Why Flashcards Work (When They Work)
Flashcards are not magic. They are a delivery mechanism for two of the most reliable findings in the science of memory: the testing effect and the spacing effect. Strip those two ideas away and a flashcard is just a fact you wrote down twice.
The testing effect, also called retrieval practice, is the finding that pulling information out of your memory strengthens it more than putting it back in. Roediger and Karpicke (2006) ran the experiment that made this famous. They had students study a passage, then either restudy it or take a recall test. On a final test a week later, the students who had practiced retrieving the material remembered far more than the students who simply read it again. Restudy felt more productive in the moment. It wasn't.
This is the core reason a flashcard works and a highlighter alone doesn't. When you flip a card and try to answer before checking, you force a retrieval. When you re-read a highlighted passage, you recognize it and feel a comfortable hum of familiarity that has little to do with whether you can produce the idea on demand.
The spacing effect is the second pillar. Hermann Ebbinghaus mapped the forgetting curve in 1885, showing that we lose new information fast at first, then more slowly. The fix is not to study harder in one sitting. It is to space your reviews out over time. Cepeda et al. (2006) pulled decades of studies into a meta-analysis and confirmed it: distributed practice produces dramatically better long-term retention than massed practice. Reviewing something five times in one night is close to useless next month. Reviewing it five times across five weeks is close to permanent.
Put the two together and you get the engine behind every serious flashcard system: retrieve the answer, and space the retrievals out at expanding intervals. The card is just the container. If you want the deeper version of this argument, we wrote a whole piece on spaced repetition for readers.
The Friction Problem AI Just Solved
Here is the dirty secret of flashcards: the science is settled, and almost nobody sticks with them. The reason is rarely motivation. It's friction.
Making good cards by hand is slow, fiddly work. You read a chapter, decide what matters, rephrase it into a question, write a clean answer, and repeat fifty times. By card twenty you are exhausted and your cards have gotten lazy. Most people who try Anki quit in the first two weeks, and the quitting almost always happens during card creation, not review. The reviews are easy. The manufacturing is the wall.
AI knocked that wall down. In 2026, a wave of AI flashcard generators (RemNote, AnkiDecks, Quizlet AI, and others) can take a PDF, pasted notes, a photo of a textbook page via OCR, or a YouTube video, and produce a deck in seconds. Many schedule reviews with FSRS out of the box. What used to take an hour now takes the length of a coffee refill.
This is genuinely good news. The thing that made most people quit is gone. But there's a catch that's the heart of this article: AI is fast at making cards and bad at making good ones. It removed the friction that protected you from a worse problem. A deck of 200 mediocre cards is not a win over 30 hand-made ones. It's often a loss, because every bad card costs review time and can teach you the wrong habit of recognition instead of recall.
So the new skill is not "how do I make cards." AI does that. The new skill is "how do I make AI make good cards, and how do I fix the ones it gets wrong."
FSRS vs SM-2: The Scheduler Actually Matters
When you review a card and the app decides when to show it next, an algorithm is making that call. For roughly two decades, the default in Anki and most spaced-repetition tools was SM-2, an algorithm Piotr Wozniak designed for SuperMemo in the late 1980s. It works, and it's simple, but it's old. It treats every card with a fixed set of multipliers and doesn't really learn the shape of your personal memory.
FSRS, the Free Spaced Repetition Scheduler, is the modern replacement. It's open-source, research-backed, and now built directly into Anki, where you can switch it on in the settings. Instead of applying rigid multipliers, FSRS models three things about each card: how retrievable it is right now, how stable the memory is, and how difficult the card is for you specifically. It then schedules the review for the moment you're predicted to be on the edge of forgetting, which is exactly where retrieval practice does the most good.
