The Numbers Behind the Headlines
What Confirmed Displacement Actually Looks Like in 2025
The gap between projected and confirmed AI job losses is vast — but structural signals suggest the reckoning is building, not arriving.
When ChatGPT launched in November 2022, the initial public discourse veered rapidly between two poles: techno-utopian promises of infinite productivity, and apocalyptic warnings of mass unemployment within years. Three years later, the empirical record is more nuanced — and in important ways, more alarming — than either camp predicted.
The confirmed, directly attributable AI job displacement figures for 2025 are, by macro-economic standards, modest. According to outplacement firm Challenger, Gray & Christmas, approximately 12,700 U.S. jobs were directly attributed to AI in 2024, rising to roughly 55,000 in 2025. ✓ Established [8] Even incorporating broader independent estimates — which attempt to count AI-driven displacement not explicitly flagged in corporate announcements — the total U.S. AI-attributable displacement in 2025 runs to between 200,000 and 300,000 positions, representing 0.13 to 0.20 percent of the nonfarm workforce. ◈ Strong Evidence [8]
The Yale Budget Lab, reviewing the labor market through mid-2025, found no discernible broad disruption 33 months after ChatGPT's launch. ◈ Strong Evidence [5] The Budget Lab's analysis is consistent with the historical record: widespread technological disruption characteristically unfolds over decades, not months, as diffusion lags and institutional inertia slow adoption curves well below what laboratory demonstrations suggest.
Goldman Sachs Research, publishing in August 2025, arrived at a similarly measured conclusion: under current AI use cases expanded economy-wide, only 2.5 percent of U.S. employment is directly at risk of displacement. Even under a wide-adoption scenario, that figure rises to just 6 to 7 percent of the U.S. workforce. Goldman Sachs further estimated that any unemployment spike during the AI transition is likely transitory, dissipating within approximately two years. ⚖ Contested [6]
These macro-level readings, however, contain a critical blind spot: they measure averages across an economy of 160 million workers, obscuring sector-level and demographic-level concentrations of harm. The analogy is instructive — a city's average temperature tells you little about whether specific neighbourhoods are flooding. The aggregate numbers are calm; in specific occupational zones, the waters are already rising.
The Hollowing Middle
How AI Targets the Cognitive Middle Class First
AI is deepening a polarisation pattern already two decades old — but for the first time, it is reaching up into white-collar cognitive work rather than down into manual labour.
To understand AI's labour market impact, the correct baseline is not the industrial revolution or the computerisation of the 1980s — it is the period from 2000 to 2020, during which U.S. employment growth concentrated heavily in two segments: high-wage professional roles and low-wage service positions, while middle-skill clerical, administrative, and production jobs contracted sharply. ◈ Strong Evidence [10] This pattern — economists call it labour market hollowing, or job polarisation — was driven by the automation of routine cognitive tasks: bookkeeping, data entry, telephone switchboard operation, and basic paralegal research.
AI does not reverse this hollowing; it accelerates and extends it. What distinguishes generative AI from prior automation waves is precisely its reach into cognitive, language-based, and even creative work that prior software could not touch. The IMF's January 2024 analysis found that approximately 40 percent of global employment is exposed to AI, rising to roughly 60 percent in advanced economies — a scale of exposure without historical precedent, because earlier automation primarily affected routine manual and low-skill cognitive tasks, not the professional and managerial core. ✓ Established [2]
MIT economist David Autor, writing in a 2024 National Bureau of Economic Research working paper, placed this dynamic in its structural context: the U.S. labor market's middle-skill, middle-class core has been progressively hollowed out by automation and globalisation over two decades, and whether AI proves destructive or constructive in the next phase is not a technological inevitability but a policy choice. ◈ Strong Evidence [11] Seventy-five percent of U.S. adults, according to Gallup polling cited in Autor's paper, believe AI will lead to fewer jobs — a figure that reflects widespread intuition about this structural vulnerability even if the macro data has yet to confirm it in aggregate form.
