He asked me: Is Asia and the Pacific making social progress? As I explain the video, it’s been great news over the long term but there are massive disparities and some concerning trends in the aggregate and underneath. Check out the video below.

In this post, I will go a bit deeper into the question how well income explains social progress. I will also go beyond Asia-Pacific and plot a few interesting charts.

Income as a means, not an end

Income is a means, a tool, an instrument. It helps with the purchase of food, housing, schooling and healthcare. And it also pays for the public services and safety nets that governments provide. But what we ultimately care about is not income itself but instead about what income allows people to be and to do.

Amartya Sen put it like this: “to live long and healthy lives, to be safe, to learn, to have a voice and to choose the lives they want to lead”.

That perspective bears also on how we talk about economic growth or growth in income. I often put it this way: without growth there is no human development, but growth is only the starting point.

Some economists argue that growth is not only necessary but sufficient. That may be close to the truth for material living standards. But it does not hold for broader concepts of wellbeing. And this is then also why we need measures that track outcomes more holistically…. like the Social Progress Index published by the Social Progress Imperative.

The SPI’s 2026 edition scores 171 countries on 57 social and environmental outcomes. None of them is a monetary income variable: GDP, income and spending are all left out, and deliberately so. The 57 indicators are grouped into 12 components:

How do income and social progress line up?

Closely. About 83% of the variation in social progress across countries traces back to differences in (log) GDP per capita. The interesting part is the other 17% that income leaves unexplained.

So let us look into this by examining the residuals of this relationship. In other words, what is the difference between a country’s actual SPI and the one predicted by its income level.

Asia-Pacific

The chart shows the gap for 28 economies in Asia and the Pacific.

The results are striking. Note how Solomon Islands, Japan and Timor-Leste beat their income peers by the widest margins: by 7 to 8 points. Japan’s income per head rose just 13% between 2011 and 2025, but its social progress score never fell in any year, and in absolute terms it now is above both (Republic of) Korea and the United States.

At the other end, Singapore falls 4.6 points short of its income peers, China 2.5 and India 2.2, even though China, India and Indonesia improved faster than almost anyone over those fourteen years.

Other regions

I will show the results for Arab States, Eastern Europe and Central Asia, Europe and North America, Latin America and Caribbean, and Sub-Saharan Africa.

The dispersion is far greater in the Arab States than it is in Asia-Pacific. The State of Palestine is obviously serious affected by the conflict, which seems to have pushed down its income levels by far more than what the SPI captures.

Very large disparities as well. Low scores for Russia and Turkmenistan. High score for Kyrgyzstan.

Interesting to see how Europe dominates North America. Finland at the top.

Brazil outperforms, but Mexico is under what its income would predict:

Also in Sub-Saharan Africa we see great disparities between countries. Despite being depicted as such sometimes, this is a far from homogeneous region.

Not a verdict

The gaps are not a verdict on how far these countries have come. They show only the deviation from what is typical among income peers. But they do show how much more social progress a country’s income could support, and the indicators of the SPI show where to look for it.

For a more detailed analysis on the SPI, check out the post below.

The discussion on AI centered largely on two themes: the boom, which concerns the size of the economic pie expanded by AI, and the bubble, which concerns the risks of investing ahead of the returns.

A third theme received much less attention: how will the gains from AI be distributed across countries? Singapore’s President, Tharman Shanmugaratnam, was a notable exception. He put the risk of a new Great Divergence between advanced and developing economies at the heart of his intervention.

This post amplifies Tharman’s message. He is right to put the risk of a new Great Divergence at the centre of the AI debate. If we get this wrong, AI could pull countries further apart, reversing decades of economic convergence and deepening the consequences of between-country inequality — from slower poverty reduction to greater social and geopolitical tensions.

The stakes are therefore enormous: roughly 90 percent of the world’s population may be on the wrong side of this divide. I will explain where that figure comes from below.

AI at the Annual Meetings

The discussions offered plenty of reasons for both optimism and caution. Here is a quick rundown of what stood out to me at the Annual Meetings.

The boom

It is striking how the terms “savings glut” and “secular stagnation”, which dominated policy discussions for years, have almost completely faded into the background. As Bank of Canada Governor Tiff Macklem observed, AI now dominates the conversation, with its impact becoming increasingly visible in the data. The savings glut has given way to an investment boom, as reflected in GDP, stock and bond markets, long-term interest rates and trade. Secular stagnation has given way to expectations of a productivity boom.

Two points are relevant here.

First, even considering the lower end of expectations, the investment under way and its potential economic payoff can be characterized as remarkable. The IMF expects global AI investment relative to GDP to exceed what went into railroads, the electricity grid or telecoms, and AI hardware already accounts for more than a tenth of world goods trade. The eventual productivity dividend could add 0.5 percentage point to annual growth, equivalent to adding an economy the size of ASEAN over a decade. In advanced Asia, the projected gain of 0.5 to 1.0 percentage point could more than offset the drag from population ageing.1

Second, we are only at the very beginning. The investment boom right now reflects the laying of the tracks. The big-ticket benefits will come later with diffusion and adoption. In the Per Jacobsson Foundation Lecture, Anthropic co-founder Jack Clark painted an especially bullish picture, arguing that AI will raise the pace of idea generation, which has been a drag on total factor productivity growth for years. He mentioned that frontier models are already raising the pace of discovery in mathematical research, drug discovery and laboratory work. Imagine what the world could do if its 10 million (or so) scientists were complemented by AI and could do the work of 1 billion of them.

One Anthropic model projects that AI could raise US output to $44 trillion by 2030, which is a 50 percent increase. Clark acknowledged that this may be outlandish even though he added that it is implied by the current trajectory of progress. It may be outlandish indeed, as physical bottlenecks in the real world (eg. laboratory capacity and capacity to run clinical trials) will put a speed limit on how many and how quickly scientific breakthroughs can translate into productivity gains.

The bubble

The second theme was the concern that there might be too much money chasing too few returns. The IMF mentioned Amara’s Law: we tend to overestimate a technology’s short-term impact and underestimate its long-term effects. If AI follows this pattern, we may now be in a precarious phase. And leverage among hyperscalers and global exposure to US equities could amplify any shock if one is triggered.

There are also concerns about the scale of the investment. The data-centre build-out in the US may require over $10 trillion by 2032, equivalent to 3.6 percent of GDP annually. Major cloud companies are expected to spend more on capital expenditure than they generate in cash. And it seems that they are increasingly relying on debt, joint ventures and leases. Hyperscalers are reported to be investing in AI labs that, in turn, buy computing capacity from them. This creates circular financial exposures that magnify the risks.

Governments are not insulated. Those that are highly indebted will have little room to respond if AI investment contracts suddenly. And even if AI delivers substantial productivity gains, we must also worry that a shift in income from labor to capital could weaken tax revenue in countries where labor taxes account for a substantial share of the tax base.

The divergence

Whereas the boom could deliver extraordinary gains, the bubble might generate a painful correction. But neither tells us who will ultimately be most affected, be it positively or negatively. And, above all, neither sheds light on whether AI will help developing countries catch up or push them further behind.

President Tharman raised this latter issue, and I found his intervention one of the most inspiring at these Annual Meetings. He spoke on the very first day, at the World Bank’s opening event on Small AI. Asked to name the one goal he would set for the next five years, he answered:

“Avoid a new divergence between the developing world and the advanced world and see continued convergence, as we’ve seen in East Asia especially, using AI, using technology, and developing the best of human capabilities.”

This statement, which embodies as much an expression of hope as a stark warning, builds on a theme Tharman has emphasized before. I remember well his speech at last year’s Annual Meetings in Washington DC, where he offered another insight with direct bearing on the development challenge posed by AI. He argued:

“Industrial policies may be back in vogue, but I’ll have to say that success in job creation requires social policy on an industrial scale. It requires us to develop every human talent, deepen and upgrade skills continually, and advance social mobility in the process.”

The two messages are closely connected. Yes, sustained convergence depends on access to new technologies. But it also depends on the ability to develop human capabilities and on the institutional capacities to spread the benefits widely.

A Next Great Divergence?

In what follows, I will examine Tharman’s concern into greater depth in four steps.

• AI is landing in a deeply unequal world, much of which is still living with the consequences of the Great Divergence associated with the First and Second Industrial Revolutions. Despite some important successes, global convergence has been limited in magnitude and geography.

• Building on UNDP’s recent report The Next Great Divergence, I will argue that we need a new global conversation on AI and development. The current framing is too limiting. We need a broader lens and a different one.

• I will then examine what such a conversation might look like, through a more holistic frame that considers AI’s impacts on people, the economy and governance, and the many channels through which it can generate both dividends and disruptions.

• To conclude, I will highlight why both the upside and the downside are likely to fall unevenly and, unfortunately, systematically so to the disadvantage of countries further back from the frontier.

Convergence, but for whom?

For the past 200 years, around 90 percent of the world’s population has lived in regions with per capita incomes below 40 percent of the frontier.

That’s what the chart below shows. It tracks regional GDP per capita relative to Angus Maddison’s “Western Offshoots”: Australia, Canada, New Zealand and the United States. Every region except Western Europe has remained below that threshold throughout these last two centuries.

Three points about what this long-run time series shows us:

• The Great Divergence. In 1820, although the process had begun earlier, regional incomes ranged from roughly a third to half of frontier levels. The later part of the First Industrial Revolution and the Second Industrial Revolution widened the gap for more than a century. By 1950, East Asia, South and Southeast Asia, and Sub-Saharan Africa had fallen below a tenth of frontier income; Latin America stood at around a quarter.

• Selective and partial convergence. East Asia rose from below a tenth to nearly two fifths of frontier income, mostly since 1980. Eastern Europe and the Middle East recovered to just over a third. Elsewhere, progress has been limited: Latin America remains around a quarter, South and Southeast Asia around a seventh, and Sub-Saharan Africa, at about 6 percent, is further behind than in 1950.

• The uneven world we have today. Only Western Europe and, by construction, the Western Offshoots exceed the 40 percent threshold, yet they account for just around a tenth of the global population. The other nine in ten live in regions below it.

Regional averages tell only part of the story. Yes, averages mask the heterogeneity underneath. That’s true. But curiously the results are not too far off the 90 percent figure when we look at it through the more granular analysis of individual income or consumption at the household level.

Since such data does not go back so far, let’s have a look at 2025 only. The global income distribution in that year can be constructed from household surveys and it offers a strikingly similar picture. Using the same frontier benchmark, the population-weighted average income or consumption per person in the Western Offshoots is $92 a day. And almost 90 percent — 84 percent to be precise — of individuals globally live below 40 percent of that level.

So yes, convergence has happened, but it has reached only a minority of humanity. It has also been insufficient to offset the centrifugal forces unleashed by the Great Divergence during the First and Second Industrial Revolutions.

The key question before us today is whether AI will help more countries catch up or whether it will reinforce the advantages of those already ahead. In other words, will see more convergence around the world? Or, will there be more “divergence in convergence” which might well lead to the era of the Next Great Divergence?

A different conversation

These questions are the starting point of The Next Great Divergence, UNDP’s flagship report published in December 2025. Developed over a year with nine partner institutions spanning the US and Europe to China and India, the report examines how AI could widen inequalities between countries.

The subtext of this report is that we need a new global conversation about AI and development. It needs to change in three ways:

First, the conversation needs to be more holistic. The current emphasis is predominantly on growth, productivity and jobs. We need to look beyond these measures and consider the wider consequences for human development: health and education, social protection, environmental sustainability and public institutions.

Second, it needs to be more balanced and evidence-based. The debate on AI is currently dominated by hype and alarmism. AI will mean different things to different sectors and countries. We need to analyze its dividends and disruptions in a granular way, asking who benefits, who bears the brunt of the disruptions, and under what conditions. We need more balance and evidence.

Third, it needs to pay more attention to inequality between countries. There is an urgent need to cast the net more widely when assessing AI’s implications for inequality. So far, the debate has focused largely on inequality within countries, which is an important topic in its own right. But we must not overlook the substantial inequalities between countries and the risk that AI will widen them further.

Dividends and disruptions

Seen through that broader lens, AI brings both dividends and disruptions across three interconnected domains: people, the economy and governance. Within each domain, there are multiple channels through which AI could affect development.

People

The table below provides some examples offered in the report of how this channel might operate, both positively and negatively.

Let me repeat a few here: AI tutors and automated diagnostics could considerably expand access to education and healthcare, which may help reach people underserved by existing systems. But such gains depend on connectivity, digital skills and the quality of the services available. Some people don’t have reliable internet access. Others may not have digital identities. So they may end up being excluded from increasingly automated public services and social safety nets.

Economy

As discussed earlier, AI could generate substantial productivity gains, but these are unlikely to be distributed evenly across sectors or workers. Routine tasks face particular exposure to automation. Also, women in clerical occupations and young people entering the labour market may be disproportionately affected. Whether new activities and jobs emerge quickly enough to offset these disruptions will depend on many things, such as how firms adapt, how workers acquire new skills, and whether economies can create opportunities at scale.

And below are more examples that are examined in greater detail in the report:

Governance

AI could help governments listen to citizens, anticipate emerging needs and deliver public services more effectively. But it also clear that it could make decisions about welfare, hiring and access to services less transparent. In the process, existing biases could be reproduced or amplified. So we need institutional capacity to assess these systems. We need to be able to hold them accountable. And we also need to be able to correct their failures.

More examples, again, here, with the details in the report:

So in a nutshell, across these three domains of people, economy and governance, AI offers plenty of opportunities for development but also creates new risks of exclusion and disruption. The central question is who will benefit from the dividends and who will bear the brunt of the disruptions.

Two gaps that drive divergence

The overarching concern is that, left to itself, AI could be a driver of divergence because of two gaps: a capability gap and a vulnerability gap.

The capability gap

Turning AI into development gains goes well beyond having access to a model or an application. It depends also on reliable electricity, connectivity, data, skills, capable institutions and the financial resources to combine these ingredients.

The point here is that wealthier countries are generally better positioned to make these complementary investments, integrate AI into firms and public services, and deploy AI at scale. Countries lacking these essential foundations may therefore gain access to the technology without being able to translate that access into comparable increases in productivity or living standards.

The vulnerability gap

Even where AI raises productivity, the transition can impose substantial costs on workers and communities. Managing those costs requires many things, including effective social protection, reliable registries, reskilling programmes and institutions capable of scrutinizing AI systems and responding when they fail.

Countries with stronger fiscal and administrative capacity are generally better placed to support displaced workers, protect those at risk of exclusion and adapt their policies as the technology evolves. Countries with narrow tax bases and limited implementation capacity face many of the same disruptions with fewer means to respond.

How the two gaps compound

There is an essential distinction between diffusion and adoption. Access to AI can spread across borders far faster than countries can develop the capacity to use it effectively and manage its consequences.

Effective adoption depends on complementary investments in infrastructure, skills, institutions and policies. As the World Bank’s World Development Report 2016: Digital Dividends argued, digital technologies deliver their full benefits only when accompanied by the necessary “analogue complements”: appropriate regulations, skills and accountable institutions.

The two gaps can reinforce each other. Countries that struggle to capture the gains will have less to invest in capabilities and protection; those that cannot manage the disruptions will face deeper costs, weakening their capacity to invest further. Meanwhile, countries with stronger capabilities can move faster, capture a larger share of the returns and reinvest them. Left to itself, AI could reinforce the very advantages that have kept the global income distribution so unequal.

Where this leaves us

The two gaps are already visible in the data. The chart below plots countries against the IMF’s AI Preparedness Index, which combines measures of digital infrastructure, human capital, innovation, regulation and other factors. Countries are categorized by income group, and bubble size represents population size. Note the systematic differences between income groups. The population-weighted average (horizontal line) is about 0.7 in high-income countries compared with just 0.3 in low-income ones. Those least prepared to capture AI’s dividends are also likely to have the fewest resources to manage its disruptions.

Let’s return to Tharman’s two messages: his warning about divergence at the 2026 Annual Meetings and his call for social policy on an industrial scale at the 2025 meetings. The risk of a Next Great Divergence is real. But it is not inevitable. Policy will shape the outcome, and social policy must be a central part of the response.

East Asia offers important lessons. Its convergence towards frontier living standards rested not just on adopting frontier technologies, but also on mass education, continuous skills upgrading, and institutions that helped spread the gains. AI now raises the bar and makes these complementary investments all the more urgent. Countries need to build these foundations if they are to turn access to technology into broad-based gains and help people navigate the disruptions.

The boom will continue, and the bubble will keep the financial pages busy. Divergence unfolds more slowly, but its effects could last for generations — as the First and Second Industrial Revolutions showed. With 90 percent of the world’s population living in regions that remain far below frontier income levels, a new divergence could leave billions further behind.

China, India and Indonesia improved faster than almost anyone. And, strikingly, Japan kept on climbing up despite its slow-growing economy. The United States, the richest large economy on earth, was among the eight countries whose score fell during a period where its income per head rose by a quarter.

One worrying trend: scores for rights and voice, which is one of the components of the index, deteriorated in 121 countries, even as almost everything else improved.

The broader lesson? Economic growth helps deliver social progress, but not reliably so and not on every dimension. And the opposite can also be true. The SPI offers examples of countries achieving social progress despite sluggish growth, and others experiencing rising incomes without commensurate social gains.

In other words, progress without growth and growth without progress.

What is the Social Progress Index?

The Social Progress Index, published by the Social Progress Imperative, scores 171 countries on 57 social and environmental outcomes. None of them is an economic variable: variables such as GDP, income and spending are deliberately left out. And this is what makes this index truly unique… and valuable.

The indicators are organized into twelve components sorted by three dimensions: (1) basic human needs, (2) foundations of wellbeing and (3) opportunity. The details are shown below.

Why am I a fan? Three reasons:

• The SPI offers a dashboard, not just an index. This allows us to look at progress in a granular way — and, I must admit, in a much more granular way than our own Human Development Index at UNDP.1 The SPI traces what is actually happening in ways that policy can then address. Eg housing, safety, press freedom, air quality, university access, … all of these are concrete areas that policymakers of different stripes can relate to.

• The SPI offers a purely non-monetary perspective on development. We live today in a one-sided world where GDP is king despite its narrow and instrumentalist character. We also live in a world where progress on the non-monetary dimensions of development is less than stellar. The slow progress rate on the SDGs is the best example of that. So a focus on the non-monetary dimensions of development is very welcome.

• The SPI is furthermore ordered in the way needs are ordered: basic needs first, then the foundations of wellbeing, then opportunity. The sequence mirrors how societies typically build on each other’s gains, from food, water, shelter and safety, through schooling, connectivity and health, to rights, inclusion and advanced education. To be clear, this ordering is merely a reading aid rather than a ranking or sequencing of importance: the three tiers carry equal weight in the index.

Some striking results

The chart below traces the evolution of the SPI against GDP per capita — let’s just call it income per capita. There is a clearly positive relationship between the two: richer countries tend to have higher social scores, and vice versa. At the same time, the relationship is not mechanical and we see considerable variation in terms of how growth “correlates with” social progress.2

The big story here is an Asian story.

The three most populous countries of the region added a 7-9 points to their score and also managed to make progress in the rankings. China climbed up 17 places to 73rd position, Indonesia 11 places to 78th, and India 6 places to 109th. The three countries are home to about three billion people, so it is remarkable that each of them managed to improve by 1.5-2x the world average of 4.8 points. Compare this with Brazil, which added far fewer points during a period of slower and more volatile growth, even though, to Brazil’s credit, the social progress curve was very steep.

The experience of Japan is especially interesting. Its income per head rose by just 13 percent in fourteen years. But its social progress score never fell in any year and its rank improved from 16th to 14th. Japan is near the top of the scale, so extra points are arguably harder to earn so this is a very significant achievement. Japan is now five points above the US (having been more or less level with it around 2011) and it has also managed to stay above Korea (which gained 3 points and overtook the US around 2017).

The chart dissects what drove social progress in the countries mentioned before. I disaggregate overall changes in the SPI into a customized set of four components: the first two dimensions of the index (basic needs and foundations of wellbeing) and the third dimension of the index split into two (rights & voice and other opportunity-related components).

A few patterns emerge from this decomposition. The improvements in India and Indonesia leaned more heavily on basic needs than anyone else’s, whereas China’s gains came mostly from the foundations block and Brazil’s almost entirely so. Japan and Korea, already near the ceiling on basic needs, drew their gains from the foundations and opportunity blocks in roughly equal measure. Of concern is that rights and voice fell in every one of the seven countries, except Japan. And the United States lost on three of the four blocks at once.

The next chart goes one level down to the components that moved most in each economy.

Three things stand out:

• Connectivity has been a common source of good. Information and communications is the largest mover in China, Indonesia and Brazil. It is the second largest driver in India.

• India gained on shelter and water, Indonesia on shelter, China on advanced education, while Japan and Korea gained on inclusiveness, freedom, safety and health.

• The US has been the mirror image. Its three largest movers are all declines, led by rights and voice, which is also the biggest drag in India, Indonesia and Korea.

The global picture

The chart below paints the broader picture looking at all 171 countries. And within the chart, I highlight the ten countries that gained most since 2011 and the ten that gained least.

The ten biggest gains appear to be around the middle of the income distribution between roughly 3,000 and 25,000 dollars per person. Their paths run steeply upward. Saudi Arabia is the one rich climber, moving almost vertically.

The weakest ten are, surprisingly, at both ends. Afghanistan, the Central African Republic and Syria slid down and to the left, losing income and social progress together. The four rich countries in the inset moved sideways or down with more income, especially the US.

In what follows, the next two sections look at each group in turn.

Where progress was the greatest

None of the ten fastest climbers is a large economy, and only one is rich.3 Looking at the size of the absolute improvement in the index, Fiji is leading the pack followed by Saudi Arabia, Moldova and the Gambia, And also Armenia, Uzbekistan, Viet Nam, Eswatini, Georgia and Nepal. All managed to boost the SPI by about 10 points. Note that the starting points of these countries differed significantly. For example, Armenia started from 63 points in 2011 whereas Eswatini’s initial value was just 40.

The chart below shows that the gains were broad across the four dimensions identified. Note that the Gambia, Fiji, Uzbekistan and Armenia saw significant gains on rights and voice, which places them in a small global minority. Georgia, Eswatini and Viet Nam show the opposite pattern: large gains on material components coupled with a small loss on rights.

At the component level, the 10 fastest climbers look remarkably alike. Information and communications is the largest single mover in nine of the ten (only in the Gambia is it displaced, by rights and voice, which rises by +28). We saw improvements in advanced education in Fiji, Saudi Arabia, Uzbekistan, Viet Nam and Georgia. We saw good progress on shelter in Viet Nam, Eswatini, Georgia and Nepal. And basic knowledge improved in Moldova and the Gambia. As noted, rights and voice was among the top three improvements in Fiji, the Gambia, Uzbekistan and Armenia.

Where progress was the weakest

The bottom ten splits into two groups that have almost nothing in common except for weak progress in the SPI. The first is a group of countries that have suffered instability and conflict. This includes Venezuela, Syria, Afghanistan, South Sudan, the Central African Republic and Lebanon. The second group is rich and at peace. This includes the US and Canada which saw a decline as well as the United Kingdom and New Zealand which barely moved (they added 0.2 and 0.5 points in fourteen years while the world average rose by nearly five). Of course, we need to caveat these results as making progress closer the top is more difficult as some variables in the SPI are hitting their saturation points.

One component is red in all ten bars: rights and voice fell in every country on this list. For Venezuela and Afghanistan it is the single largest item of change. In Syria the damage is in basic needs, which reflects the direct footprint of war. In Afghanistan basic needs actually improved, but rights and opportunity collapsed. The rich countries saw losses in basic needs and rights and voice. Foundations of wellbeing rose in all four, and other opportunity rose in three.

