How to Learn a Language with AI (2026)

The Science AI Inherited: Krashen, i+1, and the Affective Filter

Before the chatbots, there was a hypothesis. In the early 1980s the linguist Stephen Krashen argued that we don't really learn a language by memorizing rules. We acquire it, the same quiet way children do, by understanding messages.

His Comprehensible Input hypothesis, laid out fully in his 1985 book The Input Hypothesis, makes a specific claim: acquisition happens when we receive large quantities of input that we can mostly understand, pitched just one step beyond our current ability. Krashen called that level "i+1," where i is what you know now and +1 is the slightly harder material your brain stretches to comprehend.

Two things follow from this, and both matter for how you should use AI.

First, the bottleneck was never information. It was comprehensible information at the right level. A native news article is not i+1 for a beginner; it's noise. A children's book might be i minus 3 for an intermediate learner; it's boring. For decades, finding a steady supply of material at your exact edge meant a tutor, a teacher, or a lot of luck.

Second, Krashen's Affective Filter hypothesis says that stress, anxiety, and low confidence physically block acquisition. Input you're too embarrassed or nervous to engage with doesn't stick. Anyone who has frozen mid-sentence in a high school French class knows the feeling.

Hold those two ideas. An on-demand source of level-matched, low-anxiety input is exactly the thing language learners have wanted for forty years. That's the gap 2026 AI walks into.

Input Is Not Enough: Swain and the Output Hypothesis

Krashen's input-first view was influential, and it was also contested almost immediately. The same year The Input Hypothesis appeared, the applied linguist Merrill Swain published her Output Hypothesis (1985), based on a striking finding from French immersion programs in Canada.

Those students got years of rich, comprehensible input. Their listening and reading were excellent. Yet their speaking and writing stayed stubbornly off. Swain's argument: you also have to produce language. The act of forming a sentence forces you to notice the gaps in what you actually know, to test hypotheses about grammar, and to move from vague recognition to precise control.

So the honest synthesis, the one most of the field has settled into, looks like this:

• Input builds comprehension and feeds your subconscious model of the language. Mostly reading and listening.
• Output builds production, fluency, and the ability to retrieve words under pressure. Speaking and writing.
• Feedback corrects errors before they fossilize into permanent habits.

For most of history, getting all three was expensive. Input you could scrounge. Output and feedback required a patient human who would talk to you and gently fix your mistakes for hours. That person was the scarce resource. Keep that in mind as we look at what AI changes.

What AI Actually Does Well in 2026

Strip away the marketing and AI's contribution to language learning comes down to a few concrete capabilities, each mapping onto the science above.

On-demand comprehensible input. Ask a model to retell a news story "for a beginner, using only the present tense and the 500 most common words," and you get instant i+1 input on a topic you care about. The same tool can take a real article that's slightly too hard and simplify it one notch, which is the literal definition of pulling something down to i+1.

A conversation partner with infinite patience. Voice mode lets you talk, stumble, repeat, and ask "how would a native say that?" without a human's social clock running. This is where the Affective Filter point pays off: lower anxiety means more output, and more output is what Swain said you were missing.

Instant, targeted feedback. Paste what you wrote and ask for corrections plus a one-line explanation of each. That tightens the output-feedback loop from "next week's lesson" to "right now."

Translation and explanation on hover. Reading foreign text while glossing unknown words in place keeps you inside the input instead of bouncing to a dictionary and losing the thread.

Spaced repetition that actually schedules itself. The vocabulary you mine has to be reviewed, and the research-backed way to do that is FSRS (the Free Spaced Repetition Scheduler), the modern algorithm now built into Anki. It predicts when you're about to forget a card and shows it then, which is far more efficient than rereading.

Here's how the pieces line up against the underlying theory:

The throughline: AI is not a new theory of learning. It's a delivery mechanism for an old, well-supported one.

