I don't understand what theoretical basis can even exist for "I don't know" from an LLM, just based on how they work.

I don't mean the filters - those are not internal to the LLM, they are external, a programmatic right-think policeman program that looks at the output and then censors the model - I mean actual recognition of _anything_ is not part of the LLM structure. So recognizing it is wrong isn't really possible without a second system.