"What is the total number of students who like football?" asks Rori, a maths tutor in Ghana. You carefully study a bar chart but, despite being a whip-smart reader of African Business, you get the answer wrong.
"Oops! Looks like a mistake," says Rori cheerfully, then gives you some advice and encourages you to have another go. The tutor is a chatbot, not a human; the entire conversation is happening on WhatsApp. It is still early days, but this exchange is a glimpse of how artificial intelligence might change education in Africa.
Early trials of educational AI in African countries show encouraging results, with effects equivalent to months or even years of additional learning. Many of these tools try to give students an approximation of one-on-one teaching. Others are helping teachers plan lessons or mark assignments, and even writing textbooks. Sometimes AI is delivered through bespoke tools, like Rori; often, teachers and students are just using Gemini, ChatGPT or MetaAI (which is embedded in WhatsApp).
All this comes with risks. Over-reliance on AI could encourage "cognitive offloading", in which students hand over the hard work of thinking to a machine. Educational inequality could widen, as students without access to AI get left behind. There are concerns too about how to keep children safe online and how their data gets used.
"Certainly we have seen an uptick of the use of AI, but of course most of it is not really guided by any institutional policies," says Admire Mare, head of the Department of Communication and Media Studies at the University of Johannesburg and co-author of AI in African education, a report for ECDPM, a European think-tank.
Raising the bar in schools
The origins of Rori go back to the Covid-19 pandemic when Rising Academies, an educational network which works with 1,400 government schools in Africa, used radio and text messages to help students learn even when schools were closed. Later the idea evolved into a chatbot. Questions are written by maths experts, while AI helps interpret students' answers. Most of the material is text-based, so that it can be accessed on low-end smartphones.
The potential for this kind of tool was demonstrated by a preliminary evaluation, conducted with researchers at the University of Oxford and J-PAL, a research network. About 500 Ghanaian school students were split into two groups - one that used Rori for an hour a week, and one that did not - then tested twice over an eight-month period. Those who used Rori gained the equivalent of at least a year's worth of learning, relative to the control group. A larger trial, spanning five districts, is under way.
"Students work through problems, they don't paste a question and get an answer back," explains Hannah Horne-Robinson, Rori research and assessment manager. "Rori interprets what they wrote and gives feedback, and if they're stuck, they get a hint first and a chance to try again."
Those findings were echoed in a different trial in Sierra Leone, where school students used the Guided Learning feature of Google's Gemini chatbot. After just eight weeks, they had made as much progress in maths as they normally would over 1.2 to 1.7 years. The actual "effect size", the underlying measure of gains, is smaller than in the Rori study, but still impressive. More than 90% of students' conversations with Gemini were about building their understanding, rather than just asking for the answer. The chatbot used pedagogical techniques like "scaffolding", in which a teacher helps students master new concepts by giving structured support, like breaking a process down into simple steps.
"If you just give students access to a chatbot, it could actually have detrimental effects, because it just becomes a tool for finishing your homework really quickly," says Mathias Esmann, director of partnerships at Fab AI, a not-for-profit organisation that helped run the study.
"However, if you then incorporate scaffolding, and you try to make the AI tool do what a good teacher would do or a good tutor would do, then it can actually produce quite a bit of learning."
In short, the principles of good teaching are the same, whether it is delivered by a person or a computer. The researchers in Sierra Leone worked carefully with teachers to train them in how to use AI with their students. Fab AI has also developed a benchmark to assess how much the models themselves know about pedagogy, the art of teaching: in other words, says Esmann, "does an AI model know enough to teach?"
Three years ago, the models did not do very well on this test. Now the best models ace it - in English at least.
In theory, an AI tutor can offer students the kind of bespoke tuition that teachers would love to give, but are too overworked to deliver. That chimes with the proven method of "teaching at the right level", whereby classes and content are tailored to how much students already know. Esmann envisages AI tools that can assess how well children can read, then write texts at just the right level of difficulty to help them reach the next stage. He also anticipates that many more students will soon be interacting with AI tutors through voice, not just text, and in their own mother tongue.
Some of those ideas are already on the way to reality. RobotsMali, a Malian NGO, uses ChatGPT and the DALL-E image generation app to quickly produce Bambara-language texts for children. The Edutech Institute, a South African company, is bringing Indian educational software to Africa. M-Shule, in Kenya, uses AI to generate quizzes and recaps, which can be sent by WhatsApp or text message.
Kytabu, another Kenyan start-up, is trying to develop AI-powered smart glasses for teachers that give "real-time lesson prompts, student engagement insights, and voice-activated classroom support".
The limits of technology
A well-known problem in educational research is that promising results in trials are not always replicated at scale. The Sierra Leone study, for example, not only provided devices to use, but also hired "field monitors" who ensured the devices were charged and resolve any technical issues - the kind of support that is unlikely to exist outside of a trial.
For most African children, an AI tutor remains a distant dream. Less than a third of Africans are connected to mobile internet, and barely half have electricity. Most studies so far have been conducted in or near cities, where smartphone access is greatest, and in some cases researchers have handed out devices to students.
Some worry about the suitability of models trained on data that mostly comes from other continents.
"There are challenges obviously around algorithmic bias, datasets that are not representative, AI tools that are not really nuanced in understanding the African dynamics," says Mare.
He is also concerned about the way that African data might be handled by big tech companies. "Unfortunately most African countries are just coming to terms with these issues now, and most of the data protection laws really are designed with social media in mind, and not with AI media in mind."
Teachers themselves tend to emphasise the human bond they share with their students, which a machine can never hope to replicate. Horne-Robinson agrees.
"Teachers and students will still be at the centre," she argues.
But innovative tools might help them to work together in new ways, she continues.
"What I hope is that in ten years, the school you happen to attend doesn't determine whether you get quality, personalised support."