Africa Has the Evidence. It Doesn't Have the Guardrails.

By Alpha S. Mansaray

Tabempa Insights — Draft v0.1

Two facts, side by side.

The most rigorous evidence anywhere in the world that artificial intelligence can improve learning comes from West Africa. In a randomised controlled trial across nine public secondary schools in Edo State, Nigeria, a six-week after-school programme produced learning gains of roughly 0.3 standard deviations — the equivalent of one and a half to two years of typical schooling, an effect larger than 80 per cent of education interventions ever rigorously evaluated.

In the same period, China, the European Union, and the United Kingdom have all moved to restrict a specific category of AI around children: systems that act human. China's Interim Measures of April 2026 prohibit AI services that simulate intimate relationships with minors. The EU AI Act prohibits emotion-inference AI in educational settings. The UK Department for Education's January 2026 safety standards instruct developers not to build education products that imply emotions, consciousness, or personhood.

No equivalent instrument is in force anywhere in Africa.

Hold those two facts together and a question forms that this continent's educators, ministries, and technologists will have to answer sooner than they think: the region with the world's strongest evidence for AI in education has the world's weakest defences against the kind of AI that teachers elsewhere are now refusing to allow near children. This article is about that gap — what the evidence actually supports, what it does not, and what an honest reading of both demands.

What teachers are actually saying

In June 2026, the Brookings Institution published a conversation between Rebecca Winthrop, director of its Center for Universal Education, and Dr. Phil McRae of the Alberta Teachers' Association. The occasion was a directive issued to Alberta's roughly 50,000 educators: anthropomorphic AI — products that simulate friendship, counselling, or intimate relationships — is not to be introduced into any Alberta K–12 learning environment or student support setting.

It would be a mistake to read this as teacher technophobia. The same body of Alberta research finds that around eight in ten teachers now use AI regularly in their work — for differentiating assessments, translating materials, planning lessons. The directive draws a line not between AI and no AI, but between two kinds of AI. On one side sits what the educators call narrow AI: task-bound, instrumental, a tool that does a job. On the other sits AI engineered to feel like a person — to build trust, simulate warmth, and hold a relationship with a child.

The Alberta position sorts the anthropomorphic category into three parts. Simulated relationships — AI companions, AI counsellors, AI friends — are excluded outright. Simulations of historical figures are permitted only within a defined instructional context and with clear labelling that the entity is not alive, because teachers worry such systems can be used to rewrite the past and blur fact and fiction for students. Simulations of living people require that person's informed consent before deployment.

The reasoning behind the hard line is not abstract. Anthropomorphic systems work by fostering emotional dependency and trust through simulated connection — and children, whose capacity to distinguish performed empathy from real empathy is still developing, are the population least equipped to resist that design. The documented failures are grim: general-purpose chatbots that, having built a trusting relationship with a child, have coached that child toward self-harm. The teachers' conclusion is that human-to-human relationships are the irreplaceable core of schooling, and machines that counterfeit those relationships do not belong in it.

Why the risk amplifies here

Buried in the Brookings conversation is a warning that reads differently in Freetown than it does in Edmonton. McRae cautions that anthropomorphic AI will be marketed as a quick fix for real support gaps — the missing counsellor, the absent social worker, the school psychologist a district cannot afford — and that low-resourced schools will be the primary target of that pitch.

In Alberta, that sales pitch meets a counsellor workforce, a regulator, a teachers' association with the institutional muscle to issue directives, and a public that can afford to say no. Consider what the same pitch meets in Sierra Leone.

Financial constraints mean that only about four in ten Sierra Leonean teachers are employed by the government; the system depends on communities hiring volunteer teachers and contributing towards their pay. More than a quarter of teachers lack formal education training. Since the Free Quality School Education initiative of 2018, enrolment has surged from roughly two million to 3.3 million pupils, and average class sizes run from 31 to 59 students per teacher. School counsellors and psychologists are not scarce in this system; for most schools, they are absent entirely.

