The World Bank is urging developing countries to kick-start the creation of “small AI” models, compact systems trained for specific purposes that can run on a mobile phone with weak connectivity. Paschal Donohoe, the Bank’s chief knowledge officer, made the case on the sidelines of the World Bank and IMF annual meetings in Bangkok, in comments reported by the Guardian on October 18.

What “small AI” means

Donohoe defined small AI as specialized models trained on narrower datasets for a particular job. They would be available locally, work in low-connectivity settings and operate “in the local language of the community,” typically on a mobile phone. This is a policy framing rather than a technical standard, but it overlaps with a real technical trend: small language models and on-device models that fit in a phone’s memory, which we explain in our guide to model compression.

He contrasted this with Silicon Valley’s focus on “ever more sophisticated consumer products,” arguing AI can also accelerate economic development in emerging economies, but only if adapted to local needs. “We argue that economic development in the countries that we serve and partner with that could have taken centuries to make happen, the potential is there for that economic development now to happen within decades,” he said. That is an aspiration, and no evidence for the timeline was reported.

The farmer example

Donohoe’s illustration is agriculture. In the World Bank’s AgriConnect scheme, Indian farmers need soil analysis in their own language. “There’s not going to be enough scientists and agricultural experts to do that for every small farm in India,” he said. A phone-based small AI could deliver that expertise at the scale needed. He suggested such products could emerge from collaboration between the World Bank, governments and the private sector. Our article on AI for smallholder farmers looks at what has already been tried and what the evidence shows.

Preconditions: power, networks, skills

The same meetings carried a more cautious message. IMF managing director Kristalina Georgieva, speaking on a Monday panel, called on governments to provide the preconditions for AI to succeed: access to electricity, digital infrastructure and people prepared to adopt the technology. Those basics are not universal. A model that runs on a phone still needs a phone, a charge and, for updates, occasional connectivity. AI’s effects on jobs and growth were a central theme of the annual meetings, according to the Guardian.

Why it matters

Most advanced AI is built for English and a few other major languages, and runs in data centers that require reliable power and fast internet. Many people in low-income countries are left out, a gap we examine in the AI language divide. Small, local models are one way to narrow it: they are cheaper to run, can protect privacy by keeping data on the device, and can be tuned to local crops, health practices or legal systems. The risks are equally concrete. Models trained on narrow data can be wrong in ways users cannot detect, building high-quality datasets in minority languages is hard, and a mistake in soil or health advice has real consequences. Responsible deployment needs evaluation, human oversight and pathways to correct errors.

There is also a strategic dimension. Countries that rely on imported foundation models depend on foreign companies for access and pricing. Locally owned small models can be part of a broader digital sovereignty approach, though building them requires data, talent and funding that the World Bank suggests it can help to mobilize.

What to watch next

  • Concrete World Bank-funded pilots of small AI beyond the AgriConnect example.
  • Public datasets and benchmarks for low-resource languages.
  • Outcomes data from phone-based agricultural advice tools.
  • Commitments from the private sector and governments announced around the Bangkok meetings.
  • Evaluation standards for AI advice that farmers or patients act on.

Key takeaways

  • The World Bank wants developing countries to build narrow, phone-sized “small AI.”
  • Donohoe says development could take decades instead of centuries; no evidence for that was given.
  • The first example is soil analysis for Indian farmers in their own language.
  • The IMF stresses electricity, digital infrastructure and skills as preconditions.
  • Quality control and local data are the main risks.

Frequently asked questions

Is small AI the same as a small language model?

They overlap. Small AI in Donohoe’s description is any specialized, narrowly trained model that runs on low-power devices; a small language model is one technical form of it.

Why not just use big chatbots?

Large models need good connectivity, often work poorly in local languages and can be costly. Small models can run offline on a phone.

Who would build these models?

The Bank suggests collaboration among itself, governments and private companies.