explainx / blog
A comprehensive guide to using AI for language learning in 2026: why Duolingo plateaus at B1, how to use Claude and ChatGPT as conversation partners, the best AI pronunciation tools, comprehensible input at scale, and a recommended stack by fluency goal.
Jun 25, 2026
Claude Tag puts Claude inside Slack as a persistent, proactive team member. The @ClaudeDevs team shared how they use it to write 65% of their own code — including most of what built Claude Tag itself. Here are 15 specific use cases across engineering, ops, data, support, and documentation that show exactly what delegating to @Claude looks like in practice.
Jun 25, 2026
AI meetings are full of terms that sound technical but are not that hard once someone explains them plainly. This guide covers the 50 concepts that come up most in real AI strategy conversations — what they mean, why they matter, and how to use them correctly. Grouped by theme so you can read the sections most relevant to your work.
Jun 24, 2026
BirdClaw is an open-source CLI and web app by Peter Steinberger that imports your Twitter archive, syncs live updates, and stores everything in a local SQLite database. Daily AI digests, bookmark search, follow graph queries, DM triage, and blocklist management — all offline-first, no ads, no algorithm.
Language learning has always had two expensive bottlenecks: enough input and enough output practice with real feedback. For most of history, getting both required living abroad, hiring tutors, or spending years finding native speaker conversation partners willing to correct your mistakes. AI has broken both bottlenecks simultaneously — and in 2026, that changes everything about how you should structure your study.
This guide covers what actually works, what the research says about each approach, and how to build a stack matched to your specific fluency goal. If you are learning for travel, read AI for travel planning next; for creative output in your target language, see AI for creative hobbies.
| Goal | Primary Tools | Expected Timeline |
|---|---|---|
| Casual / A1–B1 | Duolingo + Claude conversation practice | 6–12 months to B1 |
| Reaching B2 | Speak + graded input + Anki + weekly italki tutor | 12–24 months |
| Professional fluency (C1+) | Human tutor + intensive input + AI domain vocabulary | 24–48 months |
| Pronunciation improvement | ELSA Speak (English) / Speechling (multi-language) | Ongoing |
| Vocabulary retention | Anki with AI-generated cards from your own reading | Ongoing |
In the 1980s, linguist Stephen Krashen proposed the input hypothesis: we acquire language not by memorizing grammar rules but by understanding messages slightly above our current level — what he called i+1 (input plus one). You internalize grammar subconsciously by encountering it in context, repeatedly, at the right level of challenge.
The problem with i+1 has always been supply. Getting thousands of hours of content precisely calibrated to your current level — not too easy, not so hard you lose meaning — historically required either expensive tutors or years of patience sifting through native media that was mostly incomprehensible.
Merrill Swain's output hypothesis adds the other half: you also need to produce the language, receive feedback on that production, and notice the gap between what you said and what you meant to say. Silent reading and listening alone leaves learners who understand Spanish television but freeze when ordering at a restaurant.
Real output practice with real feedback used to mean a human interlocutor — a tutor, a language exchange partner, or an immersion environment. Finding the right person, scheduling sessions, and staying accountable was genuinely difficult.
In 2026, AI makes both abundant and cheap:
The result: the two most expensive inputs into language acquisition are now effectively free and unlimited for anyone with an internet connection.
Duolingo is the world's most-downloaded education app and, within specific constraints, genuinely effective. Its streak system and daily reminders have gotten millions of adults to study a language every day who otherwise would not. That habit formation is real and valuable.
The plateau problem is also real. Multiple studies of Duolingo users find that consistent learners typically reach A2–B1 (elementary to lower intermediate) after 12–18 months of regular use — and then stop progressing. The core reason: Duolingo's exercises rarely force authentic production. Most tasks are translation, word matching, or sentence assembly from word banks. You're recognizing correct answers rather than generating language from internal knowledge.
A user who has completed all Spanish Duolingo content can typically read simple texts and understand short sentences, but struggles to hold a spontaneous five-minute conversation. The app never required them to actually speak, improvise, or produce sentences without scaffolding.
In 2023, Duolingo launched Duolingo Max, adding two AI-powered features using GPT-4:
These are genuine improvements. The roleplay feature in particular addresses the output gap — it forces production in a low-stakes environment. However, the conversations are still contained within Duolingo's structured scenario library and are briefer and more scaffolded than real conversation practice. They are closer to controlled drills than free conversation.
Assessment: Duolingo Max is a meaningful step up from the base app. For A1–A2 learners, it provides a valuable introduction to conversational patterns. For learners already at B1 who want to reach B2, it is insufficient — the scenarios don't adapt deeply enough to your errors, and the feedback loop is not tight enough to drive rapid improvement.
