Six weeks of AI tutoring improved English results in Nigerian public schools

A school computer lab became a place for repeated, personalised English practice in a six-week experiment in Edo State, Nigeria. Students offered the after-school programme achieved better assessment results than a control group continuing its usual schooling.[2]

The intervention ran in June and July 2024, using the then-free Microsoft Copilot powered by GPT-4. Across nine public secondary schools, students worked in pairs while teachers introduced curriculum topics and kept the conversations productive. The central opportunity was to give many learners a responsive practice partner within a shared lesson.[2]

How the practice worked

The teaching toolkit supplied a starting structure, then left room for students' questions. Its lesson plans covered clauses, essays, comparisons, letters and creative writing. Prompts identified the learner's level, asked for explanations and encouraged follow-up. Pair work also had explicit rules: take turns asking, examine the output together and let the teacher know when sharing was not working.[4]

Consider an illustrative exercise in subject–verb agreement. A pair could submit an incorrect sentence, ask why the verb needs changing, then request a fresh example to solve themselves. The educational value comes from the next attempt: can they apply the explanation? This is an example of the practice the design enables, rather than a transcript collected from the trial.

What improved on the tests

The lottery assigned 657 interested students to the programme and 671 to control. The final assessment was completed by 422 and 337 respectively. The main adjusted English analysis used 654 students and accounted for previous examination performance and school differences.[1]

A standard-deviation effect expresses a score difference relative to the spread of marks. The regular school examination adds a useful check on whether the gain appeared beyond the programme’s own assessment.[1]

Missing final assessments were substantial and differed between groups. The authors' bounding and weighting checks retained positive effects. The comparison nevertheless estimates the whole after-school package against usual schooling; it cannot isolate Copilot's contribution from extra practice, teachers and peers, or establish how long the gains lasted.[1]

The practical details behind the result

The implementation account describes students first learning to navigate computers and create accounts. Power failures, disrupted internet access and heavy rain interrupted sessions. Backup connectivity, local support and monitors helped keep the programme moving. These details matter because the time available to learn begins only after the equipment and accounts are working.[3]

Teachers also formed informal groups to exchange what they were learning about the tool. That creates a useful route for improvement: a prompt that explains a confusing point well can become available to another class, while a misleading answer can be discussed before it recurs. The authors describe that exchange as part of the programme's implementation experience.[3]

A programme other schools can examine

The World Bank subsequently released the teaching materials, assessment data and analysis scripts in a reproducibility package. Its Development Impact analytics team reports checking computational reproducibility. That checks whether the supplied analysis produces the reported exhibits; it is separate from repeating the intervention in another school system.[5]

For educators, the useful inheritance includes actual lesson plans and reflection routines. Another school can inspect how the sessions were organised, adapt the materials and evaluate its own results. The Nigerian experiment therefore leaves both a measured learning gain and a concrete teaching design that others can build on.[4][5]

Report an error or suggest a correction