Learn from an example, then try the next problem yourself
Use a worked example to understand a method, then apply it to a new problem. Check your reasoning before asking AI for further help.
Learn one, transfer many needs a gap between study and proof. Unrestricted AI chat collapses that gap: it returns the solution to the exact item in front of you. The corner is already painted. There is nothing left to bring back.
Field evidence on ChatGPT-like tutors fits that picture. Assisted practice can look strong while unassisted exam scores drop; hint-only tutors largely blunt the harm. The product lesson is not “ban AI.” Let AI teach on a twin or nudge with a hint — not finish your sit.
Where it usually goes wrong
The model solves your exact stem and you transcribe — almost no encoding. Or the explanation is so fluent you feel mastery without ever retrieving. Or the model keeps score for you, and your own judgment of learning gets soft.
AI that still leaves you the transfer
The Example drawer invents a parallel worked problem and keeps your keyed answer hidden (how it works). Rubi hints are Socratic nudges mid-working — not the final value. Rubric feedback marks after you produce writing or a sit, then queues drill. In each case you close the scaffold and produce on a fresh stem.
Two honest questions
- Can I solve a twin with different numbers with the chat closed?
- Would my score hold if the model vanished tomorrow?
If either answer is no, you practised assisted fluency. That is not learn one, transfer many.