An IMF Working Paper published in September 2024, examining U.S. commuting zone data from 2010 to 2021, found that areas with higher AI adoption experienced a stronger decline in the employment-to-population ratio over that period. Crucially, the negative employment effect was borne primarily by manufacturing, low-skill services, middle-skill workers, and non-STEM occupations — not by the high-income professional tier that attracts most media coverage. ◈ Strong Evidence [9] The regional granularity of this finding matters: AI's labour market effects are not distributed uniformly across geography, and communities whose economic base is concentrated in routine cognitive service work face qualitatively different risk than metropolitan areas anchored by high-skill professional employment.
The Human Capital Leadership Review, synthesising this evidence in December 2025, identified what it termed a bifurcation threat: a division between AI-augmented super-workers — professionals who deploy AI to multiply their output and command premium wages — and a marginalised middle tier whose tasks are automated away without access to the new augmentation premium. ◈ Strong Evidence [10]
Jobs at Risk — The Evidence-Based Taxonomy
From Interpreters to Junior Analysts
The empirical evidence identifies a clear hierarchy of AI exposure — and the occupations at the top are not the ones most people expect.
Microsoft Research's July 2025 analysis, drawing on 200,000 anonymised Microsoft Copilot conversations recorded between January and September 2024, constructed perhaps the most granular occupational exposure map yet produced. Interpreters and translators top the list, with 98 percent of their work activities overlapping with AI capabilities — an extraordinary figure that reflects the near-complete digitisation of language mediation. Historians, writers, and sales representatives also appear among the most highly exposed roles. In total, Microsoft Research identified 8.4 million U.S. workers concentrated in the 40 most AI-impacted occupational categories. ◈ Strong Evidence [7]
The St. Louis Federal Reserve, publishing in August 2025 an analysis of unemployment trends by occupational AI-exposure score from 2022 to 2025, found a statistically significant correlation: occupations with higher AI exposure showed steeper unemployment rises over that period, with a correlation coefficient of 0.57. Computer and mathematical occupations — carrying an AI exposure score of approximately 80 percent — registered some of the steepest unemployment gains of any occupational category. ◈ Strong Evidence [3]
Erik Brynjolfsson, Michael Chan, and Michael Chen of the Stanford Digital Economy Lab published findings in August 2025 documenting substantial employment declines for early-career workers in AI-exposed occupations, specifically software development and customer support. Economy-wide employment continues to expand, but growth for young workers significantly lags behind that of older cohorts. This entry-level contraction is analytically distinct from aggregate displacement: it reflects AI's replacement of the training-wheel tier of knowledge work — the junior analyst, the junior developer, the first-line customer support agent — rather than wholesale elimination of entire professional fields. [4]
The customer service sector offers one of the clearest empirical windows into what confirmed, measurable displacement looks like. The Site Selection Group documented a decline of approximately 80,000 customer service positions in the U.S. between 2022 and 2024 — a contraction concentrated in the period of most rapid AI customer-support deployment. ◈ Strong Evidence [8] This is not projection or modelling; it is observed headcount decline in a specific sector during a specific window of AI adoption.
| Occupation / Category | AI Exposure Level | Evidence Assessment |
|---|---|---|
| Interpreters & Translators | Microsoft Research: 98% work-activity overlap with AI capabilities (Copilot dataset, Jan–Sep 2024) | |
| Computer & Mathematical Occupations | St. Louis Fed: steepest unemployment rises 2022–2025 of any major occupational group | |
| Writers & Authors | Microsoft Research top-40 most-impacted roles; generative AI directly substitutable | |
| Customer Support / Service Reps | ~80,000 U.S. positions eliminated 2022–2024 (Site Selection Group); Stanford early-career declines confirmed | |
| Clerical & Administrative | Brookings/GovAI: 6.1M workers with high exposure and low adaptive capacity concentrated here | |
| Junior Financial Analysts | IMF: high-skill cognitive work uniquely exposed; entry-level tasks most automatable |
Jobs That Are Safe (and Why)
Skilled Trades, Healthcare, and the Primacy of Embodied Labour
The safest jobs share a common property: they require physical presence, dexterous embodied action, or high-stakes interpersonal trust that AI cannot replicate or deploy remotely.
