AI in local and community context

A tool designed around constant broadband, foreign place names, or a single language can fail people who use it differently. Fit is not an optional extra; it is part of whether a system works.

Farmers, teachers, health workers, elders, and students hold knowledge that a generic model may miss. Ask them to test ideas and explain what a tool cannot see.

A language model's ability in any language is roughly proportional to how much text in that language it saw during training, and this single fact explains most of what African users experience with AI tools. Training data comes overwhelmingly from the internet, and the internet's text is wildly unbalanced: English accounts for a huge share of web content, while Shona and Ndebele — spoken by millions — appear in comparatively tiny amounts, much of it Bible translations, news fragments, and social media rather than the broad range of textbooks, encyclopaedias, and technical writing available in English. Researchers call these low-resource languages, and the term describes the data, not the languages themselves, which are as expressive and complete as any on earth. The mechanism has layered consequences. A model with thin Shona data learns shallower patterns: it may manage greetings and simple sentences but produce stilted phrasing, miss idioms — a literal rendering of chakafukidza dzimba matenga loses exactly what matters — and, worse, hallucinate more, because sparse patterns leave more gaps to fill with invention. Even tokenisation, the way models split text into pieces, is fitted to data-rich languages, so the same sentence can cost more computation in Shona than in English. Work through a practical example: a Mutare farmer asks a chatbot in Shona about fall armyworm treatment. The model must operate in its weakest language on specialised agricultural content — a double thinning of its training signal — so the answer is likelier to be vague or subtly wrong than the English equivalent, and the farmer least equipped to cross-check in English bears the highest risk. The situation is improving through deliberate effort rather than accident: projects such as Masakhane, a grassroots network of African researchers, are building datasets and translation models for African languages, and major labs now report on hundreds of languages. But the misconception to correct is that this gap closes automatically as models grow. Scale amplifies whatever distribution the data has; without intentional collection of African language text — which requires funding, local expertise, and fair treatment of the communities providing it — bigger models mostly get better at data-rich languages. For users the practical rules follow directly: expect stronger performance in English, verify local-language answers on anything important, and prefer tools that state which languages they actually support rather than merely accept.

Zimbabwe and its neighbours have already run one of the world's most instructive experiments in technology adoption, and its lessons apply directly to AI. When mobile money services like EcoCash spread across the region, they did not follow the Western sequence of bank branches, then cards, then online banking. They leapfrogged it: millions of people whose villages never had a bank branch went straight to sending, receiving, and storing money on basic phones. The mechanism of this success is worth stating precisely, because none of it was accidental. The service was built on infrastructure people already had — ordinary phones and airtime networks, not smartphones and broadband. It worked through a human network of agents in kombis ranks, growth points, and township shops, who converted cash to digital value and back, and who explained the system face to face. It solved a locally urgent problem — moving money safely between town and rural areas — rather than importing a foreign use case. And it priced in small units matched to how people actually earn and spend. The result transformed daily economic life in ways Silicon Valley did not predict and did not build. Now apply the same mechanism as a checklist for AI in a local context. Does the tool run on the devices people have, tolerate expensive and intermittent data, and ideally work offline for its core function? Is there a human layer — an extension officer, a teacher, a clinic worker — who can mediate between the technology and people encountering it for the first time? Does it target a problem the community actually ranks as urgent, and does its cost structure match local incomes? A worked contrast makes the point: an AI crop-advice app that requires video streaming over 4G to a recent smartphone fails every test, while an SMS or WhatsApp-based service answering questions in local language through low-bandwidth text, with extension officers verifying advice, passes most of them. The misconception to correct is the fatalistic one — that advanced technology is invented elsewhere and Africans simply receive whatever version arrives. Mobile money shows the opposite: the region set the global standard, and Kenya, Zimbabwe, and their neighbours taught the world how digital finance works in cash economies. Leapfrogging has limits — it still requires electricity, connectivity investment, and supportive regulation — but the deeper lesson stands: local constraints are design inputs, not disqualifications, and the people who understand those constraints are the ones best placed to shape the tools.

A school wants a chatbot to help with study support, but some learners have limited evening data. A responsible response is: Offer downloadable materials and non-AI alternatives alongside the chatbot. Correct. This reduces exclusion caused by unequal connectivity. Use this case to separate what the technology contributes from what people contribute. The team should compare the intended outcome with a baseline where applicable, record important assumptions, and keep a clear route to correct or stop the process.

Ask whether the examples, data, and assumptions fit your school, company, or community. Generic demos often miss local names, laws, connectivity, and languages.

Farms, towns, and languages differ. Generic answers may miss your weather, crops, or community — add local detail and check with local experts.