SourceYouTube

Published March 4, 2026

TED

Hosts:Nanjira Sambuli

This is an AI-generated summary. The source video includes demos, visuals and context not covered here. Watch the video → · How our articles are made →

In Brief

Nanjira Sambuli, a policy researcher and Fellow at Carnegie Endowment for International Peace, argues that Africa is not just caught in the crossfire of the global AI race but offers a fundamentally different approach to building technology. Speaking at TED2025, Sambuli introduces "ubuntech," a framework rooted in the Bantu philosophy of ubuntu ("I am because you are") that treats data as representing lives and communities rather than raw material to extract. She backs this up with concrete examples: Lelapa AI's small language model that outperforms larger competitors, and Masakhane, a grassroots research community spanning over 30 African countries.

For related coverage, see From Estonia to India: Schools Bet Big on AI, Why Banning AI in Classrooms Misses the Point, and Bittensor's Bet: Cheaper AI Through Decentralization.

The central claim

Sambuli opens with an African proverb: "When elephants fight, it's the grass that suffers" (0:31). The elephants represent tech giants and nation states competing for AI dominance. The grass represents the people, geographies, and ecologies treated as resources to exploit. Africa, she argues, is "caught in the middle" as a "key battleground" over natural and human resources powering the intelligent age (1:11).

But Sambuli's talk is not about victimhood. Her central argument is that African communities are already building AI differently, guided by ubuntu, a value system from the Bantu people meaning "I am because you are" (1:40). Ubuntu is not just a philosophical concept. It is an active design principle that reshapes how data is governed, how AI products are built, and how communities organize around technology.

Data as lives, not oil

The dominant AI industry frames data as an abundant natural resource. The phrase "data is the new oil" captures this mindset. But Sambuli points out that this approach is already hitting its limits as high-quality training data dries up (2:40).

Through ubuntu, data is conceptualized differently. It represents lives, cultures, and communities. Data governance becomes about "meaningful participation, informed consent, self-determination and community ownership" (3:02). This thinking has inspired the concept of data justice in African policy frameworks, which means that communities who contribute data are represented and visible in the systems built from it.

Small models, big results

When told their languages are "low-resource," African practitioners responded by resourcing them. Sambuli highlights how conventional AI demands large language models (LLMs), massive AI systems trained on billions of text examples. But African developers are "making do with little language models" (3:59).

Lelapa AI developed InkubaLM, a small language model (SLM) named after the dung beetle, which can roll up to 250 times its body weight (4:26). The model is trained on just 0.4 billion parameters and reportedly outperforms larger models in sentiment analysis, the ability to determine whether text expresses positive, negative, or neutral opinions. It also displays "remarkable consistency across multiple languages" (4:36).

Building together across borders

Masakhane, which translates to "building together," operates across over 30 African countries (4:52) to strengthen natural language processing (NLP), the field of AI that teaches computers to understand and generate human language. The community's grassroots approach aims to show that low-resourceness "is not a data problem, but a societal one best solved through participation" (5:02).

Masakhane has also developed a non-traditional authorship model that includes all contributors in published papers, whether they contributed data, lived experience, code, or coordinated research. Through this approach, they have published translation results for over 38 African languages (5:27).

Sambuli calls this collective effort "ubuntech," which she defines as "artificial intelligence powered by ancestral intelligence" (6:06).

How to interpret these claims

Sambuli's talk presents an inspiring vision, but several questions deserve consideration before accepting it at face value.

Scalability beyond community projects

The examples Sambuli highlights are genuinely impressive at a community scale. InkubaLM outperforming larger models in sentiment analysis is a meaningful technical achievement. But sentiment analysis is one narrow task. The talk does not address whether small, community-built models can compete with trillion-parameter frontier models on broader capabilities like reasoning, code generation, or complex dialogue. The "small is mighty" framing works for specific tasks but may not generalize.

Economic sustainability

The talk is aspirational but light on business models. Community-driven, participatory approaches to AI development are admirable. How they sustain themselves financially remains unclear. Open-source NLP communities rely heavily on volunteer labor and grant funding. Whether ubuntech can scale into economically self-sustaining AI development is an open question.

The "data is the new oil" critique

Sambuli frames the ubuntu approach to data as a departure from the exploitative "data is the new oil" mindset. This critique is well-established in academic and policy circles. Data sovereignty, informed consent, and community ownership are already central themes in European data regulation (GDPR), Indigenous data sovereignty movements, and digital rights advocacy worldwide. The ubuntu framing adds cultural depth, but the underlying policy proposals overlap significantly with existing frameworks.

What stronger evidence would look like

Independent benchmarks comparing InkubaLM against established small language models on a range of tasks (not just sentiment analysis) would strengthen the technical claims. Longitudinal data showing how Masakhane's participatory model affects research output quality and community outcomes would validate the governance model. And case studies showing ubuntech principles adopted beyond Africa would show the framework's universal applicability.

Practical implications

For AI developers

Sambuli's talk offers a concrete challenge: consider whose data you are training on and whether those communities benefit from the result. The ubuntu principle that data represents lives, not just training material, is a useful lens for any team building AI products, regardless of geography.

For policymakers

The concept of data justice, where communities who contribute data are represented in the systems built from it, aligns with growing global momentum toward data sovereignty. Policymakers working on AI governance could look to African frameworks like Masakhane's participatory model as a complement to top-down regulatory approaches.

For organizations building multilingual AI

The InkubaLM example shows that smaller, purpose-built models can outperform larger ones on specific language tasks. For organizations serving non-English-speaking populations, investing in small, targeted language models may be more effective than fine-tuning massive general-purpose ones.

Glossary

Sources and resources