A better example would be to use LLMs to generate passwords or secret keys. Then even if it looks random to human, the inherent bias would make it a security disaster.