AI & Work
What the New BLS AI Exposure Categories Actually Mean
How to interpret the BLS Low, Moderate, High, and Very high relative AI exposure categories without turning them into predictions of automation, job loss, or career safety.
September 14, 2026 · Dr. Morgan Reed
The Bureau of Labor Statistics now publishes a relative AI exposure category for each detailed occupation in its Employment Projections program. The four labels are Low, Moderate, High, and Very high.
Those labels are useful, but only if we resist giving them meanings they do not have. A category is not the percentage of a job that will be automated. It is not the probability that a worker will lose a job. It is not a career-safety grade.
The categories answer a narrower question: compared with other occupations, how much does the work overlap with the AI capabilities and observed AI activity represented in the sources BLS combined?
That distinction changes how the data should be used. Exposure can tell you where to look more closely. It cannot tell you what will happen next.
What BLS combined
For the 2025–2035 Employment Projections release, published August 27, 2026, BLS combined five external measures of occupational AI exposure.
Three are theoretical measures. They examine whether AI capabilities relate to occupational abilities or whether large language models could assist with tasks or reduce the time needed to perform them. The studies use different methods, including human ratings and ratings produced with specific language models.
Two are current-evidence measures based on observed interactions with AI systems. One maps Claude usage to O*NET occupation-specific tasks. The other maps Microsoft Copilot activity to O*NET Intermediate Work Activities and then links those activities to tasks and occupations.
"Observed" needs careful interpretation. BLS notes that these measures do not directly observe whether workers in a particular occupation used AI on the job. They observe activity within particular AI systems and map that activity back to occupational work. Early adopters, product users, and the tasks that people bring to those systems may not represent every workplace.
BLS converted each source's results to percentile ranks so unlike measures could be placed on a common relative scale. It then calculated one median rank across the three theoretical sources and another across the two observed sources. A clustering method grouped occupations using those two dimensions and produced the four categories.
This is why the labels are relative. They compare occupations within this method. They do not establish an absolute amount of exposure, and the distance from Moderate to High is not a fixed unit that can be interpreted like ten percentage points.
Reading the four categories
BLS gives the endpoints the clearest plain-language interpretation.
Low generally indicates that an occupation's requirements do not match current AI-model capabilities well and that large language models have not been observed performing many of its tasks in the included usage data.
Very high generally indicates that, compared with other occupations, a larger fraction of tasks can be completed or assisted by represented AI technologies and that large language models have been observed performing some of those tasks.
Moderate and High identify occupations between those ends in the clustering results. They are comparative groups, not claims that a specific portion of the work can be automated.
A category also applies to an occupation, not uniformly to every person with that title. Employers divide work differently. Industries use different systems. A senior worker may spend more time on judgment and coordination, while an entry-level worker may spend more time producing or processing information. The BLS category cannot describe every one of those arrangements.
Exposure is not an employment forecast
The most important reading rule is simple: keep relative AI exposure separate from employment projections.
BLS explicitly says that an exposure category is not a forecast of employment growth or decline. It is also not a wage forecast, an estimate of AI adoption, a worker-replacement estimate, or a productivity forecast. The categories do not distinguish automation from augmentation.
That last limitation matters. If an AI system can assist with a task, an employer might use it to reduce effort, increase output, improve access to information, change review responsibilities, or redesign a service. The exposure measure does not decide which organizational response will occur.
The labor market can also move for reasons beyond AI. Population change, consumer demand, public policy, investment, industry growth, retirements, and changes in business organization all shape occupational employment. Exposure and projected growth can therefore point in combinations that a simple risk ranking would miss.
For example, JobShift's source-backed profile places Accountants and Auditors in the Very high category, while BLS separately projects 5.0% employment growth from 2025 to 2035. The work includes preparing and analyzing records, but it also includes reconciling discrepancies, evaluating controls, checking compliance, reporting findings, and advising organizations. Exposure to document and analysis capabilities does not settle demand for the whole occupation.
Secondary School Teachers are in the High category. Their O*NET tasks include preparing materials, grading work, maintaining records, and planning instruction, alongside adapting teaching to students, maintaining classroom order, observing development, and conferring with families and colleagues. A tool may assist some preparation or information tasks without reproducing the full classroom relationship or assigning accountability for students.
Automotive Service Technicians and Mechanics are in the Moderate category. Their work combines diagnostic information and technical manuals with inspecting vehicles, testing components, discussing problems with customers, and physically repairing and verifying systems. The category invites questions about diagnostic support and information retrieval; it does not imply that the physical and safety-sensitive parts of the workflow stand still.
These examples are not a ranking from worse to better. They show why a category should lead back to tasks, work settings, and responsibilities.
The method has a date and a boundary
The 2026 BLS release is a substantial new source, not a permanent measurement of a fast-moving technology.
BLS notes that the theoretical studies represent AI capabilities available no later than mid-2023. The observed measures may lean toward early users of specific products. Most of the source measures focus on language models and do not include image or video generation. Results can also be affected by occupational crosswalks, source coverage, missing-value estimates, model prompts, and the choices used to create the four groups.
There is another structural limit: the five sources generally treat occupations as bundles of independent characteristics, such as tasks, abilities, or work activities. They do not fully model the dependencies inside a real workflow. A generated draft may still require access to reliable records, review by someone with domain knowledge, approval under a policy, communication with a customer, and responsibility for the final action. Speeding up one step may expose a bottleneck somewhere else.
These limits do not make the categories useless. They tell us what kind of evidence we have. The measure is strongest as a broad comparison that directs attention toward occupational work. It is weaker when treated as a precise forecast about a person, employer, or future technology.
How JobShift uses relative exposure
JobShift Atlas keeps several evidence layers visible instead of compressing them into one score.
The BLS category is a source fact. BLS employment projections, annual openings, wages, and preparation fields are separate source facts. O*NET tasks, skills, knowledge, and work activities describe the work from another authoritative source. JobShift's plain-English explanation is an editorial interpretation of those facts, not a new government measure.
That separation lets you ask better questions:
- Which tasks in this occupation overlap with the capabilities represented in the BLS sources?
- Would a tool assist the worker, automate a repeatable step, or change a handoff?
- What review, context, communication, physical action, or accountability surrounds that step?
- What do employment projections and annual openings show separately?
- What preparation and occupation-specific knowledge would still be needed?
- What has actually changed in the workplaces you are considering?
The JobShift guide to relative AI exposure explains the method in more detail. The tasks and workflows guide provides a practical way to trace what happens before and after a tool-assisted step. For broader education about AI capabilities and their changing role in work, AI Revolution Atlas is the relevant next stop.
What this means for you
If your occupation is labeled High or Very high, do not read the category as an instruction to leave. Inspect the work. Identify the tasks that are changing in your setting, learn how outputs are checked, and notice whether responsibility is moving toward review, exception handling, communication, or domain judgment.
If your occupation is labeled Low or Moderate, do not read the category as protection. Technology can affect scheduling, documentation, diagnostics, customer expectations, equipment, and business demand even when current language-model measures show less overlap with core tasks.
In either case, pair the category with evidence about the occupation and with facts from the setting you actually face. A national classification cannot know your employer's tools, local demand, credentials, experience, or priorities.
The useful takeaway is narrower and more durable: AI exposure describes where represented AI capabilities may touch work. It does not tell you whose job will disappear.
Sources and further reading
Keep exploring the evidence
What should I verify before applying this article’s ideas to my own occupation? ↗