Study groups, with and without the machine

Why groups outperform solitary study, when they do

A study group works for one reason: somebody has to explain, and somebody has to be unconvinced. Everything else about them (the sharing of notes, the moral support, the biscuits) is secondary to that exchange.

The evidence is strongest for a specific structure. In peer instruction, developed by Eric Mazur for physics teaching, students answer a conceptual question alone, then discuss it with a neighbour who chose differently, then answer again. The proportion answering correctly rises substantially after discussion, including in groups where nobody initially had it right, which tells you the gain comes from articulating and challenging reasoning, not from copying the right answer off the strongest student.

That structure survives the arrival of AI intact, and is worth protecting, because the default group behaviour now is for everyone to consult the same assistant and converge silently.

Running peer instruction with three friends and a free model

1. Generate ten conceptual questions from the material: the kind with a common wrong answer, not recall. Ask for questions where the most popular wrong answer reveals a specific misconception.
2. Everyone answers question one alone, in writing, with a confidence rating. No discussion.
3. Reveal answers to each other. If you disagree, argue. If you all agree, argue anyway: one of you explains, another must try to break it.
4. Answer again.
5. Only now check against the source or the model.

The order matters. A group that checks first has an answer and stops thinking. A group that argues first produces the reasoning, which is the thing being trained.

The role worth rotating

Appoint one person per session as the sceptic. Their job is to be unconvinced and to ask why until either the explanation reaches bedrock or it breaks. Rotate it, because it is the position that learns the most and it is uncomfortable.

This is also the role a model can play when you are alone, using the sceptic prompt from Module 2, but with a group you get something the model cannot give: a person who is confused in a way you did not anticipate, and whose confusion is real rather than performed.

Dividing work: the thing that quietly ruins groups

The most common group arrangement is to split the topics: you do chapters one and two, I do three and four, we share notes. This is efficient and almost worthless for learning. Everybody ends with genuine understanding of half the material and someone else’s summary of the other half, which is exactly the two-tier outcome the fluency lesson warns about.

If you split anything, split the teaching rather than the learning: everybody reads everything, and one person is responsible for leading the discussion on each part. Preparing to teach a topic is one of the strongest study conditions known, and it only works if the rest of the group is prepared enough to ask hard questions.

Where a shared assistant helps

• Settling factual disputes fast so the argument can move to the interesting part. Verify anything that matters.
• Generating the question set, which nobody enjoys writing.
• Playing an adversary for a presentation rehearsal: paste the argument, ask for the hardest questions a panel would ask, take them in turn.
• Translating between the group’s languages. In a multilingual group, having each person explain in the language they think in and translating for the others is faster and more accurate than everybody working in a second language. This is one of the excellent uses.

The group agreement

Write it in one message, at the start, before the first deadline:

For this project: AI for question generation, critique and translation. No AI-generated prose in the submitted document. Everyone opens every source they cite. We each write our own sections. If anyone uses it differently, they say so before submission.

Group work is where academic-integrity cases most often begin, and they usually begin because nobody discussed it. You are jointly submitting a document, and I did not know he generated that section is a poor position to be in, both procedurally and with your friend. Fifteen minutes at the start prevents it.

When there is no group

Distance students, part-time students, people with jobs and families: the group may not exist. The substitutes, in descending order of value: a single study partner on a video call once a week; an asynchronous partner who reads and challenges your explanations in writing; and, last, the model playing sceptic. The last is the weakest, because it will never be confused, and being confronted with real confusion is a large part of what makes the exchange work.