Looking more deeply into the components, rights and voice is the largest single loss in seven of the ten countries. In Syria, the biggest losses are inclusiveness and basic knowledge, in South Sudan personal safety, in New Zealand shelter. Even here the material components mostly kept rising: connectivity improved in nine of the ten, Afghanistan being the exception.

Three takeaways

I would like to make three final points.

Humanity got more connected but became less free

Between 2011 and 2025, scores for rights and voice fell in 121 of the 171 countries covered by the SPI. In 61 countries, the decline exceeded five points. The global score for this component dropped from 53 to 47, even as the overall index rose from 59 to 64. Over the same period, the connectivity score climbed from 49 to 69.

In other words, humanity got online and lost its voice at roughly the same pace.

This contrast speaks to a broader tension between the material and civic components of the index, which each respond to different things. The material components, such as housing, water, schooling, connectivity and health, are more responsive to economic growth and investment. But the civic components depend on political institutions, accountability, civil liberties and the ability of people to shape the decisions that affect them.

All countries are developing countries

The SPI reveals enormous variation in both levels and rates of social progress across countries at different income levels. One thing that is abundantly clear is that rich countries have no monopoly on progress. High and very-high income countries need to actively protect past gains, address emerging weaknesses and keep pace with changing expectations and possibilities.

Some countries make rapid advances. Others stagnate or suffer outright setbacks. Even countries whose scores improve can fall behind if others move faster.

This connects to a point I often make: all countries are developing countries. Every country faces a moving frontier. Every country must work to preserve past gains while pushing towards new ones. No country has arrived.

Non-monetary dimensions are the linchpin of development

Over recent decades, global development has delivered substantial gains in income and monetary poverty, but progress across non-monetary dimensions has been slower and more uneven. See my recent article summarizing how to see development through that lens.

The Millennium Development Goals offered an early illustration. The world achieved the target of halving the global extreme poverty rate five years ahead of the 2015 deadline. Performance across non-monetary goals, covering health, education, water and sanitation, was much more mixed. The Sustainable Development Goal era risks repeating the pattern: according to the latest UN assessment, fewer than one-fifth of SDG targets are on track for achievement by 2030.

We need to move beyond the mindset that growth alone defines development. Growth has been shown to be essential for raising material living standards, and some would argue it is even sufficient. But material living standards capture only a part of what ultimately matters, which is a social measure of wellbeing.

This is why the SPI which focuses purely on the non-monetary dimensions of development is so valuable. It helps us see how societies are really doing. And it helps us pinpoint where we need to do better.

The world today counts 1,198 million people in multidimensional poverty across 108 developing countries. This is 18.9 percent of their population and half of them are…. children.

The MPI explores what they lack: whether a child in the household is out of school, whether anyone is undernourished, whether the family cooks on wood or dung, has a toilet, has electricity, has lost a child. A person is poor if deprived in at least a third of ten weighted indicators. And he or she is deemed severely poor if deprived in half or more of these indicators. That makes it a different lens from the World Bank’s $3.00 international monetary poverty line, and often a harsher one.

The MPI update is released in the same week as the IMF-World Bank Annual Meetings in Bangkok, where I am writing from. I noticed that the programme includes just one event with poverty in its title and wrote about that silence separately. Suffice it to say here that the update is timely since the global picture is concerning: global poverty has been rising, so the last thing we need is for the international community to place poverty on the back burner.

1. Half the world’s poor live in five countries

Multidimensional poverty is highly concentrated. India, Nigeria, Pakistan, Ethiopia and DR Congo together hold 612 million of the 1.2 billion, just over half. Add the next seven and we have the result that merely twelve countries account for 69 percent of global poverty. India alone has 238 million multidimensionally poor people, more than the next two countries combined.

In thirteen countries, over half of the population live in poverty. In Chad, 84 percent of people are multidimensionally poor and 65 percent are severely multidimensionally poor, which means they are deprived in half or more of the indicators at once. In the Central African Republic and Niger, four in five people are poor and more than half severely so. All thirteen are in Sub-Saharan Africa, except for Afghanistan.

The headcount understates how many people live within close reach of poverty. A further 952 million are vulnerable, with deprivation scores between 20 and 33 percent. This means they are but a shock away from crossing the line. Poor and vulnerable together come to 2.15 billion people, which corresponds to a third of the population of the developing world. Note that South Asia has more people on the brink than Sub-Saharan Africa, but it is Africa that has far more who are already severely poor.

2. Rates fall, numbers rise

The global increase is not exactly a story of countries going backwards. Of the 89 countries with two comparable surveys, only three saw their poverty rate rise between their first and latest surveys. But in 26 the number of poor people went up anyway, because the total population grew faster than poverty fell. These are countries where the transition to lower fertility rates has been slow. For example, DR Congo cut its poverty rate from 77 to 62 percent but still added 20 million poor people. Nigeria added 13 million, Mozambique, Yemen, Chad and Mali added between 3 and 4 million each. Afghanistan is the one exception where both the rate and the number rose in tandem over the whole period.

The MPI update also shows us the hopeful picture of how fast poverty can fall when it actually does. Thirty-four of the 89 countries with comparable surveys halved their MPI between surveys. Timor-Leste cut its poverty rate by 3.5 percentage points a year, Sierra Leone and Congo by 3.1. India’s fall from 55 to 16 percent over fifteen years took roughly 400 million people out of multidimensional poverty — and this is the largest reduction on record.

3. What poverty looks like

What I particularly like about the MPI is that it is an index that you can take apart and examine what aspects are actually contributing the most to overall deprivation. The chart below shows what this looks like.

One further interesting result is that where poverty is the widest it is also likely to be the deepest. The chart below plots each country’s poverty rate against the average deprivation score of its poor. A clear diagonal pattern emerges. In countries where under a tenth of people are poor, the poor are deprived in about 35 to 40 percent of the indicators. In Chad, Niger and the Central African Republic, where most people are poor, the poor are deprived in 56 to 62 percent. This is well past the severe threshold. This suggests that the poorest countries are both poor more often and poor more intensely.

There are significant differences across regions. In Sub-Saharan Africa, living standards account for 48 percent of the index: fuel, sanitation, housing, electricity. In South Asia, the three dimensions are more in balance, even though health weighs more. In Latin America, health deprivations are the largest share of the index.

4. What these data don’t show

It is important to understand the limitations to these data as the 1.2 billion global figure is in fact a stitched-together number. Each country’s headcount comes from its latest survey and is applied to the 2024 population. Only 13 of the 108 countries have a new survey in this year’s update; for the other 95, the count changed because their populations grew (which by itself added 16 million poor people). For 38 percent of the poor, that survey was taken before 2020. For example, China’s 55 million number comes from a 2014 survey. Pakistan goes back to 2017/18 and Ethiopia to 2019, which means that the effects of the pandemic, the conflict in Ethiopia and Sudan, and the floods in Pakistan are not captured.

It is also important to recognize that the MPI captures different aspects of the lived realities facing the poor than what other monetary metrics of poverty show. The chart below lines up the MPI figure with the national poverty line of countries as well as the international extreme poverty line of the World Bank. As you can see, the differences can be huge.

Three takeaways

From the above, three takeaways emerge: go beyond jobs, don’t confuse percentages with absolute numbers, and gather more and better data.

Go beyond jobs

Jobs, especially good jobs, are a sure way out of poverty and growth is instrumental to generate jobs. But the relationship between growth and employment is not automatic. Too often, growth is jobless, or generates too few jobs to absorb a growing workforce. And when jobs are created, they are often informal, concentrated in small-scale, low-productivity activities that offer limited opportunities for upgrading, rising productivity and higher wages.

There is a second reason why a jobs-centred approach is insufficient. Half of all people living in multidimensional poverty are children. And there are also the elderly as well as others who cannot work and will depend on other types of support to relieve their deprivations. Jobs may matter indirectly here, but again they would do so imperfectly.

The bottom line here is that we need to broaden the lens. Job creation must remain central to poverty reduction. But let us complement it by greater investments in human development, basic services and social protection. The objective goes beyond putting people to work. It is also to ensure that people have the capabilities, opportunities and security to escape poverty sustainably.

Don’t confuse percentages with absolute numbers

The geography of multidimensional poverty is shifting towards Sub-Saharan Africa, which now accounts for an estimated 616 million people living in poverty. The region is also home to almost all countries where multidimensional poverty affects a majority of the population. Rapid demographic growth, combined with insufficient progress in reducing poverty, is driving up the absolute numbers.

This alone makes a compelling case for greater attention to poverty reduction in Africa. But it does not, in any way, justify diminishing attention to other regions where poverty remains widespread.

I have seen the same elementary mistake repeated many times. One example is where a map of global poverty rates is presented, where much of Africa is invariably shaded dark red and much of Asia light green. This conveys the message that poverty is an African problem, while Asia has largely moved on.

But we should not confuse percentages with absolute numbers. Progress in bringing down rates does not necessarily mean that the absolute scale of the problem has disappeared. South Asia and East Asia together still hold half a billion multidimensionally poor people, and 660 million more who are vulnerable. On top of that, even if the scale were small, we would still have the moral duty to support those who remain deprived.

In other words, our focus should be on the poor, wherever they are. Leave no one behind.

Gather more and better data

Our global headcount of multidimensional poverty relies on household surveys, many of which capture conditions from well before the pandemic. The surveys are done at different times and the data we have often lag current conditions by years.

Without timely data, we are acting in the dark. We don’t see where poverty is most concentrated, where progress is made or regressions occur, and which groups are most affected.

We thus need more frequent surveys. We also need better coverage of vulnerable and hard-to-reach populations. And we need stronger systems for tracking changes between survey rounds. In the meantime, we can use administrative records and other data sources to complement household surveys where possible.

Twenty-seven years later, they do. But the question has returned in a more threatening form, this time against the backdrop of the rapid roll-out of artificial intelligence. Will AI make central banks obsolete, ineffective, or both, as new financial instruments emerge that are too complex for anyone to understand or control?

I will argue that central banks will not vanish, because they produce the one thing AI cannot: trust.

Trust may sound like an old-fashioned concept in a modern age, but it is what will fundamentally underpin the three circles where AI meets the central bank: the economy the central bank steers, the financial systems it guards, and the institution it is. Preserving trust in these three circles will require central banks to adapt.

In what follows, I develop three points:

• The world seems stuck in a recurring cycle of tech hype and fear. Looking ahead calls for more balance, and for more humility, given the long record of technological mispredictions.

• AI brings new opportunities and new risks to central banking through all three circles. But central banks remain relevant because they generate trust, a scarce asset AI cannot replace. I will refer to three types of trust: credibility, confidence, and legitimacy.

• To preserve trust, however, central banks must themselves adapt. I set out three areas for action: adapting policy frameworks, strengthening oversight capabilities, and balancing financial stability with innovation.

Central Banking in the Age of AI - Philip Schellekens

1.87MB ∙ PDF file

Download

Download

2,500 years of tech fear and hype

Before any forecast about AI, a note of humility. Our record of predicting what new technologies will do is poor and it has been marked by hype and fear in equal measure. It is therefore fair to ask whether today's prophecies about the future of central banking and finance suffer from the same.

From Socrates and Edison

About 2,500 years ago, Socrates feared that writing would weaken memory and leave only the appearance of understanding. In Plato’s Phaedrus, he has the Egyptian king Thamus tell the inventor of letters that his gift will make people trust the marks on the page rather than their own minds, which will have adverse consequences:

“For this invention will produce forgetfulness in the minds of those who learn to use it, because they will not practice their memory. Their trust in writing, produced by external characters which are no part of themselves, will discourage the use of their own memory within them. You have invented an elixir not of memory, but of reminding; and you offer your pupils the appearance of wisdom, not true wisdom, for they will read many things without instruction and will therefore seem to know many things, when they are for the most part ignorant and hard to get along with, since they are not wise, but only appear wise.”

Did Socrates’ prediction hold true? Writing in fact allowed humanity to deepen its understanding and indeed become more wise. And it is incidentally thanks to Plato, who wrote down his teacher's fears using those external characters, that we know of them today. Socrates had a point, since every new technology may provoke a degree of deskilling, but he misjudged the balance. Take it as a case of technological fear “avant la lettre”.

Now take the opposite case of hype rather than fear. In 1913, Thomas Alva Edison told a New York newspaper that books would soon be obsolete in the public schools. Pupils would be taught purely through motion pictures, and the school system would be completely changed within ten years.1

Was Edison right? The books survived well beyond the horizon he had in mind, and so did the classroom, the teacher and the exam. But he did have a point as film and video did find a place in teaching, even though a century on they continue to supplement the textbook and have not replaced it.

Socrates and Edison offer two examples of the ways we can misread a new technology: fear and hype. Both held some truth in ways that still resonate today, but both also misjudged the scale and the timing of technology’s impact on society.

… to King and Harari

The same two reflexes have since reached central banking. Mervyn King’s version came in 1999, in the closing section of his Jackson Hole paper, as a carefully hedged question that nonetheless insinuated that technology could make central banks obsolete. A central bank’s power rests on its monopoly over base money, the asset in which payments are finally settled. If real-time electronic transactions let two parties settle a trade on the spot, with pre-agreed algorithms choosing which assets change hands, final settlement would no longer need a central bank, and “the successors of Bill Gates would have put the successors of Alan Greenspan out of business.”

Twenty-seven years on, the successors of Alan Greenspan are still in business. King had a point, since real-time settlement on programmable ledgers has now arrived, but his own speech explains why it has not displaced central banks. He ends by placing his faith in Walter Bagehot, who wrote in 1873 that if the Bank of England were swept away, "generations must elapse before at all the same trust would be placed in any other equivalent." King still allowed that extinction could not be ruled out. Take it as a case of technological hype (that central banks may vanish), tempered by its own author (that the trust they carry is not easily replaced).

Next is an example of fear applied to the finance. In 2023, Yuval Noah Harari warned that AI could create financial devices that only AI can understand, which would leave a financial system that no human can regulate. He has since sharpened the argument, mentioning that laws and financial contracts are made of words, that AI has mastered words, and that therefore AI will soon be taking over both.

Whether this will happen is unclear. In the run-up to the 2008 crisis, very few people understood back then what was inside a CDO-squared. These complicated financial instruments were built by humans. Imagine what AI could do! But let us remember: a financial instrument exists only because a number of conditions are satisfied: (1) there is a law that recognizes it, (2) there is an institution that agrees to clear and settle it, and (3) there is a central bank that stands behind the money in which it is denominated. In other words, there are still people and institutions behind the creation of such financial instruments, and they can either make those decisions or refuse to make them. Harari concedes as much: at the BIS in 2024 he said that the business of finance is to create trust, and that only human institutions can keep people in the loop.

King and Harari bring hype and fear to central banking just like Edison and Socrates brought them to learning. They describe scenarios that extrapolate certain trends that some might find plausible. They also both refer to trust as the solution to their scenarios. But in the end, central banks have so far survived. They managed to do so despite the arrival of the internet, e-money, mobile money and crypto because every new form of money ended up settling in the one asset everyone trusted.

Three circles where AI meets the central bank

We talked before about the three circles. Most conversations about AI in central banking start with the tools: the chatbot, the nowcasting model, the supervisory dashboard. Those are in the smallest of the three circles: the institution a central bank is. Around it is the financial system it guards. And around that is the economy it steers, a task which AI has made appreciably harder by confounding the signals central banks rely on.

The takeaway here is that AI reaches all three circles of influence at once. And that each of the circles is also a place where trust can be earned or lost. In the economy, it is trust that monetary policy will preserve the value of money, which economists call policy credibility. In the financial system, it is confidence in the safety of money and the reliability and finality of payments. In the institution, it is trust in the authority and judgment of those who decide: legitimacy.

These three types of trust are also nested like the circles themselves: a central bank that is not believed as an institution will have a hard time keeping confidence in the financial system, and without that it will be hard for the central bank to anchor expectations in the economy.

For each circle, I look at what AI is doing there, the opportunity it brings, the risk it carries, and the kind of trust at stake.

The economy a central bank steers

AI has become a macroeconomic force, which has so far become most visible on the demand side of the economy. Global AI-related investment is about half a trillion dollars today, and the industry expects three to four trillion by 2030. Spending on data centers and chip plants has reached 0.8 percent of GDP in the US and over 1 percent in Australia. The BIS estimates that AI investment added about a percentage point to US growth in 2025 and drove upward revisions to 2026 forecasts across many economies.

The opportunity lies on the supply side, which is where the larger prize lies, but also the larger uncertainty. The chart below collects a dozen estimates of the annual productivity gain from AI over a ten-year horizon. Every one of them is positive, so we can say there is a broad “directional consensus” about the impact of AI on aggregate productivity. The magnitudes differ hugely, however, from a few tenths of a percentage point a year to nearly three. The estimates published in 2025 and 2026 are shown in red. Interesting to note that these have moved to the lower end. The BIS now places the median of the literature at about half a point of total factor productivity growth a year. ECB staff puts it at 0.3 to 0.4 points for the euro area. Despite many examples of large gains at the task level, the open question remains how much reaches aggregate productivity, and how fast.

So the demand effects are large and present, but the supply effects are potentially larger but uncertain in size and timing.

What about labor markets? History suggests that new technology has never been a source of unemployment in the long run and in the aggregate. But in the near and medium term it has, and underneath the aggregate it has. Whether AI is different is, so far, hard to tell.

What is clear is that labour markets are feeling three forces at once:

• Substitution: Among the economies the BIS tracks, those in the more AI-prepared half saw unemployment rise by about three quarters of a percentage point between 2023 and 2025, while it stayed flat in the other half. In the United States, employment in the sectors most exposed to AI grew about 0.8 percentage points less than in the least exposed over the same period. And almost 80 percent of listed firms are currently signaling plans to automate in their earnings calls (BIS Bulletin No 130, July 2026).

• Augmentation: Field experiments in customer support, writing, consulting and software development suggest that AI may produce time savings of 20 to 50 percent by taking over parts of non-routine cognitive tasks. It also appears to be the case that the least experienced workers gain most, with AI narrowing the gap between novices and veterans. Whether these gains reach the aggregate is not clear yet. It depends on how firms reorganize themselves around the technology. In time, however, AI may lead to vastly cheaper goods and services that may in turn lead to a rise in real incomes.

• Innovation: The table below illustrates that new technologies also create new types of jobs that no one would have anticipated in advance. In fact, 60 percent of today's US occupations did not exist in 1940. The job title “airplane designer” entered the Census index in 1950 and “computer application engineer” in 1970. “Artificial intelligence specialist” already entered in 2000, whereas “cybersecurity analyst” entered in 2018. The list of new professions will go on as AI continues to develop and roll out.

The risk for labor markets is mainly one of timing. The substitution mechanism may show up first while innovation will have its effect much later. Therefore, it is likely that for a while labor markets see job losses in exposed sectors that co-exist with job shortages in other sectors. Adjusting policy rates will have little effect on this structural shift and will remain an overall blunt tool if the labor market is restructuring rather than just slowing — a point made by Governor Lisa Cook of the Federal Reserve in February.

What is at stake here is a central bank’s policy credibility. Policy credibility is what keeps inflation expectations anchored, which in turn means that firms and workers do not engage in an inflationary wage-price spiral because they expect the central bank to ensure that inflation deviations from target are temporary aberrations. The benefit of credibility here is that it will take smaller moves in interest rates and less lost output to get inflation back to target. Another benefit is that it buys a central bank time and provides it with the luxury of freedom to wait for evidence instead of having to react to every single surprise.

AI is entering our world at a time where central bank credibility has been heavily tested. Isabel Schnabel of the ECB noted that the inflation surge from 2021 to 2023 has amplified institutional distrust. This implies that another misjudgement would likely come with greater costs than the first because policy credibility is weaker to begin with.

The financial system a central bank guards

A key task for a central bank is to protect confidence in the financial system. This amounts to making sure that people believe that their money in the bank is safe and also that payments once made can be trusted to reach their final destination.

Today we can see that AI has already moved into the core of the financial system, with widespread applications in trading, credit risk scoring, and scam detection. Trading of course had been automated long before AI, with about 70 percent of US equities volume executed by algorithms. But that also meant that the introduction of AI has been all the easier. The same applies to credit risk scoring and scam detection where many applications are being rolled out.

This rapid move of AI into the financial system presents two types of opportunity: to reach people the system could not see, and to better protect the ones it already serves. On inclusion, the beneficiaries may be those the financial system was not able to assess before as there was no history on them. When risk is priced on information rather than on a borrower’s record, the premium on small firms and new borrowers falls. On resilience, many credit card issuers are now able to detect scams more easily. AI systems can now learn across a whole network rather than with respect to just one account at the time.

The risks presented by AI run through the balance sheets of financial systems as well as through their security.

• As AI firms’ capital spending outran their cash flow, financing shifted to debt. The chart below shows two channels: bond issuance by AI firms tripled from 79 billion dollars in 2023 to 243 billion in 2025, and private credit to AI-related firms went from near zero in 2016 to 200 billion dollars, reaching about eight percent of all private credit. The market value of the firms at the centre of the AI boom has skyrocketed. By Goldman Sachs’ count, AI-linked firms were about 25 percent of the S&P 500 when ChatGPT launched in late 2022 and nearly 45 percent by mid-2026.

• Cyber risks have increased. A cyber attack with a frontier model, from finding a weakness to extracting value, now costs five to ten thousand dollars. Clearly, frontier models strengthen our defenses too. But as the BIS puts it, the costs are unfortunately asymmetric favoring the attackers: an attacker needs to spot and exploit just one weakness, whereas a defender must protect the whole system.

One can therefore say that the financial system is exposed to AI twice: in what its institutions do with the technology and in what they lend to and invest in. As a result, confidence is exposed twice as well. People will lose trust in AI when a scam succeeds or an automated decision is unreasonable and poorly explained. People will also lose trust when AI valuations turn out to be falsified either because returns are less than expected or arrive later than expected.

The institution a central bank is

Finally, the third circle: the institution of the central bank itself. In a recent survey by the SEACEN Centre, 17 of the 19 responding central banks or monetary authorities in Asia-Pacific are already using AI. The leading use is economic analysis and forecasting, at 95 percent. Supervision, anomaly detection and risk assessment are around 60 percent. Payment systems seem to be lagging at under half.

There is enormous opportunity for AI to contribute to efficiency in central banking. Almost everything a central bank produces — think of forecasts, statistics, supervisory assessments and policy communication — is made out of language and data, where AI can make that work cheaper and faster.

An example from BIS research: a generic language model asked to read central bank sentences and classify the policy stance conveyed by the text as hawkish, dovish or neutral gets it right about 55 percent of the time. This is just a little better than a coin toss. But when the same model was retrained on central bank speeches and papers, it would get it right at about 90 percent of the time. That is useful for tracking how expectations about the policy stance are formed and allows the central bank to also explain itself better, thus contributing to both credibility and legitimacy.

The table below reports what worries central banks among SEACEN member countries most. Note how highly information security, third-party dependence, and model opacity rank. Talent shortages are also mentioned which echoes the wider finding that nine in ten central banks globally report that recruiting for cyber security, IT, fintech and AI roles has become harder.

Dependence on outside providers is a significant issue as supply is highly concentrated in a few hands. When it comes to chips, NVIDIA holds about 80 percent of accelerator revenue in 2026 and 9 percent of the S&P 500. In cloud infrastructure, three firms (Amazon, Microsoft and Google) held 63 percent of a market that was worth 143 billion dollars in the second quarter of this year alone. The five hyperscalers accounted for a third of all capital expenditure in the S&P 500 last year. They are on course for about 800 billion dollars this year.

Legitimacy is what is at stake here. Central bankers are unelected. And their institution holds powers on the understanding that their leadership and staff are competent technocrats, who are instrument independent. But they remain goal dependent and should be able to explain and answer for their decisions.

Three areas for central banks to adapt

Because AI cannot produce trust, central banks are unlikely to vanish. But it is also why they must adapt. The three circles point to three areas for action to preserve trust by: adapting policy frameworks to protect credibility; strengthening oversight capabilities to protect legitimacy; and preserving financial stability while supporting innovation to protect confidence.

