BIGGER PICTURE Learning & work
Six weeks of guided practice changed what pupils could do
In Benin City, a chatbot became part of an after-school English programme. Teachers, a curriculum and protected practice time turned a general-purpose model into a tested learning intervention.

- Programme duration
- Six weeks
- English outcome
- +0.23 standard deviations
- Combined assessment
- +0.31 standard deviations
How the idea works
Set a curriculum goal
Teachers frame the English practice.
Practise through dialogue
Students receive responsive explanations and feedback.
Assess independent learning
A separate assessment compares randomized groups.
Extra practice with someone who responds
The programme worked with first-year senior secondary pupils in Benin City, Nigeria. During six weeks in 2024, students practised English with Microsoft Copilot, then powered by GPT-4. The intervention created additional opportunities to ask questions and receive feedback within a school programme, rather than assuming pupils could independently turn unrestricted chatbot access into learning.
The teachers designed the learning situation
Teachers connected sessions to the national curriculum and supported students in using the tool purposefully. The World Bank’s account emphasizes preparation, access to equipment and guidance as essential parts of the programme. This is a concrete design choice: AI provides repeated conversational practice while the school supplies the learning objective, structure and human support.
Sources: [3]
The test happened away from the chatbot
After the intervention, students completed a pen-and-paper assessment. The updated research record reports a 0.23-standard-deviation improvement in English, the primary subject, and 0.31 across the combined assessment of English, AI knowledge and digital skills. Random assignment makes this stronger evidence than satisfaction surveys or a count of how often students opened the tool.
A programme effect, not magic in the model
The comparison measures the supported after-school programme as a package. It cannot separate the model’s contribution from extra study time, teacher guidance or the structured activities. Nor does a short trial establish durable gains everywhere. The encouraging finding is that a deliberately designed human–AI learning setting improved assessed skills in this specific resource-constrained context.
The original sources
- World BankFrom Chalkboards to Chatbots: reproducibility record
- World Bank research teamTransforming learning in Nigeria, one prompt at a time
- World BankPeople-centered AI in education: Five lessons from the Global South