Where should AI in education begin in low-income countries?

AI in education… man. I find this one difficult. Done well, it could yield extraordinary results. But it feels like we’re still far away from that: so far, much of what we’ve gotten is a lot of students avoiding the productive struggle that makes learning stick, and educators dealing with the consequences without enough time, training or institutional support. One of the key questions I keep coming back to when I think about AI – particularly in global development – is whether a sector is bottlenecked by skilled labour. If it is, AI has huge potential in augmenting existing professionals and helping citizens directly. Education clearly is bottlenecked by skilled labour (similarly to health). Good teachers matter enormously, and most education systems do not have enough of them. On Wednesday, as part of Education World Forum in London, I joined a breakfast event on AI in Education in low- and middle-income countries, focused on fresh evidence from Sierra Leone – a leader and pioneer in improving primary and secondary education in low-resource settings in the last few years. To test if AI could improve learning in this setting, the Government of Sierra Leone, Google DeepMind and Fab AI (AI-for-Education.org) ran an RCT with 1,763 secondary students across 48 maths classrooms in Sierra Leone (i.e., 37 students per class on average). For 8 weeks in the fall of 2025, teachers used Gemini’s Guided Learning feature for half of maths lessons. Students used tablets or desktop computers, usually two students to one device. The Honourable Minister Conrad Sackey’s excellent and moving remarks emphasised that there is no substitute for government leadership in education. Irina Jurenka presented the impressive results (see below). Afterwards, Dayoung Lee and the Global AI Learning Alliance stress-tested their public good and ecosystem efforts to scale AI for learning. A few reflections on the broader question of AI in education in LMICs: 💡 Where to start? Classroom or homework use by students are very different; as are teacher use for coaching, lesson preparation, or assessment, or GenAI used for school administration. In low- and middle-income countries, I’m *much* more bullish on scaling teacher-facing use cases to start with, or homework support on parent-owned devices. 💡 The Sierra Leone pilot is encouraging, and also raises hard questions: Is this the highest-ROI, most scalable AI intervention for learning? Is the classroom the right setting to start in? How can governments actually select the tools that are best for their students and teachers? How many devices can we actually make available? How many classrooms have reliable power and internet? What training do teachers need? What happens to assessment? Lots of urgent work to do for students everywhere! Thanks to Mathias Abawi Esmann, Paul Atherton, Benjamin Piper, Serene Murad, Charlie Habershon, and Wongani Grace Nkhoma Sankofa, PhD for a thought-provoking discussion!