What BLS’s 2025–35 AI Exposure Categories Leave Out

What do BLS's 2025–35 AI exposure categories add—and leave out?

If your occupation receives a high AI-exposure label, treat it as a prompt to inspect your work, not a verdict about your job. BLS adds a comparison across 831 U.S. occupations, combining theoretical estimates of AI capability with observed AI activity mapped to occupational tasks. They do not measure how much of your own week is exposed. “2025–35” names the projection cycle accompanying the product. It does not forecast AI adoption throughout those years or predict that a worker will be replaced. In “Artificial Intelligence (AI) exposure categories,” BLS says exposure does not imply job loss, productivity gains, automation probability, or wage effects. Its observed measures map AI interactions to tasks; they do not establish whether workers in each occupation use AI on the job. “Employment Projections: 2025-2035 Summary” presents the categories alongside a national outlook, not a personal risk estimate. A high label prompts you to ask which recurring tasks could change and whether your workplace is changing them. Low is no promise of protection. Choosing a proportionate next move requires your task mix, evidence of actual workflow change, and labor information relevant to your location and constraints. The category starts that inquiry; it cannot finish it.

Sources: Artificial Intelligence (AI) exposure categories; Employment Projections: 2025-2035 Summary

What did BLS add to the 2025–35 projections?

For the 2025–35 projections, the U.S. Bureau of Labor Statistics (BLS) added a four-category comparison of relative AI exposure across 831 detailed occupations. Instead of relying on one capability estimate, BLS combines three theoretical measures with two measures built from observed AI activity mapped to occupational tasks or work activities. This gives readers a structured place to begin comparing occupations. It does not show how much of one worker’s week is exposed or whether an employer will adopt AI. In “Artificial Intelligence (AI) exposure categories,” BLS says the product complements career planning, while ruling out interpretations as job-loss, adoption, productivity, or wage estimates. The theoretical inputs ask different questions about potential fit. One relates AI applications to abilities used at work; two assess whether large language model capabilities could reduce time spent on occupation-specific tasks. These are not interchangeable readings of one quantity. BLS brings these results into its framework as source-specific rankings, rather than treating their original scores as directly comparable. The other inputs bring observed product activity into the picture. Anthropic’s measure uses Claude activity, while Microsoft’s uses Copilot activity. BLS describes both as mapping activity to O*NET tasks or work activities, then aggregating mapped results to occupations. These details add useful provenance: the classification includes evidence that particular systems have been used in ways associated with work tasks, alongside estimates of theoretical task fit. BLS makes the sources comparable. It percentile-ranks each occupation within each source, takes one median across the three theoretical ranks and another across the two observed-use ranks, then clusters occupations using those dimensions. The categories—low, moderate, high, and very high—are relative groupings, not measured percentages of exposed work; they combine a capability-related dimension with a mapped-activity dimension. They do not combine a survey of workers in each occupation with a census of employer deployments. BLS says its current-evidence measures do not directly observe whether workers in a particular occupation used AI on the job. That distinction matters when reading the labels. Observed interaction is evidence of activity in the source data, not proof that a tool performs an entire production workflow, has employer approval, or has changed staffing. A theoretical score does not establish routine workplace use either. The method makes the occupation screen more evidence-mixed and transparent about its inputs; workers still need to ask which recurring task is involved, who checks the output, what exceptions remain, and whether the task has changed in their workplace.

Sources: Artificial Intelligence (AI) exposure categories

What do the four labels mean mathematically?

