memori
The short answer
Run seven checks at roughly ten seconds a card: true, atomic, answerable cold, not self-answering, gradable, worth knowing, not already known. Expect to delete a third to a half of any batch. Most deletions won't be errors — they'll be accurate cards covering things that never needed memorising.
The framing that makes this efficient: a card is not a piece of content, it's a ten-year subscription. Each one you keep costs roughly ten retrievals spread across its life. Judged as content, a mediocre card is harmless. Judged as a subscription, it's a recurring charge for something you didn't want.
The seven-check list
In priority order. The first three eliminate most of what should go.
1. Is it true?
Check against the source you generated from, not against your memory. Numbers, units, doses, dates and named entities are where errors concentrate — and they're also where errors matter most, because a wrong number is indistinguishable from a right one once you've rehearsed it fifty times.
2. Is it one thing?
Scan for "and", "or", commas separating list items, and any answer over about twenty words. A card testing two facts fails whenever either half slips, so you re-review the half you already knew indefinitely. This is the fastest route to a leech and the most common structural defect in generated output.
3. Can I answer it in six months, cold?
Models write cards with the source document in mind, so they leave out context that was on the screen at the time. "What is the second step?" is answerable while you're looking at the chapter and meaningless in nine weeks. If a card only makes sense in a sequence, it needs the context inside it.
4. Does the question give away its own answer?
Over-specified questions are the signature failure of generated cards. Every extra qualifier is a clue, and after twenty reviews you're pattern-matching the sentence rather than retrieving the fact — which is recognition wearing recall's clothes.
5. Can I grade it unambiguously?
When you flip the card, can you say cleanly whether you got it? If the answer has five components and you produced three, there's no honest button to press — and inconsistent grading degrades the scheduler's model of the card, which then schedules it badly.
6. Do I want to know this?
The trivia filter, and the one that removes the most cards. Models can't distinguish importance from presence: if a fact is in the source, it's a candidate. You have to supply the judgement about what's worth a decade of reviews.
7. Do I already know it?
Cards for things you know cold are pure cost. Generated batches are full of them, because the model doesn't know what you know. Delete on sight — you can always add it back if you turn out to be wrong.
Checks 6 and 7 are last in the list but they remove the most cards, so in practice it's faster to run them first on a quick pass, then apply checks 1–5 to whatever survives. Verifying the accuracy of a card you're about to delete is wasted effort, and it's the reason people find the review step tedious enough to skip.
Four failures specific to AI cards
Human-written cards fail too, but they fail differently. These four are characteristic of generated output, and knowing the shapes makes them recognisable at a glance.
Fluent but hollow
What is the significance of the Treaty of Westphalia?
It marked a turning point in European history by establishing important principles that shaped the modern international system.
Which principle of state sovereignty is conventionally dated to the Peace of Westphalia?
That states have exclusive authority within their own territory — no external interference in internal affairs.
The left card reads well and contains nothing. It'll be graded correct every time on a vague gesture at the right area, which means it consumes reviews and teaches nothing. This is the failure hardest to spot, because fluency is exactly what models are best at and what readers use as a quality signal.
Uniform difficulty
Generated batches tend to be flat — every card pitched at the same level, usually the level of a definition. A real subject has facts of wildly different importance and difficulty, and a good deck reflects that. A twenty-card batch where every card is "what is X?" has captured the vocabulary of a topic and none of its structure.
The fix is at generation time: ask explicitly for a mix — a few definitions, several relationships between concepts, several applications, and the specific things that are easy to confuse with each other.
Interference blindness
What does the accessory nerve innervate?
Sternocleidomastoid and trapezius.
Which cranial nerve innervates trapezius — and which nearby nerve is it most often confused with?
CN XI (accessory). Confused with CN X (vagus), which exits through the same jugular foramen.
A model generating cards one at a time has no view of your other cards, so it produces sets of near-identical items that then compete in memory. You end up knowing that one of four similar answers is right and being unable to say which — a phenomenon that feels like the deck failing and is actually the cards colliding. Human deck-builders learn to write the distinction into the card; generators don't, unless told.
