What AI Exposure Indexes Cannot Tell You About Your Job

What can an exposure index tell you about your current job?

An AI exposure index can flag that tasks linked to your occupation overlap with capabilities described in a measure, or activity recorded by a provider. It cannot tell whether your employer will adopt a tool, which duties fill your week, how much checking the output requires, or whether staffing or demand will change. Treat the score as a prompt to inspect work, not as a forecast about your job. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains that exposure indicators describe technological susceptibility through a view of tasks. They do not establish whether automation is profitable, whether workflows will be redesigned, or what happens to employment, wages, or demand. The U.S. Bureau of Labor Statistics makes the boundary explicit in “Artificial Intelligence (AI) exposure categories”: its categories do not directly observe whether workers in an occupation use AI at work, and exposure does not itself imply job loss or productivity effects. Consider a reporting task as an illustration, not a study: a score might prompt you to inspect a summary. Your next questions are whether inputs are approved, a tool is available, and how much correction and sign-off remain. An occupation-level result cannot supply those local facts. A high score earns attention, not alarm; a low score does not establish immunity. Before a costly change, inspect one task and the evidence around it. Decide whether to monitor, test a permitted workflow, or investigate a broader move. An index helps choose where to look; your task mix and workplace evidence must show what is changing.

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; Artificial Intelligence (AI) exposure categories

Which kind of index are you looking at?

Capability-based indexes estimate how closely AI capabilities match the tasks, abilities, or work activities associated with an occupation. Observed-use measures start from activity visible in a defined product or dataset and map that activity to occupational work. A combined index does not make their values interchangeable: one may describe potential, another observed activity, and a third a relative category. The OECD AI exposure measure is explicitly forward-looking. Its 2026 paper maps capabilities across nine cognitive, social, and physical domains to occupational requirements, producing an AI Capability Gap index intended to help assess possible effects over five to ten years. A smaller gap means current AI capabilities are closer to the profile required by an occupation. The OECD reports closer matches for routine information processing and codifiable tasks, and larger gaps for work requiring contextual judgment, interpersonal understanding, complex decisions, and responsibility. This is a capability comparison with a stated horizon; the publication also says actual effects depend on adoption, regulation, organizational change, and social choice. It does not measure current use in your workplace. The ILO’s Generative AI and Jobs: A Refined Global Index of Occupational Exposure takes a more task-informed route. Its 2025 study combined a representative sample of tasks from Poland’s occupational classification, a survey of 1,640 employed people across ISCO-08 groups, international expert discussion, and model-assisted scoring. The publisher reports 52,558 data points covering perceived automation potential for 2,861 tasks, then groups results into four increasing exposure gradients. These global estimates still summarize occupations and populations, not how one worker’s duties are divided. Anthropic’s Labor Market Impacts measure illustrates observed use. It counts tasks as covered when they are theoretically within LLM capability and have sufficient work-related use in Claude traffic; it gives full weight to automated uses, half weight to augmentative uses, then averages by estimated time spent on tasks. The study reports 33% coverage for tasks in Computer and Math occupations under this method. It describes mapped use in one provider’s data, not representative employer adoption or proof that tasks have been removed from jobs. The BLS AI exposure categories deliberately combine evidence types: three theoretical sources and two measures of observed product use mapped to occupational tasks or activities. BLS converts each source’s occupation scores to within-source percentiles, combines theoretical and observed dimensions, and groups occupations into four relative categories. The result compares occupations within that U.S. system; its category is not an absolute exposure quantity. Before reading any score, check the measure, data owner, date, geography, task mapping, and whether the output is a raw value, percentile, or category. That small audit tells you what question the index can help you ask.

Sources: The OECD AI Exposure Measure; Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Labor Market Impacts of AI: A New Measure and Early Evidence; Artificial Intelligence (AI) exposure categories

