The Promise of Artificial Intelligence

The World Development Report 2026, titled The Promise of Artificial Intelligence, is the first comprehensive assessment of what AI means for the 6.8 billion people living in low- and middle-income countries. Its argument is that AI is a historic opportunity for those countries, one that could let them solve development problems that have resisted solution for decades, and that the binding question is not whether the technology arrives but whether the foundations for using it are in place when it does. Because so little was known about how AI is actually used outside high-income countries, the Report set out to gather that evidence itself, documenting how businesses and governments in developing countries use AI today and what stops them from using it more. The World Bank team was led by Gaurav Nayyar as Director; Susan Athey of the Stanford Graduate School of Business and Ufuk Akcigit of the University of Chicago served as the academic leads.

AI is a general-purpose technology, which puts it in the company of steam power, electricity, and the computer and the internet. Each of those transformed whole economies rather than single sectors, and each was slow and uncertain at first. AI is following the same pattern on a much shorter timeline. The Report notes that the steam engine took about 80 years to reach lower-income countries, electricity about 40, and the internet about 20, while middle-income countries accounted for half of ChatGPT's global traffic within six months of its launch. The historical lesson it draws is pointed. The countries that gained most from earlier general-purpose technologies were rarely the ones that invented them. They were the ones that had built the infrastructure, skills, and institutions those technologies required, and that moved quickly to build applications on top of them. The countries that did not invest in those foundations gained least and absorbed the most disruption.

The analysis turns on three things that determine what AI is worth to a developing economy. The first is capabilities. Unlike the computer or the internet, AI performs cognitive tasks that ordinarily require human expertise, such as diagnosing disease, forecasting weather, or advising a farmer, and that expertise is precisely what is scarce in most developing countries. The second is concentration. The most advanced models, the chips they run on, and the data centers that house them are controlled by a small number of firms in a few economies, which creates dependency risk and shapes whether the available tools fit developing-country needs, but which also means a country can customize a capable model without spending billions to build one. The third is complements, the foundations that let AI be used safely and productively: reliable infrastructure, good schools, capable institutions, and data that reflects the local context. Ten years ago the World Development Report 2016 called these the analog complements to digital technology. This Report argues that AI needs strong analog and digital complements both.

Its central prescription is adopt and adapt, then advance. Advancing, meaning building frontier models and the infrastructure behind them, is the hardest and costliest path, and the Report judges it unrealistic for most developing countries in the near term, noting that the largest US AI companies are expected to spend more than $750 billion on AI infrastructure in 2026 alone, more than the annual output of many economies. A few countries have made inroads, among them India's Sarvam AI, the United Arab Emirates' Falcon, and InkubaLM in Sub-Saharan Africa, each built to work better in local languages. Adopting existing tools is the place to start, but the Report is explicit that adoption alone is not enough, because models trained in high-income settings often fail elsewhere. In Nigeria, a tool trained on high-income data systematically over-recommended laboratory tests, having learned medical norms better suited to Los Angeles than to Lagos. Adaptation is where the Report expects the largest gains, through open models such as Llama, Mistral, Gemma, and Qwen, and regional efforts such as Latam-GPT and SEA-LION.

The case for optimism rests on gains that are already measurable, and on the observation that the useful tools are frequently small rather than frontier. Small AI, whether small in model size or narrow in scope, is grounded in a specific problem and built for the complements that already exist in low-resource settings, which is why it can reach people through text messages, voice calls, and basic handsets. In the Indian state of Telangana, an AI weather forecasting service led smallholder farmers to change their behavior in response to specific risks, with some increasing farm spending by as much as a third and others saving as much as $560 each. Kenya's judiciary used AI to assign more than 10,000 court cases a year to more than 1,500 mediators, cutting a large backlog. In Bangladesh, AI-generated medical imaging raised the number of patients screened each day for a diabetic eye complication by nearly 40 percent. In Ukraine, a machine-learning tool built by Transparency International Ukraine flags suspicious procurement patterns and routes citizens' complaints to law enforcement. That reach matters, because in low-income and lower-middle-income countries nearly 90 percent of people work in firms of fewer than ten and more than half work for themselves.

On risk the Report is deliberately unexcited. AI is not yet displacing large numbers of workers in most developing countries, and the exposure figures cut against the headlines: about 4.5 percent of existing jobs in low- and middle-income countries are amenable to automation by generative AI, against 14.2 percent in high-income countries, while 16.2 percent are amenable to being complemented by it. The pressure is concentrated rather than general. Business process outsourcing, the call centers and back-office work that helped India and the Philippines build employment, is where automation bites first, and one large freelance platform reported that work outsourced to developing countries fell 39 percent in 2025, though the wider evidence is mixed. The social risks are bias from training data that does not reflect local people and languages, and cheaper fraud, cybercrime, and false information running against thin defenses: in 2025 Africa had 66 active fact-checking organizations serving 1.5 billion people, against 139 in Europe serving 750 million. The political risk is dependence on systems controlled elsewhere.

What the Report asks of governments falls in three parts. Keep investing in the analog foundations, which remain weak: in Sub-Saharan Africa nearly a third of rural schools still lack reliable electricity, more than two-thirds lack dependable internet, and nearly nine in ten ten-year-olds cannot read a simple text. Build the digital building blocks, meaning affordable computing and shared local-language data, helped by the fact that developing countries now account for about 40 percent of new foreign direct investment projects in data centers worldwide. And treat evidence and trust as policy problems in their own right, since cheap coding tools have produced thousands of pilots and the hard question is which of them work, and since regulators with limited capacity should start from voluntary industry standards on impact assessment, auditing, and testing, applying existing law to concrete harms while they develop rules that do not obstruct use. The closing warning is about timing. More than 80 countries had published national AI strategies by mid-2026, but only one of the world's 25 low-income countries, Rwanda, had done so. The countries with the most to gain are today the least prepared, and the Report's advice to them is to be blinded neither by the hype nor the hysteria, and to adopt, adapt, and then advance without delay.