Every language app makes the same promise, and it is a good one: this will fit into your day. Babbel says fifteen minutes. Duolingo built its whole lesson design around five. Busuu is the most honest about it, because it lets you pick a daily goal from a menu, and on that menu ten minutes a day is labelled regular while twenty-five minutes is labelled intensive.

Nobody is lying. Those are the doses the products are designed around. The trouble only starts when you hold them up against the size of the job, and almost nobody does, because the numbers that get repeated about language learning are repeated far more often than they are checked.

So let's check them.

Somebody did the arithmetic, and it is brutal

In January 2023 the German computing magazine c't ran a cover story on this by Nico Jurran, called Sprachfindung. Three years on, nothing better has been written about why language apps stall. Jurran is one of those rare technology journalists who refuses to review a product on its own terms. Everyone else was asking whether the lessons were fun. He asked the only question that matters, which is whether the thing can actually deliver what it sells, and then he did the work to find out.

His cleverest move was just arithmetic, which is what makes it so good. He took the hour counts the US Foreign Service Institute uses for Spanish, Italian and French, converted them into fifteen-minute app sessions, and got more than 1,900 of them. At one session a day, that is over five years. The advertised dose is not a bit optimistic. It is out by an order of magnitude. A whole industry had been quoting those FSI hours for years without anybody bothering to divide.

Then, instead of stopping at the headline number, he went through where the hours you do put in fail to land. This is the part of the piece that shows how carefully he had looked, and if you have ever finished a course and still frozen at a dinner table, it will feel personal:

  • You are picking, not producing. Most app exercises hand you the words, as tiles to drag or a list to choose from, because that is what fits a phone screen. Recognising a word is far easier than producing it, and producing it is what speaking is. Nobody stands next to you in a bakery holding the correct forms on little tiles.
  • Everything comes in single sentences. So you never build the thing listening actually needs, which is the knack of holding a long stretch of speech and sorting the important bits from the filler. That is the difference between following a conversation and going blank at the first gap.
  • The audio is too clean. Same voices, same tempo, studio quiet. Real ears need different speakers, different speeds and a bit of background noise.
  • Speaking is barely trained. Repeating a sentence you were just given, at a speech recogniser, trains your pronunciation and nothing else.
  • You almost never write anything freely. Not because it would not help, but because a machine cannot mark it and a competent human is expensive.
  • There is nobody to ask. So a structure that did not land stays not landed.

He also did the thing good reporters do and went to somebody who studies this for a living, rather than to another app's press office. The expert in the piece is Dr Nicola Würffel, professor of German as a Foreign Language at the Herder Institute in Leipzig. She called the small-task format unterkomplex, under-complex, for the harder skills: fine for drilling forms, not enough for the deeper processing that free use of a language needs. Her summary of how that feels is the sentence every stalled learner recognises. I am apparently learning successfully, and I still arrive at no active, free use of the language.

That was three years ago. The rest of this piece is us checking whether it still holds, which is the highest compliment you can pay a piece of journalism: it is worth going back to.

600 hours, and what that number actually means

The Foreign Service Institute figures get quoted constantly, usually second hand and usually without the conditions that make them mean anything.

Here they are properly. The FSI sorts languages by how long they take an English speaker. Category I, which includes French, Spanish, Romanian and Dutch, is put at 24 to 30 weeks, or 600 to 750 class hours. German sits a category above that, at around 30 weeks. The target is roughly B2 or C1.

Now the conditions, which matter more than the number. Those are class hours, with homework on top. The students are adults chosen partly for aptitude, studying full time, in small groups, with a professional teacher who answers questions as they come up and corrects mistakes before they set. If you are learning in the evenings after work, on your own, you do not get to use the FSI figure. Read it as a floor.

Turn 600 hours into fifteen-minute sessions and you get 2,400 of them, which at one a day is six and a half years. Set Busuu to intensive, twenty-five minutes, and you are still looking at 1,440 days.

