If this pillar's evidence had to compress into one behavioural rule, it is this: commit to your own answer before the AI speaks. The rule operationalises the generation effect, retrieval practice's well-replicated superiority over restudy, and the guardrails finding that unguarded answer access harms later independent performance; it defends against the novice paradox by ensuring you hold a position fluent explanation must displace rather than a vacuum it fills; and it is free, portable across every tool, and enforceable tonight. Test-enhanced learning has shown durable benefits in medical education specifically, repeated testing with feedback outperforming repeated study on retention months later, with retrieval improving subsequent clinical application in standardised-patient work. This page turns that evidence into a protocol.
The six-stage protocol
Stage one, commit: choose your answer and write it down, an actual commitment, not a lean, because the generation effect requires production and because a recorded answer cannot be quietly revised into "what I really meant". Stage two, record confidence: high, medium or low, one second's work, and the raw material of calibration, since confidence-accuracy gaps are where the most instructive errors hide. Stage three, state the reasoning: one sentence on why, which converts a guess into an inspectable claim and gives any tutor, human or AI, something to diagnose. Stage four, compare with feedback: now, and only now, open the explanation or the AI, reading it against your recorded reasoning rather than absorbing it cold. Stage five, explain the error in your own words: if wrong, close the explanation and reconstruct why the right answer is right and where your reasoning broke, the reconstruction, not the reading, is where the learning happens. Stage six, reattempt later, unseen: days afterwards, a new question on the same concept, unaided, because delayed unassisted transfer is the only measurement the examinations respect and the only one the fluency illusion cannot fake.
Running the protocol in a general-purpose AI
ChatGPT-class tools will happily wreck stages one to three unless instructed, so instruct them, with a standing prompt: do not answer or narrow options until I commit; after I commit and give my reasoning, ask me one question about it; then hint on request; explain fully only when I say reveal; then set me a new question on the concept. It works, imperfectly, the tool sometimes forgets, and your discipline does the enforcement, which is the structural weakness of running pedagogy on willpower: the product is optimised for frictionless answers and you are optimising against it. The guardrails argument in full is at /blog/chatgpt-is-not-an-ai-tutor-educational-guardrails.
Running the protocol in an integrated tutor
The same six stages, engineered so they run by default, are what an integrated question-bank tutor is for, and it is where our own design shows its hand, stated plainly rather than smuggled: iatroX's flow is attempt-first by construction, the question comes before any help; the Socratic Tutor attaches to your wrong answer and asks about your reasoning before explaining, stage three and four fused; spaced repetition schedules stage six automatically, resurfacing the concept for unassisted retrieval; and unseen items on the same topic supply the transfer test that exact-item repetition cannot. Whether that implementation improves your learning is a claim our forthcoming experiments must earn; that it implements the evidenced sequence is checkable in the product in two minutes.
Two edge cases the protocol handles
When the keyed answer appears wrong: the protocol has already captured your answer and reasoning, so challenge from that record, check the cited source, and report through the question's correction route rather than silently absorbing either version; a bank's response to a documented challenge is governance data. And distinguishing memory failure from reasoning failure: if stage five reconstruction shows you knew the concept but retrieved the wrong fact, the fix is spacing and retrieval volume; if the reasoning itself was malformed, the fix is tutoring and worked contrast, and knowing which you are fixing is half the efficiency of the whole method.
Frequently asked questions
Isn't this slower than just reading explanations?
Per item, slightly; per retained concept, dramatically faster, which is the desirable-difficulty trade the evidence keeps finding, and the five-minute self-test at /blog/illusion-of-learning-ai-fluency-vs-recall will demonstrate it on your own head within a week.
Does confidence recording actually matter?
Yes, cheaply: high-confidence errors are the highest-yield review targets you own, and no other one-second act finds them.
Can the protocol run on paper?
Entirely: answer, confidence, reasoning, check, reconstruct, diary a reattempt; software automates the scheduling and supplies unseen items, the cognition is portable.
What if I keep breaking the protocol under time pressure?
Shrink it rather than skip it: commit and confidence take five seconds, and even that truncated version preserves the generation effect; the full six stages are for the topics that matter most, not for every item of a hundred-question day.