When the Ladder Was Pulled Up

There is a comfortable intuition about artificial intelligence and work, and it is wrong. It says that rich countries, with the most office jobs and the most to automate, will be hit hardest, while poorer countries with fields, factories and informal work will escape. That is only half the picture. Rich countries are more exposed, but they also have capital, infrastructure and social protection. Poorer countries are less directly exposed, but they have much less capacity to capture the gains and much less shock absorption when development paths disappear.

The IMF estimates that roughly 40 percent of global employment is exposed to AI. In advanced economies the share is around 60 percent, in emerging markets around 40 percent and in low-income countries around 26 percent. At first, that sounds like good news for the poorest countries. But the real risk is not only how many existing jobs can be automated. It is whether the next step upward still exists.

When exposure, preparedness, social protection and demography are put together, a more troubling rule appears: the labour shock from AI is inversely dangerous to demographic need. Where the workforce is shrinking, as in the Nordics, Japan, Korea and parts of China, AI can replace people who are becoming fewer anyway. Where the workforce is young and growing, as in much of Africa, South Asia and the Middle East, new jobs are needed at very large scale. That is where the same technology can become most dangerous.

The countries that least need to replace labour are best equipped to do it. The countries that most need to create jobs are the ones most at risk of having the ladder pulled up.

That is why the word preparedness has to be split in two. The IMF's AI Preparedness Index measures important things: digital infrastructure, human capital and labour-market policy, innovation and economic integration, and regulation and ethics. It is a measure of the ability to use the technology and capture the upside. It is not a complete measure of a society's ability to carry people through the shock. An economy can be well prepared for AI investment and still be socially brittle when jobs, status and local livelihoods disappear.

The United States shows the difference. It has enormous AI capability, capital markets, universities, cloud companies and model labs. But it also has thinner social protection than the Nordics and a culture in which work is often tightly linked to human worth. Two countries can therefore look almost equally ready in an upside index and still face very different political and social consequences when AI presses on the labour market.

The gender dimension — who within the region is hit

Exposure is not distributed evenly between genders. The ILO's March 2026 research brief (Gen AI, occupational segregation and gender equality in the world of work) shows that female-dominated occupations are almost twice as exposed to generative AI as male-dominated ones — 29% versus 16% — and that the gap widens in the highest automation-risk class, where 16% of female-dominated occupations fall compared with only 3% of male-dominated ones. The reason is occupational segregation: women are overrepresented in clerical, administrative and support roles with routine tasks that generative AI can most easily take over, while men dominate construction, manufacturing and physical occupations that are harder to automate. The pattern holds in 88% of the countries studied and is strongest in high-income countries. It adds a gender axis on top of the geographic map: not only which countries, but who within them bears the transition — and it reinforces that a fast shock requires targeted, not gender-neutral, measures.

Source: ILO research brief, March 2026. The ILO's GenAI-specific exposure measure differs from the IMF's 60/40/26 — keep them separate.

The ladder

For seventy years, many developing countries had an imaginable ladder. People moved from agriculture into simple manufacturing, then into more advanced production, exports, education and services. South Korea, Taiwan and later parts of China showed that the ladder was difficult but real. It was not fair, not automatic and not open to everyone, but it existed.

AI threatens to pull up two rungs at once. The first is the industrial rung. Robotics, cheaper automation and geopolitical reshoring make it harder to build development on cheap labour in factories. The second is the service rung. Generative AI reaches into translation, customer service, basic programming, document production, analytical support and the office-adjacent outsourcing that became a route for India and other service economies. What used to be labour arbitrage becomes technology and compute arbitrage.

Dani Rodrik called this premature deindustrialisation even before the generative AI wave: countries get less time inside the productivity machine of industry and begin to deindustrialise at lower income levels than earlier industrialisers did. AI adds another layer. It threatens not only factory jobs but also the alternative service path for countries that never received a full industrial window.

India becomes a key case. Much of its global middle-class promise was built on services, IT and administrative knowledge work. That is not the whole economy, but it is an important development story. When AI makes parts of that work cheaper, faster and less labour-intensive, it affects not only existing jobs. It affects the expectation of young educated people who already feel that a degree no longer gives the security it promised.

