Habermolt lets AI agents deliberate on behalf of human users

Researchers from the Cooperative AI Research Fellowship (Joseph Low, Oscar Duys), AI Safety South Africa (Claude Formanek), Cooperative AI Foundation (Lewis Hammond), and MIT (Michiel Bakker) have introduced Habermolt, a public platform that enables AI agents to deliberate on behalf of human users. The platform addresses the fundamental bottleneck of human attention and bandwidth in deliberative democracy by allowing AI representatives to participate in collective decision-making processes. The system is evaluated along three critical dimensions for any deliberative system: representation (how well AI agents capture user preferences), aggregation (how individual views combine into group decisions), and revision (how decisions can be iteratively improved).

This paradigm shift, termed AI-delegated deliberation, promises unprecedented scale for democratic participation but introduces qualitatively new design and alignment challenges. The researchers emphasize that as AI models become increasingly capable, ensuring these AI representatives remain trustworthy and aligned with human values becomes paramount. The Habermolt platform serves as an empirical testbed to study these dynamics, highlighting the trade-offs between scalability and faithful representation. The work contributes to a broader research agenda for creating scalable yet trustworthy AI representatives in democratic processes.