Identifying which jobs are protected from AI displacement requires understanding what AI cannot do — not what it struggles with today, but what is structurally beyond a system that processes and generates language, code, and images. The answer converges on two categories: embodied physical work requiring dexterity and situational judgment in unstructured environments, and high-stakes human relationships where the value of the interaction is inseparable from the human delivering it.
Skilled trades — electricians, plumbers, HVAC technicians, construction workers — occupy the first category. Their work requires physical manipulation of complex, variable environments, real-time problem-solving in situ, and liability-bearing judgment that cannot be delegated to a language model operating remotely. These occupations carry low AI exposure scores and strong projected growth trajectories. Bureau of Labor Statistics projections, cited in the ALM Corp synthesis, show nurse practitioners growing by 52 percent between 2023 and 2033 — driven by demographic demand from an ageing population that no AI deployment can redirect. ◈ Strong Evidence [8]
Personal care and food service occupations occupy a different but equally protected position. Bureau of Labor Statistics projections indicate food preparation and service jobs are expected to add more than 500,000 positions by 2033. ◈ Strong Evidence [8] These roles require physical presence, emotional attunement, and are priced in a labour market where robotic substitution remains economically uncompetitive at current wage levels.
The distinction between exposure and displacement matters here. The IMF's January 2024 analysis found that in advanced economies, roughly 60 percent of jobs may be impacted by AI — but impact is not synonymous with elimination. ✓ Established [2] Roughly half the exposed jobs in advanced economies are assessed as likely to benefit from AI as a productivity tool — meaning the worker retains their role but with AI augmenting their output. The other half face wage depression or elimination. The critical variable determining which outcome materialises is adaptive capacity — and it is the distribution of adaptive capacity that constitutes the most important and least-reported dimension of AI's labour market impact.
The Hidden Victims: 6.1 Million
The Adaptive Capacity Gap Nobody Is Talking About
High AI exposure plus low ability to adapt creates a distinct and largely invisible vulnerable cohort — overwhelmingly female, overwhelmingly clerical, and almost entirely absent from public discourse about AI's costs.
The most consequential finding in the recent empirical literature on AI and employment is not about aggregate displacement rates or sectoral unemployment trends. It is a distributional insight published by the Brookings Institution on February 12, 2026, in collaboration with GovAI, measuring not just who faces AI exposure but who faces AI exposure combined with low capacity to adapt.
The study begins with a foundational observation: of the 37.1 million U.S. workers in the highest AI-exposed jobs, 26.5 million have above-median adaptive capacity — meaning they possess educational qualifications, occupational transferability, financial resources, or geographic mobility that give them meaningful options for transition. These workers dominate the headlines. They are the Silicon Valley engineers and the financial analysts whose sophisticated jobs AI is supposedly targeting. ✓ Established [1]
But the study's most important finding concerns the remaining cohort: approximately 6.1 million U.S. workers who face both high AI exposure and low adaptive capacity. These workers — concentrated overwhelmingly in clerical, administrative, and routine cognitive service roles — lack the credentials, financial buffers, and occupational flexibility to navigate a forced transition. They are not career-changing. They are trapped. ✓ Established [1]
The adaptive capacity gap is the central analytical frame through which AI's labour market harm must be assessed. The conventional framing — white-collar professional workers face AI exposure, therefore white-collar professionals are AI's primary victims — commits a logical error: it conflates exposure with harm. Exposure without adaptive capacity is categorically more damaging than exposure with it. A software engineer at a San Francisco firm who loses their role to AI automation has, in most cases, qualifications, savings, professional networks, and geographic mobility to navigate the transition. The administrative assistant at a regional insurance company in Ohio, performing the same data-entry and scheduling functions that AI can now handle, has none of those resources.
This distinction is further sharpened by the credentials barrier in new AI-era employment. According to SSRN research cited in the ALM Corp synthesis, approximately 77 percent of newly created AI-related roles require master's degrees. ⚖ Contested [8] Even if the World Economic Forum's net positive projection — 92 million jobs displaced, 170 million created, a net gain of 78 million by 2030 — proves accurate, the structural credential mismatch means that displaced clerical workers cannot simply walk into newly created AI-operations or prompt-engineering roles. The jobs being eliminated require a secondary school diploma. The jobs being created require a graduate degree. The arithmetic of net creation is, for this cohort, entirely beside the point.
