Labour Market Polarisation: How AI Is Rewriting Who Wins and Who Loses at Work

For decades, technology mostly automated routine, repetitive jobs — factory assembly lines, data entry, basic bookkeeping — while higher-skilled, judgment-heavy work stayed safely human. Economists called this pattern ‘labour market polarisation’: middle-skill jobs hollowed out while both high-skill and low-skill service jobs grew.

Generative AI is complicating that story. Unlike earlier automation, it is proving capable of handling tasks long considered ‘high-skill’ — drafting legal documents, writing code, analysing data, producing marketing copy — while many low-wage, hands-on service jobs (cleaning, caregiving, skilled trades) remain largely untouched for now. Early evidence shows US employment growth slowing sharply in white-collar fields most exposed to generative AI, including marketing, graphic design, and office administration, even as demand surges for AI-literate technical roles paying a substantial wage premium.

The result may be a new kind of polarisation — one that does not automatically reward traditional ‘high-skill’ credentials the way past waves of automation did. Whether this trend widens or narrows inequality depends heavily on policy choices: how retraining is funded, whether AI's productivity gains are shared with displaced workers, and how quickly new job categories emerge to absorb workers pushed out of shrinking occupations.

The classic ‘skill-biased technical change’ story of the 1990s and 2000s described a labour market splitting into two growing tiers — highly-paid professional and managerial jobs, and lower-paid personal service jobs — while middle-skill, middle-wage occupations (manufacturing, clerical work, routine administration) shrank as computers and industrial robots absorbed their tasks. Wages for the surviving high-skill jobs rose steadily, rewarding a college-education premium that defined labour economics for a generation.

Generative AI breaks from this pattern in an important way: its comparative advantage lies in cognitive, language-based, and pattern-recognition tasks, which happen to overlap heavily with exactly the professional, white-collar work that previously sat safely on the ‘high-skill, high-wage’ side of the divide. Recent labour market data bears this out. Employment growth in the United States has slowed markedly in occupations most exposed to generative AI — marketing and consulting, graphic design, office administration, and call centre work — even as job postings for AI-related technical roles have surged. In South Asia specifically, AI-related postings now offer wages roughly 30% above comparable white-collar roles, reflecting a scramble for scarce AI-literate talent even as adjacent white-collar categories soften.

Meanwhile, many lower-wage, physically-grounded service jobs — caregiving, cleaning, skilled trades, hospitality — remain largely insulated from generative AI's current capabilities, which do not yet extend reliably into physical manipulation or unstructured real-world environments. This is a break from the pattern predicted by ‘routine-biased technical change’ models, which assumed automation would climb the wage ladder from the bottom up. Instead, the exposure pattern looks closer to a ‘hollowing from the top-middle,’ pressuring mid-to-upper white-collar roles whose tasks are cognitively routine even if formally credentialed as ‘skilled.’

Distributional consequences are already visible along lines beyond just occupation. Analysts flag that women, younger workers, and those in entry-level roles face disproportionately higher exposure to AI-driven task automation, partly because entry-level white-collar positions often consist of exactly the structured, judgment-light tasks — first-draft writing, basic research synthesis, routine analysis — that generative AI now performs competently. This raises a specific policy concern distinct from past automation waves: if AI erodes the traditional ‘entry-level’ rung of white-collar career ladders, it could disrupt the pipeline through which workers historically developed into senior professionals, with effects that only become visible years later.

Not every forecast is pessimistic. Several labour economists and international bodies, including a recent assessment from the European Training Foundation, argue that AI is ‘far more likely to change jobs than eliminate them outright’ — reshaping task composition within occupations rather than eliminating entire job categories wholesale. Under this reading, the critical variable is not whether AI displaces workers, but whether displaced workers can move into newly-created roles quickly enough, and whether wage growth in AI-complementary occupations is broad-based rather than concentrated among a narrow technical elite. Which outcome prevails depends heavily on policy: funding for retraining and reskilling, portable benefits that do not lock workers into declining occupations, and whether productivity gains from AI adoption are captured primarily by capital owners or shared more broadly with labour through wage growth — a distributional question that will likely define labour economics debates through the rest of this decade.

