The Global AI Distribution Gap

by

Why access to AI will converge while its benefits diverge

The distribution of the benefits of advanced artificial intelligence will be one of the defining questions of the next decade. The evidence on adoption and early uses is already substantial, even though many longer-term consequences remain uncertain.

My argument is this: basic access to AI will commoditize quickly, but the benefits will not be distributed evenly along with it. I expect access convergence with benefit divergence. The divides that last will sit higher up the stack—in who has frontier capability, who has the skills and institutional support to use it, and above all who builds, owns, and directs the systems. This essay examines what the gap looks like today, how it affects ordinary people, and which forces might bend the trajectory in either direction.

The gap is real, measurable, and currently widening

As of mid-2026, about 18.8% of the world’s working-age population uses generative AI. Microsoft’s telemetry-based measure puts usage at 28.8% in the Global North and 16.2% in the Global South; adoption is growing faster in absolute terms in the North, widening the percentage-point gap [1]. These figures are indicators of observed use, not a complete census of all AI activity.

Usage also tracks income closely. Anthropic estimates that a 1% increase in GDP per capita is associated with roughly 0.7% higher AI usage per capita, and finds no evidence yet that low-use countries are catching up [2]. What drives that relationship? Price is the weakest candidate: capable free tiers exist, and the income gradient persists in measures adjusted for connectivity and device ownership. Income is better read as a proxy for three structural factors.

The first is task mix. Wealthier economies employ more people in cognitive work that current AI handles well. Within the United States, where income is a weaker predictor of usage, the share of technical workers alone explains nearly two-thirds of cross-state variation [2]. The second is human capital: more sophisticated prompts tend to elicit more sophisticated responses, so education and experience increase the return a person gets from AI. The third is complements—reliable electricity, adequate connectivity, and model performance in local languages [3].

The distinction matters. A gap driven primarily by price should close as access cheapens. A gap rooted in what people do for a living, what they’ve been taught, and the infrastructure and institutions around them is stickier. Free tiers alone won’t erase it. These non-digital complements—electricity, schooling, and functional institutions—are the essay’s analog foundations. They shape not only whether people can reach AI, but whether they can make effective use of it.

The divide is therefore better understood as layered than as “has AI / doesn’t have AI.” There are four layers: access (electricity, connectivity, and affordability); capability tier (frontier reasoning models versus free or smaller models); fit (language, local context, and cultural norms); and absorptive capacity (education, institutions, and the business environment) [3]. These layers produce different effects and will not close at the same speed. Verification capacity cuts across them: people and institutions need the expertise and time to assess whether a system’s output is reliable.

Those four layers describe the consumption side. Production is a separate, more concentrated dimension. The World Bank’s 2026 World Development Report notes that advanced layers of the AI value chain—chips, data centers, and frontier models—are concentrated in a few countries and, often, a handful of companies [3]. A country can become proficient at using AI without gaining much influence over how it develops, whose interests it serves, or where the resulting economic value accumulates. Who builds and owns the systems is not simply another access question; it is a question of economic power and strategic control.

How this affects the general population

Six effects matter most—five structural risks that threaten to widen the gap, and one powerful counterweight pulling in the opposite direction. Diverging national income trajectories and the skill multiplier are the load-bearing economic claims. Exposure to AI’s harms without access to its benefits is the most underappreciated. The equalizing channel is the strongest counterweight. Labor disruption and epistemic concentration are real concerns, but their effects are more conditional or slower-moving. The sections move from economics, to ownership and influence, and finally to the case for optimism that complicates both.

1. Diverging national income trajectories—the “kicked ladder” problem. The IMF estimates that roughly 40% of jobs worldwide are exposed to AI, and about 60% in advanced economies—which are also better positioned to capture the benefits [5]. These are estimates of task exposure, not forecasts of job losses. Still, they point to a divergence risk: rich countries may capture more of the productivity dividend while poorer countries receive less, even if their immediate exposure to displacement is lower. The World Bank similarly warns that without complementary investments, AI could widen productivity gaps between developing and high-income countries [3].

There is also a subtler threat to the classic development model. Countries such as India and the Philippines built substantial middle-class employment on exported knowledge work—BPO, back-office processing, and entry-level coding—that includes tasks AI can increasingly perform or assist with. These economies are facing a dual squeeze: locked out of the capital-intensive hardware and compute layers at the top, while the entry-level cognitive work they rely on at the bottom is being automated away. If that ladder weakens before these economies have climbed it, the effect would be more than slower growth: it could narrow a path into higher-value employment. The ladder may not disappear. It may be replaced by one whose first rung is higher, requiring skills, infrastructure, and capital that are less widely available.