Adapting policy frameworks

AI does not alter the mandate of central banks. But it does alter the economy quite significantly that central bankers are expected to interpret. AI affects demand and supply simultaneously but it does so in different ways. The demand-side effects are clearly visible in the investment boom, the wealth effects and the AI-related exports, whereas the supply-side ones remain more murky as they are subject to considerable uncertainty about the magnitude and timing of productivity impacts.

This generates several ambiguities. For example, does strong activity reflect overheating or higher productivity? Does a weaker labour market reflect slack or restructuring? And what about the level of the neutral rate of interest? There is a risk of miscalibration from both sides. Policy could be too loose if the supply-side gains are overestimated; and it could be too tight if the demand impulse is underestimated.

The uncertainties become especially large as we consider the longer term. The chart below is a stylized illustration of the implications of the four BIS scenarios for the neutral rate, ranging from business as usual, to a moderately higher level as productivity rises but remains bounded, a much higher level if AI turns out to be truly transformative, but also a lower level if AI produced demand-side bottlenecks, such as where automation diverts income away from workers who are also consumers. It is hard to determine in real time which of these paths the economy is on.

How can central banks adapt? In several ways.

First, central banks could monitor more closely cost indicators that are specific to AI. This would include such things as prices of chips and equipment, data-centre construction costs, and electricity and water tariffs. These variables may be the new leading indicators of inflationary pressure.2

Second, they may consider communicating estimates of the neutral rate and potential output as ranges. It is best not to convey a false sense of precision by listing just point estimates. This also serves to protect credibility should there be a need to revise the central estimate. Along the same vein, greater use could be made of scenarios which could each have a separate path for productivity, the neutral rate and inflation.

Third, by adopting a gradual, data-dependent approach. It would be important in this time of uncertainty to proceed in smaller steps, which are then conditioned on evidence that is observed as opposed to just anticipated. Alternative data can be an important input to policy, such as payments data for real-time spending, web data for prices and vacancies, and satellite data in instances where official statistics are sparse.3

Strengthening oversight capabilities

The second area is oversight capability. This pertains to the central bank's own use of AI as well as its supervision of AI in the firms that it oversees.

Inside the central bank. Leading central banks already apply AI in their own work. Examples include BOKI by the Bank of Korea (a sovereign language model run internally), BISTRO by the BIS (an open-source forecasting model that performs particularly well for emerging economies), and the ECB's Athena and the Federal Reserve's LEX (which use AI to process every single supervisory filing).4

It would be important for central banks to make sure they are ahead of the curve. This would require giving staff access without delay to approved AI tools. A small on-premise model could be used for sensitive work, so that confidential data remains confined to the institution.5 Central banks can also consider adopting existing tools (such as BISTRO) before developing new ones by themselves. Throughout all of the above, it remains important that all significant AI-assisted decision have a named owner. Human oversight needs to be built into the process by design.

Towards supervised firms. Central banks in 2026 appear to prefer applying existing frameworks in a technology-neutral way, and then supplement this with guidance. Few central banks are seen to be introducing separate AI-specific legislation.6

Rather than waiting for a dedicated regime, central banks could instead make sure to apply existing model-risk, outsourcing and operational-resilience rules to AI. And to assess concentration risk, all supervised entities could be asked where AI is used and what providers they depend on. Where capacities are limited at the national level, regional bodies could come in. They could make available shared libraries of supervisory tools. A common taxonomy could be developed for AI incidents. And procurement of models and data could be undertaken jointly.

Pooling matters all the more in this current period of receding multilateralism. There is a great deal that can be achieved at the regional level without waiting for global action.

Preserving stability while supporting innovation

The third area concerns confidence in a financial ecosystem that is increasingly digital and interconnected. Two areas matter greatly here: linked payment systems and common providers.

Linked payment systems. National instant-payment systems are increasingly being linked to one another, which allows for a customer in one country to pay a merchant in another in seconds. Such links depend on a high degree of mutual trust: each central bank must be confident that payments cleared through another jurisdiction are sound. But greater connectivity also creates a shared vulnerability. A fraud ring in one jurisdiction can reach customers across several others with much greater ease.

The same connectivity, though, also creates opportunity for stronger collective safeguards. The BIS Innovation Hub’s Project Aurora tested anti-money-laundering detection on payment data. It found that a rule-based system of the kind most banks use today detected 17 percent of money launderers. But once a machine-learning model was applied to data pooled across borders, it was far more effective. It detected 79 percent, with far fewer false alarms.

There is the opportunity for central banks to more quickly share fraud intelligence across borders. They could also consider piloting privacy-preserving data pooling among those central banks willing to partner. In addition, joint stress tests and cyber exercises could be conducted across connected systems.

Common providers. Risks become systemic when they become correlated. A key issue here is commonality, such as when banks make model errors and are all using the same model from the same provider. Supervisors need to be able to document such risks better but this requires that banks report which providers they depend on. They could also force firms to switch. And they could rehearse the failure of a large provider.

Agents, stablecoins and tokenised money

Three recent developments bring King’s earlier questions into sharper focus.

The first is agentic commerce. AI agents are beginning to transact on behalf of people and firms. They can search, negotiate, pay and settle without human intervention at every step. This introduces a new form of financial risk. Agents could move deposits and execute payments at machine speed and a bank run could unfold now within minutes.

The second is stablecoins, privately issued tokens pegged to a currency. Stablecoins are one candidate for the settlement asset agents might use. Yet, as the BIS argued in its 2026 Annual Economic Report, stablecoins fall short on three properties of money: singleness, because different stablecoins can trade at different prices rather than at par; elasticity, because their supply does not expand on demand in the way bank credit does within a central bank-backed monetary system; and integrity, given their use in illicit financial flows.

The third is tokenised money and assets. Deposits, securities and reserves recorded on programmable ledgers allow settlement and conditional transfers to occur within a single transaction. This comes close to the architecture King envisaged: money and transactions operating within a digital system rather than through separate layers.

For central banks, the key question is where the regulatory perimeter is. AI agents will need a settlement asset they can trust. If that asset is a private stablecoin, the unit of account and the ultimate backstop are partly outside the central bank’s perimeter. If instead settlement takes place in tokenised central bank money, whether through tokenised reserves, wholesale central bank digital currency or tokenised deposits settled in central bank money, innovation stays anchored within the existing monetary framework. The most trusted asset remains at the centre of the system.

In this context, central banks could offer tokenised reserves or other forms of wholesale central bank money for tokenised markets, giving agents a trusted settlement asset. They could set clear conditions for stablecoins. Regulation should establish requirements for reserves, redemption and governance so that stablecoins in circulation remain reliably redeemable at par. They could require interoperability with public payment systems. Private innovation should connect to public infrastructure rather than develop as a parallel monetary system. And they could establish accountability for AI agents. Rules should specify who is responsible when an agent acts on behalf of a person or firm, ensuring every payment remains attributable to an identifiable principal.

Trust, an old-fashioned concept in a modern time

Trust is the one thing central banks provide that AI cannot. This is why they are unlikely to disappear.

When King asked in 1999 whether technology might render central banks unnecessary, he turned to Bagehot’s observation that trust in such an institution takes generations to build and generations to rebuild. The observation remains relevant. Technology changes how money moves, but not the institutional foundations on which confidence in money rests.

The implication for central banks is to adopt new technology with assurance, keep accountability firmly in human hands, and cooperate across borders. In an interconnected financial system, trust is shared. The system is only as strong as the weakest link.

A bigger base, a slower rate, a larger addition

One Korea every other year

China’s annual increase in GDP, measured in dollars as a ratio to Korea’s GDP of the same year, has averaged about half a Korea a year since 2005. But the series is quite volatile because it reflects also price and exchange rate fluctuations.

During the post-pandemic rebound in 2021, the ratio of China’s GDP increment to Korea’s GDP was a whopping 165 percent. Later, in 2023, this figure became close to 0. By 2025, it had returned to 41 percent. On the IMF’s April 2026 projections (which will be updated very soon), Chinese growth is expected to ease to 3.3 percent by 2031. But despite slower growth, this would mean that China still adds 60 percent of a Korea a year, in other words about one Korea every other year.

Let’s look at these same data through the lens of five-year windows to smooth the noise. Measured relative to Korea’s economy at the end of each window, China added between 2 and 3 Koreas during every five-year period after 2006. On the IMF’s projected path, it would add 2.8 Koreas into the next five-year period over the projection horizon. Once again, China’s additions remain very substantial despite slower growth.

Eleven Koreas since 2000

We can also look at this cumulatively by considering China’s expansion since 2000 up until a given year T. For example for T=2025, growth since 2000 produced 9.8 Koreas, whereas for T=2030 this is expected to become 10.9. The slope was steep in the 2000s, but it became mostly flat in 2022 to 2024. If the IMF has it right, the slope will resume previous patterns of a steeper slope over the forecast horizon.

Four to seven percent of world GDP every five years

The earlier comparisons take Korea as the denominator which helps frame the additions but has the disadvantage of being subject to idiosyncratic variations specific to Korea.2 It is therefore useful to also look at China’s additions relative to a broader aggregate. So let’s just look at China relative to the world.

Against the world rather than Korea, the same increments correspond to between 4 and 7 percent of world GDP per five-year window since 2006. It would be 4.2 percent for 2026 to 2030.

Real, nominal and the wedge between them

Three measures

Since 2010, China’s real GDP has grown 2.5 times, its nominal GDP in yuan 3.3 times, and its GDP in dollars 3.1 times. The three measures (real, nominal in yuan, and nominal in dollars) are quite in sync until 2021. Then the dollar line goes flat for two years (0.4 percent) while real output rises 8.7 percent.

A decomposition of dollar growth

Dollar growth is real growth plus the change in the price level and the change in the exchange rate. In the late 2000s, the second and third added nine points a year to real growth of 10.7 percent. But in 2023 to 2025, they took three points away from a real growth rate of 5.0 percent: 0.7 from falling prices and 2.2 from a weaker renminbi. Real growth barely changed between 2016 and 2025 according to the published statistics, but the dollar number swung from 5.7 to 10 percent and then to just 2.1 percent.

The chart below shows the evolution of the wedge in yet another way. Nominal GDP per capita in yuan has grown more slowly than real GDP per capita since 2023. This is the first sustained stretch that this has happened since the late 1990s. The shading between the two lines captures movements in the deflator: for two decades we had a positive wedge and for three years a negative one.

Why the deflator is rising again

In the second quarter of 2026, the GDP deflator started rising again after twelve quarters of decline. This was mainly the result of developments in imported commodity prices (especially energy and non-ferrous metals) due to the Middle East conflict. Much less so did it reflect an inflationary domestic demand environment. Consumer price inflation has remained under 1 percent and consumer-goods producer prices are still falling.

Manufacturing prices played a key role in the reversal as the chart below shows. Note that China’s statisticians deflate industry value added with output prices alone. so when input prices rise faster than output prices (as with the Middle East crisis), this method tends to overstate value-added inflation and understate real manufacturing growth. Note as well that this is exactly the mirror image of 2023.

What China’s additions mean for everyone else

Overcapacity is a savings-investment story

Overcapacity reflects fundamentally a savings-investment imbalance. China saves more than 40 percent of GDP and invests almost as much. Household consumption is under 40 percent of GDP, which is 15 to 20 points below what is typical in high-income economies.

Is the high-income economy comparison fair? China is expected to reach high-income status this year, so one could argue that the comparison is relevant. At the same time, China will be at the bottom of that group, with much room left to grow and develop. Hence, not a totally fair comparison.

Of course, China’s imbalances have drawn considerable attention given their size and implications for the rest of the world. The surplus of production over domestic spending is about 4 percent of GDP. The goods surplus is above a trillion dollars.

As the two charts below show, the imbalance reflects as much the rest of the world as China. China’s surplus was 0.6 percent of world GDP in 2025, which was close to its 2008 peak. The US deficit has hovered around 1 percent of world GDP since 2000. Japan, Korea, Taiwan, Singapore and Hong Kong together have a surplus similar to China’s. The rest of the world spends more than it produces and China produces more than it spends. The IMF expects the pattern to persist over the forecast horizon.

China’s side: consume more, invest better

Consumption growth in China has been lackluster. Retail sales grew just 1.3 percent in the first half of this year. The 15th Five-Year Plan’s target of 60 trillion yuan by 2030 implies nominal retail growth of about 3.7 percent a year.

Some rebalancing has however already happened. Total consumption (both private and public) has risen from 50 percent of GDP in 2010 to 57 percent in 2025, while investment has fallen from 46 to 39 percent.

Raising the consumption share further will require transferring income to households (such as through social security, healthcare and pensions) and letting the price level normalize so that nominal incomes grow with output.

It of course does not mean that investment should stop. China’s capital per worker is still less than half the US level. With the population falling since 2022, future growth will have to come from additional capital deepening and productivity. The gap is closing but remains considerable as the chart below shows, as measured at purchasing power parity. At market exchange rates, the gap would be wider still.

Looking ahead, it would be important that investment proceed with respectable long-term return in sectors where China does not already hold a third of world capacity. Investment that raises productivity is needed to finance the pensions of a shrinking workforce. Conversely, if investment is primarily directed to where overcapacity already exists, it will simply add to deflationary pressure.

The rest of the world’s side: capabilities, competition, ecosystems

For other countries, there is tremendous opportunity to tap into the additional Chinese income that is generated. Chinese growth will reflect in part a growing and deepening middle class. This generates possibilities for overseas exporters in terms of, eg, food, differentiated goods and services.

So far, Asia has found it hard to take advantage of the opportunity. As the chart below shows, Japan, Korea, Taiwan, ASEAN, India and Australia have run a surplus against China, as a group. But since 2023, this is no longer the case.

The chart below looks at the trends country by country. By December 2025, exports to China were below their 2021 level for Australia, India, Japan and Korea. Only ASEAN has been holding steady. The rebound that we saw in the first half of 2026 was mainly related to Korea and Taiwan selling chips into the AI boom, while Australia and Indonesia benefited from higher prices for commodity exports to China.

Boosting exports to China requires three things on the part of other nations:

• Upgrading capabilities is the starting point. In order to adopt a stronger outward orientation, countries will need access to talent, technology, finance and a solid infrastructure. These are the bread and butter foundations for being able to offer products that satisfy evolving Chinese demand.

• Exposing enterprises to greater forces of competition, internally and externally, is also important. Competition is the driving force of innovation. Firms that are sheltered at home will face greater difficulties in selling abroad competitively. Protection keeps costs high and quality low. It also removes the pressure to improve. The same goes for making sure home markets are contestable to foreign investment.

• Building ecosystems is the next important aspect so that a larger domestic footprint can be developed. This can be summarized in the need of “connecting globally, while developing locally”, so that economies of scale and scope can emerge in the context of clusters — much along the lines of how China developed its own industrial success.

In short, China’s continued expansion will present opportunities for all but tapping into the “next Koreas” will require other countries to also do their part and join the global race to improve competitiveness more actively and more effectively. There is no other way around it.

This is one of my favorite quotes. Its point is that talent is distributed much more widely than achievement, and that the gap between the two represents an enormous social cost in wasted potential and missed opportunity. And the cost is global: an Einstein lost in one country is an Einstein lost to the world.

Top performers in PISA 2025

I was reminded of Gould’s quote while going through the newly released results of PISA 2025, the OECD's Programme for International Student Assessment. PISA tests 15-year-olds on reading, mathematics and science, and among other things it measures how many students perform at the highest levels. In a sense, PISA gives us a “census of discovered talent.”

The comparison is admittedly imperfect. PISA tests only a narrow set of subjects, and its top performers are unlikely to be at the Einstein end of the spectrum. But I think Gould’s point applies more broadly to talent is never discovered, never nurtured, or never put to use. The PISA lens is a good starting point for a discussion around it.

PISA also doesn’t offer a truly global picture, even though it is the best we have. Ninety-one education systems participated in the 2025 exercise (see table above). This included 82 countries and nine sub-national or other economies. In total, more than 760,000 students sat the test. Each of them stood in for many others through the sampling design of their education system, so the results are weighted up to the population they represent: some 33 million 15-year-olds or about a quarter of all 15-year-olds worldwide.

PISA estimates that about 1.9 million of those 33 million perform at the top level in science, which is around one in eighteen. Their distribution, however, is highly uneven. The chart below shows two ways of looking at this uneven distribution.

Four Chinese provinces (Beijing, Shanghai, Jiangsu and Zhejiang), which are home to about a seventh of China's population, account for roughly half a million of them. This is more than the entire United States. Note also that Singapore has one of the highest shares of top performers anywhere, but with a small population that amounts to only about 11,000 students. At the other end, Indonesia, the Philippines and Malaysia have far larger cohorts of 15-year-olds and yet contribute very few to the global pool.

If Gould is right as my gut tells me he is, then the near-absence of top performers in some countries tells us VERY little about the amount of talent within their borders. Can we accept that nature hands out talent in these proportions? It would seem highly unlikely.

Undiscovered

The first source of loss is talent that is never discovered.

PISA tests only 15-year-olds who are in school. This is 88 percent of the age group on the OECD average, but only 43 percent in Cambodia. And many countries are not in the survey at all. India, by far the largest omission, is an obvious example.

Let us now look beyond PISA and widen the frame. Gust, Hanushek and Woessmann (2024) combine PISA with regional assessments and impute skills for children out of school, covering 159 countries and 98 percent of the world’s population. They estimate that the countries without comparable learning data are home to about half the world's youth. The bars in the above chart show the share of children who fail to reach basic skills (defined as PISA Level 1 or below) and the diamonds show the share not enrolled in secondary school at all.

The chart shows that worldwide, two-thirds of youth never reach basic skills. Also, a third of them never do so because they are not in secondary school. Look also the large disparities across regions. In Sub-Saharan Africa the figures are 94 percent and about two-thirds; in South Asia, 89 percent and about 40 percent. East Asia and the Pacific stands at 35 percent and looks good only by comparison.

Then there are the many children who are not in school because they are at work. In 2024, some 138 million children were in child labour. And three-fifths of them were in agriculture. Gould’s cotton fields are, unfortunately, quite literal.

Even inside school, talent is often not discovered.

Gifted programmes in most places rely on parents and teachers to nominate children for testing, but nomination has its biases. One could follow instead take the alternative approach of simply testing every second-grader. That is what a large Florida school district did for its second-graders. Card and Giuliano (2016) tracked the results of this exercise. They found that the share of Hispanic children identified as gifted tripled from 2 percent to 6 percent, whereas the share of Black children from 1 percent to 3 percent.

PISA offers a national-scale version of the same experiment. Take Cambodia. Over the past decade, the country brought many more 15-year-olds into school. This led to a boost in PISA coverage by more than ten percentage points. While newly enrolled students have tended to be weaker on average, Cambodia's mean scores rose in all three subjects as did the scores of its top quarter. In other words, by widening the search to include more of the 15-year-old population, Cambodia found more talent.

The policy implication is that governments can discover more talent when they make sure children are kept in school through their teens, and beyond. In poorer countries, this means enforcing child labour laws and paying families to keep children enrolled. It is also important to measure learning for everyone rather than for just the few who get nominated. Joining PISA or a regional assessment is a cheap and easy way to start. Yet, half the world's youth live in countries that have not yet done so. And within schools, universal screening costs only little more than the tests themselves.

Undernurtured

Being in school is not the same as being taught. And this refers to the second source of loss: talent that is not developed.

Across the OECD, advantaged 15-year-olds scored some 85 points above disadvantaged ones in science. Since the OECD considers a year of learning equivalent to 20 points, this is about 4 years of schooling. The chart breaks it down further in terms of material deprivation (a lack of basic goods at home), social deprivation (a lack of support and connections) and immigrant background. Unsurprisingly, the gaps compound at the bottom, with some 37 percent of disadvantaged students being low performers in science compared to just 12 percent of advantaged ones.

Beliefs reinforce the gaps. PISA asked this year whether students think their status in society can change. It turned out that disadvantaged students were 12 percentage points less likely to say yes. In Thailand, the Philippines and Indonesia, fewer than 40 percent of all students think it can.

The chart below adds to this picture by plotting each system's mean science score against the share of that score's variance explained by socio-economic status.

This shows that family background explains under 4 percent of the variation in science scores in Macao but in Viet Nam it is much higher at nearly 20 percent. The seven systems highlighted in red are considered by OECD to combine fairness and inclusion with above-average results in all three subjects. Four of them are in East Asia: B-S-J-Z, Japan, Korea and Macao. And these also have some of the largest shares of top performers. So fairness and excellence do not have to be trade-offs. In other words, a system that develops more children is likely to find more exceptional ones.

What the fairer systems have in common is that they put resources where the disadvantage is. This means more experienced teachers in weaker schools, support for students who fall behind early, and less sorting of children into separate tracks at a young age.

But there are also other things that matter which go beyond the classroom:

• The first is exposure. Bell, Chetty and co-authors found that American children who grow up near inventors are far more likely to become inventors themselves in the same fields. This suggests that mentoring and role models are a key part of how talent is developed.

• The other is expectation. Parental and peer aspirations account for a large share of the socio-economic gap in what students hope to do, so careers guidance and contact with people who do the work should be paid adequate attention to.

Underutilized

The third source of loss is talent that is not put to good use.

This loss actually begins at school. As the chart below shows, of PISA’s 1.9 million top science performers, 55 percent expect to be working in a STEM occupation at 30. In B-S-J-Z, the share is only about a third, which is half the US rate. Asia produces 43 percent of the world’s top teenage scientists and 31 percent of those who want a scientific career.

Agarwal and Gaulé (2020) tracked participants in the International Mathematical Olympiad, which are teenagers whose talent had been recognized on a world stage. They found that those participants from low-income countries went on to publish 34 percent less and would be cited 56 percent less than equally scored peers from rich countries.

Talent needs demand, and talent is more productive next to other talent. So we can see internationally that the right tail moves to where the mass is. This is demonstrated in the fact that 83 percent of Chinese and 86 percent of Indian science PhDs on temporary US visas are still in the United States five years after graduating. From a global perspective, their migration need not be a loss. If a highly talented individual moves seeks out more rewarding opportunities overseas and providing that innovations transmit across borders, the world still benefits from what he or she goes on to achieve.

This third loss is one that education ministries cannot fix on their own because it depends on whether there is work at the right level. Countries that want to keep their top performers, or bring them back, need research funding that is predictable, universities and firms that can hire at international rates, and enough of a cluster that a returning scientist has colleagues to work with. China's experience since the 2000s shows that this can be built, and that once it exists the flow reverses. For countries that cannot yet offer this, the practical policy is not to block the exit but to stay connected: diaspora networks, joint research programmes and a clear route home for those who want it.

What it costs

The demographic reality is that fertility rates are plummeting around the world. In China and Korea today, births this year are about half of what they were in 2010 which is when the 2025 PISA cohort was born. We can therefore conclude that if the share of top performers stays where it is, the absolute number of top performers in East Asia will halve by 2040. We also see that across the OECD, average scores have fallen to their lowest levels since PISA began, with much of the decline among the strongest students. As more talent will therefore have to come from somewhere, the obvious source is children who are already here but have never had a proper chance.

That is what Gould’s quote asks of policy. Does policy reach the child who is working instead of studying? Does it keep family background from deciding how much she learns? And once she is ready, is there work for her to do? For Southeast and South Asia, which have large young populations but few top performers so far, these questions matter more than for most.

The areas for action follow from the three losses: getting children enrolled and measuring what they learn, screening every child rather than the nominated few, directing resources to weaker schools, exposing students to people who do the work, and building an economy that can absorb the talent it produces. With cohorts shrinking, the countries that discover, nurture and utilize the talent they already have will be the ones that pull ahead.

Annex: Snapshot of PISA 2025 performance

Below are a few tables taken from the PISA 2025 report.

