Is multiple choice good for learning vocabulary?

It is the easiest exercise, which is not the same thing. Across 68,376 answers from 790 learners, picking a word from options was correct 97.7% of the time in a median of 2.7 seconds, while typing the same kind of word from memory was 92.4% correct and took 4.2 seconds. Recognition is useful for meeting a word, but a high score on it is not evidence you can produce it.

Why does recall feel harder than recognition?

Because it is, on both measures at once. Producing a word from memory is answered correctly less often than picking it from options and takes longer per answer, so a session of it feels slower and more error prone. Choosing from options only requires eliminating the wrong answers, and the right one is on screen throughout.

How long does it take to answer a Japanese vocabulary question?

Hear it and answer: a median of 3.4 seconds. Fill the gap in a sentence: a median of 5.5 seconds. Pick it from options: a median of 2.7 seconds. Type it from memory: a median of 4.2 seconds

Kumi data report

Multiple choice feels like progress because it is easy

Across 68,376 answers from 790 learners, picking a word from options is answered correctly 97.7% of the time in a median of 2.7 seconds. Typing the same kind of word from memory is 92.4% correct and takes 1.6 times longer.

Behavioural data from Kumi's learners, aggregated. Sample sizes are on every row.

Is recognising a word the same as knowing it?

No, and the gap is measurable in two directions at once. Recognition is more accurate, and it is faster. Both of those make it feel better to do, and neither makes it a stronger test of memory.

The reason is structural rather than motivational. Choosing from options only requires you to eliminate the wrong ones, and the answer is on the screen the whole time. Producing a word requires retrieving it with nothing to lean on, which is the thing you actually need when you are reading or speaking.

How often each kind of answer is right, and how long it takes

The exercise asks you to	Answers	Correct	Median time
Hear it and answer	5,939	99.4%	3.4s
Fill the gap in a sentence	2,691	97.7%	5.5s
Pick it from options	27,603	97.7%	2.7s
Type it from memory	14,335	92.4%	4.2s

Why does harder practice feel like it is going worse?

Because on any single day, it is. Producing from memory has a lower success rate and a longer pause before each answer, so a session of it feels slower and more error prone than the same time spent on multiple choice. The signal you get in the moment runs opposite to how much the practice is worth.

This is the well-documented gap between how well practice feels like it is going and how much of it survives to next week. What these numbers add is the size of the comfort difference inside one app, on the same words, for the same learners.

What to do with this

Not "never use multiple choice". Recognition is how you meet a word for the first time, and it is the right tool early. The trap is staying there because the success rate is pleasant, and mistaking a high score on the easy exercise for knowing the word.

It is also why Kumi does not let a word count as learned on recognition alone. Each way of knowing a word is tracked separately, so picking it from four options never quietly stands in for being able to produce it.

Method

Drawn from 68,376 answers by 790 learners across 2,472 concepts. Every figure is an aggregate, nothing is per learner, and the smallest group is thousands of answers.
Only exercise types whose accuracy is stable month over month are included. Speaking is excluded because a defect drove its skip rate from 13% to 80% between May and July 2026, and writing is excluded because its accuracy drifts downward over the same window without a confirmed cause. Publishing either would describe our releases rather than learners.
Skipped and abandoned answers are excluded rather than counted as wrong, since declining an exercise is a different behaviour from failing it.
Times are medians, not means, because answer times have a long tail from tabs left open. Answers faster than 0.2s or slower than 120s are dropped as not representing a real attempt.
These are Kumi's learners rather than a random sample of Japanese learners. Computed live each day; last generated 2026-08-22.

Kumi tracks recognition, recall, listening and production separately for every word, because being able to do one of them is not evidence you can do the others.