Explaining it back

Who is doing the talking

There is a decisive difference between asking a model to explain something and explaining it to the model. The first is reading. The second is producing, and producing is where learning happens.

The research name for the second is self-explanation: while studying, saying out loud how each step follows and why. Studies since the late 1980s find that students who spontaneously self-explain while working through examples learn considerably more from the same material than those who do not, and that prompting students to self-explain raises outcomes for everybody.

A language model is an unusually good self-explanation partner, for one boring reason: it will listen to eight hundred words of confused reasoning at eleven at night and respond specifically. No human tutor is available on those terms.

The prompt, and the discipline it needs

I am going to explain [topic] to you from memory, as if you know nothing about it. Do not help me. Do not fill in gaps. Do not tell me I am doing well. When I finish, tell me: which parts were vague, which were actually wrong, which I skipped, and which I appear to have memorised without understanding.

Then write it. Properly, in full sentences, without opening your notes. Ten minutes.

The discipline: do not look anything up while writing. The moment you check, the exercise becomes copying. The gaps are the output. If you cannot continue, write I do not know how the next part works and carry on from the other side.

The distinction that makes the feedback useful

The fourth item in that prompt (memorised without understanding) is the one that earns its place. It asks the model to distinguish between two things that look identical on paper: a correct statement of a rule, and a grasp of why the rule holds.

The tell is usually that you can state the result and not the condition. The mean of the sample estimates the population mean: correct, and it does not tell anybody whether you know what randomisation is doing there. Asking the model to probe for that catches a very common state: fluent, correct, and hollow.

Follow it up with the question that settles it:

Ask me three questions that would distinguish someone who understands this from someone who has memorised the standard statement.

Elaborative interrogation: the why chain

A close relative, with its own evidence base. For each fact, ask why is that true? and answer it. Then ask why of the answer. Three levels is usually enough to reach either bedrock or the edge of your understanding, and both are useful destinations.

Metals conduct electricity. Why? Delocalised electrons move freely. Why are they delocalised? The outer electrons are weakly held and the lattice of positive ions shares them. Why does that lattice form?

Models are good at being the why in this chain and terrible at being the answer if you let them, because after two levels they will start producing plausible depth. Keep answering yourself and use the model to ask.

Explaining to a specified audience

Changing who you are explaining to changes what you have to know.

• To a 12-year-old. Forces you to drop jargon, which is where hidden reliance on terminology shows.
• To a sceptical expert in a neighbouring field. Forces you to justify rather than assert.
• To someone who holds the opposite view. Forces you to know what the disagreement is actually about.

Ask the model to play one of these and to stay in character, including asking the questions that audience would ask. It does this well, and the questions from the sceptical-expert version are usually the ones your marker would ask.

The version with no technology

All of this predates the tools and works without them. Explain it to a friend, a family member who does not know the subject, an empty room, or the voice recorder on your phone. The value is in the production, and the model is a convenience: a listener who is always awake and will tell you where you were vague. If your data has run out, talk to the wall. It works, minus the feedback.

Where it fails

Self-explanation works on material you have already met once. It does nothing on material you have not encountered (there is nothing to explain) and students sometimes conclude they are bad at it when they have skipped the reading. Encounter first, briefly. Then explain. Then read again to fill what the explanation exposed.