It's also not a bug because this is how LLMs are meant to work—to produce some output based on their corpus of information, with a dial for how much randomness to throw in.

They set the temperature too high and it made a nonsensical connection that broke the illusion of a conversation.

But the connections are always nonsensical, even when it sounds like on topic English.