Why we built PathMosaic
We started with a common worry: Will AI take my job?
The headlines were either panic or hype. So we went to the research and to the U.S. government's own data, and built a place to see it, occupation by occupation.
What we found
What AI could do
Several respected studies estimate how much of an occupation's work AI is capable of doing, based on the tasks or abilities the job requires. Some ask whether AI could cut the time for each task by at least half without lowering quality, with or without additional software built on top of it (Eloundou et al., 2024; Eisfeldt et al., 2023). Another links progress in common AI applications, such as image recognition and language modeling, to the abilities each occupation needs (Felten, Raj and Seamans, 2021). Either way, the result is an exposure score for each occupation. All of these measure tasks or abilities, not jobs. None of them estimates how many jobs will be lost.
What AI is actually used for
Newer research looks at how people really use AI at work, by sorting hundreds of thousands of Microsoft Copilot conversations and millions of Anthropic Claude conversations into the work activities and tasks of each occupation (Tomlinson et al., 2025; Massenkoff and McCrory, 2026).
Where change is already visible
Stanford's "Canaries in the Coal Mine," updated in August 2026, found that employment of 22- to 25-year-olds in the most AI-exposed occupations is about 19% below where it would be had it kept pace with similarly aged workers in less-exposed occupations (from November 2022 to June 2026, it fell about 11% in the most exposed occupations while growing about 10% in the least exposed). The authors call these "descriptive patterns, not causal estimates of the effect of AI." The gap holds after leaving out tech companies and allowing for interest rates and remote work, but it narrows when education is taken into account, and part of it began before AI tools were widely used. Read the research.
What the research doesn't settle
Whether the time saved turns into more output (more software, more care delivered) or fewer workers; how fast it happens; and whether the gains are broadly shared. Researchers disagree on all three, and the evidence is still early and mixed.
Our reading our interpretation, not a finding
Work built mostly on reading, writing and routine analysis is under pressure, especially at the junior level. Work where hands-on skill, judgment or accountability is the core is more likely to change than to disappear. Laws, policy and how employers choose to use AI will also shape what happens.
How to think about AI and your career
We can't tell you what will happen. So we gather the best public data, show where each figure comes from and what's missing, and leave the decision to you. The future is uncertain. The best anyone can do is decide with the best information available.
Three ways to use PathMosaic
What you'll find on PathMosaic
Official projections
The Bureau of Labor Statistics (BLS) uses the three studies of what AI could do and the two of how people use it to sort 831 occupations into four exposure groups. It says exposure "does not imply job loss." It also projects job growth to 2035 for the same 831 occupations. Since its 2023–33 projections, it has adjusted individual occupations for AI "conservatively, where there is convincing evidence for a change," and it notes that its methods "are not designed to capture extremely rapid technological change."
Based on that data, most occupations in every exposure group are projected to grow. Counted by jobs, Very high is the largest group, with 31% of 2025 jobs, and it is projected to grow 2.0% by 2035 (all occupations: 3.5%).
| BLS exposure group | Occupations projected to grow | Share of 2025 jobs | Jobs by 2035 |
|---|---|---|---|
| Low | 134 of 213 | 18% | +2.5% |
| Moderate | 151 of 206 | 26% | +6.2% |
| High | 165 of 206 | 25% | +3.1% |
| Very high | 141 of 206 | 31% | +2.0% |
| All occupations | 100% | +3.5% |
BLS AI exposure categories (August 2026) and Employment Projections 2025–35. Jobs by 2035 is the change in jobs summed across each group.
What AI could do, and what people use it for
Each career card puts two kinds of evidence side by side. First, the occupation's tasks, marked where research rates AI as able to halve the time (Eloundou et al.). Second, which of its work activities people actually bring to AI today, from Microsoft's Copilot and Anthropic's Claude conversations. For software developers, AI could halve the time on 15 of 17 tasks, and 5 of their 15 work activities are among the 50 people most often ask AI for help with.
Read together, they show where things stand: what is possible, and how much of it is already happening. Across all occupations, the most common request in Copilot conversations is "Gather information from physical or electronic sources," about 23 in every 100. And in 53% of the occupations Anthropic reports on, people more often work through tasks together with Claude than hand them off entirely.
Eloundou et al. (2024); Microsoft "Working with AI" data, v1.1; Anthropic Economic Index, May 2026 conversations.
Where people actually moved when they changed careers
If you're thinking about a switch, one of the most useful things to know is where people in your occupation actually went. Each card and report shows this from millions of online resumes (2002–2016).
People change careers in every direction. Of moves out of the most AI-exposed occupations, 42% went to a less exposed one. From every exposure group, between 68% and 74% of moves went to occupations BLS now projects to grow.
Two cautions: these moves happened before today's AI tools, and 17% of them went to older occupation groups that no longer match an exposure group, so they aren't counted.
Schubert, Stansbury and Taska occupational mobility data (resumes, 2002–2016), with today's BLS exposure groups and 2025–35 projections.
Also on every career card
Median pay and where it ranks among all occupations, pay by state, openings each year, the typical education to start, recent conditions (unemployment and hiring), the education people in the occupation hold, pay by degree, and the programs and credentials that lead to it. The Guide explains every figure →
How we built it
- Describe, don't judge. No score of you and no prediction. We show published figures and let you decide.
- Every figure names its source, and the Sources page lists the release in use.
- Missing data is shown as missing, never filled in with a guess.
- Nothing you enter is stored. It's free, with no sign-up.
Who we are
PathMosaic was built by Coen Chou, with input from multiple individuals. Coen has researched employment and pay among computer science graduates; his paper "Why Nominal Wages Didn't Fall When Computer Science Unemployment Rose" has been conditionally accepted by IJHSR.
Spotted an error, or have a question? Email hello@pathmosaic.com.
References and data releases are on the Sources page.