Performance in science, reading, mathematics and computational problem-solving

Socio-economic disparities and immigrant background in science performance

Yet the same institution publishes a full income distribution for North Korea. Building on the approach developed in my earlier post on the global distribution of living standards, that distribution suggests something very different: North Korea looks more like a middle-income country, perhaps even an upper-middle one, than the low-income country implied by its World Bank classification.

This raises a puzzle: how did the World Bank arrive at the low-income classification in the first place, and how should we interpret distributional data when the underlying national income data are missing?

But there is also a broader question. What do we do when the data we need simply do not exist? When gaps have to be filled, how much should we infer from geography, neighboring countries or other variables that happen to correlate with the missing information?

In what follows, I first briefly explain the new perspective on global inequality in living standards developed in my earlier post. I then identify the countries for which distributional data are missing, examine what this means in the case of North Korea, and ask whether a different approach to missing data can produce better inferences about living standards around the world.

A new perspective on global inequality

In my last post, I asked: for every country, how many of its people live at the standard of living that is typical for its country income group?

Take Brazil. As an upper-middle-income country, how many Brazilians live at the standard typical of upper-middle-income countries? How many live more like people in low-income, lower-middle-income or high-income countries?

To answer this, I used the World Bank’s distributional data from the Poverty and Inequality Platform (PIP) and let the distributions speak for themselves. The exercise produces three “typicality lines”: $3.8, $9.2 and $26.9 a day.

So if you’re a Brazilian living above $26.9 a day, you would be considered as having a living standard more typical of a high-income country than of an upper-middle-income country.

Applied to the world, this shows that four in ten people globally live at a standard more typical of a different income group from the one assigned to their country. In upper-middle-income countries, the worst-fitting category, this number becomes one in two.

What happens when there is no data?

There is a catch. The exercise needs a household distribution for every country. For most of the 218 economies (the count includes territories), there is one, from a survey. But for 46, there is none.

So the Bank fills the gap by imputing a distribution. Note that this distribution is not borrowed (luckily) from the neighbors. Instead, it predicts one from a small set of variables that exist for almost every country, such as GDP per capita where it is available, the income group, child mortality, life expectancy and the rural population share.

Geography plays a role as it enters through the spread of the distribution, which is set by region. The 46 imputed economies share just four inequality profiles, one for Europe and Central Asia, one for Latin America and the Caribbean, one for Sub-Saharan Africa, and one for everywhere else.

But what oddities does this produce?

Take North Korea

North Korea is the clearest case. As noted, the Bank classifies it as low income, on an internal estimate of its income that is never published. The World Development Indicators publish no data on GDP or GNI per capita for the country at all. The Bank’s distributional data, meanwhile, show a country with a median of $7.88 a day and a third of its people at an upper-middle standard, a profile almost identical to Sri Lanka’s, two income groups up, and close to Indonesia’s, if these data can be trusted.

Where does North Korea’s profile come from? With no GDP figure, the model falls back on the income group and the health statistics. For a low-income country, the income group adds nothing. This means that the level is set by child mortality and life expectancy alone. North Korea reports the child mortality and life expectancy of a middle-income country, which is taken at face value, entered into the model and translates to a median of $7.87 a day.

Of all 218 economies, North Korea has the largest gap between the group its people are placed in and the group its country is filed under. The one outside estimate, from the Bank of Korea, lies between the two: 1.72 million won per head in 2024, roughly $1,260, just above the low-income line.

Cambodia and Liechtenstein show the same thing in different ways and in milder form. Cambodia’s imputed distribution looks like Indonesia’s, one group up. Liechtenstein’s, one of the richest places on earth, looks like Türkiye’s, one group down. In Liechtenstein’s case the culprit is the rural population share: the UN counts most of the principality as rural, and therefore the model treats a wealthy micro-state as a farming country.

Is there a better way?

The typicality lines suggest one. One could give each missing country the living-standards distribution typical of its income group, which seems to be the one fact known about every economy. That would be transparent and could never contradict the label.

But it would also assume the label is right, which for the exercise in my last post means assuming the answer.1 That is why the imputed countries were excluded from my previous post as a robustness check, which based on the 172 surveyed economies only did not change the results. This would hand North Korea a blend of DR Congo and Ethiopia rather than anything North Korean. The Bank’s regression is more informative on average, but it seems that its weakness is that for a few countries its inputs are as unobserved as its output.

A more modest fix would be a consistency check: should the imputed median end up more than one income group from the label, flag the case instead of publishing a distribution. PIP already withholds these countries from its website for this reason. The binned data could do the same.

Beware of the borrowed peer

In development analysis, we often look for a peer country. Geography is the obvious place the start. Countries in the same region share histories, institutions, trading relationships, demographics and, often, similar development paths.

But a regional peer is not necessarily a statistical peer.

This might be termed adjacency bias: the tendency to assume that because countries are neighbors, they should resemble one another in ways that matter for development analysis. The Bank’s imputation avoids the crude version, since it does not hand North Korea its neighbours’ living standards. But it still assumes that a country’s inequality is its region’s, and that its welfare can be read off a few correlates that were never measured there either.

This bias seeps through our concepts and our comparisons in more ways than we might notice. Stay tuned for a post on this soon.

This is Post 6 in the The Moving Escalator series on the World Bank’s country income classification, with thanks to Jorge Araujo, Daniel Gerzson, Luis Felipe Lopez-Calva and Espen Beer Prydz for comments and suggestions. Data and code are available at GitHub.

Belíndia, a fifty-year-old fable

In 1974, the Brazilian economist Edmar Bacha wrote a sharp little fable about a not entirely imaginary kingdom called Belíndia: a small, prosperous Belgium inside a vast, poor India.1 The name has come to stand for a country whose average income looks quite respectable but many of its people live far away from that average. A thin upper layer lives as people do in Belgium, while the vast majority lives at the standard of India.

Now translate this fable through the lens of the World Bank’s country income classification. This is the four-rung ladder of low, lower-middle, upper-middle and high income countries that so many conversations in development rely on.2 And ask yourself: when the Bank calls Brazil an upper-middle-income country, how many Brazilians actually live at an upper-middle-income standard? Is the fable of Belíndia still applicable?

It is, and not only for Brazil. The world today is made up of many Belíndias, and the matrix below shows just how much.

Along the bottom of this chart is the income group of a person’s country; and up the side is the standard they actually live at. The blue diagonal then shows the number of people whose living standard aligns with the typical standard of their income group. Off-diagonal, gold means people who are living at a higher standard than their country’s label suggests, while red means people who are poorer. The paler the shade, the further the living standard is from the typical standard for the person’s income group.

Only 61 percent of the 8.2 billion people on this planet are on the diagonal of the matrix and therefore fit their country’s income label. Some 22 percent live at a standard below their country’s label, while some 17 percent live at a higher one.

Why the label matters

A label that misses four people out of ten globally would be a curiosity if nothing depended on it. But a great deal does:

• The labels carry money. Concessional finance, trade preferences and global health funding all use the income classification, directly or as an input, in determining who qualifies. So if a threshold is crossed by a country average, support may be switched off for everyone inside that country. Take for example the 640 million people living at a low-income standard inside lower-middle-income countries. They may end up on the wrong side of the switch if a decision is made on the basis of an average.

• Labels also shapes narratives, for better or for worse. Graduation from one group to the next is often presented as an achievement, and failure to graduate as a trap of some kind. I’ve just visited Sri Lanka, which was declared an upper-middle-income country back in July. Yet, over half of the population continues to live at standards below what is typical for this income group. There can be a large disconnect between the narrative of achievement at the level of a country average and the lived reality of most of its people.

• And the labels also matter as analytical units. The stigmatizing middle-income trap asks why countries in that band struggle to leave it, a question about country averages that is oftentimes interpreted as a question about their people. The changing geography of poverty found that most of the world's poor have moved from low-income to middle-income countries. But much of that movement was relabelling: when India and China crossed a threshold, hundreds of millions of poor people changed category without their circumstances changing at all.

In every one of these uses, the country income label doesn’t adequately capture the living standard of people that live in the country. That is the gap this post measures.

The rest of the post takes three steps. It builds a yardstick for what living at a given income group’s standard actually means for a person. It examines every country’s distribution against that yardstick, which gives the matrix above. And it asks what the matrix has done over four decades, and what happens to it next year.

How to measure a “typical” living standard?

First, a word about data. The World Bank now publishes, for every country, a distribution of “daily welfare” (the term for income or consumption per person, in 2021 purchasing-power dollars). This measure is sliced up into 1,000 pieces of a tenth of a percent of the population of each country, and subsequently lined up to a common year.

These data, which are also behind the global poverty statistics, can be used to examine what is typical of countries within a certain country income group. I will refer to living standards in terms of people so as to keep this apart from the income that classifies a country, where the former comes from surveys and the latter from the national accounts.

So what does an upper-middle-income standard mean for a person?

The obvious route would be to translate the Bank’s dollar thresholds into household terms. This is workable but subject to numerous assumptions, each of which introduces room for error (see Annex 1).

There is, however, a much cleaner way to look at the problem, and the two give broadly similar results.3 It works like this: pool everyone who lives in an upper-middle-income country and draw the curve of their daily welfare. Do the same for the other three groups. At any dollar level, one curve is highest, and the corresponding group is where a person at that level is most typically found. At $5 a day you are most likely in a lower-middle-income country. At $50, you are most likely in a high-income one.

The chart below shows the four distributions on a log scale. The “typicality lines” (for lack of a better term) are exactly where the distributions cross.

For 2025, the four curves cross three times at $3.80, $9.21 and $26.86 per person per day. Below the first line, a person lives at a standard most typical of low-income countries. Above the last line, at a standard most typical of high-income countries. And the middle line separates what is typical in lower-middle and upper-middle income countries. With the lines now established, every country’s distribution can be compared against them.

Who fits the label in 2025?

The matrix above showed the headcounts; the chart below shows the shares.4

The blue segment of each bar is the share of the group’s population for whom the label fits. Red is the share living “poorer than the label” and gold is the share “living richer.”

At the global level, 61 percent (5 billion people) match their label, 22 percent (1.8 billion) live at a poorer standard and 17 percent (1.4 billion) at a richer one.

Group by group, the picture is lopsided:

• The high-income label fits best, at 84 percent (1.2 billion), in part because it can only miss downwards. The low-income label can only miss upward, and it scores relatively highly at 65 percent (500 million).

• The middle is where the label falls apart in two directions. The upper-middle label fits just 52 percent of its people (1.6 billion): 30 percent (920 million) live poorer than the label, most of them in China, Indonesia and the Philippines, and nearly 20 percent (550 million) live richer. The lower-middle label fits 58 percent (1.7 billion), with 22 percent (640 million) at a low-income standard, most of them in India, Nigeria and Pakistan, and 20 percent (580 million) above.

Poor people in middle-income countries are a familiar, well-documented story.5 The striking story is that affluent people in middle-income countries are just as numerous, 1.4 billion of them. And high-income countries hold an even larger number of people living affluently and well above the high-income line, whom the classification cannot distinguish because there is no further line to cross.

The misses are heavily concentrated. China and India alone account for 30 percent of everyone living away from their label. The five largest contributors (China, India, Indonesia, Nigeria and Pakistan) account for 45 percent.

Four decades of fit

In 1990, the world was poorer and simpler. Measured against today’s standards, the label fitted 70 percent of people in 1990. Three billion of the five billion classified people lived in low-income countries, and 84 percent of them lived at a low-income standard.

The fit then fell gradually, with at times a step-like pattern due to populous countries crossing a line. India became lower-middle income in 2007 with most of its people still at a low-income standard. China became upper-middle income in 2010 with most of its people still at a lower-middle standard. By 2010 the label fitted only 48 percent of humanity.

Fast growth in China and India has since pulled their people up toward the label. The fit has recovered to 61 percent.

Between 1990 and 2025, the low-income group has shrunk from 3.1 billion people to 0.8 billion. Its label fitted less well in 2025 because the countries that remain low-income hold many more people living above the standard. Meanwhile, the middle groups have grown from 1.1 billion to 6.0 billion. Their labels fit better than in 1990, but still only about half the time. The high-income group expanded from 0.8 to 1.4 billion people, but the fit of the label barely moved.

That is about to change.

In 2026, the high-income label loses much of its meaning

In a previous post, I argued that the world as a whole is expected to cross the high-income threshold in 2026. On the same nowcast, China is expected to become (or have become) a high-income country in 2026, lifting its 1.4 billion people into that label.

This transition has large consequences for the high-income label. Measured against 2025 standards, 61 percent of Chinese live at an upper-middle standard, 20 percent at a lower-middle one and 19 percent at a high-income one. Move them all into the high-income group and the group doubles, from 1.4 to 3.0 billion people, while the share whose label fits falls from 84 to 53 percent. Over a billion people in high-income countries will live at an upper-middle standard.

For the world, the fit drops from 61 to 54 percent in a single year, the sharpest fall in the series. Nobody's living standard has fundamentally changed. The shifts are mostly the result of relabelling.

Ten countries per income group

The charts below show, for each income group, the ten most populous countries and how their populations are distributed across the four living standards. The letter at the right of each bar indicates the underlying survey method: consumption (c), income (i), or nothing (n). In the last case, the distribution is imputed from neighboring countries.

A note first on survey methods.

The choice of survey method helps explain why people appear to live at a low-income standard in high-income countries like the United States, Germany and Japan. This is a feature of annual income surveys, which record zero or negative income for the self-employed in a loss year, students, people between jobs, and people living off savings.

Consumption surveys do not produce the same result because consumption has a natural floor. Income distributions also have a much wider spread than consumption distributions because people save or draw down savings. The upshot is that a country measured by income will look more Belíndian at both ends than the same country measured by consumption.

High-income countries

The high-income country chart shows a high degree of “label fitness.” Two factors explain much of this. First, high income is a terminal category, so there is no higher group for people to fall into. Second, most of the countries shown have GNI per capita well above the high-income threshold.

Russia is the exception. It is placed just above the threshold and, given its level of inequality, it is no surprise that 30 percent of its residents live at an upper-middle-income standard.

Upper-middle-income countries

The upper-middle-income group is where the label fits worst. Its ten most populous members show why.

Türkiye, Brazil and Mexico are measured by income and they show the largest tails on both sides. For Türkiye and Brazil, the high-income segment is very large. This in part reflects that Türkiye’s GNI per capita is already above the high-income threshold and the World Bank decided to keep Türkiye classified as upper-middle under its transition rules. But part of the result also follows directly from using income surveys.

The other countries use consumption surveys. Remember that these are percentages of the population, so China’s 20 percent at lower-middle income standard means 280 million people, which is more than the whole population of Indonesia. Indonesia and the Philippines have very large shares at the lower-middle standard since both countries are just above the threshold.

South Africa is an outlier of a different kind because of its high degree of within-country inequality: a quarter of its people live at low-income standards and a tenth at high-income standards. Its median is only 53 percent of its mean, the widest gap of any large country.

Lower-middle-income countries

India has its own version of Belíndia. The label fits 72 percent of the population, which is the best fit of any large country the group (unsurprisingly, given how heavily India itself influences the lower-middle-income standard) but the remaining 28 percent is 410 million people: 150 million at a low-income standard and 250 million at an upper-middle one.

Morocco and Egypt are at the top of the band, which explains the lower shares of people fitting the group. Conversely, Tanzania, Nigeria, Myanmar and Pakistan are closer to the bottom, and therefore a still-large share of their populations live at a low-income standard. Kenya and Angola have a GNI per capita well inside the band, but their medians are dragged down by inequality, so they too have a large share living at low-income standards.

Low-income countries

The low-income group is small and much smaller than it once was. Those who are left in this group tend to fit the label well by default, with between 68 and 88 percent of the population in Niger, Uganda, Madagascar, Yemen, the DRC and Mozambique at a low-income standard.

Ethiopia and Sudan are the exceptions. Both are near the threshold, with roughly half their people at a lower-middle standard.

The two countries without a survey have had a distribution imputed. Afghanistan’s imputed distribution puts 63 percent of its people at a low-income standard, which is plausible. But North Korea is shown as having 89 percent of its people at a lower-middle or upper-middle standard, with a median of $7.90 a day, higher than India’s.

This most likely reflects the Bank imputing an East Asian regional profile onto a country for which there are no data. Either the distribution is wrong, or North Korea does not belong in the low-income group. The classification and the imputation do not add up.

Where this leaves us

The World Bank’s country income classification does what it is built to do. It sorts countries using a single measure, GNI per capita, offering a one-dimensional and distribution-blind lens on development. Within those parameters, it does so transparently, consistently, and comparably across countries.

It does a good job of capturing the gap in average living standards between income groups. Differences between the four income groups account for 60 percent of global inequality on their own. On this dimension, the label is a good summary.

By design, though, the classification says nothing about how income or consumption is distributed within countries. That becomes a problem when the classification is used in ways that implicitly treat the country average as representative of its people. This happens in determining eligibility for concessional finance, trade preferences or health funding, in country narratives that treat the average as the norm, and in analytical frameworks that produce misleading conclusions, such as the notion of the middle-income trap.

Three changes would help.

• Build a distributional element into the classification itself. The precedent is well established: the inequality-adjusted Human Development Index discounts each of its three components by the inequality in its distribution, including per capita income, which is the first component.6

• Publish the matrix alongside the classification. Every July the Bank announces the new classification. The same release could report, for each country, the share of its people living at the group’s standard, at a poorer one and at a richer one. This would be straightforward to implement.

• Encourage responsible use. Labels need warning labels and regular reminders of what they capture and what they leave out. The point has been made in this series before, on not being a “trappist” and on warning labels for development frameworks.

Two annexes follow. Annex 1 sets out the methodology and tests how much the results depend on it. The lines and the fitness share survive every robustness check: varying the smoothing, dropping India and China from the estimation, correcting for underreporting at the top, and re-estimating the lines year by year all leave the 2025 global figure within about a point of 61 percent. Annex 1 also works through the alternative of translating the Bank's own GNI thresholds into living standards, which gives 56 percent. That route requires a chain of conversion assumptions, which is why the post leads with the typicality lines. Annex 2 gives the Belíndia matrix for every country in the dataset, organized by UN region and subregion.

Finally, I would be interested in your views. Is an adjusted classification worth some loss of simplicity? Which of the three changes would you prioritize? And have you seen the income label used in ways this post has not anticipated?

Annex 1. Methods and robustness

Description of welfare aggregates

The distributional data come from the World Bank’s 1000 Binned Global Distribution (PIP vintage of March 2026). They are expressed in 2021 purchasing-power parity dollars and cover every year from 1990 to 2026. For each country and each year, the dataset supplies 1,000 bins where each bin holds exactly a tenth of a percent of the country’s population. The bin contains the mean daily welfare of the people in that bin. These numbers are subsequently lined up to a common year by extrapolating or imputing on the basis of the growth rates observed, or projected in case of nowcasts, in national accounts aggregates. For consumption-based surveys, the convention is to allow for a 70 percent passthrough from growth in real Household Final Expenditure per capita. For income-based ones, there is full passthrough from real GDP per capita growth. The 2025 and 2026 distributions for countries whose last survey is older are therefore nowcasts.

Country income groups come straight from the World Bank’s classification updates per the OGHIST database. The latest update of July 2026 corresponds to FY27 and is based on the GNI per capita data (nominal USD, Atlas method) of 2025. This post shows 2026 nowcasts which are based on a mechanistic extrapolation of each country’s median annual growth rate of GNI per capita (nominal USD, Atlas method) observed over the previous decade. Full details are available in the Great Income Inversion post at Moving Frontiers. Note that all 218 countries with a published country income group are included but 46 of them have no household survey. They enter the bins with distributions imputed by the Bank from regional profiles, which are marked by “n” in the survey columns in this post.

The welfare aggregate is household consumption per person per day where the survey measures consumption (this is most of Asia and Africa, 5.9 billion people) and household disposable income per person where the survey measures income (Latin America and nearly all high-income countries, 2.1 billion).

One key feature in the distribution of welfare aggregates is skewness. Let gamma be, for each country, the median divided by the mean of daily welfare. In 2025, this variable ranged from 0.49 (Namibia) to 0.96 (Slovakia), with a median across countries of 0.79 and a population-weighted value of 0.80. Country income group medians are calculated as 0.77, 0.78, 0.77 and 0.83 from low to high income. Note that income surveys show higher gamma (median 0.83) than consumption surveys (0.78), partly because they are concentrated in less unequal countries but also because income distributions extend further at the bottom, where zero and negative incomes pull the mean down relative to the median.

The world as a whole is more skewed than any country in the dataset. Its own gamma is 0.43 and 74 percent of people live below the world mean of $22 a day against a median of $9 (see above chart). This results from pooling very different distributions where the inequality between countries is added to the inequality within them and the combined tail becomes longer than either. The gammas for each income group are about 0.66, as the chart below shows.

Curves, lines and people

In order to construct the “typicality lines” for each income group, we first need to draw the distributions of each income group. To draw them, I lay a grid over log welfare that runs from $0.10 to $2,000 a day in steps of 0.05 (or about five percent of welfare per cell). For each country, the share of people at or below each grid edge can be taken directly from the sorted bins. For each group, the share of its people in each grid cell is the population-weighted average of those country shares.

The next step is to divide by the cell width to get a density in log welfare and make sure the density integrates to one. Each curve is normalized to the unit area, which ensures that we compare the shapes of the curves of the four groups and not their sizes. In other words, a group of 3 billion people and a group of 800 million get the same weight. The raw densities show as steps, so I smoothed them (with a Gaussian kernel of bandwidth 0.2 in log welfare, which amounts to a 20 percent spread). The choice of bandwidth is an assumption in the method which will be checked for robustness later on.

At any level of welfare, the most typical group is the one whose smoothed curve is highest. The first line is the lowest welfare level at or beyond the peak of the low-income curve where the most typical group becomes lower-middle or higher; the second and third lines are defined in the same way from the peaks of the lower-middle and upper-middle curves. Within the grid cell where the switch occurs, the crossing is placed by linear interpolation on the difference between the two curves involved. For 2025 the lines are $3.80, $9.21 and $26.86 per person per day. The charts below show the mechanism on both a linear and a log scale.

Once the lines are fixed, each person is assigned to the standard of the segment they fall in. For every country the share of people below each line is then derived from the sorted bins, with linear interpolation between adjacent bin means. The four shares are multiplied by population and cross-tabulated by the country's income group. That is what ultimately gives the Belíndia matrix shown in the main body of this post.

Robustness of results

The share of the world’s population that fits the country income label in 2025 is 61.0 percent. In what follows, I will check the robustness of the method outlined above by considering a few alternative specifications.

a. The extent of curve smoothing

The only free parameter in our method is the kernel bandwidth. This robustness check asks whether the fitness share is much affected by changes in this parameter. The table below varies it from 0.05 to 0.40 in log welfare (this is an eightfold range running from about 5 percent of welfare to about 50 percent). It also adds an Epanechnikov kernel scaled to the same variance as the baseline Gaussian.

As the table shows, the lines move very little and the fitness share moves even less. The lower-middle boundary runs from $9.13 to $9.40 across the whole range, the high-income boundary from $26.72 to $26.93, and the low-income boundary from $3.70 to $3.88, so in each case the deviation is under 4 percent. The share of people whose label fits goes from 60.8 to 61.3 percent, the share living poorer than their label from 21.7 to 22.1, and the share living richer from 17.0 to 17.1. Swapping the kernel shape for an Epanechnikov of the same variance also makes little difference.

b. Exclusion of China, India and others from the estimation of the line

The lower-middle curve is dominated by India and the upper-middle curve by China because of their sheer population weights. A reasonable worry therefore is that the lines describe those two countries rather than their groups. The table below tests this by leaving each of them out of the curve estimation while still counting their people in the matrix.

