The Labour Question: What Happens When AI Does the Work?

A worker at a General Motors assembly plant — call him Joe — is in his early thirties, earning enough to own a house, to raise three children, to take a vacation every summer. He has a union card, seniority, and the reasonable expectation that if he keeps his head down and does his job, he will retire with a pension in thirty years. His father worked in the same plant. His brother-in-law works down the line.

Detroit, Michigan. 1979. A worker at a General Motors assembly plant — call him Joe — is in his early thirties, earning enough to own a house, to raise three children, to take a vacation every summer. He has a union card, seniority, and the reasonable expectation that if he keeps his head down and does his job, he will retire with a pension in thirty years. His father worked in the same plant. His brother-in-law works down the line.

The robots arrive in 1981. Not all at once — gradually, one station at a time, as the industrial automation that has been transforming manufacturing throughout the century accelerates. The robots do not replace Joe immediately. They change the work around him, create new jobs maintaining them, reshape the plant’s labour needs in ways that play out over years.

But by 1990, GM has cut its Detroit workforce by half. Joe has survived two rounds of layoffs but his brother-in-law has not. The union has negotiated severance packages and retraining programmes, but the retraining has not gone well — the skills required for the new economy are different from the skills the plant workers have, and the new jobs, when they exist, pay less.

Joe retires in 2009, just as GM goes bankrupt. His pension is partially protected by the federal bailout. His children do not work in manufacturing.

This is not a story about AI. It is a story about what happens when technology transforms the labour market — the specific, granular, human reality of an economic transition. The story of what AI is doing to work will rhyme with this story, in some ways. In other ways, it will be different. Understanding which ways are which is the most important economic question of the current era.

---

The Historical Pattern: Technology and Work Through Time

The worry that technology destroys jobs is as old as industrial technology. The Luddites — English textile workers in the early nineteenth century who destroyed machinery they believed threatened their livelihoods — gave their name to a specific kind of anti-technological resistance. The automaton weavers and spinning frames of the industrial revolution were the GPT-4 of their era: extraordinarily capable at specific tasks, threatening to specific categories of workers, and genuinely transformative of the structure of employment.

The pattern since the industrial revolution has been consistent at the aggregate level: technology has continuously displaced specific categories of work while creating new categories of work, and the total volume of employment has generally grown rather than shrunk. The agricultural workers displaced by mechanised farming found work in factories. The factory workers displaced by automation found work in services. The service workers facing AI automation are finding new work in — well, that is still being determined.

At the aggregate level, the historical record supports optimism: every previous wave of technological automation has been followed by the creation of new work, and employment and living standards have generally improved over time. Technological unemployment at the aggregate level has not, historically, proved permanent.

But the aggregate level conceals important realities. The transition from one type of work to another is not smooth, not fast, and not equally distributed. The agricultural workers who moved to factories moved over decades, not years. Many of them did not move — they aged out of the workforce, or moved into poverty, or died before the new jobs appeared. The communities that lost their economic base did not bounce back automatically; many did not bounce back at all.

The historical lesson is not that technology creates as many jobs as it destroys and that workers should not worry. The historical lesson is that technology creates different kinds of jobs than it destroys, in different places and requiring different skills, on a timescale that does not automatically align with workers’ lifetimes or communities’ economic viability. The transition is real, the costs are concentrated, and policy choices determine how severe and how equitably distributed those costs are.

---

What Makes AI Different: The Cognitive Dimension

Previous waves of automation primarily displaced physical labour. The industrial revolution mechanised repetitive manual tasks — weaving, forging, assembling. Agricultural mechanisation displaced farm labour. Factory automation displaced assembly work. These transformations affected workers whose primary contribution was physical — strength, dexterity, endurance — and the new jobs they created were often cognitive: designing machines, programming computers, managing organisations, providing services.

AI automation is different in a specific way: it displaces cognitive labour. The tasks that AI systems can now do — writing, analysing, coding, translating, advising — are tasks that require not just physical capability but thinking. This is the aspect of AI automation that is historically novel and that makes the standard reassurances about technology creating new jobs somewhat less reliable as a guide to the current transition.

The argument that new technology creates new jobs has historically relied on the observation that displaced workers can move into new roles that require cognitive capabilities that machines cannot replicate — creativity, judgment, social intelligence. This argument assumes that cognitive work is the refuge from automation. If AI can do cognitive work, the refuge becomes smaller.