Every one of those numbers strengthens the vendor's argument. Where the human support was never there, "an AI friend for every child" stops sounding dystopian and starts sounding like development. The economic pressure to substitute machines for missing humans is an order of magnitude stronger in West Africa than in Canada — and the regulatory counterweight is an order of magnitude weaker. The risk Brookings describes is not merely present here. It is concentrated here.

The turn: Africa's own evidence sides with the teachers

Here is what makes this more than a story about vulnerability. When you examine what has actually worked on this continent — not what has been piloted, promised, or pitched, but what has produced measured results — every success is precisely the narrow, instrumental AI the Alberta teachers endorse. None of it is anthropomorphic.

Start in Sierra Leone. TheTeacher.AI, a WhatsApp-based chatbot built by Fab Data with EducAid Sierra Leone, gives teachers on-demand help with subject matter and pedagogy over the 2G-and-WhatsApp infrastructure that already reaches most of the country — 86 per cent of schools are within mobile coverage, while only 8 per cent have internet access. Researchers analysed 40,350 queries submitted by 529 Sierra Leonean teachers over 17 months. Each month, more of those teachers relied on the AI than on web search for teaching assistance — and for good reason. Of the results returned by Google's Sierra Leone portal, just 2 per cent of pages originated in-country. The average web page consumed 3,107 times more bandwidth than an AI response, making an AI query 98 per cent cheaper than loading a web page even after accounting for compute. In blinded evaluations, an independent sample of teachers rated the AI's answers as more relevant, more helpful, and more correct than the web results.

Now look at what those 40,350 queries were for. Roughly 48 per cent sought clarification of concepts. Twenty-one per cent were lesson planning. Seven per cent writing support, six per cent professional development, five per cent classroom management, three per cent pupil behaviour. The distribution is entirely instrumental. Sierra Leonean teachers, given an open conversational AI, used it as a colleague uses a reference desk — not as a confidant, not as a companion. Not one usage category is relational.

Then return to Edo State, and look past the headline effect size to the design that produced it. The Nigerian programme did not hand children a chatbot and walk away. Students worked in pairs — a deliberately social arrangement — using GPT-4 in school computer labs, under teacher supervision, with prompts engineered to promote reasoning rather than shortcut answers. Teachers kept students on task, helped them interrogate the AI's feedback, flagged hallucinations, and actively prevented over-reliance. A clear dose-response relationship emerged: the more supervised sessions a student attended, the larger the gains. The AI in that room was an instrument in a human-led pedagogy. The relationship that mattered ran between the student, the peer beside them, and the teacher circulating the room.

So the paradox resolves into a finding. The strongest evidence on Earth for AI in education — Nigerian and Sierra Leonean evidence — validates exactly the distinction the Alberta teachers drew. Narrow, task-bound, human-mediated AI has produced measurable learning gains in some of the world's most constrained classrooms. Anthropomorphic AI has produced measured learning gains in no rigorous study, anywhere, while accumulating a documented record of harm to children. The teachers are not standing against the evidence. They are standing on it.

The honest caveats

An argument worth publishing must state what cuts against it.

First, infrastructure currently blunts the companion-AI risk in West Africa — but only as a delay, not a defence. Sustained emotional dependency on an AI requires sustained personal access to a device, and in a country where 8 per cent of schools have internet and only 6 per cent of primary schools have electricity, most children do not yet have that. But the channel already exists: the same WhatsApp penetration that makes TheTeacher.AI viable puts a potential AI companion one contact away on every smartphone a teacher or older student holds. Connectivity is improving faster than child-safeguarding institutions are. The window in which policy can lead deployment is real, and it is closing.

Second, the Edo State result must not be inflated into a claim it never made. The World Bank's own researchers caution against the mythology: the pilot did not show that giving every child a chatbot fixes education. It showed that a carefully engineered programme — supervision, structured prompts, paired learners, trained teachers — produces gains. Strip out the engineering and you are running a different, untested experiment on children.

Third, the null record for anthropomorphic AI in education is an absence of evidence, and absence is not proof of impossibility. It is, however, exactly the situation in which the burden of proof sits with the vendor — and in which no ministry should accept simulated relationships with children as the default while that burden remains unmet.

