What Do Increasing AI Capabilities Mean For Global Health and Development?

Executive summary

What could increasingly capable AI mean for low- and middle-income countries (LMICs)? Much consideration has gone to how AI may supercharge existing global health and development (GHD) programs, but a fuller answer depends on which constraints actually hold developing countries back, and whether AI represents a help or hindrance against those constraints. We examine the question through three traditions or ‘lenses’ from development economics.

Viewing AI through three lenses from development economics

1. Geographic constraints: AI looks most promising

AI offers promising ways to alleviate many burdens that disproportionately affect tropical economies, including AI-driven drug and vaccine discovery, AI-assisted diagnostics for neglected diseases, accelerated crop breeding for heat- and drought-resistance, and improved weather forecasting.

Such gains represent meaningful welfare improvements but may not drive long-run economic convergence with higher-income countries. Furthermore, they still require the absorptive capacity that many LMICs lack (e.g., well-functioning delivery systems, digital infrastructure, and trained workers).

2. Institutions and state capacity: AI is a double-edged sword

Institutions have risen to be seen as a critical constraint on development. AI may help increase efficiency and broad state capacity in countries with already well-functioning and inclusive institutions (e.g., tax administration, procurement oversight). In extractive or authoritarian contexts, AI may actively worsen governance by lowering the cost of surveillance and information control relative to delivering public goods, eroding citizens’ bargaining power as automation reduces elites’ dependence on labor, fuelling resource curses through surging demand for critical minerals concentrated in fragile states, and driving skilled emigration from deteriorating political environments. AI alone will not reform how extractive elites behave, and may entrench them.

3. Structural transformation and trade: AI may disrupt a historical development path

Manufacturing-led industrialization has been the most reliable historical route to economic growth. AI-driven automation in high-income countries threatens to erode the low-wage labor advantage that enabled this path, while other AI applications also challenge credible alternatives, such as services. It will become increasingly difficult to confidently direct industrial policy in a rapidly changing world. Meanwhile, value in AI supply chains will continue to accrue to early leaders such as the US and China.

Implications for GHD strategy

The constraints AI is most unambiguously set to relax may not be the binding constraints on long-run development. Meanwhile, AI poses genuine risks to the institutional quality and trade dynamics that development economics identifies as more fundamental. Negative impacts on governance and structural transformation could override real but partial gains in circumscribed domains, such as health and agriculture.

There is now a window to shape how AI interacts with fundamental development constraints at an early stage, but it requires engaging in ways beyond conventional GHD approaches. Funders and the GHD community should:

• Support cross-disciplinary work that bridges GHD and AI policy, such as investigating whether differential access regimes could empower benevolent institutions and deter harmful applications of AI, and how to balance state technological sovereignty with restrictions on nefarious uses of AI.
• Raise IP/access regimes, LMIC regulatory capacity, and mineral supply chain governance on the development agenda to enable capability building and spillovers rather than dependency and extractive dynamics.
• Elevate trade and industrial policy as central in GHD discussions about AI, not just AI applications in existing or circumscribed interventions. Formal modeling of different scenarios of AI-impacts on trade and the global economy may prove vital in informing such discussions
• Treat political will as a key investment criterion for AI governance solutions: Technology won’t compensate for its absence, and it risks creating the illusion of good governance without its substance.

Traditional GHD programming remains an important lever for improving the lives of people in LMICs. When it comes to how AI intersects with these programs, we think better outcomes will result when funders and GHD specialists:

• Bundle AI tools with delivery system investments: Co-invest in the local infrastructure and human capital that make tools usable.
• Maintain evidence standards: AI-driven interventions should not be prioritized on novelty alone.

Introduction

There is growing enthusiasm about AI and global health and development (GHD), ranging from near-term and focused AI-driven improvements in many areas of global health to expansive visions of prosperity in a world reshaped by massively transformative AI. What developments in AI mean for low- and middle-income countries (LMICs) depends not just on what AI capabilities arise, but crucially on what constrains those countries, and in turn whether AI proves beneficial, neutral, or detrimental in tackling such constraints. These constraints, and how they might interact with AI, are the core subjects of this essay.^[1]

Development economics offers multiple lenses on the constraints preventing developing countries from becoming more prosperous. In this essay, we employ three of these lenses to understand the implications of increasing AI capabilities in GHD: geographic constraints, state and institutions, and trade and development. The geographic tradition emphasizes environmental and health burdens in tropical economies (such as disease, heat, poor soils, and climate risk) as the primary drags on productivity and growth. The institutional tradition argues that extractive governance and misaligned elite incentives prevent any resource or intervention from translating into sustained growth. Finally, the tradition focused on export-led structural transformation stresses that labor-intensive manufacturing and integration into global value chains have been the most reliable historical mechanism for income convergence–the process by which developing economies close the productivity gap with richer ones.

