Canadian Language Learner gives professionals daily French and English practice in the format of the Second Language Evaluation, the government standard, with AI built in throughout, from the speaking tutor to your daily plan and progress reports for managers.
Levels shown are practice estimates. Official SLE results come only from the Public Service Commission.
Speak with an AI partner that stays in character, and get writing and pronunciation feedback in seconds.
Try a role-play →Each day, the AI picks your practice from the 12 methods based on your recent errors, your target profile and the time you have.
See the 12 methods →For every learner, the AI turns practice into a readiness estimate, the main gaps to target and a suggested next step.
See the manager view →Every bilingual position has a language profile: one letter each for reading, writing and oral. The platform estimates yours as you practise and aims each session at the gap.
Simple questions and answers on familiar topics.
Describe, explain and give instructions on concrete work topics.
Defend a position, handle hypotheticals and complex or sensitive subjects.
Always in this order: reading, writing, oral.
Pick a skill, then a method. Each skill mixes SLE-format practice with other ways to learn, in sessions of 3 to 30 minutes. Each day, the AI builds your plan from them. See our strategy
Fill-the-blank and text questions, timed like the real test.
SLE formatMemos and bulletins. Tap any word for its meaning.
Guided reading60 seconds per text to find the key point.
Timed drillFlashcards of words you looked up, spaced so they stick.
Spaced repetitionFind the mistake in workplace emails and notes.
SLE formatChoose the right word: grammar, agreement, formal tone.
SLE formatDraft a short email. The AI marks and explains each error.
AI feedbackOne rule, three examples, five checks, drawn from your errors.
Short lessonWorkplace scenarios with a caller, colleague or employee.
ConversationAnswer surprise hypotheticals and defend a position.
Timed promptsRepeat after Canadian French or English clips. Get a pronunciation score.
Listen and repeatStarts easy, gets harder, ends with a readiness report.
SLE formatManagers see each team member's estimated profile against their position requirement, without chasing anyone for updates. The AI flags who is ready, who is stuck and what to assign next.
Profiles, targets and practice time for the whole team on one screen.
Trends by skill and mock interview results, with an AI summary of gaps and next steps.
Shared gaps, such as oral hypotheticals, so training goes where it counts.
| Learner | Estimate | Target | Status |
|---|---|---|---|
| Maya T. | BBB | CBC | Oral gap |
| Daniel O. | CBC | CBC | Ready to test |
| Priya S. | BBA | BBB | Oral gap |
| Luc G. | CBB | BBB | Maintaining |
Behind every row in the table is a page like this one: progress over time, practice habits and what stands between the learner and their target.
Conditional after “si” (6 slips in last mock) · short answers to hypothetical questions · few linking words.
AI suggestion: 3 “What if?” drills this week, then a mock interview on October 14.
Suggestions sorted by level, so the language becomes part of the week and not only the sessions.
Radio-Canada news and public affairs.
Daily police drama at a natural Montreal pace.
Michel Rabagliati. Everyday dialogue with pictures to help.
Jean-Marc Vallée. Informal speech and Quebec expressions.
Kim Thúy. Short chapters in careful prose.
Small daily exposure that adds up.
The application, recordings and AI services run in Canadian cloud regions.
Architecture designed to Protected B controls, with encryption, logging and continuous monitoring.
Single sign-on through your organization's identity provider. No new passwords.
A fully bilingual interface, built to WCAG 2.1 AA.
Per-learner, per-year pricing with volume rates for larger organizations. Prices are provided on request.
We set up one cohort, measure progress and hand you a report you can take to your leadership.
We use AI where research shows it helps learners: speaking without an audience, getting instant feedback and practising every day. Every design choice traces back to a study.
Speech recognition tuned for Canadian French and English feeds an AI partner that stays in character and pitches its language to your level.
ResearchNearly half the studies in a 2024 review found chatbot practice eased speaking anxiety.Every answer updates your A, B or C estimate for each skill, and the next session is aimed at the gap to your target.
ResearchLearners often misjudge their own progress: effortful practice feels less productive even when it works better.The AI flags errors and offers hints. It never rewrites your work, so the thinking stays with you.
ResearchStudents with an unguided chatbot scored 17% lower once it was removed. A tutor with guardrails largely avoided that drop.Numbered notes explain each error in plain language, and you make the fix yourself before resubmitting.
ResearchAI and teacher feedback produced similar score gains, but pasting in corrections improves the essay more than the writer.Word-by-word scoring flags only the sounds and stress patterns that make you harder to understand.
ResearchHeavily accented speech can be fully intelligible (Munro and Derwing, 1995). Clarity is the target, not a native accent.Words you look up return as flashcards just before you would forget them, and harder cards come back sooner.
ResearchRecalling a word with effort strengthens memory far more than rereading it.0.48 to 0.80 Effect sizes reported by five meta-analyses: learners who practise with chatbots improve more, with mostly medium-sized gains. See the analysis
Read the research behind the productWhat the research actually says about using AI to learn a second language, and how those findings shape what we build.
For anxious language learners, AI chatbots are becoming a rehearsal room. The research says they help, with one important catch.
Read the articleAI can make your messages in a new language look flawless. New research suggests that's exactly what can keep you from learning it.
Read the articleNearly every language app now promises AI-powered results "backed by research." Here's what that research actually shows, and how to read it for yourself.
Read the articleResearchers have put AI writing feedback head to head with human teachers. The results are closer than you might think, and more complicated.
Read the articleAI pronunciation apps can now give you feedback on every sound you make. The research says they help, as long as you point them at the right goal.
Read the articleCanadian Language Learner is a practice platform for the Second Language Evaluation, the government standard for official languages.
