Learn Better With AI

Methods · Dialogue, argument & inquiry

C3

Productive failure

Attempt a problem beyond your current ability before instruction, so the lesson lands on gaps you already felt, instead of on a blank slate that feels clear and teaches less.

When to use this
First time through the material
Learning impact
Strong change
Learning evidence
Credible with caveats
Where you are with the topic
Building understanding
How often to use it
Worth repeating often
What AI is doing (1)Create practice, examples, or schedules
What AI is doing (2)Ask questions that make you think

When to use it

  • You are about to learn a method and are tempted to watch the solution first.
  • Past lessons felt clear in the moment and vanished by the next practice set.
  • You can spare fifteen to twenty-five minutes to struggle before any teaching.

What to ask AI

Type something like: “Give me one challenging problem that needs the target idea, but do not name the method or give hints.”

Submit a full attempt and a note on where you stuck before asking for instruction.

Only then ask it to contrast the canonical approach with what you tried.

How it helps

The standard sequence is merciful: watch the method, then try.

It also removes the condition that makes instruction discriminating.

When you have not struggled, every step of the explanation looks equally important.

When you have failed at a well-posed challenge, the explanation answers a question you are already asking.

Productive failure is that sequence: generate, flounder with structure, then receive the canonical method.

Not all failure is productive. Random impossibility demoralises.

The problem must be adjacent: prior knowledge can generate partial representations even if the full solution is not there yet.

How to run it

  1. Pick a target method you have not been taught cleanly yet.
  2. Get one challenge problem a prepared student would solve with that method, with the method name hidden from you.
  3. Time-box the attempt (about 15-25 minutes). Invent representations, try paths, write where you stuck.
  4. Only then ask for instruction aimed at your attempt, contrasted with the canonical approach.
  5. Re-attempt a near-twin immediately after instruction.
  6. Name in one line what the struggle revealed. That line is the learning residue.

Example exchange

You

I’m stuck on related rates. Can you give me one hard problem first and wait while I try, no method name, no hints?

AI

A balloon’s radius grows at 2 cm/s. How fast is the surface area changing when the radius is 5 cm? Take your time; tell me where you stuck.

You

I wrote A = 4πr² and differentiated, but I don’t know what to do with the 2 cm/s.

AI

Look at your attempt: you treated r like a fixed number. If A depends on r, and r is changing with time, how should dA/dt relate to dr/dt?

Copyable prompt

I want a productive-failure sequence for [TOPIC/SKILL] at [LEVEL].
Phase 1: Give me ONE challenging problem that requires the target idea, but do
NOT name the method or give hints. Wait for my full attempt and my note on where
I stuck.
Phase 2: Only after I submit, teach the canonical approach by contrasting it
with what I tried. Then give a near-twin to solve cold.

The Tell

If you “failed productively” but cannot say what partial model you built, it was only frustration.

Peeking at the method mid-struggle converts the session into an ordinary worked example.

Skip the post-instruction twin and the struggle never consolidates.

Principle evidence

Strength of the underlying learning idea, not a claim about AI products.

The underlying learning idea is rated moderate. Productive failure (struggle before instruction, then learn the canonical method) has a growing literature with clear design boundaries. Effects depend on problem design and delayed instruction.

AI delivery evidence

Whether an AI tutor delivers this method well is a separate question.

An AI must not rescue you early. Whether models hold that constraint well is delivery-speculative.

Related methods