Getting corrected without killing the conversation
Ask a friend to correct your mistakes and they will do it for about ten minutes. Then they will stop, because correcting someone is unpleasant and they like you.
Every learner has had this conversation. "Please correct me, I want to learn." The other person agrees enthusiastically. They correct three things in the first ten minutes, then stop entirely for the rest of the year.
They are not being lazy. Interrupting someone mid-sentence to fix their grammar is socially expensive, it breaks the flow of what you were talking about, and after the third time it starts to feel unkind. Human beings are very bad at sustaining it, and no amount of asking changes that.
The result is the plateau that a lot of confident intermediate speakers sit on for years: fluent, comprehensible, and quietly repeating the same twelve errors because nobody has been willing to mention them since 2019.
Correction has to be cheap to be sustainable
The problem is not that correction does not work. It is that correction delivered by a person has a social cost, and anything with a social cost gets rationed.
Remove the cost and the frequency changes completely. Software has no feelings about pointing out that you used the wrong particle for the fifth time, and you have no feelings about being told by it. That is a genuine structural advantage, and it is most of the reason automated feedback is useful at all.
What a useful correction looks like
Not all correction is equal, and the common kinds are mostly the weak kinds.
"That's wrong" is useless. You knew.
A grammar rule is better but arrives in the wrong format — abstract, general, and disconnected from the sentence you were trying to build. You will nod, and you will not apply it.
A better version of your own sentence is the useful one. Your sentence and the fixed sentence, side by side, differing in one or two places. The difference is visible rather than described, and it is a difference in something you personally produced thirty seconds ago and cared about.
That is what Amiko shows: your line, and under it the natural version, with the changed words picked out so you can see exactly where the two diverge. Not a rule — a diff.
Why the "why" should be optional
There is a real tension here. Explanations are valuable, and explanations also stop conversations dead. A paragraph about verb transitivity in the middle of ordering coffee ends the coffee-ordering.
The resolution is to make the explanation available but not automatic. The correction appears inline, quietly, and does not demand anything. If you want to know why, you tap and get the reason. If you are mid-flow, you register the shape of the fix and keep going.
Most of the time, noticing the difference is enough — you see that ています was wanted rather than てます, and you file it. Occasionally the difference is baffling, and then the explanation earns its interruption.
Writing in your own language is allowed
One under-used move: when you genuinely cannot say something, say it in your own language anyway.
You will get back the target-language version of exactly what you meant — not a generic phrase, but your sentence, the one you cared enough to want. That is one of the highest-value pieces of input available, and it costs you nothing but the admission that you were stuck.
People avoid this because it feels like giving up. It is the opposite: it converts a moment where you would have said nothing into a personalised model answer.
What to do with the correction
Seeing it is not learning it. The step almost everyone skips is the one that works: use the corrected version, on purpose, within the next few turns.
Say the fixed sentence back. Reuse the pattern two exchanges later. The correction has to pass through your own production once, or it stays in the same passive pile as everything else you have understood and cannot say.
You can see the whole loop running in the browser — write something slightly wrong on purpose and watch what comes back under your own message.
