What Jobs Can AI Not Replace?
Key Takeaways
- No job is guaranteed to be immune from AI. The honest question is not which jobs are “safe forever” but which kinds of work resist automation and why.
- The best available evidence comes from three sources: Microsoft Research’s study of 200,000 real Bing Copilot conversations, Anthropic’s Economic Index measuring observed Claude usage, and the World Economic Forum’s Future of Jobs Report 2025.
- Jobs with the lowest AI applicability are physical, hands-on and on-site: equipment operators, roofers, nursing assistants, cooks, lifeguards and farm workers.
- Jobs built on licensed judgement, human trust and accountability — therapists, trial lawyers, surgeons, senior managers — also resist replacement, because someone has to be answerable for the decision.
- Around 30% of US workers score zero on Anthropic’s observed exposure measure: their tasks barely appear in AI usage data at all.
- The counter-argument is real: “cannot replace” is a claim about today’s AI, not a permanent guarantee. Robotics, better models and cheaper deployment can move the line. The durable strategy is to build skills AI complements rather than competes with.
Why This Question Is Being Asked Now
“What jobs can AI not replace?” is trending as a search query, rising about 120% under the broader “ai jobs” topic. The timing is easy to explain. 2026 has been a year of conflicting signals: AI tools have become dramatically more capable, companies have announced AI-related job cuts, and at the same time the economy-wide evidence of mass displacement has remained thin.
Consider the tension. Anthropic’s chief executive Dario Amodei predicted in 2025 that AI could eliminate half of all entry-level white-collar jobs within a few years — a forecast that was widely reported and widely disputed. Meanwhile, the Stanford Digital Economy Lab’s August 2026 update to its “Canaries in the Coal Mine?” study, using ADP payroll data through June 2026, found that the employment gap for workers aged 22 to 25 in highly AI-exposed occupations had widened to 19% compared with their peers in less-exposed jobs — yet the researchers said they did not see widespread, economy-wide job displacement associated with AI. The Yale Budget Lab’s AI Labor Market Tracker, updated in September 2026 with August 2026 survey data, likewise found no clear evidence of labour-market disruption associated with AI.
So the picture is neither “AI is replacing everyone” nor “nothing is changing”. Young workers in exposed occupations are finding it harder to get hired, while the overall labour market has not collapsed. That is the context in which people ask what jobs AI cannot replace — and it is why the answer needs evidence, not reassurance.
The Three Studies That Actually Measure This
Most lists of “AI-proof jobs” are opinion dressed up as analysis. Three research efforts are worth taking seriously because they measure real behaviour rather than guessing.
1. Microsoft Research: “Working with AI” (July 2025)
Microsoft Research published “Working with AI: Measuring the Occupational Implications of Generative AI” in July 2025. The team analysed 200,000 real, anonymised conversations with Bing Copilot from January to September 2024, and mapped the tasks in those conversations to the US O*NET occupational database. From that mapping they built an “AI applicability score” for each occupation — a measure of how much of a job’s daily work AI is already performing successfully.
The headline finding: information gathering, writing, and communicating with others are the most common things people ask Copilot to help with. Occupations with the highest applicability scores are knowledge work and communication-focused: interpreters and translators scored highest at around 0.49, followed by historians, writers, sales representatives, customer service agents and some teaching and administrative roles. Computer and mathematical occupations scored 0.29; office and administrative support scored 0.26; management scored 0.13.
At the other end, the occupations with the lowest applicability scores were physical, hands-on and on-site. The study’s lowest-scoring occupations included dredge operators, roofers, logging equipment operators, orderlies and nursing assistants. The authors explicitly warned that a high applicability score does not predict whether an occupation will disappear — it measures where AI can already do the tasks, not whether employers will eliminate the job.
Two caveats matter. First, the study measures language-model capabilities in conversation, not robotics, sensors or specialised automation — so a job that scores low today could still be affected by a robot tomorrow. Second, the data is a snapshot of 2024 capabilities. Newer models can do more, so the scores would look somewhat different if the study were repeated today. The study is the best available map of where generative AI touches work, but it is a map of the present, not a forecast.
2. Anthropic Economic Index: observed exposure (2026)
Anthropic, the company behind Claude, published research in March 2026 introducing a metric called “observed exposure”, which tracks how Claude is actually being used in real workplaces — not what AI could theoretically do. The index draws on the O*NET database covering roughly 800 US occupations, Claude’s own usage logs, and an academic framework that scores whether an AI can cut a task’s completion time in half.
