The Feynman technique is a self-study routine in which a learner explains a topic in plain words to an imagined beginner, notices where the explanation breaks down, and goes back to the source. No controlled study under that name was found, but the idea is close to two researched practices: self-explanation, which a 2013 review rated moderate utility, and learning by teaching, where a 2009 Stanford study had 62 eighth-graders teach a computer character. An AI chatbot can play the beginner, and early studies of that setup are small, with at least one still a preprint.
Key takeaways
- The popular four-step method is a modern summary. No text was found in which Richard Feynman lists the steps himself.
- A 2009 Stanford experiment found that students who believed they were teaching a software character spent more time learning and learned more, with the largest effect for lower-achieving students.
- Self-explanation was rated moderate utility in a 2013 review of 10 study techniques, mainly because the evidence was limited.
- Two 2024 and 2025 studies of a ChatGPT-based learner reported gains, but one is a preprint and the other had 28 participants.
- A chatbot that answers fluently can hide gaps instead of exposing them, so the explanation must be checked against a textbook or teacher.
What the technique is, and where the name comes from
Richard Feynman shared the 1965 Nobel Prize in Physics with Sin-Itiro Tomonaga and Julian Schwinger for work on quantum electrodynamics, and he was widely known as a clear explainer. The study routine that carries his name is usually described in four steps: choose a concept, explain it in simple language as if to a newcomer, find the places where the explanation stalls and revisit the source material, then simplify and use analogies.
That description is a later summary, not a method Feynman published. A search for a primary text in which he sets out these steps turned up none, and popular pages that credit him with inventing the method give no citation. The label is best read as a convenient name for a practice that education research studies under other names.
Research neighbor one: self-explanation
Dunlosky and colleagues reviewed 10 learning techniques in a 2013 monograph in Psychological Science in the Public Interest. Self-explanation, where a learner explains how new information relates to what is already known or how a step was reached, was rated moderate utility, as were elaborative interrogation and interleaved practice. The review placed it below practice testing and distributed practice mainly because the evidence was limited, not because it was shown to fail. Summarization, which resembles a written version of the Feynman routine, was rated low.
The distinction matters. Explaining why a step follows is different from restating a passage in simpler words, and only the first has the stronger research behind it.
Research neighbor two: teaching a student who is not real
The closest experiment to an AI version is a 2009 paper by Chase, Chin, Oppezzo, and Schwartz in the Journal of Science Education and Technology. In Betty's Brain, students instruct a character that reasons from what it was taught. In the first study, 62 eighth-graders learned about fever mechanisms. Everyone used the same software, but one group believed they were teaching their own character and the other believed they were learning for themselves. The teaching group spent more time on learning activities such as reading and learned more, with the benefit strongest for lower-achieving children.
In a second study with fifth-graders, those who taught the character treated it socially, attributed mental states to it, and were more likely to acknowledge errors. The authors suggest that a sense of responsibility may motivate learning and that errors are easier to face when they belong to the character. They also caution that classes were assigned intact, which limits generalizing the result.
What studies of chatbot learners report
A December 2024 preprint by Chen and colleagues tested ChatGPT as a teachable agent in programming education. The authors report improved knowledge gains and programming ability, along with better self-regulated learning. Error correction improved little, which they attribute to the chatbot usually producing correct code and leaving students fewer chances to debug. An April 2025 paper from a CHI workshop gave 28 Chinese university students either a teachable agent or self-directed materials for music analysis. The agent group scored higher on the post-test and reported lower cognitive load.
Both studies cover narrow subjects, the music study had only 28 participants, and the first is listed as a preprint, not a peer-reviewed article. They suggest a direction. They do not establish that explaining to a chatbot beats explaining to a classmate or to a blank page.
A practical routine and what to verify
| Step | What to do | What to verify |
|---|---|---|
| Choose | Pick one narrow concept from the course | It matches the syllabus wording |
| Explain | Write or say it for a beginner, then ask the chatbot to play that beginner and ask questions | Whether the explanation uses reasons, not only definitions |
| Probe | Note every question you could not answer | The gaps against the textbook or class notes |
| Return | Reread the source and revise the explanation | That the chatbot's own statements were not accepted as fact |
| Test | Close the notes and explain again, then answer practice questions | A score on questions the chatbot did not write |
Student names, grades, and school details do not belong in a prompt. Public chatbots may store conversations, so only the concept needs to be shared.
FAQ
Did Richard Feynman invent the Feynman technique?
No source was found showing that he wrote down the four steps. The name is attached to a study habit that others described afterward.
Is there proof the Feynman technique works?
Not under that name. Related practices, self-explanation and learning by teaching, have supporting studies, though a major review called the evidence for self-explanation limited.
Can a chatbot replace a human listener?
The evidence is too early to say. The Stanford study used a character that reasons from how it was taught, while a general chatbot may fill gaps with its own knowledge and hide the learner's mistakes.
Does it work for every subject?
Unknown. The studies cited cover biology, programming, and music theory, with different ages, and one of the samples had only 28 people.
Related reading
- Limits of AI as a study tool
- OpenAI's education products and student rules
- A decade of AI in special education
- Current trends in AI and education
More in AI in education.
Last verified: October 2026
Sources
- Teachable Agents and the Protege Effect: Increasing the Effort Towards Learning, Chase, Chin, Oppezzo, and Schwartz, Journal of Science Education and Technology, 2009
- Improving Students' Learning With Effective Learning Techniques, Dunlosky, Rawson, Marsh, Nathan, and Willingham, Psychological Science in the Public Interest, 2013
- Learning-by-teaching with ChatGPT: The effect of teachable ChatGPT agent on programming education, Chen, Wei, Le, and Zhang, arXiv preprint 2412.15226, December 2024
- Exploring the Impact of an LLM-Powered Teachable Agent on Learning Gains and Cognitive Load in Music Education, Jin, Lin, Hong, Zhang, and So, arXiv 2504.00636, April 2025
- Richard P. Feynman, Nobel Prize in Physics 1965