Over the past two years, few technologies have provoked as much economic speculation as generative artificial intelligence. Its rapid diffusion into workplaces, from software development and finance to education and professional services, has reignited long-standing debates about productivity, growth, and inequality. While the dominant narrative has emphasised the promise of efficiency gains and innovation, a quieter but more consequential question has emerged beneath the surface: who, exactly, will benefit first, and at what cost to existing global economic balances? A recent working paper by the Bank for International Settlements offers one of the clearest empirical answers yet, suggesting that generative AI is likely to widen the gap between advanced and emerging economies in the short run, not because of destiny or policy failure alone, but because of differences in economic structure and technological readiness that are already deeply entrenched.
The paper examines whether the early growth effects of generative AI differ systematically across countries, focusing on the period between 2022 and 2023, when adoption accelerated sharply following the public release of large language models. Rather than treating AI as a uniform shock, the authors approach it as a technology whose impact depends on where it meets the economy: which sectors dominate production, how exposed those sectors are to cognitive automation, and whether countries possess the infrastructure, skills, institutions, and regulatory frameworks required to translate potential into output. In doing so, the paper departs from speculative macro projections and instead grounds its conclusions in observed differences across 56 economies and 16 industries.
At the core of the analysis is a simple but powerful idea borrowed from earlier work on finance and growth. Just as industries that depend heavily on external finance grow faster in countries with deep financial systems, sectors that are more exposed to generative AI should grow faster in countries that are better prepared to adopt it. Exposure, in this sense, is not about the presence of robots or physical automation, but about the degree to which tasks within an industry rely on cognitive, informational, and knowledge-intensive activities that generative AI can augment. Preparedness, meanwhile, reflects national characteristics that make adoption feasible and productive, including digital infrastructure, human capital, innovation capacity, and regulatory quality.
To measure sectoral exposure, the authors rely on an industry-level index developed using US data, which captures how extensively generative AI can be applied to occupational tasks within each sector. Finance, education, and information services emerge as highly exposed, while agriculture, transport, and construction remain largely insulated. Although the index is derived from the US economy, the paper treats it as a technological benchmark, assuming that the underlying nature of tasks does not differ fundamentally across countries. This assumption mirrors earlier empirical strategies in development economics and allows the authors to isolate the role of national conditions rather than sector-specific idiosyncrasies.
Country-level preparedness is measured using the International Monetary Fund’s AI Preparedness Index, which aggregates four dimensions: digital infrastructure, human capital and labour market policies, innovation and economic integration, and regulation and ethics. The resulting picture is stark. Advanced economies cluster at the upper end of the distribution, with a median score around one-third higher than that of emerging and developing economies. Yet the dispersion within emerging economies is large. Singapore scores higher than many advanced countries, while others, such as Nepal, remain far behind. These differences matter because preparedness is not a theoretical construct but a reflection of whether firms can access data, deploy models, integrate AI into workflows, and operate within predictable legal environments.
Using these measures, the authors estimate the growth rate of real value added for each country–industry pair between 2022 and 2023. The empirical results point to a consistent pattern. For a given increase in AI preparedness, sectors with higher exposure to generative AI grow faster than those with lower exposure. The magnitude is economically meaningful. Comparing industries at the top and bottom of the exposure distribution, a one standard deviation increase in preparedness is associated with a growth differential of around two percentage points in real value added. This result holds even after controlling for country and industry fixed effects, as well as for the initial size of the sector.
Importantly, the paper does not conflate generative AI with earlier waves of automation. To address concerns that the results might simply reflect long-standing trends driven by industrial robots, the authors explicitly control for the stock of robots per employee across sectors and countries. Robot adoption is indeed associated with higher sectoral growth, particularly in manufacturing and utilities, but it does not weaken the estimated effect of generative AI. If anything, the distinction sharpens the analysis: robots primarily automate physical and routine tasks, whereas generative AI operates through cognitive channels, complementing white-collar labour rather than replacing it outright in the short term.
When the sectoral results are aggregated to the country level, the implications become clearer. By weighting predicted sectoral growth by each country’s production structure, the authors estimate the overall short-run growth effect of generative AI. On average, advanced economies experience an increase in real value-added growth that is about 0.6 percentage points higher than the country with the lowest estimated impact in the sample. For emerging and developing economies, the average gain is closer to 0.4 percentage points. While these numbers may appear modest, they are meaningful in the context of annual growth rates, particularly when concentrated over a short period.
The distribution of gains is uneven even within income groups. Luxembourg stands out among advanced economies, reflecting the outsized role of finance in its economy. The United States and the United Kingdom also rank among the stronger beneficiaries, consistent with their large professional and information-intensive sectors. By contrast, countries with significant mining activity or a heavy reliance on extractive industries see smaller gains. Among emerging economies, Hong Kong and Singapore rival or even surpass many advanced economies, again underscoring the importance of sectoral composition and institutional readiness. Economies dominated by low-skilled, labour-intensive manufacturing benefit far less.
Taken together, the findings suggest that generative AI, at least in its early phase, is not a leveller but an amplifier. It rewards countries that already possess the institutional and structural conditions needed to exploit it, while offering more limited immediate benefits to those that do not. This does not imply that emerging economies are doomed to fall behind indefinitely. Historical evidence on technology diffusion shows that adoption lags have shortened over time, and that latecomers can eventually catch up. However, the short-run dynamics matter, particularly in a world where relative growth differences compound quickly and shape investment flows, fiscal capacity, and political expectations.
The paper is careful to delimit its claims. It does not attempt to model long-run equilibrium effects, changes in labour supply, or general equilibrium spillovers across sectors. Nor does it claim that generative AI will permanently increase inequality between countries. Instead, it highlights a transitional phase in which productivity gains accrue unevenly, shaped by pre-existing differences in readiness and structure. In this sense, the results echo earlier debates about globalisation and financial integration, where benefits were real but uneven, and where adjustment costs fell disproportionately on those least equipped to absorb them.
For policymakers, the implications are uncomfortable but clear. Investments in digital infrastructure, education, and regulatory capacity are not optional add-ons but prerequisites for participation in the next phase of technological growth. At the same time, the paper cautions against simplistic narratives that frame AI as an automatic engine of convergence. Without deliberate efforts to broaden preparedness, the early gains from generative AI are likely to reinforce existing hierarchies rather than dismantle them.
For advanced economies, the findings offer both reassurance and warning. While they are well positioned to benefit in the short run, the concentration of gains in specific sectors raises questions about distribution within countries, particularly between high-skilled and low-skilled workers. Moreover, reliance on cognitive productivity gains does not eliminate the need for complementary investments in physical capital, energy, and infrastructure. For emerging economies, the challenge is sharper. The risk is not that AI will destroy jobs en masse, but that it will bypass large segments of the economy altogether, leaving growth trajectories increasingly dependent on sectors that are less exposed to productivity-enhancing technologies.
The paper’s contribution lies in its restraint. In an environment saturated with sweeping predictions and speculative forecasts, it offers a disciplined empirical assessment rooted in observed data and transparent assumptions. It shows that generative AI is neither a miracle cure nor an existential threat, but a technology whose economic effects are mediated by structure, policy, and preparedness. The question it leaves open is not whether AI will shape global growth, but whether countries will act quickly enough to ensure that its benefits are shared more broadly. That question, unlike the short-run growth effects measured in the paper, remains decisively political.