Many roles require a specific profile, such as BBB or CBC, to work in a bilingual position or supervise a team. Classes help, but levels move in the practice between classes, and it is hard to find a patient speaking partner at 7 a.m. or over lunch.
We built the platform around the format of the real tests, with an AI tutor for the part most people find hardest: speaking under pressure. Managers get a clear view of who is on track, so training budgets go where they are needed.
No one should meet an SLE question type for the first time on test day.
Our AI flags errors and gives hints. It does not rewrite your work, because the struggle is where the learning happens.
Read the articleAn AI partner lets nervous learners rehearse without an audience, then carry that confidence into real conversations.
Read the articlePronunciation feedback targets what makes you hard to understand, not whether you sound like a native speaker.
Read the articleOur levels are estimates. Only the Public Service Commission issues official SLE results.
Data stays in Canada, and content reflects Canadian French and real workplace situations.
We start with a pilot for one cohort and a report you can take to your leadership.
See how a pilot worksWorkplace texts with four-option questions, timed like the real test and ordered from short and simple to long and complex.
Objet : Fermeture temporaire du stationnement
À compter du lundi 3 novembre, le stationnement souterrain de l'édifice sera fermé pendant environ six semaines en raison de travaux de réfection. Les employés qui détiennent un permis pourront utiliser le stationnement municipal situé au 150, rue Laurier, sans frais supplémentaires. Pour obtenir un laissez-passer temporaire, veuillez communiquer avec l'équipe des installations avant le 31 octobre.
Nous vous remercions de votre compréhension.
7Selon cette note, que doivent faire les employés qui ont un permis ?
Correct.Not quite. The memo asks permit holders to get a temporary pass before October 31.
8On peut conclure que la fermeture…
Correct.Not quite. “Sans frais supplémentaires” means permit holders will not pay more. The other options contradict the text.
Sample item written for Canadian Language Learner in the official format. It is not an official test question.
Subject: Temporary closure of the parking garage
Starting Monday, November 3, the building's underground parking garage will be closed for about six weeks for repair work. Employees who hold a parking permit may use the municipal lot at 150 Laurier Street at no extra cost. To obtain a temporary pass, please contact the Facilities team before October 31.
Thank you for your understanding.
7According to the memo, what should permit holders do?
Correct.Not quite. The memo asks permit holders to get a temporary pass before October 31.
8It can be concluded that the closure…
Correct.Not quite. “At no extra cost” means permit holders will not pay more. The other options contradict the text.
Sample item written for Canadian Language Learner in the official format. It is not an official test question.
Real-world memos and bulletins. Tap any highlighted word for its meaning and an example, then save it for review.
Le nouveau formulaire de lundi. Les demandes seront à l'employé. Veuillez vos reçus originaux et le tout à votre gestionnaire.
Tap a highlighted word to see its meaning.
Words saved this week: 11
The new form Monday. claims will be to the employee. Please your original receipts and everything to your manager.
Tap a highlighted word to see its meaning.
Words saved this week: 11
Sixty seconds per text to find the key point. It trains the speed you need to finish the real test.
Afin de réduire les délais de traitement, la Direction met à l'essai un système de prise de rendez-vous en ligne pour les clients. Pendant les trois mois du projet pilote, les agents noteront les problèmes signalés et proposeront des améliorations. Une décision sur le déploiement à l'échelle nationale sera prise au printemps.
Quelle est l'idée principale du texte ?
Correct.Not quite. The text is about a three-month pilot of online booking. Spring is only when a decision will be made.
To reduce processing times, the branch is testing an online appointment booking system for clients. During the three-month pilot, officers will log the issues reported and suggest improvements. A decision on a national rollout will be made in the spring.
What is the main idea of the text?
Correct.Not quite. The text is about a three-month pilot of online booking. Spring is only when a decision will be made.
Flashcards of words you looked up, shown again just before you would forget them.
Find the incorrect part in workplace sentences, or confirm there is no error, exactly as the written test asks.
Choose the bolded part that contains an error. If there is no error, choose E.
12Les documents hier à la directrice, .
Correct.Not quite. With avoir, the past participle agrees with a direct object placed before it: “que” stands for “les documents”, so “envoyés”.
13 le rapport terminé, nous encore approuver.
Correct.Not quite. No error: “bien que” takes the subjunctive (“soit”), and “le faire approuver” is correct.
Sample item written for Canadian Language Learner in the official format. It is not an official test question.
Choose the bolded part that contains an error. If there is no error, choose E.
12The documents yesterday to the director, .
Correct.Not quite. The subject is “the documents”, which is plural, so the verb is “have been forwarded”.
13 the manager the analysts available .
Correct.Not quite. No error: with “neither… nor”, the verb agrees with the closer subject, “the analysts”, so “were” is correct.
Sample item written for Canadian Language Learner in the official format. It is not an official test question.
Choose the right word for the context: grammar, agreement, verb tense and formal register.
Choose the word or phrase that best completes the text.
4Nous vous prions de bien vouloir nous faire parvenir les documents requis ______ le 15 novembre.
Correct.Not quite. “Au plus tard” (at the latest) is the standard phrase for a deadline.
5Si le client ______ plus tôt, nous aurions pu régler le dossier.
Correct.Not quite. A past hypothetical uses si + plus-que-parfait, then the past conditional: “si le client avait appelé…, nous aurions pu…”.
Sample item written for Canadian Language Learner in the official format. It is not an official test question.
Choose the word or phrase that best completes the text.
4Please send us the required documents ______ November 15.
Correct.Not quite. “No later than” is the standard phrase for a deadline.
5If the client ______ earlier, we could have resolved the case.
Correct.Not quite. A past hypothetical uses if + past perfect, then could/would have: “if the client had called…, we could have…”.
Sample item written for Canadian Language Learner in the official format. It is not an official test question.
