While large language model (LLM) foundation models can handle language and imagery at unprecedented breadth, recent research has underlined that they may exhibit many of the same cognitive biases that humans possess. Therefore, widespread interaction with biased LLMs could reinforce harmful stereotypes that we should aim to eradicate, as well as increase polarized viewpoints regarding divisive topics. This risk is augmented by the fact that most present-day LLM users come from various backgrounds and are not trained to be critical consumers of LLM-produced content, in contrast to earlier human–automation interaction. Therefore, to facilitate responsible human–AI interaction that mitigates the risk of exacerbating harmful stereotypes and increasing polarization, it is ever more critical to understand the nature of the biases possessed by LLMs from multiple perspectives. We adopt a dual perspective approach that examines the representation biases possessed by LLMs from both an inward and outward perspective. The inward perspective investigates the underlying structural patterns that characterize representation biases within the cognitive architecture of LLMs, while the outward perspective investigates ways in which representation biases manifest in social settings among conversational agents with a theory of mind, including their impact on opinion evolution. The investigations from both the inward and outward perspectives apply a complex network science approach, diving into the emergent nature of cognitive phenomena that occur both “within the mind” and “between minds.”