The four labels are relative groups, not percentages. In “Artificial Intelligence (AI) exposure categories,” the Bureau of Labor Statistics (BLS) explains that each source’s occupational scores were converted to percentile ranks among the occupations covered by that source. A percentile rank answers a comparative question: where does this occupation sit against others in this particular dataset? It does not preserve the original score’s units or tell you what fraction of a worker’s week could be assisted or completed by AI. That conversion makes unlike inputs easier to place on a common 0-to-1 scale, but it changes what can be inferred. The underlying sources do not all measure the same thing: some assess how AI capabilities relate to occupational abilities or tasks, while others map observed Claude or Copilot activity to work tasks or activities. After ranking, BLS takes the median of the three theoretical-source ranks and, separately, the median of the two observed-evidence ranks. Each occupation therefore has two summary positions, one for each dimension. A clustering algorithm then groups occupations using those positions into low, moderate, high, and very high relative exposure categories. A cluster is a compact label for a profile across those dimensions. It is not a threshold such as “half of this job is automatable,” and two occupations in one category need not have identical theoretical and observed profiles. The word “very high” may sound absolute, but BLS says the category is relative to other occupations and depends on the selected occupation universe, sources, crosswalks, imputation, normalization, and categorization choices. The page also cautions that the label is not a worker-replacement estimate, an adoption probability, a productivity forecast, or a wage or employment forecast. The data handling is visible enough to inspect. BLS reports 4,155 possible occupation-source combinations across 831 occupations and five sources; 3,944 values were observed and 211 were imputed, affecting 75 occupations. The agency says missing ranks were predicted from other available source data and the occupation’s two-digit Standard Occupational Classification major group. This is a methodological choice to fill incomplete coverage, not evidence that every affected category is wrong. It is a reason to remember that a category summarizes both measured inputs and, in some cases, an estimated rank. For an individual decision, do not read “very high” as “most of my week is exposed,” or treat “low” as assurance that your work will not change. If the category matters to a choice, look up the occupation’s two underlying summary dimensions and then examine your own recurring tasks. The label can help choose where to look; it cannot resolve how much of your role consists of those tasks, how they connect, or whether your workplace uses AI for them. Those questions require task and workplace evidence, not finer-sounding interpretation of the cluster name.

Sources: Artificial Intelligence (AI) exposure categories

Which parts of AI capability and use are represented—and dated?

The dates answer two different questions. BLS released its 2025–35 employment projections and AI exposure categories on August 27, 2026. The exposure page says its theoretical measures conceptualize capabilities available no later than mid-2023. The projection horizon and capability reference point are different. Treating those dates as one timeline would make the category sound more forward-looking than its method is. The inputs also represent different kinds of evidence. Three theoretical sources assess whether AI abilities relate to occupational abilities or whether a large language model could reduce time on listed tasks. Two “current-evidence” sources use observed Claude and Microsoft Copilot activity mapped to occupational tasks or work activities. BLS cautions that the mapped activity does not directly show that workers in a named occupation used AI on the job. A mapped interaction is activity in the data, not a count of employers approving tools or teams relying on output without review. And the capability measures are not live tests of every current product in every workplace. BLS identifies further limits on what is represented. Most sources focus on language modeling and omit image and video generation. The observed-use measures may lean toward early adopters of specific models. The page also says results from language models depend on the particular model and prompt. These limits make the classification a dated, selective screen; they do not establish that later tools perform more tasks reliably or that older evidence is useless. A newer model’s existence alone cannot fill the gap: applicability depends on the task, quality threshold, data access, review burden, and the organization’s controls. Those are questions about a current workflow, not conclusions supplied by this occupation-level category. For an individual decision, use the label to select tasks for a closer look, then check the current approved workflow in your own setting. Ask whether a recurring output is actually produced with AI, what a person still verifies, and whether handoffs or accountability have changed. If the tool is unavailable, prohibited, or unsuitable for the task, theoretical capability does not show workplace adoption. Conversely, the older capability window is a reason to verify present conditions rather than assume the category settles them. As the BLS “Employment Projections: 2025-2035 Summary” makes clear, 2025–35 is the projections period; “Artificial Intelligence (AI) exposure categories” gives a bounded comparison assembled from specified inputs. Read the first as a national outlook and the second as a prompt for task review. Neither says what your employer will do next.

Sources: Artificial Intelligence (AI) exposure categories; Employment Projections: 2025-2035 Summary

Why can an occupation label miss the way work actually fits together?

