AI for underserved classrooms: Powerful learning with no internet

An AI maths tutor on WhatsApp boosts learning a year for $5/child, proving AI for underserved classrooms doesn't need fast internet or expensive tech.

In this episode:

• An AI maths tutor named Rori on WhatsApp is achieving a year of learning gains for just $5 per child, demonstrating powerful AI for underserved classrooms without needing fast internet.
• Solutions like FoondaMate and Juza AI prove that offline AI education and low-resource AI learning can effectively support millions of students in developing countries.
• A 2025 UNESCO paper advocates for an 'offline-first' principle for scaling AI in African schools, highlighting the importance of resilient edtech in developing countries.
• Digital Public Infrastructure (DPI) can provide shared resources like local-language corpora and governance rails, enabling more effective and equitable AI for underserved classrooms to scale.
• Pedagogy and teacher trust are non-negotiable; AI tools must align with how children learn and enhance, not replace, the teacher's role for successful implementation.

Chapters:

• 00:00 — Cold open & welcome
• 00:30 — Rori: AI for underserved classrooms on WhatsApp
• 01:30 — Challenging assumptions: AI doesn't need perfect infrastructure
• 02:45 — The paradigm shift to low-resource AI learning
• 03:45 — Evidence of AI in African schools: FoondaMate, GlobeDock, Juza AI
• 04:45 — Building for constraints: Offline AI education and edge computing benefits
• 05:45 — The missing scaffolding: Education Digital Public Infrastructure (DPI)
• 07:15 — Non-negotiables: Pedagogy and the teacher's role
• 08:30 — Strategic investment for AI in African schools
• 09:00 — Conclusion: AI for the world as it is

How can AI support learning in underserved classrooms without internet?
AI tools like Rori, FoondaMate, and Juza AI demonstrate that effective learning can occur using simple phones, minimal data, or entirely offline 'AI in a box' solutions by syncing data periodically or running models locally.

What are the key benefits of low-resource AI learning for students?
Low-resource AI learning provides accessible, affordable, and high-quality educational support to a broad range of students, particularly the 'middle 80%' who may not have access to traditional tutors or high-tech solutions.

What is Digital Public Infrastructure (DPI) and how does it help scale edtech in developing countries?
DPI creates a shared, public layer of educational resources such as curriculum-aligned content, local-language data, and governance standards, allowing local developers to build scalable, safe, and interoperable AI solutions.

Featuring: Dan Fitzpatrick, WhatsApp, Rori, FoondaMate, World Bank, Messenger, GlobeDock Academy, Juza AI, UNESCO.

Follow AI in Education with Dan Fitzpatrick for more on AI in education.

