Whose Intelligence? Building Agentic AI without Frontier Model Dependency
The promise of agentic AI is enticing: autonomous systems that perceive, plan, and act, thereby reducing the friction between human intent and outcome. For Africa, however, that promise rests on models built almost entirely elsewhere. Agentic AI is arriving in Africa on someone else’s terms. Building agentic systems today almost inevitably means building on frontier models trained predominantly on Western-centric data, optimized for Western languages, and aligned to Western values, yet deployed to negotiate, decide, and act on behalf of African users. Every API call, every default behavior, quietly embeds a worldview that was not sufficiently built with Africa in mind. This paper therefore takes a positional stance: the current default is insufficient and, in some cases, harmful. Through a workshop hosted by the ML Collective (Africa)1 community at Deep Learning Indaba 2026,2 we pressure-tested four tensions (AI safety and policy, infrastructure and architecture dependency, cultural alignment, and value extraction and the capability trap), working toward a position on sovereign agentic AI for Africa. Each tension trades something away: a say in safety priorities, ownership of infrastructure, cultural fit, and the capability that accumulates with whoever builds the model. We recommend seven priorities, led by funding talent and by adopting continued pretraining of open-weight models as a stated strategy.
Authors
- Tobi Olatunji (ORCID: https://orcid.org/0009-0009-7214-2132)
- Ayomide Odumankinde
- Sarah Luger
- Chinasa Okolo
- ML Collective
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24224513
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint