An app has to mark your answer without a person present, so every exercise it contains must have a small closed set of correct responses. That constraint is not a shortcut anybody took. It is the defining property of the format, and everything an app is good at and everything it cannot do follows from it directly.
Which means the boundary is predictable rather than mysterious. Anything with one right answer is territory an app can own completely. Anything where several answers are correct and the difference between them is judgement is territory no app can enter, however much effort goes into it.
Eight honest facts about apps
- an app is a habit tool, and that is a real contribution. Daily contact is genuinely valuable and apps are unusually good at producing it.
- the streak is the product. The design goal is retention of the user, which overlaps with retention of vocabulary and is not the same thing.
- gamification substitutes for progress. The feeling of advancement is manufactured and can persist for months without a corresponding rise in ability.
- apps rarely pass B1. They are built for the levels where the material is finite and gradable, and there is little above that.
- they do not teach speaking. Speech recognition scores pronunciation crudely and does not produce conversation.
- use one for vocabulary only, and be explicit with yourself that this is the job it is doing.
- fifteen minutes is the useful dose. Beyond that the return falls sharply against reading or listening to real material.
- pair an app with real material from the first week, so the app is the warm up rather than the session.
What that buys, and it is real
Recognition has one right answer, so apps drill it well. A writing system has one right answer, so an app is a genuinely excellent way to learn an alphabet or a syllabary. The first thousand words have one right answer each, and the frequency ordering that a good app uses is better than the ordering most learners would choose for themselves.
They also solve the problem nobody talks about, which is starting. A daily habit at a time when the habit is the difficulty, with a low enough barrier that it survives a bad week, is worth more than a better method that gets abandoned. That is not a small contribution and it is worth saying before the rest of this page.
Where the format ends
Anything open-ended. Whether your sentence is one an English speaker would produce. Whether a word fits the register of the room you are in. Whether the thing you said answered the question that was asked. All of those have many acceptable answers and a large space of near misses, and marking them requires judgement rather than comparison.
That is also why an app can correct the answer it asked for and never the sentence you were trying to build. It never saw the sentence you were trying to build, because asking for one would have produced something it could not score.
Where it stops: the streak
The stall is the most successful retention mechanic ever built for a study product. A streak measures attendance and nothing else, and it is extremely good at what it does, which is keeping you in the format.
The damage is not that streaks are manipulative. It is that they keep somebody in a format long after the format has stopped paying, because the number is the only measurement available and it keeps going up. A learner with three hundred consecutive days has genuinely done something and has no way to tell whether the last two hundred of them were worth anything.
The test is separate from the app and takes five minutes: say something out loud that you have not said before, or read a paragraph nobody wrote for learners. If those have not moved in three months, the streak is measuring attendance at something that has finished.
The one part of an app that behaves like retrieval
The review or practice mode, as distinct from the lesson path, is the part worth using deliberately. It brings back what you got wrong instead of advancing you, which is the only behaviour in the whole product that resembles retrieval practice rather than exposure.
It replaced doing the next lesson because it is next, which is what the path is designed to make you do and which optimises for progression rather than for retention.
Its limitation is the constraint at the top of this page. It is still bounded by what can be marked automatically, so it drills the closed-answer half harder and does not extend the range by a single item.
What it looks like when it goes wrong
The recognisable failure is a long streak and no speech, which is not a failure of effort and is a failure of format. The exercises were done, the answers were correct, and nothing in a year of them ever asked for an unpredictable sentence.
The second is a learner who tries a different app when the first one stops working. All of them share the constraint, so the second one has the same ceiling in the same place, and the six weeks of renewed progress at the start are the material already known.
What an app cannot catch, and how to read a stall as a fact about a method rather than about yourself, is the stalls every learner hits, and the way through each, and the thing that has to be added when the closed-answer half is finished is finding someone to speak to, and what to do in the hour.
The open question is what to do with the daily habit an app was carrying, because that part was genuinely valuable and dropping the app usually drops it too: streaks, restarts and what to do after a bad month.