The specific tasks most clearly within AI’s current capabilities are cognitive tasks that are rule-governed, text-based, and do not require physical presence or embodied understanding. Writing, research, analysis, coding, translation, customer service, data entry, basic legal research, content moderation — these are cognitive tasks, but they are cognitive tasks that can be specified in terms of patterns in text and that do not require the kind of embodied, contextual, relational judgment that characterises the most demanding professional work.

The question is whether AI will stop at these tasks or whether it will continue to advance into the domains of professional judgment that the historical argument assumes are safe. The evidence from GPT-4’s medical and legal benchmark performance suggests that the boundary between rule-governed cognitive tasks and judgment-requiring professional tasks is not as clear as the standard argument assumes.

---

The Productivity Paradox: Who Benefits from AI Productivity Gains?

The productivity improvements from AI are real and significant. Studies consistently show that workers who use AI assistance produce more output, faster, across a wide range of tasks. The question is not whether AI creates productivity gains but who captures those gains.

In a perfectly competitive labour market with full employment, productivity gains would flow to workers through higher wages. If workers can produce twice as much with AI assistance, employers competing for workers would bid up wages to capture the more productive workers. Workers would share in the productivity gains.

The actual distribution of productivity gains from technology has historically been less equal than this. The gains from mechanisation and computerisation have disproportionately accrued to capital owners — the companies that own the machines — and to highly skilled workers who complement the technology. Workers who compete with the technology — whose specific tasks are automated — tend to see wages fall or employment contract.

The AI case is more complex than previous automation waves because AI simultaneously complements some types of work and substitutes for other types of work. A lawyer who uses AI for research can handle more cases — AI complements the lawyer’s work and increases productivity. A paralegal whose primary job is legal research may find that AI substitutes for their work, reducing demand for paralegal services. The same technology — AI-assisted legal research — has different effects on different workers in the same industry.

This complementarity-substitution divide is one of the most important economic features of AI’s labour market impact. Workers whose skills complement AI — who use AI to amplify their judgment, creativity, and relationships — will tend to see productivity and wage gains. Workers whose skills substitute for AI — whose primary contribution is the kind of cognitive work that AI can now do — will tend to see wage pressure or employment reduction.

The division is not permanent — the specific boundary between complementarity and substitution will shift as AI capabilities change. But it is a real division, and it is not uniformly distributed across the workforce.

---

The Evidence So Far: What We Know and Don’t Know

The economic research on AI’s labour market impact is still in early stages — the technology has been widely deployed for too short a time to see the full effects in employment and wage data. But the research that has been done, and the patterns visible in the data, provide important evidence about what is happening.

The productivity gains are robust across different studies, different types of tasks, and different types of workers. This is the clearest and most consistent finding in the AI labour market research.

This finding has important implications for the debate about AI and inequality. If AI assistance disproportionately raises the productivity of less skilled workers, it could reduce within-job inequality. But this effect needs to be weighed against the substitution effects — if AI also reduces demand for lower-skilled workers overall, the productivity gains for those who remain employed may not offset the employment losses for those displaced.

The data that is available is mixed. Technology sector layoffs in 2022-2024 were large and widely attributed (at least in part) to AI-enabled productivity gains allowing companies to maintain output with fewer workers. Content creation industries — journalism, advertising copywriting, some types of creative work — have seen contraction in employment. Call centre and customer service employment has come under pressure in companies that have deployed AI customer service tools.

In other areas, employment has been more resilient. The professions that were initially most concerned about AI displacement — law, medicine, finance — have largely not seen significant employment contraction, in part because the deployment of AI in these professions has so far been primarily as assistance to professionals rather than substitution for them.

---

The Occupation-Level Analysis: Who Is Most Exposed

Economic research on AI’s labour market impact typically uses “exposure” measures — estimates of what fraction of each occupation’s tasks can be performed by current AI systems — to identify which occupations are most affected.

The economists Eloundou, Manning, Mishel, and Rock published a widely cited 2023 paper — “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” — that estimated exposure for all US occupations. The paper found that approximately 80% of US workers had some fraction of their tasks exposed to AI automation, and that approximately 19% of workers had 50% or more of their tasks exposed.

The occupations with the highest exposure tended to be white-collar, text-intensive roles: legal services, financial analysis, accounting, writing and editing, programming, data analysis, and customer service. The occupations with the lowest exposure tended to be either highly physical (construction, farming, plumbing) or highly relational and contextual (nursing, teaching, social work).

