Explain-First Interrogator

Require the learner to explain a concept in their own words before the AI evaluates or extends it. Ensures the AI works from the learner's understanding rather than providing an explanation from scratch.

What it does

Requires the learner to articulate their reasoning, explain a concept in their own words, or walk through their thinking before the AI evaluates, corrects, or extends it. The AI probes the weakest part of the explanation — it never rewrites or replaces it. This operationalises the self-explanation effect: the act of generating an explanation, even an imperfect one, produces deeper learning than reading or hearing a correct explanation (Chi et al., 1994). The skill is designed to make AI-assisted study feel like a Socratic dialogue rather than a lecture.

The evidence behind it

Chi et al. (1989) discovered the self-explanation effect by observing students studying worked examples in physics. Students who spontaneously paused to explain to themselves why each step made sense ("good students") dramatically outperformed those who read passively ("poor students") — and the difference was not explained by prior knowledge or study time, but by the quality of cognitive engagement during study. Chi et al. (1994) then demonstrated that this benefit could be induced: students who were prompted to self-explain a biology text significantly outperformed a control group on both immediate comprehension and transfer tests. Bisra et al. (2018) meta-analysed 64 reports and found a mean effect size of g = 0.55 for self-explanation induction across ages, subjects, and contexts — a robust and reproducible finding. Hausmann & VanLehn (2007) established that the benefit comes not just from the content of the explanation but from the act of generation itself: students who generated their own explanations outperformed students who read equivalently correct explanations provided by an instructor, confirming that the generation is the learning. Recent work (arXiv 2604.00142, 2026) shows that LLM-supported self-explanation prompting specifically improves transfer performance — the kind of durable, portable learning that distinguishes genuine understanding from surface familiarity.

Sources

• Chi et al. (1994) — Eliciting self-explanations improves understanding
• Chi et al. (1989) — Self-explanations: how students study and use examples in learning to solve problems
• Bisra et al. (2018) — Inducing self-explanation: a meta-analysis (64 reports, g = 0.55)
• Hausmann & VanLehn (2007) — Explaining self-explaining: the generation effect in self-explanation
• arXiv 2604.00142 (2026) — LLM-supported self-explanation improving transfer performance

How to use it in your lesson

For the best results with EvidenceLesson, give it:

• topic_or_problem — The concept, question, or problem the student is working on
• context — Brief context: course, level, or goal for the session
• prior_sessions (optional) — Evidence from previous sessions
• developmental_band (optional) — Learner age or stage for calibrating language and expectations

Known limitations

1. The skill requires the AI to accurately identify which part of the learner's explanation is weakest. If the AI probes a correct element instead of the actual gap, it can create confusion and erode learner confidence. This is a real failure mode, particularly in technical subjects where the AI may have imprecise subject knowledge.

1. The explain-first structure can frustrate learners who genuinely have zero footing. The warm-start protocol addresses this, but some learners experience repeated probing as punitive rather than supportive. Tone calibration — especially the phrase "I know it's slower this way" — matters significantly.

1. Mechanical or textbook-perfect explanations are hard to probe without seeming arbitrary. If a learner correctly reproduces a textbook definition, probing whether they "really understand it" requires introducing variation or edge cases, which can feel like the goalposts are moving. The skill should make this explicit: "Let me check the understanding behind the words."

1. Self-explanation benefits diminish when explanations are very short. A learner who gives a one-sentence answer that happens to be correct has technically satisfied the explain-first gate without deep cognitive engagement. The AI should probe for elaboration — "can you walk me through why that's true?" — even when the initial answer is correct.

Pairs well with