AI job displacement: who is exposed, and what actually helps

Estimates of US jobs at high risk from automation range from about 9 percent to 47 percent. Frey and Osborne (2017) put it at 47 percent, an OECD study (Arntz, Gregory and Zierahn, 2016) at 9 percent, and PwC (2017) at 38 percent by the early 2030s. The spread comes largely from method: whole occupations versus individual tasks.

EasWrk is a hosting platform, not a research institute. This essay summarises published sources, which are listed at the end. For a related case, see autonomous vehicles and the driving workforce; for the small-business side, see small business owners and AI; the full set is on the research hub.

AI Replacing Jobs: The Statistics and Why They Differ

Few questions in economics generate more heat and less light than the question of what automation will do to employment. On one side, technology optimists point to centuries of economic history in which technological change created more jobs than it destroyed, raised living standards, and expanded the variety of human activity that could be called productive work. On the other side, a growing body of economic research suggests that the current wave of automation, driven by machine learning and generative AI, is different in kind from previous waves, and that the standard reassurances may not hold.

This paper does not attempt to resolve that debate definitively. The evidence does not yet support definitive resolution. What it does attempt is something more practical: to identify which workers face the greatest exposure to automation-driven displacement, to examine honestly why the policy responses to past waves of automation have failed the workers who needed them most, and to sketch what a more adequate response might look like. The stakes of getting this wrong are not abstract. They are measured in family incomes, community stability, health outcomes, and political legitimacy.

AI replacing jobs: the research papers behind the numbers

The best-known study is Frey and Osborne's 2013 analysis, published in the journal Technological Forecasting and Social Change in 2017. It estimated that 47 percent of U.S. employment was at high risk of automation. The figure has been widely cited and widely contested. Arntz, Gregory and Zierahn (2016) analysed the tasks within each occupation instead of treating whole occupations as automatable, and found that about 9 percent of jobs in the US and across 21 OECD countries were at high risk. PwC (2017) used its own method and estimated that about 38 percent of US jobs were at potential high risk by the early 2030s, compared with about 30 percent in the UK. The variation comes mainly from whether job-level or task-level analysis is used, and from assumptions about how quickly capability becomes deployment.

The disagreement about magnitude conceals significant agreement about direction. Virtually all serious analyses find that automation is already displacing workers from certain types of tasks, that this displacement is unevenly distributed across occupations, skill levels, industries, and geographies, and that the transition costs are being borne primarily by the workers displaced rather than by the firms or consumers who benefit from the productivity gains. The question is not whether displacement is happening but who bears its costs and whether any serious attempt is being made to address that distribution.

Which Jobs Are Most Exposed to AI

The Autor, Levy, and Murnane framework, developed in their landmark 2003 paper in the Quarterly Journal of Economics, provides the most durable theoretical foundation for understanding which jobs automation is likely to affect. They distinguished between routine cognitive tasks, routine manual tasks, non-routine cognitive tasks, and non-routine manual tasks. Automation and computerization have historically been most effective at replacing routine tasks, whether cognitive (bookkeeping, data entry, certain forms of analysis) or manual (repetitive assembly operations, sorting, material handling in structured environments). Non-routine tasks, those that require judgment, adaptation to changing circumstances, or physical dexterity in unstructured environments, have been less susceptible.

This framework, sometimes called the task-based model of technological change, generates predictions that map onto observed employment trends. Middle-skill, routine-intensive jobs in manufacturing, administrative support, and back-office financial services have declined significantly. Employment has polarized, growing at the top (high-skill, non-routine cognitive work) and at the bottom (low-wage, non-routine manual service work) while hollowing out in the middle.

Generative AI complicates this picture in important ways. The task-based framework was developed to describe the effects of computing technology that is good at rule-based processing but poor at pattern recognition in unstructured contexts. Large language models and multimodal AI systems are increasingly capable in domains that were previously considered safe from automation: complex analysis, creative work, coding, certain forms of legal and medical judgment, and customer interaction. This has extended the automation frontier significantly upward in the occupational hierarchy.

AI replacing jobs: examples by occupation

The ratings below are this essay's qualitative judgement, informed by the sources listed at the end. They are not measured figures, and the timelines are rough guesses.

Acemoglu and Restrepo (2020) studied the rise in industrial robot use between 1990 and 2007 across US local labor markets. They estimated that one more robot per thousand workers lowered the employment-to-population ratio by 0.2 percentage points and wages by 0.42 percent in the affected area. A later paper by the same authors (2022) found that between 50 and 70 percent of changes in the US wage structure over four decades are accounted for by relative wage declines among groups specialised in routine tasks in industries that were automating quickly. These findings challenge the standard technology-optimist story, which holds that productivity gains from automation create enough new demand to employ the displaced workers. The mechanism may work in aggregate over long periods, but it can fail in specific local labor markets, in specific time horizons, and for workers who lack the mobility or skills to reach the new jobs that automation creates elsewhere.

