Using AI to Prepare Lessons: What to Delegate and What to Keep

Lesson preparation has always been the invisible half of teaching: the worksheets, examples, explanations, and extension tasks assembled long before any student walks in. AI assistants are genuinely useful here — not because they teach, but because they draft. The teachers getting real value from them have worked out a clear division of labour: the AI produces raw material quickly, and the human decides what is true, what is appropriate, and what fits this particular class.

What AI does well in preparation

The strongest use cases share a pattern: tasks where volume and variation matter more than originality, and where you can verify the output faster than you could write it yourself.

• Drafting first versions: a lesson outline, a set of worked examples, a homework sheet, or a plain-language explanation of a concept. A rough draft in thirty seconds beats a blank page.
• Differentiation: ask for the same worksheet at three difficulty levels, or the same explanation rewritten for a struggling reader. Producing variants is exactly the kind of mechanical rewriting AI handles well.
• Fresh contexts for old skills: ten percentage-change problems set in hawker stalls, MRT fares, and phone plans instead of the textbook's usual scenarios.
• Anticipating misconceptions: asking what errors students typically make with a topic often surfaces genuinely useful distractors and discussion points.
• Administrative writing: parent update messages, lesson summaries, and topic overviews that need to be clear rather than clever.

What to keep human

Some parts of preparation cannot be delegated, because they depend on knowledge the AI does not have: your students. An assistant has never seen Wei Ling freeze on word problems or noticed that the back row stops listening after twenty minutes. Sequencing decisions — what this class needs next, given what happened last lesson — are judgements about real people, not about content.

Accuracy is the other boundary. AI-generated material is fluent and confident regardless of whether it is correct, and errors in teaching materials are worse than errors elsewhere because students absorb them trustingly. Every fact, every worked solution, and every syllabus claim needs a human check before it reaches a student. In Singapore this matters doubly: AI tools trained largely on overseas material will happily produce questions in the wrong format for local exams, use unfamiliar terminology, or miss what the current syllabus actually requires.

A workflow that holds up

1. Specify tightly: state the level, syllabus, topic, question count, and format. Vague requests produce generic material that takes longer to fix than to specify.
2. Generate more than you need — twelve questions when you want eight — so you can discard weak items instead of repairing them.
3. Verify everything: work through solutions yourself, check facts against a trusted source, and confirm terminology matches what students see in class.
4. Adapt: swap contexts, adjust numbers, insert the specific error your class made last week. This is where the material becomes yours.
5. Bank what survives. A verified question or explanation is an asset; over a term, the bank compounds and preparation time genuinely falls.

Two honest examples

A secondary maths tutor preparing a lesson on simultaneous equations asks for eight word problems in Singaporean contexts at two difficulty levels, then solves all eight herself. Two have awkward numbers and one has an ambiguous setup; she fixes the numbers, discards the ambiguous one, and has a usable worksheet in fifteen minutes instead of forty-five.

An English teacher asks for a model discursive essay to critique with his class. Rather than presenting it as exemplary, he uses it as it is — competent but flat — and has students identify where the argument is thin. The AI's mediocrity becomes the teaching point. That is perhaps the healthiest framing of all: AI output as raw material to think with, never as a finished product to hand over.

The time savings are real — many teachers report saving several hours a week on drafting. But the saving only holds if verification stays in the loop. Skip it once, teach a hallucinated fact, and you will spend far longer repairing the damage than you ever saved.