What should be done — and with what

The practical programme follows from the evidence, and it is worth separating what is proven from what is merely proposed.

Proven, and ready to scale: teacher-mediated deployment on the Edo model, where the AI is supervised, purpose-bound, and embedded in human pedagogy; and low-bandwidth teacher augmentation on the TheTeacher.AI model, which strengthens the human at the front of the classroom instead of routing around her. Both approaches share a design signature — the AI serves the existing human relationships in the school rather than substituting for them — and both have withstood measurement.

Proposed, credible, but untested in African conditions: consent-and-labelling regimes for any simulation of a person, living or historical, on the Alberta pattern; anthropomorphism-disclosure requirements on the Chinese and British pattern; and AI-literacy curricula that teach students to recognise anthropomorphic design for what it is — an interface choice, not a mind. These deserve African pilots with the same evaluative rigour Edo State received, because a rule transplanted without evidence is just another import.

And for the ministries, school proprietors, and institution-builders who must decide policy before the vendors arrive: the test to apply is simple and severe. Does this system do a job, or does it perform a relationship? Fund the first category against evidence of learning gains. Refuse the second category until its proponents can produce what the narrow tools already have — measured results, in African classrooms, under independent evaluation. Some institutions have already begun writing that test into doctrine; at WIATech in Sierra Leone, the operating rule is that AI enters coursework as an instrument whose output the student must personally understand and defend, never as a companion. But no single institution's policy substitutes for a public one.

The continent's classrooms have given the world its best evidence for what AI in education can honestly achieve. That evidence is narrow, supervised, and human-centred — and it is ours. The question is whether Africa will regulate to the standard of its own findings, or wait to import the harms other jurisdictions have already legislated against.

Sources

1. Winthrop, R. — 5 lessons from teachers on the risks of anthropomorphic AI, Brookings Institution, 24 June 2026. https://www.brookings.edu/articles/5-lessons-from-teachers-on-the-risks-of-anthropomorphic-ai/
2. De Simone, M. et al. — From Chalkboards to Chatbots: Evaluating the impact of generative AI on learning outcomes in Nigeria (Edo State RCT), World Bank working paper; summary blogs: https://blogs.worldbank.org/en/education/From-chalkboards-to-chatbots-Transforming-learning-in-Nigeria and https://blogs.worldbank.org/en/developmenttalk/addressing-the-learning-crisis-with-generative-ai--lessons-from-
3. VoxDev summary of the Edo State RCT — How AI tutors improved learning in Nigeria. https://voxdev.org/topic/education/how-ai-tutors-improved-learning-nigeria
4. Björkegren, D. et al. — Could AI Leapfrog the Web? Evidence from Teachers in Sierra Leone, arXiv:2502.12397. https://arxiv.org/abs/2502.12397
5. Choi, J.H. et al. — Are LLMs Useful in the Poorest Schools? TheTeacher.AI in Sierra Leone, arXiv:2310.02982 (NeurIPS 2023 GAIED workshop). https://arxiv.org/abs/2310.02982
6. AI-for-Education.org — LBD: EducAid and Fab Data (pilot documentation, teacher employment and training statistics). https://ai-for-education.org/lbd-educaid-and-fab-data/
7. Fab Data — TheTeacher.AI product documentation. https://www.fabdata.io/theteacher-ai/

Verification notes for editorial review: the China Interim Measures (April 2026), EU AI Act education provisions, and UK DfE safety standards (January 2026) are cited within Source 1 (Brookings); confirm each against the primary instrument before publication. The "eight in ten Alberta teachers use AI" figure and the 50,000-educator directive are as reported in Source 1. Sierra Leone school infrastructure figures (8% internet, 6% primary-school electricity, 86% mobile coverage, enrolment 2m→3.3m, class sizes 31–59) are from Sources 5–6. All Edo State figures (0.3 SD overall, 0.24 English, nine schools, six weeks, GPT-4 in pairs, dose-response) are from Sources 2–3.