What are the best methods to measure the faithfulness of simulations?

Examines design choices behind computational frameworks for finding common ground across collective preferences, situating AI-assisted preference elicitation within the historical context of opinion polls. Emphasizes that preferences are shaped by context and seldom objectively captured, explorin...

Presents an approach for building AI “general” agents that can predict human behavior in novel settings without requiring extensive setting-specific training data. The agents use theory-grounded natural language instructions combined with existing empirical data and knowledge from language model ...

Several have explored versions of this idea, under the keywords ‘AI-as-representative’ (Collins, 2023), ‘voting avatars’ (Grandi, 2018), and ‘virtual democracy’ (Kahng et al., 2019), ‘plurals’ (Ashkinaze et al., 2024) and ‘simulated deliberative democracy’ (Leike, 2023).

Proposes autonomous AI agents (“voting avatars”) that debate and vote on behalf of citizens, addressing the cognitive burden of direct democracy in complex societies through compact preference representation. Umberto Grandi argues these systems would leverage AI research in multiagent systems and...

Examines which voting rules are robust to prediction errors in “virtual democracy” systems that learn individual preferences and aggregate predicted votes. The research proves that the classic Borda count rule is robust to prediction errors, whereas any voting rule belonging to the wide family of...

Proposes using imitation learning on large language models trained on recorded deliberative democracy sessions to answer value-laden questions at scale. The approach involves recording human ‘mini-publics’ deliberating complex value questions with AI assistance, then training models to simulate t...