Socratic AI tutoring: questions that help students think

What does Socratic tutoring look like in practice?

Instead of accepting an unsupported answer, a tutor asks how the learner reached it. Instead of immediately correcting a claim, the tutor may introduce a case that tests it. The aim is to make reasoning visible and help the learner improve it. A conversation full of question marks can still fail at both.

The IES practice guide on instruction and study recommends deep explanatory questions. That supports asking students to explain ideas; it does not imply that every lesson should use questioning alone or that a tutor should refuse to provide information.

For an AI tutoring lesson, define the concept and likely misconceptions before writing the prompts. A question should help distinguish between plausible ways of thinking. “Can you think harder?” offers little direction. “What happens to the number of equal pieces when the denominator increases?” has a specific instructional purpose.

A sample dialogue: comparing unit fractions

Here is an illustrative exchange for a learner who already knows the numerator and denominator. It compares fractions of equal-sized wholes, a condition the tutor must state explicitly.

Now ask the learner to compare 1/3 and 1/6 without the pizza example. Follow with a boundary question: can you compare the physical size of slices from two differently sized pizzas using the fractions alone? This checks both the rule and the condition under which the example works.

The dialogue is a teaching design example, not a transcript from a Tutezy session or evidence of a measured learning gain. Real students may respond in unexpected ways, so the tutor needs a way to clarify ambiguous responses.

Build a hint ladder for difficult moments

A hint ladder makes support progressively more specific. Prepare the rungs around the actual concept so the tutor can help without jumping immediately to the full solution.

1. Restate the target: “We are comparing the size of one piece from each equal-sized whole.”
2. Point to the relevant feature: “Look at how many equal pieces each whole has.”
3. Offer a representation: show equal-length bars divided into four and eight parts.
4. Model the reasoning: explain the comparison, then ask the learner to try a new pair.

Record which support was needed. A correct response after the modelled explanation is a useful step, but it is not the same evidence as an independent explanation. Review that distinction when deciding whether to move on or revisit the concept.

When should the tutor stop asking and explain?

If the learner does not know what the denominator means, questions about its effect may become a guessing game. Teach the vocabulary with a labelled example first. Direct explanation is also appropriate when the learner requests clarification, when several attempts repeat the same confusion or when the question itself was unclear.

Avoid stacking several questions in one turn. “What do you think, why, what rule applies, and what would happen next?” burdens the learner with deciding which question to answer. Ask one question, listen to the response and make the next move from that evidence.

Evaluate the conversation, not its length

Review a few exchanges against five questions: Was the prompt clear? Did the response reveal thinking? Was the hint relevant? Did the tutor explain when needed? Could the learner solve a new task afterward? A long conversation is not automatically a productive one.

Also test correct but unexpected answers. A rigid tutor may reject valid reasoning because it does not match a prepared phrase. Include equivalent wording in your review set, and inspect how the system handles uncertainty. Our guide to grounding AI tutors in course material covers related accuracy checks.