In March 2024 the Education Minister announced that primary school pupils would be taught the basics of artificial intelligence from 2027, as part of the revised school curriculum. Digital fluency is one of the competencies the new curriculum is built around. The ministry has since described work on an AI-enabled digital learning platform and pilots of AI-supported classrooms in selected urban and rural schools.
Parent and teacher groups have urged caution, particularly at lower primary level, and commentators keep returning to the same two conditions for success: infrastructure and teacher training.
There is a third condition that gets less attention. Content.
The content problem
Every subject, every level and every chapter needs practice questions, worked answers and revision material. It has to match the syllabus exactly. It has to exist in English and in Bahasa Melayu. For mathematics and science it needs equations and diagrams. And it reaches students directly, so a wrong answer key is a wrong answer taught to thousands of children.
Ask a general-purpose AI model for twenty Form 2 science questions and you will get twenty plausible questions in seconds. Some will be off-syllabus. Some will have two defensible answers. One or two will be confidently wrong. A teacher who has to check all twenty carefully has saved very little time.
The value of AI in education content is not generation. It is generation plus a review process cheap enough that experts can actually do it.
How a production pipeline is put together
We built a content platform for Tavis, an online tuition provider for SPM students. It is in production, producing bilingual questions and learning cards. The design has four stages, and the model is the least interesting part of each.
1. Ground it in the syllabus. Subject teams upload the actual textbooks and past papers. The system extracts chapters, existing questions and figures into a knowledge base. Generation draws on that material, tagged by subject, level, chapter and subtopic. The model is not asked what it knows about photosynthesis. It is asked to write questions from this chapter.
2. Review before a person sees it. Each generated item goes through an AI review stage that casts one or more independent votes on it and proposes corrections. The number of votes is configurable, and a strict majority must judge the item good for it to pass. Anything else is routed to an editor.
3. Show corrections as a diff. When a review pass suggests a change, the editor sees the old and new wording side by side, word by word. Nothing is changed silently. Approving takes seconds because the editor is checking a specific change, not rereading the whole item.
4. Publish deliberately. Approval and publishing are separate steps. Only an explicit push sends content to the live learning platform.
Around this sit the unglamorous parts that decide whether the system is trusted: an equation editor, figure redraws with an automated check against the source figure, a full history of who approved what, and prompts that the content team can tune per subject without waiting for a developer.
What this means for schools and education companies
Do not measure AI by how fast it writes. Measure how many minutes of expert time each approved item costs. That number is what determines whether the content gets made.
Keep the syllabus in the loop. Material generated without the source text will drift from it. Upload the text.
Make review the product. A good review screen is worth more than a better model.
Keep a person accountable for what is published. The curriculum is about to teach children what AI is. The content they learn from should show them how it is used responsibly: as a tool that drafts, with a human who decides.