Nick Miller, founder of Teach Me 2, the South African platform connecting more than 10,000 families with vetted tutors, argues that AI is removing the wrong bottleneck in education. Information was never the scarce resource, he says. Getting a learner to do the hard, unglamorous work of practising and persevering still is.

New research from a randomised trial of nearly a thousand high school maths students backs him up, and more sharply than the argument alone suggests.

Miller's case comes from operating experience and from home. He recently used AI to build a clock game for his son in half a day, because he could not find an existing resource that taught time the way he wanted. "But it still didn't magically make him learn to tell the time," he says. "When the work became difficult, it was still difficult. I still had to sit with my son, decide when to help and when not to help, judge how much effort he had already expended."

None of that context, he points out, is available to the AI. It does not know the child had a bad morning or that his confidence is low. The trial data says the same thing in numbers.

Students given a standard ChatGPT-style interface scored 48% better on practice problems, and those given a version with teacher-designed guardrails scored 127% better. Then the researchers took the AI away. The unrestricted group scored 17% worse than students who never had access at all, having used the tool as a crutch rather than working through problems. The guardrailed version, which gave hints instead of answers, removed the damage entirely.

The study by Hamsa Bastani, Osbert Bastani, Alp Sungu and colleagues is published in the Proceedings of the National Academy of Sciences. Stanford’s Tutor CoPilot trial by Rose Wang, Ana Ribeiro, Carly Robinson, Susanna Loeb and Dora Demszky points the same way from the other direction: giving around 900 human tutors an AI suggestion panel made their students 4 percentage points more likely to master topics, rising to 9 points for the lowest-rated tutors, at about $20 a tutor a year. Those tutors asked more guiding questions and gave away fewer answers. It is the clearest evidence yet for Teach Me 2's position, which Miller puts plainly: "The most effective model, in my view, is not an AI replacing a tutor. It's an excellent human educator with extraordinarily powerful AI tools."

The same model produced a 127% improvement and a 17% deficit depending only on how it was configured, which is the finding worth carrying well beyond education. Any business deploying AI to people who still need to build skill faces the identical fork: a tool that supplies answers lifts output now and erodes capability underneath, while one built to prompt thinking lifts both.

Miller's question for parents generalises to managers without much adjustment. He suggests starting with what is actually stopping progress, whether that is a missing explanation, a knowledge gap, motivation or accountability, because if someone already knows what to do and is not doing it, another source of information will not fix it. It also sharpens the AI safety debate filling feeds this week.

The useful question is not whether the technology is powerful, but whether the specific implementation in front of your team or your child is building capability or quietly removing it. That is a design decision somebody made, not a property of the model.

You might also like our piece on TeachMe2 making TIME's edtech list, the Injini AI for Education programme backing local edtech founders, and ChatGPT for Teens and who sets the rules for AI used here.

Get more SA tech and business news and subscribe to The Open Letter.