What the facts and figures represent
In the order the pages show them, starting with BLS's exposure groups. Every rule here comes from a published source, except where it says the choice is ours.
The four AI exposure groups
BLS puts each of 831 occupations in one of four groups of about a quarter each, ranked against the others.
Exposure “does not imply job loss”: BLS projects most occupations in every group to grow.
- Low
In the bottom quarter of occupations for how much of the work AI could do or help with. 134 of 213 projected to grow.
- Moderate
In the second quarter from the bottom for how much of the work AI could do or help with. 151 of 206 projected to grow.
- High
In the second quarter from the top for how much of the work AI could do or help with. 165 of 206 projected to grow.
- Very high
In the top quarter of occupations for how much of the work AI could do or help with. 141 of 206 projected to grow.
About the occupation
1. AI exposure groups (Low, Moderate, High, Very high)
BLS combined five research measures of how much of an occupation's work AI could do or help with. Three estimate what AI could do (Felten, Raj and Seamans; Eloundou, Manning, Mishkin and Rock; Eisfeldt, Schubert and Zhang) and two measure what people are doing with it (Anthropic, from Claude; Microsoft, from Copilot). BLS ranked occupations on each measure and grouped all 831 into four groups of about a quarter each. So the groups are relative: Very high means more exposed than most other occupations.
What it doesn't tell you: Whether jobs will be lost, wages, or timing. BLS: exposure “does not imply job loss.”
Source: BLS AI exposure categories, August 2026
2. Projected job growth (Jobs by 2035)
BLS's projected change in the number of jobs in the occupation from 2025 to 2035. ▲ means more jobs, ▼ fewer. BLS adjusts its projections for AI “conservatively, where there is convincing evidence for a change,” and notes its methods “are not designed to capture extremely rapid technological change.”
What it doesn't tell you: What will happen to any one person's job, or to the occupation if AI changes work much faster than past technologies did.
Source: BLS Employment Projections 2025–35; Machovec, Rieley and Rolen, Monthly Labor Review (2025)
3. Median pay, openings, typical education, experience and training
Median pay is the middle wage: half earn more, half less. It leaves out the self-employed, and the page says how many are when BLS publishes it. Openings a year is the average number of jobs BLS expects to open each year to 2035, including those left by people retiring or changing work. Typical education, experience and training are what BLS finds is usual to start in the occupation.
What it doesn't tell you: What any one employer asks for, or pay where you live.
Source: BLS Employment Projections 2025–35
4. How its pay compares with all occupations
The share of occupations with a lower BLS median wage, out of every occupation BLS publishes one for. The marker sits by rank, not by dollars, so a few very high salaries don't squeeze everyone else to one end.
What it doesn't tell you: How much more one occupation pays than another, or pay where you live (that's the next entry).
Source: BLS Employment Projections 2025–35
5. Pay by state
BLS's Occupational Employment and Wage Statistics survey estimates pay and jobs for each occupation in each state, from employers. The map shades each state by its median wage, from the lowest state to the highest (a territory below that range takes the lightest shade). Tap a state to see its median, the range the middle half of workers earn (25th to 75th percentile) and how many jobs there are; every state is also in a table you can open. The site doesn't ask where you live: the choice stays in the page, and on a career card in its address so a shared link keeps it. A state BLS publishes no estimate for isn't listed, and a figure it withholds says “not published”. BLS writes very high wages as a top band, shown as “$239,200 or more”. A few occupations, such as actors, are paid by the hour only and are shown that way.
What it doesn't tell you: Pay in a city or region within a state, or what you'd earn: it is the middle of what employers in the state report paying. Living costs differ too.
Source: BLS OEWS May 2025, state estimates
6. Recent conditions: unemployment, and hires per job opening
Two recent figures beside the ten-year projection, each with the period it covers; neither is today's. Unemployment, for a calendar year, is the share of people in the occupation's labour force who were out of work and looking, counted in the occupation of their last job, from the Census Bureau's American Community Survey; the comparison is all occupations in the same survey. Hires per job opening comes from the BLS Job Openings and Labor Turnover Survey, which counts by industry, not occupation: it is the past twelve months' hires divided by their openings in the industry that employs the most people in the occupation (by BLS's industry counts), beside the same figure for all industries. The pages show it as hires for every 10 openings (0.79 is 7.9), one box per opening. Neither more nor fewer is simply better. More can mean openings get filled, or that people leave often and are replaced (restaurants, retail and warehousing hire all the time); fewer can mean employers can't fill their jobs, which gives job seekers leverage, or that they advertise but hire slowly. Read it with the unemployment figure beside it, and remember it describes the industry, not the occupation.
