Ashraf Johaardien: Intelligence does not arrive on its own. It carries the choices, assumptions and absences of the people who make it, and they answer for it.

I like Benjamin Rosman before I have met him. He lands as a guy. He is a dad, like me, and he apologises for the delay in getting back to me when he is a day ahead of my deadline.

There is a verb that gets used when a system does harm. Emerged. The behaviour emerged, as though it arrived on its own, out of the weather.

Rosman grants the first half. Complex behaviour does come out of large systems in ways that are hard to predict. It is the second half he refuses.

“People choose the high-level objectives, the data, the architectures, the incentives, the deployment context and the safeguards.”

“Those choices shape what kinds of intelligence become possible, and whose interests are prioritised.”

He is a Professor of Computer Science at Wits, where he runs the Robotics, Autonomous Intelligence and Learning Laboratory and directs MIND, the Machine Intelligence and Neural Discovery Institute. His keynote, in the Imagination Engines, is at 10:20 on Thursday 15 October in the Chris Seabrooke Music Hall, twenty minutes after Prof. Ghebreab.

“This helps keep responsibility with institutions and people rather than allowing us to attribute outcomes to some mysterious property of the technology.”

“We still have to ask who built it, who deployed it, and who was empowered to question it.”

Ask the same of the systems everyone now uses and the answer gets longer and more uneven. Vast amounts of human language, images, software and labelled data, held up by compute, capital and human labour. The power to decide what is collected and what gets optimised sits with a small number of countries and companies.

He calls this provenance, and he will not let it stay a tracing exercise.

“Provenance therefore is not only about tracing a source. It is about understanding whose knowledge became infrastructure, whose assumptions became defaults, whose labour became hidden, and whose worlds are missing altogether.”

“Those absences can materially change how a system behaves.”

A language that was never collected is a conversation the system cannot have.

The part usually offered to Africa in this story is a supporting one. Data source. User base. Deployment setting.

“The problem is not only that such a role is unfair. It is that it is scientifically and strategically limiting.”

Different environments expose different assumptions, he says, and a researcher who only works at the application layer loses the chance to shape the foundations of the technology itself.

“The point is not that Africa should build a separate AI ecosystem. It is that globally important ideas should be able to originate here, be tested here, and influence the direction of the field rather than merely adapting to decisions made elsewhere.”

Two institutions make the case for him. He was involved in founding Lelapa AI, which builds for African languages, and he now leads MIND, which works on the underlying science of how intelligent agents learn, reason and adapt.

“If you only adapt an existing model at the end, you inherit most of its choices about representation, language, objectives and what counts as useful behaviour.”

African languages and contexts, he says, can require rethinking much earlier in the pipeline. Sometimes adapting a powerful existing system is the right answer. Sometimes a problem cannot be solved properly without changing the foundations.

At home, it becomes a question for an institution deploying AI in South Africa.

“An African institution would need to treat provenance as part of governance, not as an optional technical appendix.”

The record he describes is practical. Which models are in use. What data they were trained or adapted on. Which licences and outside services they depend on. What local testing found, where the known failures are and who is most likely to be hurt by them.

“Procurement matters enormously here: institutions should demand enough transparency from vendors to evaluate systems rather than simply accepting performance claims.”

And someone has to own it, to keep the record, review changes, respond to failures and decide when a system should not be deployed at all.

Then there is the money. The question, he says, is whether it goes to buying tools or to the long work.

“That means funding people, research programmes, shared compute, high-quality African datasets and institutions that can sustain fundamental work over many years.”

“We should certainly deploy useful technologies now, but deployment alone does not create agency.”

“Building durable research capacity is slower than buying tools, but it gives us something more valuable: the ability to decide what should be built, to understand it deeply, and to negotiate from a position of genuine technical strength.”

“The next two years are key because capabilities are consolidating quickly.”

After the two years comes the ten.

“The strategic question is whether, ten years from now, African institutions will be able to shape the models, standards and scientific ideas that underpin these systems, or whether we will still be choosing between products designed elsewhere.”

Intelligence is made. The open matter is who does the making.

Ashraf Johaardien is the Writer in Residence at the Fak’ugesi African Digital & Innovation Festival 2026.

His views are his own and do not represent those of the festival or its partners.

Ashraf Johaardien
ashrafjohaardien@me.com
M\e. Creative
http://www.fakugesi.co..za

Disclaimer: Artslink.co.za encourages freedom of speech and the expression of diverse views. The views published do not necessarily represent the views of Artslink.co.za.