Speak is the most purpose-built AI speaking tool on the market in 2026. Available for Spanish, French, Korean, Japanese, German, and English, it combines:
Speak's biggest advantage over using ChatGPT for voice conversation is that the feedback is targeted and tracked. It knows you consistently confuse the Spanish subjunctive in conditional sentences and will resurface that pattern. The pronunciation engine is also more specialized for language learning than general speech recognition.
What Speak doesn't do well: The AI conversational partner, while natural, occasionally lacks the ability to go deep on complex topics. For advanced learners wanting to discuss abstract ideas in their target language, the conversation can feel slightly constrained compared to a human tutor or an unconstrained LLM.
italki remains the dominant marketplace for human tutors (both professional teachers and informal community tutors), but has added AI features including conversation warm-ups, grammar explanations, and scheduling assistance. The platform's core value is still its human tutor marketplace, and the AI features are supplementary rather than transformative. For learners who budget one or two tutor sessions per week, italki's scheduling tools and session structure support are genuinely useful.
The general-purpose LLMs — particularly Claude (Anthropic) and ChatGPT (OpenAI) — have emerged as remarkably effective free conversation partners when configured correctly. They do not have Speak's pronunciation analysis or tracked progress, but they have capabilities no specialized app matches:
The key is configuring them correctly — which the next section covers in detail.
Most learners who try using LLMs for language practice get limited results because they start a conversation in English and switch to the target language only occasionally. The following strategies extract substantially more value.
At the start of any session, set this instruction:
"For this entire conversation, respond only in [target language]. Do not use English under any circumstances, even if I write in English or ask you a question in English. Respond to everything in [target language]."
This forces you into genuine immersion. When you don't know a word, you can ask in the target language — "How do I say X?" — and you'll practice that meta-communication skill that is essential in real conversations.
Add to your instruction:
"After each of my responses, add a brief correction section at the bottom marked '✏️ Corrections:' listing any grammatical errors, unnatural phrasing, or word choices that a native speaker would not use. Explain each correction in one sentence."
This gives you the tight feedback loop that distinguishes effective output practice from mere production. You're not just speaking into a void; you're getting immediate, specific correction on everything you write.
Paste any piece of text and ask:
"Rate this text on the CEFR scale (A1, A2, B1, B2, C1, C2) and explain what features of the vocabulary and grammar place it at that level. Then rewrite it at B1 level for a language learner."
This is invaluable for finding and grading authentic content. You can paste articles, song lyrics, film transcripts, or passages from novels and immediately know whether they're in your comprehensible input zone.
Instead of drilling isolated word lists, use this prompt:
"I want to practice these 10 Spanish vocabulary words: [list]. Have a casual conversation with me about [topic — movies, food, travel] and naturally use all 10 words over the course of our conversation. Underline each target word when you use it."
This is dramatically more effective than flashcard drilling because you encounter words in natural semantic context, connected to meaning and conversation, which is how long-term retention actually works.
LLMs have deep knowledge of register, formality, and cultural subtext. Use prompts like:
"In Japanese, explain the difference between using ます/です form versus casual form, and when using casual form with someone you've just met would be rude. Give examples."
Or:
"This Spanish phrase I used — 'Estoy muy emocionado' — is it natural? Would a Mexican Spanish speaker use it differently than a Spanish speaker from Spain?"
These questions are difficult to answer from textbooks but LLMs handle them well, drawing on broad exposure to natural language use across regions and registers.
Speech recognition has become genuinely good. In 2024–2026, the underlying models powering pronunciation tools achieved near-human accuracy on clear speech in major languages, and the better tools now go beyond "did I understand you" to "this specific phoneme was off and here's why."
ELSA (English Language Speech Assistant) is the most mature pronunciation tool for English learners. It uses deep learning trained specifically on non-native English speech to:
ELSA's curriculum approach means it's not just reactive — it actively builds your phonemic inventory over time. For non-native English speakers, particularly in corporate or academic contexts, it is the most evidence-backed pronunciation tool available.
Speechling takes a hybrid approach: you record yourself speaking target-language sentences, and the recordings are reviewed by human coaches who provide audio feedback on your pronunciation. In 2026, the platform added an AI layer that provides instant automated feedback between human review cycles, with human coaches still serving as the accuracy benchmark.
This combination — AI for immediate feedback volume, humans for nuance — works particularly well for tonal languages (Mandarin, Cantonese, Vietnamese) where subtle pitch distinctions require human ears trained on that specific phonological system.