The findings are striking in both directions. Computer programmers top the list at around 75% task coverage, with usage leaning toward full automation rather than just assistance. Customer service representatives come second, and data-entry workers third at around 67% coverage. Overall, AI could theoretically handle about 90% of tasks in office and admin roles, but observed usage covers only about a third of computer and maths jobs — the most penetrated category.
And then the other side: about 30% of US workers score a flat zero. Their tasks simply do not appear in AI usage data at any meaningful level. The zero-exposure occupations highlighted in the research include cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, dressing-room attendants, agricultural workers, and courtroom lawyers. The common thread is physical presence, sensory judgement, and reading a room in real time. A language model has no body, no hands and no instincts. Those still matter.
Anthropic’s data also lines up with official statistics: the US Bureau of Labor Statistics’ employment projections through 2034 show that for every 10 percentage point increase in a job’s AI coverage score, projected growth for that role drops by 0.6 percentage points. That is a modest effect, but it points in the expected direction — more exposure, slower growth.
3. World Economic Forum: Future of Jobs Report 2025
The World Economic Forum’s Future of Jobs Report 2025, published in January 2025, surveyed more than 1,000 companies across 22 industries and 55 economies. It is a survey of employer expectations, not a measure of observed employment, and that distinction matters. Employers expect 92 million jobs to be displaced by 2030 and 170 million new jobs to be created — a net gain of 78 million. The same report lists clerical and secretarial roles among the categories with the largest expected net decline, and cashiers and administrative assistants among the fastest-declining roles to 2030.
On the growth side, the WEF reports employer expectations of growth for delivery drivers, care roles, educators and farmworkers — roles that involve moving through the physical world or caring for people. The International Labour Organization and the Polish research institute NASK, in a 2025 index, found that one in four jobs worldwide is potentially exposed to generative AI, but concluded that few jobs consist of tasks current AI can fully automate — so job transformation, not elimination, is the most likely outcome. The IMF’s research, restated by Managing Director Kristalina Georgieva in October 2026, finds that roughly 40% of jobs worldwide are exposed to AI, rising to about 60% in advanced economies.
The Four Kinds of Work That Resist Automation
Across all three studies, the same four categories keep appearing. They are not job titles; they are kinds of work. A job is resistant to AI to the extent that it depends on one or more of them.
1. Physical dexterity in unpredictable environments
This is the strongest and most consistent finding. Jobs that require a body moving through a messy, unpredictable physical world are the hardest for today’s AI to touch — because today’s AI is mostly software, and software cannot lift, climb, feel or react to a physical surprise.
Roofers work on sloped surfaces in changing weather, handling materials that bend and break. Plumbers diagnose a leak hidden inside a wall and improvise a fix with the parts at hand. Electricians trace a fault through a building’s wiring, where every installation is slightly different. Cooks judge doneness by touch, smell and timing, and adjust in real time. Lifeguards scan water and execute physical rescues. Agricultural workers prune trees and operate machinery outdoors, where the ground is uneven and the weather does not cooperate.
The Microsoft study puts these occupations at the very bottom of its applicability ranking, and Anthropic’s index gives many of them a zero exposure score. The reason is not that these jobs are simple — they are cognitively demanding in ways that are easy to underestimate. It is that their difficulty lives in the body and in the environment, not in text. A language model can describe how to fix a pipe. It cannot crawl under a sink.
The honest caveat: robotics is the wildcard. A humanoid robot with good enough dexterity and judgement could eventually do some of this work. But capable, affordable, general-purpose robots for messy environments do not exist yet at scale, and the gap between a laboratory demo and a robot that can reliably rewire a house is large. Physical work is the safest category for the foreseeable future — measured in years, not forever.
2. Human trust and emotional connection
The second category is work where the value lies in the relationship itself, not just in the task performed. A therapist’s job is not to produce correct sentences about mental health; it is to be a trusted person in the room over months and years. A nurse’s job is not only to administer treatment; it is to notice when something “just doesn’t feel right” with a patient, and to offer comfort during one of the hardest moments of that person’s life. A teacher’s job is not only to deliver content; it is to manage a lively classroom, spot when a pupil is struggling at home, and build the confidence that helps a young person thrive.
AI can already do parts of these jobs. It can draft a lesson plan, summarise a patient’s notes, or suggest a therapeutic exercise. What it cannot do is be the person a patient trusts, or the adult a child feels safe with. Trust is not a feature of the output; it is a property of the relationship, and it depends on the other party being a person who understands what is at stake.
