Addressing the learning crisis with generative AI: lessons from Edo State in Nigeria

In developing countries, the education sector is facing a learning crisis. Despite increased enrollment rates, the quality of education remains low, with many students unable to achieve basic proficiency in reading and mathematics. The learning poverty rates, exacerbated by the COVID-19 pandemic, reach up to 70% in some regions. The learning deficit tends to accumulate over time, and manifest in secondary school through very low achievement rates, which creates important challenges for youth to access quality jobs. In Nigeria, where educational challenges are particularly acute, we explored whether generative AI models could serve as effective after-school tutors. Our recent working paper: From Chalkboards to Chatbots : Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria, provides insights into this innovative approach.
 

Background and Context

Nigeria's education system struggles with low learning outcomes, particularly in public schools. To help address this challenge, we attempted to use generative AI, specifically ChatGPT, to improve educational experiences and outcomes by providing personalized tutoring that adapts to individual student needs.
 

What was the program?

We conducted a randomized controlled trial in Edo State, Nigeria, involving public senior secondary school students. Over six weeks, students participated in an after-school program, randomly assigned to either a treatment or control group. Students in the treatment group attended sessions twice a week, interacting with ChatGPT using carefully crafted prompts.

Each session began with a teacher introducing a topic aligned with the official English language curriculum, guiding students to engage with the AI in a way that encouraged reasoning rather than shortcuts. The AI served as a tutor, helping students think through concepts, offer examples, and ask follow-up questions. Teachers supported the interaction, ensuring students stayed on task and understood the feedback, including identifying AI "hallucinations", and preventing over-reliance. [We described some lessons from the implementation here, including a video with voices from beneficiaries.]
 

What were the results?

The results were striking. Students in the treatment group outperformed those in the control group by a significant margin, with overall learning gains amounting to approximately 0.3 standard deviation and 0.24 in English. This is roughly equivalent to 1.5 to 2 years of typical learning progress in similar settings, surpassing the effects of 80% of rigorously evaluated education interventions globally. A clear dose-response relationship emerged: the more sessions students attended, the greater their gains. Despite challenges like intermittent internet and inconsistent electricity, the program's effects were robust. [You can check some of the highlights of the results here and  here.] 

Figure 1: Distribution of Assessment Scores (combined) by Treatement Condition

So, what explains such strong effects?

The success of the program lies not in the technology alone but in its application. The intervention was highly structured, and rooted in what we know works in education. Sessions followed a lesson guide aligned with the national curriculum, with prompts designed to engage students meaningfully. Teachers were trained ahead of the program to guide and monitor student interactions, ensuring the AI complemented their efforts. The AI provided students with a responsive tutor, adapting to each student's level and offering real-time feedback—something many public-school systems struggle to provide at scale.

Recommendations for Policymakers

Scaling this intervention presents challenges, particularly in regions lacking basic infrastructure like computers, electricity, and internet. Closing this gap will require serious investment. But it’s a challenge worth taking on. The learning crisis is too deep, and traditional approaches simply cannot deliver the transformational pace needed. Our cost-effectiveness analysis suggests a low marginal cost per student and a favorable cost-benefit ratio. As AI models improve, they could deliver an even greater impact at lower costs. Policymakers should consider exploring innovative approaches to address the learning crisis, and collaboration among researchers, the private sector, and governments is crucial.
 

Conclusion

We are careful to interpret the results as reflecting the full package: structured prompts, curriculum alignment, teacher support, peer learning, and the AI itself. We cannot isolate the effect of the chatbot alone. But we have good reason to believe it played a central role. Other tutoring interventions in the region—often with smaller group sizes and more instructional time—tend to show much more modest effects. The effects tend to be especially low in after-school programs and for secondary school. Here, students spent an average of 13 hours total over six weeks. Yet they learned far more than business-as-usual would predict.

There is also much we still don’t know. Will the effects last beyond the six-week window? Can the model help with subjects beyond English? Are students using what they’ve learned in other contexts? And what happens when these tools are embedded into regular classrooms, not just after-school programs? These are questions that require more research and experimentation.

What we do know is that generative AI, thoughtfully deployed as part of a holistic program and embedded in strong instructional design, can enhance student learning. While not a silver bullet, it offers a promising tool in the educational toolkit. The future of learning in low- and middle-income countries may depend not just on whether we use AI, but how. Now is the time to test, iterate, and innovate, ensuring AI's integration into education is grounded in good pedagogy and practical realities.