Using AI to practise English speaking: what works, what does not

The single hardest logistical problem in learning to speak English has always been finding a patient partner at a convenient hour. AI removes that problem completely, which is a bigger deal than most of the marketing around it suggests — and smaller than the rest of it claims.

What AI is genuinely good at

Unlimited, unembarrassing volume. You can have the same conversation ten times. You can stop mid-sentence, restart, and try a different phrasing. No human partner offers that without social cost, and social cost is the main reason learners avoid practising.

Answering the question behind the question. Is "I am having a car" wrong, and why? A grammar book gives a rule; a good model explains the distinction between stative and dynamic verbs in the context of the sentence you actually produced.

Generating targeted material. Ten sentences drilling the /v/–/w/ contrast, using vocabulary from your job, at your level. That used to require a teacher's preparation time.

Rehearsal. Interviews, presentations, a difficult phone call. Run it five times before the real thing.

What it is not good at

Judging your pronunciation. Text-based models cannot hear you at all. Speech models can, but their assessment of accented English is unreliable in both directions — sometimes flagging a perfectly intelligible Indian English vowel as an error, sometimes passing a substitution that genuinely changes the word. Do not take pronunciation verdicts from software as final.

Telling you what to prioritise. An AI will answer whatever you ask. It will not notice that you have been avoiding the past perfect for three weeks, because it has no model of your history unless you build one.

Being honest by default. Ask "was that good?" and you will usually get encouragement. Ask "list every grammatical error in what I just wrote, with corrections, and rank them by how much they affect comprehension" and you get something useful. The quality of the feedback is almost entirely determined by the specificity of the request.

A practice loop that works

1. Speak for five minutes on a topic, recording it.
2. Transcribe it — most phones do this natively now.
3. Paste the transcript in and ask for errors ranked by impact on comprehension, not by formality.
4. Pick the single most frequent error. Ask for ten practice sentences on it.
5. Re-record tomorrow, watching only for that one error.

Step 3 is the one that separates useful practice from chatting. Step 4 is the one people skip.

Where a human is still necessary

Three places. Pronunciation you cannot hear in yourself. Knowing which of your fifty errors actually matters for your goal. And the accountability of someone expecting you on Tuesday.

The sensible arrangement is AI for volume and a teacher for direction — which is how the LangVR courses are built, and why the platform pairs AI practice with live classes rather than offering one instead of the other.