Survey evidence supports the intuition. Reed.co.uk’s research found that 84% of the hiring decision-makers it surveyed agree employers should prioritise human workers over AI wherever possible, and 60% of people planning a job move believe AI is reducing job opportunities in the UK. Those are attitudes, not measurements — but they signal that employers and workers alike still place a premium on human presence in care, education and support roles.
The caveat here is subtler than in the physical category. AI companions and AI therapists already exist, and some people prefer them for some purposes. The question is not whether AI can simulate empathy — it can, convincingly — but whether people will accept a simulation when the stakes are high: a mental health crisis, a child’s education, end-of-life care. The evidence so far suggests that for high-stakes human connection, most people still want a person. That preference may soften over time, but it has not collapsed.
3. Accountability and licensed judgement
The third category is work where someone has to be answerable. A surgeon who makes a decision in the operating theatre is legally and professionally responsible for it. A trial lawyer who stands up in court is accountable for the argument and for the client’s interests. A judge’s ruling carries the authority of the state. A senior executive who signs off on a strategy owns the consequences.
AI can inform all of these decisions. It can summarise case law, flag risks in a contract, or model the outcomes of a business strategy. But it cannot take responsibility. There is no legal mechanism by which an AI system can be sued, struck off, or held to account — and there is no professional body that can discipline it. That is why the Anthropic index lists courtroom lawyers among the zero-exposure occupations, and why senior management scores low on Microsoft’s applicability measure despite being heavily involved in information work.
This is also the category most likely to be misunderstood. “Accountability” sounds like a legal technicality, but it is really about trust in institutions. A society that lets an AI system make final decisions about someone’s liberty, health or money — with no human to answer for the outcome — would be a society that has given up on a basic idea: that decisions affecting people should be made by people who can be questioned. That idea is not a technical limitation. It is a choice, and it is one that institutions are likely to keep making for a long time.
That said, the boundary is moving. AI already drafts legal documents, reviews contracts and assists with diagnosis. The judgement-heavy core of these professions is resistant, but the routine work around it is not — and a profession can shrink even if its core survives. The Palantir chief executive Alex Karp has given blunt career advice on exactly this point, arguing that young people should build skills AI cannot easily replicate; our coverage of Alex Karp’s AI career advice sets out his argument in full.
4. Novel judgement in ambiguous situations
The fourth category is work where the right answer is not in the training data, because the situation is new. Entrepreneurs deciding whether to enter a market that does not exist yet. Diplomats negotiating a deal where the other side’s red lines are unknown. Scientists designing an experiment to test an idea nobody has had before. Crisis managers responding to an event that has never happened.
AI systems are, at their core, pattern matchers trained on the past. They are extraordinarily good at situations that resemble something they have seen. They are much weaker at situations that do not — where the information is incomplete, the stakeholders disagree about the facts, and the decision has to be made anyway. That weakness is structural, not a matter of more training data: a genuinely novel situation, by definition, has no precedent to learn from.
This is also the category where AI is most useful as a partner rather than a replacement. A founder who uses AI to analyse a market, model scenarios and draft a plan is more capable than one who does not — but the judgement call at the end, the decision to bet the company on an unproven idea, remains human. The same logic applies to creative direction: AI can generate a hundred options, but choosing which one is right requires taste, context and a sense of what the audience will feel. That is why creative directors and design leads score relatively high on AI resistance even though AI can produce images and copy on demand.
The Jobs Most Exposed — and What That Tells Us
It is worth naming the other side of the picture, because it sharpens the answer. The roles most exposed to AI are routine, text-heavy, entry-level white-collar jobs: data entry, basic copywriting, junior coding, customer service, translation, and administrative support. Microsoft’s study puts interpreters, translators, writers and customer service representatives at the top of its applicability ranking. Anthropic’s index puts computer programmers at 75% task coverage and data-entry workers at 67%.
That pattern is consistent with what is already happening in the labour market. Aiera.blog has covered the AI impact on the labour market and the worst-hit jobs in 2026, and the NatWest case — 1,000 roles cut while new AI roles were created, as covered in our report on the NatWest AI job cuts — illustrates the same dynamic: jobs are not simply disappearing; they are being restructured, with routine tasks automated and new roles created around the technology.
The lesson is not that white-collar work is doomed. It is that the routine parts of white-collar work are the most automatable, and the people whose jobs are mostly routine tasks are the most exposed. The defence is to move up the value chain within your own profession: from producing output to exercising judgement about it, from executing tasks to owning outcomes, from knowing things to knowing what to do when the facts are unclear.
The Honest Counter-Arguments
A good answer to this question has to include the reasons it might be wrong. Here are the three strongest objections.
