Artificial Intelligence as an Accelerator of Global Development

Across the 120 countries for which the Anthropic Economic Index publishes country data, the median usage per capita index is 2.40 in high-income countries (usage 2.4 times what population alone would predict), 0.66 in upper-middle-income, 0.31 in lower-middle-income and 0.11 in low-income countries, where only five countries (population 158 million of 767 million) clear the publication floor. The countries at the top are Australia (6.4), Singapore (5.8), Switzerland (5.0), Luxembourg, New Zealand and Canada; at the bottom, Tanzania (0.07), Mozambique, Angola, Madagascar, Bangladesh and Uganda (0.11). The median of the top ten is 41 times the median of the bottom ten.

High-income countries generate 57.4% of the usage, upper-middle-income countries 16.4%, lower-middle-income countries 12.6% and the five published low-income countries 0.2%; another 13% comes from countries below the publication floor or unattributed. For comparison, high-income countries have 17% of the world's population, upper-middle-income countries 38%, lower-middle-income 36% and low-income 9%. Publication is itself informative: countries with too few conversations to publish are, by construction, ones where use is rare, so the low-income figures are an upper bound on the true picture.

Microsoft's independent telemetry-based series tells the same story. In the second half of 2025, 24.7% of working-age people in the Global North and 14.1% in the Global South used generative AI (world 16.3%), a gap that widened from 9.8 to 10.6 points. By the first quarter of 2026 it was 27.5% against 15.4% and by June 28.8% against 16.2%, with the world at 18.8%. Both series agree that AI is diffusing everywhere and that the gap is widening, not closing; the first report of this series, on the architecture of diffusion, explains why.

Globally, 43% of Claude.ai conversations are work, 40% personal and 16% coursework. The mix shifts with income. In high-income countries coursework is 13% of use and personal use 45%; in lower-middle-income countries coursework is 24% and personal use 34%; in the five published low-income countries coursework is 35% and personal use 27%. Work use is stable across groups, at 39–45%. In topic terms, education and learning accounts for 11% of requests in high-income countries, 16% in upper-middle-income, 20% in lower-middle-income and 25% in low-income countries; software development rises from 9% to 14–16%; hobbies and lifestyle falls from 12% to 4%.

The picture is of AI as a tutor and a skills tool: where schooling and professional training are thin, people use it to learn and to code. That is consistent with the Nigerian tutoring evidence, and with a development case for AI as a substitute for scarce instruction. The index's own estimates also show a larger gain: the estimated ratio of human-only time to time with AI is 6.3 in high-income countries and 10.2 in low-income countries, suggesting that the tasks done in poorer countries are ones where AI saves proportionally more time; but these are model-estimated times and should be treated as indicative.

What is missing is as informative. Farming, fishing and forestry occupations account for 0.04% of usage globally, against agriculture employing 59% of workers in the median low-income country; healthcare practitioners account for 3.9% of use in high-income countries and 2.0% in low-income countries. These are data about one assistant's consumer interface, and farmers and health workers may reach AI through apps, voice, SMS and clinical tools that this dataset cannot see. But it means the sectors with the greatest capacity gaps are not yet showing up in general-purpose AI usage, and that a development strategy cannot assume that access to a chatbot is access to AI-enabled services.

A cross-country regression of the log usage index on income, the diffusion index (connectivity and service coverage, from report 2026-75), secondary enrolment, government effectiveness and an English-language indicator explains 85% of the variation (117 countries). Income alone explains 80%, with an elasticity of 0.94: a country with twice the income uses AI about 1.9 times as much per person. Shapley decomposition splits the 85 points as income 26, institutions 21, connectivity 19, skills 18 and English 0, with the caveat that income, schooling, institutions and connectivity are highly collinear, so their shares are shared.

Two results are worth stressing. First, English is not a significant driver: its coefficient is −0.01 (standard error 0.04), and the median usage index is lower among countries with English as an official language (0.80) than those without (1.06), partly because of income mix. Language support in modern assistants appears to have broken the old link between language and access. Second, the IMF's AI Preparedness Index, which combines digital infrastructure, human capital, innovation and regulation, correlates 0.91 with the log of usage (0.86 with usage itself), but it tracks income too: preparedness and wealth are almost the same thing at this level of resolution.

