AI in Education: Personalised Learning Without Losing the Teacher

In 1984, the educational psychologist Benjamin Bloom published what became known as the 2 Sigma Problem, reporting that students tutored one-to-one performed dramatically better than students taught in conventional classrooms. The challenge he posed was how to achieve similar results at a cost societies could afford. Later research has debated the exact size of the effect, but the core insight, that personalised, responsive instruction is enormously powerful, has held up.

In Nigeria and across Africa, the gap is stark. Classrooms are large, qualified teachers are stretched, and millions of young people want digital and technical skills faster than institutions can deliver them. Through BuildWithMMO, where I mentor developers, I see both sides: AI tools can make a motivated learner remarkably productive, and the same tools can let a learner skip the struggle that actually builds understanding.

Knowledge tracing and adaptive sequencing

Adaptive learning platforms model what a learner knows. Classic Bayesian Knowledge Tracing estimates the probability that a student has mastered each skill based on their answers. Deep Knowledge Tracing uses recurrent neural networks to model learning trajectories more flexibly. The system then chooses the next exercise to sit in the learner's zone of proximal development: hard enough to stretch, not so hard as to discourage.

Large language model tutors

Transformer-based language models can explain concepts in multiple ways, generate practice problems, give feedback on essays or code and answer follow-up questions in natural language. The most effective tutoring designs constrain the model with a system prompt and pedagogy: ask guiding questions instead of giving answers, check understanding, and encourage the learner to explain their reasoning. Retrieval-augmented generation ties responses to a specific curriculum or textbook so that the tutor teaches what the course actually covers.

Automated assessment and feedback

AI can grade short answers, provide rubric-based feedback on writing and run automated tests on code with natural language explanations of failures. For teachers, the gain is time; for learners, the gain is speed of feedback, which is one of the strongest predictors of improvement.

Speech and language technologies

Speech recognition and text-to-speech support reading practice, language learning and accessibility for learners with visual impairments or low literacy. Progress on African languages is uneven but improving, which matters for learners taught in a second language.

• Teacher copilots. Drafting lesson plans, generating differentiated worksheets and producing quiz banks saves teachers substantial preparation time while keeping them in control of content.
• Coding education. AI pair programmers help learners get unstuck, understand error messages and explore alternative solutions. In mentoring, I encourage learners to ask the AI why before they ask it what.
• Low-bandwidth delivery. Tutoring over SMS, WhatsApp and low-data apps reaches learners who cannot rely on broadband.
• Language support. Explaining a difficult concept in a learner's first language before reinforcing it in English can unlock understanding.

UNESCO's work on AI in education, including its guidance on generative AI, emphasises a human-centred approach in which AI supports rather than replaces teachers, a framing I strongly agree with.

• Outsourced thinking. If a model writes the essay or the code, the learner may get the grade without the skill. The productive struggle of working through a problem is where much learning happens.
• Confident errors. Language models can explain incorrect material fluently. Learners, by definition, may not know enough to spot the mistake.
• Assessment integrity. Take-home assignments are increasingly easy to complete with AI. AI-text detectors are unreliable and can unfairly flag non-native English writers, so they should not be the basis for disciplinary decisions.
• Privacy of minors. Educational platforms collect sensitive data about children. Data minimisation, parental consent and clear retention limits are essential.
• Equity. If only well-resourced schools can afford good AI tools and devices, AI widens rather than narrows gaps. Cultural relevance matters too; examples and references should reflect learners' contexts.

1. Design for learning, not answers. Configure AI tutors to scaffold: hint first, then partial solutions, and only then full explanations. Require learners to attempt problems before help unlocks.
2. Change assessment. Emphasise oral explanations, in-class work, projects with checkpoints and reflection on process. In developer mentoring, I ask learners to walk me through their code; understanding becomes obvious quickly.
3. Ground the content. Use RAG over vetted curriculum materials and show sources so learners and teachers can verify.
4. Keep teachers in the loop. Dashboards that show where learners struggle let teachers intervene where human attention matters most.
5. Build for constrained environments. Offline caching, small models and messaging interfaces extend reach.
6. Evaluate learning outcomes. Measure pre- and post-test gains and retention, not just engagement minutes.
7. Teach AI literacy. Learners should understand how these models work, where they fail and how to use them honestly. That is now a core skill in its own right.

AI will not fix education systems on its own; teachers, curricula, infrastructure and funding still matter most. But for the first time, a responsive, patient tutor that can explain an idea ten different ways at midnight is within reach of a learner with a basic phone. If we design these tools around how people actually learn, and we keep teachers and mentors at the centre, Bloom's challenge looks less like an impossible ideal and more like an engineering problem we can make real progress on.

For edtech founders in Nigeria, the most defensible advantage will rarely be the underlying model, which is increasingly a commodity available through APIs or open weights. It will be deep alignment with local curricula such as WAEC and NECO syllabuses, trusted relationships with schools and teachers, content in the languages learners actually think in, and evidence that learners improve. Those assets take years to build and are hard for global platforms to replicate quickly.

• Personalised tutoring is highly effective; AI makes some of its benefits affordable at scale.
• Knowledge tracing, LLM tutors with pedagogical constraints, and automated feedback are the core techniques.
• The biggest risk is outsourced thinking; design tools that scaffold rather than simply answer.
• Rethink assessment toward process, explanation and in-person demonstration of skill.
• Protect learner privacy, design for low-bandwidth contexts and measure real learning outcomes.

1. The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring — Educational Researcher
2. Deep Knowledge Tracing — arXiv
3. Artificial intelligence in education — UNESCO
4. OECD.AI Policy Observatory — OECD

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.