The practical payoff is that FSRS tends to show you fewer reviews for the same retention, or higher retention for the same number of reviews. It also lets you set a target retention rate (say, 90%) and optimizes your schedule toward it, adapting as it gathers data on your actual performance.
| SM-2 (the old default) | FSRS (the modern scheduler) | |
|---|---|---|
| Origin | SuperMemo, late 1980s | Open-source, research-backed, 2020s |
| How it schedules | Fixed multipliers per card | Models retrievability, stability, and difficulty |
| Personalization | Minimal | Adapts to your real review history |
| Target retention | Not directly tunable | Set a goal (e.g. 90%) and it optimizes for it |
| Review load | More reviews for same retention | Fewer, better-timed reviews |
| Availability | Legacy default | Built into Anki, opt-in in settings |
If you use Anki and you haven't turned on FSRS, that's the single highest-leverage change you can make today. It takes one toggle. The scheduler won't fix a bad card, but it will make a good card cost you less time.
How AI Flashcard Generators Work
Under the hood, most AI flashcard tools follow the same rough pipeline, and knowing it helps you steer the output.
First, ingestion. The tool takes your source: a PDF, pasted text, a screenshot run through OCR, or a YouTube transcript. The cleaner the source, the better the cards. Garbage in, garbage cards out.
Second, concept extraction. A language model reads the text and identifies what looks like a testable fact, definition, or relationship. This is where the quality lottery begins. The model has no idea what you already know, what your exam covers, or which sentence is a throwaway aside versus a load-bearing concept. It guesses based on text structure and surface salience.
Third, card formatting. The model turns each extracted concept into a question and answer, or a cloze deletion (a sentence with a word blanked out for you to fill in). Many tools default heavily to cloze because it's the easiest format to generate automatically from any sentence.
Fourth, scheduling. The cards drop into a spaced-repetition engine, increasingly FSRS, and the review loop begins.
The thing to internalize is that steps two and three are where AI is weakest. Extraction is a judgment call about importance, and formatting is a craft with real rules. The model is fluent and fast, but fluency is not the same as good pedagogy. That gap is the quality trap.
The Quality Trap: How AI Wrecks Retention
AI-generated cards fail in three predictable ways. Once you can name them, you can spot and fix them in seconds.
1. Recognition instead of recall. The most common failure. AI loves to produce multiple-choice questions or cards where the answer is half-given in the prompt. Recognition is easier, and it feels like you know the material, but it builds far weaker memory than free recall. When you pick the right answer from four options, you might be pattern-matching, not retrieving. A card that asks "Which of these is the capital of Australia?" trains a much shallower memory than "What is the capital of Australia?" Free recall forces the full reconstruction; recognition lets you off the hook.
2. Knowledge dumping. AI will happily stuff an entire paragraph onto one card. "Explain the causes, consequences, and timeline of the French Revolution" is not a flashcard. It's an essay prompt. When a card asks for too much, you'll fail it for the wrong reasons (you got two of five sub-points), the scheduler gets confused about whether you know it, and you dread the review. Big cards are how decks die.
3. Cloze overload. Because cloze deletions are trivial for a model to generate from any sentence, AI tools over-produce them. You end up with sentences riddled with blanks, or clozes that are guessable from context without any real retrieval ("The {{capital}} of France is Paris"). A good cloze hides the one thing worth knowing. A lazy cloze hides a word you'd guess anyway.
There's a quieter fourth problem: AI cards can be subtly wrong or context-stripped. A sentence pulled out of a chapter can lose the qualifier that made it true. You won't notice until you've memorized the wrong version. This is exactly why AI cards need a human pass, and why starting from your own highlights helps: you already read the surrounding context.
The Rules of a Good Card
The canonical guidance here predates AI by decades. Piotr Wozniak, the creator of SuperMemo, wrote "Twenty rules of formulating knowledge," and it remains the best card-quality checklist ever written. Two principles do most of the work.
The minimum information principle. Each card should test the smallest possible piece of knowledge. Not because small is cute, but because small items are easier to schedule, easier to review, and easier to keep or discard. If you half-know a big card, the algorithm can't help you. If you half-know a deck of small cards, the algorithm knows exactly which three to drill.
