What Edzai Zvobwo's critique of AI asks of Africa
There is a moment about eight minutes into Edzai Zvobwo's conversation with Kyle Fraser on the Never Think Alone podcast (video below) where the whole argument turns on an old story. Zvobwo, a mathematician from Zimbabwe who now works on causal AI and advises across the tech and creative sectors, reaches for Plato's allegory of the cave. Prisoners who have only ever seen shadows cast on a wall build a coherent, workable account of trees and cars and people. They can describe those shadows beautifully. They can predict which one comes next. But they have never been outside.
That, he argues, is what our language models are doing. They have read how other people experienced reality and they are projecting an interpretation of it back at us. A chatbot that diagnoses illness has not sat with a patient. It has absorbed what doctors wrote down. The gap between those two things is not a technical detail waiting to be closed by more compute. It is a difference in kind.
The philosophy: correlation is not understanding
The intellectual spine of Zvobwo's position comes from Judea Pearl, the Turing Award winner whose book The Book of Why set out a ladder of causation. At the bottom rung sits association, noticing that A tends to appear with B. Above it sits intervention, knowing what happens when you actually change something. Above that sits counterfactual reasoning, understanding what would have happened had you acted differently.
Zvobwo's claim is that deep learning lives almost entirely on the bottom rung. It is correlation at scale. It knows A relates to B without knowing which one moves the other, or whether some third thing moves both. His point is that this is not a small gap for anyone trying to run a business. Every marketer who has fought with attribution modelling already understands the problem intuitively. Did that person buy the shirt because they saw the ad? The industry has spent decades and enormous budgets circling that question. A system built on correlation cannot settle it, no matter how fluently it discusses it.
From here comes the most useful idea in the episode, and the one I would take into any boardroom. Data, Zvobwo says, is inconsequential. Not worthless, but consequential in the literal sense: it is a consequence, a byproduct of activity. He uses his Apple Watch. You can stare at the step count for weeks and learn very little. Understand the process that generates it, the alternation of legs and the friction of the ground, and you can start manipulating outcomes. Increase the frequency and walking becomes running. You can even imagine your way past the edge of the data into territory it has never recorded.
This lands hard against a decade of "data is the new oil" talk. If data is a residue of process, then hoarding it without understanding the process that produced it is a category error. You end up with a lake full of exhaust fumes and no engine.
Technologies do not win because they are best
The turn from philosophy to political economy happens when Zvobwo asks why this particular version of AI won.
His answer is that it was the sellable one. Neurosymbolic approaches required mapping actual processes, which is difficult and does not scale neatly. Causal methods require proving causation, which is harder still. Correlation at scale could absorb billions in capital and produce visible results on a quarterly cadence, so that is what got funded. He puts it plainly: the school of thought that won, won because it could raise money.
I would push back a little on the strong version of this, because scale also won on merit. It produced capability jumps that the alternatives did not. But the softer version is hard to argue with, and it matters more. Once tens of billions are sunk into one architecture, the field's collective sense of what counts as promising research becomes partly an artefact of what is already funded. Directions get starved regardless of their merit. Zvobwo's plea is modest and reasonable: fund the hard, unglamorous, long-horizon work too, because the people doing it are not the ones grabbing headlines.
Who is this technology actually for?
Then comes the part of the conversation that South African and broader African readers should sit with longest.
The northern hemisphere has an ageing population and a shrinking workforce. Automation there is a solution to a real shortage. You need automated nurses because there are not enough nurses coming through. The technology is answering the question that its builders' societies are actually asking.
Now look at us. Cameroon's median age is somewhere around seventeen or eighteen. Africa's working age population is expanding faster than anywhere on earth. Our problem is not too few workers. It is too few jobs. And the current generation of AI does not solve unemployment. On the most plausible reading, it makes it worse.
So we have imported, at speed and largely uncritically, a technology optimised for a demographic situation that is the mirror image of our own. That is not a conspiracy. Nobody set out to disadvantage the Global South. It is simply what happens when the direction of technological development follows the problems of whoever is paying for it.
Zvobwo's second move is sharper still, and it is the one I keep returning to. He argues that we have been reduced to super users of technology built elsewhere. Build a food delivery app and celebrate it, and then ask where it lives. It sits on Apple's store or Google's, which means every innovation we produce pays rent upward to infrastructure we do not own. He notes that the celebrated African unicorns are frequently Delaware companies, and that the founders may sit in Joburg or Lagos while the entity, the tax base and the eventual exit sit elsewhere.
This reframes an old debate. Where the land question dominates our politics as the means of production, he suggests the actual means of production in this era are foundation models, app stores, cloud infrastructure and payment rails. We do not own any layer of that stack. There are no African foundation models of consequence, which is one reason a request for an image of a woman still tends to return a white one. The corpus was not built with us in it.
He adds a further constraint that has nothing to do with mindset. If South Africa and the United States each borrow a hundred dollars, we pay materially more for it. That premium compounds across every attempt to build capital intensive infrastructure. It is a structural disadvantage before a single line of code is written.
Where the argument gets thinner
Two cautions, offered in good faith.
The first is that Zvobwo attributes much of our position to framing and self limitation. He invokes Roger Bannister and the four minute mile: once one man ran it, everyone did, so the limit was collective and mental. It is a lovely story and a weak piece of evidence. Better tracks, better shoes and interval training did real work in bringing that barrier down. Compute access, energy infrastructure and the cost of capital are not collective mental blocks, and no amount of unlearning dissolves them. His own point about IMF borrowing spreads proves this against him.
The second is his 60 percent figure for the contraction of the advertising industry. He offers it explicitly as a back of the envelope guess, and the video description then promotes it to a prediction. The underlying logic is sound: when undifferentiated production collapses in price, genuinely distinctive work becomes scarce and expensive, like a handmade Rolls Royce or an organic mango in a Woolworths aisle. But scarcity only converts into premium pricing where buyers can verify provenance and care about it. That is true of a mango. It is much less obviously true of a banner ad.
What to take from it
The philosophy and the political economy are the same argument in two registers. In both cases the claim is that we mistake a shadow for the thing itself. In the technical case we mistake correlation for understanding, and data for the process that produced it. In the economic case we mistake being an enthusiastic user of a technology for having any stake in it.
The response in both cases is the same too. Go and find the generating process. Ask who built the thing, for which problem, and whose question it was actually answering. Then decide whether it is your question as well.
Source: AI Scientist: 60% of your industry is about to disappear!
