A Decade Instead of a Century: Inside the World Bank’s AI Report, and the Stanford Economists Who Shaped It

The steam engine took about 80 years to reach lower-income countries. Electricity took about 40. The internet took about 20. Within six months of ChatGPT’s launch, middle-income countries accounted for half of its global traffic.

Those four sentences, which open the main messages of the World Bank’s new World Development Report 2026, carry the argument of the whole 380-page document. If the technology that matters most this decade is reaching poor and middle-income countries faster than anything before it, then the familiar story, in which the rich world gets the gains and everyone else waits a generation, does not have to repeat. It also does not repeat on its own.

The report, titled “The Promise of Artificial Intelligence,” is the bank’s flagship annual publication, and this year’s edition has an unusual amount of Stanford in it. Its acknowledgments name Susan Athey of the Graduate School of Business as one of the report’s two academic leads. Two more Stanford economists, Nicholas Bloom and Erik Brynjolfsson, sat on its academic advisory committee. A strand of its thinking about growth draws on a 2026 paper by GSB economists Charles Jones and Christopher Tonetti. What follows is a reading of the report itself, the full text of which the bank has posted, alongside what its Stanford contributors have said about it in Stanford GSB Insights.

Who wrote it

The acknowledgments are specific. The report “was prepared by a World Bank team led by Gaurav Nayyar, who was the Director.” Athey and Ufuk Akcigit of the University of Chicago “served as the academic leads,” Akcigit through June 30, 2026. Joyce Antone Ibrahim was the report’s manager. Overall guidance came from Indermit Gill, then the bank’s senior vice president and chief economist, and Somik V. Lall, director of strategy in the development economics vice presidency. Gill also wrote the foreword; Stanford GSB now describes him as the bank’s former chief economist and a distinguished visiting fellow at the Hoover Institution.

The advisory committee reads like a roster of the people who have defined the economics of AI over the past decade. Alongside Bloom and Brynjolfsson are MIT’s David Autor, Toronto’s Avi Goldfarb, Princeton’s Arvind Narayanan, Berkeley’s Ziad Obermeyer and Chicago’s Michael Kremer, among others.

According to GSB Insights, the report took two years of research and consultation with academics, policymakers and business leaders in more than 50 countries. Athey described her own job in an interview with the school: “My role was to develop frameworks that would encompass all of these ideas and then to bring my knowledge of the different strands of research together.”

The central claim: optimism, with conditions

The report organizes its main messages under three imperatives: be optimistic, be strategic, be pragmatic. The optimism is stated plainly. “AI can do in a decade what might otherwise take a century,” reads one heading. The reason offered is not that AI will make poor countries rich by itself, but that it can do cognitive work “that guide[s] decision-making and usually require[s] human expertise, which is often in short supply in developing countries.”

That framing, AI as a substitute for missing expertise rather than for existing workers, recurs throughout. The bank frames its analysis around three words: capabilities, concentration and complements.

Capabilities are what the technology can do. The report traces AI’s evolution “from making predictions to generating content to now autonomously executing complex tasks,” and notes that many of the gains already visible in public services come from older, predictive algorithms rather than chatbots.

Concentration is who controls it. “A small number of companies in just a few economies control the most advanced AI models, the chips they rely on, and the data centers that run them,” the report says. That creates dependency risks. But the report draws a second, less obvious conclusion: concentration “also allows countries to customize AI models without having to spend billions of dollars creating the most advanced AI systems from scratch.”

Complements are what the technology needs around it to work: reliable infrastructure, good schools, strong institutions and data in local languages. “Developing countries often have shortages in all of these areas,” the report notes. “Without the right complements, AI’s impact on productivity and growth will be slow to materialize.”

Nayyar, in comments to GSB Insights, credited Athey with this last emphasis. “Susan really emphasized the importance of complements,” he said, “going back to these very basic things, the nuts and bolts of how economies combine technology with markets and institutions.”

Adopt, adapt, advance

The report’s practical framework is a ladder with three rungs. Countries can adopt existing AI tools, adapt them to local conditions, or advance the frontier by building their own models and the infrastructure behind them. “Each step requires more investment, skills, and infrastructure than the last,” the report says.

