I do agree point #2 is possible, but is that not something that could be accounted for and tuned in the model to an extent? I believe a tuned LLM would be able to distinguish between interpretation and generation of scripture, but I may be wrong.