AI in language learning: what really changes for learners and teachers

AI in language learning matters first for one simple reason: speaking time. An AI tutor lets every learner speak every day, get corrected while they speak and rehearse the situations of their own job. The teacher does not disappear: they steer, explain and adjust, with progress data they never had before.

AI in language learning: what really changes

For years, technology mostly improved what already worked: grammar drills, vocabulary flashcards, video lessons. Useful, but rarely decisive. For most adults the sticking point is somewhere else. They read English, write a decent email and follow a presentation. They freeze when they have to speak, live, to someone who will not wait.

This is where generative AI, combined with speech recognition and speech synthesis, brings something genuinely new. For the first time, a learner can hold a real conversation with a virtual partner that understands what they say, answers naturally, keeps the conversation going and points out their mistakes. It is no longer a fill-in-the-gap exercise. It is an exchange.

Three things follow. People speak far more. They are corrected at the right moment. And the content can follow each person's real life, instead of teaching the same lesson about the weather to a lawyer and to a nurse.

Speaking practice at scale

Do the maths. A 90-minute class, 25 students. Even if the teacher never said a word, each student would get a little over three and a half minutes to speak. In reality it is far less. Companies see the same pattern: one group lesson a week leaves little room for individual practice.

An AI tutor does not solve everything, but it removes that arithmetic limit. Every learner can practise alone, whenever they like, for as long as they like. They can get it wrong, try again and rephrase, with no audience. For many people this matters as much as the volume: it is easier to take risks in front of a machine that does not judge.

Progress in speaking depends on something classrooms rarely provide: time for each person to talk, often.

Real-time feedback: correcting without interrupting

Speaking a lot is not enough if the same mistakes keep coming back. The other contribution of AI is instant feedback. The learner has just said "I am agree": the correction arrives straight away, with the right form and a short explanation, while the sentence is still fresh.

Research points the same way, with nuances worth knowing. A meta-analysis by Ngo, Chen and Lai, published in ReCALL in 2024, pooled 15 studies (38 effect sizes, published between 2008 and 2021) on speech recognition for English pronunciation. Its findings:

• the overall effect is medium (g = 0.69), so it is real;
• it is stronger when corrective feedback is explicit, meaning the tool says what is wrong and how to fix it;
• it is large for individual sounds and smaller for intonation and rhythm;
• training of four weeks or less did no better than teaching without speech recognition;
• practising with peers produced a larger effect than practising alone.

Those last two points are valuable. AI rewards regularity, not a two-week sprint. And it works better inside a human programme than on its own.

Personalising by profession and by level

Personalisation is often described as adjusting difficulty. That is the easy part. The Common European Framework of Reference for Languages (CEFR) provides the levels, from A1 to C2, and a good tool adapts its pace, vocabulary and follow-up questions to the learner's level.

The personalisation that really changes results is about context. A salesperson does not need the same words as an engineer. A law student does not prepare for the same situations as a medical student. A general course teaches you to order a coffee; a learning path built for a profession teaches you to defend a price, explain an outage or present a clinical case.

Here is what that looks like in practice, in two scenes from the learning paths of Alyx for Business and Alyx for Universities.

Scene 1: a sales manager negotiates a contract renewal in English

She works for a software company, around B1 to B2 in English. Opposite her, the AI Teacher plays a client's head of procurement: "Your renewal proposal is 12% above what we paid last year, and our budget is flat. Why should we accept this increase?"

She answers out loud. Two mistakes slip through. The first is "I am agree", copied word for word from French, Spanish or Portuguese. The second is "eventually", a false friend for speakers of French, Spanish or Portuguese, which means "in the end" rather than "possibly". The correction appears during the conversation without cutting it short: "I agree the increase is significant", "we could potentially reduce the increase". The negotiation carries on. At the end she gets scores for grammar, pronunciation, comprehension and vocabulary, plus a precise note: in "significant", the stress falls on the second syllable.

Tomorrow she will run the scene again. Next week, with a tougher objection.

Scene 2: a student prepares for an internship interview

A master's student in marketing, level B2. The AI Teacher plays a recruiter at an international company and asks her to introduce herself. She says "I study marketing since five years", and the correction suggests "I have been studying marketing for five years", with a short explanation of the present perfect and the difference between for and since.

