Learning a language as an adult usually breaks at the same place: finding someone to talk to. Not the grammar. Not the word lists. It is the awkward hour with a person who may be judging your accent.
AI has made that part easy. Voice mode in ChatGPT, Gemini Live, and apps built specifically for conversation practice will hold a dialogue at one in the morning, on a bus, in your kitchen. Nobody laughs. Nobody switches to English after your second sentence. That alone keeps many learners going.
The other advantage is volume. You can ask for twenty sentences using the word you keep forgetting, all at the level you actually are. You can paste an email and ask what a native colleague would change. You can generate a dialogue for one specific situation, such as registering with a doctor in Austria or disputing a bill in French. Textbooks cannot do that, and teachers rarely have the time.
Where the tools get it wrong
Invented grammar
Models produce explanations that sound right and are wrong. Ask why a German verb takes the dative in one sentence and the accusative in the next, and you may get a confident answer that contradicts your coursebook. Treat every explanation as a first draft and check it against a dictionary or a grammar reference. When you ask a chatbot for a source, it may invent one.
Textbook register
These systems default to the language of manuals and news anchors. They rarely volunteer the particles, contractions and small everyday phrases that make speech sound natural. Learn only from a chatbot and you can end up speaking correct, stiff, faintly odd German.
The fluency illusion
A model understands you when a human would not. It fills the gaps, guesses the meaning and answers politely, so you never see the puzzled face that tells you your word order is off. A conversation can feel successful long after it stops being useful.
Pronunciation feedback
Speech recognition is not pronunciation scoring. Many apps use recognition built on models such as OpenAI's open-source Whisper, which transcribes what it thinks you said. If the system recognises a word you pronounced badly, you get no correction. Accuracy also varies with accent, and the training data is heavy on American and British English, so a Czech-accented speaker is likely to be transcribed less reliably than a Californian one.
Nothing is at stake
A chatbot has nothing to lose. It will not hand you the wrong dish, ask you to repeat yourself or look confused. Small risks in a real conversation are what push a learner to fix mistakes.
A workable weekly routine
Keep two tools: one app for structured material and repetition, one chatbot for speaking. Add more and you will spend your evenings choosing instead of practising.
- Daily, 10 minutes. Review vocabulary you met yourself, in menus, emails or at work, using spaced repetition. Words from a generic list disappear within a week.
- Three times a week, 15 minutes. Speak out loud in voice mode. Pick one topic and stay on it. Ask for corrections at the end, not during, so that you keep talking.
- After each session, 5 minutes. Ask for your five most frequent mistakes, each with the correct version and a one-line explanation. Save the list. It becomes your next review deck.
- Once a week. Write 150 words by hand, have them marked, then rewrite the text from memory with no help.
- Once a week. Record 60 seconds of yourself on a familiar topic, listen back and compare with a native recording of the same topic. No AI can do this part for you; training your ear is the point.
- Once a month. Talk to a human. A teacher, a colleague, a tandem partner, anyone.
The prompt below removes the most common failure, which is a chatbot that corrects you every ten seconds until you stop speaking altogether.
"Be my conversation partner in Spanish at B1 level. Speak only Spanish, in short sentences. Do not correct me while I speak. After ten minutes, list my errors with corrections, ordered by how often I made them."
Two instructions carry most of the value. Forbidding corrections during the conversation keeps you talking for fifteen minutes. Asking for the list afterwards turns a pleasant chat into a study session.
What it costs in Europe
| Option | Published price | What you get |
|---|---|---|
| Duolingo, free tier | €0 | gamified drills, ads, limited mistakes |
| Duolingo Super | roughly €7 to €13 a month, depending on billing period | no ads, unlimited mistakes |
| Duolingo Max | US list price around $30 a month; European prices are set per market | AI role-play, explanations of your answers |
| Babbel or Busuu | from about €6 a month on a longer plan, more month to month | structured course, speech practice, live lessons at higher tiers |
| ChatGPT Plus | $20 a month plus local VAT in the EU, commonly around €20 to €24 charged | voice conversation, corrections, rewriting of your own text |
| Gemini | free tier with limited live conversation; paid tiers available | voice practice inside Google's assistant |
| italki tutor | typically €10 to €30 an hour | human correction, direction, accountability |
Those are the vendors' published prices, and they move. Companies set different prices per country, app stores often charge more than a website, and consumer prices in the EU include VAT. Some voice modes also cap daily use, so check the limits for the tier you are considering. All the services in the table can be used from the EU.
The arithmetic that matters: ten hours of chatbot practice in a month costs one ChatGPT Plus subscription, roughly €20 to €24 including VAT. The same ten hours with a tutor at €15 to €30 an hour costs €150 to €300. That is the case for using AI on the tasks it handles well, and for paying a person only when the task demands it.
Your voice data and EU rules
Speaking practice means sending audio of your voice to someone's servers. Under the GDPR, that recording is personal data. If a service uses it to identify you, it counts as biometric data and stricter rules apply. You can ask a company what it holds about you, and you can ask for deletion.
In practice, open the privacy settings of whichever app you use and look for the option that keeps your chats or recordings out of model training. ChatGPT, for example, has a switch for this in its data controls. Business and team accounts are excluded from training by default; consumer accounts are the ones worth checking.
Since 2 August 2025, general-purpose AI models used in the EU come under transparency obligations in the EU AI Act. Providers have to publish summaries of their training data and label AI-generated output. For a learner the effect is modest but real: it should be clear whether you are talking to a machine or a person. Systems used to grade students or decide who gets into a course are listed as high-risk and face much heavier duties. A conversation partner is not in that group.
If keeping audio inside Europe matters to you, there are European options. Mistral's Le Chat, built in France, covers the main European languages and has a free tier. DeepL, based in Cologne, states that its paid translations are processed in European data centres. Running a model on your own laptop is the strictest option, because nothing leaves the disk. It is slower and less fluent than the cloud services, and the setup takes an evening. Our AI Arena tests compare how open models behave on ordinary consumer hardware.
When a human is the better buy
Three situations justify the hourly rate.
Pronunciation at the start. A wrong sound drilled for three months costs more to repair than a few early lessons would have cost.
Exam preparation. Goethe-Zertifikat, DELF and Cambridge exams have fixed formats and time pressure. A teacher who has prepared candidates knows the traps. A chatbot can rehearse your topics, but it cannot rehearse the format.
Register and culture. Which greeting to use with a landlord, how direct you can be with a colleague, when a joke lands. Language models are weakest here, because they were trained mostly on written text rather than on a Tuesday morning in an office.
Accountability counts as well. A tutor notices when you cancel.
Common questions
Can I learn a language from zero with AI alone?
You can reach a basic reading level, and apps handle vocabulary and simple sentences well. The weak spots are pronunciation and the first few hundred words, where a human ear catches what speech recognition does not. Most learners who reach a solid B1 do it with some structured course plus regular speaking practice, with AI covering the hours in between.
Do these tools work for smaller European languages such as Czech, Estonian or Irish?
Less well. Quality follows the amount of training data, and English, Spanish, French, German and Italian are far better resourced than Czech, Estonian, Maltese or Irish. In smaller languages you should expect more invented grammar explanations and rougher pronunciation feedback. You can test this in ten minutes by asking the model to explain one grammar rule and comparing the answer with a printed reference.
Should my child use a chatbot for language homework?
Start with the rules. Under the GDPR, consent for online services is valid from 16 unless a member state has lowered the age, in some countries to 13. Providers set their own minimum ages and require parental consent for younger users. In practice, a child practising conversation with a patient machine is useful. A child letting the machine write the assignment is not.
