On this page
Three friends, one laptop, a kinematics test tomorrow. The easy move is to paste every question into a chatbot and nod along. The better move: give the AI one fixed seat, give each person a job, and let a human try every question first.
Why should the group explain before asking AI?
Talking a problem through with friends does real work. In an undergraduate genetics course, Smith and colleagues found, in a 2009 Science paper, that peer discussion improved understanding even when none of the students in a discussion group originally knew the correct answer. They tested this by giving each student a second, similar question to answer alone after the discussion.
Explaining out loud helps the person explaining too. In a 1994 study by Chi and colleagues, 14 eighth-grade students who were asked to explain to themselves after reading each line of a text on the human circulatory system gained more from pre-test to post-test than 10 students who simply read the text twice.
The risk runs the other way when AI goes first. In a field experiment with nearly a thousand high school maths students, Bastani and colleagues reported in PNAS that students using a standard ChatGPT-style tutor got 48% higher grades on practice problems, but 17% lower grades than students who never had access once the AI was taken away. A version with prompts designed to safeguard learning largely avoided that drop. We cover more of this in our look at whether leaning on AI weakens learning.
What roles should each person take?
For a group of three to five, use these four roles and rotate them after every question, so everyone explains at least once a session.
Explainer: attempts the question aloud, step by step, before anyone else speaks.
Questioner: asks "why that step?", "what if the mass doubled?" and "where did that sign come from?" The job is to find the weak step, not to be polite.
Checker: compares the explanation with the textbook or NCERT, and later tests the AI's answer with units, limiting cases and a rough estimate.
Keeper: holds the laptop and the dispute list. Only the keeper types to the AI, and only questions the group has agreed on.
With three people, the keeper also acts as checker.
What does one round look like?
Here is a ten-minute round on a Class 11 physics question: a ball is thrown straight up, so what are its velocity and acceleration at the very top?
Explainer (3 minutes): "Velocity is zero at the top, so acceleration must be zero too."
Questioner (2 minutes): "If acceleration is zero at the top, why doesn't the ball just stay there?"
The group can't agree. The keeper writes the dispute in one line: "Is acceleration zero at the top of the throw?"
Checker (1 minute): reads the textbook section on motion under gravity aloud.
Keeper types a narrow prompt the group wrote together: "Don't give the answer yet. Ask us one question that tests whether acceleration can be non-zero when velocity is zero."
Everyone answers the AI's question on paper, then compares. The explainer re-explains the whole thing in their own words.
The prompt in step 5 matters. Asking the AI for a hint or a question keeps the thinking inside the group; asking for the answer hands it over.
How do you keep a list of disputes to settle?
The dispute list is one page. For each item, write the disagreement in one sentence, what each side thinks, and how it was settled: textbook, AI, teacher, or still open. Settle in that order, book first, then AI, then a teacher for anything left.
AI answers need checking too. Chatbots can slip on arithmetic and signs, so the checker should run the five-step check for AI maths mistakes before any item is marked settled. If the dispute is about method rather than concept, a worked solution helps: the Fermi Problem Library has worked problems with a step-by-step solver that explains the reasoning rather than printing an answer, so the checker can see which step the group got wrong.
Open items become the first thing the group explains next time.
What should the group avoid?
Asking the AI to "solve this" before anyone has tried.
Letting the fastest student always explain.
Marking a dispute settled because the AI sounded confident.
Splitting up so each person asks the AI separately. One shared screen keeps the conversation in the group.
To close the session, each person writes two lines from memory on what they learned, then tries five quick ways to check you understood an AI explanation. For study between meetings, Fermi's AI tutor works out what a student already knows and builds a plan for the session, so you can bring a sharper question to the next group.