Draft a short work email. The AI marks each error and explains it, and you make the fix yourself.
Task: Write an email to your team to postpone Thursday's meeting (80–100 words).
Bonjour à tous,
Je vous écris pour vous informer que la réunion de jeudi est reporté1 à mardi prochain. Je suis désolé pour l'inconvénient2. Si vous avez des questions, n'hésitez pas de3 me contacter.
Merci de confirmer votre présence4 d'ici vendredi.
Cordialement,
Maya
“La réunion” is feminine, so the participle takes -e.
“Désolé pour l'inconvénient” copies the English phrase. This version sounds natural in a work email.
The verb “hésiter” takes “à” before an infinitive.
Clear, formal request. Keep this phrase.
Task: Write an email to your team to postpone Thursday's meeting (80–100 words).
Hi everyone,
I am writing to let you know that Thursday's meeting is postponed at1 next Tuesday. Sorry for the inconvenient2. Please confirm your presence3 by Friday.
Thank you for your flexibility4.
Best regards,
Julien
You postpone something to or until a new date, not “at”.
“Inconvenient” is an adjective. The noun is “inconvenience”.
“Confirmer votre présence” translates word for word. In English, you confirm your attendance.
A natural, polite closing line. Keep it.
One rule, three examples, five checks. Each lesson is picked from the errors you make most.
Avec avoir, le participe passé s'accorde avec le complément direct placé avant le verbe.
1Les courriels que j'ai ______ ce matin.
Correct.Not quite. The direct object comes before the verb, so the participle agrees.
2J'ai ______ la lettre hier.
Correct.Not quite. The direct object comes after the verb, so no agreement.
3Cette erreur, nous l'avons ______.
Correct.Not quite. The direct object comes before the verb, so the participle agrees.
4Combien de dossiers avez-vous ______ ?
Correct.Not quite. The direct object comes before the verb, so the participle agrees.
5Les notes qu'elle a ______ sont claires.
Correct.Not quite. The direct object comes before the verb, so the participle agrees.
Score : 0 / 5
Use the simple past for a finished time (yesterday, last week, two days ago). Use the present perfect for a period that is still open (this week, since, yet, ever).
1I ______ the file last Monday.
Correct.Not quite. A finished time (last Monday, two days ago) takes the simple past.
2She ______ in this unit since 2021.
Correct.Not quite. An open period (since, yet, ever) takes the present perfect.
3We ______ the new policy yet.
Correct.Not quite. An open period (since, yet, ever) takes the present perfect.
4They ______ the contract two days ago.
Correct.Not quite. A finished time (last Monday, two days ago) takes the simple past.
5______ you ever written a briefing note?
Correct.Not quite. An open period (since, yet, ever) takes the present perfect.
Score : 0 / 5
Workplace scenarios with a caller, a colleague or an employee. Speak out loud and get feedback when you finish.
Vous êtes agent·e au service à la clientèle. Un client appelle parce que sa demande a été refusée. Expliquez la décision et les options qui s'offrent à lui.
Bonjour, j'ai reçu une lettre qui dit que ma demande est refusée. Je ne comprends pas pourquoi.
Bonjour, je comprends votre frustration. D'après votre dossier, il manquait une pièce justificative.
Mais je l'ai envoyée ! Et si je la renvoie aujourd'hui, est-ce que ce sera réglé cette semaine ?
Si vous la renvoyez aujourd'hui, nous pourrons réviser votre dossier d'ici dix jours ouvrables.
You are a client services officer. A caller's application was refused. Explain the decision and the options available.
Hi, I got a letter saying my application was refused. I don't understand why.
Hello, I understand your frustration. According to your file, a supporting document was missing.
But I sent it! If I send it again today, will it be sorted out this week?
If you send it today, we will be able to review your file within ten business days.
Answer surprise hypotheticals and defend a position, the tasks that separate level C from level B.
Si votre équipe perdait la moitié de son budget l'an prochain, quelles activités garderiez-vous en priorité ? Pourquoi ?
If your team lost half its budget next year, which activities would you keep as priorities? Why?
Repeat after short Canadian clips and get a word-by-word pronunciation score.
Nous avons pris la décision de prolonger la période de consultation jusqu'à la fin du mois.
Nous‿avons pris la décision de prolonger la période de consultation jusqu'à la fin du mois.
In the product you hear the clip and record yourself. This preview highlights the words instead of playing audio.
We have decided to extend the consultation period until the end of the month.
We have decided to extend the consultation period until the end of the month.
In the product you hear the clip and record yourself. This preview highlights the words instead of playing audio.
A full practice interview that starts easy, gets harder and ends with a readiness report.
« Dans quelle mesure pensez-vous que l'intelligence artificielle changera la façon dont votre équipe sert le public ? Et si vous deviez la mettre en place demain, par quoi commenceriez-vous ? »
Most answers held at B. Opinions and hypotheticals are where level C is decided, so your plan now weights “What if?” drills.
“To what extent do you think artificial intelligence will change the way your team serves the public? And if you had to introduce it tomorrow, where would you start?”
Most answers held at B. Opinions and hypotheticals are where level C is decided, so your plan now weights “What if?” drills.
Every language learner knows the moment. You have studied for months. You know the words for "coffee" and "please" and "to go." Then the barista looks up, waits, and something in your head goes quiet. Your face warms. You point at the menu, or you say it in English, and you walk out holding a drink and a small sense of defeat.
It's tempting to treat that freeze as a personal weakness, a sign you aren't cut out for languages. Researchers see it differently. To them it is one of the most thoroughly documented obstacles in language learning, and it has a name. Increasingly, it also has an unlikely remedy: a conversation partner that isn't a person at all.