The theoretical starting point for understanding labour market polarisation is the task-based framework developed by economists such as Daron Acemoglu and David Autor, which models occupations not as fixed skill bundles but as collections of discrete tasks, each with different degrees of automatability. Under this framework, technology does not automate ‘jobs’ wholesale but substitutes for specific tasks within jobs, reallocating remaining tasks toward workers whose comparative advantage lies in judgment, creativity, or physical dexterity that machines cannot yet replicate. The routine-biased technical change (RBTC) literature that emerged from this framework in the 1990s-2010s explained polarisation as a U-shaped employment response: routine middle-skill task bundles were disproportionately automated by industrial robotics and enterprise software, while non-routine cognitive tasks and non-routine manual tasks both grew as a share of employment.

Generative AI complicates the RBTC framework because large language models exhibit comparative advantage in a task category the framework did not clearly anticipate: non-routine cognitive tasks that are nonetheless language-based, pattern-recognition-heavy, and reproducible from training data — first-draft writing, code generation, data synthesis, image creation, customer correspondence. These tasks were previously assumed safe from automation precisely because they require ‘judgment’ in some informal sense, yet large language models can now perform serviceable, if imperfect, versions of them at a fraction of the marginal cost of human labour. Recent theoretical work formalises this using an ‘AI automation exposure index,’ applying frameworks derived from Moravec's Paradox — the long-standing observation in robotics and AI research that tasks easy for humans (physical manipulation, sensorimotor coordination) are often hard for machines, while tasks hard for humans (formal logic, pattern recognition across large datasets, drafting from templates) are often comparatively easy for machines. Generative AI is, in effect, automating precisely the tasks Moravec's Paradox identifies as ‘AI-easy’ — which happen to be concentrated in professional, credentialed, historically well-paid white-collar work.

Empirically, this produces occupational employment data that looks different from the classic hollowing-out-the-middle pattern. Employment growth has slowed in marketing consulting, graphic design, office administration, and call centre roles — occupations that sit toward the upper-middle of the wage distribution and require post-secondary education, not the blue-collar middle-skill occupations RBTC models predicted would be most exposed. This has prompted labour economists to describe an emerging ‘reverse polarisation’ or at minimum a modification of the traditional U-shape, where the most AI-exposed occupations cluster in a band that RBTC-era models treated as relatively automation-resistant.

The wage and inequality implications operate through at least three separate channels. First, a direct displacement channel: workers in highly AI-exposed occupations face reduced labour demand, pushing down wages or employment in those specific roles, particularly for entry-level positions where task content is most standardised. Second, a complementarity channel: workers who can direct, verify, and integrate AI outputs into higher-value work see rising demand and wages, creating a new, narrower category of ‘AI-complementary’ skilled labour distinct from the broader category of credentialed professional labour. Evidence of this channel is visible in the roughly 30% wage premium reported for AI-related technical postings in South Asian labour markets relative to comparable white-collar roles. Third, a capital-labour distribution channel operates at the macro level: to the extent AI substitutes for labour and raises capital's share of value added, aggregate wage growth could lag productivity growth economy-wide, a dynamic economists studying the distribution of income between capital and labour have flagged as a first-order concern.

Demographic exposure compounds these channels unevenly. Analyses consistently find women, younger workers, and entry-level employees disproportionately exposed to AI-driven task automation — partly a function of occupational segregation and partly a function of career-stage task content, since entry-level roles across many professions consist disproportionately of the structured, judgment-light tasks generative AI performs competently. If this pattern holds, it raises a structural concern for career-ladder economics: professions have historically trained senior practitioners by having them perform junior, task-routine work under supervision before advancing to judgment-intensive responsibilities. If AI absorbs the junior tier of tasks across law, consulting, journalism, and software engineering simultaneously, the traditional apprenticeship-style pipeline into senior roles could narrow, with consequences for skill formation that would only become measurable in aggregate data five to ten years hence.

Policy responses under discussion internationally include portable, sector-agnostic retraining funds, stronger social insurance for displaced professional workers who previously had limited exposure to unemployment risk, and — more speculatively — mechanisms to ensure AI-driven productivity gains are shared with labour through wage growth or reduced working hours rather than accruing predominantly to capital owners and a narrow technical elite. Which combination of policies gains traction will likely differ substantially by country, shaped by existing labour market institutions, union density, and the political salience of AI-driven displacement relative to other economic concerns competing for policy attention in 2026.