Here’s a prediction you’ve probably encountered, in various forms, across the last decade.
Robots and artificial intelligence will automate most existing jobs within a generation. The middle-class careers today’s students are preparing for will largely cease to exist. Millions of workers will be displaced. Either society will figure out a radical solution — universal basic income, massive retraining, entirely new economic arrangements — or large portions of the population will be left permanently unemployed, their skills obsolete, their role in the economy eliminated.
This prediction has been widespread, and influential, and largely — on the current evidence — wrong. Or at least, wrong in the way it was commonly framed. What’s actually happening in the labour market is more complex, more interesting, and arguably more important to understand honestly. The future of work is genuinely changing, but not in the sweeping all-or-nothing way the popular narrative suggests.
What the research actually shows
The most influential pessimistic estimate came from a 2013 paper by two Oxford researchers, Carl Benedikt Frey and Michael Osborne, who estimated that 47 per cent of US jobs were at high risk of automation within about two decades. The paper received enormous attention and shaped much of the subsequent public discussion.
More careful subsequent research has produced more modest and more nuanced estimates. A 2018 OECD study by Ljubica Nedelkoska and Glenda Quintini revisited the Frey-Osborne methodology using a different approach — examining specific tasks within jobs rather than treating entire jobs as single units. Their finding was that about 14 per cent of jobs across OECD countries were at high risk of automation, with another 32 per cent likely to see significant changes in task composition. This is still substantial, but it’s a third of what the Frey-Osborne figure suggested, and the mechanism is different.
The emerging consensus, across much of this research, is that jobs don’t typically disappear wholesale. The tasks within jobs change. Some tasks become automated; other tasks within the same job remain, or expand in importance. New tasks emerge that didn’t exist before. The job title may persist while the day-to-day work within it transforms substantially.
The American economist David Autor at MIT has done much of the most careful empirical work on this. His broader finding, across several decades of labour-market research, is that technological change has historically produced job polarisation rather than wholesale job elimination — growth in high-skill and low-skill employment, with hollowing out of middle-skill routine jobs. The displaced middle-skill workers have often moved into expanded low-skill service work, at lower wages, rather than being eliminated from employment entirely. The pattern has been severe in its consequences for specific groups of workers, but it doesn’t match the dramatic mass-unemployment narrative.
What automation actually automates
Understanding which kinds of work are vulnerable requires understanding what automation does well.
Tasks that are routine and codifiable — following specific rules, processing structured information, executing standardised procedures — have been vulnerable to automation for decades. This is why manufacturing employment has fallen so dramatically in advanced economies (not primarily because of offshoring, though that contributed, but because of automation of routine production tasks). It’s also why back-office clerical work has been reduced — routine data entry, routine document processing, routine calculation.
Tasks that require genuine judgement under uncertainty, interpersonal skills, creative problem-solving, or physical dexterity in unstructured environments have been much harder to automate. A nurse’s work of assessing a patient’s condition, responding to emotional distress, and adapting care to individual circumstances is far from automation. A plumber’s work of diagnosing an unfamiliar problem in an unfamiliar house is far from automation. A teacher’s work of noticing which students are struggling with what and adapting accordingly is far from automation. These aren’t jobs that require extraordinary skills; they’re jobs that require ordinary human capacities that are extraordinarily hard for machines to replicate.
The recent emergence of generative AI has complicated this picture. Some tasks that were previously thought safe from automation — routine writing, analysis, certain creative work, software engineering — have turned out to be partially automatable in ways that weren’t anticipated a decade ago. But the emerging evidence, discussed in the earlier article on AI and thinking, suggests that generative AI typically augments rather than replaces workers, at least in the current generation of technology. Productivity rises; skill distributions compress; specific routine cognitive tasks within jobs get automated. The wholesale replacement of workers has not materialised, even in the fields most directly affected.
The specific pattern that matters
A particular research finding worth understanding, from researchers including the American economist Daron Acemoglu, is that the effect of automation on workers depends heavily on whether the technology is primarily replacing human labour or complementing it.
Replacement technologies take over tasks humans used to do, reducing the demand for human workers in those tasks. Complementary technologies make human workers more productive, increasing the value they produce without eliminating their role. Both kinds of technology have always existed. Whether an era is good or bad for workers depends substantially on the balance between them.
Acemoglu and his collaborator Pascual Restrepo have argued that the current period has tilted too heavily toward replacement technologies, in ways that have specific policy causes — the tax system favouring capital investment over labour, research priorities that emphasise replacing workers over augmenting them, the incentives of firms to reduce headcount rather than raise productivity per worker. This is, at least in principle, a reversible situation. Different policies could tilt the balance toward more complementary technology and better labour-market outcomes.