We suggest that AI speaks directly to constraints emphasized in the geographic tradition, potentially providing new medicines or greatly increasing the efficiency of existing interventions. However, AI may also worsen institutions by placing surveillance tools in the hands of entrenched elites, and could threaten the manufacturing path through automation. A cross-cutting theme is that the prior skills, quality of institutions, and supportive infrastructure that countries already have will likely strongly influence whether that country can or will convert new technology into productivity growth, rather than more circumscribed welfare improvements, or even into more negative and freedom-reducing ‘innovations’. An uncomfortable possibility is that the types of constraints AI is most unambiguously placed to address may not be the main factors holding countries back.

Geographic constraints

Geographic and climate factors impose real productivity costs on tropical economies that institutional improvements alone do not fully offset.^[2] Location and climate suppress output through multiple channels: heat reduces both physical and cognitive labor productivity; disease burdens deplete the workforce, crowd out educational investment, and deter foreign capital; leached tropical soils underperform temperate ones; landlocked countries pay heavy transport penalties. Climate change may render many of these problems even more pressing. While ample research shows that geography is not destiny, the underlying costs of such burdens remain substantial and unevenly distributed.

AI has promising applications against natural and geographic constraints

AI’s near-term applications address geographic constraints directly. Disease burden presents a compelling example. A recent Rethink Priorities report explored the landscape of current AI applications in healthcare, finding that AI-assisted diagnostics (image recognition for tuberculosis, retinal screening, clinical decision support for health workers in under-resourced settings) offer a way to extend specialist expertise to places that cannot sustain it in person. For neglected tropical diseases, which our prior research indicates account for roughly 12% of the global disease burden but less than 1% of R&D spending, AI methods, including protein-folding prediction and in silico compound screening, could compress vaccine and treatment development timelines that have historically taken decades.

AI applications in agriculture may also alleviate typical geographic constraints through crop breeding, precision forecasting, and rural finance. AI-assisted plant breeding using genomics and machine learning could accelerate the development of heat- and drought-resistant tropical crop varieties, addressing a persistent skew in R&D toward temperate-zone crops. AI-improved weather forecasting could address the documented accuracy gap for the tropics. For example, our prior research found SMS-based weather services increased yields by up to 66% in sub-Saharan Africa. AI-based credit scoring using alternative data (mobile usage patterns, transaction history, satellite-derived farm productivity) could expand access to capital for the unbanked, directly addressing the investment constraint that sustains poverty traps.

Why these benefits could fall short

Access to better tools alone is not sufficient to transform the many geographically clustered burdens facing developing countries. Ultimately, one also hopes to achieve total factor productivity (TFP) gains that extend beyond more circumscribed welfare improvements. Technology can improve welfare without shifting productivity growth, and mobile money in East Africa illustrates the pattern clearly. M-Pesa access reduced poverty and improved consumption resilience during shocks, but evidence for mobile money as a driver of aggregate productivity growth was considerably weaker, with most users employing accounts for transactions rather than capital accumulation, representing a real benefit but not a transformation.

The standard GHD focus on health and agriculture is not misplaced, as it addresses widespread suffering through interventions backed by strong evidence. However, many such applications operate primarily at the level of welfare rather than promising long-run convergence with higher-income countries. To unlock even these more circumscribed (but meaningful) gains, technology adoption requires absorptive capacity, or the prior skills, institutions, and infrastructure that allow a country to productively use new technology. New diagnostics still require delivery systems, well-trained workers, and functioning supply chains that AI is unlikely to fully replace. AI drug discovery will continue to prioritize commercial disease markets unless specific push-pull mechanisms redirect R&D toward LMIC disease burdens (a policy and institutional challenge that better technology alone cannot resolve). Agricultural applications require digital infrastructure and device access that is unevenly distributed, and AI-delivered guidance faces the same adoption and trust challenges that have hampered conventional agricultural extension for decades. As our prior livelihoods research documents, barriers to the uptake of agricultural technologies are persistently rooted in trust, risk aversion, and structural market failures. Financial inclusion applications similarly depend on digital financial infrastructure that many low-income countries still lack.