The workers most likely to be harmed by AI displacement are not the ones generating the most alarm. The engineers and analysts facing high exposure have the resources to adapt. The 6.1 million with both high exposure and low adaptive capacity are, in structural terms, the most vulnerable — and they are almost entirely absent from the policy conversation.
— Brookings Institution / GovAI, February 12, 2026The Gender Fault Line
Why 86% of the Most Vulnerable Workers Are Women
AI's adaptive capacity gap has a stark gender dimension that labour market analysis has been slow to foreground — and that policy frameworks have almost entirely failed to address.
The Brookings/GovAI finding that approximately 86 percent of the 6.1 million low-adaptability, high-exposure workers are women is not a peripheral data point. It reflects deep structural patterns in how gendered occupational segregation has distributed both risk and resource within the U.S. labour market. ✓ Established [1]
Clerical and administrative work — data entry, scheduling coordination, basic record-keeping, customer correspondence, office administration — has been disproportionately performed by women for much of the past century. This occupational concentration was not technologically inevitable; it was the product of a labour market in which women were channelled into roles classified as supportive, auxiliary, and routine rather than strategic or professional. The irony is now vivid: the same classification system that historically undervalued these roles is now rendering them the primary target of AI substitution.
The gender dimension of AI displacement is compounded by the adaptive capacity metrics themselves. The Brookings/GovAI study's adaptive capacity index incorporates factors including educational attainment, occupational transferability, access to financial safety nets, and geographic mobility. Women in clerical roles are systematically disadvantaged on multiple dimensions of this index: they are more likely to be geographically immobile (due to caregiving responsibilities), more likely to lack STEM credentials required for technical transition roles, and more likely to be in part-time or irregular employment arrangements that provide fewer protections during displacement.
The IMF's broader analysis adds a global dimension to this finding. In advanced economies, the IMF found that AI uniquely impacts high-skilled jobs — a pattern distinct from prior automation waves. But within that broad exposure landscape, the distribution of adaptive capacity follows gender lines that reflect pre-existing labour market inequalities rather than technological neutrality. ✓ Established [2]
Labour force participation data reinforces the structural vulnerability. Bureau of Labor Statistics figures cited in the ALM Corp synthesis show that prime-age workers without a college degree saw their labour force participation decline from 88.6 percent in 1990 to 82.3 percent in recent years — a long-run erosion that AI is poised to accelerate for the specific cohort of women in routine cognitive service work. ◈ Strong Evidence [8]
The Institutional Variable
Why AI Displaces More in the U.S. Than in Nordic Countries
Displacement outcomes are not technologically determined — they are institutionally mediated, and the gap between Nordic and Anglo-American labour markets demonstrates this clearly.
One of the most policy-relevant findings in the recent literature is that AI's labour market impact is not a fixed technological output but a variable one, shaped substantially by the institutional environment in which AI is adopted. The AB Academics analysis, published in January 2026, found that in Nordic countries with strong union representation and active labour market policies, AI adoption has been associated with narrower wage gaps. In the United States and United Kingdom, by contrast, AI adoption has coincided with sharper wage polarisation. ◈ Strong Evidence [12]
This divergence is not incidental. Nordic labour markets are characterised by centralised wage bargaining, comprehensive active labour market programmes (retraining subsidies, job-search support, wage insurance), strong union involvement in corporate technology adoption decisions, and social insurance systems that substantially cushion displacement shocks. When a Danish firm automates administrative functions, the displaced worker enters a well-funded transition system with high income replacement rates, state-subsidised retraining options, and legal obligations on the employer to consult worker representatives before restructuring.
Nordic Labour Market Model
U.S. / U.K. Labour Market Model
The implication is direct and politically uncomfortable: the degree of harm AI inflicts on the 6.1 million low-adaptability workers identified by Brookings/GovAI is not a function of how powerful the technology is. It is a function of whether those workers have access to institutional protections and transition resources. In the U.S. context, they largely do not. The American workforce transition infrastructure — community college retraining programmes, Trade Adjustment Assistance mechanisms, unemployment insurance — was designed for a manufacturing-era displacement pattern characterised by geographically concentrated factory closures, not the diffuse, sector-spanning displacement of routine cognitive work.
