Testing LLMs as Discursive Agents in Deliberative Processes

Deliberative processes face a persistent tension between scale and inclusion. While small groups enable participants to listen to one another and engage in reasoned exchange, they rarely capture the full diversity of perspectives present in society. Discursive representation offers a response to this challenge by shifting attention from who participates to how arguments, values, and modes of reasoning are represented within a deliberative setting. Building on this approach, we examine whether large language models (LLMs) can function as deliberative discursive agents. Using data from real citizens’ assemblies, we model LLMs on empirically observed human discursive profiles and assess their ability to reproduce characteristic patterns of reasoning and argumentation. This allows us to test whether LLMs can simulate perspectives that are underrepresented or absent in small-scale deliberative forums, and to evaluate the epistemic and normative limits of such simulations.