AI in education: what the evidence says, and what K-AI should learn from it

By Kinlo Tangiri, Founder, KET Foundation · 1 November 2026 · 7 min read

KET runs an AI assistant for students, called K-AI. Before we build more on it, we want to know what the research says about AI in learning. This post reads the recent evidence, asks what it means for students in Cameroon, and sets out three things K-AI should learn from it.

The short answer: AI can help students learn, but only when it is designed to teach. Used without guardrails, it can make learning worse.

What the evidence says

AI tutors built to teach have helped

Three recent studies found gains when students used an AI tutor that was set up for learning.

In Edo State, Nigeria, a randomised trial gave students an AI tutor based on GPT-4 for six weeks in June and July 2024. Students in the programme did better than a comparison group, by 0.31 standard deviations overall and 0.23 in English (De Simone et al., 2025). A standard deviation is a common unit in education research for comparing results across studies. The World Bank team that ran the trial also names its limits: attendance was uneven, the long-term effects are not yet known, and the role of teachers is still an open question (World Bank, 2025).

In Ghana, about 1,000 students in grades 3 to 9 across 11 schools used Rori, a maths tutor that runs on WhatsApp. They had two 30-minute sessions a week for eight months. The study reports an effect size of 0.37 (Henkel et al., 2024). Rori works on basic phones and with low bandwidth. The results cover the first year only, and the paper is a preprint, so it has not yet been peer-reviewed.

At Harvard, 194 physics students took part in a crossover trial. Each student learned some topics with an AI tutor and others in an active-learning class. The median learning gains with the AI tutor were more than double (Kestin et al., 2025). The tutor was built with careful, custom prompts. The authors do not argue that AI should replace teaching. This was also an elite university setting, far from most classrooms.

Open access without guardrails can harm learning

A study with about 1,000 high school students in Turkey tested two versions of a chatbot in maths. One was a plain version of GPT-4 ("GPT Base"). The other, "GPT Tutor", had safeguards built in to guide students rather than hand them answers (Bastani et al., 2025).

While students had access, both groups did better on practice work. Grades rose by 48% with GPT Base and by 127% with GPT Tutor. The surprise came after access was removed. Students who had used GPT Base then scored 17% lower than students who never had access at all. The safeguards in GPT Tutor largely prevented that harm. The authors conclude that "unfettered access to GPT-4 can harm educational outcomes" (Bastani et al., 2025).

The lesson is simple. When a tool does the thinking for the student, practice scores go up and learning goes down.

The conditions around the tool matter

UNESCO's guidance on generative AI calls for "a human-centred vision of these new technologies". It asks for regulation, protection of data privacy, and an age limit for independent use (UNESCO, 2023).

UNESCO's Global Education Monitoring Report on technology warns that "many solutions proposed may also be detrimental". It names access, governance and teacher preparation as the conditions that decide whether technology helps (UNESCO GEM Report, 2023).

What it means for low bandwidth and shared phones

Connectivity shapes what any tool can do. DataReportal estimates that Cameroon had 12.6 million internet users at the end of 2025, or 41.9% of the population. It also estimates 29.0 million mobile connections, equal to 96.4% of the population, though some of these are voice and SMS only (DataReportal, 2026).

Read together, these estimates suggest that a phone is far more common than a data connection. A tool that needs a constant, fast connection will miss many students. A tool that works in short sessions, on a basic phone, on an app students already use, has a better chance. That is one reason the Ghana study matters to us: Rori ran on WhatsApp, with low bandwidth, and still showed gains (Henkel et al., 2024).

We should be careful not to stretch these results. The students in Nigeria and Ghana are not GCE candidates. Their ages, subjects and schools differ from ours. The studies show what is possible. They do not tell us what K-AI will achieve.

What K-AI does today

K-AI is four AI assistants that know the GCE syllabus:

• Study Partner: a peer-style tutor that breaks a topic into parts and asks follow-up questions.
• Academic Counselor: study techniques, careers, time management and goals.
• Virtual Ambassador: KET's programmes, resources and events.
• Scholarship & College Assistant: scholarships, admissions, financial aid and application documents.

K-AI also explains exam questions. After a mock exam, a student can ask for an explanation of any question they got wrong or left blank. The explanation streams in on demand. It is kept under 200 words, written to be encouraging, and may end with a short memory tip. A question in French gets an answer in French. The explanations cover GCE O Level and A Level, BEPC, Probatoire and Baccalauréat.

KET has not yet published an evaluation of K-AI. We do not yet know how accurate its answers are, or whether it improves results. We make no claim that it does.

Three things K-AI should learn from the evidence

1. Hint first, answer second

The Turkey study is the clearest warning. A tool that gives answers can raise practice scores while weakening real learning (Bastani et al., 2025). The Study Partner already asks follow-up questions. The exam explanations, by design, explain the right answer once a student asks. A next step to test is a hint mode: a first nudge, then a second, and only then the full explanation. We would test this before making it the default.

2. Build for the phone students actually have

The Ghana results came from short sessions on WhatsApp, on basic devices (Henkel et al., 2024). K-AI's answers are already kept short. KET also offers learning on WhatsApp, which is light on data. The research suggests we should keep pushing in that direction: short, text-first help that works on a shared phone and a weak connection.

3. Measure before we claim, and keep teachers in the loop

Every positive study above measured learning against a comparison group. None asked people to take the tool on trust. The Harvard authors do not argue for replacing teaching (Kestin et al., 2025), and the Nigeria team calls the teacher's role an open question (World Bank, 2025). UNESCO puts teacher preparation among the conditions for success (UNESCO GEM Report, 2023).

For K-AI, that means three steps. First, check a sample of its explanations against the Board's syllabus with subject teachers. Second, compare students who use it with students who do not, on the same tests. Third, publish what we find, including what does not work.

What comes next

This post is the start of KET's research on AI in learning. Next, we plan to write in plain words about what K-AI can and cannot do, and about learning with low bandwidth and shared phones. If you research AI tutoring in African classrooms and would like to compare notes, we would like to hear from you through ket.academy. New posts and study tips also go out on the KET Foundation WhatsApp channel.

All posts