Spaced repetition with AI: cards made from what you were just taught
A flashcard is worth making only after a concept has passed a test, and the system that taught the concept should be the one that writes the card.
Spaced repetition with AI works when the AI writes the card after you have passed a test on the concept, and writes it to one rule: one question, one answer a grader can mark. Most AI flashcard generators skip that order and cut cards from a PDF you have not yet understood.
Spacing is one of two techniques Dunlosky's 2013 review rated high utility, and the forgetting curve it counters was replicated in 2015. Make two to four cards per concept, order the queue so a day spans several concepts, and read the intervals and lapses in Anki as the measure of retention.
Spaced repetition with AI works when the cards come after understanding
Spaced repetition with AI works when the AI writes each card after you have passed a test on the concept. Each card then carries only what one question can grade. A card is a retrieval cue for something already in your head. Write one from a page you have not understood and the cue points at text you can only recognize.
Dunlosky and colleagues (2013) reviewed ten study techniques and rated two of them high utility: practice testing and distributed practice. A flashcard schedule runs both at once, which is why the format survives every review of the evidence. It runs both only when the answer on the back is a fact you already reconstructed once without help. Spacing a card you never understood spaces your confusion, on a growing schedule, for months.
The order that fixes this is cheap: read, get tested, then write the cards. The same order governs the other techniques in how to learn with AI.
Why spaced repetition works
Spacing works because most forgetting happens in the first hours and the first day. A review placed inside that fall rebuilds the memory for a fraction of the cost of learning it again. Murre and Dros (2015) replicated Ebbinghaus' forgetting curve and recovered the same shape: a steep early drop, then a long shallow tail. Their data probably carries a small upward step at 24 hours, which several authors attribute to sleep. Each review resets the curve, and the fall that follows a review is slower than the one before it.
Cepeda and colleagues (2006) synthesized the distributed practice literature. Spaced study beats massed study across retention intervals, and the best gap between sessions grows with the retention interval you care about. A schedule built on that finding widens its gaps as a card survives: days, then weeks, then months. Dunlosky's review put distributed practice in the top rating band with practice testing, while rereading, highlighting and summarization all landed in the lowest band.
The problem with AI flashcard generators
Most AI flashcard generators take a PDF, a lecture transcript or a folder of notes and hand back fifty cards in a few seconds. The cards are cut from text you have not yet understood, so what you memorize is the phrasing of a source rather than the concept underneath it. The first rule in Wozniak's twenty rules (1999) is to build on understanding before you memorize, because a card you cannot explain becomes a string you recite.
The generator also reproduces the failure Bastani and colleagues (2025) measured in a field experiment with nearly 1,000 high-school students in Turkey. A ChatGPT-like interface raised practice scores 48% and lowered scores on the later unassisted exam by 17% against students who worked with no AI. The students copied answers and did not perceive the harm. A deck generated from a document you have not read makes the same trade: the work looks finished, and the retrieval never happened.
Generated cards fail a simpler test as well: read a front on its own, away from the source. When the question only makes sense beside paragraph three of the document, the card is a bookmark, and a bookmark on a growing schedule teaches nothing.
Make cards only from concepts that passed a test
Write a card for a concept after it has cleared two checks: a quiz on the day you read the lesson, and an explain-back the next day, after sleep, graded against that lesson. Roediger and Karpicke (2006) gave students prose passages, then had them either take recall tests or restudy. On a final test five minutes later, restudying won. At two days and one week, testing produced substantially greater retention, and restudying was the condition that raised students' confidence. Karpicke and Blunt (2011) found the same ranking a week out: retrieval practice scored 0.67 against 0.45 for concept mapping, and 101 of 120 students did better after retrieval practice, though 75% had predicted concept mapping would do at least as well.
A concept that fails a quiz needs teaching again and testing again the next day, and no card until it passes.
The grader and the card writer should also hold the same lesson. A system that taught the concept, quizzed it and read your explain-back knows which sentence you got wrong, and can aim the card at that error. A separate flashcard app holds only your document. The explain-back graded against the lesson is the harder of the two checks. Speed measured at the retest is the subject of how to learn fast and remember it.
What a good AI flashcard looks like
A good card asks one question with one answer a grader can mark right or wrong. Wozniak called the underlying rule the minimum information principle: one card carries one simple fact, because a card holding four facts fails whenever you miss any one of them, and you then relearn all four.
Five rules follow from it. Ask for a single item, so a card never presents a set or an enumeration. Keep yes/no fronts out, because a coin flip records as recall. Use the negative form for purpose, since "what breaks without X" has one answer while "what is the purpose of X" accepts any fluent paraphrase. Use the field's own terminology on both sides, so the term you rehearse is the term you meet in a paper. Put a number on a card only when the number is the concept itself.