Atomicity. One card, one fact. If your card has an "and" in the answer, it probably wants to be two cards. Atomic cards review fast, fail cleanly, and give the scheduler honest signal.
Here's what the difference looks like in practice:
| Bad card (what AI tends to make) | Good card (what you want) |
|---|---|
| Q: Describe the testing effect, its discoverers, and its implications for study. | Q: What does the "testing effect" say about retrieval vs restudy? A: Retrieving info from memory strengthens it more than re-reading does. |
| Cloze: The {{forgetting}} curve was mapped by {{Ebbinghaus}} in {{1885}}. (three blanks) | Q: Who first mapped the forgetting curve, and roughly when? A: Ebbinghaus, 1885. |
| MC: FSRS is (a) a scheduler (b) a fruit (c) a country (d) a font | Q: What does FSRS optimize that SM-2 mostly doesn't? A: It models your personal retrievability, stability, and difficulty to time reviews. |
| Q: Everything about spacing. | Q: Does cramming 5 reviews in one night beat spacing them over 5 weeks? A: No. Spaced reviews retain far better (Cepeda et al., 2006). |
A few more rules worth keeping from Wozniak's list: prefer asking for the thing over asking for its description, avoid sets and enumerations (they're the hardest to retrieve), and write cards in your own words. That last one matters with AI especially. A card phrased in the model's generic voice is harder to bond with than one phrased the way you actually think. Editing the wording is half the value of carding. If you want to go deeper on the retrieval side, see our piece on active recall.
A Concrete Workflow: From Highlight to Review
Here's the part that ties it together. The hardest question in flashcards isn't "how do I make a card." It's "what deserves a card at all." AI can't answer that for you, because importance is personal. But you've already answered it, every time you highlighted something while reading.
Highlighting is the encoding step. Carding is the retrieval step. Treating them as one pipeline solves the "what to card" problem for free.
Step 1: Highlight while you read. As you read an article, a paper, or a book, mark the passages that actually matter to you with Glasp's web highlighter. The act of selecting forces an evaluative judgment, which is itself a light form of encoding. You're not carding yet. You're deciding what's worth carding.
Step 2: Export and triage. When you finish a source, export your highlights into one place. Now you have a curated, personal shortlist instead of a 40-page PDF. This is a far better input for an AI generator than the raw document, because you've already done the importance filtering the AI can't do.
Step 3: Generate atomic cards. Feed the highlights into your flashcard tool and ask for atomic question-and-answer cards, not clozes, not multiple choice. A prompt as simple as "turn each highlight into one short recall question with a one-line answer; no multiple choice; split anything with an 'and'" dramatically improves the output.
Step 4: Edit, ruthlessly. This is non-negotiable. Read every card. Delete the redundant ones, split the fat ones, rephrase in your own words, and kill any that test recognition. Expect to throw away a third of what the AI produced. That's normal and good.
Step 5: Schedule with FSRS and review. Drop the survivors into an FSRS-backed scheduler and review when it tells you to. Trust the timing. The whole point is that you don't decide when to review; the forgetting curve does.
The same pipeline works for other sources. Your Kindle highlights become cards from the books you actually read, not a generic deck. A long lecture or talk becomes a tight set of cards when you run it through YouTube Summary and card the key takeaways instead of the whole transcript. And before a review session, you can use Glasp's AI chat to quiz yourself on your own highlights, which is a softer, lower-stakes form of retrieval practice that surfaces the gaps worth carding. We walk through the Kindle path in detail in Kindle AI flashcards.
| Source | Capture step | Best card type |
|---|---|---|
| Articles & papers | Web highlighter | Definitions, key claims |
| Books | Kindle highlights | Concepts, memorable lines |
| Videos & talks | YouTube Summary | Takeaways, frameworks |
| Your own notes | Highlight export | Connections, your phrasing |
Honest Caveats: What Flashcards Are Bad At
Flashcards are a tool, not a religion, and they're genuinely bad at some things. Pretending otherwise is how people waste months.
