Jobs least likely to be replaced by AI
The safest jobs in 2026 require a body in the room, a skill that can’t be written down, or a relationship built over years. Nurses, surgeons, therapists, electricians, plumbers, HVAC technicians, phlebotomists, massage therapists, skilled-trade equipment operators. Microsoft measured 200,000 Copilot conversations and found under 5% overlap between what AI actually does and what these jobs require. Dallas Fed data from January 2026 shows hiring in those roles held steady while AI-exposed jobs are shedding young workers. The counterintuitive part: the most exposed workers aren’t factory hands. They’re writers, translators, customer service reps, and data analysts. This is the first automation wave to hit desk workers harder than people who work with their hands.
Most lists of AI-proof jobs are guesses. The better ones are built on measurements: what AI is actually doing at work, where it fails, and which jobs have shown no employment drop since ChatGPT launched.
Three studies published in the last eighteen months reach the same conclusion. This guide pulls their findings together, names the specific jobs, and explains why some work is structurally resistant to AI while other roles are already shrinking.
The research
Microsoft Research (July 2025) analyzed 200,000 anonymized Bing Copilot conversations and built an “AI applicability score” for every occupation in the Labor Department’s O*NET database. They measured overlap between what AI actually does and what each job requires, rather than guessing what a future model might handle. The bottom 40 occupations on that list are the empirical core of any honest answer to this question.
Anthropic’s study (March 2026) tracked how Claude is used in real workplaces and cross-referenced it with Current Population Survey data. Its most striking finding: no measurable rise in unemployment among high-exposure workers since ChatGPT launched. The disruption is showing up in hiring, not layoffs.
The Dallas Fed (January 2026) was more specific: workers aged 22–25 in the most AI-exposed jobs have seen a 13% employment drop since 2022, almost entirely because fewer people are entering those roles, not because current workers are losing them. A job that feels safe at 45 may be quietly closing its door to 22-year-olds.
What makes a job actually resistant to AI
Across all three datasets, resistant jobs share four characteristics. These matter more than any job title list, because titles change but the traits don’t.
Notice what’s missing: “creativity.” Writers and translators rank among the most exposed workers in Microsoft’s data. If your output lives on a screen, AI is already competing with you.
The 20 jobs with the lowest AI exposure
These roles appeared consistently at the bottom of Microsoft’s applicability rankings (under 0.05 in most cases) and showed no unusual employment drop in the Dallas Fed data.
These aren’t glamorous jobs, and that’s the point. The assumption that automation hits manual work first has been wrong for this cycle. The current wave is compressing desk-based knowledge work — the opposite of every prior automation era.
Resistant jobs that are also growing
Psychologists and therapists are projected to grow roughly 22% through 2030 (WEF, 2025). AI can produce empathetic-sounding text; it can’t run a therapy session, respond to a patient in crisis, or build a clinical relationship over years. Surgeons and specialist physicians are projected at around 15% growth. Nurse practitioners are the fastest-growing major low-exposure occupation, at 45.7% projected growth by 2032 per the Bureau of Labor Statistics.
Electricians, plumbers, and HVAC technicians don’t appear on most AI-safety lists because they’re not knowledge work, but every characteristic of AI resistance applies. They also have a structural tailwind: the workforce is aging out and fewer young people are entering trades.
A 2025 Linkee study put chief executives at 14% automation risk, architects at 18%, and event planners at 20%. These aren’t roles where judgment is occasional. Negotiation, contextual taste, and real-time adaptation to shifting stakeholder positions are the whole job.
Jobs most exposed to AI by 2030
Microsoft’s highest-scoring occupations on AI applicability — the roles where AI overlaps most with actual work:
The WEF estimates 92 million jobs displaced globally by 2030, mostly in routine admin work and information-processing roles a chatbot handles in seconds. But 170 million new roles are projected to appear, for a net gain of 78 million. The real question isn’t whether your job disappears. It’s whether the parts AI handles shrink your headcount from five people to one.
What the researchers want you to know
Every team behind this data published the same warning. A low AI applicability score today doesn’t guarantee security in five years. Microsoft’s data captures what Bing Copilot is doing right now, not what embodied robotics will do next decade. Anthropic’s team noted they’d detect a doubling of unemployment in exposed occupations, and it hasn’t happened. But they flagged the early signal: not layoffs, just closed doors for young workers entering the field.
The Dallas Fed put it plainly. Employment for 22–25-year-olds in the most AI-exposed jobs is down 13% since 2022. It’s not that people are getting fired. It’s that the entry door is closing. The job feels safe if you already have it. It’s just harder to get.
What this means for a career decision
Physical work, relational work, and work in unpredictable environments is more insulated from current AI than any desk-based information job. That’s what the data says. Within exposed fields, the people holding on aren’t the ones competing against AI — they’re the ones using it to do work that used to require a whole team.
“Creativity” and a degree are not protection. Both studies found the opposite. Bachelor’s-plus knowledge workers are more exposed than workers without degrees. Writers, translators, and historians sit near the top of the exposure list.
The more useful frame: is the reason AI can’t do your job a capability gap that closes with the next model, or something structural — a body the job requires, a relationship it can’t inherit, a situation it has no playbook for? The first kind of safety has a short shelf life. The second doesn’t.