A further challenge raised by AI warrants attention with respect to geographic constraints: AI compute imposes substantial energy demands that contribute to global carbon emissions. Without a shift in the energy mix, AI may accelerate the climate risks it is also being deployed to mitigate. While these energy demands will likely incentivize increasing energy efficiency, and AI might even be used to facilitate such advances, this will not necessarily reduce energy consumption and could even increase it (i.e., Jevons’ paradox). This is one instance of a broader pattern: AI performs well against geographic constraints that are primarily information problems (missing knowledge, absent early warning, or scarce diagnostic capacity) but has much more ambiguous implications for constraints defined by infrastructure gaps, implementation capacity, or distributional politics.

The state and institutions

The core claim of the institutional school in development economics, spearheaded and exemplified by Acemoglu, Johnson, and Robinson, is that the formal and informal rules governing political and economic life are the true binding constraints on long-run development.^[3]

Questions of institutional quality and state capacity are particularly pressing in the 2020s, as institutions degrade globally,^[4] and AI’s effects on institutions are among the least-examined channels in the GHD literature. We argue that AI represents a double-edged sword when it comes to institutions and state capacity, with the potential for adverse outcomes driven through multiple channels, including surveillance that reshapes the incentives of authoritarian governments, new resource rent pressures concentrated in institutionally fragile countries, and the erosion of accountability mechanisms through which citizen pressure has historically driven institutional improvement. Where institutions are already functional and beneficent, AI tools can help build state capacity for the common good. However, that threshold may be high, and AI tools could be ineffective or even damaging without the political will to use them for good.

Why institutions and state capacity are important for growth

Inclusive institutions (those that broadly secure property rights and constrain elite power) create incentives for investment and innovation; extractive ones concentrate power and redirect incentives toward rent-seeking. Institutions shape who captures any potential windfall from AI: aid and technology introduced into extractive contexts tend to strengthen extractors rather than transform the system for the better.

State capacity drives total factor productivity growth through three interconnected channels: fiscal capacity (revenue collection and public goods), legal capacity (property rights and contract enforcement), and collective capacity (bureaucratic competence to implement policy). State capacity in these domains ultimately enhances productivity: property rights lengthen investment horizons, contract enforcement expands effective market size, and competent administration reduces transaction costs throughout the economy.

In the “developmental state” model, governments deliver economic performance in exchange for political compliance. This “developmental bargain” creates positive feedback between growth and legitimacy, and generates political incentives to build state capacity. Evans showed that the developmental states that most successfully promoted industrial upgrading, such as South Korea and Taiwan, combined bureaucratic insulation from capture (autonomy) with dense, coordinated ties to the private sector (embeddedness). Autonomy without embeddedness produces unresponsive bureaucracy, whereas embeddedness without autonomy produces capture (e.g., corruption and patronage). It was the rare combination of embeddedness and autonomy, not merely the available technology, that drove change for these countries.

Promising AI-driven efficiency gains, such as automated enforcement and algorithmic procurement oversight, can plausibly reinforce state autonomy (reducing discretionary decision-making and making some forms of capture harder), but do not necessarily promote embeddedness, nor is an AI program that makes a tax authority more automated the same as a well-coordinated developmental state. The trust-based public-private coordination at the heart of successful East Asian industrial policy took decades of relationship-building, and it is not obvious that AI can accelerate such linkages.

Mass surveillance and information control

One of the most concerning paths through which AI could damage institutions runs through political economy rather than technical failure: AI-facilitated control of information combined with mass surveillance could lower the cost of suppressing dissent relative to the cost of delivering broad-based public goods. This shift would reshape the fundamental bargain at the heart of developmental states, or indeed between governments and citizens more broadly.

Research suggests that some authoritarian regimes have already shifted from overt coercion to managing citizen beliefs through information control. This was well underway even before AI arrived, but advancements in AI are set to accelerate such capabilities, as facial recognition, predictive policing, and automated social media monitoring make belief management less costly and more precise. The result is the decoupling of regime survival from economic competence, thus removing one of the most important growth-forcing pressures on governments.

