Spaced repetition is one of the few things in learning with genuinely strong evidence behind it. Testing yourself beats rereading; spacing the tests out beats massing them; expanding the interval as an item gets stronger is more efficient than a fixed one. None of that is controversial and none of it is what this post is arguing with.
The argument is with a specific implementation detail that nearly every flashcard tool shares, and that quietly breaks the whole mechanism: you grade yourself.
What a self-graded card actually measures
The standard loop is: see the front, try to recall, reveal the back, then press one of four buttons saying how well you did.
Consider what happens at the reveal. You see the answer, and the instant you see it you recognise it. It feels familiar. It feels like something you knew. And because the sensation of "I knew that" is indistinguishable from the sensation of "I recognise that now I am looking at it", you press Good, and the scheduler pushes the card out a week.
This is not carelessness. It is a well-documented effect — the feeling of knowing is a poor guide to whether you could actually have produced the answer, and it is systematically inflated by having just been shown it. Every person using a flashcard app is running on a signal that is biased in one direction, and the scheduler has no way to know.
The consequences compound, because the scheduler's whole job is to act on that signal:
- Words you half-know get long intervals. They were "fine", so they go out a week, then three, then two months. You never actually learn them, and they stay permanently at the level where you recognise them on a card and miss them in speech.
- Your review load stops reflecting reality. The deck says eighty per cent mature. Your listening says otherwise.
- The failure is invisible. There is no moment where the system tells you the grade was wrong, because the system's only source of truth was you.
The deeper problem is that recognising a word on the back of a card is not the skill you want. You want to hear it in a sentence at speed and understand it without noticing you did. Self-grading measures the wrong thing, and then measures it optimistically.
The fix is not more discipline
The usual advice is to be harder on yourself — press Again if you hesitated at all, be honest, hold a high bar. It works for a while, for conscientious people, and it degrades in exactly the situation where it matters: at the end of a long session, when you are tired and there are forty cards left.
A system whose accuracy depends on the user's self-discipline at their least disciplined moment is not a system. It is an intention.
The actual fix is to stop asking. If the question has a right answer, the software can grade it, and then the grade is a fact rather than a self-report.
What "the software grades it" has to mean to be worth anything
Auto-grading is easy to do badly. A multiple-choice question with one plausible answer and three absurd ones is worse than self-grading, because it produces a confident wrong signal — you answered it right by elimination and the scheduler now believes you know the word.
To be honest, a generated question needs the distractors to be real competition:
Matched on part of speech. If the answer is a verb and the three wrong options are nouns, the question is a grammar puzzle you can solve without knowing the word.
Matched on frequency. If the answer is a common word and the distractors are obscure ones, the question rewards recognising which option looks familiar. Distractors have to be words you are about as likely to know.
Drawn from real material. The strongest question format is the sentence you actually heard with one word removed, and four candidates to fill the gap — because that tests the word in the context the word lives in, rather than as a dictionary entry.
And there is a fourth rule, which is the one that costs something: where four honest options do not exist, do not ask. Early in a language, or for a word with no close neighbours, there may be no set of distractors that is fair. A tool that always produces a question will produce a bad one, and a bad question is worse than none because it writes a false grade into the schedule.
We enforce all four. The questions are built server-side from the catalogue, matched on part of speech and on frequency band, and where no fair question can be built, the word keeps the old style of card and the app says so rather than pretending.
The format should climb with the word
There is a second thing self-graded decks get wrong, which is that a word is one card forever.
Meeting a word for the first time and having known a word for three months are different states and deserve different questions. A ladder we have found works:
- First meeting: recognition. You see the word, you pick the meaning from four. This is the easy direction and it is the right one to start with, because failing at the hard direction on day one teaches nothing.
- Once recalled: production in context. You see the line you heard it in, with the word taken out, and four words to fill the gap. Harder, and much closer to what listening asks of you.
- Later: by ear. You hear it and answer. This is where a word stops being vocabulary and becomes language.
The difficulty climbing with the item is the part that is missing from most decks, where a card's front and back are fixed at creation and the only thing that changes is the interval.
And the deck should not be only the words you noticed
One last failure, related and worth naming.
If your deck is built from the words you chose to save, it is built from the words that stood out. Words stand out when they are unusual. So a self-assembled deck is systematically biased towards rare vocabulary, while the words carrying most of the language sit at half-known and never get reviewed, because they never felt like words worth saving.
The fix is to let something else choose as well. Ours asks in frequency order off a corpus built from our own transcripts — commonest words first — and a word you miss there enters your deck by itself, whether or not you would have noticed it. The words you tap are still in there. They are just no longer the only source. The reasoning behind the frequency ordering is in how many words you need to understand a language.
What to take from this if you use Anki
Nothing here says to stop. Anki is a superb piece of software and the scheduler is excellent. Three adjustments make the signal into it far better:
Use cloze deletions from real sentences rather than word-to-definition cards. It is the single biggest improvement available, and it is a format change, not extra work.
Grade on production, not recognition. Decide, before you reveal, whether you could have produced the answer. If you hesitated, it is Again. The honest version of this is brutal and it is also the version that works.
Add a listening card for anything you care about hearing. Audio on the front, nothing else. It is a different skill and a deck full of written cards does not train it.
And if you want the general form of the rule: a grade should be a fact, never a self-report. Every place your learning system asks how you did instead of checking is a place where the data is optimistic, and optimistic data produces a schedule that is wrong in the one direction you cannot detect.