Removing India lowers the middle line from $9.21 to $8.19, whereas the fit becomes 60.7 percent. Removing China lowers the top line from $26.86 to $24.90 and the fit to 60.5 percent. Removing both gives 60.0 percent. Restricting the curves to the 172 economies with an actual survey, which drops the 46 imputed distributions, leaves the lines unchanged at 61.0 percent. The bottom line is that the giants do shape the lines, but the fitness share does not move all that much (at most one point).

c. The missing rich

Household surveys famously undercount the top of the distribution. To assess the implications of this, we can scale the welfare of every country’s top decile. The exercise below does so by a factor that rises linearly from one at the 90th percentile to two at the 100th percentile. The transformation raises group means by 9 to 17 percent, which roughly corresponds to the size of the survey-to-national-accounts gap that is commonly attributed in the literature to the missing rich (see also Kimura and others, 2026).

Despite these transformations, the lines don’t move much and the fitness share falls to 60.6 percent. This is because all three lines are well below the top decile of each group. Lifting the top decile therefore does little beyond moving people who are already above their country’s label further above it.

d. Fixed, yearly and pooled lines

We can think of three approaches with respect to applying the lines over time. The lines can be held at their 2025 values, re-estimated for each year using that year’s own group curves, or estimated once from all years pooled together.

The post uses the first approach so we can keep the yardstick constant and see what happens with the data measured against it. If instead we re-estimate the lines for every year, the fitness share would be 73 percent in 1990, 52 percent in 2010, 61 percent in 2025 and 57 percent in 2026. Note that the 2026 figure exceeds the fixed-line value of 54 percent because China’s arrival in the high-income group shifts the group’s curve and lowers the top line from $26.86 to $18.76. This helps absorb part of the shock. Pooling all years gives lines of $3.78, $7.72 and $23.29 and a 2025 share of 59 percent. All three series have the same shape.

e. Income versus consumption

Income distributions are more dispersed than consumption distributions because dissaving allows the bottom to stretch and saving helps stretch the top. A country measured by income will therefore look more Belíndian than the same country measured by consumption. The pattern shows in the results: among upper-middle countries, the 39 measured by consumption fit their label 54 percent of the time and the 16 measured by income only 47 percent.

Estimating all three lines from a single survey type is not possible. That is because the low and lower-middle groups are almost entirely consumption-based and the high-income group entirely income-based. The only two lines that can be compared, however, land close to the pooled ones: consumption countries alone lead to a lower-middle boundary at $8.97 against $9.21, whereas income countries alone produce a high-income boundary at $25.07 against $26.86.

In the table a different route is taken. Each income-measured distribution is compressed towards its own median, narrowing its spread to something that is closer to a consumption survey. Following this, the lines are re-estimated.

It should not surprise that the compression produces a smaller gap between the two kinds of upper-middle countries. At an elasticity of 0.8, the two types fit their label at 54 and 55 percent (compare this against 54 and 47 in the baseline). So much of that seven-point gap likely relates to the measure rather than the living standard per se.

The global fitness number rises with the adjustment, from 61.0 percent to 62.4 at an elasticity of 0.8 and 63.1 at 0.7. However, only the top line moves (from $26.86 to $27.62), because compression mostly affects the high-income group where every survey is income-measured. The mixing of survey types therefore biases the fitness share towards more mislabelling, and we can safely conclude that the baseline figure mixing the two is conservative.

One visible artifact remains: 27 million residents of high-income countries who were recorded at a low or lower-middle standard. These are people for whom annual income surveys record zero or negative income: the self-employed in a loss year, students and people between jobs. Consumption surveys have no such tail.

Inequality decomposition

Global inequality in living standards can be split exactly into three levels using the mean log deviation (MLD):

• Differences between the four income groups

• Differences between countries inside the same group

• Differences within countries.

The first term is the MLD of the four group means, the second is the population-weighted average, across groups, of the MLD of country means within each group, and the third is the population-weighted average of each country’s own MLD. The three add up to the MLD of the pooled distribution of all classified people. The chart below shows them stacked for 1990 and 2025.

What do these three levels mean for the classification? The classification is about country income labels so it is almost tautological that the inequality between a typical resident of a high-income country and a typical resident of a low-income one is recorded well by the classification. But the classification cannot detect the second and third levels of inequality. Variation within an income group doesn’t show because the GNI per capita levels are collapsed into labels. And, of course, the classification cannot capture a country’s internal income distribution.

Note that in 2025 the total was 0.69. Differences between income groups account for some 60 percent. Differences between countries inside the same group account for just 9 percent — this is small because the groups are actually narrow bands on the a log scale and most of the distance between countries is distance between bands. Differences within countries account for a much more substantial 31 percent, which has been the main subject of this post. In other words, the classification captures three-fifths of global inequality, but is blind to the remaining two-fifths and almost all of that remainder is inequality within countries.

The picture has shifted quite significantly since 1990. The total MLD was 0.99 then, of which 71 percent was due to inequality between income groups, 6 percent between countries within groups and 23 percent within countries. So the whole of the decline has been due to falling between group inequality, the subject of the previous post.

The chart below turns the same decomposition into distributions. Placing every person at their country’s mean removes all within-country inequality and leaves a two-humped curve consisting of an Asian mass around $7 a day, the rich world around $60, and a trough between them where few countries are. Rescaling every country’s distribution to the world mean removes all between-country inequality instead. This leaves a single hump. The actual distribution is between these two extremes, but closer to the first, with 74 percent of people living below the mean.

Sources of the off-diagonal

The last question is where the 39 percent of people who live away from their label come from. Two factors are responsible and their contributions can be separated with counterfactual distributions produced against the 2025 lines.

The first is skewness. Imagine we were to remove all within-country inequality by placing everyone at their own country’s mean welfare. That would raise the fitness share from 61 to 81 percent. The second is position. Imagine we were to remove the effect of a country sitting near the edge of its band by rescaling every country’s distribution so that its mean equals its group’s mean. That in turn would raise the share to 66 percent.

Removing both raises the fitness share to 91 percent rather than 100. The residual is the low-income group, whose mean of $3.89 a day lies just above the low-income line of $3.80. As a result, even if everyone is at the group mean, the whole group counts as lower-middle.

Because the two effects overlap, their sum exceeds the total. If we were to average over the order in which they are removed when attributing the effects, the 39 off-diagonal percent translates roughly into 22 percent due to within-country skewness, 7 percent due to position and 9 percent due to the group-level offset.

Of the two factors, skewness therefore accounts for about three-quarters and position for about a quarter. Belíndia is mostly a story of distribution within countries, and only secondarily one of where countries sit within their bands.

Position is concentrated in a specific set of economies. Forty-eight of them, holding 1.5 billion people, have a mean welfare outside the segment between their own group’s typicality lines. Fifteen of the 25 low-income economies are among them, with 443 million people, all with means above the low-income line. They are the countries where the label misses one way only.

Alternative approach of translating the thresholds directly

The obvious alternative to the “typicality lines” is to convert the World Bank’s own GNI per capita thresholds into daily welfare, so that each line is the living standard of an economy sitting exactly on a threshold. The conversion factor, call it rho, is an economy’s 2025 mean welfare times 365, divided by its 2025 Atlas GNI per capita. The factor absorbs many things that separate the two measures: the consumption share of national income, the gap between survey aggregates and national accounts, and the difference between PPP and Atlas exchange rates.

The chart below plots it for the year 2025. Three patterns emerge. First, rho falls when GNI per capita rises. It is around 1.5 in the poorest economies and below 0.5 in the richest ones. It turns out that survey welfare is a much smaller share of national income in richer economies. Second, rho is very dispersed. At any level of GNI per capita, the ratio varies by a factor of two or more across economies. Third, the thresholds are in the middle of that cloud, so the line will depend on which economies are in the band that is used to define it.

The method followed here is to take the median rho among the surveyed economies whose GNI per capita lies within a factor of two of each threshold. This yields respectively 35, 60 and 50 countries at the three lines (these are the bubbles contained in the three shaded bands visualized above). Multiplying them by the threshold yields the median ratios of 1.47, 0.87 and 0.67, and the threshold lines of $4.74, $11.01 and $26.19. Using these lines, the fitness share is 56.2 percent.

The drawback of this approach is that it involves many steps, and with each step there is room for error. The table produces a robustness check on the size of the window and the choice between mean and median welfare. By and large, though, the fitness shares do not vary dramatically.

There is one further complication. The numerator is welfare in constant 2021 PPP dollars and the denominator is GNI in current Atlas dollars, so the two sides have to be put on the same price basis. Once again, there is significant room for error. If we pair a survey-year GNI with a current threshold, the lines will be inflated by whatever dollar inflation has occurred since the survey. For a 2016 survey, this would have been about 17 percent.

The typicality approach needs none of this. The only free parameter that is a choice is the level of smoothing, which as shown earlier doesn’t make much difference. This is why this post leads with the typicality lines.

Annex 2. Country tables

The tables below provide the full results for countries, grouped by UN region and subregion. The bold number indicates the income group to which the country currently belongs.

The center of gravity in the country’s strategy needs to shift now.

This post calls for three shifts:

• The first is from stabilization to growth and transformation. Sri Lanka needs to move, as it were, from a stability programme with a development element to a development strategy with a stability component.

• The second is to give the development strategy a much stronger outward orientation, while making sure that greater global integration simultaneously translates into broader domestic gains. I describe this as the mutually reinforcing objective of “connecting globally, developing locally.”

• The third is from setting priorities to delivering them. The constraint to progress lies less now in identifying what needs to be done than in implementing it. Sri Lanka needs stronger state capability to implement its priorities consistently, comprehensively and continuously.

Foreign investment offers a useful lens on these three shifts. So this post starts from a narrow question: how much foreign investment should Sri Lanka be attracting?

And I will argue that the difference between actual and potential speaks to the distance the country itself still has to travel as it transitions from stability to development.

Fifty years on a low plateau

Over the last fifty years, foreign direct investment into Sri Lanka has averaged a meager 0.91 percent of GDP. There has been some volatility around this number, such as the spike to 2.8 percent in 1997 due to privatization. But the overall picture is one of a low plateau. The latest information we have is that by 2024 FDI had fallen to 0.8 percent, below the long-term average.

It is widely recognized that FDI is below what Sri Lanka could be attracting. Familiar explanations for the underperformance are internal conflict, limited size of the domestic market, changes in policy, weaknesses in factor markets, currency volatility, and of course the recent debt crisis.

Sri Lanka’s low FDI rate hurts its development prospects. This is a country of 22 million people, which makes it difficult to generate sufficient scale from the domestic market alone. Growth needs connections to markets that are larger, richer, or both. And it typically is foreign investment that helps build those connections through exports, technology, management practices and participation in cross-border production networks. It seems, though, that Sri Lanka has been moving the other way. Exports of goods and services have fallen from 39 percent of GDP in 2000 to 19 percent today. And the country’s export basket remains highly concentrated in garments, tea and rubber, much as it was a generation ago.

How much investment should Sri Lanka attract?

Against Sri Lanka's own fundamentals: three times more

Since comparisons across countries can be misleading due to country differences, it is important to establish a benchmark. I construct one from an annual panel of 130 economies from 1996 to 2024. This produces some 3,100 country-year observations, with the sample and its exclusions described in Annex A1.

Investment is explained by the typical variables suggested in the literature: market size, income level, growth, trade openness, inflation, the quality of governance, the presence of conflict, and the size of surrounding markets. This last variable sums each economy’s GDP and discounts it by distance. A Sri Lanka-specific dummy is then added to measure the country’s “conditional shortfall.” This captures how much less investment Sri Lanka attracts relative to other countries with similar characteristics. For full details, see Annex A2.

The estimates suggest that the Sri Lanka dummy is just -1.85 percentage points of GDP. Against a predicted level of about 3 percent, actual inflows averaged just 1.1 percent over the period.

In other words, Sri Lanka attracts roughly a third of what its fundamentals on average predict.

Against Sri Lanka’s frontier: five times more

A different question is what the best outcome is that countries with similar fundamentals have achieved.

The stochastic frontier method is suited to answer that. It splits the unexplained part of the regression into random noise and a one-sided shortfall relative to a frontier. Pooling 125 countries over the 1996-2024 sample period, the estimates suggest that 89 percent of the unexplained variation represents a one-sided shortfall from a frontier (Annex A4).

Furthermore, Sri Lanka attained only 21 percent of the frontier over the sample period. Compare this with the average of 41 percent across the sample’s 125 countries.

The best attainable frontier value for the FDI ratio of a country with Sri Lanka’s size, income, growth, openness, macroeconomic conditions, governance and location is 6.8 percent of GDP. Since actual inflows averaged 1.14 percent over this period, this translates into a five-fold gap.

This places Sri Lanka in the bottom decile of the sample, ranking 114th out of 125.

Against actual country peers

The benchmark and the frontier are in fact artificial constructs informed by a global dataset. These measures are Sri Lanka-specific and not actual countries.

So how would Sri Lanka compare against actual countries? The following offers two types of comparison against structural peers and regional peers.

The left panel shows “structural peers.” These are economies that most resemble Sri Lanka on size, income per head, governance, market access and openness. Costa Rica reaches 63 percent of its frontier and the Dominican Republic attains 60. These are both small economies that built export platforms. Peru and Ghana are above 50 percent, whereas Sri Lanka is at 24 percent, the third from the bottom.2 Moreover, the twenty-five structurally closest economies average 3.36 percent of GDP in foreign investment against Sri Lanka’s 1.14.

The right panel shows a number of regional comparators that often arise in policy conversations. While several of these are structurally quite unlike Sri Lanka, the general picture doesn’t change much. Cambodia, China, Indonesia, India and Viet Nam are all well above Sri Lanka’s 24 percent, which is between Bangladesh and Pakistan.

Why the gap has narrowed and didn’t close

In the previous discussion, we looked at the gap over a long period of time. But the gap itself varies from year to year, as the chart below shows.3

These year-by-year data show that actual inflows have been below the benchmark in 19 out of 21 years. This strongly suggests that the gap is not an artifact of a few bad years.

The chart also shows that the two lines converge. Why? It turns out that the reason is not that Sri Lanka attracted more investment. Two things happened instead.

The war ended. As a result, the shortfall halved from 2.27 points of GDP during the period 1996 to 2009 to 1.06 points in the fifteen years since (Annex A3). Note that the benchmark itself fell. Between 2007 and 2024 the investment the model expects Sri Lanka to attract dropped by about four points. Nearly a full point is the result of trade openness falling from 69 to 42 percent of GDP.

The crisis then “finished the job.” In 2022, with growth at minus 7.3 percent and inflation near 50, the benchmark fell to its lowest level of 0.83 percent. As the chart shows, there were two instances where Sri Lanka exceeded the benchmark. But beating the benchmark in this case meant that expectations were falling toward the low level of investment, not the other way round.

What explains the gap that remains? It seems that conflict variables did not reduce the Sri Lanka dummy in the estimates and none are significant. The post-war shortfall is 1.06 points with the default years included and 1.07 without them, which suggests that the gap was fully in place before the crisis. After fifteen years of peace, Sri Lanka still suffers a gap of a full point of GDP, which amounts to roughly $900 million a year.

Where the crisis appears to have pulled the benchmark down, this mechanism seems to be operating now in reverse. Recent indicators suggest improvement with growth back at 5 percent and reserves at their highest since 2020. The benchmark responded to this news by climbing to 1.74 percent in 2024 against inflows of 0.76. The gap therefore widened to nearly a point.

The recovery, however, can do only as much. To begin with, the coefficients on growth, inflation and governance are relatively small. And the IMF also projects weaker growth in 2026 and beyond (of about 3 percent), which should take the benchmark back to 1.4.

What would make a bigger difference is openness. Regaining the 27 points of trade-to-GDP lost since 2000 could raise the benchmark by about 0.9 points. Reconnecting Sri Lanka to world trade therefore seems an obvious priority in what comes next.

Three shifts after stabilization

From stabilization to development

Sri Lanka has weathered the pandemic, a first-ever sovereign default, the Russia-Ukraine war, cyclone Ditwah, and now the Gulf conflict. The cyclone alone did $3.4 billion of damage according to the Post-Disaster Needs Assessment. This is roughly four years of the investment shortfall in a single week.

Against that background, the recovery has exceeded expectations: growth of 5 percent in 2025, record merchandise exports and remittances, and usable reserves at their highest since 2020.

Despite this success, the stability agenda requires continued attention. As noted earlier, slower growth is predicted for 2026 and also for the two years that follow. External vulnerabilities persist and soon Sri Lanka will face the prospect of capital repayments. In light of this, macroeconomic discipline must remain central.4

But the center of gravity needs to shift for two reasons.

One reason why the focus needs to rebalance toward development is that the aggregate recovery hasn’t sufficiently reached people. Sri Lanka rejoined the upper-middle-income group in the World Bank’s country income classification this year, which attests to the positive momentum in per capita income growth. But this lift of the average has not changed the fact that a large part of the population continues to live in deprivation. Poverty reduction since the crisis has continued to be slow.

We can measure poverty against the $8.30 upper-middle-income poverty line that goes with the country’s new income class. Under this measure, poverty rose from 48 percent in 2017 to 71 percent at the 2023 peak and is projected to fall only to 63 percent by 2028. At the $3.00 international poverty line, the crisis multiplied extreme poverty more than fivefold, from under 2 percent to nearly 11. For 2028 a poverty rate of 6.2 percent is projected. We also know that about a third of households experienced moderate or severe food insecurity in 2025, nearly three times the 2019 level.

A second reason to rebalance from stabilization to development is that Sri Lanka urgently needs faster growth. Having just reached upper-middle income status, the journey isn’t over. The country has a long way to go to reach the next stage on its development journey and that is the stage of the high-income economy. Rapid real economic growth coupled with real appreciation of the currency against rich-country currencies is the only way to get there.

The above chart places Sri Lanka’s path to high income into historical perspective.

• The left panel extrapolates Sri Lanka’s growth experience over the last decade (2016-25) which saw median per capita income growth in nominal US dollars of 2.8 percent a year. Under this scenario, the country would not reach high-income status until 2098.

• The right panel offers a more positive assessment by extrapolating to the future the median growth pace of what was obtained a decade earlier (2006-15), when growth in GNI per capita was 13.2 percent a year. Under these conditions, Sri Lanka would reach high-income status by 2036.

The near-term outlook appears closer to the slow panel as the IMF projects growth of about 3 percent (in real GDP) for 2026 through 2028, in a labour market where roughly 70 percent of employment is informal. Meanwhile, some 300,000 people left the country in 2024 alone. The record remittances celebrated in the external accounts appear as a mirror image of a growing stock of skilled migrants abroad.

It is quite clear that faster growth needs more and better investment. For an economy the size of Sri Lanka’s, a significant part of that investment will need to come from abroad. This is where the one percent in the very first chart poses and will continue to pose a significant constraint on Sri Lanka's development trajectory. Addressing this constraint will require another shift in strategy.

Connecting globally, developing locally

That second shift will be to pay far greater attention to fostering global integration while ensuring domestic development. Connecting globally while developing locally is what Sri Lanka needs at an accelerated pace to advance its longer-term development agenda.

Connecting globally

A sustained push to attract export-oriented firms and deepen integration into regional and global value chains is essential. This requires efforts related to trade reform, investment-climate improvements, infrastructure and skills, all focused on sectors with a plausible competitive position.

The analysis has shown that openness is the lever that raises the benchmark the most. The trade-to-GDP ratio has halved since 2000, and the analysis found that each point recovered is worth about three-hundredths of a point of predicted FDI.

Adjustments in trade policy could be considered to address any anti-export bias embedded in the tariff structure. Para-tariffs and duties on intermediate inputs are raising the cost of producing in Sri Lanka for sale abroad. Such practices are not compatible with the ambition to raise Sri Lanka’s participation in cross-border production networks, where components typically cross borders several times before final assembly and thus could be exposed to tariffs multiple times.

Investment facilitation deserves attention as well. Long-standing tax holidays and duty exemptions have not yielded the expected success in raising inflows. This might be consistent with the earlier finding of red tape playing a key role. Think of the number of approvals an investor must obtain, the difficulties of securing serviced land, or the speed of customs clearance.

The above chart shows the 2011 and 2025 rounds of the World Bank's Enterprise Survey for Sri Lanka. The share of firms reporting licensing and permits as a major obstacle apparently rose from 19 to 43 percent. The share that listed tax administration as a major obstacle went up from 18 to 44 percent. On the positive side, electricity and access to land improved.

Colombo Port City represents a major inward investment opportunity. It has its own commission and regulatory regime and is therefore in effect like a controlled experiment that allows us to test whether Sri Lanka can supply what investors say they are elsewhere missing (eg a single point of approval, a stable tax and legal framework, a credible dispute resolution system, and clarity on exchange control and cross-border settlement). If the special regime works, the case for extending it to the rest of the economy will become hard to resist.

Various sectoral opportunities exist where Sri Lanka could raise its global profile. Electronics are seeing growth. Ports and logistics offer potential especially if port and airport capacity can be aligned with traffic flows and progress is made with modernizing customs and moving to a single window. Tourism remains however below potential and plenty of untapped opportunities exist in the digital economy. Sri Lanka could also take greater advantage of its proximity to India.

In all of the above, industrial policy has a role to play. This is especially so where coordination failures pose bottlenecks. But at the same time it is important to recognize that support without discipline may end up protecting activities that never have a chance of becoming competitive. Assistance should therefore be time-bound, making use of sunset clauses and published performance conditions rather than open-ended exemptions that are renewed by default. The objective is to promote internationally competitive firms, not uncompetitive national champions.

Developing locally

Developing locally is in part a precondition to connecting globally, but it is also the objective of Sri Lanka’s growing global integration: to spread the gains of its external interactions more broadly across its economy and society.

The agriculture sector is worth mentioning in this regard as it is one of the sectors with the potential to support a large part of Sri Lanka’s economy while also employing a large part of its people. Agriculture employs around a quarter of the workforce and it produces about 8 percent of GDP.

The sector, however, faces several challenges. These range from the small scale of holdings and the low levels of mechanization to the thin access to capital. There is also a web of input subsidies which, combined with import protection policies, keep production viable at the expense of productivity. It may be helpful to conduct a review of land use, crop mix and the rules governing paddy land. Rural incomes would benefit.

Tourism is another sector that affects many parts of Sri Lanka’s domestic economy, including lagging regions. In 2025, Sri Lanka received 2.36 million tourists. This is the highest on record, exceeding (finally) the peak of 2018. But arrivals fell close to a fifth year on year in March 2026 as the Gulf conflict disrupted long-haul routes.

Aside from boosting volume, opportunities also exist to move towards a higher-value tourism model. In November 2025, average spend was $148 per visitor per day according to the Sri Lanka Tourism Development Authority. This is very low for a destination with Sri Lanka's range of assets. The low value also reflects that the country still scores weakly on international assessments of tourism development. Efforts could be considered to expand the geographical footprint of tourism beyond the southwest coast into the North, East, North-Central and Uva provinces. There are also opportunities to raise safety, service and hygiene standards island-wide.

Other priorities to foster local development pertain to vocational skills, incentives for firm-level training, transport connectivity to growth centres, and factor markets. Commercial land is scarce, which creates opportunities to activate state land through long leases on commercial terms. Given the low levels of female labor force participation (below 30 percent), tackling practical constraints that women experience with respect to childcare, safe transport and school-compatible hours are low-hanging fruits. Outmigration of people poses a challenge as the country loses skilled workers steadily while offering few avenues for skills to come in.

Expanding state capability

The investment shortfall of Sri Lanka is not market size, income, growth, openness or macroeconomic stability, because all of these are inside the benchmark.

It is also not government effectiveness as the model controls for this dimension of governance. Also, Sri Lanka's score is not weak enough to account for a gap this size and the indicator captures the general quality of the bureaucracy rather than regulatory practice or contract enforcement. It is not market access (Annex A2 explains why). It is not the war or the crisis, for the reasons given above.

