Nigeria tested AI tutoring with teachers, not instead of them. English scores rose

In Benin City, Nigeria, the experiment was not to hand pupils a chatbot and leave them alone. The intervention combined computer-based conversation, English practice, peer collaboration and teachers' supervision. That arrangement matters because it is the actual arrangement a randomized study tested—not a hypothetical world in which AI replaced a classroom.[1]

A reanalysis published in Education Sciences in October 2026 revisited a six-week programme across nine public secondary schools. Of 1,328 pupils assigned by lottery, 657 were offered an extra twelve sessions of 90 minutes each using Microsoft Copilot, English-language exercises and paired work. The remaining 671 continued with regular instruction. Among pupils for whom the researchers had the required final measurements, the programme was associated with a 0.310-standard-deviation improvement on a combined assessment, and a 0.238-standard-deviation gain in English.[1]

Those numbers are not photographs of transformed classrooms. They are standardized comparisons of test performance between randomized groups. But they are meaningful: in a setting with limited educational resources, a structured intervention produced measured short-term learning gains rather than simply more chatbot use or more enthusiastic survey responses.

What was actually delivered

The programme used a general-purpose AI assistant as one component of supervised teaching. Students practised English concepts with curriculum-aligned prompts, collaborated in pairs and had adults available to guide or correct them. The twelve sessions represented up to eighteen additional hours of organised learning during six weeks. A bare smartphone with a chatbot and no support would be a different intervention.[1]

The 2026 paper describes 657 students assigned to the treatment and 671 to the comparison group. The overall adjusted estimate used 654 pupils with the final assessment and required covariates. Researchers reported a 95% confidence interval of 0.176 to 0.444 standard deviations around the 0.310 estimate. For the English component the estimate was 0.238, with an interval of 0.104 to 0.372. The intervention also raised measured knowledge of AI, by 0.309 standard deviations.[1]

Crucially, the effect did not depend solely on a test designed around the tutoring activities. A school-administered English examination that covered broader content showed a 0.206-standard-deviation gain in an analysis of 636 pupils. That strengthens the case that the programme helped with more general English learning, though a school exam is not the same as a long-term national learning outcome.

Digital skills provided a useful counterpoint. The estimated gain was 0.139 standard deviations, but its confidence interval included zero. The evidence therefore does not establish a reliable improvement in that outcome. A successful educational evaluation should have room for results that were encouraging, results that were clear and results that remained uncertain.[1]

The data problem that success cannot erase

Of the 1,328 randomized pupils, final assessment records were linked for 759: 422 in the AI-tutoring group and 337 in the comparison group. Expressed as shares of the originally allocated groups, that is roughly 64% and 50%, respectively. Pupils with observed endline data are not necessarily representative of everyone randomized, especially when the likelihood of follow-up differs between groups.

The researchers adjusted for baseline measurements and school factors and ran sensitivity checks. One inverse-probability-weighted estimate for the overall assessment was 0.298 standard deviations, close to the headline 0.310. This makes the finding more credible under the weighting assumptions; it does not magically recover the missing results or prove the estimated effect would apply to pupils who never joined the assessment.

A further interpretive trap is attributing the whole effect to the language model. The intervention combined extra teaching time, supervision, peer practice, suitable prompts and AI feedback. Randomization estimates the effect of offering that package relative to ordinary instruction. Without an equally intensive non-AI tutoring arm, it cannot isolate how many points came from the model rather than from having structured extra help.[1]

An encouraging result with a usable blueprint

The study's particular contribution is not that generative AI can teach better than a person. It is that schools can organise AI-assisted practice in a way that fits a real curriculum and yields measurable improvement in a constrained setting. It also shows what must be budgeted if others want to repeat it: devices, access, electricity, instructor time, training, prepared activities and enough attendance to give the intervention a chance.

The World Bank's original 2025 working paper described the same programme, and the 2026 analysis re-examines its data; these are not two independent trials. Treating them as one evidence base prevents a common mistake in fast-moving AI reporting: counting republication or reanalysis as replication.[1]

There is no evidence here of a lifetime benefit, a nationwide rollout or an AI tutor capable of replacing teachers. There is evidence of pupils gaining ground on tests after six weeks of extra, guided learning. For families, schools and policymakers, that measured gain is valuable enough without turning a practical educational result into a technological miracle.[1]

Report an error or suggest a correction