Edo State secondary schools (World Bank evaluation)
Edo State, Nigeria: Copilot after-school tutoring lifts English scores in a trial
Nine public secondary schools in Benin City, Edo State, ran a six-week after-school English programme in which students worked with Microsoft Copilot under teacher guidance. A World Bank randomized trial reports students scored 0.31 standard deviations higher overall, at about $48 per pupil.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
0.31
standard deviation
Gain in total score on the end-of-programme test for students selected for the programme versus the control group (English, digital skills, AI skills)1
0.238
standard deviation
Gain in the English score, the main outcome, for students selected for the programme versus the control group1
0.21
standard deviations
Gain on the regular third-term English exam, which covered the whole year's content1
72%
Average attendance among students assigned to the programme1
The problem
Nigeria, like much of Sub-Saharan Africa, faces teacher shortages and weak learning outcomes in secondary school. One-to-one tutoring is known to help students, but few education systems can afford it at scale. The World Bank team asked whether a free, general-purpose AI chatbot could act as a tutor in a low-resource setting.
What they built
An after-school English programme for first-year senior secondary students, typically about 15 years old, in nine public schools in Benin City, Edo State. Students met in school computer labs and used Microsoft Copilot, which was powered by GPT-4 at the time, as a virtual tutor for the national English curriculum.
How they did it
The programme ran from June to July 2024 and offered up to two 90-minute sessions a week, twelve in all, with about 30 students per session. Students worked in pairs on shared computers. Each session started with a teacher-supplied prompt tied to a topic from the regular English course, followed by free interaction with Copilot while teachers moved around the room.
Teachers did not teach directly. They had one three-day training and a toolkit with learning resources, an AI literacy handbook and session guides with suggested prompts. The prompts were written to make the chatbot act as a tutor that adapts to the student, using local names and examples. Monitors tracked attendance and session quality in each school.
Students were picked at random from those who volunteered, which allowed a randomized controlled trial against a control group.
Results
Students selected for the programme scored 0.31 standard deviations higher on the final test than the control group, with English up 0.238 standard deviations. They also did better on the regular third-term English exam, which covered the full year (0.21 standard deviations). Gains were larger for girls and for students with stronger earlier results, and each extra day attended added to the effect. Average attendance was 72%.
The authors put the pilot cost at approximately $48 per pupil. They also model the English gain as equal to 1.5 years of typical schooling in Nigeria, which is an estimate rather than a measured result.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Gain in total score on the end-of-programme test for students selected for the programme versus the control group (English, digital skills, AI skills) | 0.31 standard deviation1 | |
| Gain in the English score, the main outcome, for students selected for the programme versus the control group | 0.238 standard deviation1 | |
| Gain on the regular third-term English exam, which covered the whole year's content | 0.21 standard deviations1 | |
| Average attendance among students assigned to the programme | 72%1 | |
| Cost per pupil of the six-week pilot for 657 students | approximately $481 |
Lessons for your project
Keep teachers in the room. In this programme teachers set each session's opening prompt and supervised pairs of students, and the authors read the programme as a whole, including the teacher guidance and prompts, as driving the results.
Write the prompts for tutoring, not answers. Lesson guides told Copilot to act as a tutor and use local names and customs, so it asked questions instead of just giving answers.
Plan for power and internet cuts. Outages and the rainy season disrupted sessions, so backup power and connectivity matter before scaling.
Budget for content. Content development was 72% of fixed costs, so a second round should cost less per pupil.
Brief users on risks. Teachers and students were warned about wrong answers, bias and over-reliance in the first session.
Caveats
- The evaluation was done by World Bank staff who also helped design the pilot. The sources are the World Bank's own working paper and blog posts, with no outside confirmation.
- A World Bank blog thanks the Edo State Commissioner for Education and the Edo State Secondary Education Board for collaborating on the pilot, and the authors are World Bank staff. The sources do not say who funded or ran each part of the programme, so the case is attributed to the participating Edo State schools and the World Bank evaluation.
- The deployed date is the start of the six-week programme (June to July 2024). The published date is the December 2025 update of the working paper first published in May 2025.
- Of the students assigned, 657 to the programme and 671 to the control group, only 422 and 337 completed the final test. Attendance among those assigned to the programme averaged 72%.
- The paper's abstract rounds the English effect to 0.23 standard deviations and its introduction gives 0.24; the cost section gives 0.238, which is used here. The blog says 800 students took part; the paper's counts are used here.
- Claims of 1.5 to 2 years of schooling equivalent and of cost-effectiveness are the authors' modelling, and are not listed as metrics. It is a six-week pilot of English only and long-term effects are untested.
- Copilot was a free tool described as powered by GPT-4 at the time. The paper says the tool required only student registration.
Sources
- From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria
World Bank Policy Research Working Paper 11125 · 2025-12 · Research paper · documents.worldbank.org
- From chalkboards to chatbots: Transforming learning in Nigeria, one prompt at a time
World Bank Blogs (Education for Global Development) · 2025-01-09 · Company blog · blogs.worldbank.org
- From chalkboards to chatbots in Nigeria: 7 lessons to pioneer generative AI for education
World Bank Blogs (Education for Global Development) · 2024-09-18 · Company blog · blogs.worldbank.org
Summarized in our own words from the sources above. Spotted an error? See corrections.