Large Language Models as Misleading Assistants in Conversation