I would be interested in an LLM with a bicameral mind: one stream generates text and another one judges the text generated by the first and chooses whether or not to censor it and so forth.

What would such an LLM's threshold for censorable text be?

Well, right now the LLMs are pretty damn trigger happy about censorship after RLHF. Ideally, they would be slightly less prone to do things like tell me they can't help me sort my own data because it has a copyright.