In 1986, Elaine Horwitz, Michael Horwitz and Joann Cope published a paper that shaped decades of research. They argued that the anxiety people feel when learning a new language isn't just general nervousness wearing a different outfit. It is its own distinct experience, rooted in the strange position of trying to express yourself with a fraction of your usual vocabulary (PMC).
They traced it to three sources. One is communication apprehension, the ordinary discomfort of talking in real time. Another is test anxiety, the worry about being graded. The third, and arguably the most corrosive, is fear of negative evaluation: the sense that the teacher, the classmate, or the native speaker across the counter is quietly judging how badly you sound (Wikipedia).
That third fear creates a trap. The natural way to avoid being judged is to stop speaking. But speaking is precisely the skill that needs practice. Researchers studying conversational AI note that anxiety can push learners to avoid speaking altogether, which is the surest way to stop improving (arXiv). The quieter you get, the harder it becomes to speak, and the more anxious you feel when you finally must.
This is where a chatbot changes the equation. The feature that makes it seem hollow as a conversation partner, the fact that nobody is on the other end, turns out to be its biggest advantage for nervous learners. There is no face to read for disappointment. No classmate is listening. You can start a sentence three times, abandon it, and try again, and nothing about the exchange will be remembered as embarrassing.
A team of researchers in Turkey framed it in the language of the original theory. If fear of negative evaluation is a central driver of speaking anxiety, then a partner that doesn't evaluate you in the social sense should lower that anxiety (Frontiers in Psychology).
There are practical benefits too. A chatbot is available at eleven at night, on a lunch break, or in a parked car five minutes before a job interview. You decide what to talk about and how fast to go. For people whose schedules or nerves make a weekly class feel impossible, that flexibility matters.
The idea is intuitive, but it has also been tested, repeatedly, and the results are encouraging.
In one four-week experiment in Wuhan, China, researchers divided 62 university students into two groups. Both practiced speaking topics drawn from the IELTS exam twice a week. One group practiced with Doubao, an AI chatbot developed by ByteDance that holds context-aware conversations and gives instant feedback. The other practiced without AI. By the end, the chatbot group's anxiety had fallen significantly more than the comparison group's (Springer).
The Turkish team took a different approach. Forty-eight first-year English teaching students each sat two comparable speaking exams, one conducted by an AI chatbot and one by human instructors, and filled out an anxiety questionnaire before each. Anxiety was measurably lower before the AI-run exam (Research Square).
Other studies suggest the benefits extend beyond feelings. When researchers added an AI chatbot to a common classroom technique called think-pair-share, students reported less anxiety and more enjoyment, and they scored better on a speaking test (Computer Assisted Language Learning). A study of the language app Mondly found that learners valued being able to practice privately, at their own pace, on topics of their choosing, and that those very features helped ease their nerves (Nature Humanities and Social Sciences Communications).
Zoom out and the pattern holds. A 2024 systematic review of research on chatbots for English speaking practice found that nearly half of the studies it examined pointed to anxiety relief as a benefit (ScienceDirect).
If the story ended there, it would be a tidy advertisement. It doesn't.
A review of chatbot studies published between 2010 and 2024 identified a couple of cases in which learners' anxiety actually rose after working with a chatbot. The authors linked it to context: when chatbot practice was tied to a graded course, students worried about failing, and the bot became one more thing to be evaluated by (ReCALL). Even in the Turkish exam study, where anxiety dropped overall, some students described the AI as unsettling because it felt strict and unfamiliar (Research Square).
The lesson is subtle but important. A chatbot doesn't lower anxiety simply by being a machine. It lowers anxiety when it feels like a safe place to experiment. The moment it starts to feel like a test, with scores and stakes and a sense of being watched, much of the old fear can return.
For learners who want to try it, the research points toward a particular way of practicing.
Start by making the conversation easy enough to succeed at. Tell the chatbot your level and ask it to speak slowly and simply. A conversation you can keep up with builds confidence; one that races past you does the opposite. Use voice mode rather than typing, since the fear most people need to work through lives in speaking aloud, not in writing.
Then rehearse the situations that actually make you tense. Ask the AI to play a waiter, a landlord, a job interviewer, or a doctor's receptionist, and run the scene several times until it feels routine. It can also help to ask the bot to hold its corrections until the end rather than interrupting you mid-sentence. Constant correction can feel like the very judgment you're trying to escape, and a short list of fixes afterward is easier to absorb.
Most importantly, decide in advance when you'll move on. The purpose of practicing with a machine is to become able to talk with people. After a week or two of rehearsal, book one real conversation, whether with a tutor, a language exchange partner, or a patient friend. The bot has done its job when you no longer need it to feel brave.
The freeze at the coffee counter is real, common, and well understood. For a long time, the standard advice was simply to push through it. What AI chatbots offer is a gentler on-ramp: a place to make mistakes without an audience, to build the habit of speaking before the stakes are high.
The research suggests that for many learners, it works. Just remember what the rehearsal is for. The chatbot is the practice room, not the stage. Use it to warm up your voice, and then go find someone to talk to.
Imagine a learner named Maria, six months into a year abroad in Madrid. Her Spanish messages are impressive. Her texts to her host family are warm and idiomatic. Her emails to her landlord are polite, precise and free of errors. She has never been more productive in Spanish.
Then the landlord calls about a broken water heater. There's no time to open a chatbot, and no box to paste a draft into. Maria reaches for the words and finds that they aren't there. Six months of writing in Spanish, she realizes, have not made her much better at Spanish. An AI has been doing the hard part.
Maria is a composite, but her problem is real, and it is becoming common. Generative AI makes it effortless to produce polished text in a language you barely know. That is a genuine convenience. It is also, according to a growing body of research, a trap for anyone whose actual goal is to learn.
