KoboshiPeople learning languages with ChatGPT, Gemini and Claude mostly talk to them, in the car or on a walk. They also design their own quizzes and ask for line-by-line explanations of manga, games and new scripts. The 15 public posts below, from February 2025 to September 2026, show three things. Speaking with an AI is easy to start and carries no embarrassment. Pronunciation feedback is its weakest part. Words met in a chat fade unless the learner moves them somewhere they will be reviewed.
When do people practise speaking with an AI?
Mostly while driving or walking, with the phone in a pocket.
A French learner at A2 to B1, with a human tutor one hour a week, uses ChatGPT’s conversation mode (Hacker News, 2025-05-21): “I tend to use it in the car, just asking about random facts or ideas or playing 20-questions.”
A Spanish learner, on the ChatGPT app hands-free: “I’ve used this to carry an hour long foreign language practice in Spanish while walking my husky.” Their tip: “you have to give it a topic at the beginning otherwise it won’t be able to drive the conversation forward and will stick to banal pleasantries.” So they paste in a short news item first (Hacker News, 2025-06-26).
A native Afrikaans speaker in the United States, with few people nearby who speak it: “I’ve been using Gemini to live chat in Afrikaans and do impromptu Afrikaans grammar lessons during my solo drives around town.” The result: “It’s been really, really nice getting to speak my language regularly again” (Hacker News, 2026-09-15).
Why does talking to an AI help learners speak?
The AI asks, the learner has to answer, and no person is watching.
A Japanese learner at about JLPT N4, driving 30 to 60 minutes a day, tried ChatGPT, Grok, Claude, Gemini and several tutor apps in voice mode: “I could summarise my experiences by saying they all suck tremendously and none of them sound remotely like any Japanese person I’ve ever spoken to in my life.” Still: “regardless of what the LLM is outputting, I am forced to come up with a response in Japanese. It’s great practice and there’s zero embarrassment or awkwardness” (WaniKani Community, 2026-01-13).
A Mandarin learner on ChatGPT and Gemini names the gap: “I can study vocab and flash cards all day but the second that voice starts speaking sentences and I need to understand in real time, I freeze up.” And the limit: “they quickly lose the plot regarding their role as a tutor” (Hacker News, 2025-12-24).
What else do learners ask the AI to make?
Their own drills, and explanations of material they care about.
A Hindi learner who started with Pimsleur, an audio course, uses ChatGPT three ways: planned voice role-plays (“With the new GPT Voice, I have basic, planned conversations. Let’s go to a restaurant.”), self-designed quizzes, and reading practice in Devanagari script: “I ask it to write childrens stories for me, but I ask it to write each line in the hindi script, then phonetic spelling of the hindi script, and then in english” (Hacker News, 2025-09-12).
A Japanese learner pastes dialogue from manga and the game Persona 5 into ChatGPT for a translation and breakdown: “It will explain things word for word and why it’s said the way it is. It even explains slang and more colloquial style speech” (WaniKani Community, 2025-02-07).
A learner of Mandarin and Japanese, studying with ChatGPT an hour three times a week: “I used voice mode on ChatGPT to learn the tones for mandarin”. They then checked their progress with a tutor on italki: “having it confirmed by a tutor was nice” (Hacker News, 2026-06-22).
Where does AI voice practice go wrong?
Mostly on pronunciation and interruptions.
- It forgives bad pronunciation. “ChatGPT voice mode seems to have too much tolerance for mispronouncing,” one learner wrote; “it even says your pronunciation is correct if asked” (Hacker News, 2025-06-26).
- The model voice can drift from native. A Korean learner, after a voice update: “Korean now often sounds more like an English speaker pronouncing Korean.” Why it matters: “The pronunciation produced by the tutor is the pronunciation we are being trained to imitate” (OpenAI Community, 2026-08-31).
- It interrupts while you search for a word. One user asked OpenAI to bring back hold-to-talk, because the model keeps “interrupting you while thinking (which happens ALOT learning a language)” (OpenAI Community, 2026-07-09).
What happens to the words after the chat?
Learners who keep them copy them into a review app, usually Anki (a free flashcard app), and that step takes effort.
An Anki user studying two languages: “Now I ask ChatGPT to generate me a whole bunch of example sentences for a particular topic or grammar point that I want to master, and I bulk import them into Anki in one go.” Then a check: “I don’t blindly trust ChatGPT’s output.” A native-speaker spouse reviews one language; other AI models check the other (Hacker News, 2025-11-02).
A Greek learner living in a Greek-speaking country, who used ChatGPT for practice and to read text from photos: “I also maintained decent retention by converting everything I learned into Anki cards” (Hacker News, 2026-02-13).
A Portuguese learner in Lisbon, on their routine with a human tutor before they built a tool to automate it: “After every lesson I’d sit down and make flashcards from my notes and screenshots. Spaced repetition works, but making the cards took manual effort each time. Most days I just didn’t do it” (Hacker News, 2026-04-13).
A Japanese learner wanted the reverse, a chat that knows the words already studied: “an api could call your learned vocab, which Chat GPT could try to use more frequently in Japanese conversation”, “Rather than manually updating it on what vocab you know” (WaniKani Community, 2025-07-01).
How many people use AI to learn, and how many pay for it?
Many people already use AI to learn; few pay for it.
AI for learning. In a Pew Research Center survey of 5,123 U.S. adults, run from 2025-02-24 to 2025-03-02, 26% had used ChatGPT for learning, up from 8% in March 2023; among adults under 30, 46% had used it to learn something new (Pew Research Center, 2025-06-25).
