I spend most of my working life building AI tools for education. Recently I ran a simple experiment: I asked ChatGPT to create a Grade 9 case study about entrepreneurship in South Africa.
What I got was a tech startup in a co-working space, raising venture capital from angel investors, scaling through digital marketing. The entrepreneur had an MBA and a MacBook.
I tried again. “Make it about a South African small business.” This time: a coffee shop in Cape Town’s waterfront, complete with an Instagram marketing strategy and a point-of-sale system.
I shared this with a group of EMS teachers. They laughed. They had all been there. One teacher told me she had given up on AI tools entirely because she spent more time fixing the outputs than it would have taken to create the resource herself. Another said he just uses it for admin now, never for actual teaching content.
These are experienced teachers. They know exactly what a good entrepreneurship case study looks like for their learners. They know that spaza shops, hawkers at taxi ranks, and township salons are where their learners actually see business happening. They know that financial literacy means stokvels and SASSA grants and airtime as currency, not 401(k) accounts and credit scores.
The AI does not know any of this. And that is the real problem.
The Expertise Gap
The conversation about AI in education tends to focus on whether teachers have the right skills to use these tools. Are they trained? Do they know how to write good prompts? Can they tell when the AI gets something wrong?
These are valid questions. But they miss something fundamental: the biggest gap is not in teachers’ ability to use AI. It is in AI’s ability to recognise what teachers know.
A teacher who has spent years developing examples that resonate with their learners, who understands the economic realities of their community, who knows which analogies land and which fall flat, that teacher carries an enormous amount of professional expertise. None of it exists inside ChatGPT.
Every time a teacher opens an AI tool and types a prompt, they start from zero. The tool does not know their grade, their curriculum, their learners, or their context. So the teacher has to re-explain everything, every time. And even then, the outputs default to American assumptions because that is what the training data reflects.
This is not a prompting problem. It is a design problem.
What Gets Lost
When AI tools ignore teachers’ expertise, the consequences go beyond inconvenience.
Content drifts toward generic. A case study about “a small business” could be from anywhere. The specific details that make a lesson connect, the local examples, the familiar contexts, the currency amounts that make sense, those get smoothed away in favour of something that sounds professional but feels foreign.
Teachers adapt rather than create. Instead of building on their own knowledge, teachers find themselves editing and correcting AI outputs, stripping out American references and substituting local ones. The AI sets the starting point, and the teacher works from there. Over time, the teacher’s own expertise becomes the afterthought rather than the foundation.
Confidence shifts. When the tool consistently produces polished content that does not match your context, the subtle message is that your context is the exception. The default is what the AI knows. Your local knowledge is the thing that needs to be added on top, if there is time.
Your Expertise Is the Missing Piece
Here is what I have come to believe after working on AI in education across multiple countries: the problem is not that teachers need to get better at using AI. The problem is that AI tools have no way of learning what teachers already know.
Think about what you carry into your classroom every day. You know which topics your learners struggle with. You know which real-world examples spark a conversation and which ones get blank stares. You know the difference between what CAPS says on paper and what actually works in practice. You know your community’s economic realities, the businesses your learners walk past, the financial decisions their families navigate.
That knowledge is not a nice-to-have. It is the thing that makes education work. And right now, it is completely absent from the AI tools teachers are being told to adopt.
What Could Be Different
The practical tips are worth knowing: be specific in your prompts, give the AI your own examples to work from, always check outputs against what your learners would actually recognise. These help.
But I think the bigger question is worth sitting with. What if AI tools could actually learn from teachers, rather than the other way around? What if your expertise, your local knowledge, your understanding of your curriculum and your learners, could be built into the tool itself, so you did not have to re-teach it every session?
That would mean teachers shaping AI, not just consuming it. It would mean South African classrooms getting tools that reflect South African education, not adapted American templates.
Whether that is possible, and what it would actually take, is something I think is worth exploring seriously. But it starts with recognising that the most important knowledge in any classroom is not inside the AI. It is inside the teacher.