Question ladders
A single good question is useful. A sequence of them is what actually moves understanding, and the sequence has a shape: establish what is known, test whether it is understood, then push until it breaks.
The third rung is the one people skip. Deliberately looking for the case where your own explanation fails is uncomfortable and is where nearly all the value sits.
Working against yourself
Ask the question you would rather not: How would you check whether this AI response is actually correct? Then argue that case properly rather than dismissively.
In practice, across AI-assisted study and question scaffolding, this is what separates someone who has memorised a method from someone who knows when it does not apply.
Transfer, which is the real test
Understanding that only works in the form it was learned in is not understanding — it is pattern matching with good manners.
So the exercise is to take the idea into an unfamiliar context and see what survives. What breaks tells you which parts were scaffolding.
Self-assessment without kidding yourself
The honest check is whether you can teach the idea to someone who will interrupt with awkward questions, without notes.
The dishonest check is recognition: reading an explanation and feeling that it makes sense. That a fluent AI answer is a correct one is the same error wearing a different hat.
Practising deliberately
Set a standard before you start, not after: asking the model to argue the opposite case, then judging which argument holds. Reviewing against a standard you chose afterwards is not review.
You will know it is working when the learner prompts for reasoning and sources rather than conclusions.