Transcript

SPEAKER_00 0:00

If this episode makes you think, please let us know in the comments and support us by subscribing and leaving a review. Thank you. Today we are exploring a really eye-opening article titled AI Doesn't Need Perfect Classrooms, written by Yamaka Rivon on Yemenam, an advisor at Africa Practice, and published by Brink, a publication now part of Africa Practice. What really caught my attention right from the start and what I think will grab yours too, is the story of a maths tutor that has no face, no fee, and no need for fast internet, producing learning gains worth roughly an extra year of schooling. At a marginal cost of about $5 a child. It's a powerful counter narrative to almost everything we assume about AI in education. See, when most people think about what artificial intelligence needs to work, they conjure up images of massive data centers, blistering fast internet, powerful computers, and almost always everything being in English. And you know what? That picture has pretty much become common sense. But what a marker's article makes crystal clear is that this widely accepted view might actually be out of date. For a growing number of classrooms around the world, especially those we often think of as underserved, a much subtler, more resilient approach to AI is already making a huge difference. These are AI for underserved classrooms, working on cheap phones with weak signals and in local languages. It's a powerful lesson for all of us, no matter where our schools are located. First up, let's challenge some core assumptions. The article describes a teenager in Ghana opening WhatsApp on what could be a $10 phone, starting a maths lesson with an AI tutor called Rory. No laptop, no broadband, no specialist teacher in the room. This isn't just a pilot, it's a reality reaching more than 40,000 learners, and as I mentioned, it's achieving learning gains equivalent to an entire extra year of schooling for just five US dollars per child. That's simply incredible, isn't it? It completely flips the script on the idea that you need to solve all your infrastructure problems before intelligent tools can add value. For too long, the narrative, especially in places like Africa, was infrastructure first, then connectivity, then devices, and only then the clever software. And what the article reveals is that this sequence meant the places that needed tech the most always ended up at the back of the queue. What this article highlights is a paradigm shift. AI built on the premise that it should meet people where they are, not demand they arrive where the technology is most comfortable. It's designed to run on the edge or work completely offline with little or no connection. Think about it, it's built for the languages and the cheap devices people actually have. This focus on low resource AI learning is profoundly important. We often focus on the top 20% of high achievers or the bottom 20% needing intervention. But AI like this can truly serve the middle 80%. That often invisible majority of students who can greatly benefit from accessible, low-cost, quality learning support. The second thing that really stands out is the sheer volume of evidence presented here, suggesting this isn't just a hopeful theory. The article mentions a study across six African countries where the consistent finding was that AI already works across the entire connectivity spectrum. Beyond Rory, there's Fundimate, for instance, a study tool leveraging WhatsApp and Messenger, now helping over a million students on cheap handsets and with minimal data. In Ethiopia, Globed Doc Academy runs an AI-powered offline first platform reaching more than 200,000 secondary learners across 200 towns, syncing data through SD cards and local caching where the network just won't cut it. And in Kenya, JUSA AI packages a curriculum-aligned tutor into an offline AI in a box that runs on a low power computer in classrooms. Even a 2025 background paper for UNESCO's Global Education Monitoring Report goes so far as to make Start Offline, its first principle for scaling AI in African schools. Now why does this idea of building for constraints matter for all of us, not just those in EdTech in developing countries? Well because these established engineering choices that prioritize resilience and reliability have benefits everywhere. Edge AI, where models run on local devices like phones or school servers, can be faster, cheaper and more private, even in places with reliable connectivity. Smaller models are less expensive to run and easier to govern. Offline first software which syncs only when a connection is restored, ensures the tool remains useful under real-world conditions. It's about building technology that works whether the internet link is strong, weak, or completely gone. This isn't a wager against connectivity, it's a strategy for robust, dependable technology that puts purpose over technology. Ensuring the tool serves the educational need first and foremost. Here's why this idea of missing scaffolding is so powerful for school leaders and educational system designers. The article argues that the question isn't whether these tools can work, but why so few of them move beyond successful pilots to become widespread systems. The recurring challenge isn't the ingenuity of the tools or the talent of the developers. It's the architecture. The shared scaffolding that allows a good tool to outlast its grant funding. What's missing, the author contends, are the shared public goods that let any good EdTech solution scale. Think about it, curriculum aligned content repositories developers can build upon, local language corpora with clear licensing, lightweight model baselines tuned to local contexts, and horizontal AI infrastructure that every application can draw on. This is the case for an education digital public infrastructure, or DPI. Imagine a public governed layer of educational goods sitting within the national system that schools, teacher colleges, and developers can all draw from. The article describes it as a stack. At the base are the applications learners and teachers actually touch, micro-tutoring workflow support, multilingual learning. Above those are the technical rails, the corpora, the model baselines, the offline ready architectures, and above Chinese those whole sit the governance rails, privacy standards, safeguarding protocols, curriculum alignment and procurement rules. Get this stack right and you empower a thousand local builders to innovate on top of it, ensuring solutions are safe, interoperable, locally grounded and curriculum aligned before they ever reach a child at scale. This aligns perfectly with my own seven lessons for AI adoption framework, particularly the need to align and customize for institutional context, rather than just throwing tools at problems. If you're finding these conversations helpful and want to stay ahead of the curve in AI for education, please consider following this podcast for more insights and discussions that will help spark your imagination to innovate. And this brings me to the two non-negotiables that the article firmly identifies, and honestly they resonate so deeply with my own philosophy. The first is pedagogy, a technically elegant tool that's pedagogically empty, the author warns, will simply teach a child how to game it. The content, they argue, has to be rooted in evidence about how children actually learn and how cognition works, sequenced for real curricula and vetted by the people who train teachers. Otherwise, we risk automating poor instruction at scale, which would be far worse than the gap we started with. This is all about purpose over technology and design and learning that cannot be faked because it demands depth, care, and imagination. It speaks to the idea of cognitive stretch, creating tasks that require application and judgment, not just recall. The second non negotiable is the teacher, wherever a teacher exists. Every serious finding in their research and every honest founder they interviewed pointed the same way. If an AI tool adds to a teacher's workload, it dies. If it fails to earn a teacher's trust, it never scales. This is a crucial reminder that teachers are not resistant to change. They just need time and space. We need to design for the teacher as carefully as we design for the learner. Teachers remain the highest leverage point in the educational chain, the actor who can truly turn a clever app into genuine learning. This reinforces the core pillar of enhancement not replacement and keeps the human in the loop. Even in places where teaching capacity is severely stretched, the principle is that technology should strengthen the human support around learning, whether that's a teacher, a parent, a community center, or even the learner working independently. What this article really asks us to consider is that the money now being mobilized, like the ten billion US dollars the African Development Bank and the UN Development Program are aiming to raise by 2035 for AI initiatives, needs to land strategically. It needs to build those shared rails, the corpora, the safeguards, and the teacher tools that will compound for a generation, rather than scattering across more pilots that each rebuild the same foundations alone. Ultimately the bigger point is this a learner in an under-resourced environment doesn't need us to wait for perfect conditions, nor to import tools built for a different world. She needs small, capable, locally grounded AI, the shared infrastructure that makes it dependable and crucially where available, a teacher whose hand it strengthens. This is AI for underserved classrooms. Built for the reality of the learner, not some idealized future. It's about evolution, not revolution. The real opportunity is to build AI for the world as it is, not the world we keep promising. That's all for today. Thanks for listening.