The Weekly Stack: Input, Output, Review

Capability lists don't teach anyone a language. A routine does. Here's a concrete weekly stack that respects the input-output-review structure and fits in roughly an hour a day. Scale the minutes to your life; the proportions matter more than the totals.

Daily input (20-30 min). Read or listen to something slightly above your level on a topic you'd consume in your native language anyway. Cooking, football, a TV recap, whatever keeps you engaged. If it's too hard, ask AI to simplify it one notch rather than abandoning it. Adults read nonfiction at roughly 238 words per minute and fiction near 260 in their native tongue (Brysbaert's 2019 meta-analysis); in a new language you'll start far slower, and that's expected. Volume over speed early on.

Output every other day (15-20 min). Talk to an AI voice partner or write a few paragraphs. Pick a real scenario: ordering at a restaurant, describing your weekend, arguing a mild opinion. Push slightly past your comfort so you hit the gaps Swain cared about. Then ask for corrections.

Review daily (10 min). Run your spaced-repetition deck. Cards should come from words you actually met in your input, not a generic top-1000 list. Context-rooted vocabulary sticks better because you have a memory hook for it. This is straight active recall plus spacing, the two highest-leverage study techniques there are.

Weekly human checkpoint (optional but valuable). A tutor session, a language exchange, or a class. This is where you catch the things AI quietly gets wrong, and where real cultural and social feedback lives.

A useful way to see the same week is by which skill each block trains:

If you want a deeper comparison of how AI study features map onto learning modes, see our breakdown of AI study modes compared.

Turning the Web and YouTube into Comprehensible Input

The hardest part of Krashen's model in practice is the supply problem. Where does a steady stream of interesting, level-appropriate input come from once you've exhausted the textbook dialogues? The answer in 2026 is the open web and video, made comprehensible.

Start with reading. The foreign-language internet is the largest free input library ever assembled. Recipe blogs, sports forums, fan wikis, opinion columns. The trick is staying inside the text. Use Glasp's web highlighter to highlight the words and phrases you don't know as you read, so you're marking real gaps in context instead of copying isolated words into a dictionary. When a phrase is genuinely confusing, Glasp's AI chat can explain why it's structured the way it is, right where you found it.

Video is where most learners stall, because native-speed speech is brutal at first. This is exactly where transcripts rescue you. Run a foreign-language video through YouTube Summary to pull the transcript and key takeaways, then read along while you listen. Suddenly the firehose of speech becomes comprehensible input you can pause, reread, and mine for vocabulary. A travel vlog in Spanish or a cooking channel in Japanese turns into a structured lesson without anyone writing a curriculum.

This pairs naturally with how we've written about learning from YouTube more broadly: the video is the raw material, and the transcript plus your highlights are what convert watching into actual acquisition.

The point of all this is repetition with variety. Krashen's i+1 isn't a single magic sentence; it's a flood of slightly-challenging input across many topics, so the same grammar and vocabulary recur in different clothes until your brain stops noticing them as foreign.

Building a Personal Input Library You Can Review

Input you understood once and never saw again is mostly wasted. The learners who actually progress treat their input as an asset to revisit, not a stream to consume and forget. This is the bridge between Krashen's input and the review half of your weekly stack.

Every highlight you make while reading foreign articles becomes part of a personal comprehensible-input library: a searchable, growing collection of real language at your level, on topics you chose. Over months that library becomes a far better record of your learning than any pre-made deck, because every entry has a context you remember.

Books belong here too. If you read foreign-language books on a Kindle, your Kindle highlights sync into the same library, so a novel you're working through in Italian feeds the same review pipeline as the articles and videos.

Then close the loop. Export your highlights and turn them into flashcards for your spaced-repetition system. A highlight already comes with its sentence, which means your cards have built-in context instead of bare word pairs. That context is what makes spaced repetition for readers so effective: you're not memorizing "manzana = apple," you're re-meeting a sentence you once understood, which is far closer to how acquisition actually works.

The workflow in one line: read or watch, highlight what's at your edge, export to flashcards, review on an FSRS schedule, repeat. Input becomes retained memory instead of a pleasant afternoon you forget by Thursday.