Its recommendation is to start at the bottom and put the most effort into the middle. Adopting tools that already exist “is the best place to start.” But adapting, “rather than merely adopting AI, will deliver the biggest benefits.” A tool trained in high-income countries “may provide recommendations not aligned with local norms,” and tools should be designed for people’s real constraints, the report says, such as “AI advice delivered through voice calls on basic mobile phones for those who cannot read or afford smartphones.”

The top rung gets a blunt assessment. Building frontier models “is by far the hardest and most expensive path,” requiring chips, large data centers, huge training datasets and world-class researchers. “For most developing countries, building frontier AI models is not a realistic goal in the near future.”

Athey compressed the idea into four words at a one-day event on the report held at the GSB, as quoted by Insights: “Adaptation enables adoption.”

Small AI

The report’s most concrete material is in a box titled “Small AI, big impact.” Small AI, it explains, can mean a model with fewer parameters or a tool built for “one narrow task, such as identifying a crop disease or screening a medical image.” Either way, it “is built to run where the complements it depends on, such as power, internet connectivity, and digital devices, are scarce.”

The examples are specific, and the report cites a study for each:

Nuru, a phone app that can work offline, “outperformed agricultural extension officers at diagnosing a diseased leaf.”

Cough Against TB, built by Wadhwani AI with India’s Central Tuberculosis Division, “correctly flagged 9 in 10 people while screening for tuberculosis from the sound of a cough.”

In a randomized trial across 40 Kenyan health facilities, mothers using the PROMPTS chatbot “gained knowledge and sought care for their children and themselves more often.”

In Ghana, Rori provides math tutoring by text message on basic phones. According to the report, it “generated nearly one full year of learning from two 30-minute weekly sessions over eight months at a cost of about five dollars per student.”

Elsewhere, the report describes Kenya’s judiciary using AI “to assign more than 10,000 court cases each year to more than 1,500 mediators, helping reduce a large backlog of cases,” and teachers in Sierra Leone who “increasingly used an AI assistant on WhatsApp instead of a standard web search because it provided cheaper and more useful advice on matching lessons to students’ skill levels.”

Brynjolfsson told Insights the emphasis matched his own findings: “The biggest gains often go to the people who previously had the least access to expertise.” The report cites his study of customer support agents, which GSB Insights describes as covering roughly 5,000 workers, mostly in the Philippines, and finding that generative AI helped the newest and least-skilled workers most.

Jobs: a different picture from the American debate

In the United States, the AI conversation tends to start with job losses. The report starts somewhere else, because the structure of work is different. “In low-income and lower-middle-income countries, almost 90 percent of workers are employed in businesses with fewer than 10 people, and more than half of workers are self-employed,” it notes. With most people in manual rather than cognitive work, the report’s chapter summary concludes, “the immediate impact of AI adoption is to complement rather than displace workers.”

Adoption is already further along than many readers might assume. Enterprise surveys conducted for the report found that among formal firms with more than five employees in India, Jordan, Kenya, Mexico, Nigeria and Thailand, “on average, about 20 percent have recently used an AI chatbot in their business operations, compared with 34 percent in the United States.”

The report does not wave away displacement. It points to evidence of “nontrivial displacement effects” of generative AI in China, concentrated among entry-level, high-wage and highly educated workers, and to emerging evidence that in South Asia generative AI “has already reduced monthly job listings by about 20 percent for the most exposed white-collar occupations that have ready substitutes.” It flags Kenya, where knowledge-intensive services such as ICT, professional services and finance accounted for “only 2 percent of jobs but … 19 percent of growth between 2006 and 2019,” as an example of how a small, exposed sector can matter out of proportion to its size.

Its scenarios turn on whether AI moves into physical work. Under a moderate scenario, the share of employment exposed to AI in low- and middle-income countries reaches only 25 percent by 2035. In an advanced scenario with breakthroughs in physical AI, “three-quarters of occupations become exposed,” and the exposed share of employment in those countries is projected to exceed 50 percent.