An occupation label summarizes a bundle of work; it does not describe the sequence by which a particular worker produces a result. The BLS methodology says all five input sources treat occupations as bundles of independent tasks, abilities, or activities. They do not account for bottlenecks, complementarities, or other features in which one task depends on another. That is a meaningful boundary: a task can appear technically suitable for AI in isolation while the surrounding steps determine whether it can be used in a real workflow. The category offers no estimate of how often that mismatch occurs or how much it matters, so the practical consequences below are an interpretation of the stated limitation, not a measured effect. Consider an illustrative example, not an observed case. Two people share the same occupation label and both prepare client-facing reports. One spends much of the week assembling standard material from structured records, then routes it for review. The other spends more time resolving unusual cases, reconciling conflicting inputs, and explaining why a recommendation should be trusted. A broad occupation-level category cannot tell us that their task shares are different. Nor does it show how much time either person spends checking a draft, answering questions, or taking responsibility for the final decision. Their titles match, but the composition and order of their work may not. The distinction between task capability and end-to-end performance matters here. A system might assist with a repeatable preparation step, yet the result still needs a person to verify sources, handle exceptions, meet a required approval, or explain an error. If that review is slow or consequential, it can constrain how much of the workflow is practical to change. On the other hand, if a routine step becomes easier, the amount or value of connected review and communication could also shift. Those are possibilities, not outcomes established by the BLS categories. The methodology identifies the omission of dependencies; it does not model a particular handoff, error cost, approval rule, or redesign. This is why a category should prompt a closer description of work rather than a conclusion about a worker. For a representative week, write down several recurring outputs and the steps that produce them: what is gathered or drafted, what is checked, which cases require judgment, where another person must approve or respond, and who is accountable when the result is wrong. Note exceptions and handoffs as well as routine steps. Keep the account factual and brief; confidential records or customer details are unnecessary. The aim is to notice the shape of the work, not to turn minutes or tasks into a new exposure score. That inventory can help frame a useful workplace question: has an approved tool actually changed a step, the review burden, the number of exceptions, or the responsibility attached to the output? A BLS label cannot answer that for an individual role. A task list cannot guarantee what an employer will do next either, but it makes the discussion more specific and gives the worker something observable to check. The label points toward a bundle; the workflow description shows where the bundle holds together and where a change might matter.

Sources: Artificial Intelligence (AI) exposure categories

What does a task-level comparison reveal that an occupation cluster cannot?

A task-level comparison can reveal differences that disappear when a whole occupation receives one label. The International Labour Organization’s 2025 study, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” builds its assessment from tasks documented in the Polish occupational classification. It used a representative sample of that system’s 29,753 tasks and survey input from 1,640 employed people across one-digit ISCO-08 groups. The researchers collected 52,558 data points about perceived automation potential for 2,861 tasks, compared those inputs with surveys and Delphi-style discussions among international experts, then used the task knowledge and model predictions to score tasks in ISCO-08 technical documentation. The publisher presents four increasing exposure gradients. This layered design brings together worker perceptions, expert judgment, and model estimates, rather than treating one capability score as representative of every task. These are study inputs, not a census of workers’ duties or a measure of employer installations. That finer grain can make variation within a broad occupation easier to see. A title may cover different task combinations; task assessments can distinguish a repeatable information-handling step from work whose content or context needs separate evaluation. A gradient across task evidence gives a more specific prompt: which activities are being assessed, and which are grouped together in the occupational description? It still does not show how much time a particular worker spends on each activity. Nor does assessed task potential establish that a system can complete an end-to-end workflow reliably, that an employer will adopt it, or that staffing will change. Those are separate questions. The comparison with the U.S. Bureau of Labor Statistics’ 2025–35 categories is about method. BLS combines three theoretical exposure measures with two observed-use measures mapped to occupational tasks or work activities. It percentile-ranks each source, summarizes theoretical and observed inputs separately, and clusters occupations into four relative groups. The ILO approach foregrounds task-level evidence and reports gradients; BLS supplies an occupation-level cluster from different inputs. Their geographies, occupational systems, data sources, and outputs differ. ISCO-08 task documentation and the Polish sample are not interchangeable with BLS’s U.S. projection occupations. An ILO gradient cannot be converted into a BLS label, and neither result is a personal replacement estimate. The ILO record describes a global framework, but its task ratings still do not account for one worker’s workplace. BLS cautions that its inputs treat occupations as bundles of independent tasks, abilities, or activities and do not model bottlenecks or complementarities. Task-level comparison can sharpen what to inspect while simplifying how duties connect. A worker can use either instrument to frame a question, then list recurring responsibilities and identify where actual tools, approvals, review, and handoffs have changed. That inventory is local evidence, not a new score or forecast. Occupation-level screening identifies where to look; task detail specifies what to ask; workplace observation tests whether change is happening in the role. This improves the next question, not certainty about the outcome.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Artificial Intelligence (AI) exposure categories

Why does exposure stop short of adoption, productivity, and displacement?