2. Uneven labor disruption, but not the obvious one. The ladder problem points to a counterintuitive pattern in who may be affected first. Because a larger share of employment in many developing economies is in agriculture, manufacturing, and other manual or in-person work, some workers may be less exposed to current generative AI than workers in knowledge-intensive services [3, 5]. That does not mean developing economies are insulated: administrative, financial, educational, retail, and service work is also substantial, and many of those activities include tasks AI can automate or reorganize.

The near-term effect may therefore be uneven within countries as well as between them. A subsistence farmer in Malawi may be less exposed to current generative AI than a junior accountant in Manila or a call-center agent in Bangalore. Over a longer horizon, robotics combined with AI could reach more manual work. Workers in countries with fewer savings, weaker safety nets, and less access to retraining will have fewer resources to absorb such transitions.

3. Stratification within societies—the skill multiplier. The same dynamic operates inside countries. Anthropic reports a very strong relationship between prompt sophistication and response sophistication (r > 0.92) [2]. That finding needs careful interpretation: it does not establish that less-educated users are incapable of receiving useful or sophisticated assistance. It does suggest that people who already know how to frame complex problems are better positioned to extract value from the technology.

The difference matters. An experienced professional may use AI as a collaborator that extends existing expertise, while a less-experienced user may rely more heavily on it for answers or task completion. The second user can still benefit substantially, but has less ability to recognize omissions, challenge assumptions, or improve the result. Anthropic’s analysis also finds that high-adoption countries use AI more for augmentation, while lower-adoption contexts skew more toward delegation and coursework [2]. AI does not automatically level skill differences; it can compound them, as skill-biased technologies have often done.

There is an important counterargument: AI can help users develop the very skills that make them better at using it. That possibility is real, particularly in education. But whether it happens depends on how tools are designed and used, and whether people receive guidance that helps them learn rather than simply outsource the work. Add frontier-tier subscriptions priced beyond most households, and the potential for cognitive stratification sits on top of existing economic stratification.

4. Exposure to AI’s harms without access to its benefits. People do not need to use AI to be affected by it. Populations with limited access still inhabit information ecosystems shaped by AI-generated content, face AI-scaled scams and disinformation, and are subject to decisions made by systems owned elsewhere—including in credit scoring, hiring, border screening, and surveillance.

The UN has warned that less-developed countries risk being left further behind as production shifts toward AI-driven systems [6]. A country can become a source of data, a testing ground for content and products, and a market for someone else’s models while capturing relatively little of the economic value. That is one dimension of what critics call AI colonialism. Another is political and informational influence: systems developed elsewhere can shape what people see, how institutions make decisions, and which perspectives are treated as authoritative. These are related but distinct forms of power: one concerns who captures economic rents; the other concerns who shapes information environments and the decisions made within them.

5. Epistemic and cultural concentration. This is the ownership problem seen from the user’s side. When a small number of labs’ models mediate how billions of people write, learn, and make decisions, populations whose languages, norms, and knowledge are poorly represented receive less capable or relevant service. They also encounter assumptions and priorities embedded in systems trained largely on data from other societies.

The World Bank emphasizes that tools developed in high-income contexts can produce recommendations misaligned with local realities. It also notes that adaptation—such as voice interfaces on basic phones, local languages, and local data—can deliver substantial benefits [3]. Access without fit leaves a population nominally included but effectively underserved.

There is a further concern. When a few organizations control both the systems that gather and interpret information and the channels through which that information can be presented persuasively, economic concentration becomes informational influence. That does not mean every system is designed to persuade, or that cultural influence automatically amounts to manipulation. But it raises questions about transparency, data governance, and the separation of legitimate assistance from undisclosed persuasion. Those safeguards matter more as AI systems mediate more of the information people rely on.

6. A genuine equalizing channel: cheap assistance, reliability at a premium. Five effects of concern don’t make the whole picture. The optimistic case is real. AI appears to diffuse unusually fast compared with earlier general-purpose technologies: the World Bank contrasts roughly 80 years for the steam engine to reach lower-income countries, 40 for electricity, and 20 for the internet with the rapid spread of ChatGPT traffic in middle-income countries [3]. Stanford’s 2026 AI Index reports that generative AI reached 53% population adoption within three years, faster than the PC or the internet [4]. These measures track different things and are not directly comparable, but together they suggest rapid diffusion.

