The Cognitive Architecture of Learning

This lecture names the Safety Gap: the distance between what an AI-assisted product appears to demonstrate and what a student can independently do. It explains why the same fluent output that extends an expert's thinking can short-circuit a novice's practice — novices lack the mental models needed to check and calibrate — and why usable understanding is built through retrieval and correction. The design response is a Cognitive Sanctuary: a bounded course moment that protects one essential act of student thinking and makes hidden capability visible without surveillance.

A working distinction between productive struggle (difficulty that builds learning, like retrieval, spacing, interleaving, and elaboration) and needless friction (difficulty that only blocks access, like fighting formats or decoding unrelated jargon). AI can remove both, so the design task is to protect the first and repair the second — the lecture calls the second move an Access Repair and shows why it raises rather than lowers standards. A useful test: would removing the difficulty also remove the evidence of learning? The equity stake is direct, because needless friction falls hardest on students already facing access barriers.

Three postures AI can take in an assignment — Oracle (gives answers; fast but high-risk for novices), Tutor (questions and hints that keep the student retrieving and explaining), and Adversary (produces flawed output for the student to critique). The lecture shows the two common mismatches, novice-plus-Oracle (illusion of competence) and novice-plus-Adversary (shallow critique), and argues that every posture needs one short, observable boundary to keep it from collapsing back into answer-generation. You leave with one posture instruction, one boundary rule, and one sentence naming the thinking being protected.