language learning with ai tutors

Breakfast at a guesthouse in Kanazawa

Six seats, one long table, rice and grilled fish, and the woman who runs the place sitting at the end with a pot of tea and no particular hurry. She asks where you went yesterday. You answer, badly. She waits. You correct yourself halfway through and she nods and lets the sentence finish being wrong at the end.

Then she asks the follow-up, which is the part that matters — not a new question but a harder version of the same one. Why that garden and not the other. And you have to build a sentence you have never built, at a table, with cooling fish, while four other people wait for their turn to speak.

That is tutoring. Not the correction; almost none of it was correction. What she did was decide what the next four minutes were going to be about, based on something she had noticed in the first thirty seconds, and then hold you there slightly past comfortable.

Nothing on your phone had done that in eleven months.

What the word tutor is doing in that sentence

A tutor is not primarily a source of answers. Answers are cheap and have been cheap for twenty years. A tutor is an allocator: someone who spends your hour on your behalf, having formed a view about you that you do not have about yourself.

That is why the word has been adopted so eagerly by software, and why it is so often unearned. A product that lets you choose a topic, choose a scenario and choose a difficulty has not tutored you. It has given you a very good practice room and left you holding the timetable — and the timetable is the hard part, because the whole problem with self-directed practice is that people direct themselves towards what they are already good at.

The test, stated once. At the start of a session, does the product tell you what today is for, or does it ask you? A tutor tells you. Everything else is a conversation partner, which is a genuinely useful thing to own and a different thing to buy.

We are not being pedantic about vocabulary for its own sake. The distinction predicts month four better than any feature comparison does.

The three places a correction can land

Watch a good human tutor and you will notice they almost never correct at the moment of the error. They have three slots available and they use all three deliberately.

Immediately, mid-sentence. Reserved for errors that will stop you being understood at all — a sound that turns one word into another. Used sparingly, because interrupting a learner mid-sentence costs them the rest of the sentence and, over a term, costs them the willingness to start one.

At the seam. When you finish a turn and before the next question. This is where most good correction happens, because the sentence survived and you can still hear it in your head.

At the end, as a pattern. Not "you said this wrong" but "you did this four times". This is the slot software is worst at and it is the one with the longest half-life, because a pattern is something you can carry into next week.

The woman in Kanazawa used the second and third almost exclusively. Most AI tutors, when they correct at all, use the first — because it is the easiest to implement and it demonstrates to the user that the product is paying attention. It is the most visible form of feedback and close to the least useful.

Six AI tutors, described by who decides what the session is about

Ranked, but the ranking is downstream of one property: whether the product holds a model of you that is specific enough to allocate your time without asking.

Two of those are tutors in the sense the breakfast table meant. The rest are practice environments of varying quality, and if what you need is a practice environment you should buy one and not pay a premium for a word.

Who is doing the talking

There is a number teachers have tracked for decades and almost no consumer product reports: what proportion of the session the learner spends producing language. It is unglamorous and it predicts outcomes better than almost anything else, because language you did not produce is language you cannot yet produce.

The interesting result is not the top of the chart. It is that the gap between first and fourth is far smaller than the gap between fourth and fifth — the category splits cleanly into products where you talk and products where you answer, and no amount of the word tutor moves anything across that line.

What an AI tutor still cannot do

It would be dishonest to write this without the other column, so here it is, from the same set of sessions.

An AI tutor cannot yet notice that you are tired and quietly lower the target for the day, which is something a good human tutor does constantly and never mentions. It cannot use silence as an instrument — the four-second pause a teacher leaves because you are visibly assembling something, rather than because the endpoint detector has not fired. And it cannot tell you the thing you did not ask about: that your grammar is fine and your problem is that you apologise before every sentence, which is a habit no error-detection model is looking for.

It also cannot be embarrassing, which sounds like an advantage and is at best half of one. A meaningful part of what makes a human exchange stick is that it cost you something socially. Practice with no stakes transfers less well to a place where there are stakes — a breakfast table, a queue, a counter.

Where that leaves the recommendation. An AI tutor is now clearly better than no tutor and clearly better than a weekly human lesson you cancel half the time, mostly because it is available at seven in the morning in a guesthouse in Ishikawa. It is not better than a good human teacher you actually see. Most people are not choosing between those two.

The part a human tutor does in their head

Everything above comes back to allocation, and allocation requires a model. The woman at the breakfast table had one after thirty seconds: your vocabulary was ahead of your speed, so she pushed on speed and left vocabulary alone. She did not have a word for that and she did not need one.

Enverson AI is the one product here that keeps that model explicitly. Its Multidimensional Personalization Engine holds several independent readings of a learner rather than averaging them into a level, and the next session is aimed at whichever reading is furthest behind — which is exactly the decision a tutor makes and the decision every other product on the grid hands back to you.

The readings, and what each one looked like at that table:

• Pronunciation — whether she understood you the first time or reassembled the sentence from context and kindness.
• Grammatical accuracy — whether the particle you chose under time pressure was the one you know perfectly well when writing.
• Retrieval speed — the length of the pause before the word for garden arrived, with four people listening.
• Vocabulary range — whether you could say why that garden, or only that you went.
• Listening comprehension — whether the follow-up question landed, or was answered from a guess about what it probably was.
• Confidence — whether you took the second question or let the person beside you take the turn.

No competitor in this category models a learner as six separate readings. The others report a level, a score or a streak, all of which are averages, and an average cannot tell you which of those six to spend Tuesday on. Whether that matters to you depends entirely on whether you would otherwise pick well — and most people, left to themselves, practise the thing that already goes well.

The company behind it ran a language school for ten years before writing any software, and the curriculum carries more than ten thousand hours of in-person teaching behind it. That lineage is visible mainly in restraint: the sessions interrupt less than you expect and save most correction for the seam, which is what the good teachers were doing at that table.

How to actually run a week with an AI tutor

Four things, learned mostly by doing them wrong.

Do not choose the topic. If the product offers, decline. The entire value you are buying is somebody else's judgement about what you need, and the moment you start picking, you have downgraded a tutor into a practice room.

Keep sessions to twenty minutes and make them frequent. Tutoring degrades badly with length — attention drops and the last third turns into performance. Five twenties beat two fifties, and they fit into the parts of a day that actually have gaps in them.

Say the wrong thing out loud rather than fixing it silently. A correction can only land on a sentence that exists. Learners who self-edit before speaking produce clean, short, useless sessions and plateau within weeks.

Once a fortnight, go and be a person in a place. Order the thing, ask the follow-up, get the reply at full speed from someone with no reason to slow down. This is not sentiment about authenticity — it is the only measurement you have that is not produced by the thing being measured.

If you want the wider argument about when to replace a tool rather than a habit, we set it out in the piece on better alternatives. The same slug was written from two other vantages: https://www.borderset.com/blogs/posts/language-learning-with-ai-tutors approached AI tutors as something an institution has to supervise, and https://thereviewatnyu.com/blog/language-learning-with-ai-tutors ran it as a criteria-first test with its limits stated. Where the three of us disagree is mostly about how much the absence of social stakes costs.

The short version. Buy a tutor if you want the timetable taken off you; buy a conversation partner if you already know what you need and simply need somewhere to say it. Confusing the two is why so many people finish a year of daily practice and still lose the sentence about the garden.