That's their problem. If they are using an LLM and cannot verify the output they shouldn't be using an LLM