Jobs

You will hear two confident stories about AI and work. One says the robots are taking every job. The other says relax, technology always creates more jobs than it destroys. Both stories are lazy. The honest picture is messier, more interesting, and worth getting right, because you are about to build a career inside it.

Tasks change before jobs do

A job is a bundle of tasks, and AI rarely swallows the whole bundle at once. It takes the tasks first: drafting the routine email, summarising the document, writing the boilerplate code, producing the first version of the design. A lawyer still exists, but the hours of document review inside that job shrink. Sometimes that means the same work needs fewer people. Sometimes it means each person handles more clients. Which way it goes depends on demand, management choices and how much of the job was routine to begin with.

Why does the same automation sometimes shrink a workforce and sometimes grow it? Because of what happens to demand. When a task gets cheaper, the thing it produces gets cheaper, and people usually buy more of a cheaper thing, so if demand grows faster than the efficiency saving, the industry ends up employing more people, not fewer. The classic example is the ATM. When cash machines spread through banking, it looked obvious that bank tellers were finished, because the machine did the teller's defining task. Instead, teller numbers in the United States kept growing for decades afterwards, because each branch now needed fewer tellers, which made branches cheaper to open, so banks opened many more branches, and the tellers who remained shifted from counting cash to service and sales. The task disappeared. The job changed shape and multiplied. Nobody can promise AI will follow the same path in every industry, but the mechanism is the one to understand: the effect on jobs depends on demand and on decisions, not on the technology alone.

A worked example: one firm, two futures

Run the arithmetic on one illustrative case. A small accounting firm has ten staff, and suppose routine work, the data entry, the standard returns, the first drafts of client letters, fills half of everyone's hours. An AI tool arrives that handles that routine half well. On paper the firm now needs five people. But walk through what the owner actually faces. In one future, the firm has a waiting list of small business clients it currently turns away, so it keeps all ten staff, serves nearly twice the clients and undercuts competitors who did not adopt the tool. In the other future, the local market is saturated, no new clients exist, and the owner quietly stops replacing people who leave. Same technology, same firm, opposite outcomes, decided entirely by demand and by choices. Multiply that fork across every industry and you have the honest answer to whether AI destroys jobs: it depends, and it depends on things humans decide.

Who is most exposed

The pattern so far is uncomfortable for students in particular. The tasks AI handles best are exactly the ones entry level knowledge workers used to cut their teeth on: research summaries, first drafts, basic analysis, routine correspondence. Routine knowledge work is more exposed than most trades, care work or anything needing hands and presence in the physical world. That flips the old advice that a desk job was the safe option, and it raises a real question employers have not answered: if AI does the junior tasks, where do juniors learn?

That ladder problem deserves its own moment, because you are the person standing at the bottom of the ladder. Expertise in most fields is built by doing routine work under supervision until judgement develops: the junior lawyer reads a thousand contracts and slowly learns what a bad clause looks like. If a model reads the contracts instead, the firm saves money this year and discovers in ten years that it has no seniors, because it never grew any. Some employers will solve this deliberately, by treating junior work as training rather than output and paying for it as such, and others will not. For you the practical consequence is blunt: do not skip the reps. When you hand AI your first draft, your assignment or your analysis, the output improves but you do not, and the thing employers will eventually pay for is the judgement that only builds when you do the work yourself often enough to know what good looks like.

What history actually says

The optimists are right about the long run. Farm mechanisation, electricity and computers each destroyed huge categories of work, and each time new industries eventually employed more people than were displaced. Nobody in 1990 imagined app developers or social media managers. But the honest version includes the second half: eventually can take a generation, and the person whose trade disappears at 45 rarely becomes the person hired by the new industry. Transitions are real, painful and unevenly shared. Both facts are true at once, and pretending otherwise is how people get blindsided.

What this means for the choices in front of you

None of this decides your career for you, but it should inform the choices you are already making. The old hierarchy that placed university desk jobs above trades has quietly inverted for many students, because an electrician or a nurse works in the physical world with real responsibility, exactly the combination current AI handles worst, while some of the graduate desk roles those degrees fed into are among the most exposed work in the economy. That does not make university a mistake. It means the question to ask about any degree, pathway or apprenticeship is no longer what job does this lead to, but what tasks will I be trusted with, and how many of those tasks need a human's judgement, presence or accountability. A subject that teaches you to think, argue and take responsibility travels well. A subject that only teaches you to produce routine output is training you to race a machine that does not sleep.

The skills that hold their value

• Judgement: deciding what should be done when the answer is not in any document, and owning the call.
• Responsibility: someone must be accountable when things go wrong, and it will not be the model.
• Relationships: trust between people is built by people. Clients, patients and teammates want a human who knows them.
• Physical skill: electricians, nurses, chefs and mechanics work in the physical world, where AI progress is much slower.
• Taste: knowing which of ten plausible outputs is actually good. The more machines generate, the more this is worth.

This lesson leans on its neighbours more than it may seem. Who captures the gains from automation, whether the people displaced are supported, and whether anyone stays accountable when a system takes over a task badly are all versions of the question this whole topic keeps asking: who decides, and who bears the cost. Keep that question in your pocket, because it works on every AI story you will ever read.