ChatGPT writes questions. It does not run the loop.

It can do the first 10 minutes of it and then it falls apart. Drop a lecture PDF into a chat, ask for 20 MCQs on it, and you get 20 MCQs. That is a one-pass quiz, not active recall. Active recall is a closed loop: every fact you miss has to come back, the stem cannot be the same wording on revisit (or you pattern-match instead of pulling the fact), the distractors have to be discriminating (or wrong answers are obviously wrong and the question does not actually test recall), and on a miss you need the original slide quoted at you so you can update the memory against the source. A chat window does not carry per-question retention state across sessions, does not run a rubric gate on distractors, does not rephrase the stem on revisit, and does not anchor the explain to a specific slide line. Those are four structural gaps and they all matter by week two.

Partially, and partially does not save you. You can paste the same lecture and ask for the same 20 cards reworded, which fixes the rephrase axis on one document. You cannot ask ChatGPT to remember that you got question 7 wrong yesterday and surface it back today, because per-conversation memory is fragile and chat-spanning memory does not key on a fact-level identifier. You cannot reliably ask it to score its own distractors against a four-axis rubric before showing them to you, because it will say it did and then ship the same kind of synonym-of-the-right-answer distractor it would have shipped without the rubric. The blind eval data is unambiguous on this: on a held-out three-document set, a generic generator scored 57.8 of 100. A rubric-gated pipeline scored 81.3 on the same three documents.

Yes, two. First, single-pass concept explanations: ask it to explain why the loop of Henle counter-current multiplier works the way it does, and you get a coherent explanation you can then test yourself against. That is not active recall; it is the explain step. Use it freely. Second, generating a free-response prompt cold ('write me ten questions on glomerulonephritis, do not include the answers') is a good cold-retrieval drill because you cannot accidentally recognition-match anything. Those two uses are fine. The thing it cannot do is run the loop over weeks across thirty lectures.

Same story on the structural axes. Any general-purpose chat assistant is a one-pass writer of questions, not a retrieval loop, because retrieval requires persistent per-fact state and a rubric gate that lives between generation and surfacing. The model is not the bottleneck; the loop is. Switching to a smarter model improves question quality on a single shot and does not fix miss-tracking, rephrase, or distractor discipline.

Each fact in the system has a topic-pin that survives across encounters. The stem text does not. On revisit #2 of the same underlying fact the system re-rolls the stem wording and reshuffles the option order. By revisit #5 you have seen five different sentences asking the same question, which forces five genuine retrievals instead of one retrieval and four pattern matches. The cognitive-science version of this is the testing effect under varied retrieval conditions. The product version is that you cannot memorize the first three words of the question and call it studying. ChatGPT cannot do this because the chat does not store a stable topic-pin across sessions; it stores a transcript, and a transcript is a wording, not a fact.

It is the single most useful thing on a miss. When you get a card wrong, the explain panel quotes the bullet line on the slide it came from. Not a paraphrase, the bullet your professor wrote. That gives you something concrete to update against and lets you re-pull the slide for context if you want. ChatGPT will paraphrase the source from training data and occasionally invent a citation. The difference shows up most on the slides where the professor's wording is the wording your TA will mark against on the exam.

Factual correctness (is the answer the right answer for this slide), clarity (would a peer reading the stem cold understand what is being asked), distractor quality (do wrong answers discriminate from the right answer, or are they synonyms or obviously wrong), and question-type coverage (does the set include recall, application, case stems, and image-occlusion rather than only one shape). Candidates failing any axis are regenerated before output. On a held-out three-document eval (a microbiology lecture, an internal medicine deck, a pharmacology PDF) the gated output scored 81.3 of 100. The next best generator scored 78.0, the one after that 68.0, and an un-gated generator scored 57.8. Source: Jungle internal admin Quality Comparison panel, 2026-04-24.

Yes. The case-style generator runs on every fact slide regardless of source, so a 90-slide cardiology deck produces around 50 case-style stems alongside the MCQs and free-response cards. Case stems are the highest-leverage format for board-style exams because they mirror the question shape on the actual test. You can filter the queue to case-style only when you want to drill clinical reasoning instead of recognition. AnKing, Zanki, and other community decks remain the right choice for the boards content itself. The class-deck side is where the chat-vs-loop comparison matters most.

Marginally and not in a way that matters. ChatGPT Plus is $20 a month. Studyly has a free tier, no card on file, and a paid tier. The cost question is the wrong one because the bottleneck on cramming season is not dollars per month, it is whether you finish the loop. The chat-window approach loses on whether you actually drill the lectures more than once, which is where retention is paid for.