The next question is harder: describe a disagreement within a team. She politely asks for the question to be rephrased, then answers. The feedback goes beyond language. It notes that she picked up the thread after a difficult question, that her answer followed a clear structure (situation, action, result), and suggests a pause rather than "erm" before answering. That is exactly what a good interview coach would say, if they had time to listen to her twenty times.

AI for language teaching: the teacher's role

The question comes up at every training day: will AI replace language teachers? Our answer is no, and the research quoted above hints at it too: fully solitary practice has a smaller effect than supported practice.

The division of labour is fairly clear. AI handles what a person cannot provide at scale: repetition, availability at any hour, correction of every sentence. The teacher keeps what a machine does badly: understanding why a learner is stuck, choosing this month's priority, explaining a misunderstood rule another way, encouraging, following up with the learner who is drifting away.

In the programmes we build, teachers see what each learner gets right and what they need to work on. They assign targeted activities and adjust the study plan. Freed from basic correction, class time goes to what matters.

For language schools and independent teachers, it goes further: AI can extend their own method. On the Alyx platform for language schools and teachers, the AI Teacher is created from the teacher's own face, voice and way of teaching, inside an app in the school's branding. Between lessons, students keep practising with "their" teacher.

Measuring progress instead of assuming it

For a long time, language training was evaluated with attendance sheets and satisfaction surveys. You knew who had turned up, rarely who had improved.

Because every conversation is analysed, AI makes a different kind of measurement possible: real practice time, scores session after session, progress by skill, recurring mistakes by team or cohort. A learning and development manager can see that the sales team now handles contract vocabulary but still trips over certain false friends. A teaching team can spot, in a cohort of several hundred students, the ones who have stopped practising.

This measurement has a limit worth stating: a session score is not a certification. To certify a level, recognised exams (TOEIC, TOEFL, IELTS, Cambridge) remain the reference. The value of AI data lies elsewhere: steering continuously, instead of discovering the results at the end of the year.

Risks and good practice

Enthusiasm should not hide the risks. Three deserve particular attention.

Data. A speaking session produces voice recordings, transcripts and error logs. That is personal data. Ask where it is hosted, how long it is kept, who can access it and whether it is used to train third-party models. UNESCO's guidance on generative AI in education and research stresses data privacy and the need for institutions to validate tools for their ethical and pedagogical appropriateness. At Alyx, for example, data is hosted and processed in France, speaking session recordings are deleted automatically after 30 days, and teachers see the performance report, not the student's audio (details on our AI and data page).

Quality. A language model can be wrong, accept an awkward phrase or "correct" a sentence that was fine. Good tools frame the conversation with a scenario and a level, and a teacher can still check. In Europe the rules are also taking shape: the EU AI Act classes as high risk the AI systems used in education that may determine access to education or the course of someone's professional life, such as exam scoring. According to the European Commission's page on the regulation, the obligations for this category apply from December 2027.

Over-reliance. A learner who only ever talks to an AI gets used to a partner that is patient, predictable and always available. Real conversations are rougher. AI should prepare people for them, not stand in for them: sessions with a teacher, group work, real situations.

How to choose an AI tool for learning or teaching a language

Whether you run a school, lead HR or coordinate languages at a university, these questions save time:

1. How much of the time is spent speaking? If written exercises still dominate, the tool is not using what AI does best.
2. When and how does correction arrive? During the conversation, with an explanation, or in a report nobody reads?
3. Does personalisation reach the profession or field of study? Ask to see a scenario from your sector, with your vocabulary.
4. What role does the teacher have? Can they follow each learner, assign activities and adjust the learning path?
5. What does the dashboard measure? Practice time, progress by skill, views by team or cohort, a report you can share.
6. Where is the data, and for how long? Hosting, retention period, any reuse.
7. Which languages and levels? Check that your target languages are covered at every level you need.

One last piece of advice: turn down the generic demo. Ask for a learning path to be built in front of you for one of your professions or fields of study. Within thirty minutes you will know whether the tool understands your reality.

Where to start

AI for languages delivers when it is used for what it does well: getting people to speak, correcting them on the spot, adapting content to their real life and making progress visible. It disappoints when people expect it to replace teaching.

Start small. Pick one profession or field of study, two or three priority situations, a pilot group. Give it a few months of regular practice, look at the data with the teacher, adjust. It is less spectacular than an announced revolution, and far more effective.