Increasing AI capabilities also open up new avenues through which the wealth of data already collected on citizens can be collated and understood, as well as enabling new forms of collection and monitoring (e.g., using AI agents to more deeply scrutinize individual citizens, synthesizing formerly cumbersome qualitative information, or even utilizing communications between citizens and an LLM to gain insight into their private lives). Concerns over the use of AI in domestic surveillance have already surfaced even in higher-income countries such as the United States, where Anthropic came under considerable pressure from the Department of Defense to allow its AI tools to be used for any lawful purpose, which could include mass domestic surveillance.

Resource curses can corrupt institutions

AI’s expansion of demand for critical minerals may concentrate new resource wealth in exactly the countries where extractive institutions are already strongest. The resource curse is the well-documented pattern by which natural resource wealth correlates with weaker governance and growth. Intuitively, one might think that a wealth of valuable natural resources would be an unmitigated boon for a country, but because resource-driven rents allow elites to generate wealth and sustain power through patronage rather than productive investment, in practice, they can represent a curse. The geography of AI-critical minerals maps closely onto regions of institutional fragility, meaning the AI mineral boom arrives where the curse bites hardest.

AI substantially increases demand for electricity as well as for the critical minerals^[5] required for computing and energy infrastructure. Such increased demand can amplify the power of extractive institutions built around it, with further potential for exploitation from powerful external actors, such as states or corporations. The Democratic Republic of the Congo dominates global cobalt supply yet remains among the poorest countries in the world; the combination of resource wealth and institutional fragility replicates the classic curse pattern at accelerated speed. Following Myanmar’s 2021 military coup, unregulated rare earth mining boomed to satisfy Chinese semiconductor demand, with revenue flowing through ethnic militias and directly financing the authoritarian regime.

In the case of minerals crucial to AI supply chains, the legal protections governing mining concessions and environmental standards may increasingly be subject to executive overreach. New mineral demand driven by AI growth can therefore create conduits for elite patronage and new forms of geopolitical leverage for dominant AI powers.

Loss of the social contract

AI-driven automation represents both the promise of productivity benefits and the risk of job loss and consequent unrest. One further, relatively neglected consideration is that such automation may weaken one of the main mechanisms through which inclusive institutions have historically expanded: elites’ dependence on the wider population for productive labor. Where labor is displaced by automation, the political incentive to invest in human capital and infrastructure may decline, and the bargaining power through which citizens have historically extracted better governance weakens. In many initially extractive contexts, gradual institutional improvement came precisely because elites needed a productive workforce. Widespread automation without corresponding job creation may circumvent this typical process.

AI surveillance, raised above, also undermines the collective action that has historically driven political reform. By raising the risks of organizing, it weakens citizens’ ability to signal preferences to domestic elites and international audiences alike – a function central to successful mobilization. The chilling effect can run even deeper than repression, deterring collective action before it occurs. This removes the feedback mechanism through which governments learn which failures require response, progressively putting institutional reform out of reach. AI surveillance capabilities are already diffusing to governments that lack the resources to build equivalent infrastructure themselves, enabling this form of low-cost information control at scale.

Political deterioration as a result of such developments can degrade both quality of life and economic opportunities within a country, reducing growth directly but also driving away human capital. In particular, such degradation could drive the emigration of tertiary-educated or otherwise highly skilled workers, depleting both private-sector productivity and positive state capacity. Many sub-Saharan African countries already lose large shares of their tertiary-educated populations to emigration, and AI-enabled political consolidation may intensify this tendency, removing another channel through which governance failures might otherwise be corrected.

The bottom line for institutions

Where institutions are already reasonably functional and beneficent, AI may offer useful tools for state capacity, such as AI-assisted tax administration, automated fraud detection, satellite-based land monitoring, and AI-improved health information systems. However, these tools require institutional strength to be used productively, rather than creating it. AI tax tools are ineffective without the political will to enforce obligations on powerful elites, and budget transparency achieves little if enforcement rests with the same actors it is meant to constrain. Without a solid foundation in which to be deployed, adoption of such tools risks superficial “isomorphic mimicry” (adopting the form of good governance without real reform or capacity).