The high-exposure occupations are also, on average, higher-paying occupations. This creates a distributional dynamic that is different from previous automation waves: AI exposure is concentrated in higher-earning occupational categories, not in lower-earning ones. The workers most exposed to AI substitution are relatively more affluent — which may reduce the political urgency of the impact but does not eliminate the genuine economic disruption.

---

The New Work Question: What Jobs AI Will Create

The standard reassurance about AI’s labour market impact is that the same technology that destroys jobs also creates new ones. This has been true of previous automation waves. It may be true of AI. But what new jobs will AI create, for whom, and at what wages?

Several categories of new work are already visible.

Whether these categories represent significant employment at the scale of the displaced occupations is not yet clear. The emerging evidence suggests that prompt engineering is less specialised than initially hoped — most users of AI systems can learn to use them effectively without specialised training, which limits the employment in this category.

---

The Inequality Dimension: AI and the Distribution of Income

The labour market impact of AI will not be uniformly distributed across the income distribution, and understanding the distributional implications is essential for evaluating AI’s overall social impact.

The evidence on this concern is mixed. The finding that AI assistance disproportionately raises the productivity of lower-skilled workers suggests a potential equalising effect within employed populations. But this effect is offset by the substitution effects — if AI reduces demand for lower-skilled workers overall, the remaining workers in lower-skill occupations may become more productive without that productivity translating into higher wages, because the bargaining power of workers is reduced when their work can be automated.

The specific industries and occupations that are most exposed to AI automation — legal services, financial services, accounting, programming — are relatively high-wage industries. If AI primarily affects high-wage workers, the distributional impact may be less severe than if AI primarily affects low-wage workers. But the high-wage workers who are most exposed are also more likely to have the resources — education, savings, networks — to navigate the transition.

The most economically vulnerable workers are those who are both exposed to AI automation and have limited resources to absorb the impact: workers in content creation, customer service, and routine data work who lack the educational credentials and financial resources to retrain for new roles. These workers are in the middle of the wage distribution — not the highest earners, who can use AI as a complement to their work, and not the lowest earners, whose primarily physical and relational work is less immediately exposed.

---

The Policy Responses: What Governments Are Doing and What They Should Do

The labour market implications of AI are creating pressure for policy responses that address the productivity gains and the displacement costs. The policy conversation is still in early stages, but several approaches are being developed and debated.

---

The Professional Dimension: Medicine, Law, and the Judgment Premium

The most consequential question about AI’s labour market impact is not what it will do to content writers or customer service representatives — it is what it will do to the professions. To medicine, law, accounting, engineering, architecture, and the other high-skill, high-wage occupational categories that have been the primary vehicles of middle-class prosperity for the past century.

The professions have been protected from automation by the specific nature of professional work — the combination of technical knowledge and contextual judgment that characterises expert professional practice. A lawyer is not just a person who knows legal rules; a lawyer is a person who can apply legal rules to specific, ambiguous, contextual situations in ways that serve the client’s interests. A doctor is not just a person who knows medical facts; a doctor is a person who can apply medical knowledge to the specific, embodied, emotional situation of the patient.

AI’s performance on professional licensing examinations — GPT-4’s 90th-percentile bar exam score, its passing score on the USMLE — has been taken as evidence that AI is approaching professional competence. But the licensing examinations test knowledge and rule application; they do not test the contextual judgment, the relationship skills, and the wisdom that characterise excellent professional practice.

The more defensible argument about professional employment is not that AI will replace professionals but that it will change the structure of professional work — automating the knowledge-retrieval and rule-application components while leaving the judgment, relationship, and contextual components to human professionals. This would allow professionals to handle more complex matters, to serve more clients, to operate at higher levels of the professional hierarchy.

---

The Global Dimension: AI Automation Across Countries

The labour market impact of AI is not uniform across countries. The specific occupational structure, wage levels, and technological infrastructure of different countries create different patterns of exposure and different capacity to adapt.

The countries with the highest exposure to AI’s labour market impact tend to be the high-income countries with highly educated workforces and significant employment in the occupations most susceptible to AI automation — legal services, financial services, programming, content creation. The United States, the United Kingdom, and the European Union economies face significant AI exposure in their high-wage service sectors.