Psychologists use the term cognitive offloading for the habit of handing mental work to an outside tool. Writing a shopping list is offloading. So is using GPS instead of memorizing a route, or a calculator instead of doing long division in your head.
Offloading isn't inherently harmful. A calculator lets a physics student concentrate on the physics. The trouble begins when the work you're offloading is the very skill you're trying to develop. And in language learning, that is almost always the case. The effort of pulling a word out of memory, assembling a sentence, or spotting your own mistake is not an obstacle between you and learning. It is the learning.
The clearest evidence for what happens when that effort disappears comes, surprisingly, from a high school math class. In a large field experiment published in the Proceedings of the National Academy of Sciences in 2025, researchers at the Wharton School gave nearly a thousand high school students access to GPT-4 during math practice (Semantic Scholar).
With the AI available, students performed better on their practice problems. The surprise came afterward. When access was taken away, students who had used a plain version of GPT-4 scored 17 percent lower than classmates who had never used it at all. The researchers concluded that students had been leaning on the AI as a crutch, and then performed worse on their own (SSRN).
There was a second group in the study, and it may be the most important part of the story. These students used a version of the tutor built with safeguards, designed to guide them toward solutions rather than simply supplying answers. That version largely avoided the harm (SSRN). The problem wasn't AI. It was AI that did the thinking for them.
A 2024 study by a separate team looked at the same phenomenon from another angle. In a randomized lab experiment, learners completed a writing task with one of four kinds of support: an AI chatbot, a human expert, writing analytics software, or nothing at all (arXiv).
The researchers gave a memorable name to the risk they were investigating: metacognitive laziness. Metacognition is the part of learning where you plan what to do, monitor how it's going, and judge whether it worked. When an AI takes over those jobs, the authors argue, offloading can harden into a habit of avoiding deliberate mental effort altogether (arXiv).
What makes this especially tricky is that it doesn't feel like a problem. One review of the research describes a study in which heavy AI use made university students feel more confident and efficient, even as it increased their dependence on the technology (arXiv). The sensation of progress and the reality of progress can drift apart without the learner noticing.
If you followed tech news in 2025, you may remember headlines about an MIT study that strapped EEG caps onto people writing essays. It deserves a mention here, along with a warning label.
The study involved 54 participants who wrote essays either with ChatGPT, with a search engine, or with no tools at all (Neuroelectrics). Those using ChatGPT showed the weakest patterns of brain connectivity and had noticeably more trouble quoting from essays they had just written (arXiv).
That finding spread quickly online, often in wildly exaggerated form. The authors themselves caution that their results are specific to essay writing in an educational setting and may not generalize (MIT Media Lab). Other researchers have questioned the study's small sample and its methods (arXiv). It's a preliminary result, not proof that chatbots damage your brain.
But one detail from the study is worth holding onto. Participants who wrote without help first, and only then used ChatGPT to revise, showed increased brain connectivity rather than less (Britannica). The order mattered. Think first, then bring in the machine.
Decades of memory research explain why this matters so much for anyone learning a language.
One of the most robust findings in learning science is that retrieving information strengthens memory far more than reviewing it. Trying to recall a word, even when the effort is uncomfortable, does more for retention than rereading it on a flashcard. In fact, the harder the successful retrieval, the stronger the resulting memory tends to be (ResearchGate).
The psychologist Robert Bjork calls these desirable difficulties: challenges that slow you down in the moment but improve what you remember over time (PMC). A 2026 study showed the effect with learners of Japanese. Those who practiced recalling kana characters with feedback remembered them better 30 and 50 days later than those who simply studied them again (PubMed).
There's a twist that makes this hard to act on. Learners are poor judges of which methods are working. Effortful practice often feels less productive than easy review, even when it produces better results (MDPI).
Put those findings together, and the trap comes into focus. When an AI supplies the word you were searching for, you skip the retrieval that would have cemented it. When it corrects your sentence, you skip the work of noticing what was wrong. The experience feels smooth and efficient. That smoothness is the problem.
None of this is an argument for abandoning AI. The math study already showed the way out: the tutor with guardrails didn't harm learning. A 2026 review of research in higher education reached a similar conclusion, finding that learners stay mentally engaged when AI offers hints, prompts and feedback rather than finished answers (PMC).
So the real question is what job you give the AI. Asking it to translate your message is offloading the skill. Asking it to quiz you on the vocabulary you'll need for that message is practicing the skill. Asking it to rewrite your paragraph hands over the thinking. Asking it to point out your three biggest mistakes, without fixing them, keeps the thinking with you.
A few habits make the difference. Try to write the sentence or recall the word yourself before asking for anything. When you're stuck, ask for a hint instead of the answer. Avoid sending AI-written text in a real conversation unless you could have written it yourself. And set aside at least one practice session a week with no AI at all, so you have an honest sense of what you can actually do.
For Maria, the fix wasn't to delete her chatbot. It was to change how she used it. She started drafting her messages herself and asking the AI only to flag errors. She had it quiz her on words before difficult conversations instead of scripting them. Her messages got a little less polished. Her Spanish got a lot better.
AI can make you look fluent long before you are. For a quick email, that's a gift. For learning a language, it's a detour. The discomfort of reaching for a word you almost know isn't a sign that something is going wrong. It's the feeling of the language becoming yours. Let the AI coach you through that moment, not carry you past it.
The routine has become familiar to millions of language learners. You finish an essay in your second language, paste it into ChatGPT, and ask what's wrong with it. A few seconds later, a tidy list appears. Your introduction lacks a clear thesis. The second paragraph wanders. You've confused two tenses in three places. The feedback is fast, free, endlessly patient, and available at two in the morning the night before a deadline.
