The rubber duck and the AI student

Recommended for: Any course, early in the semester; also works as a staff development session

Background for the instructor

Students mostly hold one mental model of a chatbot: it is a better Google, a machine you ask for answers. That model is the problem. A fluent, well-organised answer produces the same feeling as a clear lecture — the sense of having understood — while leaving nothing behind. The illusion of mastery is the risk, not the inaccuracy.

The remedy is not a warning but a replacement. Three other models put the student in the position of doing the thinking, and all three are things a chatbot happens to be good at.

Procedure

1. Take one concept from your own course that students find hard.
2. Rubber duck. Students explain the concept to the chatbot, told only to listen and ask what is unclear. The value is in having to articulate it, not in the reply.
3. The AI as your student. Students instruct the model to act as a learner who does not know the concept, teach it, and then ask it to say what it now does and does not understand. Where the explanation had holes, they show up here.
4. The question generator. Students have the model produce practice questions on the material and answer them without looking anything up, then check.
5. The answer machine, for contrast. Students simply ask for an explanation and read it. Then compare: which one felt best, and which one left them able to do something?
6. Close on the gap between the two answers to that question.

Discussion questions

• Which pattern taught you the most, and how could you tell at the time?
• When did the answer feel better than your understanding actually was?
• Which of these could you do without a chatbot at all — and is that an argument against the tool or for the pattern?

Assessment opportunity

Students hand in the transcript of the AI-as-student round together with a note on what it exposed in their own explanation.

Extension options

As a first step for teaching staff: spend fifteen minutes talking to a chatbot about your own lesson plan, using the rubber-duck pattern. It is the cheapest way to get a feel for what these tools are and are not good for.

Reference

• The Norton Guide to AI-Aware Teaching (Vee, Watkins & Bruff, W. W. Norton, 2026), chapter 13, "Using AI as a Tutor"; the framing here follows Derek Bruff's account of the chapter.
• "AI is not Google" is Leon Furze's formulation.
• The second pattern is a form of the protégé effect: people learn more when preparing to teach.