This is not meant to put you off, and it is not a reason to stop. It is a reason to throw out one particular feeling. If you finished a course, held a long streak and still could not say what you meant at a dinner table, you were not lacking discipline or talent. You did a small fraction of the work while an interface told you every day that you were on track. The gap you felt is arithmetic, not aptitude.

The word counts everyone repeats are made up

This is where checking pays off, and where we part company with our own source on one point. Jurran got the hard part right and the easy part wrong, which is the usual way round, because the wrong part is the bit everybody repeats.

The c't feature reproduced one of those CEFR tables with a vocabulary size next to each level: around 500 to 700 words at A1, 1,500 at A2, 2,500 at B1, 4,000 at B2, 8,000 at C1, 16,000 at C2. You have seen versions of it on language blogs, in course marketing, probably in your own inbox.

The CEFR contains no vocabulary sizes. It describes what a learner can do at each level, not how many words they carry. Every number in that table is somebody's later guess, and the guesses disagree with each other by thousands. It is a table that has been passed around for so long that it now looks like a source, which is exactly how this sort of thing survives.

The real research says something different, and more useful.

James Milton measured vocabulary size across the CEFR levels and found learners needing roughly 1,500 words to reach A2, perhaps another 1,500 for B1, and around 3,500 at B2. His conclusion: you need over 5,000 words for anything like fluency in a European language. His test only samples the most frequent 5,000, so the upper levels are exactly where the instrument runs out of road.

Paul Nation came at it from the other end. Instead of asking what learners have, he asked what texts demand. Using a large corpus of English, he worked out how big a vocabulary you need to cover 98 percent of a text, which is the point at which reading starts working without constant help. The answer: 8,000 to 9,000 word families for written text, and 6,000 to 7,000 for speech. Novels came out at 8,000 to 9,000. Newspapers at 8,000 plus names. Unscripted conversation and talk radio at 6,000 to 7,000.

Three honest caveats. Milton and Nation count in slightly different units, so you cannot lay the two sets of figures side by side. Both did their work on English, so the exact numbers belong to English and carry over to other languages in shape rather than in detail. And 98 percent coverage still leaves you one unknown word in fifty, which is roughly one every four or five lines of a paperback.

But look at the shape, because both lines draw the same one, and it is not the shape in the popular table. The early steps are small and shared. A thousand or so words each, the same words for everybody, which is exactly the range one curriculum can serve millions of people with. Then the amount you need to actually operate in the language, measured from either end, sits somewhere between five and nine thousand. Which means the distance from B1 to comfortable is not one more step the size of the last one. It is several times everything you have done so far.

And that is worse news for course apps than the myth was, not better. The popular table's 16,000 at C2 is easy to wave away as a translator's problem. Nation's 8,000 to 9,000 is the number for reading a novel or a newspaper, which is a thing ordinary people actually want to do, and it sits far above the point where any course still has a sensible order to teach things in. Correcting Jurran's one weak number does not dent his argument, then. It sharpens it.

What has changed since 2023, and what has not

It would be unfair to treat a 2023 piece as current, and one of those six gaps has genuinely narrowed.

Speaking has an answer now. Duolingo shipped Video Call, an AI conversation partner, rolled it out on Android in January 2025 and has added languages since. It sits behind the Max tier, and the rest of the category has followed with something similar. Luis von Ahn's framing of it is fair: this is the kind of practice that used to mean travelling somewhere or paying a tutor. For the specific problem of getting your mouth moving without anyone judging you, it works, and it is there at three in the morning. The nobody-to-ask gap has narrowed too, since a model will explain the subjunctive whenever you ask, with the usual caveat that it will sound equally confident when it is wrong.

What has not changed is anything to do with volume. An AI partner does not turn 600 hours into 60. It does not give you a noisy recording with six people talking over each other, it does not make you read a novel, and it does not work out which five thousand words are yours. Duolingo rebuilt itself around AI in 2025 and reached 58.7 million daily users by the second quarter of 2026. None of that touches the arithmetic.

