Except that the errors made by an LLM might be harder to spot then converter errors that typically are very blatant, and don't usually alter text (perhaps just

Except that the errors made by an LLM might be harder to spot then converter errors that typically are very blatant, and don't usually alter text (perhaps just drop parts of it). Also, a bug in a converter is conceptually much easier to fix than to re-train your LLM. I am not sure that AI in it's current state is useful when "high fidelity" is required.