This matters because it suggests the future of work isn’t predetermined by technology. Technology shapes what’s possible; policy and institutions shape what actually happens. The claim that AI will inevitably destroy jobs treats technology as an autonomous force outside human control. The more careful research suggests that the specific outcomes depend on specific choices that societies make about how to deploy technology, how to educate workers, how to structure labour markets, and how to distribute the gains from productivity growth.
What this suggests for individual planning
For a student making decisions about education and career, several working principles seem defensible given what’s actually known about the future of work.
Uncertainty is large, and strong predictions are usually wrong. People who confidently told students in 2005 what the labour market would look like in 2025 were mostly wrong in specific ways. The same applies to people confidently predicting 2045 from 2025. General flexibility and ongoing learning capacity are probably more valuable than betting on specific predictions about which jobs will exist.
Skills that combine human judgement with technical capability have held their value. Nursing, teaching, skilled trades, engineering, research, caring professions, jobs that require reading human situations and responding to them — these have remained valuable across decades of technological change. They are likely to remain valuable even as specific tasks within them shift.
Pure routine cognitive work is less safe than it used to be. Jobs that primarily consist of producing standard documents, executing standard analyses, handling standard enquiries — the white-collar equivalent of routine factory work — have already begun to change substantially under automation pressure. Students planning careers in these areas should expect the work to change substantially within their careers, and should build capacities that extend beyond the current routine version.
Learning how to learn is probably the single highest-leverage capacity. The specific content of what you’ve learned by age twenty-two will matter less over a career than the capacity to keep learning new things as the world changes around you. The students who do best across long time horizons are usually the ones who develop robust learning habits rather than specific fixed skill sets.
Don’t over-weight the AI-specific risks. Yes, generative AI will change many fields. No, it won’t instantly eliminate most jobs. The students most at risk are those who bet everything on tasks currently vulnerable to automation — pure routine writing, simple data analysis, entry-level coding. The students less at risk are those who build depth in judgement, relationship, creativity, or domain expertise that AI can augment but not replace.
The counter-thread worth hearing
A reasonable caveat: past patterns may not fully predict future ones. Previous waves of automation — from agriculture to manufacturing, from manufacturing to services — produced job polarisation rather than mass unemployment, because new kinds of work emerged as old kinds declined. Whether this pattern will continue under the next wave of technology is genuinely uncertain. Some serious researchers, including the economic historian Robert Gordon, argue that the economic gains from new technology may be smaller than in previous eras, limiting the growth of new jobs to replace displaced ones.
And the distributional effects matter even when aggregate employment doesn’t collapse. Job polarisation, even without mass unemployment, has produced significant harm to specific groups of workers — particularly men in manufacturing, workers in rural areas, workers with middle levels of education who lost well-paid routine jobs. The broader society’s employment picture may look stable while specific communities experience severe economic damage. This isn’t comforting if you’re in one of those communities.
So the honest picture is: mass technological unemployment is unlikely on current evidence, but significant distributional harm within overall stable employment is certain, and the specific policy responses to these dynamics are going to shape what kind of society emerges over the next few decades.
The question that remains
The deepest thing worth understanding about the future of work is that it’s being made, not predicted. Technology sets possibilities. Choices — about policy, about education, about investment priorities, about how gains from productivity are distributed — determine which possibilities actually become reality. The future isn’t something that happens to us. It’s something we make through aggregated decisions, most of which are made by institutions whose actions citizens can, at least in principle, influence.
For a student about to enter adult life, this is both empowering and sobering. Your individual career strategy matters, but it won’t protect you from decisions made by the society you live in. Your citizenship — your voice in how those larger decisions get made — matters too. The students who thrive over the next several decades will be those who build both individual capabilities and collective commitments to the kind of society they want to work in.
The question worth carrying, as you think about your own future work:
What kinds of work are most likely to still be meaningful in twenty years — and are you building toward any of them, or is the path you’re currently on one someone would have recommended decades ago based on a world that is passing?
Key research referenced: Carl Benedikt Frey and Michael Osborne’s 2013 Oxford paper on automation risk; Ljubica Nedelkoska and Glenda Quintini’s 2018 OECD study; David Autor’s research on job polarisation; Daron Acemoglu and Pascual Restrepo’s research on replacing vs. complementary technologies; Robert Gordon, The Rise and Fall of American Growth (2016).