By Victor Chege, a Governance student at AISEA.
In 2023, an AI generator famously rendered the Kenyan coast as a fantasy of giraffes walking on water and baobab trees with Wi-Fi routers. The prompt had simply been “modern life in East Africa.” Don’t get me wrong, there’s probably a market for people who want to keep scrolling even when they’re out in nature but this misfire was more than just amusing; it was revealing.
Despite the training on billions of data points, AI systems still trip over the subtleties of a region where knowledge isn’t always written down and life can defy easy classification.
For East Africa, the data is often incomplete and the meaning distorted. For instance, only about 0.02% of total internet content is in African languages. But then again connectivity is uneven. In Sub Saharan Africa, internet penetration was 43% as at 2021 compared to over 90% in the US. What this means is that data is incomplete or skewed, it’s not because AI intentionally did this; it just doesn’t have the tools to see clearly.
The Grandmother Test
A huge chunk of knowledge in East African communities isn’t written down. It’s lived, observed, and remembered. Its encoded in relationships, naming traditions, seasonal rhythms, and shared responsibilities. These forms of intelligence don’t translate into datasets but they’ve guided land use, conflict resolution, and community care for generations. AI systems trained on written and western-centric data will miss these signals not because they’re irrelevant but because they don’t conform to the formats machines are taught to recognize.
A lot of communities operate through a dense network of kinship, clan, and tradition which doesn’t translate cleanly into code. These dynamics resist flattening into data points even as AI increasingly influences decisions around healthcare, banking, employment, and policing. In Somali clans, the abbān (clan representative) negotiates trade, enforces rules, and guarantees traveler safety . Rather than impersonal contracts or credit checks, these systems rely on relational trust. AI hiring tools or dispute resolution bots built without acknowledging this risk producing decisions that feel irrelevant or unjust.
The Grandmother Test is simple: if your system cant engage with knowledge that’s spoken, relational, or remembered, it’s not ready to act on behalf of the people who hold it.
The Informal Is Normal
In Nairobi, over 80% of employment is informal. What does that mean for AI powered fintech solutions that assess creditworthiness based on digital trails? For the roadside seller who transacts with mobile money, their economic activity may never generate the kind of structured data AI relies on. Their trustworthiness, however, is embedded in a dense web of social obligations, church groups, and neighborhood watch networks. AI also struggles with the improvisational genius of East African life. In Dar es Salaam, there’s a word, kujibanza, which loosely translates to inserting yourself where you don’t quite fit and somehow making it work. From markets that reorganize themselves after every rainfall to fruit vendors that double as mobile money counters, the continent thrives on repurposing, reusing, and reconfiguring. This kind of logic borne out of patchwork resources and creativity is nearly illegible to systems trained on datasets scraped from global north infrastructure.
Languages AI Can’t Hear
Kiswahili is spoken by over 200 million people, yet AI speech recognition systems often bungle it due to underrepresentation in training data. It struggles with tonal inflections, local slang, and code switching between English, Kiswahili, sheng, and mother tongue. In rural Ethiopia, for example, over 80 languages are spoken. In Uganda, Luganda has no formalized orthography but is accepted by all dialect speakers. Try automating that. One project, the Masakhane initiative, has been pushing back by mobilizing local communities to build natural language processing tools (NPTs) for African languages. But when OpenAI’s Whisper speech-to-text tool was tested in Kiswahili, its average word error rate was worse than for Icelandic which is a language spoken by fewer than 400,000 people.
The Myth of The Neutral Machine
East Africans are not passive recipients of technology. From the first SMS-based health campaigns to citizen-led map projects like Ushahidi, people have hacked and adapted digital systems for their own needs. But what happens when the machines push back? Facial recognition systems trained on lighter skin tones have failed spectacularly in parts of Africa. In 2019, a UN report warned of biometric systems being deployed in refugee camps with no clear oversight. They were basically automating vulnerability in the name of efficiency. The technology was imported but the consequences were deeply local.
This brings us to an uncomfortable truth: AI systems don’t just reflect the world, they reshape it. When they misunderstood East Africa, it isn’t just a software bug, it’s a new layer of structural exclusion wrapped up in the authority of code.
Data Colonialism Is Real
Some of the training datasets used in large AI models include scraped content from Kenyan newspapers, Nigerian forums, and South African government websites. Few, if any, of the communities that produced the data were consulted or compensated. It’s the old extractive model made popular by the conquest of the dark continent. Meanwhile global companies are outsourcing data labeling to East African workers. Click workers are paid a few cents to categorize, moderate, and train the same algorithms that may one day replace or surveil them. One Kenyan content moderator working for a major Silicon Valley firm described the work as “watching the internet’s nightmares on repeat.” This isn’t just a labor issue, it’s an existential one. Who gets to teach the machine and under what conditions?
