Speaker: Michiel Bakkar (Assistant Professor at MIT; Senior Research Scientist at DeepMind)
Abstract: Collective deliberation is crucial for resolving complex societal issues, yet traditional methods are slow, difficult to scale, and often struggle to equitably represent diverse perspectives. In this talk, I will present a project that was published in Science last October in which we use large language models (LLMs) to facilitate consensus-building. Our system iteratively synthesizes participants' personal opinions and critiques to generate statements that capture common ground on contentious topics. Across multiple studies, including a virtual citizens' assembly with a demographically representative UK sample, participants rated AI-mediated deliberations as more informative, clear, and unbiased than human-written alternatives. Notably, discussants often revised their views post-deliberation, converging on a shared perspective. Beyond these results, I will explore how AI-mediated deliberation informs broader efforts in pluralistic AI alignment, ensuring AI systems can navigate and respect diverse human values rather than imposing a singular perspective. I will also discuss how these techniques apply to real-world sociotechnical systems like X Community Notes, where AI helps identify broadly supported insights in a real-world crowdsourced fact-checking system.