AI language tutors, compared against real tutoring

Every language app now says it has an AI tutor. The claim costs nothing: wire a chat model to a microphone and the marketing writes itself. But a conversation partner and a tutor are different jobs. A partner talks with you. A tutor changes what you can do, on purpose, and the difference between the two is not the model behind the app. It is pedagogy. So instead of ranking apps by demo feel, this post lays out a rubric borrowed from what good human tutors do, and then holds the field to it, including us. Kotobia is our product; judge the fairness of what follows accordingly.

The rubric: seven things good tutors do

Watch an experienced tutor work and a pattern shows up fast. The learner does most of the talking, 70 or 80 percent of the session, and the tutor is doing quiet, deliberate work in the gaps. That work breaks down into seven habits.

They keep vocabulary at your level, rephrasing before translating, because input you cannot parse teaches nothing. They correct by recasting: you say a sentence wrong, and the correct version appears naturally inside their reply, so the correction lands without stopping the conversation. They ask open questions, never ones you can escape with a yes or no. They let silence sit; five seconds of you struggling toward a word is where the learning happens, and a good tutor does not rescue you too early. They scaffold heavily at the start of a session and fade the help as you warm up. They keep a private list of your recurring errors, the fossilized ones, and target those explicitly rather than letting a recast wash over a mistake you have made thirty times. And they connect practice to your life, because you will fight harder for a sentence about your own week than one about a fictional airport.

A tutor, human or AI, that does these seven things will move you. One that does none of them is a talking phrasebook, whatever the model underneath.

The field, held to it

Speak comes closest on conversational craft. Its lessons scaffold well, its speech recognition is forgiving, and the production values make daily practice pleasant, which matters more than reviewers admit. Where it thins out is memory of you: the curriculum drives the conversation, your error history does not, and there is no review system underneath deciding which words deserve tonight's practice.

Duolingo Max bolts AI conversation onto the strongest habit machine ever built for language learning. The convenience is real. The sessions, though, are scripted around the unit you just finished, the questions tend to be answerable in fragments, and nothing about your speech changes what the app shows you tomorrow. It is conversation as a feature, not tutoring.

TalkPal and the wrapper apps give you unlimited willing conversation, cheap. Against the rubric they mostly fail: level control is loose, corrections arrive only on request, silence gets filled instantly because chat models abhor a pause, and no session leaves a trace on the next. As free-talk volume for an intermediate learner with self-discipline, still useful.

Langua deserves its following among serious learners. The conversations are decent, and saved words go into a review system, which is more than most manage. The link runs one direction, though: conversation feeds the deck, but the deck and your error history do not drive the conversation. You review what you saved; you are not made to say it again under pressure.

Kotobia was built from this rubric, so on paper we hit the seven: level-capped vocabulary, recasts, open questions, deliberate patience with silence, scaffolding that fades within a session, an explicit read of your fossilized errors at session start, and conversations seeded from your interests. The honest caveats: we are a small product without Speak's polish, English is the only native language supported today, and the tutor is text-scripted speech, not a human voice with a human ear for nuance. What we have that the field does not is the loop: the tutor speaks from your actual review deck, and what you manage to say feeds the schedule back. Say a due word correctly in conversation and it counts as a review. Fumble it and it comes back sooner. No one else closes that loop, and it is the entire reason the product exists.

Three questions that sort the field

If you are evaluating any of these, including us, three questions do most of the sorting. Does the tutor know what you are studying, or does every conversation start from zero? Who talks more, you or it? And does anything you say today change what the app does with you tomorrow? Apps that pass all three are doing tutoring. Apps that fail all three are charging you to talk to a model you could prompt yourself.

Our own plans and prices are on the home page. If you are coming off a long Duolingo streak specifically, the alternatives post covers that switch in more detail.