AInteraction: Social Construction of Knowledge by Humans and Machines

The project investigates how humans and large language models (LLMs) construct knowledge dialogically, and how insights from one can inform the other. Social theory has long emphasised interaction, indication, and adjustment of perspectives as mechanisms of meaning-making, yet empirical work typically measures only pre/post attitudes and neglects the fine-grained conversational processes through which understandings are formed, coordinated, and contested. At the same time, current LLMs mostly learn in a static, one-shot training regime and are subsequently used as solitary “oracles,” overlooking the potential of dialogic interaction for model learning.

To address these gaps, the project combines social-science experiments, computational modelling, and agent-based simulations within a unified framework of generalised socio-semantic hypergraphs. First, randomised controlled trials are conducted in which diverse participants engage in dyadic, text-based discussions about a novel, complex topic (e.g. AI), while conversational sequences, interaction patterns, and evolving individual understandings are tracked. Second, participants’ knowledge and its transformations are represented as nested semantic hypergraphs embedded in social interaction, and Relational Hyper-Event Models are extended to test theoretically derived mechanisms of dialogic knowledge construction operationalised in the generalised socio-semantic hypergraph framework. Third, multi-agent Dialogic AI experiments are conducted in which LLMs interact with each other to learn new concepts, enabling direct comparison between human and machine alignment, coordination, and conceptual structure change within the same socio-semantic framework. Finally, these findings are integrated into agent-based models that simulate discussions and explore conditions for inclusive, constructive dialogue.

The project is expected to deliver (1) a nuanced, empirically grounded account of dialogic knowledge construction among humans, (2) a novel paradigm for Dialogic AI, and (3) methodological advances in generalised socio-semantic hypergraph modelling and simulation for the study of complex deliberative processes.