By review #3, your flashcard is a recognition test.

The recognition trap, in one paragraph

The active recall research that flashcards are built on rests on a specific claim: retrieving a fact from memory strengthens the memory trace more than re-exposing yourself to the fact. The flashcard is the unit of retrieval. Front of the card you see a cue, you retrieve, back of the card you check. So far so good. The problem is that on review #2, the cue is identical. By review #3 it is not the fact you are retrieving anymore, it is your memory of the wording on review #1. The retrieval part has quietly stopped happening. Your brain recognizes the cue, fetches the answer it associated with that cue, and you check the back of the card and feel productive. Pattern recognition is doing the work that retrieval was supposed to do, and there is no way to tell from the inside.

The fix is uncomfortable: the cue has to change. Same fact, different words, different format, different surrounding context. Only then is your brain forced to retrieve the underlying concept rather than surface-match the sentence. AI flashcard creation is uniquely well positioned to do this, because rewording a stem at scale is a thing an LLM does cheaply and reliably. Most AI flashcard tools do not do it anyway, because their pipeline ends at "generate cards" and never touches the cards again.

One source, four recall formats

The first half of "active" is format variety. The same fact lives in four card types, each of which exercises a different recall mode. You see all four across the lifetime of a deck, not just the format that happens to be easiest to generate.

Free-response is pure recall: the prompt asks "name the structure that separates the right and left ventricles" and you have to type the answer with no options to choose from. Case-style takes the same fact and embeds it in a clinical scenario, so you have to apply the fact instead of just retrieve it. Image-occlusion is the format most AI flashcard tools quit on: if the source slide has a labeled anatomy diagram, the figure is extracted, the labeled structure is masked, and you recall the label from the surrounding context. MCQ is the fastest of the four, but the distractor quality is what decides whether the card is actually testing you (more on the rubric below).

The same fact, dressed three different ways

Below is the same underlying fact (the interventricular septum is the wall separating the ventricles) on review #1, review #3, and review #5. Stem reworded each time. Distractor pool rotated each time. Right-answer index moves around. The topic-pin stays identical, which is what the scheduler tracks.

The rubric that runs before any card ships

The other reason most AI flashcard tools produce passive cards is that there is no quality gate between generation and the user. Whatever the model wrote, the user sees. Studyly grades every card before it ships on four dimensions, all of which directly feed into whether the card actually exercises recall:

• Factual correctness. Does the answer match the source? Cards that fail this never reach you. A flashcard with a wrong answer key creates negative learning, which is worse than no card at all.
• Distractor quality. Are the wrong answers plausible? A multiple-choice card with three obviously-wrong options is a binary recognition test ("which of these is real medicine?"). Distractors have to look like the right answer at first glance and be wrong on a specific point.
• Clarity. Can the stem be read once and understood? Ambiguous cards reward students who memorize the answer-letter, not students who read the question.
• Type coverage. Is the right mix of formats produced from a given source, or is every card an MCQ because that is the easiest format? Type coverage is what guarantees the four-format spread shown above.

Each dimension is a sub-score that rolls up into the question-quality number on the eval below. The same rubric runs at revisit time as a quality gate, which is how the auto-rephrasing pass does not silently drift into nonsense.

The held-out eval, in numbers

Three source documents (a slide deck, a textbook chapter, a paper) were held out. Each tool generated cards from the same three documents. Every output was graded on the four dimensions above. Same documents, same rubric, same graders.

Higher is better. Full methodology and the rubric definitions are on /quality. The 23-point gap between the top and bottom of this list is mostly distractor quality and type coverage, which are the two dimensions that decide whether a card is genuinely testing recall or just asking you to spot the obvious answer.

When manual flashcards still beat the AI version

Three honest cases where typing your own cards is the right move.

• You are the source of the fact. Personal mnemonics, your own diagnostic shortcuts, the way your attending phrased a clinical pearl in rounds. There is no source document to upload because the source is in your head. Type it.
• The deck is short and you only need a single pass. The 60-second conversion still wins on time, but the auto-rephrasing and the four-format spread don't get to do their work on a twenty-card deck you take once and never see again. The unique value is in the loop, and a single-pass deck doesn't run the loop.
• Computational problem sets. Worked-equation cards (calculus, physics, quantitative pharmacology) need a math problem-solver, not a quiz tool. Studyly handles concept questions, not the mechanics of integrating an equation step by step.