What it doesn't tell you: Your own chances of being hired, or what will happen next. Where the Census groups the occupation with others, the unemployment figure is for the group, and the page says so.
Source: Census ACS 2024 1-year microdata; BLS JOLTS
7. Job Zones
O*NET's measure of the total preparation an occupation usually needs (education, training and experience together), from Job Zone 1–2 (little) to 5 (extensive). For Job Zone 4, in O*NET's words: “Most of these occupations require a four-year bachelor's degree, but some do not.”
What it doesn't tell you: Whether you personally could make the move. It's context on how much preparation is usual, not a verdict.
Source: O*NET 31.0
About AI and the work
8. “AI could halve the time” on a task
Researchers rated each of an occupation's O*NET tasks, in 2023, by whether AI could do it in half the time or less at the same quality. AI alone: a chat assistant on its own could. AI + software: it would take software built on top of AI. Not much: neither would halve the time. The squares show one per task; an outlined square is a task the researchers didn't rate. O*NET has since reworded a few dozen tasks slightly; they keep their rating.
What it doesn't tell you: Whether anyone is actually using AI for the task, or how much it would change the job. The ratings are from 2023.
Source: Eloundou, Manning, Mishkin and Rock, Science (2024), human ratings; O*NET 31.0 tasks
9. How tasks are ordered
O*NET asks people in each occupation how important each task is, from 1 to 5, and whether it is part of their job. Tasks are listed by importance times the share who say it is part of their job, highest first. That ordering is ours. Tasks O*NET hasn't rated come last, in O*NET's own order, and where O*NET flags a rating as uncertain the page says so.
What it doesn't tell you: Which task matters most in a particular job.
Source: O*NET 31.0 task ratings
10. Work activities
O*NET links every task to broader work activities that many occupations share, 332 in all, such as “Communicate with others about specifications or project details.” An occupation's activities are the ones its tasks fall under, and they're listed by how many of its tasks fall under each. Because the same activities appear across occupations, they show what carries over from one to another.
What it doesn't tell you: Skills or qualifications. Activities describe the work, not what a person has.
Source: O*NET 31.0
11. “N in 10,000 conversations” and “#rank of 332”
Microsoft looked at 200,000 conversations people had with its Copilot assistant in 2024 and counted which work activity each was about. “People asked AI for help” counts what the person wanted done; “AI did it for them” counts what the AI did. “6 in 10,000” means about 6 of every 10,000 conversations involved that activity. For scale, a typical activity comes up about 3 times in 10,000; the most common, looking up information, about 2,300. The rank puts it among all 332, #1 the most common; activities that never came up share last place, shown as “tied”. The summary box counts how many of the occupation's activities are among the 50 most common and names them; that cut-off is ours. Most occupations have a few: everyday activities such as reading documents or keeping knowledge current rank high for nearly everyone.
What it doesn't tell you: How people in this occupation use AI: the figures cover everyone who used Copilot. Dedicated coding tools aren't included, so programming help is under-counted.
Source: Tomlinson et al., "Working with AI" (Microsoft, 2025), data v1.1
12. How people in the work used Claude
Anthropic sorts conversations with Claude by the occupation of the task. Handed off means Claude did the task; worked on it together means the person stayed involved, refining or learning. Anthropic also labels each conversation as work, personal or coursework, and by what it produced (in its words, “the most prominent concrete output Claude produced”), such as a code fix, an app or a document; the page shows these as shares of this occupation's conversations. Where a report shows a sample-size line, it compares this occupation's share of all conversations with the average occupation's, 0.14%. Beside each task, the card shows the task's share of all Claude conversations in May 2026: most tasks are a tiny share, so “under 0.01%” is common. Anthropic publishes a figure only when enough conversations sit behind it, so “not published” means too few conversations, not zero.
What it doesn't tell you: How most people in the occupation work: these are only people who brought tasks to Claude.
