The obvious way to use AI for lesson planning is to ask it for a lesson. It is also the way that produces the least useful result, because it hands the model the one part of the job it is worst at, deciding what the lesson should achieve, while taking away the part it is best at.
Flip the workflow. Bring a lesson you already have, with set success criteria and learning outcomes and use AI on the parts of it you would improve if you had another hour.
Why "write me a lesson" underperforms
A lesson plan is a set of decisions: what these students need next, how long they'll actually take on the tricky bit, which misconception is worth pre-empting, what to cut when it runs over. Every one of those decisions depends on knowing the class and the curriculum position.
Asked to generate from nothing, the model has to invent all of it. What comes back is structurally plausible and pedagogically anonymous: the average lesson on that topic, with a starter, a main activity, a plenary, and no particular reason for any of it. You then spend longer fixing it than you would have spent writing it.
Don't ask AI what to teach. Ask it to pressure-test what you've decided to teach, and to build the pieces you've specified.
Five things worth handing over
1. Anticipating misconceptions
This is the strongest use we've seen. Give it your lesson and ask what students typically get wrong about this topic, and at which point in the sequence. The output is drawn from a very large amount of writing about teaching this concept, and it will frequently name something you had not planned for, particularly outside your specialism.
Here is my lesson on [topic] for [year level]. List the six misconceptions students most commonly hold about this, note where in my sequence each one is likely to surface and suggest one diagnostic question for each that would reveal it quickly.
2. Generating the variations you didn't have time for
More practice questions at the same difficulty. A second worked example with different numbers. Three alternative contexts for the same problem so it doesn't only make sense to students who play sport. This is mechanical production work against a specification you supply, which is exactly the right shape of task.
3. Explaining a concept several different ways
When a student doesn't get it, the useful thing is a genuinely different explanation rather than the same one repeated slowly. Asking for four explanations at different levels of abstraction (a concrete real-life analogy, a visual description, a formal definition, a worked case) gives you options to reach for in the moment.
4. Stress-testing your sequence
Ask what knowledge the lesson assumes that it doesn't teach. Models are reasonable at spotting the step you skipped because it is obvious to you, which is precisely the step that is not obvious to a student meeting it for the first time.
5. The writing you resent doing
Rubric wording. Success criteria in student-facing language. The instructions paragraph. The email to parents about the assessment. None of this is where your professional judgement adds value, and all of it takes time.
Ask whether you would recognise a bad answer instantly. Misconceptions, extra practice questions, alternative explanations: you can judge these on sight. Anything where a wrong answer would look the same as a right one is a task to keep.
A workflow that fits a real planning session
- Plan the lesson yourself, roughly. Objective, sequence, the one thing they must leave understanding. Ten minutes, on paper, no tools.
- Paste it in with your class context. Prior knowledge, timing, constraints, what you're worried about.
- Ask for the critique, not a rewrite. "What does this assume? Where will it run over? What misconception have I not planned for?"
- Decide what to act on. Some of it will be right, some will be generic advice you can ignore. This judgement is the job.
- Then ask it to build the specific pieces you've decided you need: the extra questions, the diagnostic, the success criteria.
- Read everything before it reaches a student.
The order matters. Planning first means the lesson is yours and the AI is working on your material. Planning second means you spend the session evaluating someone else's generic lesson, which is both slower and worse.
The honest limits
It will suggest activities that sound good and don't work with thirty students in a room with fixed tables. It has no idea that this class cannot be given scissors. It will confidently reference a curriculum outcome that doesn't exist, or attach the right content to the wrong year level. Check anything it claims about the syllabus against the syllabus.
And it will produce a great deal of plausible pedagogical language that means very little. "Encourage students to engage critically with the material" is not a suggestion; it is a sentence shaped like one. Learning to skim past that is part of using the tool well.
Used this way, the realistic gain is not a finished lesson in ten seconds. It is a lesson that is a bit better than the one you would have taught, produced in about the same time, because the thinking stayed with you and the typing went somewhere else.