Exposure is a signal about potential task fit or mapped activity. It does not show whether an employer will buy, approve, integrate, trust, or benefit economically from a system. The Bureau of Labor Statistics says its categories do not measure adoption probability, productivity gains, employment effects, or worker replacement. The category therefore begins an inquiry; it does not complete one. A capability may exist in a demonstration and still fail the practical tests of accuracy, data access, security, cost, or accountability in a particular workplace.

Think of the links as contingent: capability may make a task technically feasible; feasibility may lead an organization to adopt a tool; adoption may alter a workflow, including how much checking, exception handling, and escalation it requires; redesigned work may affect output and costs; and those changes may influence demand, staffing, hours, or hiring. Each arrow can stop, reverse, or produce mixed effects. If a tool is used, a worker may spend less time drafting a first version but more time verifying it. Productivity depends on the whole process and the value and quantity of its output, not simply on faster completion of one step.

The International Labour Organization brief, “Workers’ exposure to AI: What indicators tell us – and what they don’t,” makes this boundary explicit: exposure measures technological susceptibility under a static view of tasks, while omitting economic and institutional adoption constraints and leaving workflow change, wages, employment, and demand unresolved. It also notes a possible countervailing route: productivity gains can lower costs and expand demand. Exposure alone cannot tell which route will dominate.

Observed labor-market evidence also cautions against treating AI use as having one predictable employment direction. The American Economic Association’s 2025 paper “How Different Uses of AI Shape Labor Demand: Evidence from France” analyzes firm-level adoption data from France in 2017–2020. Its abstract reports difference-in-differences estimates of increased firm employment and sales after adoption, alongside different employment responses by use: positive for some uses, including information and communications technology security, and negative for others, including administrative processes. This is evidence about firms in France during that period, not a forecast for a U.S. worker today; It does show why “AI exposure” cannot stand in for a uniform labor-demand result.

For an individual, look for observable workplace evidence before concluding that capability has changed the job: whether tools are approved for the work, which handoffs or responsibilities have changed, how review time and exception volume are shifting, and what role expectations now require. They still cannot reveal future headcount plans, and absence of visible change today does not guarantee none later. Keep capability, adoption, productivity, demand, and displacement as separate questions; the BLS label answers only the first of them imperfectly and comparatively.

Sources: Artificial Intelligence (AI) exposure categories; Workers’ exposure to AI: What indicators tell us – and what they don’t; How Different Uses of AI Shape Labor Demand: Evidence from France

How should exposure be read beside employment outlook?

Read the BLS AI exposure category beside the 2025–35 employment outlook, not as a substitute for it. The category compares occupations on theoretical and observed exposure; the projection estimates national employment levels and changes over a decade. Neither forecasts local openings, pay, or a particular employer’s plans. Combining them into a personal redundancy score would give them a meaning BLS does not claim. An exposure label is a screening signal about the relationship between AI measures and occupational tasks. An employment projection is a modeled view of future labor demand across the United States. In its August 27, 2026 release, “Employment Projections: 2025–2035 Summary,” BLS estimates employment levels and changes but warns that numeric precision does not remove long-term uncertainty. It advises readers to focus on direction and relative size rather than treat each estimate as certain. The horizon is national and forward-looking; it is not a forecast for a specific city, company, or worker. Employment change also differs from annual hiring openings, which can include replacement needs. The accompanying “Factors affecting occupational utilization” table adds explanatory context to those projections. BLS says it examines occupations across industries for structural factors likely to increase or decrease demand, and notes that not every occupation-industry combination has a listed factor and that multiple factors can apply. Entries describe several pathways, including productivity changes, shifting demand, and substitution. For example, the table says AI tools may let advertising and promotions managers test and oversee campaigns while fewer managers oversee more staff; elsewhere it describes demand for computer and information research scientists tied to developing new technologies, including AI. These are projection explanations, not isolated causal estimates or evidence of an individual employer’s plans. A decline can reflect productivity, demand, substitution, or other structural changes. A projected increase can coexist with exposure if demand for related output or services also grows. The projection release itself discusses both AI-related productivity pressures in some occupations and demand for workers connected to AI infrastructure and development. These national projections do not show that AI alone caused a change or how it will be distributed among workers. Use exposure to choose what to inspect, then consult outlook evidence separately. Before a larger move, check current regional postings and compensation for the target work, its stated prerequisites, and how your experience fits. Compare those facts with your salary floor, location, training time and cost, health, and family responsibilities. For an upgrade in your current role, first look for actual changes in tools, handoffs, review work, or expectations, then identify a skill that addresses the changed task. BLS’s category can start that inquiry; the national outlook and utilization factors can add context. Neither can settle the individual decision on its own.