Source: Anthropic Economic Index, release 2026-06-26 (May 2026 conversations)
About career moves
13. Where people in your field were a year later
A national survey asks the same people about their job now and the job they held longest the year before. Where the two differ, they changed field. It can only be read by broad field, such as Computer and Mathematical, because too few people make any one detailed move. Shares are shown only when they meet the National Center for Health Statistics standard for presenting proportions.
What it doesn't tell you: Which occupation within a field people moved to; that's the next entry.
Source: IPUMS CPS ASEC, 2003–2025
14. Where people in an occupation moved next
Moves between occupations found in online resumes from 2002 to 2016: of everyone who left an occupation, the share who went to each other one, with the number of moves behind each share. The data uses 2010 occupation codes, so some destinations are older groups BLS has since split, and some newer occupations are counted inside an older group.
What it doesn't tell you: What will happen next. The moves are from before today's AI tools, and resumes over-represent younger, white-collar job-seekers and moves into management.
Source: Schubert, Stansbury and Taska (resume data, used with the authors' permission)
15. Moves into lower-exposure occupations
The moves above, into occupations BLS rates in a lower exposure group than yours. One view keeps only occupations that typically need at least your occupation's education to start. That narrowing is ours, not a published rule.
What it doesn't tell you: Whether a move is right for you, or what it would pay you.
Source: Resume data; BLS AI exposure categories and projections
16. “You already do X of its Y work activities”
The work activities two occupations share, as a count of the other occupation's activities. What carries over is what you already do; what's new is the rest, each with how often people asked Copilot for help with it. When ranking options, activities found in hundreds of occupations, like keeping records, count for less, so a match means similar work rather than common paperwork. The page always shows the plain count.
What it doesn't tell you: Licences, degrees or experience an occupation may require.
Source: O*NET 31.0
17. Other, less obvious options
Occupations outside your own field (the first two digits of the occupation code) that took under 1% of moves but share at least 4 of your work activities, ranked by how much of their work you already do, at most two from any one field, leaving out catch-all “all other” groups. Where fewer than three occupations share 4, the minimum is 3. They are ideas to look into, not proven paths.
What it doesn't tell you: That anyone has made the move. Many have no recorded moves at all.
Source: O*NET 31.0; resume data
18. Where people moved in from
The same resume data read the other way: of everyone who moved into an occupation, the share who came from each other one. These are career changes, so a first job after school isn't counted.
What it doesn't tell you: The usual route in from school, which for most occupations is education, not a move. When the resume data counts an occupation only inside an older catch-all group (“…, All Other”), the student results leave this out: the figures would describe the whole group, mostly other work.
Source: Schubert, Stansbury and Taska (resume data)
About what an occupation takes
19. The education people in an occupation hold: two surveys
Two surveys answer this, and the pages show both when both have figures. O*NET asks people working in each occupation; about 20 answer for a typical occupation, so its shares are uncertain, and a share O*NET judges too uncertain is shown as “not published”. It reports the exact occupation and counts certificates. The Census Bureau's American Community Survey reaches about one in a hundred households, so thousands of people stand behind most occupations, but it counts degrees only and groups some occupations together (the page names the group). About 40 small occupations, such as mathematicians, have no Census figures here: the public file counts them inside larger groups. Where they differ, the Census figure rests on far more people. The same survey gives median pay of full-time, year-round workers by highest degree, shown for a level held by at least 5% of the occupation's workers, with at least 30 such workers behind it. It is pay people report for the past year, so it can differ from BLS's median, which employers report.
What it doesn't tell you: What an employer requires, or what you'd need. Pay by degree describes who holds the job now.
Source: O*NET 31.0 (education); Census ACS 2024 1-year microdata
20. What an occupation takes, beside what you have
Education is the only line marked as met or not, because your education and BLS's typical entry education use the same scale. Everything else is side by side: years in the occupation and in the ones O*NET lists as related; the education people in it hold (an O*NET survey, not published for every occupation; a share O*NET judges too uncertain is shown as “not published” rather than left out, so the parts shown can add up to less than 100%); the skills and knowledge areas O*NET rates by importance from 1 to 5; and the software O*NET lists. No source names the licences or certifications an occupation asks for, so yours are listed, not compared; how many jobs require one is below.
What it doesn't tell you: Whether you'd be hired. No figure here scores you.
