Artificial Intelligence (AI) is increasingly discussed as though it were synonymous with large language models. Ask which country is leading in AI and the conversation quickly turns to foundation models, GPUs, model parameters and the companies capable of training them. Ask where Africa stands and the answer is often equally predictable: Africa is behind.
There is some truth in that assessment. The continent still faces major gaps in high-performance computing, advanced semiconductor capability, cloud infrastructure, research funding and access to specialised hardware. A 2024 review of African computer-vision research found extremely low representation in top-tier venues. (https://www.jair.org/index.php/jair/article/view/16653). These are real weaknesses and they should not be disguised.
But the conclusion also depends on where we choose to start counting. AI does not begin with a model. And intelligence does not begin with language. Human beings experience the world multimodally. We see, hear, move, remember, interpret and combine signals through context. Language matters, but it is only one part of intelligence.
That is why I think the current African AI debate is too heavily weighted toward natural-language processing. African-language technologies are important and should be supported. But many of the continent's most consequential problems are physical: agriculture, mining, transport, health, infrastructure, energy, wildlife, manufacturing and climate. Machines that operate in these environments must first be able to perceive them. They must interpret crops, roads, medical images, mineral surfaces, power lines, vehicles, animals, buildings and changing landscapes. This is the domain of computer vision.
Consider agriculture. A phone camera or drone can help identify crop disease, assess plant health or monitor livestock. Satellite imagery can support drought analysis, flood mapping, land-use monitoring and infrastructure planning. In healthcare, computer vision can assist with X-rays, pathology slides, microscopy, ultrasound and retinal imaging. In mining, it can contribute to geological mapping, equipment inspection, mineral sorting and worker safety. In conservation, camera traps and drones can reduce the burden of manually reviewing wildlife imagery. These are not peripheral uses of AI. They connect intelligence directly to the productive economy.
This is also why the phrase “the world is multimodal” matters. The future of AI is not likely to be purely textual. Useful systems will combine images, speech, text, sensor data, location, history and time. A farming system may combine satellite images, rainfall data, soil measurements and a farmer's description. A medical system may combine an X-ray, laboratory results and clinical history. A conservation platform may combine video, acoustic signals, GPS and environmental data. Computer vision also exposes a deeper truth about AI:- a machine does not see the world directly. It sees the world through the history of its training data.
A model trained mainly on environments elsewhere does not automatically understand African environments. Lighting differs. Roads differ. Architecture differs. Farming practices differ. Medical populations differ. Animals differ. Camera quality differs. Even the way people occupy public space may differ. In machine learning this is partly a problem of domain shift. A model can perform impressively in one setting and degrade sharply when deployed somewhere else. That is why local data matters; not because Africa should isolate itself technologically, but because systems used here must understand the realities they encounter here.
This should make data strategy central to African AI policy. Who collects the images? Who decides the labels? Who owns the dataset? Who defines ground truth? Who validates performance? Who benefits when the model becomes commercially valuable?
Artificial intelligence is often presented as software, but AI has a physical body. Data centres require servers, storage, networking equipment, cooling systems, electricity and backup power. Servers require semiconductors. Electronics require metals and minerals. Cameras require sensors, processors and optical components. The cloud ultimately sits somewhere on the ground. Once we understand AI in this way, Africa's place in the industry looks different. The continent is already important in parts of the upstream physical economy that supports digital technology.
The Democratic Republic of the Congo, for example, is a major producer of cobalt and tantalum and an important copper producer (https://www.usgs.gov/programs/mineral-resources-program/mineral-resources-data). These materials do not all go directly into AI accelerators, and mineral importance must not be confused with semiconductor leadership. But they are part of the wider material system on which modern computing depends. The danger is that Africa repeats an old pattern: raw materials leave the continent; processing, manufacturing, design and intellectual property happen elsewhere; and the finished products return at much higher value.
The same pattern can happen with data. African workers can annotate images. African hospitals can provide medical scans. African farms, cities and conservation areas can generate valuable visual datasets. Yet if the platforms, models and intellectual property are owned elsewhere, Africa may participate in AI without capturing much of its value. That is why I would not argue that Africa already leads AI. It does not. Nor would I claim that mineral wealth alone gives Africa technological power. It does not. A more accurate position is that Africa is unevenly positioned across the AI value chain: strategically important at some layers, seriously underdeveloped at others, and presented with an opportunity to move upward.
Computer vision could be one of those routes. African universities can build benchmark datasets around local environments. Hospitals can establish carefully governed medical-imaging repositories. Mining companies can partner with African researchers on industrial vision. Agricultural agencies can generate crop imagery. Conservation organisations can collaborate on ecological datasets. Governments can responsibly open non-sensitive geospatial and infrastructure data for research. Africa can also develop strength in efficient computer vision: models designed to operate on phones, cameras, drones and edge devices with limited bandwidth, power and compute. That would not be second-rate AI. It could become globally useful engineering knowledge because much of the world does not operate inside hyperscale data centres.
The goal should not be technological isolation. African systems should learn from global science and participate in international research. But African institutions should also help define the problems, control important datasets, understand model failure, own intellectual property and build products that can compete beyond local markets.
Africa's position in AI should therefore not be judged only by how many frontier models it trains. We should examine the full chain: the materials beneath the ground, the components built from them, the infrastructure that stores and processes information, the sensors that capture reality, the data that represents it, the people who curate that data, the models that learn from it and the systems deployed in the real world. Seen this way, Africa is neither simply ahead nor simply behind. The challenge is to connect these layers and keep more value at each stage.
The task is not to convince ourselves that Africa already leads artificial intelligence. It is to recognise that technological leadership is not determined at a single point in the value chain. Africa must move from extraction to processing, from data generation to data ownership, from annotation to research, from research to products and from consuming intelligent systems to building them.
The world is not text. The world is multimodal. If Africa wants to shape the next era of intelligence, one of its most important ambitions should be simple: teach machines to see the world from here.
Dr Sibusiso R Bakana is a Senior Computer Vision Researcher at CSIR-Defence and Security Cluster and President of Convocation at the University of Johannesburg. He writes in his personal capacity.