What Jobs Are Safe From AI? An Honest Answer
The study everyone cites as a list of jobs at risk says it measures no such thing. What the research actually found, and what it means for your job.
Search this question and you get a lot of lists. Ten jobs AI will never touch. Forty jobs at risk. The lists disagree with each other, and most of them are built on the same piece of research — which, awkwardly, says it does not measure what they claim it measures.
So here is the honest version, starting with the part nobody puts in a headline.
The study everyone cites, and what it actually says
Most "jobs at risk" lists trace back to a 2025 Microsoft Research paper, "Working with AI: Measuring the Applicability of Generative AI to Occupations". The researchers took roughly 200,000 anonymised Bing Copilot conversations from January to September 2024, worked out which tasks people were actually asking AI to help with, and mapped those tasks onto the O*NET database of what each occupation involves. The output was an "AI applicability score" per occupation.
That got reported as a risk ranking. The authors then published a follow-up note specifically to say it was not one.
Our study does not draw any conclusions about jobs being eliminated; in the paper, we explicitly cautioned against using our findings to make that conclusion.
They went further: "our data do not indicate, nor did we suggest, that certain jobs will be replaced by AI." And the line that matters most for anyone reading a list with their own job on it — "a job is far more than the collection of tasks that make it up."
That is not a technicality. It is the whole difference between "AI can help with things you do" and "AI can do what you do". The research measured the first. Every list that ranks your job by doom is quietly claiming the second.
What the research did find
Taken for what it is, the finding is useful. AI applicability clusters around information work — creating, processing, and communicating information. The occupations scoring highest were things like translators, writers, historians, sales roles, office and administrative support, and computer and mathematical work. If your day is mostly words, numbers, and messages, a lot of your individual tasks are things a model can assist with.
At the other end were jobs defined by physical work and presence: dredge operators, bridge and lock tenders, water treatment plant operators, phlebotomists, nursing assistants. Not because those jobs are more valuable, but because a chat window cannot draw blood or operate a lock — and, separately, because a physical robot capable of it is a different and far more expensive proposition than a software subscription.
The authors are also clear about the limits of the data. It reflects who was comfortable using Bing Copilot in 2024, and it cannot reliably tell work use from personal use. O*NET describes tasks, not the judgement, relationships, and accountability that make up the rest of a job.
Why "safe" is the wrong question
Almost nobody loses a whole job to a model. What happens is that some part of the job gets faster, and then the shape of the role changes around it.
A paralegal still runs discovery, but spends less time on first-pass document review. A marketer still owns the campaign, but writes fewer first drafts from scratch. The title survives. The mix of what fills the week does not. That is the actual mechanism, and it is why job-title lists mislead — two people with the same title can have very different exposure depending on which tasks fill their day.
Which means the useful question is not "is my job on the safe list". It is: which specific tasks in my week are information tasks, and what happens to my role when those take a quarter as long?
Ask which of your tasks are exposed, not whether your title is. Two people with the same job title can have completely different answers.
What the labour market data actually shows
It is worth being precise here, because this is where the scary numbers live. The employer forecasts for the rest of the decade are covered separately in what jobs AI will replace by 2030; what follows is what has actually been measured.
Challenger, Gray & Christmas, which tracks announced US job cuts and the reasons employers give for them, attributed 54,836 job cuts to AI in 2025 — more than the two previous years combined, which totalled 71,683 cumulatively since 2023. That is a real and rising number, and it is also a small fraction of total US layoffs. It counts what employers said, which is not always the whole story: "AI" is a more palatable public reason than "we over-hired in 2022" — AI and job loss goes through what the count includes and misses.
The other half of the picture is where the money went. PwC's 2026 Global AI Jobs Barometer, which analyses close to a billion job advertisements across 24 countries, found that roles asking for AI skills carry a 62% wage premium over otherwise comparable roles — up from 57% the year before. Postings for AI-skilled roles grew 69% against 9% for the wider jobs market.
Read together, those two datasets say something more specific than "AI is coming for jobs". They say the market is separating people who use these tools from people who do not, and paying accordingly.
The jobs least likely to be disrupted
If you want the list anyway, here is the honest version of it. Work is harder to hand to a model when it involves:
- Physical presence and manual skill — trades, maintenance, hands-on healthcare, operating equipment.
- Being accountable for the decision — jobs where someone has to sign their name and carry the consequence.
- Relationships as the actual product — long-term client trust, negotiation, care work, teaching a specific person.
- Unstructured, messy environments where the inputs are never twice the same.
- Work where being wrong is expensive and hard to detect, which is exactly where these tools are weakest.
That last one is worth sitting with. The failure mode of these systems is not that they break loudly. It is that they produce fluent, confident, wrong answers that look exactly like right ones — which is covered in more detail in What AI Is Actually Bad At. Jobs where nobody would notice a plausible error for six months are not safe jobs. They are jobs where the checking becomes the work.
What to do about it
Nothing on this list requires a career change.
- Write down your actual week — the tasks, not the title. Mark which ones are producing, summarising, or reformatting information. That is your exposure, and it is more accurate than any published list.
- Take the two most repetitive of those and learn to do them with AI properly. Not "I tried ChatGPT once" — properly, to the point where the output needs light editing rather than a rewrite.
- Get good at checking output you cannot verify at a glance, because that skill is what makes the rest of it safe to rely on.
- Keep the parts of your job that are judgement, relationships, and accountability. Those are the parts that were always the job.
If you want a starting point for step two, How to Write a Prompt That Works on the First Try covers the request side, and How to Check an AI Answer When You Are Not the Expert covers the verification side. Between them they are most of what "being good with AI" actually means in an office.
The short version
No credible research says your job is on a list to be deleted. The best study we have says AI is applicable to a lot of information tasks and explicitly refuses to draw the further conclusion. The labour data shows a real but modest number of AI-attributed cuts, and a large and growing pay gap in favour of people with the skills.
The safest position is not a job title. It is being the person on the team who knows what these tools are good for, where they fail, and how to check them. That is learnable, and it is a smaller job than changing careers.
Coursium teaches exactly that — the practical use, the failure modes, and the checking. Stay ahead of AI by learning the tools on your phone.