The residuals show where countries break from the pattern. Above prediction: Lebanon, Tunisia, Benin, Cameroon, Mongolia and the Palestinian territories among developing economies, plus Australia and Malta. Below: Bangladesh (usage a third of what its income and capacity predict), Angola, Tanzania, Saudi Arabia, Türkiye, Uzbekistan, India and Egypt. India's usage index of 0.30 is lower than its capacity predicts; that is also because the data cover one consumer assistant, and India's AI use may be concentrated in other tools and languages. These residuals are questions for further study, not verdicts.

The best-identified results come from randomised or quasi-experimental studies of specific deployments. Across 5,172 customer-support agents, access to a generative AI assistant raised issues resolved per hour by about 14–15%, and by about 30% for the least experienced workers (34% in the working-paper version), with little effect on the most experienced (Brynjolfsson, Li and Raymond, published in the Quarterly Journal of Economics in 2025). In Edo State, Nigeria, a six-week after-school programme using a generative AI tutor raised test scores by about 0.3 standard deviations, which the World Bank equates to 1.5–2 years of typical schooling. In 15 Nairobi clinics run by Penda Health, across 39,849 visits, clinicians with an AI decision-support tool made 16% fewer diagnostic errors and 13% fewer treatment errors, though the study found no significant difference in patient-reported outcomes; it was financially supported by OpenAI, whose co-authors ran the analysis with Penda's. A machine-learning flood-forecasting model published in Nature in 2024 gave reliable forecasts up to five days ahead in basins without river gauges, and Google reports that its Flood Hub now reaches users in more than 100 countries.

The pattern in those studies is that AI narrows gaps in skill and experience: gains are largest for novices, students and clinicians without a specialist at hand. That is the development case. But it comes with limits. The Nigerian study was short and teacher-supported; the Penda result is a reduction in errors, not in harm; the support-centre study covered one firm and chat support; and none shows what happens when the tool is scaled across a national system with unreliable power, connectivity and supervision. The results show what is possible with a well-designed deployment, not what general access delivers.

At the macroeconomic level, the estimates diverge sharply. Acemoglu's task-based calculation puts the ten-year gain in total factor productivity at no more than about 0.66% (0.53% allowing for harder tasks) and in GDP at about 0.9–1.2%, because only a small share of tasks is both exposed to AI and profitable to automate. Goldman Sachs projects a 7% increase in world GDP over ten years. The two differ in the share of tasks affected, the cost savings, and the speed of adoption, and the calculator in section 6 lets the reader see which assumptions each implies. For developing economies the question is as much about adoption as about technology: an economy where 3% of workers use AI will not see the gains of one where 40% do.

The medians tell the story. The median low-income country has 0.14 physicians per 1,000 people, 41 pupils per primary teacher, 59% of employment in agriculture, a diffusion index of 40 and an AI Preparedness Index of 0.30; the median high-income country has 3.6 physicians, 13 pupils per teacher, 3% agricultural employment, a diffusion index of 96 and a preparedness index of 0.65. Lower-middle-income countries sit between, with 0.38 physicians, 30 pupils per teacher, 36% agricultural employment and indices of 66 and 0.36.

In total, 3.8 billion people (47% of those in countries with data) live in countries with fewer than one physician per 1,000; 1.2 billion of them are in countries where the diffusion index (connectivity, power, water, vaccination, accounts) is below 60. 1.3 billion live where there are more than 35 pupils per primary teacher, of whom 0.9 billion are below the same readiness threshold. 2.6 billion live where more than 40% of employment is in agriculture, of whom 0.85 billion are in low-readiness countries. Thresholds are the report's own choices.

This is the central tension. These are the places where, on the evidence of the field studies, AI assistance could do most to stretch a thin workforce: AI-assisted triage, tutoring, advisory services. They are also places where power, connectivity, devices, skills and local content are least ready, and where the first wave of general-purpose AI has not arrived. Closing the gap requires that foundation: the diffusion architecture of the previous report. AI will not be an accelerator of development where electricity and connectivity are missing.

The calculator turns evidence into arithmetic. Health: the country's physicians (physicians per 1,000 × population) multiplied by the share assisted and the effective capacity gain gives equivalent extra physicians, which it compares with the number needed to reach one physician per 1,000. Education: the learning gain for students reached is multiplied by the share reached; about five years of schooling per standard deviation is the equivalence implied by the Nigerian trial (0.3 SD ≈ 1.5 years). Economy: a Hulten-style calculation: the increase in productivity is the share of the workforce using AI × share of tasks exposed × cost saving, and output rises by roughly 1.5 times that once capital responds (Acemoglu's GDP estimates are 1.4–1.8 times his productivity estimates). It is shown against the two reference estimates (about 1% and 7% in ten years).

