Why Learn a Language When AI Can Translate?
· 8 min read
Machine translation has largely won at transferring information. Here is what it still cannot do, and why learning a language is now easier than it has ever been.
Real-time machine translation now works well enough that the question deserves a serious answer rather than a defensive one. The honest version: for transferring information, translation has largely won. For everything else people actually use language for, it has barely competed. Here is where the line falls, and why it falls there.
The objection, stated properly
Most defences of language learning begin by understating the opposition. Let's not.
You can point a phone at a Japanese menu and read it. You can wear an earbud in a Warsaw taxi and hold a functional conversation. You can read a Hungarian news article, a Turkish contract, or a Korean forum thread in your own language, instantly, for free. Ten years of effort used to buy you access to a body of text. Now the access is a default, and the effort is optional.
If your goal is to extract information from a language, machine translation is faster than you will ever be, in more languages than you will ever learn. Someone who tells you otherwise is defending a hobby, not making an argument.
So the question is real: what is left?
Translation handles transactions. It does not handle relationships.
The clearest boundary is not technical, it's social.
Ordering a coffee is a transaction. Information goes one way, goods come back, and a machine in the middle costs you nothing. Machine translation is excellent at this and will keep getting better.
A friendship is not a transaction. Neither is a joke, an argument, a flirtation, a condolence, or the twenty minutes of aimless talk that turns a colleague into a friend. These depend on timing, on interruption, on the half-sentence that gets abandoned, on speaking at the same moment and both laughing. Insert two seconds of processing latency and a synthesised voice into that, and you have not slowed the interaction down, you have replaced it with a different kind of interaction, one in which both people are talking to a device about each other.
This is why the "AI solved it" case tends to be argued by people describing travel and refuted by people describing living somewhere. Tourists have transactions. Residents have relationships.
You cannot audit a translation you don't understand
Machine translation fails silently. When it renders a sentence wrongly, it does not tell you; it hands you a fluent, confident, plausible sentence that means something else. If you don't speak the language, you have no way to notice.
For a menu, the stakes are a disappointing lunch. For a rental agreement, a medical consultation, a negotiation, or a conversation where someone is being indirect on purpose, the stakes are different. Knowing even a moderate amount of a language changes your relationship to the tool: you stop being a passenger and start being able to spot when the output has drifted.
This is a genuinely new argument. It didn't exist in 2010, because the tool wasn't good enough for anyone to trust it blindly.
The gap is much wider than English-Spanish suggests
Public opinion about machine translation is formed on the language pairs where it works best. English with Spanish, French, German, Chinese. These have enormous parallel corpora, decades of investment, and the most attention from every major lab.
Quality degrades, sometimes sharply, as you move down the resource curve. Georgian, Wolof, Icelandic, Tamil, Welsh, Quechua: the further you go, the more you find translations that are grammatically loose, register-blind, or confidently wrong about idiom. And there is an unhappy correlation here. The languages machines handle worst are frequently the languages where speaking a few sentences yourself carries the most social weight, precisely because so few outsiders bother.
If you are learning a widely-spoken language for utility, the machine has a strong case against you. If you are learning a smaller language for connection, the machine barely has a case at all.
What a language gives you that a translation of it does not
A translation delivers the proposition. The language delivers the thing itself.
Every word sits in a network of other words, its neighbours, its register, the contexts it usually shows up in, the people who tend to use it. A competent translation preserves the meaning and discards the network. That's not a flaw in translation; it's what translation is. But it means that reading a poem, a song, an insult, or a piece of bureaucratic evasion in translation is genuinely a different experience from reading it in the original, in a way that has nothing to do with the translation being good or bad.
There is also the harder-to-defend point, which is nonetheless the one most multilingual people report: you are a slightly different person in each language you speak. Different jokes are available. Different degrees of directness feel natural. Some things are easier to say. This is not measurable and nobody should pretend it is, but it is the reason people who have done it rarely regret it.
What actually changed
The interesting shift isn't that AI removed the reason to learn a language. It's that AI removed the excuse for learning one badly.
The bottleneck in language learning was never motivation. It was material: finding enough text and audio you could almost understand, at your exact level, on subjects you actually cared about, in a language where such material may never have been produced. Beginners were handed dialogues about hotel check-in. Intermediate learners fell into the gap between textbook material and native material, and most of them quit there.
That constraint is gone. Generating unlimited comprehensible input at a precise level, in almost any language, is now trivial, which makes the input-based approach that researchers like Stephen Krashen argued for in the 1980s practical for the first time since it was proposed.
What to actually use
PlusOneLanguage is built directly on that shift. It generates comprehensible input texts at your exact level in 48 languages, including the low-resource ones where graded readers were never published, and it deliberately reuses the vocabulary you have already met across later texts, so words come back before you forget them rather than appearing once and disappearing.
Concretely, the loop looks like this:
- Read a short text at your level. You choose the topic, or let the app pick. The text is written so you understand most of it and meet a handful of new items, the classic i+1 ratio.
- Tap any word you don't know. You get a translation that takes the surrounding sentence into account, not a dictionary's first entry, so cuisine in "ma mère prépare la cuisine" is not glossed as a verb. One tap saves the word to your deck with the sentence it came from.
- The next text reuses your saved words. This is the part isolated tools skip. Your deck is not a graveyard; it is the input for the next generation. Words you saved yesterday reappear in new sentences today, in different grammatical shapes and different contexts.
- Spaced repetition times the returns. Flashcards run on an SM-2 schedule, so a word you found hard comes back sooner and a word you found easy comes back later. The scheduling decides when, the generated texts decide where.
- Turn recognition into production. Reading makes words familiar; it does not make them available when you speak. Sentence Builder makes you reconstruct sentences from your own vocabulary, and Conversation Missions push you to produce language under mild time pressure. Pronunciation audio is there for the sound-to-spelling mapping that reading alone never fixes.
A few design choices matter more than they look:
- Context is preserved permanently. The example sentence attached to a word is the one you originally saved it from, so the flashcard explanation and the example never drift apart.
- Grammar arrives inside texts, not before them. For French, rules and conjugations are woven into the generated text and explained in place, then tracked so you can see which ones you have actually met and mastered.
- Level bands, not a single level. Material is drawn from around your level rather than strictly at it, which is what keeps input comprehensible without becoming boring.
- It works for languages the internet forgot. If you are learning Lithuanian, Icelandic, or Catalan, the shortage was never your motivation, it was the material. Generated input removes that shortage.
Fifteen minutes a day of this is not the same activity as fifteen minutes of tapping through a word-matching game. One produces recognition of items; the other produces the repeated, varied, contextual exposure that acquisition actually requires.
So the honest summary is this. If you want to know what a document says, use the machine; you will not beat it. If you want to understand people, be understood by them, notice when you're being misled, or find out who you are in another language, that is still yours to do, and it is now considerably easier than it has ever been.
The machines got good at translating. They did not get good at being you.