The welfare case doesn’t require AI to outperform a good doctor, tutor, or lawyer. For someone with no access to any of them, the benchmark is the assistance actually available—often very little—and against that baseline the gains are substantial. In Nigeria, six weeks of teacher-guided AI tutoring produced learning gains the researchers equate to up to two years of ordinary schooling [7]. In Kenya, an AI safety net used by clinicians cut diagnostic errors by 16% across nearly 40,000 patient visits [8].

Both successes included a verification layer: a teacher or clinician with relevant expertise checking the model’s output before it reached the person relying on it. That layer matters. In a study of medical advice, raw chatbots did not show an established advantage over web search or people’s own judgment on the tested tasks; models can mix sound guidance with confident errors that novices struggle to identify [9].

The educational evidence points in the same direction. In one randomized trial, students using an unguarded AI tutor practiced 48% better, then scored 17% worse on the unaided exam, while believing they had learned just as well [10]. This does not mean AI tutoring is inherently harmful. It shows that a tool can improve performance during assisted practice while undermining independent learning if it encourages students to bypass the thinking they need to develop.

This exposes a variable that cuts across all four layers: the distribution of verification capacity. A well-funded school employs teachers who know how to integrate AI into instruction. A hospital relies on clinicians who know when to accept or reject an AI recommendation. A small school or underfunded clinic may lack the personnel, time, or institutional capacity to provide that oversight. The ability to distinguish useful assistance from misleading output is itself unequally distributed.

Whether AI builds human capital or erodes it is therefore not a question about technology alone. It depends on design, institutions, and who gets the verification layer. Access determines whether AI is available; capability determines what it can do; fit determines how well it serves the user; and absorptive capacity determines how effectively the user can employ it. Verification helps determine whether the resulting assistance can be trusted and used safely. It may be one of the highest-stakes distributional variables of the next generation.

Both sides of the central tension are true. AI can compound existing advantages, as the skill-multiplier argument shows, while also lifting the floor for people with very little assistance today. “Compounding stratification” and “a rising floor with a faster-rising ceiling” coexist. The availability of competent verification—and the institutions capable of providing it—helps determine who experiences which outcome.

The forces that could bend the curve

Whether that tension resolves toward the floor or the ceiling depends on forces pulling in both directions.

Pulling toward convergence

Collapsing costs and open weights. Open-weight models are capturing a growing share of API token volume, which could benefit the Global South by lowering costs and enabling local adaptation [1]. But open weights do not automatically create practical access to capable AI. Running and maintaining a model still requires computing resources, reliable electricity, connectivity, technical expertise, and capital. Cheap, capable, locally hosted models are among the strongest commoditizing forces in the system, but their benefits depend on whether countries and organizations can deploy them.

Speed of diffusion. Software-based technologies can spread through smartphones people already own, rather than requiring every user to wait for new physical infrastructure. That helps explain the rapid diffusion figures and may shorten catch-up times compared with earlier technological waves. Still, the speed at which an application spreads is not the speed at which a society develops the skills, institutions, and infrastructure needed to benefit from it.

Smaller organizations gaining leverage. Agentic systems could allow small businesses, public agencies, and individual entrepreneurs to perform work that previously required larger staffs and more organizational infrastructure. That could reduce some traditional advantages of scale and open opportunities where skilled labor is scarce. But the equalizing effect is not guaranteed. If platforms, compute, and distribution channels remain controlled by a small number of firms, organizations may gain operational leverage while becoming more dependent on infrastructure they do not own.

Pulling toward divergence

Frontier-tier concentration. The most capable models, and the compute that runs them, remain concentrated in a few firms and countries [3]. Frontier subscriptions are out of reach for many households. Commoditized basic access does not eliminate this higher-level divide, particularly when the most valuable capabilities depend on expensive infrastructure and specialized expertise.

The agentic and robotic frontier. The IMF’s exposure estimates predate the current agentic-AI wave; they measure structural task exposure, not realized displacement [5]. Robotics combined with AI could eventually reach manual work, which has so far been less exposed to generative AI. Agentic systems could also reorganize work faster than institutions can adapt. Either development could raise exposure before affected countries and workers have the savings, retraining systems, and public institutions needed to absorb the transition.

Strategic dependency. As education, healthcare, finance, public administration, and other essential services rely more on AI infrastructure, dependence on foreign-owned systems becomes more consequential. Access might be restricted, prices could change, political relationships could deteriorate, or a provider’s interests could diverge from those of the country relying on it. Technological dependency is not automatically strategic vulnerability, but the two become intertwined when a country lacks alternatives, local capacity, or meaningful control over critical systems.

