Different Countries, Different Collapses (Part 1)

The Hollow Middle becomes self-reinforcing when three things happen at once: entry pathways narrow, bargaining power weakens, and growth no longer depends on broad middle-class demand. When AI enters the room, it accelerates these pressures and changes where societies choose to absorb them.

My last instalment was built largely from US data, partly because the American version is easier to see: layoffs, weaker entry-level hiring, consumer frustration, political anger and anxiety about whether work still leads to security. But the country comparison starts from a different premise. The Hollow Middle will not appear everywhere in an American form. AI diffusion is uneven, and so are the societies into which AI arrives.

AI diffusion is not spreading uniformly. It is concentrated in economies with stronger institutions, higher governance capacity and deeper technological infrastructure.

This is consistent with the IMF’s AI Preparedness Index, which assesses preparedness across 174 countries using macro-structural indicators such as digital infrastructure, human capital, innovation and regulation. The countries adopting AI fastest are not, in general, the countries least able to absorb its consequences.

That alone should make us cautious about the simplest version of the argument: that AI adoption is already generating the same middle-class crisis everywhere.

It is not.

The more unsettling pattern is that AI is arriving after several foundations of middle-class reproduction had already begun to weaken.

Across very different economies, households are finding it harder to form, workers are struggling to claim a stable share of growth, and families are less able to assume that tomorrow’s security will be stronger than today’s. Fertility has fallen in most countries. Housing has become less affordable in many advanced economies. Where data is available, labour’s share of growth has weakened or stagnated across several economies that otherwise look institutionally very different.

These indicators should not be read as proof that AI single-handedly caused the Hollow Middle. Fertility is shaped by housing, childcare, culture, urbanization, gender norms and family policy. Housing affordability is difficult to compare globally at the same level of quality. Labour-share data is incomplete in several countries we would most want to compare.

The narrower claim is that these indicators all matter for middle-class reproduction, and several are weakening at the same time. That matters for AI because productivity can rise while the social machinery that turns growth into broad security weakens. In a task-based view of automation, the key question is whether technology only substitutes for labour in existing tasks, or whether it also creates new labour-intensive tasks, occupations and career ladders. AI’s middle-class effect will depend on that balance.

This is why the country story matters. AI is arriving into systems where parts of the social contract were already fraying. In advanced economies, it may compress entry-level professional work faster than institutions can redesign career pathways. In emerging economies, it may threaten a services-led ascent that has not yet delivered for some of the largest young cohorts in history. David Autor’s argument that AI could rebuild middle-class work if deliberately designed around worker expertise is the important counterpoint: by extending the reach of workers with foundational expertise into decision-making tasks currently reserved for elite professionals.

The pressure is broad. The exposure is uneven. The next question is where each society allows the pressure to show up.

The same pressure does not appear through the same signal

That is where the country differences begin.

The first dimension is pressure visibility: how publicly legible middle-class pressure becomes. This is a measure of whether strain becomes visible, contestable and socially recognizable. In some countries, pressure moves quickly into public language: labour-market anxiety, political anger, protest, media attention, consumer sentiment. In others, it remains more private or indirect: delayed household formation, lower consumption, migration, precautionary saving, informal work or quiet withdrawal from aspiration.

This is why no single indicator can measure middle-class pressure globally. Unemployment may be the right signal in one country and the wrong signal in another. Fertility may reveal long-cycle insecurity in one system and mostly reflect family policy, housing or gender norms in another. Reported distress may be a genuine alarm in one society and a muted signal in another. Albert Hirschman’s classic distinction between exit and voice is useful here. When people experience decline, they can demand repair, or they can withdraw.

The loudest cases are not simply the most advanced AI economies. In this sample, for instance, pressure is most publicly legible in countries where protest, distress and labour-market strain are all visible at once, such as Spain and South Africa. Japan shows the quiet version: pressure appears more through delayed adulthood, subdued consumption and demographic retreat. China is better read as a mixed or administratively muted case: some pressure appears through household caution, fertility decline and savings behaviour, while public contestation remains a less complete measure of underlying strain.

In high-visibility systems, track public and fast-moving signals: protest, reported distress, unemployment, youth exclusion, hiring, layoffs, real wages, sentiment and political salience. In lower-visibility systems, track indirect or slower-moving signals: fertility, household formation, savings, migration, consumption weakness, internal workplace adjustment and gaps between reported distress and observed behaviour.

This is the first step toward classification. Countries are first grouped by whether pressure becomes public or remains private; whether it is voiced, muted, displaced or withdrawn from view.

Adjustment channels: who carries the cost of AI transition?

The second dimension asks a different question from visibility. Who carries the adjustment when AI starts changing tasks, roles and costs? Firms, institutions, workers, households, new entrants, or those outside formal protection?

In some economies, adjustment falls more heavily on workers and households. Firms can reorganize roles, freeze hiring, reduce headcount, change contracts or shift work arrangements with fewer buffers around workers. The burden of transition is more likely to appear through job insecurity, weaker wage progression, thinner entry-level pathways, retraining costs, career resets or household financial pressure.

In other economies, more of the adjustment is absorbed or negotiated inside firms and institutions. Employment protections, collective arrangements, social insurance and workplace norms can slow the move from task change to job loss.

A third pattern appears where many workers sit outside formal protection. In these systems, firms may still adjust quickly, but the cost is often pushed onto workers more privately. AI-era disruption may appear as underemployment, unstable service work, informal work, migration, or a longer wait for secure jobs.

This is why the indicators are better read as clues about who absorbs the shock. Formal legal constraints, such as redundancy pay and notice periods, show how much responsibility firms carry when they restructure. Collective constraints, such as union density and labour-rights compliance, show whether adjustment is likely to be negotiated. Protection gaps, such as limited social protection and informal employment, show how exposed workers are when adjustment happens.

In this sample, the United States sits closer to the worker-and-household side than most advanced economies. This is because adjustment is highly individualized: firms can move quickly, workers carry more job and income risk, and career transitions often depend on personal savings, mobility and retraining. By contrast, Sweden, Germany and France sit closer to the firm-and-institution side, where adjustment is more likely to be mediated through workplace rules, social insurance or negotiated restructuring.

Emerging markets need a separate reading. In countries like India and Vietnam, large protection gaps mean workers and households may absorb much of the transition outside formal systems. Mexico and South Africa are more mixed: both have significant worker exposure, but also formal institutions that shape adjustment for protected workers. Countries like China and Saudi should also be read carefully, because adjustment may be filtered through state, firm, migrant-labour or social-protection systems.

From dimensions to archetypes

The first two dimensions explain different things. Pressure visibility asks whether strain becomes publicly legible or remains private and indirect. Adjustment absorption asks who carries the cost when AI changes tasks, roles and labour demand: workers and households, or firms and institutions.

Neither dimension says whether a country can solve the problem. A signal can be loud and still lead to the wrong response. A signal can be muted and still trigger action. A system can cushion workers in the short term while allowing career ladders to thin over time.

The following archetypes show where AI-era middle-class pressure is likely to travel first. The matrix is deliberately simple. One axis asks whether pressure becomes public or stays muted. The other asks whether adjustment is carried more by workers and households, or absorbed more by firms and institutions. Countries can move between archetypes as AI adoption, labour-market pressure and institutional choices evolve.

The archetypes that follow are best read as a map of where to look and what to fear. Part 2: Four Ways AI-Era Pressure Travels will examine the country-specific patterns.

Cover photo by Donald Wu on Unsplash