Making it accessible, understandable, and usable is another matter.
How LLMs sound is not a fundamental limitation of the technology.
The current model's poor writing style and tone are currently a main focus of research and I would expect improvements there soon.
You do not have to train on anything you do not deem up to standard. This supposed poisoning of training data remains a common fantasy.
I see no evidence that this is what the use of LLMs is accomplishing for most users. Rather, they see to get distilled answers without the depth required to fully understand the response. Partially because that's what they like, and that's what the LLMs serve. Deeper understanding is not being given by LLMs. Instead, it's the SEMBLANCE of depth and laypeople don't know the difference. It's effectively eroding comprehension for some cool knowledge dopamine hit.
1) LLM content can at best be as good as the source material it was trained on. That's the upper bound. "Out of distribution" output of LLMs is mostly unusable.
2) LLM-generated content increasingly drowns out original content, online and elsewhere.
This spells "monotonically decreasing content quality" to me.