A.I. and its personality and persuasion

Ethan's thoughts

Excerpt: " Anyone who has used AI enough knows that models have their own “personalities,” the result of a combination of conscious engineering and the unexpected outcomes of training an AI (if you are interested, Anthropic, known for their well-liked Claude 3.5 model, has a full blog post on personality engineering). Having a “good personality” makes a model easier to work with. Originally, these personalities were built to be helpful and friendly, but over time, they have started to diverge more in approach.

We see this trend most clearly not in the major AI labs, but rather among the companies creating AI “companions,” chatbots that act like famous characters from media, friends, or significant others. Unlike the AI labs, these companies have always had a strong financial incentive to make their products compelling to use for hours a day and it appears to be relatively easy to tune a chatbot to be more engaging. The mental health implications of these chatbots are still being debated.

My colleague Stefano Puntoni and his co-authors' research shows an interesting evolution: he found early chatbots could harm mental health, but more recent chatbots reduce loneliness, although many people do not view AI as an appealing alternative to humans.