How I use LLMs for language learning
Note: These notes reflect my experiences learning Spanish. I don't know the extent to which foundation models have been trained on corpora of text in other languages. My assumption is that there is a heavy bias toward English (at least in the case of Google/Anthropic/OpenAI models) and that the representation of each other language correlates with their respective prominence, but I could be way off here. I will research this topic later.
Useful Strategies
Below are a few primary ways in which I've used LLMs to augment my language learning. For context, I began learning Spanish at an early A1 level in December of 2024 and reached a mid-to-late B2 level by December of 2025. Over the course of that year I tried many types of apps, resources, books, and among these LLMs probably represent about 5% of my learning system. I outline the other 95% elsewhere.
LLM as Feedback Generator for Your Writing
This is one of my favorite use cases for LLMs in the language-learning domain. Here's a recent prompt I gave ChatGPT and the response:
Prompt
Give me feedback on this Spanish writing (grammar, register, style, etc.)
El valor de ser generalista Una persona que se considera sí misma un pensador tiene que seleccionar un punto en la escala de ser especialista a ser generalista. Seguramente hay buenos motivos por los dos lados. En el caso de ser especialista, lo más que profundizas en una temática, lo más te distingues tú mismo con respecto de lo demás. Pero podrías decir lo mismo en el caso de ser generalista, si consideras la raredad de un conjunto de competencias básicas.
Response
“se considera sí misma” → extra reflexive
In Spanish, considerarse already carries the reflexive meaning. Adding “sí misma” is usually redundant unless you want strong emphasis or contrast.
More natural: “una persona que se considera pensadora”
Acceptable with emphasis: “una persona que se considera a sí misma pensadora”
“pensador” vs “pensadora”
Because “persona” is grammatically feminine, Spanish often pulls the adjective into the feminine form in abstract or essay-style writing.
More natural: “una persona que se considera pensadora”
“tiene que seleccionar un punto”
Grammatically correct, but sounds literal and slightly stiff. In abstract writing, Spanish prefers verbs like situarse, ubicarse, or optar por.
More natural: “tiene que situarse en un punto” or “debe ubicarse en algún punto”
“la escala de ser especialista a ser generalista”
Spanish normally expresses this idea using “entre X y Y” rather than “de X a Y” when talking about a conceptual spectrum.
More natural: “la escala entre ser especialista y ser generalista”
“por los dos lados”
Understandable, but slightly unnatural in formal prose. Spanish prefers expressions like “en ambos casos” or “por ambas partes”.
More natural: “en ambos casos”
“lo más que profundizas…, lo más te distingues”
This structure is a direct calque from English (“the more…, the more…”). Spanish uses “cuanto más…, más…”.
Correct structure: “Cuanto más profundizas en una temática, más te distingues…”
“lo demás”
Here you are referring to other people, so the plural form is required.
Correct form: “los demás”
“con respecto de”
Grammatically correct but heavy and formal. “Con respecto a” or “frente a” is more idiomatic in essays.
More natural: “con respecto a” or “frente a”
“raredad”
Understandable, but stylistically weak and uncommon in this context. Spanish prefers “rareza”, “carácter poco común”, or “singularidad”.
More natural: “la rareza de un conjunto de competencias” or “el carácter poco común de un conjunto de competencias”
This is probably far better feedback than you're likely to receive from a human, particularly a human with lots of other writing to critique. LLMs do not grow tired of telling you how awful your Spanish is (unless you configure it to take it easy on you).
The best part of this is that the benefits compound. You could have prompts like these in a ChatGPT project that is preconfigured to, upon generating feedback, provide a CSV for you to import into Anki (people say it's crucial for efficient learning to create your own cards...I think the critical point is having sufficiently engaged with the content that ends up on your cards, and this satisfies that requirement).
Note: I've hooked up Claude Desktop to the Anki MCP, skipping the import step entirely.
LLM as Tireless Answerer of Very Specific Questions
Sometimes you have such question so specific that it is difficult to quickly find Google search results that answer it. More often, you don't know what the right question or search terms are that properly articulate what you're trying to figure out. Here the LLM shines, with similar spaced repetition system benefits detailed above if you set things up that way.
I use this often when reading Spanish content, like a news article from El País. I might encounter a certain grammar structure that I'm unfamiliar with, and ask: "Hey Claude, what's the deal with this Spanish phrase? Why is this tense used here? What's the semantic register and what would a less formal register use?" This really gets to the central point of the power of LLMs for learning in general. You can still be doing rigorous learning, but be doing it at a much faster pace.
LLM as Speaking Partner
I've been pleasantly surprised by the ability of ChatGPT Voice mode to maintain an arbitrarily-long conversation in Spanish, doing so in various voice modes that don't sound terribly like robots. In terms of enjoyment I would place it somewhere between speaking to oneself on the low end and conversing with real people on italki (my preferred way to find native speakers) on the high end. What has kept me from using LLMs in this way on a regular basis is these primary issues:
- I am constantly distracted by the thoughts of the environmental impact of generating obscene quantities of computationally-expensive inference requests that at best yield me a few new Anki cards and a tiny margin of improvement in my speaking fluency.
- I am constantly distracted by the fact that I am speaking to a complex arrangement of matrix multiplications.
Overall I believe this method has a promising future. I'm also quite certain that all the language learning apps are turning themselves into LLM voice API wrappers.
LLM as Study Plan Generator
This was more of a one-time exploratory use, where I wanted to try out the deep research capabilities by giving the prompt a bit of background context on my current Spanish level in the different modes of usage + instructions to research language learning + instructions to get the latest available tooling and resources. This actually proved useful and helped me escape some of the plateaus commonly associated with learners at B1-B2 (DELE Levels).
Some of the tips it gave me:
- Move from learning single vocabulary units to learning families of words, phrases, switching between linguistic registers, etc.
- Start shadowing to improve speaking style, pronunciation, flow
- A good framework for how to divide my practice time (leading me to put a lot more of the time into consuming a large volume of comprehensible input instead of spending more time on flash cards, grammar drills, and other more formal learning techniques
General Thoughts
I'm definitely an optimist in modern LLM technology's ability to accelerate language learning.
I'm bullish on LLMs accelerating learning in general when used properly.
I'm bearish on a majority using them properly for learning. Hopefully the various wrapper apps and model harnesses will take care of that.