By Dyuti Khulbe/Public Praxis Forum

For decades, technological revolutions have arrived in developing countries after the infrastructure, expertise and markets around them were already established elsewhere. The internet changed that pattern to some extent. Artificial intelligence may change it even more.

A farmer does not necessarily need access to a sophisticated research laboratory to use an AI-generated weather forecast. A health worker does not need to build a large language model to use one for medical assistance. A teacher can use an AI tool to prepare lessons without having access to the computing infrastructure on which that tool was created.

This is the opportunity at the centre of the World Bank’s World Development Report 2026: The Promise of Artificial Intelligence. The report argues that AI is spreading far more rapidly than previous general-purpose technologies and could give developing economies access to forms of expertise that have traditionally been scarce. Middle-income countries, for example, accounted for half of ChatGPT’s global traffic within six months of its launch.

The Bank’s proposition is therefore quite ambitious. Developing countries do not need to wait for AI to mature elsewhere and then gradually catch up. They can adopt, adapt and advance the technology themselves.

But there is a tension running through the report that is easy to miss when AI is presented simply as the next great development opportunity.

The technology may be new. The conditions determining whether it actually improves people’s lives are not.

The leapfrog story

The appeal of AI for development is easy to understand.

Many developing countries face persistent shortages of specialised expertise. There may be too few doctors, agricultural extension workers, teachers, engineers or public officials with the time and resources to meet demand. AI can potentially extend some of that expertise much further.

The World Bank gives several examples. AI-supported medical screening can assist health workers where specialist expertise is limited. Weather forecasting can help farmers make better production decisions. Teachers can use AI to develop lessons and educational material. These applications do not require a country to create its own frontier AI model; they require the ability to use existing technology effectively.

This is where the idea of technological leapfrogging becomes attractive. A country that lacks some of the infrastructure associated with earlier technological development may be able to move directly towards newer systems instead of following exactly the same path taken by richer countries.

The mobile phone offers an earlier example. Many places that never developed extensive fixed-line telephone networks moved relatively quickly into mobile communications. AI could produce a similar effect in certain areas, particularly where it can be delivered through relatively simple technologies.

The World Bank is particularly interested in what it calls “small AI”: tools that can deliver useful forms of expertise through text messages, voice calls and basic phones, even where access to high-end computing and reliable internet remains limited.

There is a significant development proposition here. AI does not necessarily have to reproduce Silicon Valley inside every country to become economically and socially useful.

But leapfrogging has a catch.

Using a technology is not the same thing as building the capabilities around it.

The technology may travel faster than the system

Consider a government introducing an AI system to help agricultural workers advise farmers.

The software may be excellent. Yet its usefulness still depends on things that have little to do with the software itself.

Does the system understand the local language? Is the data it relies on relevant to local conditions? Do agricultural workers have the skills to interpret its recommendations? Is there reliable electricity and connectivity? Can farmers actually access the advice? What happens when the recommendation conflicts with local knowledge or conditions that the system cannot see?

The World Bank is unusually explicit about this. Its report describes infrastructure, education, institutions and data as the “complements” that determine how much value AI can generate. Without them, the technology may exist without producing much development impact.

This is also why the report distinguishes between adopting AI and adapting it.

Adoption is relatively straightforward: take an existing tool and use it.

Adaptation is more demanding. It means making the technology useful in a particular environment, whether through local languages, locally relevant data, different delivery systems or applications designed around specific needs. The report argues that adaptation may generate some of the largest benefits for developing countries.

That distinction is important because development has rarely been about transferring a finished solution from one place to another.

Ideas travel. Technologies travel. Policies travel. But they encounter institutions and societies that were not part of the original design.

AI will be no different.

Who gets to build the intelligence?

There is another complication.

Most developing countries are not going to build the world’s most advanced AI models. The World Bank does not suggest that they should. Developing frontier AI requires enormous amounts of computing power, data, capital and highly specialised expertise, making it an unrealistic priority for most countries in the near term.

Instead, countries can use models developed elsewhere and concentrate on adapting them to their own circumstances.

That is pragmatic. It is also where questions of dependence enter the picture.

The most advanced models, chips and data centres are concentrated among a relatively small number of companies and countries. This concentration gives developing countries access to technologies they would otherwise struggle to create themselves, but it also creates dependency risks. The World Bank itself identifies this concentration as one of the central characteristics of the current AI economy.

There is an interesting contradiction here.

The same concentration that makes advanced AI accessible can also make countries dependent on the companies controlling it.

A government may therefore gain access to powerful technology without gaining much control over the technology itself. It may be able to use an AI model without understanding how its underlying infrastructure is built, how its training data was assembled, or what would happen if the provider changed its prices, terms or technical standards.

This is not necessarily a reason to reject imported technology. Development has always involved international flows of capital, knowledge and technology. The more important issue is whether access gradually creates capability—or dependence.

That distinction becomes particularly important when AI starts entering public institutions.

When governments become AI users

Governments are already experimenting with AI in areas such as taxation, healthcare, education, public administration and fraud detection. A recent World Bank survey covering governments in 60 economies found widespread experimentation, but also considerable informal use. 44% percent of reported internal government AI use consisted of individual public servants using AI for ad hoc tasks rather than through a formal institutional system.

That small distinction says quite a lot.

An employee using an AI tool to summarise a document is one thing. A ministry incorporating an AI system into a decision-making process is something else entirely.