| Card that fails | Why it fails | Card that works |
|---|---|---|
| List the four properties of ACID. | Four facts on one card. You miss one and relearn all four. | Which ACID property stops a transaction from seeing another transaction's partial writes? Isolation. |
| Does an index speed up reads? Yes. | A yes/no front. A coin flip records as recall. | A B-tree index replaces a full table scan with what? A descent to the key in logarithmic time. |
| What is the purpose of a write-ahead log? | Any fluent paraphrase passes, so it cannot be graded. | What breaks if a database commits without a write-ahead log? A crash after commit loses the transaction. |
| What stops two threads touching the same data at once? | Homemade wording. You will not recognize the term in a paper or a code review. | Which primitive gives one thread exclusive access to a critical section? A mutex. |
| What is the Raft election timeout in our cluster? 150 to 300 ms. | A tuning value that moves with every deployment. It teaches nothing about elections. | How many nodes of a five-node Raft cluster must agree to commit an entry? Three, a majority. |
How many cards per concept, and which ones
Two to four cards per concept, picked by the role each card plays rather than by how much of the lesson they cover. The first card holds the definition, which fixes the term. The second holds the trap, the mistake the lesson warned you about, because that is the sentence you will get wrong under load. The third holds one mechanism or one number, whichever carries the concept. Then add one card for every question you asked while learning it. A question you asked marks a gap you noticed yourself. The answer already sits attached to the memory of being confused, which makes it a strong cue.
Cloze deletion suits terms and traps: delete the one word that carries the meaning and leave the sentence standing. Question and answer suits mechanisms, because a mechanism has an order and a cloze hides only one step of that order. A concept that produces eight cards was two concepts under one name. Split it and test both halves.
Order the queue so a day spans several concepts
Order the queue so a single day of reviews crosses several concepts instead of running one concept end to end. Dunlosky's review rated interleaved practice moderate utility, above rereading and summarization. Mixing related items forces you to work out what a question is asking before you retrieve the answer.
A queue that deals twelve cards from one concept in a row lets each card supply the context for the next. You answer card seven because card six primed the term, and the score records a retrieval that would fail on its own the following week. Interleaving lowers your accuracy inside the session and raises it at the retest, so a mixed queue feels worse and scores better. The daily queue should test everything you have passed so far, a few cards from each concept. Dunlosky's full ranking of the ten techniques sits in the guide on the best way to learn.
Read the review history as the measure of retention
Anki records an interval and a lapse count for every card, and those two numbers are the honest measure of what you retained. An interval that keeps growing means the card is holding. A lapse resets it, and one lapse on a card is ordinary. Three lapses on the same card, or lapses spread across several cards from one concept, point at the lesson rather than the cards: the concept reached recognition and never reached recall. Teach it again, then write fresh cards. The Anki manual defines a card as mature once its interval reaches 21 days. Maturity is the earliest point at which the review data supports a claim that the concept is yours. Confidence carries no information here. Roediger and Karpicke found that restudying, the weaker condition at every delay past five minutes, made students feel more sure of themselves.
The Learning Harness writes its cards this way. Once a concept passes the same-day quiz and the next-day explain-back, it writes two to four cards into Anki. It adds one card for each question you asked, and orders the review queue so a day spans several concepts. Those rules sit alongside the rest of the method.
Questions about spaced repetition and AI flashcards
Can AI make Anki cards?
Yes. An AI can write cards and put them in a deck through an add-on API or a plain text import, and the writing is the easy part. The judgment is when to write them. A card earns its place after you have answered a question on the concept with the source closed. An AI that writes cards from a document you have not read gives you a deck of that document's sentences.
Is spaced repetition effective?
Yes, and it rests on two of the best-evidenced findings in the study literature. Dunlosky and colleagues reviewed ten techniques and rated two of them high utility: practice testing and distributed practice. A flashcard schedule runs both at once. Cepeda and colleagues found that spaced study beat massed study across retention intervals, and that the gap to leave grows with how long you need to remember.
How many flashcards should I review per day?
As many as the queue hands you, which follows from how many cards you added rather than from a target you pick. When reviews run longer than the time you have, add fewer new cards and keep clearing the queue, because a skipped day returns as a backlog of lapses. Two to four cards per concept keeps the daily load near the reading time.
What is the best AI flashcard generator?
Judge one on three tests. It should write a card after you have passed a test on the concept rather than straight from a document. Each card should ask one question with one answer a grader can mark. The system that taught the concept should be the one that writes the card, because it knows which sentence you got wrong. Most tools fail the first test.
Should I make flashcards from a PDF?
Read it, test yourself on it, then make cards from what you got right. Cards cut straight from a PDF carry the author's phrasing, so you learn to recognize a sentence instead of recalling a concept. A front that only makes sense beside paragraph three of the source works as a bookmark. Rewrite it so it stands alone, then keep it.
Sources
- Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving Students' Learning With Effective Learning Techniques. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
- Murre, J. M. J., & Dros, J. (2015). Replication and Analysis of Ebbinghaus' Forgetting Curve. PLoS ONE, 10(7), e0120644. https://doi.org/10.1371/journal.pone.0120644
- Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354–380. https://doi.org/10.1037/0033-2909.132.3.354
- Wozniak, P. (1999). Twenty rules of formulating knowledge. SuperMemo. https://www.supermemo.com/en/blog/twenty-rules-of-formulating-knowledge
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2422633122
- Roediger, H. L., & Karpicke, J. D. (2006). Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
- Karpicke, J. D., & Blunt, J. R. (2011). Retrieval Practice Produces More Learning than Elaborative Studying with Concept Mapping. Science, 331(6018), 772–775. https://doi.org/10.1126/science.1199327
- Anki Manual. Statistics. docs.ankiweb.net. https://docs.ankiweb.net/stats.html
The Learning Harness is in a closed beta and not yet available.