Speak (mentioned above in the conversation tools section) integrates pronunciation analysis directly into its conversation practice flow. Rather than isolated pronunciation drills, it flags pronunciation issues that arise during actual conversation — closer to real use conditions. For learners who find isolated pronunciation practice dry, this integration into meaningful conversation is more sustainable.
Current AI pronunciation tools have one significant limitation: they are trained on relatively clean, deliberate speech and struggle to give feedback on the prosodic dimensions of fluency — rhythm, intonation contour, speech rate, and the natural reductions and contractions of fast native speech. A learner can pass every ELSA drill and still sound slightly robotic in real conversation because their intonation is flat. Human tutors and extensive listening to authentic native content remain the best interventions for prosody.
The Dreaming Spanish channel demonstrated that massed comprehensible video input — hours of native content pitched at learner level — can produce remarkable acquisition over time. The model works; the bottleneck was always generating enough content at exactly the right difficulty level for each individual learner.
AI has removed that bottleneck entirely.
Tools like Readlang and dedicated graded reader generators can now produce stories, news summaries, and dialogue scripts at any CEFR level on any topic in under a minute. A Spanish B1 learner interested in Formula 1 can get a B1-level article about the last Grand Prix — content that connects their passion to their target language, at exactly the right comprehension challenge.
For older approaches like Pimsleur (which has always relied on comprehensible audio input in carefully sequenced dialogues), AI extends the model: learners can now generate Pimsleur-style dialogue scenarios on topics Pimsleur never covered, voiced through text-to-speech with increasingly natural prosody.
Using any capable LLM, you can request:
"Write a 400-word story in B1 Spanish about a football player moving to a new team. Use only vocabulary and grammar appropriate for B1. After the story, provide a glossary of any words that might be above B1 and their English translations."
This approach means you never have to read something boring to practice your target language. Every piece of comprehensible input can be about something you actually care about — which research consistently shows improves both attention and retention.
One underrated feature of AI-generated input is dynamic difficulty calibration. When you're reading an AI-generated story and a sentence feels too simple, you can say "make this passage more challenging — I'm between B1 and B2." The content adjusts immediately. No waiting for the next textbook chapter or the next Dreaming Spanish playlist to catch up to your level.
Anki remains the gold standard for spaced repetition vocabulary review, but the card creation process has always been its biggest friction point — most learners find manually creating thousands of cards tedious enough to abandon the system.
In 2026, the workflow has changed:
The example sentences are the critical addition. Cards with the word in a natural, memorable context produce dramatically better retention than translation-only cards. AI generates plausible, natural-sounding sentences far faster than you could find them manually.
Readlang lets you click any word in a web page or uploaded text to look it up and automatically add it to a spaced repetition review queue. In 2026, its AI integration generates context-aware definitions and example sentences at the time of lookup. For learners who do extensive reading in their target language (a core B2+ strategy), Readlang bridges the gap between reading and vocabulary review automatically.
The research-backed insight is contextual encoding: words encountered and reviewed in the context of sentences you actually read, about topics you care about, form stronger memory traces than isolated translation pairs. The AI-powered workflow above realizes this principle at scale — every card in your deck comes from content you were genuinely trying to understand, connected to a specific moment of comprehension.
Google Lens and Apple Translate's live camera mode have matured into genuinely useful tools: point your phone at a restaurant menu, a street sign, a newspaper headline, or a handwritten note, and get an instant translation overlaid on the original. For travelers and learners in an immersion environment, this removes a major friction point and allows reading authentic environmental text that was previously inaccessible.
2025–2026 saw the first consumer-grade real-time translation earpieces reach the market with sub-second latency in major language pairs. Devices from companies like Timekettle and integrations with smart earbuds using on-device models can translate spoken conversation in near-real-time. These are most useful as comprehension aids in authentic conversation contexts (business meetings, social situations) rather than as learning tools per se — passively receiving translated input is not the same as actively practicing production.
For watching foreign-language TV and film, 2026 tools can now generate AI dubs with lip-sync that maps translated audio to the original speaker's mouth movements with reasonable accuracy. Platforms have started offering this as an alternative to subtitles. For language learners, this is actually a double-edged tool: watching in your native language is comprehensible and enjoyable, but it removes the acquisition benefit of exposure to the target language. The better use case is partial dubbing — watching the first episode dubbed to orient yourself, then switching to subtitles or no subtitles as your comprehension improves.
Given everything AI enables, it is worth being explicit about where it still falls short.