MIT's David Autor made this point directly in his 2024 NBER working paper: AI's application as destructive or constructive is a policy choice, not a technological inevitability. ◈ Strong Evidence [11] The OECD's aggregate data — wage inequality rising by approximately 10 percent across member countries between 2000 and 2022, driven by routine cognitive automation — represents an average that obscures the wide institutional variance between countries. ◈ Strong Evidence [12]
Hype vs. Harm
Where Goldman Sachs, Yale Budget Lab, and Stanford Disagree
The empirical debate between measured sceptics and early-alarm researchers is not a simple optimist-pessimist divide — it is a methodological dispute about what counts as evidence and over what time horizon.
The evidentiary landscape for AI's labour market impact is genuinely contested, and intellectual honesty requires presenting the disagreements as substantively as the areas of consensus. Three distinct research positions are represented in the current literature, and they are not reconcilable by splitting the difference.
The first position, associated with the Yale Budget Lab's 2025 analysis, holds that the 33 months since ChatGPT's launch have produced no discernible broad disruption in the U.S. labour market, and that historical precedent consistently shows widespread technological disruption unfolding over decades rather than years. ◈ Strong Evidence [5] This position is methodologically conservative and historically grounded: it correctly notes that predictions of rapid mass technological unemployment have been made and falsified repeatedly over the past century.
Goldman Sachs Research's August 2025 analysis reinforces this measured view with sector-level modelling: only 2.5 percent of U.S. employment at direct displacement risk under current conditions, and any transition-period unemployment spike likely transitory within two years. ⚖ Contested [6]
The Measured Sceptic Position
The Early-Alarm Position
The third position, represented by Brynjolfsson, Chan, and Chen at the Stanford Digital Economy Lab in their August 2025 paper — titled, revealingly, Canaries in the Coal Mine? — holds that while aggregate employment remains healthy, occupation-level and career-stage-level data shows clear early signals of concentrated harm. ◈ Strong Evidence [4] The canaries metaphor is chosen with care: early-career workers in AI-exposed occupations are the canaries — the first to experience the harm that will, if unchecked, propagate upward through the occupational hierarchy.
The methodological resolution to this dispute is not purely empirical — it is partly a question of what unit of analysis matters morally. If the relevant unit is aggregate U.S. employment, Yale and Goldman Sachs are correct that the picture remains stable. If the relevant unit is early-career software workers in 2024, or the 6.1 million low-adaptability clerical workers in 2026, the picture is materially different. Both are accurate descriptions of different slices of reality. The policy question is which slice commands more urgent attention.
The 2027–2030 Inflection
Why Today's Modest Numbers May Be Structurally Misleading
Current displacement figures reflect early-adoption conditions — the compounding effects of AI capability improvements and corporate adaptation cycles are expected to materialise with greater force between 2027 and 2030.
The relatively modest confirmed displacement figures of 2024–2025 — 55,000 direct AI job cuts, 200,000 to 300,000 total AI-attributable positions — must be contextualised within corporate adoption timelines rather than interpreted as a stable long-run equilibrium. Enterprise AI deployment follows a well-documented diffusion curve: pilot projects in 2022–2023, selective deployment at scale in 2024–2025, and deep operational integration reshaping entire functions in 2026–2028. The numbers visible today represent the leading edge of deployment, not its matured state.
The IMF's broader assessment is relevant here: roughly half of the jobs in advanced economies exposed to AI risk wage depression or elimination — not immediately, but as AI capabilities expand and corporate cost structures adapt. ✓ Established [2] The timing of that materialisation is contested — the ALM Corp synthesis, drawing on multiple research streams, places the major compounding effects between 2027 and 2030 — but the structural forces driving it are not. ⚖ Contested [8]
The St. Louis Federal Reserve's August 2025 analysis provides a valuable leading indicator: the correlation coefficient of 0.57 between AI adoption rates and unemployment change across occupations from 2022 to 2025 represents a measured signal, not a crisis alarm. ◈ Strong Evidence [3] But correlation coefficients measured at early-adoption stages tend to understate the eventual effect magnitude, because enterprise adoption is non-linear: it accelerates as software costs fall, as worker familiarity increases, and as competitive pressure forces lagging firms to match early adopters' AI-driven cost structures.
