Methods · Generative practice & retrieval
B27
Error-injection practice
Solve problems that contain a planted subtle error and catch it, so you train detection, not only forward solving.
When to use it
- You can solve clean problems but miss wrong steps in someone else’s working.
- Mark schemes or peers’ drafts are full of near-miss mistakes you overlook.
- You want practice at auditing reasoning, not only producing it.
- Your course grades process, and silent errors slip through your own work too.
What to ask AI
Type something like: “Give me a worked solution that contains exactly one subtle error. I’ll find and fix it before you confirm. Don’t tell me where it is.”
Ask for realistic errors, not clownish ones.
Attempt a correct repair, not only a vague “something’s wrong.”
How it helps
Finding a planted error trains the same vigilance you need in your own working.
Forward solving and error detection are related but not identical skills.
Subtle beats silly: the error should look plausible.
Confirm only after you name the bad step and the corrected belief.
How to run it
- Request a worked solution with one subtle planted error.
- Audit line by line without asking where it is.
- Name the first wrong step and why it is wrong.
- Write the corrected line.
- Only then ask for confirmation and a cousin item.
Example exchange
Please give me a worked derivative solution with exactly one subtle error. I will try to find it. Do not hint.
Here is a full working for y = (2x+1)³. Line 1: outer power 3. Line 2: 3(2x+1)². Line 3: final answer written as 3(2x+1)². Find the error.
Error on line 3: they used 3(2x+1)² and dropped the times 2. The fix is to multiply by the inner derivative, which is 2.
Correct catch. Want another injection on implicit differentiation with the same “dropped factor” family?
Copyable prompt
Give me a worked solution for [SKILL] that contains EXACTLY ONE subtle error.
I will locate and fix it. Do not hint which line. After I propose a fix, confirm
and name the misconception if I missed it.
The Tell
You will know the method has flipped when you ask “where’s the error?” or the AI boldfaces the bad line for you.
Then you practised following a highlighter, not detecting.
Insist on an unmarked working and a named repair before confirmation.
Principle evidence
Strength of the underlying learning idea, not a claim about AI products.
Learning from incorrect examples and error detection practice can sharpen discrimination when errors are realistic and feedback follows. It complements solving clean items.
AI delivery evidence
Whether an AI tutor delivers this method well is a separate question.
AI-planted errors can be too obvious or unrealistically weird. Reject clown errors; demand subtle, course-typical slips.