AI Tutoring in Schools: What Trials in Nigeria, Turkey, Ghana and the US Show

The use case

A generative AI tutor is a chat interface built on a large language model and instructed to guide a student through a problem, explain a concept or give feedback, rather than simply supply the answer. This paper separates four things that are often grouped as "AI tutoring". The first is the student-facing generative tutor, which is the subject here. The second is generative AI that coaches a human tutor instead of the student. The third is adaptive learning software and rule-based tutoring, which has been studied for decades. For example, the World Bank authors cite a gain of 0.37 standard deviations in maths after 4.5 months of technology-aided after-school instruction in India¹. Those results are context, not evidence about generative tutors. The fourth is extra study time and human teaching, which produce gains of their own. Automated marking and plagiarism detection are outside the scope of this paper.

Where it is running

Almost all measured deployments are research pilots, not system-wide services.

• Nigeria. A World Bank team ran a six-week after-school programme in nine public secondary schools in Benin City, Edo State, from June to July 2024. It used Microsoft Copilot (GPT-4) for English with about 15-year-olds, and 657 students were assigned to the programme¹.
• Turkey. One large high school ran four 90-minute maths sessions in autumn 2023. They covered about 1,000 students in grades 9 to 11 and used GPT-4 tutors built for the study².
• Ghana. Rori, a maths tutor delivered over WhatsApp, was tested with about 1,000 students in grades 3 to 9 in 11 schools. Students used it for two 30-minute sessions a week over eight months⁶.
• United States. Eighteen Tennessee middle schools used Khan Academy with its Khanmigo tutor in daily remedial maths sessions over two years³. Two unnamed high-poverty districts tested an AI literacy tutor with elementary pupils⁵. A separate trial put an AI assistant in the hands of human tutors working with about 1,800 students in grades 3 to 8⁷.
• Estonia. The AI Leap programme plans a learning app that does not give direct answers. It is aimed at 20,000 students in grades 10 and 11 in 2025/26, with 38,000 more students and 2,000 more teachers planned for 2026/27 (operator-reported plans)⁹. I found no published outcome results.

What the evidence shows

Nigeria. In a randomised controlled trial among volunteers, students offered the programme scored 0.31 standard deviations higher on an end-of-programme test. They scored about 0.24 higher in English, the main outcome, and 0.21 higher on the school's own third-term English exam (all measured against a control group, May 2025 working paper)¹. The control group had no extra sessions, so the result combines the tutor, the teacher facilitation and about 18 extra hours of study. The authors' "1.5 to 2 years of schooling" and their projection of 1.2 to 2.2 standard deviations over a full year are modelled figures, not measurements¹.

Turkey. In a preregistered trial published in June 2025, students with a standard GPT-4 chat scored 48% higher on practice problems but 17% lower on the later closed-book exam than students with no AI. Students with a tutor prompted to give hints scored 127% higher in practice and the same as the control group on the exam².

Ghana. Rori students gained 0.37 standard deviations over eight months, measured in a school-level trial. This was year one only, and the authors call it preliminary⁶. I did not verify how much of Rori's dialogue is generated rather than scripted, or who ran the evaluation.

Tennessee. The two-year cluster randomised trial (August 2026) found assignment to Khan Academy with Khanmigo raised maths achievement by about 0.06 to 0.08 standard deviations per school year. The implied effect of a full year of active participation is 0.14. Gains "resemble those from Khan Academy practice without AI assistance"³. The comparison group used other digital programmes plus small-group teacher instruction⁴.

Elementary literacy. In an AI reading tutor trial (reported June 2026), students barely used the tool, and there was no meaningful difference in reading scores⁵.

Tutor-assist. Students whose tutors had the AI assistant were 4 percentage points more likely to pass lesson exit tickets (62% to 66%), and students of lower-rated tutors gained 9 points. The study found no significant gain on end-of-year tests (randomised, two months, 2024)⁷.

No trial I opened compares a generative tutor with a human tutor given the same time. The Nigeria authors argue that the gain is unlikely to come from teacher time alone, but that is an argument, not a test⁸.

What it takes: cost, integration, governance

Cost. The Nigeria pilot cost about $48 per pupil, with a marginal cost of about $9. A four-quarter version was estimated at $124 per pupil (planning and budget data, modelled). Content development was 72% of fixed costs¹. The tutor-assist tool cost about $20 per tutor per year, based on usage in the trial⁷. Operator figures for the commercial products were not available in the sources I opened.