What the task list can miss inside one occupation

An occupation label is a useful filing system, not a description of every worker’s week. Two people with the same title may spend different amounts of time drafting, checking, coordinating, resolving exceptions, or explaining results. An occupation-level score cannot show which tasks dominate your role or which can be transferred reliably in your workplace. Consider an illustrative reporting workflow. One task might be turning approved, structured figures into a first draft. Others might include reconciling conflicting records, tracing an unusual variance, explaining what changed, or signing off on the report. A tool that helps with a draft does not establish that the rest of the workflow can be handed over; review and responsibility remain part of the work. A task list is therefore a proxy. The International Labour Organization’s *Workers’ exposure to AI: What indicators tell us – and what they don’t* says exposure measures rely on static lists of existing tasks and omit constraints such as economic conditions and institutional barriers. The list can describe possible susceptibility, but not how a particular team divides work or whether a firm can change its workflow. The ILO also notes substantial variation within occupational groups, cautioning against treating an occupation average as every worker’s task bundle. The ILO’s *Generative AI and Jobs: A Refined Global Index of Occupational Exposure* shows that richer task evidence still has a defined scope. For its 2025 index, the study combined a representative sample from Poland’s 29,753-task occupational classification with a survey of 1,640 employed people across ISCO-08 groups. It gathered 52,558 data points on perceived automation potential for 2,861 tasks, This detailed basis still produces occupational exposure gradients, not observations of each reader’s duties, employer, or team. Mappings add another layer. The U.S. Bureau of Labor Statistics explains that source measures are crosswalked to occupational classifications, and that source coverage and crosswalks can affect category assignments. It also says the inputs treat occupations as bundles of task or ability characteristics without accounting for bottlenecks or complementarities. A task may depend on context, upstream exception handling, or an authorized approver; naming it does not capture those dependencies. Translate the result into a small inventory: list repeated tasks, note their inputs and how often exceptions arise, then mark who reviews and who is accountable for sign-off. It makes the next question concrete: which part of your workflow is changing, and what evidence at work would confirm it?

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Artificial Intelligence (AI) exposure categories

Why can two indexes agree on direction and still give different scores?

Two AI exposure indexes can point toward the same occupations while disagreeing about the size of exposure. In “Labor Market AI Exposure: What Do We Know?”, The Budget Lab compares seven measures across 710 U.S. occupations covered by every measure. It finds broad agreement about relative direction, but greater disagreement among occupations rated highly exposed. They signal where AI could matter more consistently than they quantify how much. Computer programmers illustrate the gap. In the comparison, programmers rank near the high end across measures, but their placement varies: one measure puts them at the 99th percentile of exposed occupations, another at the 88th. The Budget Lab reports programmer score variance of about 0.5, near the 88th percentile of occupation-level variance; plumbers are near the bottom, with variance close to zero. The broad direction may be more stable than the distance implied by the scores. A worker comparing occupations separated by a few points should not treat that gap as a precise difference in likely workplace change. The measures differ in what they count and how they classify work. The Budget Lab comparison includes task ratings, ability ratings, patent-to-task links, and observed Copilot conversations. It crosswalks different occupational classifications to a shared U.S. system, but the measures remain distinct. Measures also began on different scales, and the appendix notes one outlier was excluded from later tests. Results therefore depend on the selected data and method. The Bureau of Labor Statistics provides a separate example of why a composite needs context. Its 2026 categories combine five external sources, convert source values into within-source percentiles, impute some missing occupation-source values, and assign four relative categories. A percentile describes position within that source and occupational universe; it is not a universal unit directly comparable with another index’s score. BLS documents these steps so readers can interpret the output, not mistake it for a direct measurement of an individual worker. Read each score within its method and date. Check what was measured, which occupations were covered, and whether the result is a rank, percentile, standardized score, or category. If a decision turns on a small difference between occupations, inspect the task descriptions and uncertainty first. Disagreement reduces confidence in the size of a gap; it does not erase the screening signal that recurring digital tasks may deserve closer attention.

Sources: Labor Market AI Exposure: What Do We Know?; Artificial Intelligence (AI) exposure categories

What does a polished or official score still leave out?

Transparency can make a measure easier to audit without making it personal. The U.S. Bureau of Labor Statistics (BLS) says its 2026 AI exposure categories combine five external sources: three theoretical measures and two based on observed AI use mapped to occupational tasks or activities. BLS groups occupations into four relative categories. It states that the result does not directly observe whether workers in an occupation used AI on the job, and that exposure does not imply job loss, productivity gains, automation probability, or wage effects.

The category is built through choices, not read directly from a worker’s day. BLS maps source results to its occupational classification, converts scores into percentiles within each source, combines theoretical and observed dimensions, then clusters occupations into low, moderate, high, and very high groups. Percentiles put source scores on a shared relative scale; they do not make the inputs measurements of the same thing. “High” means relatively high within the comparison, not an absolute amount of work that will be automated or a forecast that employment will fall.

Coverage has boundaries too. BLS notes that some inputs do not cover every occupation. Of 4,155 possible occupation-source combinations across 831 occupations and five sources, 3,944 were observed and 211 imputed, affecting 75 occupations. The agency identifies the occupation universe, sources, crosswalks, imputation, normalization, and category choices as limits. These disclosures make the measure inspectable: readers can see where values came from and what processing shaped them. They do not reveal which duties fill a particular worker’s week, how often exceptions arise, or who must check and approve the result.

That distinction matters for a current-role decision. The International Labour Organization’s brief, “Workers’ exposure to AI: What indicators tell us – and what they don’t,” explains that exposure measures generally rely on static task lists and omit adoption constraints such as economic conditions and institutional barriers. An occupation category cannot show a team’s policy, whether an approved tool fits its workflow, the cost of correcting errors, the quality threshold for a deliverable, or how a manager allocates tasks. These factors can affect whether a capability becomes routine work, occasional assistance, or is not used in that setting. This is an inference from the measures’ stated boundaries, not a report about any particular workplace.