Lower-income countries with large manufacturing sectors, agricultural sectors, and service sectors oriented toward physical and relational work may be less immediately exposed to AI automation of the current generation’s capabilities. But they may also be less equipped to benefit from AI productivity gains, lacking the technological infrastructure, the AI talent, and the capital required to deploy AI at scale.

The global dimension also involves the specific question of who captures the productivity gains from AI. If AI systems developed in the United States or China are deployed globally, the productivity gains accrue to the companies that own the AI systems — primarily American and Chinese companies — while the employment displacement is distributed more broadly. The international distribution of AI’s economic impact is a significant governance challenge that current international institutions are not well equipped to address.

---

The Historical Parallel That May Not Apply

The standard reassurance about AI’s labour market impact relies on the historical parallel: every previous wave of automation created more jobs than it destroyed, and the AI wave will be the same.

The parallel deserves scrutiny, because several features of AI automation differ from previous automation waves in potentially important ways.

These differences do not prove that the historical reassurance is wrong. They are reasons to be less certain that it is right, and reasons to invest in the governance and policy frameworks that could ensure the transition is as equitable as possible — whether or not the historical parallel holds.

---

The Moral Dimension: Who Owes What to Whom

The labour market question is not just an economic question. It is a moral question: who owes what to whom in a society undergoing a technological transition?

The workers whose jobs are displaced by AI automation have, in many cases, spent years or decades developing the specific skills that AI is now automating. They made investments — in education, in training, in building careers — based on the reasonable expectation that those skills would remain valuable. If AI renders those investments worthless, the harm is real and the question of who bears that harm is not answered by saying that technology always creates more jobs than it destroys in the long run.

The companies that have benefited most directly from AI — the technology companies that have deployed AI to reduce their labour costs, the investors whose returns have come from AI-enabled productivity gains — have an argument for bearing a larger share of the transition costs. The productivity gains from AI represent a transfer of economic value from workers to capital owners; the moral case for taxation of those gains to fund the transition is straightforward.

The governments that have created the regulatory environment within which AI development and deployment have occurred — and that have chosen not to create the regulatory requirements that might have slowed deployment or required investment in worker protection — also bear responsibility for the transition costs. The choice to allow rapid AI deployment without strong labour market protections is a policy choice, not a natural law, and the costs of that choice should be distributed equitably.

The workers who are displaced are not the architects of the technological transition that is affecting them. They are its subjects. The moral case for ensuring that the transition is managed in ways that protect them from its worst costs — through strong social insurance, through meaningful retraining opportunities, through policy that ensures the productivity gains are broadly shared — is strong and has not been adequately answered by the policy responses that have been implemented so far.

---

The Fundamental Question: Is This Time Different?

The question that underlies the entire labour market debate about AI is whether this time is different — whether the pattern of job creation following job destruction that has characterised previous automation waves will hold in the AI era.

The honest answer is: we do not know. The historical evidence supports cautious optimism — the pattern has held for two centuries. The features of AI automation that distinguish it from previous automation waves give reasons to be uncertain that the pattern will hold. The transition will be managed better or worse depending on the policy choices made. The question is whether this time is different — and the answer is still being written.

The honest answer is: we do not know. The historical evidence supports cautious optimism — the pattern has held for two centuries. The features of AI automation that distinguish it from previous automation waves give reasons to be uncertain that the pattern will hold.

What is certain is that the transition will be managed better or worse depending on the policy choices made. The historical transitions that went best — in terms of maintaining living standards and distributing the benefits of productivity gains broadly — were the transitions where strong policy responses were in place: strong labour unions that bargained for workers’ share of productivity gains, strong social insurance that cushioned displacement, strong public investment in education and infrastructure that created new economic opportunities.

The historical transitions that went worst were the ones where policy failed to keep up with technology: where workers bore the costs of displacement without support, where the productivity gains accrued entirely to capital, where communities built around specific industries were left to decline without investment in alternatives.

The AI transition is still early. The policy choices that will determine whether it is managed well or badly are still being made. The workers who will bear the costs of displacement — or benefit from the productivity gains — are still, in many cases, in careers that have not yet been significantly affected by AI.

The window for making good policy choices is open but not indefinitely. The pace of AI adoption, and the pace of the capability improvements driving it, suggests that the window is narrow. The question of what happens when AI does the work is the most important economic question of the current era, and the answer is still being written.

---

Further Reading

---

---