Devices and connectivity. The Nigeria sessions needed school computer labs and routers, and power and internet cuts were frequent¹. Rori ran on basic phones over low-bandwidth networks⁶.

Teacher and supervision. Nigerian teachers had three days of training and a toolkit. They circulated during sessions and monitors tracked fidelity¹. The Nigeria authors say supervised, structured use is what separates these gains from benefits that stay with already-motivated students⁸.

Curriculum and guardrails. Each Nigerian session followed a curriculum topic with prepared prompts. The Turkish hint-giving tutor was fed teacher-written solutions and common mistakes for every problem, which the authors call labour-intensive². Estonia's app is built to coach rather than answer⁹.

Language. Nigerian prompts used local names and customs¹. Estonia's app is checked by the Institute of the Estonian Language⁹. Tutor-assist suggestions were sometimes "too smart" for the grade level⁷.

Rules. In the EU, emotion recognition in education has been prohibited since February 2025. Chatbot transparency rules take effect in August 2026. High-risk education systems, such as exam scoring in the Commission's example, face obligations from 2 December 2027 under the AI Omnibus, which entered into force on 27 July 2026¹⁰. Whether a given tutor counts as high-risk depends on its function, and I did not verify that classification. England's Department for Education publishes product safety standards for developers and suppliers; I did not review their detail¹¹. I did not find child-data or age-rule sources, which is a gap.

Where it falls short

Low use. In Tennessee, 96% of students tried Khanmigo, but the median student messaged it on a third of practice days and in 17% of sessions where they made a mistake. Most messages were bare answers or clicks on suggested prompts³. In the literacy trial, students averaged two to ten minutes a week, and a human helper did not raise logins. Those who did log on were more likely to be higher-performing⁵.

Answer-copying and errors. In Turkey, standard GPT-4 gave the correct answer to a practice problem only 51% of the time. Students mostly copied answers, and they did not perceive any loss in learning².

Uneven effects. In Nigeria, gains were larger for students with higher prior attainment and higher socio-economic status. The effect for girls is driven partly by one girls-only school¹.

Implementation. Nigeria had account-setup problems and some control-group contamination¹. Tutor-assist raised session results but not end-of-year test scores⁷. I found no reports of paused or withdrawn programmes, and no privacy incidents, in the sources I opened. Tutor-assist authors note that de-identification of names does not remove all identifying details⁷.

Challenge test: why this may look less successful than presented. The strongest figure, from Nigeria, comes from one six-week pilot with volunteers. It had a small sample and differential attrition: 422 of 657 treated and 337 of 671 control students completed the endline. It has not been replicated, and the authors say replication is needed¹. Authors of the Nigeria study were also involved in implementation¹. Nothing I found measures gains after the programme ends. The Nigeria authors list durability as an open question⁸, and the Turkey study measured short-term outcomes only². The students who need most help may benefit least: Nigerian gains rose with prior attainment and income¹, and low-engagement students in the US trials were those with lower performance⁵. Much of the effect depends on teachers and design. The Nigeria result came from a tutor plus teacher facilitation plus structured prompts, and the Turkey harm disappeared only with teacher-written content¹². The largest US trial found little added value from the AI itself³. A tutor-assist tool also helped weaker human tutors more, which suggests that the human setting matters⁷.

Stage: Pilot with results — why

The hypothesis holds. There are several randomised trials with measured gains, but they are short, mostly single-site and unreplicated, and the one at scale (Tennessee) shows modest effects driven by non-AI components. Estonia's rollout is planned, not yet measured. Student use of general chatbots is widespread (operator-reported as 64 to 90% of Estonian students before the programme)⁹, but that is unsupervised use, not tutoring programmes with evidence.

What to watch next

• 2 December 2027: EU high-risk obligations for education AI systems begin¹⁰.
• 2026/27 school year: Estonia's expansion to vocational schools, with University of Tartu researchers measuring long-term impact; no results date is published⁹.
• Redesigned Khanmigo: Khan Academy says it has redesigned the interface to push students to engage⁴. An independent trial of the new version would show whether engagement and gains improve.
• Nigeria replication or follow-up: a repeat trial or a follow-up on durability would address the gaps named above⁸.

Questions to ask before adopting

• Does the vendor's evidence come from a randomised trial with a control group that received equal time?
• Is the tutor set to give hints, and is its content checked for accuracy against our curriculum?
• Who will supervise sessions, and what training and time does that need?
• How will we measure real use by the weakest students, not just access?
• What child data does the tool collect, where is it processed, and does it meet our legal rules?