The weakest link

The report’s caution comes from an idea economists call bottlenecks, and here the Stanford connection is direct. Among the sources the report cites is a 2026 paper by Jones and Tonetti. As GSB Insights summarizes their work, economic output is limited by its worst bottleneck, much as a chain is only as strong as its weakest link.

“It’s really important to try to understand that there are barriers to technology adoption that are longstanding,” Tonetti told Insights, “and those same barriers might affect the adoption of AI just like they affected adoption of previous technologies.”

The report gives the idea a vivid illustration. It recounts a Google Health deployment of AI-assisted diagnosis in clinics: once the clinics had the tool, “the limiting factor shifted to the quality of lighting, which affected the clarity of the images that the clinics’ nurses took.” Uploads that took seconds in the lab “took more than a minute when it was deployed in the clinics,” and in one clinic “the internet was down for two hours on one of the days in the study.” Its conclusion: “accruing the full potential of AI requires managing a series of moving bottlenecks that may take a long time to clear.”

The numbers behind those bottlenecks are stark. In low-income countries, according to the report, “less than half of the population has access to electricity, only about one-quarter has access to the internet, and the use of smartphones is far from universal.”

Athey put it in a phrase at the GSB event, as quoted by Insights: “AI is going to expand capacity in some dimensions, but in the end, the overall impact is determined by the tightest part of the bottle.”

Too many pilots

One problem the report wrestles with will be familiar to anyone who has watched development aid: lots of experiments, little evidence about which ones work. The report’s chapter on public services finds that “the lack of evidence for selecting, procuring, and evaluating AI applications, especially at the front end, limits AI’s full potential.”

Athey, according to Insights, heard the problem stated directly while visiting India: “The last thing we need is another agricultural advisor app. What we really need is one good one.” Lall, speaking at the GSB event, described the goal as moving “away from doing a million pilots to figuring out what sort of pilot is useful, what are the metrics to use to evaluate the success of some pilots, how to kill bad ones, and how to scale up the good ones.”

The report’s answer is to make governments better buyers. As users of AI, it says, governments “should use their buying power to test, evaluate, buy, and expand successful AI solutions in uses that matter most for public services: for example, agricultural extension, health care, and education.” As regulators, they should start from voluntary industry standards and use existing law against concrete harms, “while leveraging international cooperation to develop AI regulations that avoid creating unnecessary barriers.”

The risks the report names

For an optimistic document, the report is direct about the downside. Its chapter on political impact warns that control of frontier models, advanced chips and data centers by a few companies “creates dependency risks, but pursuing ‘AI sovereignty’ comes at a high cost and with uncertain benefits.” It cautions that incentives to attract AI investment “squeeze public budgets yet provide uncertain returns,” that AI “can benefit owners of capital more than people because of the concentration of market power,” and that it “can spread propaganda and enable the surveillance of citizens.” Chapter 1 notes simply that “AI can also enhance the harmful capabilities of malicious actors,” a point Athey echoed at the GSB event, noting that bad actors face fewer impediments to adopting AI.

Why it matters here

Stanford has produced plenty of AI research about rich countries. This report is a reminder that some of the university’s most consequential economics is being written for the 6.8 billion people GSB Insights says live in low- and middle-income countries, and that the questions there are different. They are less about whether a chatbot will replace a paralegal and more about whether a farmer with a basic phone can get a forecast, or whether a clinic’s lighting is good enough for the software to work.

Gill’s foreword, as quoted by Insights, frames the stakes: “Until now, concrete evidence about AI’s impact on the developing world has been scant. This report starts to fill that void.” The bank says it will use the report to guide policy advice in the countries where it lends.

Whether the “decade instead of a century” promise holds will depend on the dull work the report keeps returning to: power lines, schools, data in local languages and the discipline to stop funding pilots that do not work. Athey, in her remarks at the GSB event, explained why she thinks it is worth the effort. “As a percentage change from the baseline, those benefits could be much larger in developing countries,” she said. “That’s what really gets me out of bed in the morning to work on this.”

Photo: Christopher Michel / Wikimedia Commons, CC BY-SA 4.0. Susan Athey, professor of economics at Stanford Graduate School of Business, in 2025.