In this case we have a dissonance between an intellectual understanding that these models are in some sense just predicting where the conversation should go, and our false insight that they are persons. The illusion of emotion and engagement is strong and getting stronger.
It starts with folks like Blake Lemoine, but ultimately we're all vulnerable. Just as we're vulnerable to increasingly tailored and sophisticated forms of media manipulation. An understanding of the mechanics of media narratives, broadcasting etc is not a strong enough defence.
The 'insight fallacy' is the idea that understanding the nature of a problem will be enough to solve it. In this case we have a whole bunch of people studying alignment, and attempting to understand the kind of interior models LLMs form. While simultaneously improving them and commodifying them.
It goes beyond a perverse profit incentive. As a species we have a deep desire for non-human companionship - be it in the form of spirits, aliens, A.I. or even pets. We seem determined to create it, and wildly emotionally vulnerable to it in even the most primitive forms. From ELIZA on down, when people have engaged with conversational agents, they have overridden our better judgement. False insight and false friendship, is a major social vulnerability.
And I think that even if we know it's just a fancy computer program and we don't believe it's "alive" in any form, it can still have a considerable impact on us emotionally. Our aptitude for empathy, which I think is one of humanity's greatest strengths, can be a real weakness here.
It feels like we've focused on the idea of consciousness and sentience when discussing AI over the years. But I don't think it even needs to get anywhere near there. As soon as a chatbot acts sufficiently real, it's real.
In no time we'll have Monroebots: https://www.youtube.com/watch?v=YuQqlhqAUuQ