Two explanations remain:

• The first is the friction of doing things. This relates to land acquisition, energy cost and reliability, the number of agencies an investor must deal with, and how contracts are enforced in practice. None of this is captured fully in a broad governance index. But all of it matters in decisions about where to locate a project.

• The second is predictability. Sri Lanka has repeatedly reversed direction on issues such as trade agreements, tax regimes and land rules across electoral cycles. An investor with a fifteen-year horizon will put a premium on stability before committing capital.

In light of the above, Sri Lanka needs to upgrade the capability of the state in a number of dimensions. It will be important for government to be better able to identify priorities, coordinate across agencies, unblock bottlenecks, and maintain continuity. While policymakers must retain the flexibility to change course, a steady policy stance is important to stimulate long-horizon investments in infrastructure, skills and export capability. And the country also needs a comprehensive approach to reform to ensure that bottlenecks are tackled in a coordinated, non-siloed way, across the board.

In sum

For over fifty years, foreign investment has hovered at a low of about 1 percent of GDP. Sri Lanka could do a whole lot better. Its own fundamentals over the course of the country’s last few decades point to potential investment being three times higher, i.e. at around 3 percent of GDP. Relative to the frontier constructed from what similar countries have attained, investment could be raised five times, to about 5 percent.

The framing of this post is one of opportunity. Thanks to a successful period of stabilization, Sri Lanka has restored many of the conditions for investment to take off. But the attention now has to shift from stabilization to transformation and growth. Macroeconomic discipline remains critical but development now needs to take center stage, so that the investment benchmark which has been weakened by crisis and two decades of falling openness can be raised back to historical levels or, even better, to those at the frontier.

To do so, however, will require two additional strategic shifts. The first is to enhance Sri Lanka’s integration into the global economy, while at the same time, in a mutually reinforcing way, advancing the domestic agenda. This is described in this post as the dual objective of “connecting globally, while developing locally.” And the second, perhaps most important of all, is to ask ourselves: if change didn’t happen in the past, why would it now? And on that basis to adopt a laser-sharp focus on implementation, which will require raising state capabilities so that the above two shifts are pursued consistently, comprehensively and continuously.

Annex: Estimates

A1. The sample

The sample consists of a panel of 130 economies with annual observations from 1996 to 2024, where 2024 is the last year for which indicators are consistently available. This produces a total of 3,097 country-year observations. Note that this sample reflects a number of exclusions. I have excluded economies with fewer than one million people because in small states we often find that a single transaction can move the FDI ratio by several points. I have also removed offshore financial centers so that contamination from phantom investments routed through special-purpose entities can be avoided. Finally, I have excluded outliers by trimming the 1st and 99th percentiles of the dependent variable. The rationale here was to limit the influence of a handful of natural resource and privatization booms.

Across the panel that I obtained this way, net FDI inflows averaged 3.5 percent of GDP. Note that this is already much larger than Sri Lanka’s average of 1.1 percent over the same years. To better understand why Sri Lanka deviates this way, I try a number of explanatory variables which are standard in the literature: market size (log GDP), income level (log GDP per capita), growth, trade openness, inflation, government effectiveness from the Worldwide Governance Indicators, and the size of surrounding markets, where the latter sums every other economy’s GDP discounted by bilateral distance. Note also that several conflict variables from the Uppsala Conflict Data Program are added, including a war-intensity indicator, log battle deaths and years since the last armed conflict.

Note as well that the full Sri Lanka series in the World Development Indicators database goes from 1970 to 2024 and that it produces a mean FDI ratio of 0.91 percent. The shorter series used here (1996-2024) is slightly higher at 1.14 percent, partly because of the different period and also because 8 years are excluded for missing indicators on governance (1997, 1999, 2001) and trade openness (2010 to 2014).

A2. The pooled regression

The dependent variable is net FDI inflows as a percent of GDP. The four columns in the table below show the pooled OLS specifications with year fixed effects and standard errors clustered by country. Column (1) is the baseline. Columns (2) and (3) add the conflict variables (first the war indicator alone and then the full set). Column (4) restricts the sample to the post-war period of 2010 to 2024. Note that the coefficients on the Sri Lanka dummy, which measure the FDI shortfall effect, are highlighted in red.

In the baseline the explanatory variables behave as they should. Trade openness raises FDI as does growth and governance. Log GDP enters negatively, which is standard, since larger economies tend to have proportionally smaller FDI ratios even when absolute inflows are large. The Sri Lanka dummy is -1.85 in the baseline and moves to -1.93 and -2.09 when the conflict variables are added. Note, however, that none of the conflict terms are statistically significant.5

The coefficient on log surrounding market potential is rather small and insignificant in columns (1) to (3). In the post-war sample, it turns negative and marginally significant. Taken at face value that would lower Sri Lanka’s benchmark, which is counterintuitive. The variable is however highly collinear with income and openness, so we might discount this result as being uninformative about the value of the Indian market.

A3. The gap before and after the war

The table below splits the sample in several subperiods and shows the Sri Lanka dummy coefficients that result from the regressions over these subperiods. On timing, I marked the end of the war in 2009 based on when UCDP/PRIO recorded the last armed conflict in Sri Lanka. The debt default happened in April 2022 and the IMF programme followed in March 2023, so the “pre-default” period ran 2010 to 2021.

The table and chart below show that the investment shortfall was 2.27 points of GDP during the war years and 1.06 points in the fifteen years afterwards. These results are highly significant. The chart below also shows the 95 percent confidence intervals.

The last two rows of the table allow us to test the crisis explanation. Excluding the default years leaves the post-war estimate at 1.07 rather than 1.06. Adding conflict controls to the post-war sample leaves it at 1.07. In other words, the shortfall of about one point seems therefore stable across peacetime windows, with and without the crisis, and with and without the conflict variables.

A4. The frontier

With a stochastic frontier model, one can examine whether there is evidence of a frontier in the data and also by how much a country is below that frontier. The technique works by splitting (1) the unexplained part of the regression into ordinary noise, spread symmetrically and measured by sigma-v, and (2) a one-sided shortfall below the frontier, measured by sigma-u. The variable gamma is the share of unexplained variation attributable to the shortfall rather than noise. Gamma is 0.89, which is very large. This suggests that there is a frontier with a very strong likelihood, as confirmed by the likelihood-ratio test.

The table also shows a line marked “efficiency.” This is the share of frontier FDI actually attained. Sri Lanka’s efficiency rating is 0.21, which ranks the country 114th of 125 economies. The sample’s mean is 0.41.

A5. Year by year

The table below shows how the benchmark evolves year by year. The baseline specification here is estimated on all countries except Sri Lanka. This removes any influence of Sri Lanka’s own observations on its benchmark. The coefficients that result are subsequently applied to Sri Lanka’s observed characteristics for each year.

In the gap column, there were 19 negative entries and 2 positive ones, in 2018 and 2022. In 2022, the benchmark was 0.83 percent of GDP, which was the lowest in the series because growth back then was minus 7.3 percent, inflation was close to 50 percent and governance was at its weakest in three decades. Actual inflows of 1.19 percent exceeded the benchmark.

Between 2007 and 2024, the benchmark fell by four points of GDP. Over half of that reflects global developments affecting FDI. The largest Sri Lanka-specific driver is trade openness which fell from 69 to 42 percent of GDP. This driver alone is responsible for about a full point.

Other factors include growth, inflation and governance which together account for only a few tenths of a point. This then also explains why the swing from crisis to recovery moves the benchmark less than one might expect. And it also relates to why the IMF’s projected slowdown to about 3 percent growth in 2026 would reduce the benchmark from 1.7 back to 1.4.

Through a PPP lens, China actually overtook the United States around 2015. The precise year depends on the source: 2014 based on the Maddison Project constant PPP series and 2016 based on the IMF's current PPPs.2 Today, China's economy is estimated at 1.4 times the size of the US economy, and the gap is expected to continue widening.

The above chart looks at China-US comparisons in four ways: aggregate versus per capita and, within each of these, PPP versus market prices.

• On an aggregate basis, the PPP approach shows that China crossed parity in 2016. It is expected to reach 1.5 times the US by the early 2030s, according to the IMF’s April 2026 forecasts. At market exchange rates, however, China’s GDP peaked at 77 percent of US GDP in 2021 and it has since fallen to about 64 percent.

• On a per capita basis, China remains far behind the US. In PPP terms, China is at about a third of the US and in market terms, only at about a fifth. Per capita GDP is of course the more relevant measure of living standards — although Gross National Income per capita would be even better as it accounts for net income flows across borders.3 The rest of this post will focus on just the aggregate view.

Why purchasing power parity

Market-rate GDP converts each country’s output into dollars at the prevailing exchange rate. This is the relevant conversion to capture things that are actually transacted in dollars. Think about imports, external debt, cross-border investment and, more broadly, the financial weight of an economy.

But there are two distortions that create issues when it is our purpose is to compare the productive size of economies.

The first distortion is currency movement. When the dollar strengthens, other economies shrink in dollar terms. For example, China’s dollar GDP was nearly flat between 2021 and 2024 at a time when its real output grew at a healthy 5 percent a year. This was entirely due to yuan depreciation relative to a strong dollar.

The second distortion is the difference in price levels across economies. The market exchange rate does not tell us how much goods and services a dollar buys inside China. A haircut or a taxi ride, eg, are mostly determined by domestic wages and prices and these are much lower in China than in the US.4

PPP removes these two distortions by valuing output at a common set of international prices.5 What remains therefore is just the volume of goods and services actually produced and this is why we say that this number reflects well the true physical scale of an economy. It captures well a country’s real output, the size of its industrial base and also the extent of the its infrastructural footprint. And it also captures well the domestic resource base that is available to the state, since governments spend most of their budgets at home in domestic prices and on domestic labour.

The share of China and the US in global GDP changes quite drastically when we change the yardstick as the chart above shows. Shifting to PPP not only lifts China, it also raises the share of India from 3.3 to 8.2 percent. And it also moves up Russia, Indonesia and Brazil. Conversely, Germany, Japan, the UK and France all move down. Higher-price higher-income economies look larger at market exchange rates whereas lower-price lower-income economies look smaller. That is the pattern.

Two centuries

The chart below takes the longer-term perspective of the last two centuries.

Note how China dominated the global economy at the start of the chart’s series in 1820 when it was 12.6 times larger than the US economy. It was only in 1887 that the US took over. By 1961, the US had achieved the widest lead of around 5.6x relative to China. By the late 1970s, following China’s opening up, the country embarked on a growth acceleration that it managed to sustain for over four decades. The maddison series tells us that China overtook the US in 2014 and that by 2022 (the latest year in the dataset) China’s economy was 1.4 times the size of the US economy.

Note the contrast with market exchange rates in the year 2022. In PPP terms, China was 1.4 times the US. But using market rates, the exact opposite was true: the US was 1.4 times China.

The league table

We can also look at how the global Top Five has evolved over this long period of time.

The chart shows snapshots across the decades from 1820 to 2022. See how the US commanded the top spot for the longest period within this sample. This is slightly unfair to China as extending the sample further back would overturn that statement. Even so, China has made its come-back over the last few decades. Note also how India was in the Top Five in the first half of this period, then vanished and recently came back. The UK was a fixture in this table until it dropped out after 1960.

Conclusion

So if we interpret the question in the title as “who projects the most financial weight across borders,” then the US is still the answer, and by a wide margin.

But if the question is interpreted — as I do in this post — as “which economy produces the most, which has the largest industrial base, or which government commands the most domestic purchasing power over real resources,” then the answer changed around 2015:

China became the world’s largest economy roughly a decade ago.

My answer is that it will not be easier. Not unless we reframe development itself: not by drawing up yet another list of goals, but by changing how we see development, how we imagine it, and how we shape it.

The alternative is the loop we are already in, where each round of aspirations is set high, missed, and reset. The SDG scorecard says only 18 percent of targets are on track.

Reframing is how we break the loop rather than restart it. This post makes the case in three parts.

Subscribe now

Seeing development

The first part is about seeing that three things about the state of global development are true at once.

Long-term progress

Over the long course of history, every world region has seen sustained gains in income per person, as the Maddison Project Database shows. Literacy rates have climbed and child mortality has collapsed across all developing regions as the World Development Indicators show. The Human Development Index has risen almost everywhere.

A large unfinished agenda

But the unfinished agenda is enormous. Nine in ten people still live in regions where income per person is below 40 percent of the frontier, represented here by the so-called “Western offshoots” of the United States, Canada, Australia and New Zealand.1

Asia-Pacific alone spans the world’s widest range of gross national income per person, from Singapore to Afghanistan. The spread in human development also remains very wide. As the chart below illustrates, the Human Development Index shows very large variations between income groups (horizontal bars) and within them (vertical beeswarms), let alone within individual countries (not shown here).

And the Sustainable Development Goals, the world’s development scorecards, tell us that our collective aspirations remain far unmet: only 18 percent of targets are on track, while more than a third are stagnating or regressing.

A more turbulent development landscape

At the same time, the development landscape is getting more turbulent. In the World Economic Forum’s latest risk survey, 62 percent of respondents expect the next decade to be stormy or turbulent. Only 8 percent expect it to be stable or calm.

The risks cluster into three groups: headwinds to growth and jobs (debt distress, demographic shifts, technological change, slower globalization), existential threats (climate, biodiversity, pandemics, antimicrobial resistance), and growing governance risks (democratic backsliding, state capture, polarization, disinformation). This is the risk landscape we mapped in the 2024 Regional Human Development Report.

Seeing development clearly means refusing to choose only one of these stories.

Imagining development

The second part is to look at the forces reshaping the world without the reflex of crisis. Crisis language narrows our vision. Opportunity thinking keeps more possibilities open. We need more imagination to envision opportunity.

The big picture

One generally overlooked news item is that in 2025 the world is likely to have reached Peak Middle Income. Middle-income countries are home to about 6 billion people today, and that number is now set to fall as countries graduate upward, to around 5.4 billion by 2050.

Underneath this lies what I have called the Great Income Inversion. In 1990, seven in ten people lived in the poorer half of the World Bank’s classification. By 2050, seven in ten will live in the richer half. The low-income share of world population is projected to fall from 58 percent to 7 percent.

Countries do not climb at the same pace. China moved from low income to the doorstep of high income in four decades, and will likely cross that high-income threshold in 2026. India is on a slower path, reaching upper-middle status around 2036, roughly three times slower than China over the comparable stretch.

The world getting richer, at least on average and into the longer term, will present opportunity for those who reap it.

But what about the disruptors? I would like to argue that as serious as many of them may be, each of them can be cast under a different light. I will illustrate this by providing examples of three megatrends that need a reframe

Demography

In addition to Peak Middle Income, the world is reaching also another peak: Peak Youth. The under-25 population will plateau around 2030 while the 60-plus population is projected to triple by 2100 (UN World Population Prospects 2024). The typology of pre-, early-, late- and post-dividend countries that we developed in the Global Monitoring Report 2015/16 still holds and we continue to see enormous diversity in patterns of demographic change around the world.

One reason why we need to reimagine demographics is that our instinct is to treat population aging as a burden. Yet, healthy aging and policies that expand labor supply can offset much of the demographic drag on growth. We need a new lens on aging itself. Longer lives with vitality and purpose, plus a growing health and care economy, are good things, not merely problems to manage.

It is time we stop treating demography as an economic funeral.

Technology

As argued in our recent The Next Great Divergence report on AI and widening inequality between countries, we have a long history of framing technology as hysteria and hype. Some 2,500 years ago, Socrates feared writing would ruin our memory. More recently, Edison confidently predicted film would make books obsolete. Both fears turned out to be wrong, although perhaps not entirely.

Today we are living in an era of unprecedented rapid technological change with AI taking currently center stage. But the global conversation on AI needs more balance and humility. It also needs a broader perspective and it needs greater attention to inequality between countries rather than only within them.

Furthermore, when it comes to the fear that technology destroys employment, history suggests we should place at least as much emphasis on whether demand expands as on the labor-saving effects of technology itself.

Globalization

It is not correct to speak of deglobalization. Trade as a share of GDP remains high almost everywhere, and further integration remains possible, including also in markets and market segments that have seen less globalization (such as labor markets and trade in services).

What we see is disglobalization: a slowdown in momentum, or a slowing of the pace of economic integration across borders, not a retreat in levels.

Meanwhile, the scale of opportunity keeps rising in absolute terms. China, a major driver of global growth over the last decades, has developed into an economy of about 20 trillion dollars as of 2025. Thanks to this much-larger base, slower growth still continues to add substantially to global opportunity.

Even now, China still adds a current Korea to the global economy every three years.

Shaping development

The third part is about what to do. Three facts set the stage. They are followed by three necessary redirections.

Three facts

People feel economically insecure. In the World Values Survey, large majorities of the adult population in Pakistan, Viet Nam, Myanmar, the Philippines, Indonesia and Malaysia say they worry “very much” about losing or not finding work. Any development model that does not put jobs at the center will fail the people it is meant to serve.

Growth may not be sufficient, but it is absolutely necessary.2 There is no country in the world that has managed to achieve high levels of material wellbeing without sustained growth over an extended period of time. And if we look at the sheer range of the income classification, most of the world still has a long distance to travel.

Slow and steady wins the race. The historical record since 1950 shows that rich countries do not necessarily grow faster than poor ones in good years. They tend to shrink less often and less deeply in bad ones. Predictable policy that avoids reversals is worth more than bursts of speed.

Three redirections

From these facts, among others, follow three redirections that any future development model will need to satisfy for it to fully serve the purpose of humanity:

• Place people at the heart of the developmen paradigm. Human development concepts are well developed but poorly mainstreamed. A systemic reset would enlarge people’s choices, build stronger human security through social protection, health, risk-informed development and food security, and meet obligations to future generations through the energy transition, climate resilience, nature protection and sound public finances.

• Pursue smarter forms of growth. The call for growth should be louder, not quieter. But not jobless, ruthless, voiceless, rootless or futureless growth. Export-led strategies need a change of tack. Domestic markets deserve more attention. New sources of growth lie in agriculture, services, the green, blue and care economies, and technology directed towards development outcomes.

• Emphasize, above all, the politics of reform and the art of delivery. The most common objection to any of this is “if it hasn’t happened in the past, why would it in the future?” This is why the third redirection is the most important. We need to shift emphasis from diagnostics and laundry lists of technocratic solutions to the politics of reform and the art of delivery. Real change is and always has been rooted in leadership and governance. Neither is fixed. Political will can be nurtured. And governance can be optimized to become more anticipatory, adaptable and agile, three attributes that often determine whether countries succeed or fail.3

Where this leaves us

Development seen whole is a story of considerable long-term progress, a still very large unfinished agenda, and an increasingly more turbulent development landscape. Development imagined differently without the polycrisis reflex helps us see demographic, technological and trade shifts, among others, as opportunities amidst disruption. And development shaped around people, smarter growth and a laser sharp focus on making change actually happen will give us a fighting chance of breaking the cycle of missed development goals.

The Moving Escalator is a series about four decades of movement through the World Bank’s country income classification, where the world is heading by 2050, and whether the classification itself remains fit for purpose. This is Post 5.

If you were to treat the world as one country, that country is nowcast to cross the World Bank’s high-income threshold in 2026. In other words, the planet is about to become rich.1

For the entire history of the income classification, going back to 1987, the world’s average per capita income corresponded to the upper-middle-income band. But no more: 2026 is likely to be the world’s “graduation year.”2

Is that worth celebrating?

Not really. And that’s the point of this post: the average GNI per capita figure is spread over eight billion people and tells us little about the world most people actually live in. Two in three people live in countries with per capita income levels below that global average.

The distribution-blind income classification of the World Bank can only detect one type of skew behind this global average: the differences in per capita income across countries. Inequality within countries is a second source of skew, which I will discuss in the next post. And that turns out to be far enough: the span between the richest and poorest country on earth is 600x — much larger than what can be observed within any country.

The world turns rich long before the median person’s country does

So let’s examine between-country inequality in per capita income levels. The previous chart contains a second line: GNI per capita in the median person’s country.3

We can see that while the world average is high income, the median person’s country is upper-middle income. Median per capita income saw a considerable rise until the early 2010s. It crossed the lower-middle and upper-middle income lines in 1998 and 2010 respectively, when in both cases China represented the median person’s country. This later shifted to Indonesia as China continued to grow.

Now extend both lines forward. The average crosses the high-income line in 2026. The median line does not reach it even by 2050, the end of the projection horizon. The average will have “graduated” but the median will take considerably longer.

Today the distance between mean and median is about the width of an entire income bucket and the chart below shows how the ratio between mean and median, a measure of skewness, has evolved over time.

Three results to note:

• Between-country inequality declined massively from 1991 to 2012. The mean-to-median ratio fell almost sevenfold, from 12.7x to 1.9x. China's rise through the middle of the distribution accounted for much of the decline.

• The ratio rebounded after 2015. The collapse in commodity prices reduced the dollar-denominated GNI per capita of many commodity exporters.

• The ratio now stands at around 3x and is projected to remain broadly flat through 2050.

If the world were a country, it would be a very strange one

To further illustrate the extent of between-country inequality today, let’s take a deeper look at the 2026 nowcast and treat the world as one country with an internal income distribution.

Half of the global population lives in countries which capture just 9% of world income. Conversely, the richest tenth of the global population, ranked by GNI per capita of the country they live in, captures 49% of world income. The world average, nowcast at $14,801 for 2026, is at the 64th percentile of the ranked distribution.

Being average is therefore quite clearly an “above-median experience”: 64% of humanity, 5.2 billion people, live in a country poorer than the world average. And together they hold just 17% of world GNI in aggregate terms.

The 5.2 billion below the average are far from homogeneous. Per capita income levels vary widely across the countries they live in.

Finally, let’s zoom in further to the country level to see the full range of between-country inequality. Among 197 ranked economies, Burundi is the poorest and Bermuda the richest, with a roughly 600-fold gap between them.

Think of what this means for our “newly minted high-income country” called World. No real country comes close to an internal division of this 600-fold magnitude. Across the world’s large economies, the gap between the richest and poorest region runs from about 2-fold in Germany and 5-fold in the United States, through roughly 10-to-15-fold in India and Indonesia, to the world record: the 43-fold gap between Russia’s Arctic oil districts and Ingushetia.4

A graduation nobody will notice

When a real country “graduates” to high income, things change. Concessional finance may be affected. Aid eligibility is likely to narrow. Trade preferences will be less generous.

So what kind of club is the world joining?

If we draw the four income bands in plain dollars and look for the poorer world, it is almost a struggle to find it. The low, lower-middle and upper-middle bands together cover less than a tenth of the axis. We squeeze into that sliver every country below the high-income line and with them the large majority of humanity. The high-income band takes the rest of the picture, and it has no ceiling.

The dollar chart is really what allows people to declare with a straight face that the world has become “rich.” And that’s precisely possible because the poorer world barely features in this visualization.

What if we “bend” the ruler?

Take a log scale and the bottom stretches out and the top compresses. Logs are also more natural for measuring the effort it takes to travel across the bands as equal distances represent equal growth efforts, at least arithmetically. But above all, with logs we can see the sheer diversity within the group below the line.

The bottom line is that the world crosses this year at the very bottom of the high-income band. Above this doorstep, inequality continues to run on and the classification stops measuring the differences at the top because the highest category is the “terminal” one — which is in and of itself a problem as the top is becoming too crowded and too heterogeneous.

Below the doorstep lies the part of the distribution that is barely visible on the dollar scale: the countries where two in three people live, whose incomes are a full income bucket or more below where the world now officially stands.

So the world turns rich this year. On average. But most of the world does not.

The Moving Escalator is a series on four decades of movement through the World Bank's country income classification, where the world is heading by 2050, and whether the classification remains fit for purpose. This is Post 4.

Two years ago in Washington DC, at the IMF/World Bank Annual Meetings, I heard the state of the global economy get diagnosed as:

• Low-income countries are in a poverty trap;

• Middle-income countries are in a middle-income trap;

• High-income countries face secular stagnation.