Where AI Falls Short (Read This Before You Trust It)

Honesty here protects your time. AI is a powerful delivery mechanism, but it has specific, well-documented failure modes, and the people selling it rarely lead with them.

It hallucinates grammar. Ask a model to explain a rule and it will answer confidently every time, including when it's wrong. It may invent exceptions, misstate gender agreement, or rationalize a "natural" phrasing that no native would use. For high-stakes grammar, verify against a real reference or a teacher. A fluent-sounding wrong answer is more dangerous than no answer.

Pronunciation feedback is forgiving. Text feedback is strong; the ear is weaker. Models often accept pronunciation a native speaker would flag, which can quietly lock in an accent that's harder to fix later. Real human listening, or at least a phonetics-focused course, fills this gap.

Cultural and pragmatic nuance gets flattened. Knowing when a phrase is rude, intimate, regional, or sarcastic is the hard, human part of fluency, and it's exactly what AI smooths over. It will happily teach you a textbook-correct sentence that lands wrong in a real conversation.

It can become a comfortable substitute for the scary part. Talking to a machine is safe, which is great for lowering Krashen's Affective Filter and terrible if it replaces ever talking to a person. The friction of a real conversation is part of the training.

This is why the most credible sources frame AI as most effective when paired with high-quality structured courses, cultural immersion, and real human interaction, not as a replacement for them. Use AI to multiply your reps and lower the barrier to starting. Use humans and structured courses to catch what AI can't see. The combination beats either alone.

Frequently Asked Questions

Can I become fluent using only AI?

Probably not to a high level. AI is excellent for input volume, low-stakes output practice, and review, which is most of the grind. But it under-corrects pronunciation, flattens cultural nuance, and occasionally hallucinates grammar. Treat it as the engine for daily reps and pair it with real human conversation and a structured course for the parts it can't judge.

Is comprehensible input really better than studying grammar rules?

They're not enemies. Krashen (1985) argued input does the heavy lifting for acquisition, and decades of evidence support input volume mattering enormously. But explicit grammar study and output (Swain 1985) speed up accuracy and help you notice gaps. The strongest routines use input as the foundation and add targeted grammar and speaking on top.

What spaced-repetition tool should I use?

Anki running the FSRS algorithm is the research-backed default in 2026. FSRS predicts when you're about to forget a card and schedules the review for that moment, which is far more efficient than fixed intervals or rereading. Seed your deck with vocabulary you actually met in your reading and watching, not a generic frequency list, so every card has context.

How do I make native YouTube videos understandable as a beginner?

Use the transcript. Run the video through YouTube Summary to get the full transcript and key takeaways, then read along while listening and pause freely. The transcript converts native-speed speech, which is otherwise overwhelming, into comprehensible input you can reread and mine for new words.

How much time per day do I actually need?

An hour a day, split roughly into 20-30 minutes of input, 15-20 of output a few times a week, and 10 minutes of review daily, will move most people steadily. Consistency beats marathon sessions. Two focused hours on Sunday lose to twenty minutes every day, because spacing and frequent retrieval are what build durable memory.

Conclusion

The science was settled long before the tools arrived. Krashen told us in 1985 that we acquire languages from abundant, level-appropriate, low-anxiety input. Swain reminded us, the same year, that we also have to produce language to truly own it. What was missing was never the theory. It was a cheap, patient, always-available way to deliver input, prompt output, and correct mistakes. That's the part AI finally solved.

Build the loop and let it run: comprehensible input every day, output a few times a week, review every day, and a human in the mix when you can manage it. Then make your input compound instead of evaporate.

Highlight foreign articles and videos with Glasp's web highlighter to build a personal comprehensible-input library, make tough videos understandable with YouTube Summary, pull in foreign books through your Kindle highlights, ask about confusing phrases with Glasp's AI chat, and export your highlights into spaced-repetition flashcards so the language you understood once becomes the language you keep.