The World Bank is making a case for a different kind of artificial intelligence strategy in developing economies: smaller, specialized models designed to work in local languages, on mobile phones and in places where reliable high-speed connectivity cannot be assumed.
Speaking around the World Bank and International Monetary Fund annual meetings in Bangkok, the bank’s chief knowledge officer, Paschal Donohoe, argued that this “small AI” approach could help countries translate digital capability into practical development gains. The emphasis is not on replicating the largest general-purpose systems being built by major technology companies. It is on making narrowly trained tools available where people need them, for clearly defined tasks.
AI designed for local conditions
The distinction matters because access shapes usefulness. A powerful cloud-based service may be difficult to use consistently in communities with constrained connectivity, limited computing resources or linguistic needs that receive little attention from global consumer products. A smaller model trained for a specific task can potentially run on a phone and operate with less dependence on a constant connection.
Donohoe’s description places the local setting at the center of the design. In this framing, the relevant questions are not simply how capable a model is on a broad benchmark, but whether it can deliver information in the language used by a community, through the devices that community already has, and under real-world network conditions.
That is a significant shift from much of the current AI conversation. The public contest around artificial intelligence is often dominated by increasingly sophisticated systems aimed at broad consumer use. The World Bank’s argument is that development needs may call for a more targeted model: tools that solve a limited problem well enough to be useful at scale.
Agriculture offers a concrete example
Donohoe pointed to the World Bank’s AgriConnect agricultural scheme in India. Farmers involved in the program need access to soil analysis in their own language. Providing highly tailored guidance to every small farm through a large human expert workforce would be difficult. The bank sees specialized AI as a possible way to extend the reach of that information.
The example illustrates both the promise and the limits of the proposal. The value is not in an AI system making a general claim about farming. It is in helping deliver particular information needs to particular users. For a farmer, a tool becomes relevant when it can be accessed, understood and connected to a practical decision.
Specialization could also make deployment more manageable. A narrower system can be built around a specific domain and dataset rather than attempting to answer every question. That does not remove the need for careful design, local knowledge or oversight, but it does focus investment on applications with a defined purpose.
Infrastructure remains the condition for impact
The World Bank’s pitch is not that small AI can bypass the wider requirements of development. At the Bangkok meetings, IMF managing director Kristalina Georgieva stressed that governments need to establish the conditions for AI adoption. She identified electricity, digital infrastructure, public readiness to adopt technology and government capacity to provide safety protections as essential foundations.
Those requirements are important context for the enthusiasm around phone-based models. A tool that can run with low connectivity may lower one barrier, but it still depends on power, devices, skills and an environment where people can use it with confidence. Local-language access also requires work on data, interfaces and deployment that reflects the communities being served.
In other words, the proposal is not a shortcut around public investment. It is an argument about where AI investment should be directed once countries are building the basics: toward systems that fit local constraints instead of assuming the constraints will disappear.
From capability to collaboration
Donohoe suggested that useful products could emerge through cooperation among the World Bank, governments and private-sector participants. That reflects the practical challenge ahead. Building a model is only one part of a deployment. It must be connected to a public or commercial service, maintained over time and made accessible to people who may not have extensive technical support.
The approach also implies that economic value will not be measured solely by the size of a model or the amount of computing behind it. For development-focused uses, the more meaningful test may be whether a specialized tool improves access to knowledge in a setting where expertise is scarce.
The World Bank’s underlying claim is ambitious: development advances that historically took far longer could potentially be accelerated. But the route it describes is grounded in an unglamorous premise. AI’s broadest benefit may come not from the most expansive systems, but from tools narrow enough to work on an ordinary phone, in a local language, for a real and recurring problem.
- The World Bank is advocating specialized “small AI” models for developing economies.
- The proposed systems are intended to work on mobile phones, in local languages and with limited connectivity.
- Electricity, digital infrastructure, adoption readiness and safety protections remain necessary for impact.
This article is based on reporting by The Guardian. Read the original article.
Originally published on theguardian.com