The question of whether the net global employment effect will be positive — the WEF's projected net creation of 78 million jobs by 2030 — or negative is genuinely unresolved. ⚖ Contested [8] What is not unresolved is the distributional question: whether the net effect is positive or negative, the workers with low adaptive capacity will not benefit from newly created AI-era roles. The 2027–2030 inflection is, for the 6.1 million low-adaptability workers, a countdown rather than a distant horizon.
Policy Response Gap
Reskilling, the Credential Wall, and What Governments Are Getting Wrong
Current policy frameworks are oriented toward the wrong population, operating at the wrong scale, and embedded in institutional assumptions built for a different era of displacement.
The dominant policy response to AI-driven labour market displacement — reskilling programmes aimed at helping displaced workers transition into AI-adjacent roles — contains a structural contradiction that the empirical evidence makes clear. If approximately 77 percent of newly created AI-related roles require master's degrees, ⚖ Contested [8] then a reskilling programme offering six-month vocational courses to displaced administrative workers is not a pathway into the emerging AI economy — it is a credentialing gap measured in years of graduate education that most displaced workers cannot afford, cannot access, and in many cases cannot complete while managing caregiving responsibilities that disproportionately fall on women.
The structural mismatch between the displaced cohort and the newly created role requirements is the policy challenge that current frameworks are not designed to address. The U.S. Trade Adjustment Assistance programme, designed to support workers displaced by international trade, offers a template that has been consistently underfunded and narrowly defined. Extending and reformatting such mechanisms for AI-driven displacement — with income support periods of sufficient length to enable genuine credential attainment, childcare provisions to support women with caregiving responsibilities, and geographic mobility support — would require political investments that have not materialised.
The AB Academics January 2026 analysis found that in Nordic countries with strong unions and active labour market policies, AI adoption was associated with narrower wage gaps, while in the U.S. and U.K. it coincided with sharper wage polarisation. [12] MIT's David Autor, in his 2024 NBER working paper, argued explicitly that AI's destructive or constructive application is a policy choice, not a technological inevitability. [11] The implication is that the harm currently being documented is not the unavoidable cost of technological progress — it is the cost of institutional inaction.
Three specific policy failures are identifiable from the evidence base. First, current reskilling investments are insufficiently scaled: independent estimates place total AI-attributable displacement in 2025 at 200,000 to 300,000 positions, ◈ Strong Evidence [8] and the policy apparatus to support them — community college programmes, workforce development boards, state retraining subsidies — was not designed for this volume or occupational distribution. Second, the geographic concentration of risk has not been mapped into targeted intervention: the IMF Working Paper's September 2024 finding that U.S. commuting zones with higher AI adoption experienced stronger declines in employment-to-population ratios between 2010 and 2021 ◈ Strong Evidence [9] identifies specific communities at elevated risk — communities that have not received targeted preparatory investment. Third, the gender dimension of the adaptive capacity gap has not been incorporated into policy design: programmes built without explicit attention to caregiving barriers, credential gap magnitude, and occupational feminisation patterns will fail to reach the 86 percent of the most vulnerable cohort who are women. ✓ Established [1]
The aggregate picture — 40 percent of global employment exposed to AI, a technology uniquely targeting cognitive work at a scale without historical precedent ✓ Established [2] — demands a policy response calibrated to the worst-positioned workers, not the most-visible ones. The 6.1 million low-adaptability workers documented by Brookings and GovAI in February 2026 are not generating Silicon Valley headlines or Wall Street analyst notes. They are, in the title of the Stanford Digital Economy Lab's August 2025 paper, the canaries in the coal mine. [4] Whether governance systems register their warning before the displacement wave builds to its projected 2027–2030 intensity will determine whether AI's labour market legacy is a manageable transition or a structural scar concentrated in the most economically fragile communities in the American workforce.