Having a LLM also generate your unit tests defeats the purpose of checking whether a LLM generated piece of code is sound.

There will always need a human on the end to check if the LLM didn't hallucinate anything on it, including manually writing the unit tests.