Abstract
Citizens’ Assemblies and other deliberative mini-publics are designed to foster democratic deliberation, understood as inclusive and informed processes of mutual justification among equals. AI technologies such as large language models (LLMs) have been suggested as tools to scale up the breadth and depth of Citizens’ Assemblies, i.e. increase the number of participants and the quality of outputs. To investigate whether AI technologies can fulfil these promises, we distinguish six different core functions in democratic deliberation, namely representation, perspective-taking, factual learning, co-creation, integration and decision-making and argue that the potential of AI varies depending on the function. We build on exceptional survey and observational data from a real-world case where an LLM (ChatGPT) was used to support the functions of factual learning, co-creation and integration: the Citizens’ Assembly on Energy, organised in Finland in winter 2025. Based on participant and facilitator surveys as well as observations at different stages of the deliberative process we find, first, that LLMs have some shortcomings when it comes to factual learning in Citizens’ Assemblies, especially in creating questions for expert panels. Second, we find that LLMs can ease the workload of human facilitators and deliberators in the co-creation and integration functions, for example by categorising preliminary facts and recommendations for more in-depth deliberations among small groups. Our findings contribute to the emerging debate on the optimal division of labour in deliberative processes between human deliberators, facilitators and generative AI systems.