AI in Education: What Actually Works (And What Does Not)

AI in education is a crowded space. ChatGPT can write lesson plans. Claude can explain concepts. Gemini can generate quizzes. There are dozens of AI-powered education tools, each promising to transform teaching.

We tested 12 of the most widely used AI tools in the Nigerian education context. The tools fell into three categories: general-purpose chatbots (ChatGPT, Claude, Gemini, Meta AI), education-specific platforms (Khanmigo, MagicSchool AI, Eduaide.AI, Quillionz), and content generation tools (Tome, Notion AI, Canva AI, Jasper). We asked each to generate content aligned with the NERDC curriculum, create assessments matching WAEC patterns, and explain concepts at the appropriate level for Nigerian students.

Here is what we found.

What AI Does Well

1. First-draft generation: AI is excellent at producing a first draft that an educator can refine. A lesson plan generated by AI might be 70% useful. The structure is right, the general content is appropriate, but the specifics need adjustment for the curriculum, the students, and the context. Starting from 70% is much faster than starting from 0%.

1. Content explanation: AI can explain concepts at different levels of complexity. Ask it to explain photosynthesis at the JSS1 level, and it produces a simpler explanation than at the SSS3 level. This is genuinely useful for differentiated instruction.

1. Assessment question generation: AI can produce assessment questions in bulk. For drill-and-practice exercises (not high-stakes assessments), AI-generated questions save significant time. The quality varies, but the volume advantage is real.

1. Translation and localization: AI can translate content between Nigerian languages with reasonable accuracy. More importantly, it can localize examples. Replacing "vacuum cleaner" with "broom" in physics examples, or using Nigerian currency in economics problems.

What AI Does Poorly

1. Curriculum alignment: AI does not know the NERDC curriculum. When we asked ChatGPT to "create a lesson plan for JSS2 Mathematics on quadratic equations," it produced a reasonable lesson plan. But it included content from SSS1 (completing the square) and omitted JSS2 content (graphical representation of quadratic functions). The curriculum alignment was wrong.

1. Appropriate difficulty: AI struggles to calibrate difficulty for Nigerian examination standards. When asked to create "WAEC-standard Mathematics questions," it produced questions that were either too easy (below WAEC difficulty) or too complex (above WAEC difficulty). The sweet spot, the specific difficulty curve that WAEC uses, requires training data from actual WAEC papers, which AI does not have.

1. Cultural context: AI defaults to Western examples. A physics problem about "a car traveling at 60 mph" is not relevant to a student who has never seen a speedometer. An economics example about "stock market investment" is abstract for students whose families invest in agriculture or trade. AI can be prompted to localize, but it does not do so automatically.

1. Pedagogical sequencing: AI does not understand how topics build on each other in the Nigerian curriculum. It might introduce a concept that requires prerequisite knowledge the students have not yet covered. This is because AI does not understand the curriculum as a sequence. It understands it as a collection of topics.

1. Assessment validity: AI-generated assessments sometimes have wrong answers marked as correct, ambiguous questions, or distractors that are obviously incorrect (making the question too easy). For high-stakes assessment, AI-generated questions always need human review.

What AI Should Not Do

1. Replace teacher judgment: AI cannot determine what a specific group of students needs. A teacher who knows their students are struggling with algebra should not let AI decide to move on to geometry. AI should inform decisions, not make them.

1. Bypass curriculum review: AI-generated content should always be reviewed against the NERDC curriculum. The AI might include content that is not in the curriculum or omit content that is. This is not a failure of AI. It is a limitation of current technology.

1. Serve as the sole assessment: AI-generated assessments should be supplements, not replacements, for teacher-created assessments. The teacher understands what the students have actually learned. AI only understands what the curriculum says should be learned.

1. Make pedagogical decisions: Should a topic be taught over 2 weeks or 4? Should a concept be introduced through hands-on activity or direct instruction? These are pedagogical decisions that require professional judgment. AI can provide data to inform these decisions, but it should not make them.

How Curriva Uses AI

Based on this testing, we designed Curriva's AI features around three principles:

1. AI assists, never decides: AI generates content that the educator reviews and modifies. The final decision always rests with the educator. The UI makes this explicit. AI-generated content is visually distinct from educator-authored content.

1. AI is curriculum-grounded: Every AI generation references the NERDC curriculum data. When AI creates a lesson plan, it draws from the specific learning objectives, topics, and assessment criteria in our curriculum database. This does not eliminate alignment errors, but it reduces them significantly.

1. AI is transparent about sources: When AI generates assessment questions, it tags which curriculum objectives the questions address. When it suggests activities, it links to the curriculum's suggested activities. The educator can trace the AI's output back to its sources.

The Practical Impact

In our internal testing with 8 educators across 4 subjects, those using AI assistance (with the principles above) reduced their resource creation time by approximately 40% compared to creating the same resources without AI. The test was conducted over 2 weeks in August 2026, with each educator creating 3 resources with AI assistance and 3 without.

Not because AI did the work for them, but because AI handled the first-draft generation, and the educator spent their time on review and refinement instead of starting from scratch.

That reduction is significant. For a teacher spending 8 hours on a scheme of work, AI assistance brought the time down to approximately 5 hours in our test. For a teacher spending 4 hours on a lesson plan, it reduced to approximately 2.5 hours.

The time savings are real. But they are only possible if the AI output is good enough to refine rather than rewrite. And that requires grounding AI in the curriculum data that makes its output relevant to Nigerian educators.

This is why the curriculum extraction work matters. The structured data we extracted from the NERDC curriculum is not just a reference database. It is the foundation that makes AI assistance useful instead of misleading.

Related Reading

For the curriculum data that powers our AI features, see NERDC Curriculum by the Numbers. For how AI-generated assessments align with examination patterns, see WAEC and NECO Examination Patterns. For the engineering decisions behind low-latency AI responses, see Building for Low-Bandwidth.