A more concerning possible trajectory for AI’s institutional effects is actively regressive and may further the divergence between lower- and higher-income countries: Where state capacity is weak or institutions are predominantly extractive, AI may be used to supercharge surveillance and information control, consolidating existing power structures, accountability mechanisms that have historically driven institutional improvement may be eroded, and talent flight may accelerate.

Trade and the development path

Manufacturing-led growth through “structural transformation” (the economy-wide shift of workers from low-productivity agriculture into higher-productivity industry) has been the most reliable historical route to sustained income convergence. There is a risk that AI may shut off this route before most developing countries have traversed it.

A contingent and narrowing path

From the British Industrial Revolution through to the Asian Tigers of the mid-to-late 20th century, the shift of workers from low-productivity agriculture into higher-productivity factory employment has been the primary engine of sustained growth. Export-oriented development relies on labor-intensive industries because labor is cheaper in lower-income countries than in high-income countries. This cost differential represents a comparative advantage that drives early industrialization and, as productivity compounds through learning-by-doing, supports upgrading to higher-value activities.

Because manufacturing is subject to increasing returns and learning-by-doing, productivity advantages compound over time. Manufacturing productivity in developing countries exhibits unconditional convergence toward frontier levels regardless of institutional quality. That tendency distinguishes manufacturing from most services and agriculture, making it historically the development path most accessible to countries with relatively weak institutions.

The manufacturing convergence path was already seen as weakening before AI arrived. In contrast to earlier structural transformations, manufacturing productivity convergence between countries from 1990 to 2018 was concentrated in formal-sector enclaves rather than producing wide-ranging spillovers. Furthermore, premature deindustrialization (labor moving from agriculture into informal services rather than manufacturing) was observed in many developing economies, and was a growth-reducing rather than growth-enhancing form of structural transformation. If cheap AI-driven automation is achieved, then it may accelerate the closure of this already-narrowing path.

The automation threat

Automation of manufacturing in high-income countries erodes the advantage of low-wage labor in lower-income countries: if AI-driven automation can drive productivity to match labor cost reductions expected from placing manufacturing in lower-income countries, then the incentive to offshore production to such locations diminishes, and the labor-cost route to industrialization narrows. Capital-biased technical change (the tendency of new technology to raise the returns to capital relative to labor) is already pushing formal firms in tradeable sectors toward labor-saving techniques, compressing the employment growth that manufacturing once reliably produced.^[6]

Challenges to alternative forms of productivity growth

The threat of AI to structural transformation is not restricted to factory floors. AI also threatens call center work, data entry, and basic coding tasks. These are service-sector alternatives to manufacturing that some economists had conceived as a secondary development path. The question for countries that missed the manufacturing window is whether any routes to convergence remain.

Entry into low-skill, low-margin stages of digital services is unlikely to provide the learning spillovers that manufacturing historically offered. Data labeling work has been documented as low-wage, precarious, and offering little pathway to higher-value participation in the AI economy. As AI automates routine cognitive tasks, workers moving out of agriculture may enter a services sector that is simultaneously being degraded by gig work rather than upgrading into higher-capability roles.

The concentration of value at the top of AI value chains reflects an asymmetry in development capacity. Over 90% of frontier AI compute capacity is concentrated in the United States and China, while the African continent accounts for less than 1% of global AI research output. Lead firms have retained control over core model architecture, compute, and intellectual property in ways that limit the prospect of capability upgrading by lower-income country participants, meaning developing countries lack both access to AI infrastructure and the capacity to build AI systems suited to their own economic conditions. This control over frontier AI systems can be supported with reference to the very risks of misuse we’ve highlighted as concerning in this piece, such as that some states may nefariously deploy such systems against their citizens. Still, countries that are purely consumers of AI tools developed elsewhere may receive the labor displacement without the corresponding task creation or high economic returns, becoming sources of raw behavioral data for models, the value of which accrues elsewhere.

Higher-skill tradeable services may offer an alternative convergence path, but for a narrow band of countries that AI may further restrict. Professional services and higher-skill IT-enabled exports can carry learning spillovers approaching those of manufacturing for countries with adequate educational systems. The concern is that AI is now rapidly automating the entry-level service tasks that formed the initial rungs of that upgrading ladder for exactly those economies, compressing the window before most lower-income countries can access it.