Brynjolfsson and McAfee (2014) argued that digital technology can raise productivity while leaving many workers behind. Their point is that the gains from automation are real, and that how they are shared is a separate question.

The Automation Paradox: Why Technology Doesn't Always Replace

The relationship between automation and employment is more complicated than either the optimists or pessimists typically acknowledge. Several mechanisms operate simultaneously and can produce outcomes that appear paradoxical when viewed in isolation.

The most important of these is what Acemoglu and Restrepo (2019) called the task reinstatement effect. When technology automates some tasks within a job, it often increases the value of the remaining non-automated tasks, and can create demand for new tasks that did not previously exist. The automated loom displaced hand-weavers but created demand for loom operators, textile engineers, and distribution workers. The spreadsheet is often cited as having cut routine bookkeeping while raising demand for analysts. These are commonly used examples, and the mechanism is one of the main reasons technological change has not so far produced permanent mass unemployment.

But the task reinstatement effect is not guaranteed, and its effectiveness depends on the speed and breadth of the automation wave, the availability of complementary skills in the existing workforce, and the institutional mechanisms available to help workers transition to new roles. When automation moves faster than the educational and training system can adapt, when the new tasks require skills that displaced workers lack, and when the labor market mechanisms for facilitating transition are inadequate, the reinstatement effect can be overwhelmed by the displacement effect.

There is also what economists sometimes call the demand effect: productivity gains from automation reduce costs, which reduce prices, which increase real incomes, which increase demand for goods and services, which increases employment. Again, this mechanism is real in aggregate. But it operates at the economy-wide level and on a long time horizon. Take a hypothetical worker displaced from a call center in Ohio by a conversational AI system. She does not benefit from the fact that consumers, on average, will spend their savings from reduced service costs on other goods and services. The aggregated economic expansion does not find her at her particular address with her particular skills in her particular labor market.

Standing (2011) introduced the concept of the precariat, a class of workers characterized by chronic insecurity of employment, income, and identity. The precariat is not new, but the forces that create it are intensifying. The combination of automation-driven displacement, the rise of gig and platform work as a substitute for stable employment, and the erosion of the institutional protections that previously provided workers with some buffer against economic volatility is producing a growing stratum of the workforce for whom the standard mechanisms of economic adjustment, find new work, retrain if necessary, rely on social insurance in the interim, are not functioning as advertised.

Why Government Retraining Programs Have Largely Failed

The standard policy response to technological displacement, across administrations and across countries, has been to invest in retraining programs. The logic is intuitive: workers displaced from declining industries or occupations need new skills for growing ones; government can facilitate that acquisition through funded training programs; workers emerge with marketable skills and find new employment. The theory is clean. The practice has been, by most honest assessments, deeply disappointing.

The U.S. Trade Adjustment Assistance (TAA) program, which provides retraining support to workers displaced by trade competition, is among the most studied examples. A Mathematica evaluation of the program under the 2002 amendments (2012, published by the US Department of Labor in 2013) compared participants with matched comparison workers. Participants fared significantly worse in the first two years after filing for unemployment insurance, and the gap narrowed in years three and four. By year four, participants' earnings were either about $3,300 lower than the comparison group's or about the same, depending on the estimate. The evaluation also found that the extra income payments did not fully make up for participants' lower earnings. One plausible reason for the early gap is that time in training is time out of work.

The OECD's Employment Outlook 2019 looked at adult learning across member countries. It found that participation is lowest among the groups that most need training, including low-skilled adults, older workers and people who have lost jobs, and that training only works if it is high quality and matched to labor market needs. Those findings fit the pattern in the TAA evidence: training that is disconnected from real demand for skills is unlikely to lead to jobs.

One explanation for the mismatch between what displaced workers need and what many publicly funded programs provide is cost. Employer-connected, targeted training with support services is expensive, while generic classroom training is cheap. Under budget pressure, cheap programs may win out over expensive ones. This is the essay's interpretation, not a finding from the studies above.

There may also be a mismatch between the workers most likely to be displaced and the workers retraining suits best. Workers in routine-intensive jobs often have long tenure in one industry and lower levels of formal education, and they may live where growing industries are absent. Consider a hypothetical worker who is 55, has spent 30 years on a factory floor, has a high school diploma, and lives in a rural area is not well positioned to retrain as a data analyst, a nurse practitioner, or a software developer, however good the program. The credential she could realistically earn may not match a job within commutable distance of her home.

