsibility for the result changed. For sensitive or consequential work, use approved data and supervised evaluation rather than experimenting on live decisions. The Budget Lab’s “What We Do and Don’t Know About How AI is Affecting the Labor Market” compares U.S. employment and wages in exposed and unexposed occupations with a synthetic differences-in-differences design. It reports no statistically or economically significant effects so far, while noting limits in exposure metrics and subgroup analysis. This is context, not a guarantee about a role or a substitute for local demand evidence. Let observed work guide a response: improve a workflow, build a nearby skill that uses your experience, or consider larger training only if a demonstrated gap and your salary, time, location, health, and family constraints support it. Learn verification, data handling, or workflow design when task evidence points there; do not chase a score or vendor interface.

Sources: Workers’ Exposure to AI: What Indicators Tell Us—and What They Don’t; What We Do and Don’t Know About How AI is Affecting the Labor Market

A practical next step for comparing your scores

Put each score beside its measurement contract: source, unit (task, ability, or work activity), AI scope, occupation mapping, reference date, and aggregation method. The Bureau of Labor Statistics’ “AI Exposure Categories” page shows why: its framework combines distinct theoretical and observed-use inputs and normalizes them within source. A shared category or percentile therefore does not establish shared meaning.

Next, list three to five duties that recur in your week and matter to your work. Mark each measure as covering, omitting, or leaving unclear each duty. Do not fill gaps in a source’s method by guessing. If a duty appears changeable and policy permits, choose one low-stakes step to observe: what the tool handles, what needs correction, which exceptions remain, and whether your review responsibility changes. Keep sensitive information out of unapproved systems.

If definitions and scope genuinely align, compare scores only within that matched frame. If they do not, use the difference to inspect tasks; neither reading estimates your chance of job loss. Consider a larger learning investment or career move only when task evidence, adoption or demand signals, and constraints such as income, location, time, health, and family converge. The free AI task checker offers transparent task-level change-pressure signals, not a validated displacement probability.

Sources: AI Exposure Categories

Questions readers ask