It is also, many teachers suspect, not quite as good as it looks. Over the past two years, researchers around the world have tested that suspicion, comparing AI feedback with human feedback on the writing of real language learners. What they found doesn't support either the enthusiasts or the skeptics entirely. AI turns out to be a genuinely capable reader of student writing. But how much it helps depends heavily on what kind of feedback you need, how strong a writer you already are, and what you do with its suggestions.
One of the first things researchers noticed is that ChatGPT and human teachers don't read an essay the same way.
In a 2024 comparison by Guo and Wang, teachers tended to concentrate their comments on content and on language errors. ChatGPT spread its attention more evenly, commenting on content, organization and language in roughly balanced measure (ScienceDirect). A separate study of Chinese undergraduates found something complementary. Students did better overall with their teachers' feedback, but they were particularly successful at acting on ChatGPT's suggestions about organization (Computer Assisted Language Learning).
That points to a useful division of labor. Busy teachers marking a stack of essays often focus on the most visible problems, such as grammar slips and weak arguments. An AI reader, with no time pressure, may be better placed to comment on the architecture of a piece: how paragraphs connect, whether the argument builds, where a reader might get lost.
So which one actually helps students improve more? Across several recent studies, the honest answer is: it's close, with teachers holding a modest edge.
A 2025 study in Frontiers in Education compared AI-generated and teacher-generated feedback on argumentative essays by learners at two proficiency levels. Both kinds of feedback produced significant gains in writing scores, and the difference between the two groups was negligible (Frontiers in Education).
A six-week study in secondary school classrooms reached a similar conclusion with an important qualification. Both types of feedback helped students improve, but teacher feedback produced steadier, more consistent gains over the period. The study also found that ChatGPT's scoring of the essays tracked the teachers' scores closely, suggesting the AI was reading the writing in broadly the same way a human would (APJEE).
Where teachers pulled clearly ahead was in the quality of students' revisions. In the study of Chinese undergraduates, who received feedback from both their teachers and ChatGPT, students engaged more with the teacher comments and revised more accurately as a result (CALL). And when 112 learners in an advanced writing course in Saudi Arabia received both kinds of feedback, they rated their instructor's comments as the more useful of the two (ResearchGate).
The studies also reveal some specific weaknesses, and they matter most for learners who are just starting out.
The first is accuracy with weaker writing. When teachers and ChatGPT both assessed 100 essays written by Greek learners of English, the AI had trouble consistently identifying language errors in the lower-level essays (ScienceDirect). That's an uncomfortable finding, because beginners are exactly the writers who most need their errors pointed out.
The second is overload. In a study of 25 intermediate-level students at a university in Türkiye, learners generally appreciated ChatGPT's feedback and felt it helped them improve. But they also found it too long, not always specific enough, and sometimes written in language more advanced than their own (System). Feedback you can't understand is feedback you can't use.
The third is simple incompleteness. In one small pilot study, students revised their essays with ChatGPT's help, and their teacher still found errors afterward (ERIC). A clean-looking AI review is not a guarantee that the essay is clean.
The Turkish study turned up another finding that deserves attention, because it has less to do with the AI than with the people using it.
Students were much more willing to accept ChatGPT's comments on grammar and word choice than its comments on content or organization. And when they accepted language corrections, they often simply copied the AI's corrected version into their essay (System).
That approach improves the essay. It does much less for the writer. Pasting in a correction skips the step where you notice what went wrong and produce the right form yourself, and that step is where the learning happens. It's one reason that similar score gains between AI and teacher feedback don't necessarily mean similar learning. An essay can get better while its author stays exactly where they were.
Perhaps the most consequential pattern in the research is that AI feedback does the most for the learners who need it least.
A detailed case study followed four students of different proficiency levels as they worked with ChatGPT's corrections. All four engaged with the feedback in the sense of reading it and revising. But only the students with high proficiency engaged with it deeply, using deliberate strategies to understand the corrections and apply them (Nature Humanities and Social Sciences Communications).
Combine that with the AI's weaker accuracy on beginner essays, and a clear picture emerges. For intermediate and advanced learners, AI feedback can be a strong tool. For beginners, it is both less reliable and harder to use well, which makes a human teacher all the more valuable at that stage.
Most of this research has focused on learners of English, though that is beginning to change. A study of advanced Spanish learners found that ChatGPT-4 could provide detailed, rubric-aligned feedback across areas including argumentation, grammar and structure (System). Learners of other widely spoken languages can likely expect similar strengths and weaknesses, though the evidence is thinner.
The research suggests a few ways to make AI feedback work harder for you.
Start by telling the AI who you are. A line such as "I'm an intermediate learner of French, please give feedback I can understand" addresses the most common complaint, that the comments are pitched above the student's level. Then ask it to identify problems rather than fix them. Requesting the three most important issues in an essay, without corrections, keeps the thinking on your side of the screen and avoids the copy-paste trap.
It also helps to take one category at a time. Ask about organization first, revise, and only then ask about grammar. A short list of issues is far easier to act on than a wall of comments. Once you've made your own revisions, you can ask the AI to check whether you fixed things correctly. Over time, keeping a simple log of the mistakes that keep coming up gives you something no single round of feedback can: a picture of your own patterns.
Finally, if you have access to a teacher or tutor, don't think of it as a choice. Several of the studies point toward combining the two. In one pilot, a third of students said they preferred receiving feedback from both their teacher and ChatGPT (ERIC). A sensible approach is to use AI on early drafts and save the human reader for the final one.
So, should you trust ChatGPT with your essay? Mostly, yes, as long as you understand what you're getting. It's a fast, tireless and surprisingly perceptive reader, particularly good at spotting problems with structure. It also misses errors, can bury you in comments, and delivers its greatest benefits to writers who already have a solid foundation.
Treat it as a second opinion rather than a ghostwriter. Let it tell you where to look. Then do the fixing yourself, because the revision is where you become a better writer.