Source drift
Ask for cards from your notes and you'll often get cards from the model's general knowledge, blended in seamlessly. Usually the additions are correct. Sometimes they contradict what your course actually taught — a different classification system, a different threshold value, a convention your examiner doesn't use. Being right in general and wrong for your exam is a genuinely bad outcome, and it's invisible unless you check every card against the source.
Verifying material you don't know yet
The obvious problem with "always verify AI output" is that you're least able to verify precisely when you most need the cards. The workable answer isn't more diligence — it's changing what you verify against. Constrain generation to a source you trust, then check cards against that source rather than against your own knowledge.
This turns an expertise problem into a matching problem, and matching is something you can do on day one of a subject. If a card asserts something your lecture notes don't contain, delete it — without adjudicating whether it's true. You're not trying to determine correctness; you're enforcing provenance.
Three practices that make this work:
- Always generate from supplied material, never from a topic name. "Make cards about the Krebs cycle" invites the model to draw on everything it knows. "Make cards from these notes" gives you something to check against.
- Ask for the source line with each card, at least while you're calibrating a new tool or prompt. Once you've seen that a given setup stays faithful across two or three batches, you can drop it.
- Ask what it left out. "After the cards, list anything important in the notes you couldn't turn into a good card." The gaps tell you where your notes are thin, which is more useful than the cards themselves.
For low-stakes learning — a language, a hobby, general interest — this is enough. For medicine, law, engineering, or anything where a confidently wrong card could matter, generating from anything other than an authoritative source you supplied isn't a shortcut worth taking.
What hallucination actually looks like here
Worth being precise, because the popular image is wrong. Fabricated cards are rarely absurd. In practice they look like:
- Plausible numbers. A threshold of 5.5 where the real figure is 5.0. The most dangerous category, because nothing about the card looks wrong.
- Correct-but-stale conventions. A classification that was standard ten years ago and has since been revised.
- Over-clean causal claims. Real mechanisms have exceptions and disputed steps; generated cards tend to state them as settled.
- Attributions. The right idea assigned to the wrong researcher, or the right researcher with the wrong year.
What makes this worse in flashcards than in prose is the rehearsal. A wrong sentence in an essay you read once is a passing error. A wrong card is a wrong fact you will deliberately drill until it's unshakeable — and it will then be harder to correct than if you'd never learned it, because you'll have to overwrite an existing memory rather than form a new one.
That asymmetry is the actual argument for the review step. Not that errors are frequent, but that this particular medium amplifies the ones that get through.
Fixing a batch instead of a card
Most quality problems in generated output are systematic rather than individual — the model applied one wrong assumption twenty times. So diagnose at the batch level first:
| Pattern across the batch | Cause | Fix |
|---|---|---|
| Every answer is a paragraph | No length constraint | Reformat all: answers under 15 words |
| Every question starts "What is…" | No variety instruction | Reformat: mix definitions, relationships, applications |
| Questions contain their answers | Over-specification | Reformat: strip qualifiers from questions |
| Vocabulary lacks context | Word-level extraction | Reformat: add an example sentence to each answer |
| Too many cards | No budget | Delete, don't reformat — regenerate with a maximum |
| Scattered factual errors | Unconstrained source | Regenerate from the source only. Don't patch. |
The last two rows are the important distinction: formatting problems are worth fixing, content problems are worth regenerating. Patching a batch that drifted from its source leaves you unsure which cards you actually checked.
Batch-level fixes are tedious in most tools, which is why people don't do them. In Memori this is Magic Reformat — select a group of cards, give one instruction ("shorten every answer", "add an example sentence"), and preview the before-and-after for all of them before committing. And because generated cards are shown selectable and editable before they're saved, most of this checklist happens at creation rather than as a cleanup job later.
Auditing a deck you already built
If you have 400 cards of uncertain quality, don't review them all. Let the data find the bad ones.
- Sort by lapse count. Cards you've failed four or more times are your problem set — typically 5–10% of a deck accounting for a third of the pain. Nearly all of them are defective rather than difficult.
