Deliberation Within AI: A Reproducible Pipeline for Multi-Agent Deliberation Simulation

Recent advances in generative AI have renewed interest in whether large language models can simulate aspects of public deliberation. Current deliberation simulations remain methodologically fragile: fluent conversational outputs are often treated as evidence of deliberative quality despite limited reproducibility, weak safeguards for viewpoint diversity, and limited systematic evaluation. This article argues that AI-deliberation simulations can function as credible analytical instruments only if they incorporate infrastructures for reproducibility, plurality preservation, and deliberative quality assessment. The article develops a stage-structured method for deliberative simulation combining persona validation, plurality auditing, and discourse-quality assessment within a reproducible pipeline. The framework is presented as a methodological proof-of-concept and evaluated through a simulation study on the June 2023 Swiss federal referendum on the Climate and Innovation Act together with a batch replication study comparing direct and structured deliberation architectures. The findings show that, as an integrated theory-driven architecture, the structured pipeline generates more stable deliberative dynamics than unconstrained interaction, including clearer plurality preservation and more coherent opinion updating. At the same time, the simulations remain below human minipublics in terms of contextual grounding and justificatory depth. The article positions AI-deliberation simulations as methodological infrastructures for cumulative and empirically testable research on AI-supported deliberation in digital governance.