Every income bracket gets its own affliction and we can fit the whole world on one slide.1 I remember thinking that this is too neat.

The above speaks to a broader issue. Development loves its trios. There are countless examples of them: three pillars, three pathways, three solutions… And yes, this post has three examples.

There are good reasons for this. Simple frameworks are easy to remember. After all, what good is a complicated framework if nobody remembers or understands? A simple framework also gives a messy subject some structure, which is especially appreciated in the chaotic world of international development. By abstracting from the detail, we sometimes see the underlying patterns better.

That’s the idea. But simplicity can slide into “simplism.” We may oversimplify and remove the complexity that matters. Our frameworks can become straitjackets: rather than simply helping us see, they start telling us what we should see. This happens surprisingly often: a diagnosis developed for one country at one moment in time becomes a prescription for all countries at all times.

I will illustrate this with three well-known frameworks, all tied in different ways to the World Bank’s income classification. The first is the trio of traps as in the above. The second is “Grow, Invest, Insure”, the game plan to end extreme poverty. And the third is “Invest, Infuse, Innovate”, the 3i framework of the World Development Report 2024.

To set the tone right: all three contain useful ideas. Where they go wrong, in similar ways, is when those ideas become too tightly tied to the income classification. A statistical classification should not start doing the work of a development theory.

Each income bracket is in a different trap

Three afflictions in the form of a trap, one per income bracket. What could be wrong with that?

First, it is too tidy. In 2025, the income classification contained 87 high-income, 106 middle-income and 25 low-income countries. That is a large number of countries being asked to share a diagnosis. The middle-income group, for example, spans radically different economic structures, demographics, institutions, and geopolitical circumstances. Yet, by crossing a common and rather arbitrary threshold in GNI per capita, such countries all acquire the same predicament.

Second, it is too fatalistic. I have argued in a previous post (see below) that the middle-income trap should be retired for its conceptual flaws, lack of empirical basis, and harmful side effects. Countries do climb the income ladder at different speeds, and some take ages to cross the range of the income group they’re in. But a slow climb is a policy challenge, not a trap. The question is why progress is slow and what needs to change. The same applies at the bottom and the top: low-income countries face productivity and structural transformation challenges, while high-income countries face their own problems with productivity, demographics and fiscal pressures. Above all, this language of traps risks taking agency out of the equation precisely where agency matters most.

Third, it is too narrow. The entire framework is anchored to GNI per capita. This is useful for telling us roughly how much income an economy generates, on average, but that’s it. It says nothing about inequality, institutions, learning, security, women’s opportunities, service delivery, or sustainability. I would not dismiss income as irrelevant. It matters greatly, and it also correlates with many dimensions of development. But correlation is not the same as coverage. An income average is just part of the picture. It is not the full picture itself.

Social protection on the back burner for poor countries

The second framework is the World Bank’s three-pillar gameplan to end extreme poverty: “grow, invest, insure.” I cannot claim distance from this one as it featured prominently in a report I managed some years ago. The idea is to:

• Sustain broad-based growth that creates competitive economies and income-earning opportunities for the poor and the bottom 40 percent;

• Invest in human capital through health, education and early childhood development, so that vulnerable groups are able to participate in and benefit from growth;

• Protect the poor through social assistance, safety nets and formal insurance, so that shocks do not force desperate choices with irreversible consequences,

Who can argue with this?

The problem is what happened with the third pillar.

For decades, social protection was treated as a supplement, especially in poorer countries. The 1980 World Development Report noted that the scope for such programs in low-income countries was “necessarily limited.” As Gill, Revenga and Zeballos (2016) show, this evolved into the “two-and-a-half-point strategy,” with public assistance as the half pillar.

Only in 2016 did the World Bank formally make insurance a full pillar. Even then, the promotion came with conditions attached: countries were advised to expand social insurance only once extreme poverty had been reduced to low levels. For India, it was said that the time had not yet come.

Is it right to present social protection as a “graduation prize” that one earns by growing first?

And the old sequencing has not entirely disappeared today. Universal social protection is still to be approached “gradually,” as fiscal constraints allow, with affordability and targeting continuing to shape what poorer countries are expected to provide.2 The half pillar was retired in principle but it looks like it was kept on a retainer.

Also, the emphasis has recently shifted toward jobs. But jobs cannot substitute for protection against sickness, unemployment, old age and other shocks.

We should stop looking at social protection as merely a fiscal or technical issue. It is a human right and a foundation for inclusive development. It cushions shocks for those who work, supports those who cannot, enables families to invest in health and education, and builds resilience across generations.

Growth and jobs alone will not achieve lasting poverty reduction unless anchored in systems that protect people through life’s inevitable risks.

Don’t prioritize innovation when poor

My third example comes from the 2024 World Development Report and its answer to the imaginary middle-income trap: the 3i framework of “Invest, Infuse, Innovate.” It goes like this:

• Low-income countries should invest, accelerating the accumulation of physical capital and basic human capital to build foundational capabilities.

• Lower-middle-income countries should invest and infuse, complementing investment with the adoption and diffusion of technologies and business models from more advanced economies, through trade, FDI and integration into global value chains.

• Upper-middle-income countries should invest, infuse and innovate, pushing domestic firms toward the global knowledge frontier once the gains from imitation are exhausted.

As a way of explaining how the sources of growth might change as economies develop, this is useful. The problem starts when the framework is read as a sequence prescribed by income level. Should countries really wait for their income classification to change before moving on to the next "i" they need to prioritize?

Take India. It is a lower-middle-income country that has globally competitive firms and capabilities in software, pharmaceuticals and space. Would we really want to tell those firms to focus on infusion rather than innovation because the country's average income has not yet crossed a particular threshold?

Or take a low-income country such as Afghanistan. Why should a poor country wait before adopting technologies that offer the possibility of leapfrogging, whether in mobile payments, energy or digital services?

In reality, development does not follow a linear path. The most successful countries that climb up the income classification do not add these ingredients one bracket at a time. They combine them in different proportions, depending on what their economies need at different stages and in different sectors.

That is why I would use the 3i framework as a menu, not a ladder. The balance between investment, infusion and innovation should depend on the country's capabilities, sectors, institutions and opportunities, not simply on where its GNI per capita happens to fall.

Read before use

The simple trios add value and the intent is constructive. But each risks turning an income bracket into a diagnosis, and a diagnosis into a prescriptive destiny.

The trap framework files every country under the affliction of the income group to which it belongs. Grow, Invest, Insure has for too long treated social protection as a secondary concern for countries further down the income ladder. And the Invest, Infuse, Innovate framework converts a useful description of how growth strategies evolve into a prescription about what countries at a given income level should and should not prioritize.

We should stop the use and abuse of the income classification as a theory of development. The classification merely sorts countries and it does so on the basis of just one indicator and a narrow one: income per capita. It does not diagnose, and it certainly need not prescribe.

So when we use the simple trios, let’s attach a warning label: “Consume responsibly. The contents may oversimplify.”

The Moving Escalator is a series on four decades of movement through the World Bank’s country income classification, where the world is heading by 2050, and whether the classification remains fit for purpose. This is Post 3.

We often say countries risk “getting old before they get rich.” This post is about a development concept that got old before it got valid: the middle-income trap.

The full name is the middle-income country trap, or the MIC trap. The idea is that countries get stuck in their climb from middle to high-income levels, as defined by the thresholds of the World Bank’s country income classification. They get stuck because the policies and institutions that drove earlier growth no longer work as well, while the next set of growth drivers has yet to take hold.

There is a real problem here, but I don’t think “middle-income trap” is the right way to describe it. The concept is not conceptually sound and neither does the empirical evidence convincingly establish a distinctive middle-income trap.

Middle-income countries do face genuine policy challenges in sustaining and accelerating growth. But those challenges are not unique to middle-income countries, nor do they imply a common barrier at the middle of the country income distribution.

It is time to retire the Middle-Income Trap.

I will be make the case in five steps. First, a critique of the concept itself. Second, a look at how countries actually climb the middle-income ladder. This section is a little heavier on the data, so bear with me. Third, a review of the empirical literature on whether a middle-income trap exists. Fourth, a projection of what the global country income distribution might look like by 2050. Finally, an examination of the potential harm the concept itself might do, particularly when a contested idea becomes conventional wisdom.

A concept with two typos in its name

Before turning to the data, first consider the concept on its own terms.

There are two typos in the name.

• “Trap.” A trap implies a distinctive barrier from which countries struggle to escape. But the literature does not establish such a barrier. What middle-income countries face is a development problem that is pretty familiar: growth slows when existing sources of productivity run out, the low-hanging fruits of reform have been reaped, and new growth opportunities need to be found. Countries differ enormously in how they respond. This is no trap, but a policy challenge.

• “Middle-income.” The need to renew growth drivers does not stop at the high-income threshold nor does it exclude the low-income countries. Rich countries exhaust old growth models as anyone following the productivity debates in Europe or Japan can attest. And I can think of many growth challenges that apply to low-income countries too, not least those related to managing their demographic transition well. All countries need to adjust their growth model at different points in their development journey.

The originators of the concept have conceded both points. Revisiting their 2007 creation in “The Middle-Income Trap Turns Ten”, Indermit Gill and Homi Kharas write that the term refers to “a trap that can catch middle-income countries” and is “not a statement that middle-income countries are more likely to be trapped than other countries.” They add: “In retrospect, it would have been helpful to clarify this.”

It would have.

A decade later, the World Development Report 2024 put the trap on its cover anyway.

But by then, the concept was already deeply embedded in the language of politicians, policymakers and pundits, and had generated thousands of research papers.

Countries climb at different speeds

What do actual income transitions look like?

Below I examine the two legs of the middle-income climb: from lower-middle to upper-middle income (LMIC to UMIC), and from upper-middle to high income (UMIC to HIC). For each leg, I distinguish between economies that have completed the crossing and those still under way. I restrict the sample to economies with at least 10 million inhabitants.1

As a preview, the charts show no common wall at the middle-income level. Countries cross the thresholds at very different speeds, and their paths are shaped by growth, commodity cycles, financial crises, exchange-rate movements, conflict and other country-specific shocks.

From LMIC to UMIC

Let’s start with the completed runs for the first leg.

Azerbaijan completed the run fastest, in just six years, helped by an oil boom. China followed a steep path too, taking 13 years if we start from when it first became an LMIC in 1997.2 Viet Nam, the median performer in this sample, took 16 years. It climbed fastest in its first five years and then eased gradually.

Indonesia needed 29 years and was the slowest of the completed runs in this sample: its climb was interrupted early by the Asian financial crisis before resuming. And its final decade covered just the last tenth of the distance.

Sri Lanka spent about five years close to the line (censored at 100) without being reclassified. It eventually entered the UMIC group in 2025. Ukraine had a path that resembles a sawtooth of war, recovery and renewed shocks.

The ongoing runs tell a complementary story.

Eighteen economies are currently on this leg. Cambodia, Bangladesh and Uzbekistan track close to the median completed run (thin gray line). Egypt and Honduras have been in the bracket for a quarter century, with their progress repeatedly interrupted. In Egypt's case, devaluations erased years of gains measured in Atlas dollars.

Sudan, Nigeria and Zambia have lately moved away from the line. Notice how Senegal and Cameroon are almost flat from the year of entry, remaining far below the threshold.

The slow climbers were slow from the start. Their difficulty looks less like hitting a wall at a particular income level and more like persistently weak growth, compounded by currency and conflict shocks.

From UMIC to HIC

The upper transition is the one most often described as the harder leg.

Romania is the median performer. It completed the UMIC-to-HIC crossing in 16 years, which is the same number of years as the median run on the lower leg. Russia completed the run in just 8 years. Then came Czechia in 12, Poland in 13, and Chile in 19. Argentina took a total of 23 years, crossing after a collapse followed by a long recovery. Venezuela reached the line in 2014 before falling away again.

The upper leg spans a smaller income multiple than the lower leg: the ratio between the HIC and UMIC thresholds is 3.1, compared with 3.9 between the UMIC and LMIC thresholds. It should also be noted that those entering the upper-leg run seem to be entering closer to their target line.3

So if countries complete the two runs in the same number of years, growth must have been faster during the lower-income transition. But this does not point to a middle-income trap. It is consistent with the broader pattern in which growth tends to moderate as income rises, including the regression to the mean documented by Pritchett and Summers (2014).

Among ongoing runs, note the stellar performance by China, which is about to cross the line after 15 years. Kazakhstan did well too and is now at 95 percent. Türkiye has touched the threshold twice and is on the cusp of reclassification.

Malaysia is 33 years in and has grown steadily albeit more slowly over the past decade, reaching about 85 percent. Some countries took even longer: Mexico (35 years), Brazil (36) and South Africa (37). These are among the countries most often cited as evidence of a middle-income trap.

Brazil benefited from buoyant commodity prices during 2011-14 and went in reverse later on when the boom unwound. Mexico seems to have oscillated between 60 and 90 percent of the threshold for three decades. South Africa has moved sideways since 1988.

Jordan and Tunisia have followed relatively flat paths from the time they entered the bracket, well below the HIC line.

These look like stories of persistently slow growth, commodity cycles, crises and policy constraints. The same forces appear below the middle-income thresholds and among slow-growing high-income economies.

The middle of the income ladder has no monopoly on stagnation.

What the evidence says

The empirical literature has looked for a middle-income trap for more than a decade. The results do not establish a distinctive trap.

The findings are worth reading carefully because some of the papers often cited as evidence for the trap say something narrower.

• Eichengreen, Park & Shin (2011) and (2013) are often cited as evidence for the trap. But their result is about growth slowdowns, not a uniquely middle-income barrier. Slowdowns cluster around US$10-11,000 and US$15-16,000 in 2005 PPP, with substantial dispersion. The authors themselves stress that “there is no iron law of slowdowns.” Multiple slowdowns at different income levels are a poor fit for the idea of a single middle-income wall.

• Felipe, Abdon & Kumar (2012) define being “trapped” using median transition times of 28 years in lower-middle income and 14 years in upper-middle income. They acknowledge the thresholds are arbitrary. Their follow-up paper, “Middle-Income Transitions: Trap or Myth?” (2017), leaves the question open.

• Pritchett & Summers (2014) conclude: “regression to the mean is the empirically most salient feature of economic growth.” They find this pattern to be much more robust than the middle-income trap. In their view, treating growth slowdowns as evidence of a trap risks misreading ordinary growth dynamics as a special middle-income phenomenon.

• Im & Rosenblatt (2015) likewise conclude that “transition matrix analysis provides little support for the idea of a middle-income trap.”

• Bulman, Eden & Nguyen (2016) find “no evidence for (unusual) stagnation at any particular middle income level.” They do find that the determinants of growth differ by income level, meaning strategies need to evolve. But changing growth strategies is not the same as being caught in a middle-income trap.”

• Ye & Robertson (2016) take a different approach, using time-series tests for unit roots and structural breaks. They find only around seven countries meeting the conditions for a genuine trap. If a trap exists, their results suggest it is the exception rather than the defining feature of middle-income development.

• Han & Wei (2017) use transition matrices and find that the data “clearly reject” the idea that middle-income economies have either a high absolute probability of being stuck or a higher relative probability than low- or high-income groups. Their results suggest middle-income countries reach “escape velocity” more often than poorer or richer ones.

• Alonso & Ocampo (2020) assemble the fullest book-length treatment in Trapped in the Middle? (Oxford University Press). The contributors find that few middle-income countries have completed the transit to high income in recent decades, yet the volume traces this to heterogeneous structural and institutional constraints rather than a single barrier at a common income level. The diagnosis points to many country-specific problems.

• Patel, Sandefur & Subramanian (2021) document in the Journal of Development Economics that unconditional divergence, the central fact of growth empirics, has been reversed since the mid-1990s, with poorer countries catching up with richer ones. They call the trap debate "anachronistic" for a simple reason. Middle-income countries have posted higher growth rates than all other groups since the mid-1980s. A group that outgrows everyone else makes for an odd trap.

• Kremer, Willis & You (2022) reach a similar verdict in the NBER Macroeconomics Annual. They find a trend toward unconditional convergence since the mid-1980s, culminating in absolute convergence since 2000, driven partly by convergence toward development-favored policies across countries. Convergence is the opposite of entrapment. Their debate with Acemoglu and Molina concerns the strength of convergence, and neither side defends a middle-income wall.

• Imam & Temple (2024) offer the most sympathetic recent treatment, using Markov chains and mean passage times in an IMF working paper since published in the Scandinavian Journal of Economics (2026). Even they concede there is little evidence that transitions from a middle-income category are uniquely difficult in the income data. Their candidate trap is in the lack of upward mobility in relative TFP, and lies mostly in the future.

Peak Middle Income and beyond

If the trap cannot be found in forty years of data, the natural next question is what happens when we stop looking for it.

Projecting 2050 conservatively

The projections in this section are built on that premise. Each of 197 economies is projected forward using the median of its own annual growth of GNI per capita, in US$, Atlas terms, over the decade to 2025. And the classification thresholds drift upward at 1.244 percent a year, the median pace over the same decade. Full details are provided in an earlier post.4

The choice of median growth over the last decade is conservative for two reasons. First, the median dampens the influence of extremes, such as a devaluation or a one-off commodity spike. The influence of the COVID-19 pandemic is hereby mitigated. Second, and less obviously, 2016-2025 was an unusually weak decade for GNI per capita growth in middle-income countries.

The chart shows how weak. The typical middle-income economy grew 3.6 percent a year over 2016–25, down from 5.9 percent in the exceptional decade of 2006–15 and lower than in every earlier decade in the series. Weighted by population, the drop is steeper still: from 9.3 to 5.0 percent. Only 37 percent of middle-income economies gained ground on the United States over the seed decade, down from 91 percent in the decade before.

So the projection is not an optimistic scenario. It assumes that a disappointing decade repeats itself.

Nor do I build any trap-like slowdown into the projections as economies approach an arbitrary threshold. The evidence reviewed above gives no basis for doing so. Transition-matrix and duration studies found that exit probabilities do not fall simply because a country has spent longer in an income group. Fast growers tend to remain fast and slow growers tend to remain slow, with no distinctive slowdown appearing in the middle.

There are, of course, caveats. This is still a mechanical benchmark, not a fully-fledged forecast built from the ground up based on country detail. It might overestimate or underestimate future growth, and the further out a crossing lies, the larger the margin of error.

But the question that the exercise addresses is actually quite simple: what happens if recent growth patterns persist, including whatever slowdowns are already embedded in them?

The answer is a large movement out of upper-middle income over the coming decades. China crosses the high-income line, in all likelihood, in 2026. But there are many other factors contributing. And there is a twist related to demography I will get back to at the end

Peak Middle Income by population

What does the constant-pace scenario imply for the way the world’s population is distributed across the classification?

Let’s start with the group at the center of the middle-income trap debate because this group has peaked (or is about to).

In headcount terms, the middle-income population is expected to have peaked at 6.02 billion people in 2025. This is technically a nowcast since China is projected to transition to HIC some time in 2026. China’s exit takes 1.4 billion people out of the group in one step, causing a dent that will not be recovered over the horizon to 2050. By then, the middle-income population is expected to settle at 5.40 billion.

The turning point is even clearer in population shares. Middle-income countries accounted for 75 percent of the world’s population in 2016. Today the share is 73 percent. By 2050, this is expected to fall to 56 percent.

But the middle-income story makes little sense in isolation. The rise of the middle is largely the mirror image of the decline of the bottom.

The low-income population fell from 2.82 billion in 1987 to 0.77 billion in 2025 as India, and before it China, moved upward through the classification. The high-income share of world population stayed remarkably stable for four decades, at around 16-17 percent, before China’s reclassification produces a one-off jump to roughly 37 percent. Afterward, the share settles into a new pattern.

Peak Middle Income by country count

Now switch from people to economies. How many economies occupy the middle of the classification?

By this measure, Peak Middle Income is long past.

The number of middle-income economies peaked at 110 in 1992. The spike partly reflected the Soviet, Yugoslav and Czechoslovak successor states, which entered the classification largely in the middle-income brackets.

There were 106 middle-income economies in 2025. From here, the middle is projected to empty out: 78 economies are projected to remain middle income in 2050, equivalent to 36 percent of all classified economies. At its peak, middle-income economies accounted for 54 percent of the total.

Where do the departing economies go?

They generally move upward. The high-income roster has expanded almost continuously since 1987, from 41 economies to 87 today. The projection takes the number to 124 by 2050, when 57 percent of all classified economies would be high income.

The classification therefore inverts at the country level too. In 1987, the modal classified economy was middle income and nearly one-third were low income. By 2050, the modal economy is projected to be high income, while the number of low-income economies falls from 25 to 16.

Who fills the buckets

There is a puzzle hiding in the previous two sections. If dozens of economies are joining the high-income group, why does its share of world population remain broadly stable after 2026?

The answer lies in how the buckets fill and empty.

Each income group changes size through two channels: population change among existing members, and reclassification as whole countries move between groups.

Between 2000 and 2025, India and 39 others carried 2.17 billion people out of low income into the middle. Between 2025 and 2050, China and 36 others carry 2.25 billion out of the middle into high income.

The hatched bands calculate organic population growth among economies being reclassified. They show the total effect of population growth before and after the reclassification event.

The result is that the middle-income group that remains keeps growing from within (+0.98 billion among the 69 stayers). The high-income group, however, shrinks from within once China joins: its new membership loses 0.13 billion people after reclassification.

The decomposition sharpens the argument about Peak Middle Income.

Of the 6.02 billion people in middle income today, 42 percent arrived by upward reclassification over the past quarter century. By 2050 the composition flips. Some 70 percent of its population was already there in 2025 and only 9 percent arrives from below, which reflects the fact that the low-income feeder has largely emptied.

High income sees the opposite pattern, where 63 percent of its 2050 population arrives by reclassification, and its own demographic contribution turns negative (−3 percent).

In other words, the middle increasingly becomes a club of incumbents, and the top grows through admission.

As the top bucket keeps on growing, one might ask whether the classification remains fit for purpose if its largest bucket has become the top one?

Don’t be a trappist

The word Trappist most commonly refers to a strict Roman Catholic religious order of monks and nuns, the specific style of excellent and rather strong beer brewed by them, or TRAPPIST-1, the ultra-cool dwarf star system famous for its seven Earth-sized planets.

But there is a fourth kind of Trappist: the development practitioner who practices what I call development simplism, the habit of compressing a complicated development problem into a single fatalistic label.

The previous sections have shown three things. First, the name is misleading. The “trap” is not a trap, and the problem is not specific to middle-income countries. Second, the empirical literature does not establish a middle-income trap. Third, even a deliberately conservative projection points to a world moving decisively beyond Peak Middle Income.

But there are further reasons to retire the concept.

• It breeds fatalism. “Trap” implies inevitability. That demoralizes reformers and hands underperformance an alibi, when the political economy of growth points to domestic choices, coalitions and institutions. What countries need is agency, not destiny.

• It gives arbitrary thresholds too much weight. The trap discourse hinges on the crossing of the high-income line, which has questionable origins (see my previous post) and has since its introduction in 1987 only been adjusted for inflation. The line also has consequences beyond semantics. Donors, the Global Fund and trade-preference schemes such as the US GSP tie eligibility (wholly or partly) to the income category. Cross a line, lose a lifeline!

• It risks becoming self-fulfilling. If investors, rating agencies or policymakers act on the belief a country is in a trap, that belief itself may be sufficient to dampen investment and reform momentum, making a growth stagnation all the more likely.

What to do about it

Drop the fatalism.

What the trap language describes is a policy challenge: sustaining growth as its drivers change. Every country faces this challenge, including rich ones. Middle-income countries have climbed the income ladder at different speeds for reasons that are overwhelmingly about domestic institutions, political coalitions, policy choices and shocks. Some have moved quickly. Some slowly. Some have gone backwards and recovered. There is no single trajectory.

