[nocoder]

My experience with the LLM bot is that they are really keen to appease the user and often over confident about their responses. They are prioritizing user engagement over factual nature of the response, it's as if their reward function includes the time spent on the bot. This leads to the bot swaying too much in either direction when it comes to debatable information. They essentially learn what the user prefers and so tend to reinforce those ideas. In some sense, they are like social media influencers who are too confident in their opinion because they are trying to get you to like them. I see us going further into the echo chambers where on the same topic, the bots will give different information to an users based on what the bot thinks about the user preference.