- For each, ask: is this hard, or is it broken? Broken looks like two facts in one, a question that doesn't specify what kind of answer it wants, or a near-duplicate of another card. Rewrite or delete; don't just keep reviewing.
- Sort by answer length. The longest 5% are almost all compound cards. Split them.
- Look for clusters from one source. If eight cards from one lecture are all failing, the cards aren't the problem — you didn't understand that lecture. Go back to the material.
- Check your true retention. Below about 80% with sensible settings, the deck has a quality problem rather than a scheduling one. No amount of interval tuning fixes cards you can't answer.
An hour spent on the worst 5% of a deck usually buys more than a month of extra reviewing.
When not to use generated cards at all
Cases where writing them yourself is genuinely better:
- Material you're actively struggling with. The act of deciding what the question is is the studying. Automating it removes the part you needed.
- Anything with no reliable source to constrain against. Cutting-edge, contested, or highly local material — your institution's specific protocol, a lecturer's idiosyncratic framing.
- Very small sets. Twelve cards is ten minutes by hand. The overhead of prompting, generating and reviewing isn't worth it.
- Personal or contextual material. Names of colleagues, your own reference numbers, things where you are the source.
- When you're using generation to avoid the reading. If cards are being made from material you haven't actually read, no checklist will save the result. That's the one failure mode this page can't fix.
Frequently asked questions
Are AI-generated flashcards any good?
Good at phrasing, bad at judgement. A model reliably produces a grammatical question and answer from a source; it doesn't reliably decide which facts are worth remembering or notice when a card tests recognition rather than recall. Ten seconds of review per card closes most of the gap.
What's the biggest problem with AI-generated flashcards?
Not hallucination. The dominant failure is cards that are entirely true but unlearnable — too long to grade, two facts in one, or answerable from the phrasing of the question. These pass a factual check and quietly waste hundreds of reviews.
How do I check accuracy in a subject I don't know?
Constrain generation to a source you trust, then verify against that source rather than your knowledge. If a card asserts something the source doesn't contain, delete it without adjudicating whether it's true. That turns an expertise problem into a matching problem.
How many cards should I expect to delete?
A third to a half of a typical batch, and deleting that many means the checklist is working. Most deletions aren't errors — they're accurate cards covering material that didn't need memorising.
Can I trust AI flashcards for medicine or law?
Only when generated from and checked against an authoritative source you supply. In those fields the failure mode is a confident, plausible, subtly wrong card that you then rehearse until it's unshakeable. Restricting generation to your own materials isn't optional there.
How do I fix a large batch at once?
Work out whether the problem is systematic or per-card. Systematic issues — every answer too long, every question over-specified — can be fixed across the batch with one reformatting instruction. Cards that are individually wrong should be deleted and regenerated, not patched.
Does it matter which model wrote the cards?
Less than the prompt and the source. A weaker model with a tightly constrained prompt and good source material outperforms a stronger one told to "make flashcards about X" — because the dominant failure is judgement about what to include, not fluency. The checklist is the same either way.
Should I edit a mediocre card or delete it?
Delete, unless the underlying fact is one you specifically want. Editing feels thriftier and usually isn't: the card was cheap to generate and the edit costs you more attention than regenerating. Reserve editing for cards testing something you genuinely need, where only the phrasing is wrong.
Where this comes from
- Woźniak, P. Effective learning: Twenty rules of formulating knowledge. The minimum information principle, and the case against complex and enumerated cards.
- Nielsen, M. (2018). Augmenting Long-term Memory. On cards as a long-term commitment, and on the practice of continually rewriting the ones that fail.
- Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. For the distinction between recognition and recall underlying check 4.
- Dunlosky, J., et al. (2013). Improving students' learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4–58. On practice testing, and on why fluency is a poor proxy for learning.
- Bjork, R. A., & Bjork, E. L. (1992). A new theory of disuse and an old theory of stimulus fluctuation. Why a card that's too easy to answer does very little.
- The interference discussion draws on the standard retroactive- and proactive-interference literature; the practical framing — write the distinction into the card — comes from the Anki community's handling of leeches rather than from a specific study.
