How to Use AI to Learn a New Language (Without Falling for the Hype)

Can AI Actually Teach You a Language, or Is This Just Hype?

Before you read another 2,000 words, here’s the honest answer: AI can meaningfully speed up how you learn a language, but not the way the app-store screenshots suggest. Nobody hands you fluency while you sleep. What it actually does is knock down two obstacles that quietly stop most busy adults cold: paying for regular one-on-one practice, and the plain discomfort of speaking out loud with no one around to hear you fumble through it.

I train more than 2,000 learners a year on how to use AI tools without wasting time on features that look impressive in a demo but don’t actually move the needle. I hear a version of the same story constantly in my workshops: someone proud of a months-long Duolingo streak who still freezes the first time they have to order coffee in the actual language, at actual conversational speed. It’s a common story, honestly, closer to the default outcome of the streak model than an exception to it. Let’s be honest: most language apps overpromise fluency they can’t deliver, and a badge for showing up seven days in a row is not the same thing as holding a real conversation with someone who talks at normal speed and doesn’t slow down for you.

So this isn’t a roundup of twelve apps you should download. It’s a straight answer to one question: if you’re a working professional with maybe 20 minutes a day, what should you actually use, what should you skip, and what does the research say about whether any of it works.

Most of the people I talk to who tried and quit a language app fall into one of two camps. Camp one did the flashcard grind for months, felt like they were making progress because the streak kept climbing, then froze the first time a real person spoke back at conversational speed. Camp two bought a subscription, used it for a week, and let it quietly renew until they noticed the charge eight months later. Neither of those is a tool problem exactly. It’s a mismatch between what the tool is good at and what the person actually needed. That mismatch is what this guide is trying to fix.

What Makes 2026's AI Tools Different From the Old Duolingo Owl?

For over a decade, a ‘language learning app’ meant flashcards with a game skin on top: match the word, translate the sentence, don’t lose your streak. That model teaches vocabulary and grammar patterns reasonably well. What it has never done well is conversation, because a fixed set of pre-written exercises can’t respond to what you actually say.

That’s the piece that’s genuinely new. ChatGPT’s Voice mode is free for every user on every plan, and OpenAI markets it directly for language practice, right alongside brainstorming and quick answers. Its own feature page suggests prompts like ‘Help me practice ordering an iced latte in Japanese’^[1]. You talk, it listens, it responds in the target language, and it can switch mid-conversation if you get stuck and need to fall back to English for a second.

Duolingo approached the same problem from the other direction. Its Max tier launched in 2023 specifically to add live, AI-generated conversation on top of the existing lesson structure, built with early access to GPT-4^[2]. Two features came out of that: Video Call, an unscripted back-and-forth chat with a Duolingo character, and Roleplay, a scenario-based exercise (ordering coffee, checking into a hotel) that ends with a short performance report.

The distinction that actually matters for a busy professional isn’t ‘gamified app versus AI app.’ It’s structured drilling versus open conversation. You need both. Most people who’ve tried and failed to learn a language only ever got the first one, then wondered why three years of streaks didn’t leave them able to order dinner without switching back to English.

There’s also a quieter shift worth naming: cost. A weekly session with a human tutor typically runs somewhere between $15 and $40, and that’s before you factor in scheduling around your calendar and theirs. A voice conversation with ChatGPT costs nothing extra if you’re already on the free plan, and it’s sitting there at midnight, during a lunch break, in the ten minutes before a meeting starts. I don’t think the AI teaches better than a person. Honestly, in a straight fight, a good tutor wins every time. But a tutor isn’t available at 11pm when you finally have twenty free minutes, and that gap in availability matters more than people expect.

Is Duolingo Max Worth $168 a Year?

Duolingo Max sits above the free and Super tiers and currently runs $29.99 a month or $168 a year in the US, roughly £19.99 a month or £119.99 a year in the UK^[3]. It layers two AI features on top of everything in Super Duolingo (no ads, unlimited hearts): Video Call with a character named Lily, and Roleplay, the scripted scenario exercises^[2].

