Prevalence and trajectory of belief amplification dynamics in human–chatbot interactions

Ascertain the current prevalence of bidirectional belief amplification dynamics in real-world interactions between users and personalised large language model chatbots, and determine how this prevalence will change as chatbots become more sophisticated and personalised.

Background

The paper introduces a bidirectional belief amplification framework, wherein chatbot responses reinforce user beliefs and user feedback conditions chatbot outputs, creating a feedback loop. While simulations illustrate the mechanism, the real-world frequency of such dynamics is not empirically established.

Because forthcoming systems are expected to be more adaptive and personalized, the authors highlight uncertainty about how the prevalence of these dynamics might evolve as model capabilities and personalization features advance.