The AI era ban is over — SA schools must embrace the leapfrog generation

Education in South Africa is at an inflection point. Pupils are using AI the way previous generations used calculators and the internet, except the results are faster and almost impossible to trace. Evidence is emerging that this has significant development risks. On the other hand, teachers are using AI for everything from drafting lesson plans to summarising long reading lists, and designing whole courses. Yet official policy is often out of step with reality: a parent at a top private school recently told me her daughter secretly uses ChatGPT for homework to bypass a blanket ban imposed by the school. Beneath all of this is a shared, unspoken understanding: something significant is happening, and we don’t yet fully know what to do about it.In a recent survey, 60% of US teachers admitted to using AI tools, but alarmingly, 68% had received no training in this whatsoever, and only one in five were teaching at schools with any AI policy in place. On the student side, a study at the University of Reading found that 94% of AI-generated submissions went undetected. Even the elite institutions are being caught off guard. Earlier this year, Brown University Economics Professor Roberto Serrano allowed his intermediate microeconomics class to write a midterm exam at home as a small act of compassion after a traumatic incident on campus. The class averaged 96%, against a historical average closer to 70%. Suspicious, he moved the final exam back into a supervised room. Roughly a third of the students didn’t show up, and those who did averaged 48%. This was the Ivy League, with every advantage money can buy, and the professor still couldn’t see the gap until he physically removed the technology from the equation.Serrano’s instinct of retreating to the supervised, device-free room seems to be the instinct of much of the education sector. Faced with something unfamiliar, institutions hunt for it, ban it and revert to 1985’s teaching practices. The problem with banningTo ban something, you have to be able to reliably detect it, and detection doesn’t work. In October 2025, my own institution, UCT, one of Africa’s most respected research universities, disabled its AI detection tool, citing reliability concerns. Even when a detector does appear to work, it holds up for only a few months before a new model or prompt emerges to fool it.The second problem is worse. Across seven widely used AI detection tools, Stanford researchers found that 61% of essays written by non-native English speakers were falsely flagged as AI-generated. For native speakers, the false-positive rate dropped to below 3%.In South Africa’s multilingual context, English is the medium of instruction for millions of students for whom it is a second, third or even fourth language. When we deploy AI detectors in those classrooms, instead of catching cheaters, we may be penalising students for the way they write in a language that is not their own, adding another layer of disadvantage on to learners who are already working harder to express their ideas in an unfamiliar tongue.Banning and detecting, in a context like ours, is not a neutral policy. Letting AI run wild isn’t the answer either Handing AI the wheel, using it to produce the thinking rather than sharpen it, is just as harmful.In perhaps the most ironic example, the Department of Communications’ AI policy released in April was found to contain fictitious references which had seemingly been hallucinated by AI. Nobody had been trained to catch this.At the student level, this produces essays that sound authoritative, flow well, and contain nothing the student actually thought. MIT research found that people couldn’t even recall the content of essays AI had written for them. The output is hollow because you cannot defend what you didn’t think. This is but one of a dozen real risks that come with unregulated use.These two extremes are both dead ends. But the education sector has been here before.The fears that drove abstinence-only sex education in schools were not unfounded: STIs, unintended pregnancies and sexual violence are real, and the people who worried about them were not wrong in their concerns. Their error was in believing that silence and prohibition were the most effective tools for managing the issue.The fears driving AI bans deserve the same honest acknowledgement.But here is what we know about risk and darkness: they compound each other. A student using AI secretly, without guidance, in the margins of a ban, is more exposed to every one of these risks, not less. Education does not eliminate risk, but it is the only intervention that can reduce it. Bans teach none of this. They simply move the behaviour somewhere nobody can see.What most institutional policies get wrongOne of the deepest failures of university policies is that they answer the question “How do I stop them?” and rarely the question “How do I make them good at this?”Few of these policies teach students how to use AI well, i.e. how to frame a problem, interrogate an output, or know when their own judgment should override the machine. And this reflects the people writing them too. They are not leveraging this technology effectively in their own work, they have had no training, and they are improvising policy for a tool they don’t yet use well themselves. The policy vacuum at the top is really a competence vacuum, and students can feel it.We have the focus all wrong. A good AI policy should read less like an academic-honesty clause and more like a driving manual.The developing world caseSouth Africa cannot afford to stay in the dark.For generations, the inequality in education has been structural and attempts to address it through conventional means have run up against the same wall.AI does not care which school you went to. A student in a rural school with a basic smartphone now has access to the same intellectual scaffolding. They can pressure-test an argument, get feedback on a draft and ask questions at any hour. In the long sweep of educational history, that is almost unprecedented.Provided they know how to use it.Luckily, this is no longer theoretical. In Edo State, Nigeria, researchers ran a six-week trial with public secondary-school pupils. The group met twice a week to work with ChatGPT. Each session opened with a teacher introducing a topic from the official English curriculum, then guiding the pupils into the AI using carefully designed prompts. The point was to make the model behave like a tutor that asked follow-up questions and pushed pupils to reason a concept through for themselves. The teacher stayed in the room throughout: keeping pupils on task, checking that they understood the feedback, and teaching them to spot the AI’s hallucinations. The measured learning gains were equivalent to roughly one-and-a-half to two years of ordinary schooling, placing the programme among the most cost-effective education interventions ever measured. What made it work was a teacher using AI to extend her reach and teaching critical use in the same breath.AI in education is both an equity question and a national competitiveness one. A young person in Gugulethu with a smartphone and a working knowledge of these tools can build software, write a business plan, design a product, and compete in the global digital economy. A decade ago, that path required a degree, a network, and capital. The barrier to entry for knowledge work has not merely lowered; for many fields, it has effectively collapsed. Countries that bring AI into the light and graduate AI-literate students will hold a material advantage over those that don’t. Developing countries have a genuine chance here not just to close the gap, but to leapfrog.When we ban AI and drive it underground, we do not protect pupils equally. The children of well-off families will learn to use it anyway and arrive at university fluent. The pupils who lose out are precisely the ones we want to be reaching. The sad truth is there is a version of the future where AI becomes one more axis on which privileged pupils compound their advantage, and underprivileged ones fall further behind. There is another version that holds the promise of restoring merit as a determinant socioeconomic mobility. The difference between these futures is not the technology. It is what we choose to do with it in the classroom. DM