McKinsey Global Institute (2017) estimated that between 75 and 375 million workers globally, 3 to 14 percent of the global workforce, may need to switch occupational categories by 2030, depending on how fast automation is adopted. Even the low end would be a very large transition. The sources reviewed for this essay contain no example of retraining at that scale and speed having been done successfully.

The Geography of Displacement

Automation's effects on employment are not evenly distributed across geographic space. The concentration of specific industries in specific regions means that the displacement of workers in those industries falls heavily on particular communities, creating local economic shocks that can be severe and long-lasting. The communities most exposed to automation-driven displacement are disproportionately rural, smaller metropolitan, and non-coastal areas with high concentrations of manufacturing, logistics, administrative processing, or agricultural employment.

Acemoglu and Restrepo (2020) found that commuting zones with higher exposure to industrial robots saw larger falls in employment and wages than less-exposed zones. The mechanism is geographic concentration: when a major employer in a small city or rural area automates, the workers it displaces do not easily move to other markets where the new jobs created by automation productivity are located. They remain, and the local economy contracts.

This geographic dimension of displacement interacts badly with the standard economic recommendation to affected workers: move to where the jobs are. Internal migration in the United States had fallen to about a 30-year low by 2011, across demographic groups and for moves of all distances (Molloy, Smith and Wozniak, 2011). The authors found that the housing slump and recession explained little of it and said the puzzle remained. Plausible barriers include high housing costs in productive metropolitan areas, family and social networks, the difficulty of selling a house in a depressed local market, and fewer good jobs for non-college workers even in growing cities.

The communities left behind by automation-driven displacement face a compounding dynamic that is difficult to interrupt from the outside. When workers leave, local tax revenue declines, reducing the quality of public services including schools. The declining quality of schools makes the community less attractive for new businesses and reduces the prospects of the children who remain. Economic decline in a community can be accompanied by worse health and weaker social ties. Researchers also link job loss to worse health outcomes at the individual level (Sullivan and von Wachter, 2009).

Any serious policy response to labor displacement must grapple with the geographic dimension. A response centered on individual retraining and labor mobility addresses the situation of a subset of workers, those who are mobile, younger, and sufficiently educationally prepared to retrain. It does not address the communities that are left behind when that subset leaves, nor does it help the workers who cannot or will not relocate. Place-based economic policy, which attempts to bring economic activity to distressed communities rather than move workers to activity elsewhere, has had mixed results historically but has received renewed interest precisely because the limitations of the individual-mobility-and-retraining framework have become more apparent.

Universal Basic Income After AI: Evidence and Limitations

Universal Basic Income has become the most discussed policy response to the threat of large-scale automation-driven displacement. The core idea is simple: provide every adult with a regular unconditional cash payment sufficient to cover basic needs, regardless of employment status, and allow individuals to make their own decisions about how to supplement it. UBI advocates argue that this approach is simpler to administer than the existing patchwork of means-tested programs, removes the poverty traps created by programs that phase out as income rises, and provides genuine economic security that enables workers to take risks, pursue education, and care for family members without the precarity that currently characterizes low-wage employment.

Stanford's Basic Income Lab runs a dashboard that tracks more than 30 guaranteed income pilots across the United States. The evidence from pilots is informative, though its applicability to a full-scale UBI program is contested. The Stockton Economic Empowerment Demonstration (SEED) gave 125 residents of low-income neighborhoods $500 per month for 24 months from February 2019. In the first-year results, released in March 2021, full-time employment among recipients rose from 28 percent to 40 percent, against 32 to 37 percent for a control group, and recipients also reported better financial and physical health.

The Finnish basic income experiment gave 2,000 unemployed people aged 25 to 58 a payment of 560 euros per month in 2017 and 2018, with recipients drawn at random from people receiving unemployment benefit from Kela. The evaluation by the VATT institute found that the payment had no group-level employment effect in the first year and only small effects overall, while recipients reported better wellbeing. VATT noted that a change to unemployment rules in 2018 complicates reading the second year.

Critics of UBI raise several objections that deserve serious engagement rather than dismissal. The fiscal cost of a universal payment sufficient to cover basic needs at the national scale is very large. In the United States, a payment of $1,000 per month to each of the roughly 262 million adults (Census Bureau, 2023 estimate) would cost about $3.1 trillion a year. Federal outlays were about $6.8 trillion in fiscal 2024, so the payment would equal close to half of all federal spending (author's arithmetic from the Census and fiscal figures). Funding this through progressive taxation is arithmetically possible but requires political and institutional changes that have proven extremely difficult to achieve. Funding it by consolidating existing programs risks eliminating supports that are better targeted to people with specific needs than a universal payment would be.