An official score helps organize questions when transparent. It cannot replace local evidence. After reading the category, inspect one recurring task: its inputs, judgment and exceptions, required review, and accountable sign-off. Then look for permitted, repeated use and changed handoffs. The score can tell you where to look; task and workplace context help establish what is changing in your role.

Sources: Artificial Intelligence (AI) exposure categories; Workers’ exposure to AI: What indicators tell us – and what they don’t

What observed use adds—and what it still cannot show

Observed use narrows the distance between theoretical capability and at least some real activity, but a provider’s traffic is not a measure of all workplace use or of use in your team. Even mapped product use does not establish permission, successful integration, task reassignment, or an employment effect.

In “Labor market impacts of AI: A new measure and early evidence,” Anthropic combines task ratings of what a large language model could speed up with work related Claude usage mapped to O*NET occupational tasks. Its method gives more weight to automated use than augmentative use, then weights task coverage by the estimated share of time spent on those tasks to produce an occupation measure. The report says Claude activity covered 33% of tasks in the broad Computer and Math category. That figure describes coverage in the Anthropic Economic Index usage data used for its analysis, including Claude.ai conversations and API traffic, with usage from August and November 2025. It is not the share of all workers in that category using AI, nor the share of any one worker’s week that has changed.

The measure is informative precisely because capability and observed activity do not match one for one. Anthropic notes that a theoretically feasible task may not appear in its usage data because of model limits, legal constraints, software requirements, verification steps, or other hurdles. Its analysis uses one provider’s activity, a particular mapping to task descriptions, and judgments about what counts as significant, work related use. The traffic is not representative of the full labor market or an employer’s rollout.

The U.S. Bureau of Labor Statistics adds both Anthropic’s Claude measure and a Microsoft Copilot measure to its current evidence inputs, mapping product activity to occupational tasks or work activities. The BLS page is explicit that these mapped measures “do not directly observe whether workers in a particular occupation used AI on the job.” The mapping helps compare evidence at an occupation level; it cannot establish whether your employer permits a tool, whether your team has integrated it into a recurring process, or whether it is approved for your data and output.

For a current role, separate four observations: activity in a provider’s data, an employer approved pilot, repeated use in a team workflow, and a changed allocation of tasks or review responsibility. A pilot may show assistance under set conditions; repeated use suggests integration; reassignment or changed review indicates a workflow shift. None determines staffing or future demand. Ask whether your organization permits the tool and whether the task is actually being handed off, or whether the tool supports work that you still verify and own.

Sources: Labor Market Impacts of AI: A New Measure and Early Evidence; Artificial Intelligence (AI) exposure categories

What labor-market evidence can check an index—and what can it not settle?

Labor-market studies can test whether group outcomes are changing, but cannot turn an occupation score into a personal forecast. “Labor Market Impacts of AI: A New Measure and Early Evidence” (Anthropic) found no systematic unemployment increase among workers in its most exposed U.S. occupations since late 2022. It also reported a qualified signal of slower hiring into highly exposed occupations for workers aged 22 to 25. “What We Do and Don’t Know About How AI Is Affecting the Labor Market” (The Budget Lab at Yale) found no statistically or economically significant aggregate employment or wage effects through the first quarter of 2026. Neither says what will happen to a particular worker.

Anthropic matched occupational task measures to U.S. Current Population Survey records, comparing unemployment trends for workers in the highest observed-exposure quartile with workers whose tasks had no measured coverage. Its difference-in-differences estimate of the average unemployment gap was small and statistically insignificant. The study also examined job starts among 22-to-25-year-olds, estimating a 14 percent post-release decline relative to 2022 in entries to highly exposed occupations, barely statistically significant. The authors note other explanations, including workers staying put, entering other occupations, returning to school, or survey measurement error. This signal merits monitoring, but does not show that young workers as individuals are being displaced.

The Budget Lab used synthetic differences-in-differences to compare employment shares and real hourly wages in U.S. occupations in the top third of its exposure measure with a weighted comparison group. It found no clear employment effect and no statistically or economically significant wage effect. This weighs against claims of broad displacement already visible in aggregate data. Yet an average can conceal opposing changes: demand may expand in some exposed occupations and contract in others, offsetting in the combined result. The Budget Lab says the Current Population Survey is better suited to broad groups than narrow subgroups, making it less able to detect changes concentrated among a small slice of workers. A null average does not establish that every occupation is unchanged.

Use these findings as context for monitoring, not a verdict on your role. They are U.S. group-level comparisons over an early adoption period, with different measures and limits; neither tests your employer’s workflow, task reassignment, or local demand. A high score does not become a personal danger signal because one subgroup result is suggestive, and a null aggregate result does not prove safety. Look for evidence in your work: whether an approved tool is repeatedly used on a task, whether handoffs or review duties have changed, and whether work is being reassigned. Studies show broad patterns; ground action in observable local changes.