With its default health settings (30% of health workers assisted at a 15% gain) Tanzania gains the equivalent of 4.5% more physicians, which closes under 1% of its gap to one physician per 1,000, because that gap is so large (61,000 physicians). In Nigeria, even if every physician were assisted at a 30% gain, the gap would close by only 18%. AI can stretch the workforce; it cannot substitute for training one. In the economy panel, a country where 10% of the workforce uses AI on a tenth of tasks at a 20% saving gains about 0.3% of output in ten years; to reach the 7% scenario, adoption, exposure and cost saving all have to be many times higher.

The calculator is a way to see assumptions, not a forecast. The effect sizes are drawn from studies of specific deployments, mostly in better-resourced settings; scaling them to national systems assumes away the problems of power, connectivity, devices, skills, supervision and trust that section 5 describes. It also ignores costs, displacement, distributional effects and risks of error. Its use is to show how much would need to be true, and where, for AI to be the accelerator the optimistic estimates imply.

Foundations first, and in parallel. Income explains four fifths of AI use, and connectivity, power and skills are what income buys. Electricity, affordable data, devices and schooling are preconditions for AI to be an accelerator rather than another gap. Distributed solar, shared community devices and low-bandwidth, voice and SMS interfaces can extend reach before the foundation is complete, and they are where public and donor money can lower the threshold.

Point AI at scarce capacity. The evidence favours deployments that support the workers who are scarce: tutoring for learners, decision support for clinicians, advisory for farmers, within systems that have teachers, nurses and extension agents to supervise them. Public procurement, open evaluation and shared models for local languages and contexts are public goods that the market under-supplies in poorer countries. General-purpose chat is not enough: farming barely registers in it.

Measure, publish and protect. Evidence is thin: a handful of trials, several funded by developers. Randomised evaluations of deployments in developing countries, shared data on usage and outcomes (including from providers, as the Economic Index and Microsoft's reports begin to do) and attention to risks (errors, privacy, displacement, concentration of capacity in a few firms) are conditions for turning promising pilots into policy. The question is not whether AI accelerates development, but which deployments, under what conditions, for whom.

Usage: Anthropic Economic Index release of 26 June 2026 (Claude.ai chat and Cowork; Free, Pro and Max plans; data for May 2026), country rows for 120 countries above the publication floor (World Bank classification). Metrics: usage share, usage per capita index (share of usage ÷ share of 15–64 population), use-case shares, collaboration modes, estimated human-only and with-AI times (model-estimated), request topics (major groups) and SOC occupation groups. Income-group figures are weighted by each country's usage share. The data come from one provider and one family of products, and the developer of that product also built this report's assistant; the Microsoft AI Diffusion Report is used as an independent cross-check. Population, GDP per person (PPP), government effectiveness, secondary enrolment, physicians, pupils per teacher and agricultural employment from World Bank WDI (latest year since 2010–2020); IMF AI Preparedness Index (2023); diffusion index from H Heuristics report 2026-75; English-official-language flag from GeoNames.

Regression: OLS of ln(usage index) on ln GDP per person, diffusion index, secondary enrolment (capped at 120), government effectiveness and an English-official indicator, standardised, 117 countries; Shapley decomposition of R². Scarcity thresholds: physicians under 1 per 1,000, over 35 pupils per primary teacher, over 40% of employment in agriculture; low readiness = diffusion index under 60. Calculator: health: equivalent physicians = physicians × share assisted × capacity gain, gap = max(0, 1 per 1,000 × population − physicians); education: learning gain (years) = share reached × effect (SD) × 5 years per SD; economy: ΔTFP = adoption × exposure × cost saving, ΔGDP = 1.5 × ΔTFP over ten years; the default adoption is a proxy of 18.8% (Microsoft, June 2026) × usage index, capped at 90%.

Caveats: usage data cover one assistant's consumer interface, not all AI use; unpublished countries (mostly low-income) are absent and, by construction, low-use; topic and occupation classifications are model-assigned; time estimates are model estimates; cross-country regressions are associations, and the predictors are collinear; field-experiment effect sizes come from specific settings and some from developer-funded studies; the calculator's parameters are assumptions; macro estimates differ in method and are not forecasts.