There is also a serious question about whether unconditional income, by itself, addresses the non-economic dimensions of job loss. Work provides not only income but social connection, structure, identity, and a sense of purpose and contribution. Research on involuntary job loss finds lasting harm that goes beyond lost income; Sullivan and von Wachter (2009), for example, found sharply higher mortality among high-seniority male workers after job displacement. Replacing income without replacing the social and psychological functions of work may be less effective than its proponents assume, particularly for workers who have defined themselves through their labor.

Minsky (1986), writing about economic instability rather than automation, argued for a job guarantee as a mechanism for stabilizing the economy by ensuring that any worker who wanted a job could find one at a livable wage, with the government serving as employer of last resort. Contemporary advocates have revived this proposal as a response to automation, arguing that it directly addresses the income, social, and psychological functions of work in a way that UBI does not, while also producing socially useful output. Critics point to the administrative complexity of running a large-scale public employment program, the risk that public employment crowds out private employment, and the historical mixed record of large public works programs.

Building a Safety Net for the Automated Economy

No single policy instrument is adequate to the challenge posed by large-scale, uneven, geographically concentrated labor displacement driven by AI and automation. What an honest safety net for the automated economy requires is a portfolio of policies that address the multiple dimensions of the problem simultaneously, funded adequately, and designed on the basis of what has actually worked rather than what is politically convenient.

The first element of that portfolio is early warning and transition support. Workers and communities need more lead time than they currently receive before displacement occurs. Mandatory notice requirements when firms automate at significant scale, combined with publicly funded transition counseling and planning support, could help workers begin the adaptation process before job loss rather than after. The case for early support is that help before unemployment is likely to work better than help after a long spell out of work. This is a judgement, not a measured result.

The second element is a serious overhaul of publicly funded training. The OECD finds that training works when it is high quality and matched to labor market needs. Training that is disconnected from employers is unlikely to help. The political obstacles to funding effective training at scale are not primarily technical. They are about who pays and who benefits. A levy on firms that automate, used to fund training for workers displaced by that automation, would align incentives in a way that the current system does not.

The third element is income support that does not require workers to become impoverished before they can access help. In 2023 only about 29 percent of unemployed workers in the United States received unemployment insurance, with state rates ranging from 10 to 55 percent (Minneapolis Fed, 2025). Expanding coverage, extending duration, and raising replacement rates for workers in high-displacement industries would reduce the immediate crisis of job loss and create more time for transition. Raising coverage is not a novel idea.

The fourth element addresses the geographic concentration of displacement. Place-based investment in community infrastructure, broadband access, local healthcare, and the local institutions that anchor community life is not a substitute for individual economic support but is a necessary complement to it. Communities that have been depleted by automation-driven employment decline have weaker institutions for supporting the workers who remain. Rebuilding those institutions is a precondition for effective transition at the community level.

The fifth element, more politically contested than the others, is a serious conversation about who captures the gains from automation. The productivity gains from AI and automation are real. They appear to accrue mainly to shareholders, to highly skilled workers whom AI complements, and to consumers through lower prices, though the split is hard to measure. Displaced workers receive little of the gain from the productivity improvements that cost them their jobs. Various mechanisms for redistributing a portion of those gains, including automation taxes, expanded capital ownership mechanisms, profit-sharing requirements, or enhanced capital income taxation, have been proposed. None has achieved significant political traction. The question of redistribution is ultimately a political question about whose interests the economic system is organized to serve. Avoiding it does not make it go away.

Standing (2011) argued that the precariat, and the insecurity it represents, is not an accident of market forces but a political construction: the result of deliberate choices about how labor markets are regulated, how social insurance is structured, and whose interests economic policy is designed to protect. The automated economy will produce more or less insecurity, and distribute its gains more or less equitably, depending on choices that are currently being made, largely by default, in legislatures, boardrooms, and regulatory agencies. Those choices will compound over the coming decade in ways that are very difficult to reverse. The cost of getting them wrong will be borne, as always, most heavily by those who were already least secure.

The scale of the challenge is not an argument for fatalism. It is an argument for urgency. The history of economic transitions driven by technological change contains both catastrophic failures, the industrial revolution's first century included levels of child labor, urban squalor, and working-class immiseration that are often sanitized in retrospect, and genuine successes where deliberate policy choices produced more equitable outcomes. The difference between those outcomes was not the technology. It was the political will to insist that the gains from technological progress be shared rather than concentrated, and the institutional capacity to act on that insistence. Building those institutions for the AI era is one of the main economic policy challenges of the next decade.