On this page
When you know you will have to explain an idea, you study it differently: you sort main points from details and notice the steps you can't justify. An AI chatbot can play the person you teach, asking the awkward questions a classmate would, at 11 pm, without judging you. The trick is to make it confused in one specific way.
Does teaching someone else really help you learn?
Yes, and the benefit is larger when the teaching is a conversation. In a 2019 meta-analysis in Japanese Psychological Research, Keiichi Kobayashi pooled 28 studies that compared teaching with ordinary studying. Preparing to teach gave an effect size of 0.35, and preparing and then actually teaching gave 0.56. The gains showed up for deep understanding as well as surface facts, and they held up after a delay. Teaching helped more when it was interactive than when it went one way.
There is a catch. A review by Rod Roscoe and Michelene Chi found that peer tutors tend to show a "knowledge-telling" bias: even trained tutors focus on delivering knowledge rather than developing it. The review ties tutor learning to checking your own understanding and connecting ideas, which is what a good question pushes you to do.
Why give the AI a misconception?
Because a chatbot left to itself already knows the answer. In a 2024 preprint on teaching ChatGPT, posted on arXiv before peer review, Angxuan Chen and colleagues found that students who taught ChatGPT improved their knowledge and programming skill, but it had limited impact on their error-correction skills, likely because ChatGPT tends to write correct code. A student who is never wrong gives you nothing to fix.
A planted misconception fixes this. The AI defends a wrong idea real students hold until your explanation changes its mind. If you can't shift it, you have found a gap.
Computer students are not new. In the Betty's Brain studies, Catherine Chase and colleagues found that pupils made more effort to learn for a teachable agent than for themselves. They called this the protégé effect, and it helped lower-achieving students most.
How do I ask AI to act like a confused student?
Start a fresh chat and paste this:
"You are Riya, a Class [X] student who is confused about [topic]. You sincerely believe this wrong idea: [misconception]. Do not tell me the correct answer, and do not drop your wrong idea just because I say it is wrong. Change your mind only when my explanation answers your objection. Ask one short question at a time, the kind a real classmate would ask, including 'why?' and 'what about this case?'. If I use a term without explaining it, ask what it means. When you are convinced, say so, then ask me one harder question. At the end, list the questions I struggled to answer."
Then:
Pick one concept, not a whole chapter.
Take the misconception from your own past mistakes.
Explain with your notes closed.
Each time you stall or say "it just is", log the question.
Afterwards, check every logged question in your textbook.
What does a session look like?
A short run on Newton's first law, with the misconception "a moving object needs a force to keep moving".
Riya (AI): "When I stop pedalling, my cycle stops. So motion needs a push, right?"
You: "It slows because friction and air resistance act against it. With no net force, it would keep moving at the same velocity."
Riya: "But I've never seen anything move forever. How do you know?"
This is where you are likely to stall, because "the book says so" is not an answer. A stronger reply shrinks the opposing force step by step: a ball rolls further on a smooth floor than on carpet, and a puck slides further still on ice. Remove that force completely and nothing is left to stop it.
Riya: "Then why do I have to keep pedalling to go at a steady speed?"
You: "Your pedalling balances friction, so the net force is zero and your velocity stays constant. That is the same law."
Your log has two gaps: you needed an example to justify "forever", and you first mixed up force with net force. That is your revision list.
What should I do with the gaps?
Turn each logged question into a small task. For a half-understood idea, the Fermi Concept Library gives every concept in physics, chemistry, biology, mathematics and computing its own lesson and its place in the map, so you can see where the idea fits. Then teach Riya again the next day with a different misconception on the same topic.
The Feynman technique is the solo version of this method, and once the gaps are closed, making an AI chatbot quiz you tests recall. Parents can play Riya too: ask for one idea at dinner and keep asking "why?" until you hear "I'm not sure". That moment is the point, not a failure.
If you would rather be taught than do the teaching, Fermi's AI tutor works the other way round: it notices where you are stuck, decides whether to explain, ask or show, and checks your understanding as it goes.