The current balance. Anthropic sees no evidence that low-use countries are catching up—but also none that high-use countries are pulling further away [2]. The gap is persistent rather than runaway. That can coexist with a widening absolute difference: a gap can grow in percentage points while remaining stable in proportional terms. This stability is precarious, however. It depends partly on how quickly open-weight models can be localized and deployed before frontier capabilities and the infrastructure supporting them pull further ahead.

Where this likely lands

The most likely outcome is neither utopian convergence nor catastrophic divergence, but access convergence with benefit divergence. Basic access will become substantially more common over the next few years, though the pace and meaning of “access” will vary with connectivity, affordability, language support, and local institutions. The durable divides will shift up the stack: who has frontier-tier capability, who has the education and institutional support to extract value, who can adapt models to local contexts, and who owns the compute, captures the rents, and holds the power to direct the systems.

The verification layer cuts across all of this. Having access to an AI system is one thing; having people and institutions capable of evaluating its output is another. A country may distribute AI tools widely and still leave its population exposed to confident errors, inappropriate recommendations, or decisions that no local institution has the capacity to challenge. That is why the analog foundations—electricity, schooling, and functional institutions—matter. They are not simply prerequisites for getting AI into people’s hands; they determine whether people can use it effectively, evaluate it critically, and retain meaningful control over decisions made with it.

The deepest question isn’t who can use advanced AI. It’s who directs it: whose problems are built into the systems, whose language gets first-class support, whose institutions can challenge its decisions, and whose economy captures the surplus. A world where everyone consumes AI but a few countries and firms produce and govern it could look egalitarian on the surface while remaining profoundly stratified underneath. That is a failure mode worth watching precisely because it would not announce itself as a divide in access.

One further complication deserves emphasis. We tend to measure AI’s benefits through productivity, employment, education, and income. Those measures matter, but economic convergence does not necessarily produce social or political convergence. A population could become more productive while becoming more economically dependent, less autonomous, or more vulnerable to outside influence. AI’s benefits should not be judged by how much value the technology creates, but for whom that value is created and used.

Methodological caveats

  • Vintage mismatch: The IMF exposure estimates [5] date to January 2024 and predate the agentic-AI wave; they measure structural task exposure, not realized displacement. The 2026 Microsoft, Anthropic, and World Bank sources [1–3] reflect observed usage and are the stronger empirical anchors for the current adoption picture.
  • Measurement limits: Microsoft’s diffusion metric is telemetry-based and currently undercounts some newer tools; the institute says its next revision will raise measured usage, especially in China. Anthropic’s index reflects one vendor’s user base. Cross-country comparisons should therefore be treated as directional rather than precise [1, 2].
  • Evidence versus extrapolation: Several conclusions are interpretations of observed trends, not findings directly established by the cited studies. In particular, longer-term effects of agentic AI, robotics, strategic dependency, and cultural influence remain uncertain. They are treated here as plausible risks or possibilities rather than settled outcomes.
  • Comparability: Historical diffusion comparisons and current adoption measures use different technologies, definitions, and measurement methods. They illustrate the speed of recent diffusion, but should not be read as a like-for-like ranking [3, 4].

Appendix: Evidence map

The notes below indicate which findings in each source support the claims made in the text. They are an aid for readers who want to check the evidence trail without leaving the essay; the essay’s interpretive conclusions go beyond what these sources directly establish, as noted in the methodological caveats.

1. Microsoft AI Economy Institute (2026), The continued state of global AI diffusion in 2026.

  • Supports: 18.8% global working-age usage of generative AI; Global North at 28.8% versus Global South at 16.2%, with the North growing faster in absolute terms; the growing share of open-weight models in API token volume; the telemetry-based nature of the measure and its undercounting of some newer tools, with a revision expected to raise measured usage, especially in China.

2. Appel, R., et al. (2026), Anthropic Economic Index report: Economic primitives.

  • Supports: GDP–usage elasticity (roughly 0.7% higher usage per 1% higher GDP per capita); no evidence of convergence among low-use countries, and none that high-use countries are pulling further away; the technical-worker share explaining nearly two-thirds of cross-state variation within the US; augmentation patterns in high-adoption countries versus delegation and coursework in lower-adoption contexts; the prompt–response sophistication correlation (r > 0.92). Limits: reflects one vendor’s user base.