Human connection and accountability. Learning a language is hard. The weeks when you make no noticeable progress, when everything feels like static, are real and common. A human tutor who knows you, has tracked your journey, and can tell you "your accent in the imperfect tense is actually dramatically better than three months ago" is irreplaceable as a motivational anchor. AI does not know you; it has no longitudinal memory of your struggle.
Cultural immersion. Language is embedded in culture. The idioms, the humor, the taboos, the way different social classes speak differently, the regional pride people feel about their dialect — these are acquired through living inside a culture, not through conversation drills. AI can explain cultural context, but it cannot replicate the feeling of being in a market in Mexico City or a bar in Tokyo and understanding not just the words but the full human situation.
Authentic speech variation. Real native speakers do not speak the way AI systems do. They hesitate, reduce sounds, speak quickly, use regional slang, trail off sentences, interrupt each other, and deploy intonation patterns no text-to-speech engine has fully captured. The phonological development that comes from extended listening to authentic human speech — with all its messiness and variation — produces a different kind of fluency than AI conversation practice alone.
Motivation through relationship. Many learners who have achieved high fluency credit a specific person — a tutor, a language exchange partner, a romantic partner, a close friend — as the emotional engine of their acquisition. The intrinsic motivation of wanting to communicate with a specific human being is one of the most powerful drivers in the research. AI is a partner but not a relationship.
For learners who want to enjoy basic travel conversations, understand song lyrics, or satisfy a personal curiosity without intensive study.
| Component | Tool | Frequency |
|---|---|---|
| Habit formation | Duolingo (base or Max) | Daily, 10–15 min |
| Conversation practice | Claude / ChatGPT in immersive mode | 2–3x/week, 20 min |
| Vocabulary | Anki with AI-generated cards | Daily, 10 min |
| Listening | AI-generated graded audio / Dreaming Spanish | 2–3x/week |
Expected outcome: Functional A2–B1 within 6–12 months at this cadence. Conversational enough for basic travel and media consumption.
B2 is the threshold for genuine communicative competence — holding complex conversations, understanding news media, reading authentic literature with occasional dictionary lookups. It is the target most adult learners describe when they say they want to "be fluent."
| Component | Tool | Frequency |
|---|---|---|
| Conversation coaching | Speak app (AI) | 4–5x/week, 20–30 min |
| Comprehensible input | AI-graded stories + Dreaming Spanish or equivalent | Daily, 30–60 min |
| Vocabulary review | Anki + Readlang integration | Daily, 15 min |
| Human tutor | italki (community tutor or professional) | 1–2x/week |
| Grammar | AI error correction in conversation + targeted explanations | Integrated into conversation practice |
Expected outcome: B2 in 12–24 months at this intensity, depending on starting level and target language difficulty. B2 in Spanish or French is faster than B2 in Japanese or Arabic.
Professional fluency means speaking in your target language in work settings, presenting, negotiating, writing formally, and understanding regional accents and informal speech. C1–C2 requires thousands of hours of authentic input and production.
| Component | Tool | Frequency |
|---|---|---|
| Primary tutor | Professional italki teacher (domain-specific if possible) | 3–5x/week |
| Domain vocabulary | AI conversation practice on professional topics | Daily |
| Authentic input | Native news, podcasts, books — ungraded | 60+ min/day |
| Pronunciation refinement | ELSA / Speechling (ongoing) | 3x/week |
| Writing practice | AI correction of professional writing in target language | Weekly |
| Cultural immersion | Travel, online communities, native speaker friendships | Ongoing |
Expected outcome: C1 in 24–48 months at this intensity for distant languages (Chinese, Arabic, Japanese for English speakers). Closer languages (Spanish, French, Italian for English speakers) can reach C1 in 18–30 months.
Regardless of your goal, three principles from the research apply:
1. Output without feedback is not practice. Speaking into a void builds confidence but does not accelerate acquisition. Every session of production should include error correction — which is exactly what AI enables on demand.
2. Volume matters more than most learners realize. The research on comprehensible input suggests successful B2 learners have consumed hundreds to thousands of hours in their target language. AI makes high-volume, high-quality input available — but you still have to put in the hours.
3. Motivation is the meta-skill. The best AI tool you never open does nothing. Connecting language learning to genuine reasons — a country you want to live in, a person you want to communicate with, a career you're building — predicts persistence better than any methodology. Build your stack around content you actually enjoy, not content you think you should study.
In 2026, the argument for learning a language has never been stronger and the barriers have never been lower. The input and output bottlenecks that stopped most adult learners before they reached conversational fluency have been removed. What remains is a question of time, deliberate practice, and choosing tools matched to the stage you're actually at.
Start with what you'll actually do every day. Upgrade the tools as your level demands more.
More AI-for guides
Broader context