AI for languages. Menlo Ventures surveyed 5,031 U.S. adults with Morning Consult in April 2025: “When Americans engage in learning or personal development activities, nearly half (43%) use AI for academic support, 36% use it to do research for work or school, 30% use it for learning languages” (Menlo Ventures, 2025-06-26). Also: “77% of people are actively learning, but only 18% use AI for it.”
Paying for AI. a16z’s The Top 100 Gen AI Consumer Apps, 7th Edition (2026-10-05): “In August, 4.5% of eligible consumers in YipitData’s U.S. e-receipt panel had an active paid personal subscription to at least one of ChatGPT, Gemini, or Claude, up from 2.1% a year ago” (a16z). The three figures come from different surveys of different groups, so no single ratio links them.
Voice. At Google I/O on 2025-05-20, Google said Gemini Live conversations “are five times longer than text-based conversations on average” (Google). Speak, an app that “uses voice-based scenarios to guide learners through everyday situations”, “surpassed $100 million in annualized revenue” and reached “roughly 15 million downloads since its debut”, according to Benzinga on 2025-11-20 (Yahoo Finance); Speak is private, so this is a press figure.
Can you practise speaking with an AI on a free plan?
Yes. ChatGPT gives free users limited voice access, Google made Gemini Live free on Android and iOS in May 2025, and Claude’s voice mode is in beta on its free plan. A free account is enough to start.
| App | Free plan | Paid plans | Languages | Source, date read |
|---|---|---|---|---|
| ChatGPT Voice | “Limited access to GPT-Live-1 mini. Limits may change.” | Go and Plus: 3 hours; Pro: 15 hours at $100, unlimited at $200; per rolling 24 hours | You set the language you speak most, and can ask it to speak another | OpenAI Help Center, 2026-10-07 |
| Gemini Live | “available to everyone on Android and iOS for free” (with camera and screen sharing), announced 2025-05-20; the current Gemini Live page states no price | a Pro plan gives “more access to Gemini Live” | not stated | Google, 2025-05-20; Gemini Live page, 2026-10-07 |
| Claude voice mode | “available to all plans (Free, Pro, Max, Team, and Enterprise)”, in beta, on mobile, desktop and web | same | “Support for languages other than English is in beta.” | Claude Help Center, 2026-10-07 |
Does speaking practice actually help you learn a language?
Yes, with a caveat. Research on conversation with feedback, which researchers call interaction, finds clear gains: Mackey and Goo’s 2007 meta-analysis of 28 interaction studies “found large effect sizes for learners who were engaged in interaction compared to those who were not”, reports a review in the journal Language Teaching (Loewen and Sato, 2018, Language Teaching). It adds: “there are numerous factors that impact its efficacy.”
The caveat: those studies measured conversation with people, before AI chat, so whether an AI partner helps as much is still open. Weak pronunciation feedback is one reason to keep a human tutor in the loop, as the italki learner did. For why understanding a language and speaking it are separate skills, see why you understand Spanish but can’t speak it.
How do you keep the words from an AI conversation?
Save them somewhere that brings them back on a schedule. Kim and Webb’s 2022 meta-analysis of 48 experiments with 3,411 learners found that spacing practice over time “had a medium-to-large effect on second language learning” (Kim and Webb, 2022, Language Learning). A chat thread has no schedule: the words stay where you met them.
Koboshi, if you study in the chat. Koboshi, a vocabulary app, connects to ChatGPT and Claude through one connector address, https://koboshi.app/mcp, with nothing to install (setup steps). In the chat, name the words to keep: “Save sobremesa and madrugar to Koboshi.” For a word new to Koboshi, your AI writes the meaning and Koboshi checks it before other learners see it. The words come back as quiz questions on an FSRS schedule, the review-timing method Anki also offers. Adding a connector in ChatGPT needs its $20, $100 or $200 plan; Claude’s free plan allows one custom connector, which is enough (Koboshi connector page). Koboshi has no voice tutor: you speak in your AI app, and Koboshi keeps the words you name.
Anki, if you already have decks. Anki is free and fully customisable, and many learners already use it: AnkiDroid, its Android app, had between 3.2 million and 4.0 million Google Play users during 2025 (AnkiDroid 2025 wrapped). The bulk-import routine above works today. With Anki you make or import each card yourself; some learners value that work, and others, like the Lisbon learner, skip it. Community-built connectors (MCP servers) can also link Claude to Anki while desktop Anki runs on your computer.
You can use both. Koboshi imports Anki’s “Notes in Plain Text” export, up to 500 words at a time (how to import Anki words into Koboshi).
How this was made: This piece was drafted with AI under an editor’s direction, revised after the editor read the full draft, and every quote and figure was opened at its source on the dates given. Stories were gathered on 2026-10-05 from Hacker News (9 searches of its public comment index, limited to 2025 and later), the OpenAI Community, WaniKani Community and LingQ forums (one search each), and 10 web searches. Of about 30 first-person accounts considered, 15 are quoted above; we skipped repeats by the same writer, posts about older models, product pitches, one-off tests and anyone who might be a minor. Each learner quote is copied, spelling included, from the original post, opened on 2026-10-05 and linked; usernames are left out. The voice table was rechecked on each provider’s page on 2026-10-07. Hacker News readers skew technical and committed, so these stories fit that group better than casual learners.