Randomness is difficult. I wouldn't expect any LLM to be able to reliably produce random anything, except in the cases where they have access to tools (ChatGPT Code Interpreter could use Python's random.random() for example).