Scroll through any app store and the pitch is remarkably consistent. Talk to an AI tutor that never gets tired. Roleplay ordering dinner in Paris. Video-call an animated character who adapts to your level. Speak confidently in weeks. And, almost always, a reassuring phrase underneath: backed by science.
It's a fair question whether that's true. Language learning is a multibillion-dollar industry, and "AI" has become the feature every product must have. The good news is that there's now enough research to give a real answer. The less exciting news is that the answer is more modest, and more interesting, than the advertising suggests.
Individual studies can be misleading. A single experiment with forty students might find a dramatic effect, or none at all, for reasons that have little to do with the technology. That's why researchers rely on meta-analyses, which gather dozens of studies and combine their results into an overall estimate.
Those estimates are usually expressed as an "effect size." The details are technical, but the rough guide is simple: an effect around 0.2 is considered small, around 0.5 medium, and around 0.8 large (Nature Humanities and Social Sciences Communications).
Several meta-analyses have now examined chatbot-assisted language learning, and they tell a strikingly consistent story. A 2023 analysis of 18 studies found an overall effect of about 0.53 (ResearchGate). A 2024 analysis of 28 studies by Wang and colleagues, published in the Review of Educational Research, found about 0.48 (ResearchGate). A 2025 analysis of 31 studies by Lyu put the figure at about 0.61 (Wiley), and a pooled look at 21 studies of AI chatbots for learners of English landed at about 0.65 (ResearchGate).
In plain terms, across well over a hundred studies, learners who practiced with chatbots generally did better than those who didn't, and the advantage was medium-sized. That's a meaningful result. It's also not "fluent in weeks."
The averages hide some useful patterns.
The type of AI matters. Lyu's analysis found that chatbots powered by generative AI, the technology behind tools like ChatGPT, had significantly larger effects than older, rule-based bots that follow scripted conversations (Wiley). That suggests the recent wave of AI features isn't just marketing. The tools genuinely have become more capable.
Who you are matters too. One broad review of chatbots in education found that intermediate language learners benefited more than groups of mixed ability, and that mobile apps tended to outperform web-based tools (ScienceDirect). Another analysis found a much larger benefit for learning a foreign language than for studying one's native language (ResearchGate).
And results vary widely. A 2025 meta-analysis of 29 studies reported one of the highest overall effects, about 0.80, but it also noted significant differences between the studies it included (Interactive Learning Environments). An average of 0.5 can conceal studies where AI helped a great deal and others where it barely helped at all.
There's one more wrinkle that almost never appears in promotional copy. Studies that find exciting results are more likely to be written up and published than studies that find nothing. Researchers call this publication bias, and it tends to inflate the averages that meta-analyses report.
Good meta-analyses try to correct for it, and the correction can be sobering. One analysis of chatbot-assisted learning initially found a large positive effect. After adjusting statistically for publication bias, the effect shrank to small-to-moderate (ResearchGate). The benefit was still real. It just wasn't as impressive as the raw numbers implied.
Duolingo, the biggest name in language apps, is a useful case study, partly because it publishes more research than most of its competitors. That is genuinely to its credit. It also rewards careful reading.
Consider a peer-reviewed study on Duolingo Max, the company's AI-powered subscription tier. It found that learners felt more confident in their language skills after using the AI features for a month (Duolingo). That's a real finding, but notice what it measures. It's about self-efficacy, meaning a learner's belief in their own abilities. Confidence matters for language learning, as anyone who has frozen mid-conversation knows. But feeling more capable is not the same as a test showing you can speak or understand more.
The same study also made a thoughtful design choice. It tested both existing paying subscribers and free users who were given the AI features at no cost, specifically to check whether the novelty or price of the features might be driving the results (Frontiers in Education). That's the kind of detail that separates careful research from marketing.
Duolingo has also published a 2025 research report on its AI Video Call feature, studying Japanese learners of English over one month (Duolingo). Company reports can be informative, but they aren't independent, and it's worth waiting to see whether outside researchers find similar effects.
Some already have, at least for the app in general. A quasi-experiment with two classes totaling 40 students learning English found gains in engagement and in willingness to communicate after using Duolingo (IRRODL). A study of 93 Chinese students randomly assigned to AI-based instruction using Duolingo or to traditional instruction found that the AI group improved significantly more in speaking (PMC).
Then there are the numbers that travel on their own. You may come across claims such as "78 percent of roleplay users feel more prepared for real-world conversations." Figures like this often circulate on marketing and business blogs (example) without a clear link to a published study. Even when accurate, "feel more prepared" is a survey response, not a measured skill.
You don't need a statistics degree to evaluate an app's claims. A handful of questions will get you most of the way.
First, ask what was actually measured. Test scores and recorded speaking samples are strong evidence. Reports that learners felt more confident or enjoyed the app are weaker, even if they're pleasant to hear. Second, ask what the app was compared against. Beating a group that did nothing at all is easy. Beating a textbook, a class or a human tutor is impressive.
Third, look at the size and length of the study. Much of the research in this field involves fewer than a hundred people over a few weeks. That's a reasonable starting point, not a final verdict on whether an app will make you fluent over a year. Fourth, check who ran it. Company-funded research can be perfectly sound, but it carries more weight when independent researchers find the same thing.
Finally, think about who the learners were. A large share of this research involves university students learning English. If you're a fifty-year-old learning Korean on your commute, the results may or may not apply to you.
Put it all together, and a fair conclusion emerges. AI language tools work. The benefit is real, it appears to be growing as the technology improves, and it seems strongest for intermediate learners using mobile apps.
