TL;DR
Epistemic Debt is the gap between what AI did for you and what you actually understand. The study introduces two tools for closing that gap: the Explanation Gate (explain the AI's answer in your own words before you use it) and the Consultant Stance (ask the AI "why," not just "do it").
The headline finding is a "competence collapse." You hand a task to AI, get a great result, use it. On the surface, everything's fine. But when the AI isn't available, the context gets reset, or you need to explain the solution to someone else — you can't. Not because you're lazy: without actively putting it into your own words, the information simply doesn't stick in long-term memory. You read the answer — but didn't process it. The knowledge feels like yours, but it never became yours.
The Explanation Gate solves this with a mandatory "teach-back": before applying an AI answer, explain its essence in your own words. A second AI (or the same one) grades the depth of your understanding — not what you said, but why it works that way. If the explanation is shallow, you get a leading question, not the answer. The gap closes.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
The method works as two requests in one chat:
REQUEST 1: [Your task] → the AI generates an answer/solution
REQUEST 2 (Explanation Gate):
You explain the answer in your own words
→ An AI judge grades the depth of understanding
→ Score < 3 out of 5: the AI asks a Socratic question (doesn't give the answer)
→ Score ≥ 3 out of 5: permission to use it
The loop repeats until understanding is reached.
In parallel — the Consultant Stance as a habit for every request:
❌ Contractor: "Do X"
✅ Consultant: "Explain why X beats Y and where X might break"
Example in Practice
Task: Hannah is a product manager at an EdTech startup. She asked Claude to draft a B2B go-to-market strategy: the target is corporate training, priced at $2,000/year per license. She got a slick 5-page document. The next day the CEO asks: "Why that price? What's it based on?" — and Hannah can't explain.
Prompt (two-step):
Step 1 — a regular request:
Draft a B2B go-to-market strategy for an EdTech platform.
Target segment: corporate training, companies with 500+ employees.
Price: $2,000/year.
Format: key decisions with justification.
Step 2 — the Explanation Gate (right after you get the answer):
Hold on. Before I put this strategy to work —
play the role of a strict sales mentor.
I'm going to explain the essence of the strategy to you in my own words:
[INSERT YOUR EXPLANATION — 3-5 sentences:
why this price, why this segment,
what the key risk is]
Rate my understanding on a 1-5 scale:
- 1: Restated what's written, without understanding the reasons
- 3: Explained the pricing logic and how it ties to the segment
- 5: Named the key risks and alternatives
If the score is < 3 — ask one leading question, don't give the answer.
If the score is ≥ 3 — confirm and add one nuance I missed.
Response format: {"score": X, "feedback": "..."}
Result: The AI will return a score and either ask a probing question ("You named the price, but why exactly $2,000 rather than $1,500 or $2,500?") or confirm your understanding. After 1-2 rounds Hannah can explain the strategy to the CEO in her own words — because she formulated it herself.
Why It Works
The human weakness when working with AI: Reading someone else's text creates an illusion of understanding. We read a well-written answer, it seems logical — and the brain logs "got it." But that's recognition, not knowledge. The information stayed in working memory and never moved into long-term memory.
The LLM's strength: The model is great at generating text to a pattern. Give it the role of "a strict mentor who asks Socratic questions" and it will reproduce that pattern precisely — digging into the reasons, withholding ready-made answers, grading depth.
How the method uses this: The Explanation Gate turns passively receiving an answer into actively formulating it. When you try to explain something in your own words, you find the holes in your understanding. The AI judge points them out without filling them in. This is a "desirable difficulty" — the brain works harder and locks in the knowledge.
Control levers: - The grading scale (1-5) → lower the threshold to 2 for simple tasks, raise it to 4 for mission-critical decisions - The mentor role → swap in a specific person ("a tough CFO," "a skeptical Sequoia investor") — for sharper grading - "Don't give the answer" → drop this condition if you want to learn quickly with hints - Number of rounds → add "maximum 3 iterations" so you don't get stuck in a loop
Prompt Template
# STEP 1: Get the AI's answer
{your regular request}
---
# STEP 2: Explanation Gate (paste after you get the answer)
Before I use this answer — step into the role of a strict mentor.
My explanation of your answer in my own words:
"{my explanation: why exactly this solution,
what the core logic is, what the main risk is}"
Grade the depth of my understanding on a 1-5 scale:
- 1 (surface): Restated what's written, without understanding the reasons
- 3 (relational): Explained the connections between elements and the logic of the solution
- 5 (deep): Named the risks, alternatives and edge cases
Rules:
- If score < 3: ask ONE leading Socratic question. Don't give the answer.