Extend the conversation upward.

The challenge of sustaining and accelerating growth does not end at the high-income threshold. The sky above the line is open, and the distance between the bottom HIC and its richest member today is the widest stretch of the entire income classification.

Look at the two scales.

On the linear scale, LICs, LMICs and UMICs occupy a thin sliver at the left, whereas the HIC band stretches to $140,000, where Bermuda is. On the log scale, where equal widths mean equal ratios, the HIC band is still the widest of the four.

The richest HIC is 10 times above the high-income threshold, which is nearly as large as the factor of 12 that spans the entire middle-income range from the LMIC entry line to the HIC line.

In other words, the high-income line is not a finish line. That was also the argument of the previous post “All Countries Are Developing Countries”.

In sum, retire the trap. Drop the fatalism. Look up. And don’t be a Trappist!

The Moving Escalator is a series on four decades of movement through the World Bank’s country income classification, where the world is heading by 2050, and whether the classification remains fit for purpose. This is Post 2.

Every July, the World Bank sorts the world into four income groups: low, lower-middle, upper-middle, or high. The groups are set by dollar thresholds and this year’s line for high income is $14,375 in GNI per capita.

Notice what the classification does not contain. There is no group called “developed” or “developing”. The classification simply ranks countries into four income buckets.

Yet much of the public conversation still reduces this four-part classification into a crude binary. The top group is the “developed world” and everyone else is the “developing world”. All too often, high income is treated as synonymous with developed, with the other three groups collectively labeled developing.

This post reviews a decade’s news reporting from twelve major English language outlets in Hong Kong SAR (China), India, Japan, Malaysia, Nigeria, Qatar, Russia, United Kingdom and United States. It asks: have we moved on to a new vocabulary?

The answer is: we haven’t. And that matters because the old vocabulary carries a claim that the data cannot support: some countries have arrived.

My argument is the opposite. All countries are developing countries. All are chasing a moving frontier. No one has arrived. Development never graduates.

The world in four buckets

Before examining how we talk about the classification, let us see what the classification actually says.

The chart below shows how the world is currently sorted into the four groups per the FY27 classification, based on 2025 data. Each bubble represents an economy.1 Income per person is on the vertical axis, color shows the income group, and bubble size represents population. The left panel uses dollars; the right panel shows them in logarithms.

Three observations:

First, nothing here is binary. The world is spread out across the four categories. There is no split between those who are developed and those who are not. Instead, we see four groups. And, shown in a previous post, there is and has been significant movement across the classification. For example, the 2026 data will likely place China above the high-income threshold.

Second, consider the range of the categories. The lower three categories occupy a thin sliver at the bottom. The contrast with the high-income group is especially visible on the linear scale. Unlike the other three groups, the high-income category has no upper bound, although I have censored the chart at $90,000 for visualization purposes.2 If high income truly meant “developed”, such enormous variation within the category would be hard to explain.

Third, using logarithms stretches the lower end of the distribution, and that makes the differences among poorer economies easier to see. Equal vertical distances between any two pairs of countries represent equal income ratios. But even in ratio terms, the high-income band remains the widest of the four. As a result, it takes more growth to travel across the top band than across any other.

Where the line comes from

The high-income threshold goes back to a decision made nearly forty years ago. The 1989 staff report that launched the modern classification proposed $6,000 in 1987 prices as the benchmark separating middle-income from high-income economies. The level was chosen to fix anomalies in how the Bank then listed “industrial’ economies. The benchmark was designed largely to reproduce an existing club of industrial economies rather than emerge from any theory of development.3 Every threshold since has been that number adjusted for inflation.

Interestingly, the same report informed Executive Directors that “developing economies” could remain a convenient shorthand for low- and middle-income countries. It also added a caveat: membership in the group did not mean a country was actually developing. And exclusion did not mean a country had reached some preferred or final stage of development. Exactly.

In 2016, the World Bank dropped the developed versus developing distinction from its flagship data portal, the World Development Indicatos. The UN followed in December 2021, removing the “developed regions” and “developing regions” groupings from its M49 statistical standard, while acknowledging there was no agreed definition of what “developed” meant in the first place.

While pockets of the old vocabulary survive, the institutional thinking has clearly evolved and moved away from the old distinction.4

Our words haven’t changed

I pulled ten years of story-level data from the Media Cloud news archive, covering twelve major outlets and over twelve million stories across nine countries, and counted how often they use three phrases: “developing country,” “developed country,” and “high-income country” (as well as their plural versions).

As the chart below shows, across every single outlet in the sample, “developing country” dominates. “Developed country” comes second everywhere, and “high-income country,” the term most development institutions now prefer, third and barely mentioned.

Usage of the three terms was broadly stable until 2020, when all three surged during the pandemic before declining again. The surge is easy to understand. COVID-19 brought concerns about vaccine access, debt distress and recovery finance, and this pushed the rich-versus-poor framing onto every front page.

One might hope the subsequent decline reflects the possibility that the World Bank’s 2016 decision is now finally paying off, but the ratios between the terms have barely moved. A more plausible explanation is that development itself has become less central in news coverage as geopolitics and national interests have crowded out other matters. In any case, the old terms continue to dominate the newer ones.

One result surprises. Relative to “high-income country,’ the term “developed country’ is far more prevalent in non-Western outlets. One explanation may be that “high-income” is simply a piece of technical statistical vocabulary that has diffused more prominently through Western policy circles and the outlets that cover them. Another is aspiration: in much of the world, “developed” is still the destination written into national visions, from China's mid-century modernization goal to India's Viksit Bharat 2047. A vision of becoming "developed" is undoubtedly more politically compelling than becoming "high income."

Of course, I should also note that this just covers a sample of major outfits and it focuses within them on those in English - that will likely affect the results

More than semantics

While our thinking has evolved, our words, it seems, have not. That matters for two reasons that point in opposite directions across the income line.

Nobody has arrived

The issue here is not so much that “developing countries” may be more developed than “developed countries” and therefore we should not label them as such. They often do, especially when what it means to be developed is reduced to a measure as narrow and problematic as GNI per capita in current USD. After all, development is far more than just average per capita income and a more multidimensional lens would show many examples of countries below the threshold doing better on non-monetary aspects of development (health, inclusion, sustainability, …) than those further up.

The deeper problem is that “developed” implies “arrival” and the reality is that no country or economy has arrived.

Just look up from the high-income threshold. If your goal is GNI per capita, then there is a long way up. Marginal welfare gains from additional income decline, but they do not disappear. It would be difficult to argue that the welfare gains would be flat across the range shown in the first chart of this post.

And even if countries “arrive” in the high-income bucket, the crossing can be undone: 25 out of the 93 economies that have reached high income since the series began in 1987 have subsequently fallen out of it at least once.

More fundamentally, every country is chasing a moving frontier. Every society is still failing its people in their own ways whether through poverty and disease, or loneliness, burnout, and inequality. And if poverty is measured not only against today’s standards but against the progress we know is possible tomorrow, then no country is rich.

Development has no finish line, which means all countries are developing countries. Including the richest. The frontier keeps on moving under their feet. That is also the idea behind the name of this newsletter.

The label stigmatizes, in both directions

Hans Rosling made the point better than anyone. In the 1960s, sorting the world into two clusters may have been defensible: one group of countries was rich, educated, and demographically stable with long life expectancies, and the other was poor, with short lives and large families, and there was little in between. That world no longer exists, because the middle buckets have started to fill in. The Bank’s own four groups were therefore a promising upgrade, Rosling argued.

But the perception that the world could be divided into “developed” and “developing” continued to persist, as the media analysis also shows. And that does damage. It compresses very different development contexts into one big bucket. It invites a single mental image for most of humanity. It prices risk before examining reality.

Nor does the leading replacement “Global South” solve the problem.5 It basically replaces an economic label with a geographical one that is neither geographically accurate nor analytically useful. Like “developing world” it compresses immense diversity into a single category.

But the label also matters above the line. “Developed” flatters. It tells the richest societies the work is done, at the very moment their frontiers, demographic, technological, ecological, are moving fastest. Complacency is a development trap too.

What to do about it

A few practical suggestions:

• If income is meant, say “high-income” and “lower-income.” Even better use the full vocabulary of the income classification. The terms are accurate, neutral and already official. The term “developing” could be used for what it actually describes, which is every country on the escalator.

• Newsrooms retired plenty of loaded vocabulary by writing the change into their stylebooks. More advocacy to help retire the old vocabulary would be useful.

• Consider subgroups for the high-income ladder bucket. Part of why “developed” survives is that the classification itself implies a finish line: one unbounded final band. Adding bands above the current threshold would visualize the remaining aspirations better.

The Moving Escalator is a series about four decades of movement through the World Bank’s country income classification, where the world is heading by 2050, and whether the classification itself remains fit for purpose. This is Post 1.

If you switch on the news and see nothing but crises, zoom out and take a longer-term view.

One way to do so is to look through the lens of the World Bank’s country income classification. Prompted by its annual update in July, I took a fresh look at how countries have moved across its four income groups — low, lower-middle, upper-middle, and high income — and projected what the world of 2050 might look like through this lens.

The classification is based on per capita Gross National Income. That makes it a narrow lens on development, since development is about much more than income, let alone average income. Yet it remains one of the most influential yardsticks in international development and a useful starting point for understanding long-term change.

The three charts below trace the movement of countries through the classification from 1990 to 2050. The projections extrapolate each country’s median growth over the past decade. I explain the method in more detail at the end.

The 70% inversion

The first chart shows how the global population is distributed across country income groups: what share of humanity lives in each of the four groups, and how that distribution shifts over six decades. Note that the labeled bands name the largest country in each flow, and each carries a group of others.1

The most striking result is the inversion of the global distribution of people across country income groups. In 1990, the world was heavily concentrated in the poorer half of the classification. Most people lived in countries classified as low income (red) or lower-middle income (orange). Together, these two groups accounted for 71% of humanity.

By 2050, the projections suggest that 71% of the global population will live in the richer half, in countries classified as upper-middle income (green) or high income (blue). The distribution will have completely inverted.2

That these figures are almost identical is a coincidence (the difference emerges only after the third decimal). The inversion however is no coincidence. The country income group distribution is on track to flip upside down. This is one of the most remarkable development stories of our time.

The scale of transformation

The second chart tells the same story in absolute numbers, tracing the number of people, in billions, living in each country income group between 1990 and 2050. This perspective reveals several additional patterns.

The first is the significance of population change. The world’s population is expected to almost double: from 5.3 billion in 1990 to 9.6 billion in 2050. The classification captures that growth well because it covers over 99% of humanity. And the total height of the bars traces the global increase almost one-for-one.

The second is that the climb to higher income happens despite where population growth is concentrated. Population has grown fastest in the poorest countries, where fertility has stayed relatively high while life expectancy has continued to rise. Despite this, the world has still climbed the moving escalator faster than the thresholds themselves have risen.

And this is the central point of the post: a very large absolute movement of people into richer-country categories.

China and India on the moving escalator

The projections show what a remarkable success story China has been through the lens of the classification. China rode the express escalator. It was still low income in the late 1990s, became lower-middle income in 1999, and reached upper-middle income in 2010. Based on current data, China is on track to cross the high-income threshold in 2026, with the World Bank expected to announce its reclassification in July 2027. The move from lower-middle to upper-middle income took only eleven years, while the transition from upper-middle to high income took sixteen.3

India has climbed more slowly. It entered the lower-middle-income group in 2007 and is projected to reach upper-middle income in 2036 under the business-as-usual scenario (discussed further below). That means India is likely to take 29 years to cover the same stretch that China completed in 11 years. With faster reforms, a more favorable external environment, and a measure of good fortune, that transition could, of course, arrive sooner, even though the opposite could also transpire.

Because China and India together account for over one-third of humanity, their movement through the classification has an outsized effect on the global distribution across income groups. This is clearly visible in the charts.

But the most consequential shift still lies ahead: India’s move into the upper-middle income. By then, its population will be at about 1.6 billion people. That single transition will lift the upper-middle-income group to 3.4 billion people by 2040, and it peaks near 3.5 billion in the mid-2040s. This makes it the largest any income group has ever been. By 2050, the high-income group will however overtake, reaching 3.6 billion.

Technical notes

The starting point

I consider 224 economies listed in the World Bank’s historical dataset that are classified in at least one year during the period 1987-2025. Six left the classification before 2025 so cannot be projected (e.g., USSR, Netherlands Antilles, and Serbia and Montenegro). So the universe consists of 218 economies.

The starting point is each economy’s 2025 GNI per capita level as reported in the World Development Indicators. This is used exactly as published. A country’s currently published GNI per capita can sit on the other side of the line reflected in the classification. That is because the Bank updates its classification in July of the next year and revises the previous year then. I use the reported figures as published, which leaves the discrepancy visible.

Threshold drift

Since GNI per capita is expressed in nominal terms, inflation is corrected for by allowing the thresholds to drift. The good thing is that the thresholds move in sync. Since 1990, the lower-middle-income threshold has been consistently about 4 times the low-income one and the high-income threshold has been 3.1 times the upper-middle income one.

I impose a common drift rate of 1.244% a year for all three thresholds. This rate is based on the median annual change in the published thresholds themselves over the decade to 2025. The Bank indexes all three to the SDR deflator, which is compiled from inflation measures of economies represented in the IMF's Special Drawing Rights: currently China, Japan, the UK, the US, and the Euro Area. I did a robustness check on the choice of the 10-year window, but altering the window between 10, 15, 20 and 25 years doesn’t move the drift rate outside the 1.16% -1.25% range, so 1.2% seems a good assumption.4

GNI per capita projection rates

I have taken the approach of not allowing for any downgrades in the income classification during the projection horizon. One could argue that this unreasonable. Conflict, for example, is the most likely factor driving a sustained long-term decline. But given the 2050 horizon, I am assuming that such cases will be resolved and hopefully they will. To accomplish this, I impose a lower bound of 1.2% to growth rates. Since this is also the rate at which the thresholds increase, no economy is projected to fall out of its income group. This assumption matters considerably as it affects countries that represent 14.4% of the global population.

In determining country-level projection rates, I apply a degree of smoothing at three different levels. First, the use of a median over the past growth rates of the last decade filters out unusually strong and weak years. Second, the projected growth rates are based on quantities already smoothed by the Atlas method (the exchange rate with the US$ is averaged over three years and adjusted for differences in inflation between the country and major economies). So the projections embed that smoothness into the projected rate. Third, I bound the projections from above at the 90th percentile of median 10-year growth observed across each income group: 5.6% for low income, 7.5% for lower-middle, 9.1% for upper-middle and 7.2% for high income. This is to rule out any unreasonable growth paths relative to the historical experience over the last 10 years. This assumption affects countries that represent 7.3% of the global population.

There are 21 economies (representing 1.8% of the global population) that suffer a data problem in that fewer than 8 of the 10 annual observations needed to establish the median GNI per capita growth rate over the last decade are available. These economies remain in the sample but they are frozen at their 2025 classification. Six among them are not high income: Cuba, Eritrea, Democratic People’s Republic of Korea, South Sudan, Syria and Yemen.

All in all, about three-quarters of humanity is projected on its own observed growth rate. The remaining quarter splits into 21.7% whose rate is set by a bound and 1.8% whose class is frozen outright.

Interpreting the projections

The projections should be interpreted as a simple business-as-usual scenario that extends recent experience into the future. Each economy’s GNI per capita is assumed to grow at its median annual rate over the past decade. The objective is not to forecast individual countries with precision, but to construct a simple, transparent baseline grounded in recent history. Although the past decade includes the COVID-19 pandemic, using the median largely filters out the influence of that extraordinary shock.

The scenario does not describe what is desirable, achievable, or even most likely to happen. It simply projects current trends forward under broadly unchanged policies and an external environment that stays roughly similar (or as, one might argue, similarly rough). With stronger reforms, many economies could outperform these projections. Without them, outcomes could be weaker. External developments (such as technological breakthroughs, shifts in globalization, climate shocks or another pandemic) could, of course, produce surprises in either direction.

Even with these caveats, the broad picture emerging from the scenario appears relatively robust for two reasons. First, predicting movements between income groups over a decade is easier than predicting any country's GNI per capita in a given year. Whether a country crosses a threshold early or late within the 2040s changes little to the overall picture in 2050. Second, the projections aggregate outcomes across nearly 200 economies, so individual forecasting errors tend to offset one another to some degree.

A comment on entry and exit

While I assume no such changes over the projection horizon, economies have historically entered and left the classification from time to time. This typically results from geopolitical change (most often), expansion of the Bank’s coverage (e.g., when small states like Nauru were added), or a lapse in the data needed to classify an economy in a particular year (such as Venezuela from 2020 to 2024).

As a result, the flows between columns do not always connect. Each column still sums to 100% of the classified population, but a band can terminate without a successor or begin without a predecessor. Countries that disappear simply drop out of one column, while newly created ones enter the next without a connecting band.

For example, the USSR, Yugoslavia, and Czechoslovakia were all classified as upper-middle income in 1990, with about 330 million people between them. They left the classification upon dissolution, and their successor states re-entered mostly a rung or two below: of the 22 successor economies, 13 came back as lower-middle income and 6 as low income by 1995, with only Slovenia, Croatia and Czechia re-entering as upper-middle income. The upper-middle-income band was not fully replenished until China arrived in 2010.

Download the data on GitHub

Massive progress but narrow perspective

While I am reasonably confident in the projections, the country income classification itself remains a narrow way of viewing development.

Because it relies solely on average per capita GNI, the classification is distribution-blind. A country crosses a threshold when its average crosses it, which tells us nothing about who inside the country has benefited.

The classification is also income-centric. Development is far more holistic, of course, encompassing health, education, opportunity, rights, and agency alongside income. Many of these dimensions correlate with income, but the relationship is far from perfect.

Even so, the results remain striking. A world in which 71% of humanity lives in upper-middle and high-income economies is a world with greater fiscal capacity, deeper markets, and more resources for health and education than the world of 1990.

The lens may be narrow, but through it we can still see one of the largest improvements in material living standards in human history. And that is a positive story worth remembering next time we switch on the news.

These lines organize how we think. They tell us who is ahead, who is behind, what counts as progress, and what constitutes arrival.

Here is the thing about these lines: they move.

The high-income threshold of 1990 is in a different place today. The technologies that defined the frontier a generation ago are now mere table stakes. The demographic profile that once powered East Asia’s economic miracle is inverting into its greatest test.

Even the questions move. “How do countries grow?” has given way to fundamentally harder ones: How do countries grow when the ladder others climbed is being pulled up, when populations age before they prosper, when the technological frontier races ahead faster than institutions can absorb it?

This newsletter takes its name from that double movement. Frontiers move on their own, reshaped by demography, technology, and geopolitics, among many other relevant trends. And frontiers can be moved, by countries that see clearly, bet wisely, and govern well.

Moving Frontiers is thus about both: how the frontiers of development are shifting beneath us, and how we can shift them ourselves.

Subscribe now

Reframing development

Underneath Moving Frontiers lies the idea that development needs to be reframed. The reason is that the world has delivered some of the most remarkable gains in history, yet aspirations remain unmet, human insecurity is rising, and the development landscape ahead of us looks far more turbulent than at any point in recent memory.

Accelerating progress under these conditions demands a willingness to see development, imagine it, and shape it differently.

Seeing development differently

The long view reveals extraordinary progress: over two centuries, every region has raised incomes, health, and education. But the journey remains deeply unfinished — ninety percent of humanity lives in regions below forty percent of frontier income, and not a single SDG is on track for 2030.

And the road ahead looks harder, not easier, as job insecurity hits record highs and clusters of risk converge across existential threats, fresh headwinds to growth and job creation, and growing governance risks.

Seeing development clearly means holding all three truths together: genuine progress, a large unfinished agenda, and mounting turbulence.

Imagining development differently

Too much of the development discourse collapses this turbulence into a narrative of polycrisis.

Crisis language narrows our vision; opportunity thinking keeps it open.

Aging is not merely a drag on growth but also a massive opportunity to live longer lives with vitality, purpose and connection — and to develop new engines of growth in health and care. Technology, after centuries of failed doom predictions, is delivering significant welfare surpluses beyond monetary output metrics, without necessarily costing jobs. And globalization is not retreating but slowing — disglobalization, not deglobalization — with the absolute scale of opportunity undiminished and, indeed, still expanding.

Imagining development differently means reading disruption as raw material for renewal.

Shaping development differently

Seeing and imagining are the easy parts; the harder question is what the development models of the future must contain.

I believe these are the following three things:

• Put people first: human development concepts are well developed but poorly mainstreamed, and a systemic reset must place people at the heart of strategy.

• Pursue smarter growth: the call for growth should be louder, not quieter — but it must be the right kind, not jobless, ruthless, voiceless, rootless, or futureless.

• And make change happen: the binding constraint is rarely a shortage of ideas but the politics of reform and the art of delivery. Yet both are all too often treated as afterthoughts.

Shaping development differently means people first, smarter growth and an unrelenting focus on effective leadership and governance, so change actually happens.

Combining different lenses

The title of this newsletter reflects a personal journey across institutional frontiers. I have spent my career inside global development institutions, moving from the International Monetary Fund to the World Bank to the United Nations Development Programme.

Each has its own way of seeing:

The Fund taught me to think in terms of stability: the macro-financial plumbing that must hold before anything else can be built. Hard-won gains can unravel when the foundations crack. The first discipline of development is keeping things from breaking.

The Bank taught me to think in terms of transformation: the long structural climb from farm to factory to frontier, and the evidence base that tells us which interventions actually work. As country economist for China, Brazil, and Malaysia, and later leading Global Modeling & Analytics in the research department, I saw development as a measurable, comparable, worldwide project, with all the immense power and all the distinct blind spots that measurement brings.

UNDP has taught me to think in terms of capability and possibility: development as the expansion of what human beings can be and do, pursued in real time across 34 countries in Asia and the Pacific, where I currently serve as Chief Economist. From this vantage point, the abstractions of thresholds and trajectories resolve into something concrete: governments making strategic choices under uncertainty, and the quality of governance deciding whether good analysis ever becomes good outcomes.

Stability, transformation, capability. Each lens is incomplete on its own. Held together, they change what you see.

This newsletter is where I hold these three perspectives together, or at least will try to do so.

What is coming?

The opening series on the “Moving Escalator” will ask how we measure frontiers. For many decades now, the World Bank’s country income classification has served as development’s de facto scoreboard, sorting every country into low, lower-middle, upper-middle, and high income. I will trace how the world has actually moved through that lens: who climbed, who stalled, and where the hard passages lie. Then I will turn the question around and ask whether the lens itself still serves us, and what a better one would look like.

Next comes a series on China. Whatever one’s priors, China’s development is the most consequential economic transformation in recorded history, and I will argue that we systematically look at it through the wrong frames. The series will ask how we should look at China, and what China’s experience, its successes and its unresolved tensions alike, carries as lessons for the rest of world.

After China, India: the other continental-scale development experiment, running on a different operating system, whose next two decades will do much to define what the moving frontier means.

And threaded throughout, I will examine several mega-trends: demographic change, shifts in globalization, and technological change. The temptation with each is to catalogue risks. I want to look more deeply at opportunity, more holistically at what these forces mean for countries pursuing a frontier that refuses to stand still, and above all connect the dots between the trends as they interact in very interesting, and potentially rather challenging, ways.

An invitation

Expect an essay every week or two, anchored in data and charts, written for anyone who thinks seriously about development: policymakers, practitioners, researchers, students, and the professionally curious.

One of my goals will be to pose questions that unsettle the status quo. The comfortable boxes of development thinking, the income categories, the standard sequences, the tired dichotomies of state versus market and growth versus equity, have all outlived much of their usefulness. So I will try to step outside these and cast fresh perspectives on old and new questions alike.

Hope you will enjoy the materials!

Share Moving Frontiers