Source: BLS; O*NET 31.0
21. Jobs that require a credential
BLS's Occupational Requirements Survey asks employers what each job requires, counting only what is required, not what is preferred. For 2023, the final figures, it reports the share of jobs that require any credential, and separately a license (issued by a government agency), a certification (issued by a certification body or a professional or industry association) or an educational certificate (issued by a school or training provider for completing a program of study). A job can require more than one, so the kinds don't add up to the first line. BLS writes very small and very large shares as “<0.5%” and “>99.5%”, and the pages keep its wording. A kind BLS doesn't publish for an occupation is shown as “not published”: too few jobs in the sample, not zero. ORS covers about half of occupations; for the rest, the page says it publishes nothing. Where the figures are for a wider occupation than the one shown, the page names it.
What it doesn't tell you: Which credential, or whether yours counts: BLS publishes shares only, never the credential's name. Nor what each state requires; the License Finder link covers that.
Source: BLS Occupational Requirements Survey, 2023 complete dataset
22. Looking up certifications, training, licenses and apprenticeships
Each occupation links to three CareerOneStop tools, run by the U.S. Department of Labor: the Certification Finder (certifications CareerOneStop lists for the occupation), the Training Finder (schools and programs, nationwide until you narrow the place) and the License Finder (licenses by state). Each link searches by the occupation's O*NET code, so it follows the occupation shown, and the code is all it sends. Apprenticeships are on Apprenticeship.gov, whose finder can't be linked to one occupation; the link appears only where the Department of Labor's list of apprenticeable occupations includes this one, with the name to search for.
What it doesn't tell you: Which credential an employer requires, or which program is good. The finders list what exists; we don't rank or check them.
Source: CareerOneStop (U.S. Department of Labor); Apprenticeship.gov
About the student results
23. Interest scores and career fit (Best, Great, Good)
The O*NET Interest Profiler has 60 activities, 10 for each of six interest areas. Your score in each area is how many you checked, 0 to 10. Each career's fit is how closely the pattern of your six scores matches O*NET's interest ratings for it, by correlation: Best fit .729 or more, Great fit .608 or more, Good fit above zero, as the Interest Profiler Manual sets out. The pattern matters, not the level, so checking many or few activities doesn't change the fit. If every area has the same score there's no pattern to match. People who take the profiler twice get similar scores, and people whose work fits their interests tend to be somewhat more satisfied with it.
What it doesn't tell you: What you'd be good at or succeed in. Following the list hasn't been tested as a way to choose a career.
Source: O*NET Interest Profiler Short Form (used unmodified, CC BY-ND 4.0); O*NET 31.0 interest ratings; Hoff et al., Journal of Vocational Behavior (2020)
24. Subjects it relies on
O*NET rates 33 knowledge areas, such as Mathematics or English Language, by how important each is to an occupation, from 1 to 5. The student results list every area rated 3 or more and mark the ones you ticked; a subject you ticked that rates lower is named with its rating. The cut-off of 3 is ours, to keep the list short, and all 33 are one click away. Beside the careers that fit your interests, the three of your ticked subjects each relies on most are shown, with no cut-off.
What it doesn't tell you: How good you'd need to be at a subject, or whether you'd do well in it.
Source: O*NET 31.0 knowledge ratings
25. Jobs you've had, beside a career
Each job is matched to an occupation (you pick it on the form). The results say whether O*NET lists that occupation as related to the career, and how many of the career's work activities the job's occupation shares with it, counted as in “You already do X of its Y”. Then BLS's typical work experience to start. O*NET's related list, in its three groups, sits at the end of each career: O*NET's analysts chose it, so it isn't drawn from where people moved.
What it doesn't tell you: Whether the experience would count with an employer. BLS doesn't say which occupations count as related experience, so nothing here is marked met or not.
Source: O*NET 31.0 (related occupations, work activities); BLS
26. Where a major leads, and programs that prepare for an occupation
The occupations the National Center for Education Statistics lists as ones a program prepares people for, and the reverse. Experts drew up the list; it isn't a count of where graduates went. Many people work outside the field they studied. The education people in an occupation hold comes from an O*NET survey of people in it.
What it doesn't tell you: How many graduates end up in each occupation.
Source: NCES CIP 2020 to SOC 2018 crosswalk; O*NET 31.0