AI Language Practice vs a Tutor vs Duolingo

We make an AI language partner, so read this with the appropriate scepticism. We are going to try to be genuinely useful anyway, because the comparison people usually make — AI versus tutor versus Duolingo — treats them as three options for the same job, and they are not.

What streak apps actually train

Gamified apps like Duolingo are extremely good at one thing: getting you to show up. That is not a small thing. The main reason adults fail at languages is not bad method, it is stopping after five weeks.

What they train, mechanically, is recognition and translation. You are shown four options and pick one; you are given an English sentence and assemble the target-language version from word tiles. Both are recognition tasks with the answer visible on screen. They build a real base of vocabulary and a feel for word order.

What they do not build is production under time pressure, which is the skill that makes conversation possible. This is why the very common experience of "I finished the Spanish tree and froze in Madrid" happens. It is not a failure of the app; it is a mismatch between what was trained and what was needed.

What a human tutor does that nothing else does

Three things, and they are important.

Diagnosis. A good tutor notices that your problem is not vocabulary, it is that you are constructing every sentence in English first and translating. They will restructure your practice around that. No AI currently does this well, because it requires holding a model of you across months and having opinions about what to change.

Accountability. A person who expects you on Thursday at seven is a commitment device. Software is not, no matter how many notifications it sends.

Real social stakes. Speaking to a human who might be confused, amused or bored is a different physiological experience from speaking to software. Some of what makes speaking hard is that pressure, and you cannot fully train for it without it.

The cost is money and scheduling, and those costs have a downstream effect: because a tutor is expensive per hour, you have fewer hours, and you spend some of them on things you could have done alone.

What AI conversation is actually good at

One thing, but it is the bottleneck for most people: volume of production at zero cost and zero social risk.

You can have a ten-minute conversation at 6am, be incoherent, restart, say the same sentence six times until it comes out cleanly, and there is no one on the other end whose time you wasted. For the specific problem of converting passive knowledge into active speech, repetition volume is most of the answer, and this is where AI is not slightly better but categorically better.

Second, and less obviously: graded difficulty on demand. A native speaker friend cannot reliably speak to you as a patient L1 speaker would; they will either over-simplify into baby talk or drift back to natural speed. Software can hold a level. Amiko's five levels go from short-sentences-and-common-words to slang-and-real-speed, and you can change them mid-conversation when a character is going too fast.

Third: support that does not stop the conversation. Translation under every line, a reading above every word, a suggested reply when you are stuck. With a human, asking "what does that mean?" costs a turn and breaks the flow. On a screen you can tap a word and keep going.

What AI is bad at, including ours

It will not hold you accountable. It has no long-term model of your specific weaknesses. It is too agreeable — it will understand your broken sentence when a real stranger would have said "sorry?", which means you can over-estimate how comprehensible you are. And it cannot replicate the adrenaline of an actual conversation with an actual person who is waiting.

There is also a general-chatbot caveat worth being direct about. You can practice with ChatGPT or Claude and it will work, up to a point. What you do not get is level control that holds, furigana and pinyin rendering, a suggested next line, vocabulary graded against JLPT or TOPIK, audio in a natural voice, or a character who stays in role instead of drifting into being a helpful assistant explaining grammar in English. Those are the things a purpose-built app is for, and if you do not need them, a general chatbot is a perfectly reasonable free option.

The combination that actually works

If you have money and time: a tutor every week or two for diagnosis and accountability, daily AI conversation for volume, and a vocabulary source for input. The tutor tells you what to fix; the AI is where you do the reps.

If you have neither: daily AI conversation plus graded vocabulary gets you further than either alone. Twenty minutes a day, ten of it talking.

If you are at the very beginning: learn the writing system first — hiragana, Hangul — because everything downstream is faster once you can read. Then start talking immediately, at a level low enough that it is possible. You can try that here without an account.

The one thing we would say with real confidence, from watching thousands of people practice: the learners who progress are not the ones with the best method. They are the ones who spend the largest share of their study time producing the language rather than consuming it. Whatever tool gets you doing that is the right tool.