3. World Bank (2026), World Development Report 2026: The Promise of Artificial Intelligence.

  • Supports: the four-layer framing of the divide (access, capability tier, fit, absorptive capacity); concentration of the advanced AI value chain in a few countries and firms; historical diffusion comparisons (steam ~80 years, electricity ~40, internet ~20, versus the recent rapid spread of ChatGPT traffic in middle-income countries); the role of complements (electricity, connectivity, education, institutions); the larger share of manual and in-person work in developing economies and the associated unevenness of exposure; adaptation examples (voice interfaces on basic phones, local languages, local data); the warning that AI may widen productivity gaps without complementary investments.

4. Stanford HAI (2026), The 2026 AI Index Report.

  • Supports: 53% population adoption of generative AI within three years, faster than the PC or the internet. Limits: not directly comparable to the World Bank’s historical diffusion figures or to Microsoft’s telemetry measure.

5. Cazzaniga, M., et al. (2024), Gen-AI: Artificial Intelligence and the Future of Work (IMF).

  • Supports: roughly 40% of global employment exposed to AI, and about 60% in advanced economies, with advanced economies better positioned to capture the benefits. Limits: January 2024 vintage; measures structural task exposure, not realized displacement, and predates the agentic-AI wave.

6. UN Office of the Secretary-General’s Envoy on Technology & ILO (2024), Mind the AI Divide.

  • Supports: the warning that less-developed countries risk being left further behind as production shifts toward AI-driven systems.

7. De Simone, M. E., et al. (2025), From Chalkboards to Chatbots (World Bank).

  • Supports: six weeks of teacher-guided AI tutoring in Nigeria producing learning gains equated to up to two years of ordinary schooling.

8. Korom, R., et al. (2025), AI-based Clinical Decision Support for Primary Care: A Real-World Study (OpenAI & Penda Health).

  • Supports: a 16% reduction in diagnostic errors with an AI safety net used by clinicians in Kenya, across nearly 40,000 patient visits.

9. University of Oxford (2026), New study warns of risks in AI chatbots giving medical advice.

  • Supports: the finding that raw chatbots show no established advantage over web search or people’s own judgment in the studied medical-advice tasks, because they can mix sound guidance with confident error that novices may struggle to sort.

10. Bastani, H., et al. (2025), Generative AI without guardrails can harm learning (PNAS).

  • Supports: students with an unguarded AI tutor practicing 48% better but scoring 17% worse on the unaided exam, while believing they had learned just as well.

References

1. Microsoft AI Economy Institute. (2026, September 21). The continued state of global AI diffusion in 2026 [Global AI Diffusion Report, Q2 2026]. Microsoft On the Issues. https://blogs.microsoft.com/on-the-issues/2026/09/21/the-continued-state-of-global-ai-diffusion-in-2026/

2. Appel, R., Massenkoff, M., McCrory, P., McCain, M., Heller, R., & Tamkin, A. (2026, January 15). Anthropic Economic Index report: Economic primitives. Anthropic. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report

3. World Bank. (2026). World Development Report 2026: The Promise of Artificial Intelligence. Washington, DC: World Bank. https://www.worldbank.org/en/publication/wdr2026

4. Stanford Institute for Human-Centered Artificial Intelligence (HAI). (2026). The 2026 AI Index Report. Stanford University. https://hai.stanford.edu/ai-index/2026-ai-index-report

5. Cazzaniga, M., et al. (2024, January 14). Gen-AI: Artificial Intelligence and the Future of Work. IMF Staff Discussion Note SDN/2024/001. International Monetary Fund. https://www.imf.org/en/publications/staff-discussion-notes/issues/2024/01/14/gen-ai-artificial-intelligence-and-the-future-of-work-542379

6. United Nations Office of the Secretary-General’s Envoy on Technology & International Labour Organization. (2024, July 26). Mind the AI Divide: Shaping a Global Perspective on the Future of Work. UN/ILO. https://www.ilo.org/publications/mind-ai-divide-shaping-global-perspective-future-work

7. De Simone, M. E., et al. (2025). From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria. World Bank. https://documents.worldbank.org/en/publication/documents-reports/documentdetail/099548105192529324

8. Korom, R., et al. (2025). AI-based Clinical Decision Support for Primary Care: A Real-World Study (OpenAI & Penda Health). arXiv:2507.16947. https://arxiv.org/html/2507.16947v1

9. University of Oxford. (2026, February 10). New study warns of risks in AI chatbots giving medical advice. https://www.ox.ac.uk/news/2026-02-10-new-study-warns-risks-ai-chatbots-giving-medical-advice

10. Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS. https://www.pnas.org/doi/10.1073/pnas.2422633122