If I'm going to pick up speech patterns at all, I would really rather pick them up from a native speaker of the language, since at the very least I'll make the sorts of mistakes that a human might make. I want to sound like human, not like a language model. Language models sound like the average of several humans at best, and a strange program trying to imitate human speech at worst.
Once I'm fluent in a language, enough to recognize when the language model itself is probably making a mistake, then I might become comfortable using it. But not as my first introduction to the nuances of the language, when I'm still building my own internal representation up from scratch. After all, my goal is to converse with other human speakers. Shouldn't that be my personal training corpus? I'm a neural network too, and I don't want to feed myself bad data.
This isn't really true. Open ai use heavy rlhf to make LLMs sound like that by default but they can sound like whatever. If a native speaker says it's fine then it's fine lol. You can still choose not to use it as you can choose not to do anything but then it's an irrational fear more than real concern.
This is also essentially a substitute for graded readers for language learners
So, in conclusion, I'm curious to learn whether this is a solvable problem or if LLMs are inherently not the right tool to use.