Industrial policy and investment decisions under AI disruption

AI disruption poses an additional challenge for developing-country industrial policy, which has traditionally targeted support at specific sectors expected to remain labor-intensive. Such foresight becomes increasingly difficult when AI is accelerating technological change in unpredictable ways.

The risk of “regulatory arbitrage” compounds the industrial policy challenge: AI companies facing increasing scrutiny in high-income markets may actively seek LMIC markets with weaker regulatory capacity, replicating the patterns documented in tobacco, baby formula, and extractive industries. LMICs often lack the technical capacity to evaluate AI system claims or detect failures until well after implementation. Involvement in the AI value chain, or access to AI systems, may be used as leverage to extract policies that favor AI developers at the expense of longer-term benefits to the country. Hence, dependence on foreign-controlled AI infrastructure creates structural vulnerabilities for national policy autonomy. Countries that rely on externally provided AI systems for critical services (health information, financial systems, agricultural data) may find themselves exposed to leverage that compounds existing dependency relationships. The concentration of AI infrastructure in a small number of firms and countries has no clear precedent in development policy and philanthropy, which has not previously had to account for this kind of systemic exposure.

Synthesis: Three frameworks, one uncomfortable conclusion

The three GHD lenses we’ve outlined highlight different concerns about AI. Many issues that currently dominate GHD discourse (such as disease burden, agricultural productivity, and financial exclusion) have promising AI-driven trajectories, but they may not be the most consequential for long-run development. Indeed, the rapid development of AI could entrench old barriers to growth and ultimately lock some countries out of traditional development pathways.

From the geographic perspective, AI development offers a relatively unambiguous optimistic vision: AI may deliver near-term solutions to, or amelioration of, such burdens as disease or agricultural penalties. Evidence on technology leapfrogging suggests these gains matter for welfare regardless of whether they shift long-run TFP trajectories. For countries where geographic constraints represent a substantial burden, AI could therefore be welfare- improving even without broader institutional change. Still, AI’s effects on institutions and global value chains could determine whether progress on geographic constraints can be sustained. Concerns emphasized in the geographic tradition, and which we believe have been more prevalent in discourse around AI and GHD, are real and tractable, but they address the most solvable constraints for developing countries, rather than necessarily the most consequential ones.

Considering all of the different lenses together, an optimistic case requires three conditions to hold simultaneously, and the evidence raises questions about each: geographic barriers would need to be the chief binding constraint (the weight of evidence points away from this for many LMICs); AI would need to not substantially worsen institutions (e.g., through surveillance, resource rent pressures, or erosion of accountability); and the manufacturing development path would need to remain viable or be credibly replaced. Furthermore, for AI to have a positive impact, there must be sufficient infrastructure, state capacity, and will to implement AI-driven solutions. The more likely scenario involves real but partial gains against geographic constraints, state capacity tools deployed for both good and ill, some institutions consolidating their extractive functions or becoming more extractive, with AI productivity gains concentrated in already-productive economies that hold the complementary inputs in abundance.

Implications for GHD strategy

Our analysis has direct consequences for how to deploy attention and resources in the AI era, if one is interested in improving the well-being of people in LMICs.

The most consequential near-term GHD conversations about AI will concern trade and foreign investment policy, not just development program design. Such discussions need to address the threats (and opportunities) for LMICs of AI-driven automation closing the manufacturing development path, especially if services no longer represent a clean substitute. In a world economy disrupted by AI, what are the areas in which such countries can credibly contribute and leverage a competitive advantage? Further discussion should center around whether AI infrastructure dependency creates new exploitation risks or intensifies resource curses, in turn making AI supply chain governance a GHD priority. AI intellectual property and access regimes that enable capability-building rather than dependency, support for LMIC regulatory capacity to govern AI companies operating in their markets, and attention to mineral supply chain governance in institutionally fragile producer countries all fall outside conventional GHD programming. They now belong on the agenda of any development community thinking seriously about AI.

GHD may benefit from cross-disciplinary collaboration and knowledge sharing with AI policy experts to address questions or find solutions that become increasingly important as AI capabilities increase. For example, could versions of differential access — typically used to restrict access for those likely to engage in dangerous uses, or promote access to those with benevolent uses — be considered in relation to frontier model access for states with credible commitments to use, or not use, AI in particular ways? Could such arrangements instead result in unfair external leverage imposed on recipient countries, or become conduits for bribery and corruption? Just as governments or institutions may be swayed by influential corporations, are there policies or other levers to support AI companies seeking to resist pressure to enable nefarious uses of their technology by state actors?