Flashcards excel at facts, vocabulary, definitions, formulas, and discrete relationships. Anything with a crisp, checkable answer is perfect card material. Learning a language, memorizing anatomy, holding a body of terminology in your head: this is exactly what spaced repetition was built for.
They are much weaker for deep conceptual synthesis. You cannot card your way to understanding why a historical event happened, or how three frameworks relate to a problem you've never seen, or what a novel is really about. Those require connecting ideas, arguing, and writing. A card can hold a building block, but it can't assemble the building. If your subject is mostly synthesis, cards are a supporting actor, not the lead.
The AI caveat is just as important. AI cards need human editing, every time. The model is a fast first-drafter, not an author. It doesn't know what you know, it can strip context and introduce subtle errors, and it defaults to formats that build the weakest memory. The five-minute edit pass is not optional polish. It's where the learning value lives.
And recognition is not recall. If your tool keeps generating multiple-choice cards because they're easy to auto-generate, switch the setting or switch the tool. A deck that feels easy is usually a deck that isn't working. For the bigger picture of retention beyond cards, see how to remember what you read.
Frequently Asked Questions
Are AI-generated flashcards as good as ones I make myself?
Not out of the box. AI is faster and removes the tedium that makes most people quit, but it tends to produce recognition-heavy, overstuffed, cloze-heavy cards. The winning approach is hybrid: let AI draft from your own highlights, then edit ruthlessly. A human-edited AI deck beats both a pure-AI deck and a from-scratch deck you never finished making.
Should I use FSRS or stick with the default SM-2?
Use FSRS. It's built into Anki, free, and research-backed. It models your personal forgetting curve and schedules fewer, better-timed reviews than SM-2, which uses fixed multipliers from the late 1980s. Switching is a single toggle in settings, and it won't disrupt your existing cards. The scheduler can't fix a bad card, but it makes good cards cheaper to maintain.
What's wrong with cloze deletions and multiple choice?
Nothing, in moderation. The problem is overuse. AI over-produces clozes because they're trivial to generate from any sentence, and many end up guessable from context, which means no real retrieval happens. Multiple choice trains recognition, which builds weaker memory than free recall. Use cloze to hide the single thing worth knowing, and prefer open recall questions for anything you want to remember durably.
How many cards should one source produce?
Fewer than you'd think. A dense article might yield five to fifteen genuinely useful cards; most of what you read doesn't need carding at all. This is why starting from your highlights helps: you've already filtered for what matters. If your AI tool spits out 80 cards from one chapter, that's a red flag, not a feature. Delete aggressively.
Can I make cards from books and videos, not just text I paste in?
Yes. That's the modern workflow. Your Kindle highlights become cards from books you actually read, and YouTube Summary turns a long video into a short set of takeaways you can card. The capture tool does the encoding; the flashcard tool does the retrieval. Pulling cards from your own highlights instead of raw sources is what keeps the deck personal and lean.
Conclusion: Make Fewer, Better Cards
AI changed the economics of flashcards. The friction that made most people quit (the slow, tedious manufacture of cards) is gone. That's a real gift, and you should take it. But the gift comes with a trap: the same speed that makes carding effortless also makes it effortless to flood your deck with cards that build recognition instead of recall, dump too much onto one prompt, and quietly drag your retention down while feeling productive.
The science hasn't changed. Roediger and Karpicke's testing effect, Ebbinghaus's forgetting curve, Cepeda's spacing meta-analysis, and Wozniak's twenty rules all still hold. Retrieve, space, and keep each card atomic. FSRS makes the spacing nearly automatic. The atomicity is on you, and on a five-minute edit pass over whatever the AI hands back.
The best raw material for all of this is what you already cared about enough to mark. Capture it while you read with Glasp's web highlighter, export your highlights when you finish, turn the ones that matter into atomic cards, and review on FSRS. Pull from your Kindle highlights and your YouTube Summary takeaways, and use Glasp's AI chat to quiz yourself and find the gaps worth carding. Highlight first. Card second. Edit always. Then let the forgetting curve do the rest.