We are truly dealing with immense sizes here, see [0]. I am not saying it is representative of all things, but that large corpora are sufficiently representative of words and phrases in a language (and by extension, the words and phrases assoiciated with topics and ideas). In written form you may miss some slang, pidgins and dialects, but those would be underrepresented in other media as well.

[0]: https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-...