If you have been to any business conference in this country over the last few years, you will have heard the pitch. Artificial intelligence is going to transform everything. It will add trillions to Africa's economy. It will let us leapfrog into the future the way cellphones let us skip landlines. Somewhere between the keynote and the canapés, someone will mention the Fourth Industrial Revolution, and everyone will nod.
And if you have spent any time on social media, you will have heard the other pitch. The robots are coming for your job. AI is going to end civilisation. Be afraid.
Both stories are wrong, and both are getting in the way of the conversation we actually need to have.
Start with what is real. AI is a genuinely useful technology, and it is already working here. It drafts documents, summarises reports and translates between languages, although it still stumbles over isiZulu, isiXhosa and Sesotho because it was mostly trained on English. It helps programmers write code. Local pilots are using it to help screen for TB and diabetic eye disease in clinics where specialists are scarce. Farmers are using it to stretch water and fertiliser further. None of that is hype. It is happening, and it matters.
But the same technology lies with total confidence. Ask an AI system a question and it will give you a fluent, polished answer that may be completely made up. It learnt from data that barely includes African languages, faces and realities, so when it is used to score credit applications or screen CVs, it can quietly repeat the very patterns of exclusion this country has spent thirty years trying to undo. An algorithm trained on our history will happily automate our history.
Then there is the small matter of what AI actually needs to run: enormous amounts of electricity, computing power, cheap data and skilled people. We are short of all four. The leapfrog story is seductive, but you cannot leapfrog on an overloaded grid with some of the most expensive data on the continent and your best graduates being recruited to London and Amsterdam. And in a country where more than three in ten people cannot find work, telling call centre agents and admin clerks not to worry about automation is not reassurance. It is an insult. If we want the benefits, reskilling has to be part of the deal, not a footnote.
Meanwhile, the abuse is not waiting for anyone's policy. The scams are already in your WhatsApp groups. Deepfake videos of famous South Africans punting miracle investment schemes. Cloned voices phoning grannies about fake family emergencies. Doctored audio drifting into our politics just in time for the next election. And a quieter problem that gets far less airtime: made-up facts and invented citations turning up in court papers, in research, in official documents.
Which brings us to the most South African AI story of them all.
In April this year, after six years of summits, commissions and frameworks, government finally gazetted the Draft National AI Policy for public comment. Credit where it is due: it was a sensible document in many ways. Rather than inventing a giant new AI regulator, it proposed letting existing bodies like the Information Regulator, ICASA, the Competition Commission and the Reserve Bank oversee AI in their own sectors, coordinated through a national forum. It was grounded in the Constitution and in ubuntu. It borrowed from Europe's AI Act without copying it blindly.
Two months later the Minister withdrew it. Someone had checked the reference list and found citations that did not exist. They had been generated by AI, and nobody had verified them. The policy written to govern artificial intelligence was sunk by unchecked artificial intelligence. You could not script it better. A revised version is expected early next year, presumably with every footnote checked twice.
Laugh at the irony, but do not miss the lesson. If hallucinated AI output can slip into a national policy document, it can slip into your board pack, your child's assignment, your morning news feed. The single most important habit of the AI age is embarrassingly old-fashioned: check before you trust.
While we wait for the revised policy, it is worth clearing up a misconception. South Africa has no AI law, but that does not mean AI here is a lawless frontier. POPIA already gives you rights when a machine makes a decision about you on its own. The Cybercrimes Act covers the fraud and forgery behind most deepfake scams. The Consumer Protection Act covers companies making inflated claims about their "AI-powered" products. Directors are on the hook under King governance principles. The tools exist. What is thin on the ground is enforcement capacity, and that is a funding and skills problem, not a legislative one.
So what should we actually be demanding? Nothing exotic. Tell people when they are looking at AI-generated content, and watermark it. Tell people when a machine, not a human, is deciding whether they get a loan, a job or a payout, and give them a real way to appeal. Fund the regulators properly so the rules on paper mean something in practice. Deal with data costs and connectivity so the benefits of AI do not pool in Sandton while the risks spread everywhere. And make verification of AI output a hard requirement wherever the stakes are high: in courtrooms, in research, and yes, in government policy documents.
There is also something ordinary citizens can do. When the revised policy comes out for public comment, comment. Businesses, unions, universities, individuals. The POPIA years proved that public input genuinely shapes how these laws end up working. This is one of those rare windows where the rules are still wet cement.
In the meantime, treat AI the way you would treat a charming stranger who talks a very good game. Do not fear it. Do not worship it. Use it, benefit from it, but never trust it more than it has earned. The hype is loud and the fear is louder, but the real story sits in between: a powerful tool, arriving in a country with real constraints, at a moment when the guardrails are still being poured.
Just ask the people who wrote the policy that had to be withdrawn.