Bangladesh shows the other side. Its garment industry has been an export engine and an entry point into formal work for millions of people. But textiles and apparel are also areas where automation, trade policy and demand shocks can quickly change how much labour is needed per exported dollar. The point is not that all those jobs vanish overnight. The point is that the development model becomes thinner exactly when more people need it.

Who owns the machine

It is not enough to ask which countries are exposed. We have to ask who owns the gains. AI is software at the interface, but underneath it is chips, electricity, data centres, networks, models, cloud capital and export rules. That stack is concentrated. A small number of companies and countries control much of the means of production.

This is also a form of digital extractivism: data, demand and usage are generated globally, while compute capacity and returns gather in a few places. Brookings estimates that Africa and Latin America together account for only about 3% of global AI compute capacity, while BloombergNEF reports that roughly three quarters of the 23 GW of data-centre capacity under construction in September 2025 was in the United States.

The United States' AI Diffusion Rule of January 2025 tried to make this geography explicit by dividing the world into different tiers of compute access. The rule was rescinded in May 2025 before it became the stable map many described it as. But the underlying point did not disappear with the rescission. Access to advanced chips and data-centre capacity is now geopolitics. The map is moving, but the gates are real.

That makes redistribution much harder. Robot taxes, basic income, AI funds and dividends from automation all rest on the same quiet assumption: that there are AI gains inside one's own jurisdiction to tax. For countries that do not own the models, chips, clouds or capital, that logic is not enough. They need access, investment, technology transfer and institutions that keep productivity gains from simply passing through their economies and being booked somewhere else.

The coupled transition

Work is not the only variable. AI is now coupled to energy and climate. The IEA estimates that global data-centre electricity use could more than double by 2030, reaching around 945 TWh in a base-case scenario. That means AI is not just a digital phenomenon. It needs grids, land, cooling, power generation and capital. Energy policy therefore becomes part of AI policy.

For oil states, this creates a double movement. The green transition threatens fossil revenues. AI and automation may also reduce the ability to park young citizens in public jobs or simple private services. But data centres and AI investment can also create new demand for cheap energy, especially gas and electricity. AI becomes both threat and lifeline. In such a situation, the climate transition may meet more resistance, not less.

The witness

We have seen a milder version of this before. The China shock destroyed parts of US manufacturing employment. The most important lesson was not only the number of jobs lost, but how local labour markets failed to heal in the way standard theory expected. Wages, participation and social outcomes remained weak for a long time in many affected regions. People did not move frictionlessly to the next opportunity. Many stayed where the old economy had broken.

The AI shock could be faster, broader and more cognitive than the China shock. It may hit people who invested in education, status and future promises. The optimist is often right in the aggregate: new tasks and new jobs appear. But the aggregate is not a place to live. The individual town, family and career can carry the cost long before national accounts show recovery.

What it depends on

The outcome is therefore not determined by exposure alone. It depends on three things: how fast capability rises, how broadly access is shared and who owns the gains. These variables shape three worlds. In acceleration, technology outruns safety nets and divergence grows. In friction, the world fragments into blocs, export rules, bottlenecks and regional emergency measures. In sharing, open access, investment and institutional transfers are used to put at least parts of the ladder back.

It is tempting to ask whether AI takes the jobs. The larger question is whether AI takes the development path. If young countries can no longer climb through the factory, can no longer climb through the service office and do not own the digital means of production, then AI is not just a labour-market question. It becomes a question of global convergence, migration, climate policy and political stability.

This is not a forecast. It is a space of choices. The ladder can be pulled up by markets, geopolitics and passivity. It can also be rebuilt. But first we have to stop pretending that the map shows the same development path as last time.

Sources and checkpoints

This essay is based on open sources and published indices. Some quantitative claims from the raw draft have deliberately been softened or removed where the evidence was weaker or the policy situation had changed.

Rolf Skogling writes AI-skiftet from a practical, industry-near perspective, focused on how AI is actually used in organisations and production.