The second requires procurement rules, data standards, oversight, training, accountability and institutional capacity. It also raises questions about who is allowed to challenge an AI-assisted decision and what happens when the system gets something wrong.

In other words, technology becomes part of an institution.

And once that happens, the development story becomes more complicated than a question of technological access.

A government may use AI to process applications faster, for example. That could be a genuine improvement. But speed is only one measure of a public service. If the system is difficult for citizens to understand, excludes people who lack digital access, or produces decisions that officials cannot adequately explain, efficiency may come with other costs.

The same technology can therefore look very different depending on the institutional purpose it is serving.

Development is not a single outcome

This brings us back to something older than AI.

For much of modern development thinking, progress was closely associated with economic growth, industrialisation and productivity. Later approaches expanded the field considerably, bringing poverty, inequality, human capabilities, institutions, participation and environmental sustainability into the discussion.

Amartya Sen’s capability approach was particularly influential in shifting attention towards what people are actually able to do and be, rather than treating income or resources as sufficient measures of development.

That perspective matters when thinking about AI because technological capacity can increase without necessarily increasing people’s capabilities.

An AI system might make a government department faster. It might increase a firm’s productivity. It might allow a doctor to screen more patients.

Those are meaningful gains.

But development is also concerned with what those gains enable people to do with their lives.

A faster administrative system that is inaccessible to large sections of the population does not tell the same story as a faster system that genuinely expands access to public services. An agricultural technology that increases yields while leaving farmers more dependent on expensive external inputs raises different development questions from one that strengthens their long-term resilience.

The distinction is subtle but important: AI can improve an activity without necessarily improving the wider system in which that activity takes place.

This is where the World Bank report becomes particularly interesting from a development perspective. Its own emphasis on local adaptation implicitly recognises that the value of AI cannot be separated from the context in which it operates.

The knowledge problem

There is another reason adaptation matters.

AI systems are built from data, and data is never simply a neutral representation of the world.

What gets recorded, measured and included in datasets depends on institutions, resources and decisions about what is worth collecting. Some communities are heavily represented in official data. Others may appear only partially, or not at all. Some forms of knowledge are easily converted into datasets; others are embedded in professional practice, local experience or oral traditions.

This creates an important challenge for AI in development.

A system can be technically sophisticated and still have a limited understanding of the people and environments it is being used to serve.

The World Bank recognises part of this problem through its emphasis on local languages and local data. Its recommendation to adapt AI rather than simply import it reflects an understanding that technologies trained and designed in one context may not automatically perform well in another.

But adaptation also requires deciding what counts as relevant local knowledge in the first place.

That is not merely a technical decision.

It involves researchers, governments, companies, development agencies, professionals and communities. Each may understand the same development problem differently. A ministry may see agricultural productivity as the priority; farmers may be more concerned about market access or debt; a technology company may see an opportunity for automation.

AI does not settle those differences.

It enters them.

A new technology, familiar development questions

Perhaps this is why the most interesting part of the World Bank report is not its prediction about what AI will become.

It is the way the report describes the conditions needed for AI to become useful.

Infrastructure matters. Skills matter. Institutions matter. Data matters. Local adaptation matters. The ability to work across multiple platforms and providers matters. The report even warns against allowing dependence on a single supplier to undermine countries’ choices later on.

These are not exclusively AI questions.

They are questions about development capacity.

In that sense, AI may be less of a break from earlier development debates than it initially appears. It introduces an extraordinary new technological capability, but it still has to pass through the familiar structures that determine how societies turn resources into outcomes.

The difference is that AI may move through those structures at unprecedented speed.

That could be enormously valuable. It could also make existing weaknesses more visible.

A country with strong institutions may be able to experiment, evaluate and adjust. A country with weak data systems may struggle to know whether an AI programme is actually working. A government with limited technical expertise may become dependent on private providers. A public institution with little capacity for oversight may adopt a system before it has worked out how to govern it.

The technology does not produce these institutional differences. It exposes them.

Beyond the promise

The World Bank calls its report The Promise of Artificial Intelligence. The choice of word is revealing.

A promise describes a possibility, not an outcome.

AI may allow developing countries to extend scarce expertise, improve services, raise productivity and tackle problems that have resisted conventional approaches. The evidence emerging already suggests that some of these possibilities are real.

But development has never been produced by technology alone.

It emerges through the interaction of technologies with institutions, markets, public policy, social relationships and people’s own capacities to use what is available to them. AI enters that same world, carrying enormous potential but also the assumptions, dependencies and power relationships attached to the systems that produce it.

For Public Praxis Forum, that is perhaps where the World Bank’s report becomes most useful. It gives us a contemporary example of a much older development problem: how does an idea become useful when it moves from one context into another?

The answer cannot be found in the technology alone.

AI may help countries move faster. Whether they move towards greater capability, greater dependency, broader opportunity or deeper inequality will depend on the institutions and choices surrounding it.

This also connects with an earlier PPF essay, Who Is Responsible When AI Gets It Wrong?, which looked at AI from the perspective of responsibility and accountability. Here, the concern is broader. Before asking who is responsible when an AI system causes harm, there is another layer to consider: what happens when we place that system inside a particular development context in the first place? The institutions choosing the technology, the problems they are trying to solve, the data they rely on and the people affected by its use all shape what the technology eventually does.

The promise, then, is real.

What happens with that promise is a development question.

Featured image: Photo by Kampus/Pexels