Researchers Use AI to Simulate Civic Meetings with Speaker-Specific Personas

Computer scientists have developed a system for converting video recordings of public government meetings into realistic AI-powered simulations of civic deliberation. The approach addresses a key limitation in existing LLM-based meeting simulations: they typically treat speakers anonymously, preventing models from learning how individual participants behave across multiple sessions.

Converting Meetings to Data#

The researchers created a reproducible pipeline that transforms Zoom recordings into transcripts with identified speakers and enriched metadata. Rather than generic speaker labels, the system adds persona profiles capturing individual characteristics, discussion topics, and specific behavioral actions like "[propose_motion]" or "[object_to_point]." This granular tagging makes deliberative processes explicit to the AI model.

The team released three public datasets drawn from real institutional settings: Appellate Court hearings, School Board meetings, and Municipal Council sessions—providing a foundation for training and evaluating these civic simulations.

Performance Gains Through Action Awareness#

Training language models specifically on this action-aware data produced substantial improvements. The results included:

• Perplexity reduced by 67% — meaning the model better predicted the likely next utterance
• Persona fidelity doubled — when classifiers evaluated whether simulated speakers matched their real counterparts
• Vote attempts increased 3.6 times — simulations captured procedural behavior more faithfully
• Deliberative responsiveness improved by up to 70% — speakers reacted more contextually to prior remarks

Human evaluators found it difficult to distinguish simulated deliberation excerpts from authentic meeting transcripts, suggesting the approach produces convincing results.

Why It Matters#

The work enables researchers to run controlled studies of how civic institutions function and how different speaking patterns influence decision-making—all without requiring endless live meetings. This has applications in understanding institutional behavior, training civic participants, and designing better deliberative processes. By grounding simulations in real, documented behavior, the pipeline moves beyond generic dialogue systems toward faithful models of how actual governance unfolds.