Been thinking recently about the goings-on with Duolingo & AI, and I do want to throw my two cents in, actually.

There are ways in which computers can help us with languages, certainly. They absolutely should not be the be-all and end-all, and particularly for any sort of professional work I am wholly in favour of actually employing qualified translators & interpreters, because there's a lot of important nuances to language and translation (e.g. context, ambiguity, implied meaning, authorial intent, target audience, etc.) that a computer generally does not handle well. But translation software has made casual communication across language barriers accessible to the average person, and that's something that is incredibly valuable to have, I think.

Duolingo, however, is not translation software. Duolingo's purpose is to teach languages. And I do not think you can be effectively taught a language by something that does not understand it itself; or rather, that does not go about comprehending and producing language in the way that a person would.

Whilst a language model might be able to use probability & statistics to put together an output that is grammatically correct and contextually appropriate, it lacks an understanding of why, beyond "statistically speaking, this element is likely to come next". There is no communicative intent behind the output it produces; its only goal is mimicking the input it has been trained on. And whilst that can produce some very natural-seeming output, it does not capture the reality of language use in the real world.

Because language is not just a set of probabilities - there are an infinite array of other factors at play. And we do not set out only to mimic what we have seen or heard; we intend to communicate with the wider world, using the tools we have available, and that might require deviating from the realm of the expected.

Often, the most probable output is not actually what you're likely to encounter in practice. Ungrammatical or contextually inappropriate utterances can be used for dramatic or humorous effect, for example; or nonstandard linguistic styles may be used to indicate one's relationship to the community those styles are associated with. Social and cultural context might be needed to understand a reference, or a linguistic feature might seem extraneous or confusing when removed from its original environment.

To put it briefly, even without knowing exactly how the human brain processes and produces language (which we certainly don't), it's readily apparent that boiling it down to a statistical model is entirely misrepresentative of the reality of language.

And thus a statistical model is unlikely to be able to comprehend and assist with many of the difficulties of learning a language.

A statistical model might identify that a learner misuses some vocabulary more often than others; what it may not notice is that the vocabulary in question are similar in form, or in their meaning in translation. It might register that you consistently struggle with a particular grammar form; but not identify that the root cause of the struggle is that a comparable grammatical structure in your native language is either radically different or nonexistent. It might note that you have trouble recalling a common saying, but not that you lack the cultural background needed to understand why it has that meaning. And so it can identify points of weakness; but it is incapable of addressing them effectively, because it does not understand how people think.

This is all without considering the consequences of only having a singular source of very formal, very rigid input to learn from, unable to account for linguistic variation due to social factors. Without considering the errors still apparent in the output of most language models, and the biases they are prone to reproducing. Without considering the source of their data, and the ethical considerations regarding where and how such a substantial sample was collected.

I understand that Duolingo wants to introduce more interactivity and adaptability to their courses (and, I suspect, to improve their bottom line). But I genuinely think that going about it in this way is more likely to hinder than to help, and wrongfully prioritises the convenience of AI over the quality and expertise that their existing translators and course designers bring.