Can AI strengthen foundational learning in African classrooms?

Midway through our session at the Yidan Prize Conference this year, a participant asked:

“Some countries in Northern Europe are removing AI and screens from classrooms altogether, worried about their effects on children’s critical thinking. So why are we so focused on bringing AI into African classrooms?”

That question helped us articulate why AI matters in our context and, more importantly, the conditions under which it can make a meaningful difference.

In sub-Saharan Africa, 89% of children cannot read a simple text by age 10. Although children are in school, many are not learning. The World Bank and UNESCO have described this crisis as learning poverty. 

African countries are responding by revisiting curricula, adopting more effective teaching approaches, and introducing national languages into the classroom. These reforms require investment to develop teaching materials and provide sustained support to teachers. AI is increasingly being explored as one way to accelerate this work. 

The question now is not whether AI should be used in African education. It is how to use it responsibly. Doing so requires high-quality standards, tools that work in the national languages children speak, and government leadership. 

These questions shaped many of the discussions at this year’s Yidan Prize Conference – the Foundation’s first conference on the African continent. Co-organized by the Yidan Prize Foundation, ARED, and ADEA under the auspices of Senegal’s Ministry of National Education, the event brought together around 200 ministers, researchers, practitioners, and funders to explore how to improve education outcomes across Africa.

As part of the conference, IDinsight facilitated a session on ‘AI for Foundational Learning,’ drawing on our experience building and testing AI tools for education in Senegal. The session brought together panellists from Pratham, Baamtu, Fab AI, ARED, and IDinsight. In this blog, we share the key insights that emerged from the discussion and highlight some of the tools demonstrated during the session.

Five insights on AI for foundational learning

1. Start with teachers and classroom needs

When we talk about AI in education, the focus should not be on the technology. The priority should be the people who will use it. The ultimate goal is to improve learning outcomes for children. To get there, teachers are essential. They are the ones assisting students in the classroom every day. We need to design tools that respond to their needs and fit into their daily practice. The most convincing AI tools we saw helped teachers access a lesson guide, translate a term, or hear how to pronounce a sound in a national language. They were convincing not because they were sophisticated, but because they were useful. To build tools that truly work, teachers need to be involved from the start.

2. Build for African languages and contexts

Africa has over 2,000 languages. AI works well in widely spoken languages like English or French, where abundant digital data is available. Some major African languages, like Kiswahili or Wolof, are also making progress. But for the vast majority, very little content has been digitized. And without digital data, AI cannot perform well. As one panelist put it: “We missed the bus with the last wave of technology. We cannot afford to miss it again with AI.” If we want AI to serve foundational learning in Africa, it has to work in the languages children actually speak, which requires investing in training AI models on national languages.

3. Design for equitable access

Some AI tools risk benefiting children who are already performing well while leaving the most vulnerable further behind. AI should help reduce inequalities. In practice, this means prioritizing out-of-school children who need catch-up support. It also means reaching children and teachers in remote areas, with tools that work offline and on basic devices.

4. Put governments in the lead

When it comes to AI in education, governments must be in the driver’s seat. This includes setting clear rules on what data can be used, where it is stored, and who controls it. As one panelist put it: “a government should be able to say that all data will be governed under its own rules, on its own infrastructure.” It also means safeguards. Education is too important for uncontrolled experimentation. Content must be validated, quality must be guaranteed, and governments should avoid becoming dependent on commercial solutions that are not grounded in local contexts. Ultimately, governments should be able to own and sustain these tools independently.

5. Generate evidence on what improves learning

There can be pressure to scale AI solutions quickly. But scaling a tool that has not been properly tested is ineffective and risky. Tools must be contextually relevant, pedagogically sound, validated by education authorities, and safe for learners. Building and testing good AI tools is costly.

Beyond quality, there is the question of evidence. What level of evidence should we require before deploying AI at scale? The AI Eval Playbook offers a useful framework with four progressive levels of evaluation:

1. Does the model perform as expected?
2. Do users engage meaningfully?
3. Does it change teacher practice?
4. Does it improve learning outcomes?

Each of these levels matters. But the closer we get to measuring impact on children’s learning, the more confident we can be that a tool is ready to scale.

What we saw being built

Beyond the panel discussion, our session included live demonstrations of AI tools currently being developed for education. Four organizations showed what they are building.

1. Pratham demonstrated PadhAI, an AI-powered reading assessment tool. A child reads a text aloud in Hindi, and the app analyzes the audio using speech recognition trained on children’s voices. Within seconds, it produces a detailed reading profile – words read correctly, missed, or misread – with 95% accuracy in assigning literacy levels. It works offline, on a basic Android phone, at $0.50 per assessment. For teachers who currently spend weeks assessing a class with pen and paper, this is an important shift.
   
2. IDinsight presented two tools. The first is a WhatsApp-based assistant for teachers implementing Senegal’s bilingual education reform. Through an app they already use daily, teachers can access their teaching guides, translate terms into national languages, and get pronunciation support – all grounded in Ministry-validated content. The goal is to help teachers feel more confident in bilingual classrooms where they often receive little support. The second is a content generation tool built on knowledge graph technology. The official curriculum has been encoded into a knowledge graph, allowing the tool to generate first drafts of lessons in a few minutes for content producers to review and refine.
3. ARED showed how they use generative AI to produce children’s storybooks in Senegalese national languages – Wolof, Pulaar, Serere, and Mandinka. AI generates first drafts of stories and illustrations, but each book then goes through human review: adaptation to national languages, quality control of images and content, and layout. The result is culturally relevant books in languages where very few reading materials exist for children.
4. Fab AI presented the results of a randomized controlled trial evaluating Google’s Gemini Guided Learning tool in Sierra Leone, in partnership with Google DeepMind and the Ministry of Education. Students used Gemini in pairs on tablets for mathematics. The study found improved learning outcomes, with equal benefits for boys and girls, and higher usage associated with larger gains.

Looking ahead

The distance between a demo and a tool that a ministry can run at scale is important. There is healthy skepticism around AI in education, and we share it. As the organizations building these tools, we have a responsibility to test them rigorously, to be transparent about what works and what does not, and to invest in evidence before pushing for scale. In the coming months, we will continue testing and evaluating our tools in Senegal, and we will share what we learn along the way.

The IDinsight team at the Yidan Prize Conference included: Nasita Fofana (Manager), Karimou Ba (Technical Project Director), Spero Faladé (Senior Associate), and Rebecca Sharp (CEO). IDinsight’s participation in the AI for Foundational Learning working group was supported by the Gates Foundation.