That's the thing, they use a combination of their training and past conversational context. Maybe this person pasted information from online resources directly into the conversation, since most models aren't allowed to access arbitrary links for security reasons. And also, a lot of this information isn't specific to Hive per se as much of those concepts are common in stake based economies. AIs are really good at guessing what they think is right, even when it isn't. You get good at recognizing it when you (have to) interact with lots of AI output online.