Questions people ask

No. Exposure means AI could do or help with some of an occupation's work. BLS, which publishes the groups, says exposure “does not imply job loss.” It projects most occupations in every group to grow, including 141 of the 206 rated Very high. More in the Guide →

BLS ranks 831 occupations by how much of their work AI could do or help with, and splits them into four groups of about a quarter each: Low, Moderate, High and Very high. The groups are relative, so Very high means more exposed than most other occupations, not a fixed amount. More in the Guide →

Yes, carefully. In BLS's words: “BLS generally applies adjustments based on this research conservatively, where there is convincing evidence for a change.” It also says “BLS projection methods are not designed to capture extremely rapid technological change.” So the projections are a careful estimate that may lag if AI changes work faster than past technologies did. More in the Guide →

The measures behind the groups look at an occupation's tasks and whether AI could do or help with them. They don't weigh things that also protect a job, such as licences, legal responsibility, the need to be there in person, or a client's trust. And the groups are relative, so an occupation can be High without most of its work being affected. More in the Guide →

Researchers rated every O*NET task, in 2023, by whether AI could do it in half the time or less at the same quality: on its own (AI alone), with software built on top of it (AI + software), or not much. It says what AI could speed up, not whether anyone is using it that way. More in the Guide →

Because they answer different questions. The task ratings show potential: what researchers judged, in 2023, that AI could speed up. The Copilot and Claude figures show what people are asking AI to help with now. Where both are high, AI is already part of the work; where the potential is high but use is low, the change may be still to come, or something the ratings don't capture (judgement, rules, being there in person) may be holding it back. Neither predicts what will happen to jobs. More in the Guide →

Exposure groups and projections: BLS, August 2026 (projections to 2035). Occupations, tasks and activities: O*NET 31.0. Task ratings: 2023. Copilot figures: 2024 conversations. Claude figures: May 2026. Moves between fields: a national survey, 2003–2025. Moves between occupations: online resumes, 2002–2016. Every figure names its source, and the Sources page lists the release in use. Sources →

A score of our own would mean choosing how much each piece of evidence counts, and another person would choose differently. BLS combines five published research measures and puts its name to the result. We show the pieces beside it (tasks, Copilot and Claude use, moves) so you can see where the exposure comes from.

Microsoft counted which work activity each of 200,000 Copilot conversations in 2024 was about. “6 in 10,000” means about 6 of every 10,000 conversations involved that activity; a typical activity comes up about 3 times. It is one of the few published studies of real AI use, broken down by work activity. More in the Guide →

Anthropic, which makes Claude, publishes how people used it by occupation: whether Claude did the task (handed off) or the person stayed involved (worked on it together), and what the conversations produced. We haven't found free, published data like this from other AI companies; if one releases it, we'll add it. More in the Guide →

By how important people in the occupation rate each task, times the share who say it is part of their job, using O*NET's ratings. That ordering is ours, and the page says so. Tasks O*NET hasn't rated come last. More in the Guide →

Some of it. The task ratings (Eloundou and others, Science) and one of BLS's five measures (Felten, Raj and Seamans, Strategic Management Journal) are peer-reviewed. BLS figures are official statistics. The Microsoft and Anthropic studies, the resume data and some of BLS's other measures are working papers or company reports. Each is named where it is used.

It is the only free, detailed record of moves between specific occupations, and it ends in 2016, before today's AI tools. It comes from online resumes, which over-represent younger, white-collar job-seekers and moves into management, and they count jobs held at the same time as moves. We show it with those limits rather than leave it out. More in the Guide →

O*NET links every task to work activities that many occupations share. “You already do 10 of its 29 work activities” counts the ones the two occupations have in common. It doesn't include licences, degrees or experience the other occupation may require. More in the Guide →

Those options are picked by shared work activities, not by what people did, so some have few or no moves in the resume data. They are ideas to look into, outside your own field, not proven paths. More in the Guide →

The data records moves between occupations on resumes, so a first job after school isn't a move. For most occupations, the usual way in from school is the education shown beside it. More in the Guide →

No, and no. A resume you upload or paste is sent to an AI service only to read it into fields you can check, then deleted; nothing is saved. You can enter your jobs by hand instead, and the student path needs no resume and sends nothing anywhere. Privacy →

Your answers to the O*NET Interest Profiler give a score from 0 to 10 in six interest areas. Each career's fit is how closely that pattern matches O*NET's ratings for the career, using the Interest Profiler Manual's cut-offs for Best, Great and Good. It reflects what you said you'd like doing, not ability or chances of being hired. More in the Guide →

Yes. The student path needs no resume and sends nothing anywhere, and every figure links to how it was worked out, so it can be discussed openly. It is free, with no sign-up.

PathMosaic was built by Coen Chou, with input from multiple individuals. Spotted an error, or have a question? Email hello@pathmosaic.com. About →