Sources: Labor Market Impacts of AI: A New Measure and Early Evidence; What We Do and Don’t Know About How AI Is Affecting the Labor Market

How can you test whether your work is changing before making a career move?

Pick one repeated task and compare the current process with an approved AI-assisted process, if available. Record what the tool handles and what still needs correction, exception handling, verification, or accountable sign-off. A demonstration that a system can produce an output shows capability under those conditions. It does not show that your employer permits the tool, that a team uses it repeatedly, or that anyone's assignment has changed. The International Labour Organization's brief, “Workers’ exposure to AI: What indicators tell us – and what they don’t,” says exposure measures describe potential task susceptibility, not whether firms adopt tools or how workflows and employment outcomes adjust. The U.S. Bureau of Labor Statistics likewise says its current-evidence measures map observed AI interactions to occupational tasks; they do not directly observe whether workers in a particular occupation used AI on the job.

Note how often the task arises, what information it requires, where exceptions occur, and what quality check follows. Establish a baseline: where time goes, which revisions matter, and what context the task requires. If workplace rules allow a trial, use approved tools and low-risk material. Compare finished work, not only the first output, and include checking and correction time. A faster first draft may not save time if review expands or errors create more work. This describes one workflow at one point, not future staffing or demand.

When use is restricted, data is sensitive, or an error could have serious consequences, do not route around the rule. Ask what tool, data, approval, and review process are permitted. Capability without permission or workable review is not current adoption in your role. BLS separates theoretical exposure from current-evidence inputs and notes that mapped activity does not reveal a workplace's rules or handoffs.

Look for observable changes: approved use, recurring use in the task, a changed handoff or review, and an assignment actually removed, added, or redistributed. These are practical signals, not a formula predicting job loss. If only a demonstration exists, monitor or learn. If repeated use changes your process, consider a bounded workflow upgrade and strengthen the skill needed to check results. This evidence can inform a decision before you pay for training or leave a field, but one task cannot settle what will happen across a whole role or labor market.

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; Artificial Intelligence (AI) exposure categories

When does a score justify monitoring, a workflow upgrade, or a larger move?

A score alone justifies monitoring, not a resignation, degree, or costly retraining plan. A reversible workflow upgrade becomes more reasonable when a task you actually perform is repeatedly handled with an approved tool and you can check the result. A larger role or learning move deserves investigation when important duties are being reassigned or the need for their output is shrinking, and when an adjacent path fits your income floor, location, available time, health, and family responsibilities. These are evidence thresholds for deciding what to examine next, not forecasts. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains why an index cannot set those thresholds by itself: exposure measures describe possible task change, while adoption constraints, workflow redesign, employment, wages, and demand remain separate questions. The Budget Lab’s “What We Do and Don’t Know About How AI is Affecting the Labor Market” adds a useful counterweight. Its synthetic differences-in-differences analysis of U.S. employment and wages across more- and less-exposed occupations found no statistically or economically significant effects so far. The report also cautions that occupation-level exposure measures are imperfect for identifying negative demand shocks. Neither result settles an individual worker’s situation. Match your response to evidence and reversibility. With only an occupation score, monitor for better task or workplace evidence. If your employer permits a tool, compare an approved workflow with the existing process, including corrections, exceptions, review, and sign-off. If substantial duties are repeatedly removed, or the organization needs less of the output, investigate role redesign and adjacent options before committing resources. Persistent human review or continued demand may support staying and upgrading, but neither guarantees security. Stronger task-level evidence plus verified local adoption and changed assignments is more relevant than an occupation average; even that combination cannot establish future staffing or demand. Reassess when repeated evidence changes; keep the move proportionate to your constraints.

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

What is a proportionate next move when the score is all you have?

If an occupation score is all you have, treat it as a prompt to observe one task, not a forecast about your job. Write down a recurring task, what information it uses, where exceptions arise, how much review it needs, and who remains accountable for the result. Then check which tools your workplace approves and watch whether they are actually used and whether work is reassigned. It is a prompt to revisit, not a forecast. That restraint follows the International Labour Organization brief “Workers’ exposure to AI: What indicators tell us – and what they don’t”: exposure measures describe technological susceptibility, but do not establish whether firms find automation profitable, how workflows change, or how employment, wages, and demand adjust. Its practical implication is to pair an index with local evidence, not to ignore the signal or let it decide for you. If you want help organizing the task-level signals, the free checker can support that first review; it does not provide validated odds of displacement. A larger move can wait until repeated changes in duties, review, or assignment give you something specific to compare against your income, location, time, health, and family constraints.

Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t

Questions readers ask