I've been learning languages for most of my life. English first, I loved it as a kid, I was genuinely excited about grammar, past perfect included, and it led to a CPE with an A grade at eighteen and a CELTA with an A a few years later. French came after that, and I learned it on my own, an app, a lot of Netflix, then a tutor for writing. Today I teach English, and lately I've been doing more of my own learning with an AI on the side. A few things have become clear about what these models are genuinely good at, and where they will happily lead you astray.
AI is improving fast. Everything described here is what the models do in September 2026, and some of it won't stay true for long.
Start with why
The model needs to know your goal, not just the language. "Teach me Spanish" produces generic lists. "I'm a Polish carpenter moving to Valencia to build wooden playgrounds for schools, I need Spanish for talking to the site manager about deliveries and defects" produces the words you'll actually use. When I teach, I start every course by asking a student why they're learning. AI works the same way. Most of the value sits in that first question.
Most general word lists are a trap. "The 100 most common English words" is mostly articles, prepositions, and pronouns, words like the and on that carry close to no meaning on their own. Not where you should start.
Common verbs are the exception. Being able to say be, have, go, get early lets you actually form sentences, and the verb is where most of the meaning of a sentence lives. Start with those.
Words in context
A word with nothing but its translation next to it teaches you translation, not the word. It teaches you to turn the language into Polish, and it never shows you how to use it. Ask the model for several examples for each word, especially when the word has more than one meaning.
The most natural place to meet new vocabulary is a dialogue. Ask the model to simulate two people talking about something specific to your situation, and build your list from that. There's a trick here: ask it to write a conversation between two people, not to teach you. A model in teacher mode sounds like an assistant. Two people talking to each other in a scene sounds like real language.
And don't let the method take over. You can spend forever polishing your system, the app, the prompts, and never learn a word. A rough page of words with examples, printed and stuck to the fridge, is enough to start. Learn with it for a bit, then improve. Practice shows you what to fix. And learn with someone, even if it's a different language. It's easier to keep going when you're not doing it alone.
Idioms and the untranslatable
Be careful with idiom lists. Most of them are decades out of date. "It's raining cats and dogs", "once in a blue moon", you will rarely hear a real person say either. Learn the expressions you can actually hear, like take something at face value. You'll start recognising them the moment you listen to native speech.
Some expressions resist translation entirely, not because of the words but because of the culture. When I was learning Japanese, a greeting like yoroshiku onegai shimasu kept being rendered as "nice to meet you", and that misses almost everything the phrase carries. You learn this kind of thing from full phrases and real situations, not from a card with one word on each side.
Grammar: examples from the model, rules from people
Models produce beautifully grammatical sentences. Use those as examples. But when it comes to explaining rules, they will confidently invent grammar that doesn't exist. They mislabel tenses, and they can't even agree with themselves about British versus American usage. For now, I take examples from the model and rules from people. British Council pages, a reference grammar, a real textbook. The model is the wrong teacher for grammar theory.
Pronunciation: don't imitate the machine
Here the model is at its most misleading. AI voices speak hyper-correctly. An AI saying French sounds perfect, exactly like the audio CD that came with the textbook, and nothing like the people in Paris, who mumble, swallow sounds, and run words together. Model audio also arrives without a decided accent. Imitate it and you get an accent that belongs to no one.
So pick one. Decide who you want to sound like, a Londoner, an American, a particular region, and aim at that. Then listen to real people, a lot. Exposure beats drills for everything except the individual sounds themselves. For those, minimal pairs are the tool, and you don't need AI for it. Train on words that differ by a single sound, and you learn the difference and why it matters.
An app that scores your pronunciation is worth having if you don't have a native speaker at hand. Just remember that scoring single words pushes you toward hyper-correctness, while real speech happens in the joins between words.
AI is genuinely good at one more thing here: giving you listening material pitched at your level. Native radio at A2 is mostly noise, you understand almost nothing and your motivation dies. Ask for audio you can actually follow, where you understand most of what you hear, and build up from there.
Learning from scratch: the whole plan
If I started a language from zero today, I'd do it roughly like this.
Generate a few simple dialogues about your own life. Use the model as a guide, not as a teacher. Then copy the sentences by hand into a notebook. Writing by hand makes things stick in a way reading never does, you involve more of your senses.
Then move to the real thing fast, almost embarrassingly fast. A daily-life series, a forum people actually use, the internet of that language. It's hard at first and worth it. This is where the model earns its keep: paste a screenshot of a real conversation and ask what's happening, take a post you don't understand and ask for an explanation. It's the guide you read along with, not the material you read.
When I hit a grammar question, I ask the model to name the structure, then look it up properly myself. I don't let it explain the rules. And I'd use a model from the country whose language I'm learning, as long as it's one of the good ones. Today that means American or Chinese. For Chinese it matters: a Chinese model teaches real Chinese better than anything Western, so for Chinese, use a Chinese model. There's a Polish model, Bielik, that could be the perfect teacher for Polish one day, it just isn't good enough yet. For other languages, any good American or Chinese model works.
English helps here more than people expect. The best material for almost any language is in English, so it's worth reaching a working level first, and using it as a bridge.
The through line: language lives between people. The risk in all of this is learning to talk to a machine, which is not the same as learning to talk to a person. Do the AI part as early as you like, then get yourself to real material and real conversations as fast as you can.