A student can finish a fluent conversation with an AI and still be unable to explain the idea alone. The simplest way to expose that gap is to close the chat before the final answer. The conversation becomes practice; the independent response becomes evidence of learning.
OpenAI’s Study Mode is designed to ask questions, offer hints, break material into steps and check understanding. In an example published for educators, students explore evaporation through the tool, then write their own explanation and a discussion question without referring back to the conversation.[1]
A four-stage learning loop
Use the AI, then remove it
| Stage | Student action | Evidence produced |
|---|---|---|
| 1. Attempt | Write an initial explanation, prediction or solution before opening the chat | A baseline showing what the learner already understands |
| 2. Guided dialogue | Use Study Mode for one question at a time; ask for hints and test competing explanations | A record of reasoning and revisionSource 1 |
| 3. Closed-chat exit ticket | Close or hide the conversation, then explain the concept from memory | Independent recall and organisationSource 1 |
| 4. Transfer task | Solve a new problem that changes the surface details | Evidence that the learner can apply the idea rather than repeat it |
The workflow changes the prompt from “Give me the answer” to “Help me notice where my reasoning breaks.” That is not merely a polite way to use a chatbot. It creates observable stages at which a teacher—or the learner—can see whether understanding improved.
A practical prompt and exit ticket
After the dialogue, close the chat and answer three questions: What is the mechanism in my own words? Which example shows it? What mistake would produce the wrong answer? Then solve one altered problem. For evaporation, the transfer might change clothes on a line to a shallow dish of water on a windy day.
The research supports training—but not generic scepticism
In a controlled classroom study, 116 students aged 13 to 15 completed six science investigation tasks with a generative AI system. A two-hour workshop covering model behaviour, failure modes, prompting and evaluation led students to reformulate queries more often, ask more follow-up questions, judge correctness more accurately and produce better final answers than an untrained group.[2]
A separate randomised study of high-school students found that a brief text-based literacy intervention did not significantly reduce acceptance of incorrect ChatGPT recommendations. Students accepted wrong recommendations in 52.1% of trials, and the intervention increased rejection of correct advice. The lesson is not “trust less.” It is “practise distinguishing useful from unreliable help.”[3]
What to assess
- Can the learner explain the mechanism without seeing the AI conversation?
- Can the learner solve a transfer problem with different surface details?
- Can the learner identify what evidence would change the answer?
- Can the learner distinguish uncertainty from a confident but unsupported claim?
A transcript can show participation, but it is a weak measure of ownership. The closed-chat answer and transfer task make the learner carry the cognitive load at the moment that matters.
AI can extend the period of guided practice, especially when a teacher cannot sit beside every student. It becomes more educational when the session ends with independence. Close the chat, change the problem, and ask the learner to show what remains.
Sources & further reading
- Teaching students to think and learn through questions OpenAI · 15 Oct 2026 · Accessed 15 Oct 2026
- Teaching Students to Question the Machine arXiv · 2 Apr 2026 · Accessed 15 Oct 2026
- A Brief AI Literacy Intervention Does Not Significantly Reduce Over-Reliance and Increases Under-Reliance on ChatGPT arXiv · 18 Sep 2026 · Accessed 15 Oct 2026