There is also a pattern worth naming. Every serious fix the providers have shipped, live lessons, podcasts, longer video, conversation partners, is something that is not the core app loop. Read that as an admission and it is an honest one: this works, and it does not work alone.

So what actually works

Würffel's own advice fits in a sentence: be clear about what you want to learn, choose accordingly, and combine. Here is a concrete version, roughly in the order the parts start to matter.

1. Use the course app for the beginner range, and stop waiting for it to level up. Through A1 and A2 it is doing the thing it is genuinely good at: forms, structures, your first fifteen hundred words, and the habit. Just do not read a finished tree as a finish line.

2. Switch from picking to producing as early as you can stand it. Turn off word banks wherever the setting exists. Type instead of tapping. Use the web version, which usually asks for longer answers than the phone one. Any exercise that hands you the pieces is one you will pass without learning much.

3. Add long, messy input, chosen because it interests you rather than because it matches your level. Podcasts with transcripts, graded readers, bilingual editions, radio, YouTube, films you already know by heart. Two traps to watch. Subtitles quietly turn a listening exercise into a reading exercise. And clean studio audio will never prepare you for a station announcement, so go looking for recordings with several speakers and some noise in them.

4. Get one human into the loop, sooner than feels comfortable. A tandem partner, an italki or Preply tutor, one of the providers' own group lessons. An AI partner is an excellent daily rep and a poor substitute for the small social risk that makes speaking stick. A human is also the person you can ask.

5. Own your vocabulary, and start long before you think you need to. Given the numbers above, this is the part that decides whether the other four add up to anything.

Where we fit, and where we do not

That fifth part is what Vokabulo is for. The words come from your own life instead of a syllabus: the phrase a colleague used in a meeting, the sign you walk past every morning, the clause in the rental contract you only half understood, saved with the situation you met it in. Spaced repetition brings them back on a schedule rather than a streak, you type your answers instead of picking them off tiles, and usage labels tell you whether a phrase belongs in a message to a friend or an email to a client.

The reason it works that way is the shape in the research. The first thousand or two words are shared, and a course should hand them to you. The five or six thousand after that are shared with nobody, because they are made of your job, your street, your family and your paperwork. No curriculum can pick them, because nobody knows what they are except you.

And the honest limits. Vokabulo will not train your ear on a platform announcement, it will not correct the long email you wrote, and it will not answer your question about the subjunctive. Those need input, a human and a teacher, which is exactly the combination Würffel was arguing for.

The bit that sounds harsh and is actually a relief

Jurran ended his piece with a line that reads like bad news: learning a language completely with apps is practically impossible. It is the sort of sentence a magazine that lives off technology advertising could easily have softened, and he did not soften it.

It is a relief, though, because it moves the failure out of your character. You were handed a tool that is excellent at one slice of the problem, along with an implied promise that it covered the whole thing.

And the good news is real. There is more target-language listening and reading freely available than at any point in history, there are people to talk to in every time zone, and there is a conversation partner in your pocket at any hour. The work is still six hundred hours. But they are far better hours than they used to be, and you get to choose what fills them.

More on the vocabulary half in why Duolingo doesn't stick, on the two products' different bets in Duolingo vs Vokabulo, and on the level where all of this comes to a head in breaking the intermediate plateau.

Sources

  • Nico Jurran, Sprachfindung, c't Magazin 1/2023, page 60, including the interview with Dr Nicola Würffel of the Herder Institute, Leipzig. If you read German, read it. This article would not exist without it.
  • Foreign Service Institute, foreign language training category hours.
  • I.S.P. Nation (2006). How large a vocabulary is needed for reading and listening? The Canadian Modern Language Review, 63(1), 59-82.
  • James Milton, The development of vocabulary breadth across the CEFR levels, EuroSLA Monographs.
  • Duolingo investor and product announcements, 2025 and 2026.