- If score ≥ 3: confirm my understanding + add one nuance I missed.
Format: {"score": number, "feedback": "text"}
Placeholders:
- {your regular request} — any task from step 1
- {my explanation} — you write it yourself, 2-4 sentences in your own words
🚀 Quick start — paste into the chat:
Here's an Explanation Gate template for learning through AI.
Adapt it to my task: {your task}.
Ask me questions to understand the context.
[paste the template above]
The LLM will ask what exactly you want to understand more deeply and what level of understanding you need — because without that it can't calibrate the strictness of the grading or pick the right threshold.
Limitations
⚠️ Friction slows you down: The Explanation Gate adds 15-20 minutes per session. For urgent tasks where understanding doesn't matter, it's wasted time.
⚠️ Doesn't work for mechanical tasks: Formatting a table, translating, fixing typos — there's nothing to "understand." The method is for when AI makes decisions for you, not when it handles routine.
⚠️ The AI judge can be gamed: The study recorded attempts at "gaming" — people gave formally correct but empty explanations. The defense is a stricter judge prompt that demands concrete examples.
⚠️ It helps, but doesn't eliminate the problem: The Scaffolded group showed 61.5% success versus 23.1% for Unrestricted. That's 2.5x better — but not perfect. Some debt still accumulates.
How They Studied It
78 participants — CS students and bootcamp grads — got an assignment: build a course-planner app in React in 90 minutes. None of them had React experience — which was exactly the point, to capture the "learning through AI" moment. Three groups: no AI, unrestricted AI, and AI behind an Explanation Gate (the VibeCheck plugin for Cursor IDE).
The cleverest part of the design is phase 2. After 90 minutes, everyone's AI was "taken away" and a bug was injected into their own code: the await was removed from asynchronous calls. Not a new bug — a regression in their own working code that they had "built" in phase 1. The result laid the difference bare: the Unrestricted group failed 77% of the time — people couldn't fix code they had just "written." The Scaffolded group succeeded 61.5% of the time — nearly twice as well. Especially telling: the Manual group (no AI at all) scored 69.2% — better than Unrestricted AI! High productivity with AI but without understanding turned into lower competence than just doing the work by hand.
An unexpected finding: 6 people in the Unrestricted group did manage. Their logs showed they used AI as a consultant, not a contractor — an average of 4.2 explanatory questions per session ("why do we need useEffect here?") versus 0.4 for everyone else.
Consultant Stance: Practical Prompts
The main extractable principle is changing the type of request. Here's the difference in practice:
Contractor (racks up debt):
Write the copy for an email newsletter about a new personal finance course.
Consultant (builds understanding):
Write the copy for an email newsletter about a new personal finance course.
After the copy, explain:
1. Why this particular structure (and not another one)?
2. What psychological trigger does the subject line use?
3. Where might this copy fail, and why?
Adaptations and Extrapolations
🔧 Technique: "Reverse Teach-Back" — have the AI quiz you instead of grading you
Instead of having the AI hand out a score, ask it to ask you 3 questions about the material before it gives the final answer:
Before answering {task} — ask me 3 questions
to find out what I already know about the topic.
After my answers — give the full answer,
taking the gaps in my understanding into account.
This flips the direction: not "grade what I understood," but "find out what I know" — the model simulates a diagnosis rather than a test.
🔧 Technique: The Explanation Gate for Decision-Making
Apply it not to learning but to important decisions. Before you act on an AI recommendation, explain its risks:
The AI suggested: {recommendation}
Here's why I think it might not work: {your words}
Assess: do I understand the real risks, or am I rationalizing?
If I understand — confirm. If not — ask one question.
Resources
Paper title: Mitigating "Epistemic Debt" in Generative AI-Scaffolded Novice Programming using Metacognitive Scripts
Author: Sreecharan Sankaranarayanan, Extuitive Inc. (Flagship Pioneering), Cambridge, Massachusetts
Conference: L@S '26 (The 13th ACM Conference on Learning at Scale)
Key concepts from the paper: Cognitive Offloading vs Cognitive Outsourcing (Kirschner, 2026), Script Theory of Guidance (Fischer), Desirable Difficulties (Bjork), SOLO Taxonomy, Teachable Agents paradigm