Nivelo Guide
Can ChatGPT Teach You English?
Yes, for a lot of it: explaining a rule, fixing a paragraph, playing the other side of a conversation in text. No, for the two things that decide your level: it cannot hear how you actually sound, and it cannot measure you. Here is the honest split, and how to use it well.
Can ChatGPT teach you English?
Partly, and the part it does well is bigger than most teachers admit. ChatGPT explains grammar on demand, corrects your writing sentence by sentence, gives you ten example sentences for a word, and will play a landlord, an interviewer or a barista for as long as you like. What it cannot do is the two things a good teacher does with their ears: notice that your th came out as d, and tell you honestly what level you are. Both of those need a measurement, and a chat model does not measure. It predicts text.
| What you want | ChatGPT | Why |
|---|---|---|
| Explain a rule with examples | Good | Explanations are text; it has read every grammar book |
| Correct a paragraph you wrote | Good, with a check | It fixes errors well but also “improves” correct sentences |
| Practise a conversation in text | Good | Role-play is its native mode |
| Build vocabulary in context | Good | Ask for sentences at your level, not lists |
| Tell you how you sound | No | Voice mode turns your speech into words first; the sounds are gone |
| Tell you your CEFR level | No | It guesses from text; reading, listening and speaking are not in the sample |
| Keep you honest over weeks | No | It has no memory of your first-try rate, and it praises by default |
What is ChatGPT genuinely good at for English learners?
Four things, and they are worth real money in a classroom. First, explanations at exactly the depth you need: ask why I have been living here and I have lived here both work, and you get a clear answer with examples in seconds. Second, writing correction: paste a paragraph and ask for the corrections listed one by one with the rule each one breaks. Third, text role-play: it will be a difficult customer, a job interviewer or a friend who only talks about football, and it never gets tired. Fourth, vocabulary in context: ask for a word inside five sentences a B1 learner would meet, rather than a definition.
Two habits keep it useful. Ask it to show its corrections rather than rewrite silently, because a silent rewrite teaches you nothing and often changes sentences that were already fine. And when it states a fact about an exam, a visa or a word’s register, check it against the official source. It writes with the same confidence when it is wrong.
Why can’t ChatGPT hear your pronunciation?
Because voice mode does not listen to you the way a person does. Your speech goes through a speech recogniser first, and a recogniser’s whole job is to produce the words you meant. Say cortao for cortado and the transcript reads cortado. Say tink for think and the transcript reads think. The model then answers the transcript. The sound you made, the one thing a pronunciation teacher would react to, never reaches it. This is not a setting you can change. It is how the pipeline is built.
We measured how far this goes. In September 2026 we sent 48 short clips of native-accent Spanish voices, twelve each from Spain, Mexico, Argentina and Colombia (two synthetic native voices per country reading six sentences), to a leading audio model and asked it only to name the country. It got 22 right, 46 percent, against a 25 percent chance floor, and called Colombian speakers Mexican five times out of twelve. A model that cannot reliably tell a Bogotá accent from a Mexico City one is not going to tell you that your d was too hard. Pronunciation feedback needs a scorer built for sounds, not a chat model listening through a transcript.
| You said | What the transcript shows | What a pronunciation scorer hears |
|---|---|---|
| “I tink so” | I think so | th realised as t — the sound that marks you as a learner |
| “Comfor-TA-ble” (four syllables) | comfortable | Stress on the wrong syllable, an extra syllable |
| “Ship” for “sheep” | ship (or sheep, by context) | Short i instead of long ee — the vowel you cannot hear yet |
| A mumbled ending on “worked” | worked | Dropped -ed; the past tense went missing in speech, not in text |
Can ChatGPT tell you your English level?
It will happily say “you are around B2”, and it may even be close, but it is a guess from one kind of evidence. A CEFR level covers reading, listening, speaking and writing. A chat model sees your typed sentences, which is one of the four, produced with unlimited time and a backspace key. Self-assessment has the same problem in a milder form: a meta-analysis by Steven Ross (Language Testing, 1998), pooling sixty correlations from earlier studies, found that learners’ own estimates correlate only moderately with measured proficiency, and that the accuracy varies by skill and by how the question is asked.
The honest move is to get measured on the skills the chat cannot see. A short CEFR-aligned test gives you a range rather than a pinpoint, and that is the right shape for a five-minute measurement. It is not an official certificate, and no online test is. It is a number you did not choose yourself, which is the point.
Why does chatting with an AI feel easier than it should?
Because most of it is recognition, not retrieval. When ChatGPT writes I would like a cortado, please and you read it, nod and type “thanks”, you have recognised a sentence. When you have to produce that sentence yourself, out loud, from the idea “order a coffee” with nothing on the screen, you have retrieved it. Memory research is unusually clear on which one builds skill: in Roediger and Karpicke’s 2006 experiments, students who practised retrieving material remembered far more a week later than students who re-read it, even though re-reading felt easier and more productive at the time.
This is why a text conversation can feel fluent while your spoken English stays stuck. The chat supplies the words and you supply agreement. A useful drill flips it: get the meaning in your own language, produce the English line yourself, then compare with how a native would say it. Repeating after a recording trains your mouth; working the line out from the meaning trains the part that goes blank when a real person is waiting.
How should you actually use ChatGPT for English?
| Goal | Ask it like this | What to check |
|---|---|---|
| Understand a rule | “Explain the difference between X and Y with three examples at B1 level, then give me two sentences to complete.” | Do the two sentences before reading the answers |
| Improve writing | “List every error in this paragraph, one per line, with the rule. Do not rewrite it.” | Rewrite it yourself from the list |
| Practise speaking, by text | “You are a hotel receptionist. Ask me one question at a time. Only correct me at the end.” | Answer in full sentences, not fragments |
| Build retrieval, not recognition | “Tell me in Spanish what to say in this situation. I will write it in English. Then show a native version.” | Say your version out loud before you look |
| Learn vocabulary | “Give me the word inside five sentences a B2 speaker would actually say. No definitions.” | Make one sentence of your own with it |
What can’t any AI do yet, and what should you do about it?
Three gaps, each with a fix that does not involve waiting for a better model. It cannot measure you: take a CEFR-aligned test that samples reading and listening, and treat the result as a range. It cannot hear you: record yourself saying a line, then play a native recording of the same line straight after, and listen for the one sound that differs; a dedicated pronunciation scorer can put a number on it, a chat model cannot. It will not keep you honest: track how many lines you get right on the first try, because a first-try rate moves slowly and cannot be flattered.
Used this way, ChatGPT is the best free grammar tutor and writing corrector most learners have ever had. It is just not a listener. If your goal is to be understood by real people, put the ear somewhere else: a native recording, a pronunciation scorer, or an actual conversation with something that reacts to what you said rather than what you meant.
Is ChatGPT voice mode good for speaking practice?
It is good for confidence and for keeping a conversation going, and that has real value if speaking makes you freeze. It is not good for accuracy, for the reason above: it responds to the transcript, so a badly pronounced sentence with the right words gets the same warm reply as a clear one. Use it to talk more. Use something with an ear to talk better. And whatever you use, decide your level with a measurement, not with a compliment.
Take the first step
Find out your current CEFR level in 5 minutes
Practice on your phone too: Nivelo is on the App Store.