Duolingo is a multimodal learning tool. There's some translation but there's also fill in the blank, describe from prompt, oral story interpretation, spoken descriptions, and even AI chat bot interactions in recent version.
There's AI slop (or hastily human generated slop, hard to tell) in Duolingo so I won't advocate for its quality, but I've been trying to use several different flagship models for language learning (with a native speaker on speeddial to fact check things) and they get stuff wrong a lot. LLMs are absolutely not ready to be your sole source for language learning. They seem perfectly competent at communicating in whatever language you want, and are fine at translation, but for example, explaining grammatical concepts of one language in another language they have been surprisingly incompetent at in my experience.
I and my wife used an LLM to translate something she had written, she could have done that herself but she doesn't feel up to a task like that yet (due to the target audience). And I myself am far away from being able to translate that kind of text to my native language.
In general the translation was good, but the wording felt a bit unnatural, and to my surprise it got some basic grammar wrong - specifically, using the wrong grammatical gender for some nouns (sometimes there are valid variants, but not in the cases I'm referring to), and also using pronouns where a native never would - where it's too hard to immediately see what the pronoun refers to. In the end I had to massage the output a lot before it was acceptable, and we spent hours before the output was acceptable (changing the input to try to coerce a better translation, and after that refreshing the translation manually to fix grammar errors, wording, and as mentioned, overuse of pronouns).
The two sentences do not contradict. Using LLMs alone would be bad. However they can be used with other things. Most people are get fluent in a language use several different methods to learn.
It isn't clear if LLMs are good. The formal studies cannot possibly be done so don't bother looking. (a few early studies might be done, but not enough to draw conclusions). And of course LLMs may well change in the future so even if you have a conclusion it may not apply to what we see next year.
I'm learning an admittedly fairly obscure african language, but one with tens of millions of speakers worldwide. LLM can produce intelligible but grammatically-incorrect and unidiomatic output. Is this better or worse than not helping at all? I'd argue worse.
> I'm learning an admittedly fairly obscure african language, but one with tens of millions of speakers worldwide. LLM can produce intelligible but grammatically-incorrect and unidiomatic output.
This isn't a problem with the technology; it's easy to observe that it doesn't happen with better-known languages. Your problem is that you don't have a model for your target language.
> Is this better or worse than not helping at all? I'd argue worse.
My first instincts go that way too. But note that language classes consider it desirable for the students to try to speak with each other in the target language. (And not just where they can be supervised - the more they do it, in any context, the better.)
If the only input you ever get has the grammar incorrect, your grammar will also be incorrect. But you can handle a lot of your input being incorrect without major problems.
> This isn't a problem with the technology; it's easy to observe that it doesn't happen with better-known languages. Your problem is that you don't have a model for your target language.
technically, you're correct. But I don't expect to see much given how resources for ai licenses are calculated.