Here’s the part worth knowing before you pay for it: as of January 1, 2026, Duolingo made its third AI feature, Explain My Answer, free for everyone in its most popular courses, including Spanish, French, German, Italian, Portuguese, Japanese, and Korean^[4]. That feature used to be a Max-exclusive perk and was, frankly, the most genuinely useful one, since it explains why a specific answer was wrong in plain language instead of just marking it red. So the honest pitch for Max in 2026 isn’t ‘AI feedback’ anymore. It’s specifically Video Call and Roleplay, and both are capped at a few minutes per session and stay fairly scripted^[3].

I left the ‘best overall’ column out of that table on purpose, because there isn’t one. Duolingo’s strength is the habit loop and the spaced-repetition engine underneath it, which I’ll get to below. ChatGPT’s strength is that it will talk about literally anything, in your target language, for as long as you want, for free or close to it.

One more honest note on Duolingo Max specifically: it’s only available on iOS and Android, not the browser, and Roleplay and Video Call currently cover a limited set of languages. For English speakers, Spanish, French, German, Italian, and Portuguese get both features; Japanese, Korean, and Chinese currently get Video Call only^[2]. If you’re learning something outside that list, Max’s headline features may not even apply to your course yet, which is worth checking in the app before you pay for a year up front.

How Do You Actually Use ChatGPT Voice Mode to Practice Speaking?

Most people open voice mode, say ‘let’s practice Spanish,’ get a generic response, and give up after two minutes. Nine times out of ten the tool was fine and the prompt was doing all the damage. Here’s what actually works:

Give it a scene, not a topic

‘Practice French with me’ produces vague small talk. ‘Pretend you’re a landlord in Paris and I’m calling about a leaking pipe, only respond in French, and correct me gently after each thing I say’ produces a real exercise with a goal, vocabulary constraints, and built-in correction.

Set the correction style up front

Tell it exactly how strict to be: ‘Correct every grammar mistake immediately’ versus ‘let minor mistakes go and only stop me for things a native speaker would actually notice.’ Without this instruction, it defaults to being encouraging rather than accurate, which feels nice and teaches you less.

Ask for a mistakes summary at the end

Close every session with ‘give me a list of every mistake I made and the correct version.’ This is the single biggest upgrade over just chatting, because it turns a pleasant conversation into a study document you can actually review later.

Keep sessions to 10 or 15 minutes. Speaking a second language is mentally exhausting in a way that reading flashcards isn’t, and the quality of your output drops fast once fatigue sets in. Short and daily beats long and occasional, every time.

Use the transcript, don’t just talk and forget

Every voice conversation gets saved as a full transcript you can scroll back through afterward. Most people never open it. That transcript is genuinely one of the more useful study artifacts you’ll produce this week: it’s a record of exactly the words and structures you reached for under pressure, not the ones a textbook decided you should learn in chapter four. Skim it once, pull out three phrases you want to keep, and drop them into whatever spaced-repetition tool you’re using for review.

None of this requires a paid plan to start. The free tier includes voice mode, just with tighter usage limits, so there’s no reason to wait for a subscription before trying a single ten-minute session tonight.

Why Is Spaced Repetition the One Piece of "AI Magic" That's Real?

Duolingo’s research team published the algorithm behind its practice reminders back in 2016, and it’s a genuinely elegant piece of work, not a marketing claim. Researcher Burr Settles built what he called half-life regression: a model that estimates how long a specific word is likely to stick in your memory based on your personal history with it, then schedules your next review right before you’re statistically about to forget it^[6].

The underlying idea, the ‘forgetting curve,’ goes back to 1885, when German psychologist Hermann Ebbinghaus documented how quickly memory decays by testing his own recall of nonsense syllables. Memory doesn’t fade at a steady rate: you forget fastest right after learning something, then the rate slows the longer you’ve known it^[6].

When Duolingo tested its trained model against the older, simpler review system in a live experiment with hundreds of thousands of students, it cut prediction error by more than 45% and lifted daily engagement by 12%^[6]. That’s the part of ‘AI language learning’ that isn’t hype: a quiet algorithm deciding when you’re about to forget a specific word and nudging you to review it at the right moment, instead of either too early (waste of time) or too late (you’ve already lost it).