Sources: Employment Projections: 2025-2035 Summary; Factors affecting occupational utilization; Artificial Intelligence (AI) exposure categories

Which evidence should change a worker's next move?

Use three evidence layers in order: the BLS label to choose what to inspect, recurring task and workplace observations to identify what is changing now, and separate local demand evidence plus personal constraints to decide whether to upgrade, move adjacent, or make a larger transition. They answer different questions; none gives a personal job-loss probability. BLS says its category is relative, does not measure adoption or employment effects, and should support career planning rather than stand in for it. Consider an illustrative knowledge-work bundle that includes collecting information, drafting a first version, checking sources, handling exceptions, communicating a recommendation, and being accountable for the decision. Ask which steps are recurring, which are actually changing in your workplace, and what evidence shows the change: an approved tool in the workflow, a handoff that has moved, less time spent on a step, more review or correction, or different expectations for the role. The BLS exposure categories explain that observed interactions are mapped to occupational tasks and do not directly observe whether workers in a given occupation used AI on the job. A category therefore cannot verify what is happening in your team. Then check demand for the path you might take, in the place where you can work. BLS’s “Factors affecting occupational utilization” table identifies structural factors expected to affect occupational demand within industries for the 2025–35 projections, and says factors are not identified for every industry employing an occupation. It adds national context, not local openings, pay, or personal fit. Check current regional postings, their stated prerequisites and compensation where available, and compare them with your salary floor, location, time and cost for training, health, and family responsibilities. For readers outside the United States, U.S. projections are context, not a local forecast. That evidence can support different-sized moves. An upgrade fits when work changes but your domain remains useful: learn an approved workflow, improve verification, or take responsibility for exceptions and quality. An adjacent move can preserve experience when neighboring roles use the same subject knowledge but place more weight on coordination, review, or accountable decisions. A larger change needs stronger evidence: realistic local demand, a clear prerequisite path, and a credible account of what the transition costs in time and income. Do not choose it from an exposure label alone. Match learning to the gap. A short course may help with a defined tool or workflow; a small project can show that you can apply and check a capability; deeper formal study makes sense when the target role actually requires its prerequisites. None guarantees hiring or income. First write down several recurring outputs and mark drafting, checking, exceptions, communication, accountability, and confirmed workflow changes. If the map points to a skill gap, choose one bounded learning step. If the work has not changed and the target path is unclear, gather more workplace and local-demand evidence before paying for training.

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

What is the proportionate next move for one worker?

Treat the BLS category as a prompt to inspect work, not as a verdict about your job. The page for “Artificial Intelligence (AI) exposure categories” says its clusters do not estimate an individual’s task shares, workplace adoption, or chance of replacement. A useful next step is to write down several outputs you produce repeatedly. For each, note the steps involving drafting or collecting information, checking, handling exceptions, communicating, and taking responsibility. Keep the list descriptive; it is not a new score.

Then look for evidence that the workflow is changing now: an approved tool, a shifted handoff, new review requirements, or a changed expectation about turnaround or output. If the work is changing, a current-role upgrade may address a specific tool or verification need. An adjacent move may reuse domain experience while changing the task mix. A larger transition deserves a separate check of local demand, pay, prerequisites, training time and cost, alongside health, family and location constraints. A category alone is not a reason to buy a course or leave a role.

The free task checker at /ai-job-risk-checker can organize task-level change-pressure signals; its result is not a validated probability of displacement. The category can tell you where to look. Your tasks, workplace evidence and circumstances must guide what to do next.

Sources: Artificial Intelligence (AI) exposure categories