“Cannot replace” is a claim about today, not forever. Every job on the resistant list is resistant to today’s AI. Robotics could erode the physical category; social acceptance could erode the trust category; legal reform could erode the accountability category; better models could erode the novel-judgement category. The history of automation is a history of confident predictions about what machines could never do, followed by machines doing it. Anyone who tells you a job is safe forever is selling you something.
Partial automation can shrink a profession without eliminating it. The ILO–NASK researchers warn about exactly this: even when a job survives, partial automation can lower its quality, reduce the number of people it employs, or push down its pay. If AI lets one lawyer do the work of three, the profession of law survives — but two of the three lawyers need new jobs. “AI cannot replace lawyers” and “AI will reduce the number of lawyers” can both be true at the same time.
Exposure is not evenly distributed. The IMF’s January 2026 analysis finds that high-skill and low-skill workers tend to gain most from demand for new skills, while middle-skill roles such as routine office jobs are being squeezed. The ILO finds that women are far more likely than men to hold jobs in the highest generative-AI exposure category, largely because women are overrepresented in clerical work. The question “what jobs can AI not replace” has a different answer depending on who is asking, and the people most exposed are often the people with the least power to retrain.
What To Do With This Information
If you are choosing a career, changing one, or trying to make your current job more durable, the research points to a practical strategy:
- Build skills AI complements, not skills it competes with. If your job is mostly producing text, data or routine analysis, AI will do that part faster and cheaper. The durable parts of your job are the judgement, the relationships and the accountability. Invest in those.
- Get close to the physical world or to people in high-stakes moments. Skilled trades, healthcare, education, emergency services and hands-on technical work all score low on every measure of AI exposure.
- Learn to use AI rather than ignoring it. The people most at risk are not those whose jobs AI can do — they are those whose jobs AI can do when operated by someone else. Andrej Karpathy has described a growing gap between frontier AI capabilities and everyday use; our coverage of Karpathy on the AI adoption gap explains why closing that gap is itself a valuable skill.
- Own outcomes, not tasks. A task can be automated. An outcome — a healthy patient, a winning case, a functioning building, a team that trusts its leader — requires someone to be responsible for it. The more your job is defined by outcomes, the harder it is to replace.
None of this is a guarantee. It is a strategy based on the best available evidence, and the evidence will change. The most durable career asset in an age of AI may be the willingness to keep learning — because the line between what AI can and cannot do is going to keep moving, and the people who move with it will do better than the people who stand still.
What We Could Not Verify
- Anthropic Economic Index exact figures. The task-coverage figures cited here (programmers around 75%, data-entry workers around 67%, office and admin roles around 90% theoretical coverage, 30% of US workers at zero observed exposure) come from secondary reporting of Anthropic’s March 2026 research. We did not independently verify them against Anthropic’s primary publication in this session.
- The +120% trend figure. The reported rise of about 120% for the query “what jobs can ai not replace” is given in the assignment brief. We did not independently verify it against Google Trends in this session.
- Reed.co.uk survey details. The figures cited (84% of hiring decision-makers, 60% of people planning a job move) come from Reed’s published article. The survey’s sample size, fieldwork dates and methodology were not independently verified.
- Stanford Digital Economy Lab’s 19% gap. The reported widening of the employment gap for 22 to 25 year olds in exposed roles to 19% through mid-2026 comes from secondary reporting of the August 2026 update. We did not independently verify it against the study itself.
- Dario Amodei’s 50% prediction. This is a forecast, not a fact. It was reported in 2025 and remains contested; it is included here as an attributed prediction, not as evidence of what will happen.
Sources
- Microsoft Research, “Working with AI: Measuring the Occupational Implications of Generative AI”, July 2025 — https://en.papernotes.org/NeurIPS2026/social_computing/working_with_ai_measuring_the_applicability_of_generative_ai_to_occupations/
- Anthropic Economic Index research, March 2026, via TheStreet — https://www.thestreet.com/economy/anthropic-finally-reveals-which-jobs-ai-cannot-replace
- World Economic Forum, Future of Jobs Report 2025 (January 2025), as summarised and discussed in Layer3 Labs — https://www.layer3labs.io/guides/what-jobs-will-ai-replace
- Pouya M., “How AI Will Affect Jobs: Evidence, Risks and Opportunities” — https://www.pouyam.com/guides/how-ai-will-affect-jobs
- Reed.co.uk, “Jobs AI can’t replace” — https://www.reed.co.uk/career-advice/jobs-ai-cant-replace/
- Articuler, “What Jobs Are Safe From AI? 30 AI-Proof Jobs” — https://www.articuler.ai/resources/compare/ai-proof-jobs/