A learner receives help from a chatbot and completes a difficult exercise. The response is clear. The next teaching question concerns what the learner can now do when they meet a similar problem, perhaps a week later, with less assistance.
The OECD’s Digital Education Outlook 2026 reviews emerging research in which stronger work produced with general-purpose AI does not always translate into better performance once learners lose access to the tool. It also describes promising results from uses designed around educational aims. The evidence concerns varied tools and settings; it supports examining the design of each learning activity. OECD Digital Education Outlook 2026.
Decide what the learner needs to practise
Our reading is that a teacher should begin by identifying the thinking the activity is meant to develop. That decision makes it easier to judge where AI assistance belongs. If the purpose is to learn how to construct an argument, handing the full argument to the learner changes the work they have an opportunity to practise.
A useful activity could ask the learner to make an initial attempt before seeking help. They might then request feedback on one part and explain which suggestion they choose to use. The explanation gives the teacher something to discuss with them and gives the learner a reason to consider the advice.
The amount of help should be deliberate. Some learners need a worked example before they can attempt a related problem. Others are ready for a small prompt. A teacher needs room to make that decision from the learner’s response, including the option to provide human guidance when the tool’s advice is confusing.
These are possible teaching approaches. Their effectiveness would need to be evaluated in the subject and age group where they are used.
Return to the idea after the help ends
A later activity can explore what the learner can use independently. The task should draw on the same underlying idea while changing enough of the context to require a fresh decision. Asking the learner to explain that decision can reveal where the next lesson should begin.
A short follow-up also helps the learner judge their own understanding. They can compare their explanation immediately after receiving help with what they can explain later. Where something remains difficult, the teacher can agree on further practice.
For a school evaluating an AI learning tool, this suggests building time into the pilot for those later observations. Staff should know how much assistance learners receive during the original activity and examine their subsequent work under clearly described conditions. Keeping that evidence alongside the original activity gives teachers a basis for deciding whether to change the support offered in the next lesson.