What Can’t Be Captured.
Wisdom isn’t something you hold alone. It lives across people and is passed through stories, jokes, and the beat someone takes before naming a place. AI, no matter how advanced, cannot fully hold the complexity of the East African experience. How would it classify the Luo funeral traditions that turn into week-long festivals of remembrance? Or the Tanzanian street poets who remix Quranic verses with Bob Marley lyrics? Or the subtle etiquette of how you should go about answering an elder? AI is good at detecting patterns but here the meaning is often in the exceptions.
Not Just Problems, Also Proposals
This perspective is not a lament, it’s a lens. If East Africa is often invisible to AI systems, it also offers a reminder that intelligence isn’t just statistical. It’s also relational, contextual, and embodied. In Kampala, a collective of artists is training custom AI tools to generate folk inspired visuals in Luganda and Runyankole aesthetics. In Nairobi, a women-led tech hub is teaching AI ethics through storytelling workshops rooted in local folklore. These aren’t just side notes, they are the future where AI is made with, not for, African communities.
What The Algorithm Forgets
What AI cant grasp about East Africa isn’t just a technical problem. It’s also a philosophical one. In the rush to automate; the slow, communal and deeply contextual forms of knowledge that have helped communities survive through colonialism, capitalism, and climate crisis. AI isn’t the enemy. It doesn’t choose how to learn or act. We do. It simply reflects the data it’s fed and goals it’s given. So when it misses the mark; that’s on us. The real challenge is to find ways of recording our knowledge, logic, and ways of being without stripping away their meaning, depth, or spirit.
Bibliography
Adelani, D. I., Abbott, J., Neubig, G., D’souza, D., Kreutzer, J., Ruder, S., & Emezue, C. (2021). MasakhaNER: Named entity recognition for African languages. arXiv. https://arxiv.org/abs/2103.11811
Adelani, D. I., Neubig, G., Ruder, S., Rijhwani, S., Beukman, M., Palen‑Michel, C., Sibanda, B., & Emezue, C. (2022). MasakhaNER 2.0: Africa‑centric transfer learning for named entity recognition. arXiv. https://arxiv.org/abs/2210.12391
Ade‑Ibijola, A., & Okonkwo, P. (2023). Trust norms for generative AI data gathering in the African context. Data & Policy. https://www.cambridge.org/core/journals/data-and-policy/article/trust-norms-for-generative-ai-data-gathering-in-the-african-context/7B8CDB90BB2AA5CFE34A5E91D6C750CF
Gwagwa, A. (2024). Resisting colonialism – Why AI systems must embed the values of the historically oppressed. Chatham House. https://www.chathamhouse.org/2024/06/artificial-intelligence-and-challenge-global-governance/06-resisting-colonialism-why-ai
Lanfrica. (2023). How open‑source is powering the future of African NLP. Lanfrica Blog. https://lanfrica.com/blog/how-open-source-is-powering-the-future-of-african-nlp/
Muldoon, J., Graham, M., & Cant, C. (2024, July 6). AI feeds off the work of human beings. The Guardian. https://www.theguardian.com/technology/2024/jul/06/feeding-the-machine-human-labor-ai
Nekoto, W., Marivate, V., Matsila, T., Fasubaa, T., Fagbohungbe, T., Akinola, S. O., Abbott, J., & Bashir, A. (2020). Participatory research for low‑resourced machine translation: A case study in African languages. Findings of the Association for Computational Linguistics: EMNLP 2020, 2144–2160. https://aclanthology.org/2020.findings-emnlp.195/
Shefeni, S. (2023, March 17). The invisible labour of Africa in the digital revolution. African Arguments. https://africanarguments.org/2023/03/the-invisible-labour-of-africa-in-the-digital-revolution/
Ushahidi, Inc. (n.d.). Ushahidi [Computer software]. https://www.ushahidi.com
The Star. (2023, September 29). Status of employment in Kenya from 2016 to 2021. https://www.the-star.co.ke/news/2023-09-29-status-of-employment-in-kenya-from-2016-to-2021
Babbel Magazine. (n.d.). What language is spoken in Iceland? https://www.babbel.com/en/magazine/what-language-is-spoken-in-iceland
Internet Society. (2022). The Internet Society pledges to expand internet access in Africa [Press release]. https://www.internetsociety.org/news/press-releases/2022/the-internet-society-pledges-to-expand-internet-access-in-africa/