Using AI to Reverse-Test Understanding: A Feynman-Style Teach-Back Learning Method

AI can produce explanations that sound clear, but that doesn't mean you've truly mastered them. A more effective approach is to reverse the conversation: first explain the concept in your own words, then have the AI check your reasoning against the materials you provide.

This exercise is well suited to learning programming concepts, preparing for exams, or reviewing training documentation. Its core isn't to have AI explain things for you—it's to have AI test whether you can explain things independently.

Why "understanding" isn't the same as "mastery"

AI-generated explanations tend to be logically smooth and precisely worded, creating an illusion that "I get it." But the real marker of understanding is this: close the materials, and you can explain the concept clearly in your own words and answer follow-up questions.

Teach-Back (explaining back) is a classic way to test this. Combining it with AI hinges on a role reversal—the learner produces output first, and AI only serves as a cross-checker, rather than acting as the narrator from start to finish. The original material explicitly warns: AI confidently agreeing with you does not equal independent verification. It may simply echo what you say, so you must require it to judge against the reference materials you provide.

The Generation Effect is a classic finding in cognitive psychology: compared with passive reading or listening, learners who actively generate information (such as restating, summarizing, or explaining in their own words) show significantly better memory retention and depth of understanding. The reason is that active generation requires drawing on existing knowledge to reorganize and encode it, and this process itself strengthens neural connections. Teach-Back is a direct application of the Generation Effect—the learner produces output first, exposes knowledge gaps, and then patches them in a targeted way. AI plays the role of "structured examiner" rather than "knowledge provider," which precisely avoids the trap of passive reception. Understanding this makes clear why the article emphasizes role reversal: it's not that AI explains poorly, but that passive listening itself makes it hard to form solid understanding.

The Four-Step Process

Step 1: Pick a concept and a reliable reference

Choose a specific concept and prepare a reliable reference. For example, when learning database transactions, prepare the relevant section from your course notes or official documentation. The quality of the reference determines the quality of the subsequent check—AI can only judge right and wrong within the scope of the materials you provide.

Step 2: Close the materials and explain in your own words

Close the reference materials and explain the concept from memory and understanding. Don't look anything up during this step—the goal is to expose your true level of mastery. It's fine if the explanation is incomplete or imprecise; those are exactly the things to fix later.

Step 3: Have AI diagnose against the materials

Give both your explanation and the reference materials to the AI, and ask it to distinguish three categories: correct points, existing gaps, and claims that your provided materials cannot support.

You can use the following instruction directly:

"I'm learning [topic] at [level]. My reference material is [paste permitted excerpts or notes]. Please ask me to explain the topic in my own words and wait for my answer. Compare my explanation with the provided reference material, point out what I got right, one important gap, and anything the material cannot establish. Ask one follow-up question but don't reveal the answer. Let me revise. Don't fabricate facts, citations, or mastery scores."

The last sentence—"Don't fabricate facts, citations, or mastery scores"—is important. When AI is asked to score, it tends to produce assessments that look precise but have no basis.

The line "Don't fabricate facts, citations, or mastery scores" in the prompt addresses two known weaknesses of large language models: first, hallucination—generating plausible but incorrect content when lacking grounding; second, sycophancy—the tendency to accommodate the user's statements or give the positive evaluation the user expects. When asked to score, the model may output precise yet baseless numbers like "mastery: 85%," creating a false sense of authority. Requiring it to diagnose only against the reference materials you provide essentially anchors the model's output to verifiable text, reducing room for improvisation. This is also why Step 3 emphasizes distinguishing "claims the material cannot support"—the model's own "knowledge" should not be used as a basis for grading.

Step 4: Answer one follow-up at a time, revise, then continue

After the AI poses a follow-up question, answer it first, revise your explanation based on the feedback, and only then move to the next round. Handle one question at a time to avoid information overload and keep revisions targeted.

A Concrete Example: Database Transactions

Taking database transactions as an example, a competent learner's explanation might be: a bank transfer involves two related updates that should either succeed together or roll back together.

A reasonable follow-up would be: what happens if the debit succeeds but the credit fails?

This question tests whether you truly understand atomicity, rather than just having memorized the definition. The original material specifically notes that this is only a teaching example, not a recorded real test result—it should not be cited as evidence of effectiveness.

Atomicity is one of the four ACID properties of database transactions, meaning all operations in a transaction either commit successfully in full or roll back in full—there is no intermediate "partially complete" state. In the bank transfer example, the debit and credit must be processed as a single atomic unit: if the debit succeeds but the credit fails, the system must roll back the debit, otherwise the accounts won't balance. Atomicity is guaranteed by the database's logging and recovery mechanisms, typically implementing rollback through an undo log. The key to understanding atomicity is being able to reason through the consequences of an "intermediate failure" scenario, not just recite the definition—which is exactly why the article uses follow-up questions to test understanding.

Reference Materials Are the Basis for Grading

Whether this method works depends on whether you provide a reliable reference. AI's corrections need to be checked back against the materials, and when something is uncertain, you should ask it to flag it.

Before treating AI feedback as a conclusion, ask yourself two questions: Can this correction be found in the reference materials? Did the AI add content beyond the materials that it "knows"? The former ensures accuracy; the latter helps you spot the model overstepping.

Tools and Caveats

The Fort That Holds offers a free Teach It Back example on its product page. An optional Study With a Second Brain collection includes eight guided workflows, a copyable HTML guide, and editable text for a one-time fee of $9. Note that this is a prompt toolkit designed to work with your existing AI service—not an AI subscription service.

The workflow has only undergone basic model smoke testing; no learning-effectiveness research has been conducted, and no grades or exam results are promised. In other words, it's a learning method worth trying but lacks effectiveness data, so users should keep reasonable expectations.

Use Cases and Limitations

This exercise suits self-studying programming concepts, exam review, and digesting training documentation. Its strength is turning passive reception into active output, with AI providing immediate, reference-based feedback.

Its limitations are equally clear: feedback quality is limited by the reference materials you provide; AI may still misjudge or go along with your thinking; and there is a lack of research evidence on learning effectiveness. Treating it as a structured self-check tool rather than an authoritative grading system is the sensible way to use it.

One-line Summary

Use the Feynman-style teach-back method to have AI check your explanation against reference materials, exposing real gaps in understanding.

Paragraph Summary

The article introduces an AI-assisted learning method called Teach-Back, whose core is role reversal: the learner first closes the materials and explains the concept in their own words, then has AI diagnose their explanation against reliable reference materials—identifying correct points, gaps, and unsupported claims—and guide revision through follow-up questions. The article emphasizes that AI's clear explanations create an illusion of "mastery," and that AI's agreement doesn't equal independent verification; its corrections must be checked back against the reference materials. The prompt must explicitly require AI not to fabricate facts, citations, or mastery scores. This method suits self-studying programming, exam prep, and digesting training docs, but the toolkit has only undergone basic smoke testing and lacks learning-effectiveness research, so it should be viewed as a structured self-check tool rather than an authoritative grading system.