Training LLMs based on tweets is one thing, training it on chats with users is something completely different. I don't think this actually happens. The model would degrade extremely fast.