I agree, this seems to start from the assumption that a new unit test should pass, but that isn't really the case. A unit test should assert that the code functions accordingly to it's planned design. The AI writing unit tests can't know it's planned design, so it can't write test cases that prove the code is wrong