But "works" doesn't mean "replaces everything else." Most studies are small and short, results vary a great deal, and some of the most quoted numbers measure feelings rather than skills. The research supports treating an AI app as a strong supplement: a convenient way to practice every day, alongside reading, listening and, above all, real conversations with real people.
That may not fit on an app store banner. It is, however, what the science actually says.
You say the word once. A polite, blank stare. You say it again, more slowly, leaning on each syllable. Nothing. The third time, you give up and point at the menu.
Pronunciation may be the most frustrating part of learning a language, and it's often the most neglected. Classroom time is short, and a teacher with thirty students rarely has the chance to listen closely to each one and explain exactly which sound went wrong. Many learners simply absorb the idea that they "have an accent" and that there isn't much to be done about it.
AI speech technology promises to change that, offering instant, detailed feedback on the way you sound, any time you want it. It's a genuinely useful development. But before you spend hours chasing a perfect score, it's worth reconsidering what, exactly, you're trying to fix.
A great many learners set out to "lose their accent." Three decades of research suggest that's usually the wrong target.
In a landmark 1995 study, the Canadian linguists Murray Munro and Tracey Derwing pulled apart three qualities of speech that most people lump together (Journal of Second Language Pronunciation). Accentedness is how different you sound from what a listener expects. Intelligibility is whether the listener actually understands the words you meant to say. Comprehensibility is how much effort it takes them to understand you.
These turned out to be related but only partly connected. Their most influential finding was that speech can be heavily accented and still highly intelligible (Journal of Second Language Pronunciation). Anyone who has worked in an international office knows this intuitively. Some colleagues with strong accents are effortless to understand, while others with milder accents are hard to follow.
The finding reshaped the field. Pronunciation teaching moved away from accent reduction and toward the more achievable goal of being understood (Studies in Second Language Acquisition). Munro and Derwing also pointed out a sobering corollary: it's entirely possible to change your accent without becoming any easier to understand (ResearchGate). Effort spent on features that don't affect clarity is, in a practical sense, wasted.
So the real task isn't to sound like a native speaker. It's to find the handful of sounds and patterns that actually cause confusion, and to fix those.
Most AI pronunciation tools are built on automatic speech recognition, the same technology that powers voice assistants and live captions. You speak into your phone, and the system compares the sounds you produce with a model of how the words are typically pronounced. Many tools break that comparison down to the level of individual sounds, whole words and entire sentences, and highlight where you drifted from the model.
The newest tools add another layer. Instead of just flagging a sound in red, they use a language model to explain the problem in plain words, the way a tutor might: your stress fell on the wrong syllable, or you merged two vowels that English keeps separate.
The evidence so far is encouraging. A mixed-methods study had learners of English practice with speech recognition technology combined with feedback from classmates. The researchers concluded that the combination was an effective way to improve both pronunciation and speaking, compared with traditional teacher-led instruction (PMC).
In another study, 93 Chinese university students were randomly assigned either to AI-based instruction, using an app with speech recognition and personalized feedback, or to traditional classes. The AI group's speaking skills improved significantly more (PMC). A study of SpeechAce, a tool that rates pronunciation and fluency, similarly found that students' pronunciation test scores rose after practicing with it (ERIC).
Researchers are also examining how feedback should be delivered, not just whether it works. A 2025 study in the journal Language Learning compared three conditions: a visual display that highlighted mispronounced words and sounds, that same display paired with a written explanation generated by GPT, and no feedback at all. Notably, it focused on features closely tied to intelligibility, such as which words receive emphasis and how speech is grouped into phrases (Wiley). That's a promising sign. The next generation of tools may be designed around being understood rather than sounding native.
Other work compares tools that tell you explicitly what went wrong with tools that simply transcribe what they heard, leaving you to notice the mismatch yourself (arXiv). Both approaches have value, and the second one is available to almost anyone for free.
For all their usefulness, these tools come with limitations worth understanding.
The biggest is conceptual. Most pronunciation scores measure how far your speech is from a reference speaker. That's much closer to accentedness than to intelligibility. A disappointing score doesn't necessarily mean people struggle to understand you, and chasing a perfect one can pull you back toward the accent-reduction goal that research has largely moved past.
The technology also has its own blind spots. Every learner's pronunciation is shaped by their first language, and researchers are still developing systems that can reliably explain which differences come from that influence (arXiv). In practice, some tools may simply handle certain accents better than others.
And a capable engine doesn't guarantee a good learning experience. In the SpeechAce study, even as scores improved, the researchers concluded that the interface wasn't well suited to pronunciation learners and needed to be redesigned (ERIC).
The research suggests a few ways to get the most out of these tools.
The simplest test costs nothing. Speak into any app with live captions or dictation. If it writes down the wrong word, a human listener might well hear the wrong word too, and you've found a sound worth working on. From there, prioritize the errors that change meaning, such as "ship" and "sheep" in English, over features that merely mark you as a non-native speaker.
It also pays to practice rhythm, not just individual sounds. Stress and intonation have a large effect on how easy you are to follow, so work on whole phrases and sentences rather than words in isolation. A technique called shadowing, in which you listen to a short clip of a fluent speaker, repeat it immediately, record yourself and compare, pairs naturally with AI feedback.
Finally, check in with a real person every few weeks. Ask a tutor or language partner how easy you are to understand. An AI score is a stand-in for that judgment, and a useful one, but it isn't a replacement.
AI pronunciation tools offer something that used to be rare: detailed, individual feedback on the way you sound, available whenever you have ten minutes and a phone. The research indicates they can genuinely improve how learners speak.
The key is aiming them at the right target. You don't need to erase your accent. You need the barista, the colleague and the stranger on the train to understand what you mean. Use the machine to find what's getting in the way of that, fix it, and let the rest of your accent stay. It's part of who you are, and it tells people something true: that you took the trouble to learn their language.