In the near term, circumscribed AI investments (such as drug discovery and vaccine development for neglected tropical diseases, AI-driven diagnostics and clinical decision support, or improved weather forecasting and early warning systems) also warrant continued attention, because they address documented suffering without necessarily requiring high institutional quality to deliver results. Even so, we suggest that support for such AI tools should usually be bundled with delivery system investments: co-investing in the local infrastructure that makes tools usable matters as much as the tools themselves. Furthermore, even in these domains, the evidence base is not conclusive. We should therefore resist treating AI-driven innovations as necessarily higher-priority than existing or traditional interventions. Confidence should track evidence quality and well-justified potential, rather than just novelty or largely hypothetical promise.

Medium-term bets on AI tools for state capacity (tax administration, procurement systems, health information architecture) are worth pursuing, but the investment decision must consider the institutional context, not just the technology. The question is not simply whether a tool exists, but whether the political will to use it effectively and appropriately does. Deploying such tools without that commitment risks creating the illusion of good governance without the function. The practical implication is to make demonstrated political will an explicit criterion for investment, rather than assuming AI tools will compensate for its absence.

By bringing three lenses from development economics to bear on the question of what AI means for GHD, we reach an uncomfortable but, we think, grounded conclusion: The constraints AI may most unambiguously address may not be the ones holding economies in LMICs back. Because AI represents a potential threat, or at least a double-edged sword, when it comes to state capacity, institutions, and trade, it is possible that circumscribed, yet real benefits in domains such as health or agriculture are ultimately overridden by negative impacts of AI on these potentially more fundamental constraints. The advent of advanced AI serves to make attention to these constraints more pressing, but potentially more promising too. If the GHD community can engage with and find solutions to the problems we’ve raised, then it may be possible to shape how AI intersects with such crucial constraints from an early stage, delivering real benefits to those who need them most.

Contributions and acknowledgments

Jamie Elsey developed the project concept. Jamie Elsey and Ruby Emerson conducted research and wrote the manuscript. Kieran Greig, John Firth, and Jerome Mayaud provided management and review.

1. Of course, the extent of AI capabilities themselves are also very important. For the purpose of this essay, we are principally discussing abilities that have the potential to transform various areas relevant to GHD, such as the global economy (e.g., through automation) and health (e.g., through novel treatments). While some of our arguments may hold true in relation to ‘artificial general superintelligence’, we do not focus on superintelligence scenarios. ↑
2. Tropical GNP per capita fell from roughly 70% of temperate-zone levels in 1820 to just 25% by 1992; economic productivity declines sharply above a temperate annual average temperature, with a 1°C increase cutting industrial value-added growth by over two percentage points in poor countries. ↑
3. The empirical evidence for institutional dominance on long-term economic development is compelling, and a spate of papers in the early 2000s made the case. Many “reversals of fortune” (in which regions wealthy in 1500 are generally poor today) are traceable to extractive colonial institutions (Acemoglu, Johnson, and Robinson 2002). Institutional patterns persist across centuries: Peruvian regions subjected to colonial forced mining labor are significantly poorer today than those that weren’t. ↑
4. The 2025 World Justice Project Rule of Law Index found that approximately 68% of countries surveyed saw a decline in their overall rule of law scores in 2025, up from 57% the prior year; in 63% of countries, independent auditing and oversight of government powers declined. ↑
5. These include, among others, lithium, cobalt, copper, gallium, germanium, and rare earth elements. ↑
6. Daron Acemoglu discusses a more specific version of AI automation risk: “So-so automation” describes technologies that are productive enough to be adopted and displace workers, but not productive enough to generate the new tasks and labor demand needed to compensate. Acemoglu’s 2024 analysis estimated realistic aggregate TFP gains from AI at 0.71 to 1.1% over a decade, against the 7% or more projected by AI optimists, and found that AI had, at the time of writing, operated predominantly in the so-so range. For developing countries, so-so automation in high-income economies removes the labor-cost incentive for offshoring that created manufacturing global value chain entry points over the past four decades. ↑