Which Jobs Are Safest from AI?
This page answers one question: which occupations are least exposed to generative AI, and why. The ranking uses the same public-data index as the rest of the site, so it reflects what researchers rate AI able to do and what people actually use it for today.
Safe does not mean untouched. Nearly every occupation on this list already uses AI somewhere, for paperwork, scheduling, triage or documentation. What keeps them low is that the part of the work carrying the responsibility still needs a person in the room.
The occupations that stay near the bottom share a shape rather than an industry. Skilled trades, direct care, emergency response and safety-critical operation appear together because each combines physical presence with a decision somebody has to answer for.
Read the exposure, then read the jobs outlook. A low exposure figure says little about demand: some of these occupations are projected to grow quickly, others to shrink for reasons that have nothing to do with AI. Every occupation has its own page with both figures.
Top 25 Jobs Safest from AI
| Rank | Job | Generative-AI exposure | US jobs 2025–35 |
|---|---|---|---|
| 1 | Housekeeper | 13 | +0.6% |
| 2 | Fisherman | 17 | -4.2% |
| 3 | Welder | 18 | +2.4% |
| 4 | Miner | 19 | -1.0% |
| 5 | Construction Worker | 19 | +7.3% |
| 6 | Carpenter | 19 | +3.9% |
| 7 | Elevator Technician | 22 | +5.7% |
| 8 | Aircraft Mechanic | 23 | +5.4% |
| 9 | Firefighter | 25 | +3.7% |
| 10 | Water Treatment Operator | 27 | -5.7% |
| 11 | Ship Engineer | 27 | +2.7% |
| 12 | Baker | 27 | +5.5% |
| 13 | Renewable Energy Technician | 27 | +29.5% |
| 14 | Veterinary Assistant | 28 | +9.1% |
| 15 | Plumber | 28 | +6.8% |
| 16 | Automotive Technician | 29 | +5.0% |
| 17 | Farmer | 32 | -3.4% |
| 18 | Paramedic | 32 | +5.7% |
| 19 | Dentist | 33 | +5.7% |
| 20 | Warehouse Operator | 34 | +8.9% |
| 21 | Train Operator | 35 | +0.5% |
| 22 | Electrician | 36 | +9.2% |
| 23 | Laboratory Technician | 37 | +2.7% |
| 24 | Mechanic | 38 | +17.8% |
| 25 | Power Plant Operator | 39 | -5.1% |
Exposure runs from 0 to 100 and compares occupations with each other, using public datasets (see sources). It is not a probability that the job disappears. Methodology
Sources
Exposure index calculated by ai-job-risk.net from the datasets below (data as of October 15, 2026). US labour figures are from the BLS, unchanged.
- Eloundou, Manning, Mishkin & Rock, "GPTs are GPTs: Labor market impact potential of LLMs", Science (2024); data MIT licence
- Gmyrek, Berg, Kamiński, Konopczyński, Ładna, Nafradi, Rosłaniec, Troszyński. Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140. Geneva: ILO, 2025. © ILO, CC BY 4.0
- Tomlinson, Jaffe, Wang, Counts & Suri, "Working with AI: Measuring the Occupational Implications of Generative AI", Microsoft Research; CC BY 4.0
- Anthropic Economic Index (job_exposure); CC BY
- U.S. Bureau of Labor Statistics, Employment Projections 2025–35 and AI exposure categories
The publishers did not prepare, review or endorse this site. Full citations, licences and method
What Keeps an Occupation Near the Bottom
Four conditions do most of the work here too. They are not the opposite of the ones above so much as the places where those conditions fail to hold.
The work happens in a physical place
Rewiring a panel, stabilising a patient or clearing a blocked line means being there, handling conditions nobody described in advance, with tools. Software can plan and document that work. It cannot perform it.
Someone has to carry the consequence
Where a wrong call causes real harm, clinical, legal, structural or physical, organisations keep a named person accountable. That accountability is a requirement of the role, not an inefficiency waiting to be removed.
Every case arrives different
Automation needs repetition. Roles where no two situations are alike, and where the right response depends on reading one specific person or site, give a model very little to generalise from.
Trust is part of what is being bought
In care, teaching, negotiation and crisis response the person is not a delivery mechanism for information. Their presence is what makes the exchange work, so removing them removes the thing being paid for.
What This List Does Not Mean
It is not a promise. A low position reflects the evidence available now. Capability that reached physical or high-stakes work would move these roles like any others.
It is not the whole job. Even in the least exposed occupations the administrative layer, the notes, reports, scheduling and claims, is being automated, and that changes how the day is spent.
It is not career advice. Exposure is one input among several. Pay, demand, entry cost and working conditions matter at least as much when deciding what to do.
Frequently asked questions
Q.What jobs are safest from AI?
The occupations at the bottom cluster in skilled trades, cleaning and personal services, construction, extraction and hands-on care. Each needs physical presence, and much of it needs judgment on site. The full ranking is on the rankings page, and each occupation has its own page explaining what is and is not exposed.
Q.Is any job completely safe from AI?
No, and a page claiming otherwise would be selling something. Almost every occupation on this list already uses AI for scheduling, documentation, triage or reporting. Low exposure means the part of the work that carries the responsibility still needs a person, not that the job is untouched.
Q.Are skilled trades really safer than office jobs?
On this measure, generally yes, and for a specific reason rather than sentiment. Trade work happens in places that are never quite the same twice, on equipment that fails in ways nobody documented, with consequences that are physical. Software can quote, schedule and document that work, and increasingly does. Performing it is a different problem, and progress there is measured in robotics, not in language models.
Q.How do you decide which jobs are safe?
Every occupation is measured with the same four public datasets and the same formula, so positions can be compared with each other. The sources, the formula and the known limits are set out on the methodology page. Figures change only when a source publishes a new edition.