Why this matters for you: if a tool doesn’t do spaced repetition, it’s implicitly asking you to re-learn the same words from scratch every few weeks. That’s where most of the wasted time in language learning actually happens, and it’s the one advantage a dedicated app like Duolingo has over a general chat tool, which has no memory of which words you keep getting wrong across sessions unless you explicitly ask it to track that for you.

It’s also worth knowing this science predates the current wave of generative AI by almost a decade. Spaced repetition isn’t a new AI trick bolted onto an old app to make it sound cutting-edge. It’s a well-established finding from memory research, going back through Ebbinghaus and decades of follow-up studies on distributed versus massed practice, that language apps happened to be very well suited to implement at scale^[6]. That’s a meaningfully different claim than ‘AI will make you fluent,’ and it’s the one I actually trust.

What Does a Realistic Weekly AI Practice Routine Look Like?

Most people fail at this not because the tools don’t work, but because the routine they set for themselves is too ambitious to survive a bad week. Here’s one that’s built to survive a bad week.

• Monday, Wednesday, Friday (10-15 min): Structured lessons and spaced-repetition review in Duolingo (free tier is fine) to build vocabulary and grammar patterns.
• Tuesday, Thursday (10-15 min): A scenario-based ChatGPT or Claude voice conversation. Pick something from your actual week: a work meeting, ordering food, a phone call you’re dreading.
• Saturday (15 min): Listening only, no output pressure. A podcast, a Duolingo Story, or a YouTube video in your target language at slightly-too-fast speed.
• Sunday: Five minutes reviewing your ‘mistakes list’ from the week’s voice sessions. Nothing new, just reinforcement.

That’s roughly 90 minutes a week. It sounds small next to ‘immerse yourself completely,’ but the research suggests a few hundred hours spread out over months gets you to a genuinely useful conversational level, not thousands^[7]. Consistency beats intensity here, by a wide margin.

If you miss a day, or a week, don’t restart the whole plan or punish yourself into an all-or-nothing binge to make up for it. Just pick the next scheduled day back up. The learners I watch succeed long-term treat missed days as data, not failure: if Tuesday’s voice session keeps getting skipped, move it to lunchtime instead of guilt-tripping yourself into a streak you’ll abandon in a month anyway.

What Can AI Still Not Do When Learning a Language?

Here’s my honest skepticism, and it comes from actually watching learners use these tools, not from a general distrust of AI. Voice models are very good at understanding you even when your pronunciation is rough, which is great for confidence and terrible for accuracy. If the AI understands ‘gracias’ no matter how badly you stress the syllables, it has no strong incentive to flag that your pronunciation would confuse an actual Spanish speaker in Madrid. You have to explicitly ask for pronunciation feedback, and even then, a voice model is reading transcribed text back at you, not analyzing your mouth position or airflow the way a dedicated pronunciation coach or a human tutor would.

AI also has no idea what you actually did wrong socially. It won’t tell you that the phrase you just used is technically correct but sounds stiff, or that you picked the overly formal register with someone you should be casual with, unless you specifically ask it to roleplay that nuance. A lot of real fluency is cultural instinct, not grammar, and that’s still something you build through exposure to real people, real shows, real friction.

And there’s a more basic issue: accountability. An app can remind you. It cannot make you show up. The learners I see actually get good are the ones who pair AI practice with some kind of external commitment, a weekly call with a tutor, a language exchange partner, a class, something with another human on the other end who will notice if you skip three weeks in a row. Use AI to make your practice time more efficient. Don’t use it to replace the reason you started learning the language in the first place, which usually involves other people.

Sana Mian, Co-Founder of Future Factors AI

Sana is an AI educator and learning designer specialising in making complex ideas stick for non-technical professionals. She has trained 2,000+ learners across corporate teams, bootcamps, and keynote stages. Future Factors offers AI Bootcamps, Corporate Workshops, and Speaking & Consulting for businesses ready to adopt AI without the overwhelm.

More about Sana →