The Teach-It-Back Method and How AI Evaluates It
So far in this guide, the AI has mostly played the role of teacher: it explains, it quizzes, it gives hints. This chapter flips the script. Here we hand the learner the chalk and ask them to do the explaining. It turns out that the single best way to find out whether someone truly understands an idea is to ask them to teach it. And one of the most useful things a modern AI tutor can do is listen to that explanation, spot the gaps, and gently fill them in.
Let's build this up from the ground: first the science of why teaching helps you learn, then a famous four-step method anyone can use, then how to prompt a learner to explain something back, and finally how an AI actually grades a free-form explanation without being fooled.
16.1 Why teaching something is the best way to learn it
There is a well-studied finding called the protege effect: we learn material better when we prepare to teach it, or actually teach it, to someone else. (A "protege" is a student under your wing — so the name means "the effect of having a pupil.") When you know you'll have to explain an idea, you stop skimming. You have to organize the pieces, connect them, and put them in an order that makes sense — and that act of reorganizing is exactly what moves knowledge from your short-term mental "workbench" into durable long-term memory.
Researchers Chi, Roy, and Hausmann argued that teaching works through the same engine as self-explanation — the simple habit of asking yourself "why is this true? how does this connect to what I already know?" while you study. Self-explanation is one of the most reliably effective study strategies ever measured (a large review found a strong average benefit across dozens of studies). Teaching just adds a social, motivational layer on top: you feel responsible for your pupil, so you try harder.
16.2 The Feynman Technique: explain it like the listener is twelve
The physicist Richard Feynman is associated with a four-step routine built on one blunt principle: "If you can't explain it simply, you don't understand it well enough." The steps are:
- Choose & study — pick one concept and learn it as best you can.
- Teach it simply — explain it in plain words, as if to a curious twelve-year-old (ideally out loud, to a real person).
- Find the gaps — notice exactly where you stumble, hand-wave, or fall back on jargon. Those stumbles are your knowledge gaps. Go back to the source and fill them.
- Simplify & refine — clean up the explanation and add an analogy until it flows.
This matters because of a trap we cover elsewhere in the guide: the illusion of fluency. Re-reading notes feels smooth and makes you feel like you know the material. But recognizing words on a page is not the same as being able to produce the idea from scratch. Plain-language explanation pops that illusion instantly — the moment you can't say it simply, you've found something you don't actually understand.
16.3 How to prompt a learner to explain it back
"Teach-it-back" bakes the protege effect and self-explanation into a single tutor interaction: after teaching a concept, the AI asks the learner to explain it in their own words. The craft is in how you ask. A few principles:
- Ask for the "why," not just the "what." "Explain why we flip the second fraction when dividing" reveals understanding; "What is 1/2 ÷ 1/4?" only tests a procedure.
- Name an audience. "Explain it so a younger sibling would get it" forces genuine simplification instead of parroting textbook phrasing.
- Ask for one thing at a time. A single concept per prompt keeps the request inside the learner's limited working memory.
- Invite an analogy or example. "Give me a real-world example" is one of the hardest tests of true understanding — you can't fake an apt analogy.
- Ask before telling. Get the learner's explanation first, then correct. This is a "desirable difficulty": the effort of producing the answer is what builds memory.
THE TEACH-IT-BACK LOOP
┌──────────────────────────────────────────┐
│ 1. Tutor teaches the concept │
│ 2. Tutor: "Now explain it back to me, │
│ in your own words, like I'm new." │
│ 3. Learner explains (free text) │
│ 4. AI evaluates vs. a rubric → │
│ finds gaps & misconceptions │
│ 5. Tutor gives targeted feedback on the │
│ ONE weak spot, then re-asks │
└──────────────────────────────────────────┘
▲ │
└─────── repeat ──────────┘
16.4 How an AI evaluates a free-text explanation
Here is the hard part. Grading a multiple-choice answer is trivial — match it to the key. But an explanation is free text: messy, worded a hundred different ways, partly right and partly wrong. How does an AI judge it fairly? The standard technique is called LLM-as-judge — you use a large language model (the same kind of AI that powers chatbots) as the grader, but you give it strict instructions instead of letting it grade by gut feeling.
The reliable recipe has a few non-negotiable parts:
- Write a concrete rubric. A rubric is a scoring checklist that spells out, point by point, what a good answer must contain — ideally one idea per criterion. For "explain photosynthesis," the rubric might list: mentions sunlight, mentions water, mentions carbon dioxide, mentions that sugar/glucose is produced, mentions oxygen as output. The AI marks each item present or absent.
- Set the model to its most consistent mode (engineers call this "temperature 0") so the same answer gets the same grade every time.
- Require a written justification for every score. Forcing the AI to say why it marked something wrong both improves its accuracy and gives the learner a real explanation.
- Give it a few graded examples ("here's a 5/5 answer, here's a 2/5 answer") so its standards line up with a human's.
Done this way, an AI judge typically agrees with human graders within one point about 80–90% of the time on clear rubrics — good enough to be genuinely useful, and the only way the "teach-it-back" loop can scale to thousands of learners.
The most valuable output isn't the score, though — it's the gap detection. By checking the explanation against each rubric item, the AI can pinpoint exactly what's missing or wrong: "Good — you correctly said plants use sunlight and water. But you didn't mention where the carbon comes from, and you said plants 'breathe in oxygen' — actually they release oxygen. Let's revisit that one piece." That targeted feedback, aimed at the single weak spot rather than re-teaching everything, is precisely what a great human tutor does.
16.5 The pitfalls of auto-grading free text
An AI judge is powerful but biased in known, sneaky ways. If you don't guard against these, you'll hand out unfair grades and learners will (rightly) lose trust.
| Bias | What it means | The fix |
|---|---|---|
| Verbosity bias | Longer answers seem "better" even when padded with fluff | Grade against rubric items present, not length; reward concision |
| Position bias | When comparing two answers, it favors whichever came first | Swap the order, grade twice, and average |
| Self-enhancement bias | A model rates text written by its own family of models higher | Never let a model grade its own output; use a different judge |
| Style over substance | Confident, polished phrasing scores well even if it's wrong | Tie every score to specific rubric evidence in the justification |
- Teaching is learning. The protege effect and self-explanation mean that asking a learner to explain a concept back produces deeper understanding than being explained to.
- The Feynman Technique — explain it simply, find where you stumble, fill the gap, refine — turns vague "I think I get it" into precise knowledge of what you don't yet know.
- Prompt for the "why" and an audience. Ask the learner to explain in plain words to a beginner, give an example, and ask before telling.
- Grade with a rubric, not vibes. Use an AI judge with concrete present/absent criteria, a required justification, consistent settings, and validation against human grades.
- Watch for bias and gaming. Auto-grading free text is fooled by length, polish, and keyword-stuffing — control for these, and the gap-detection feedback becomes the real prize.