Today 2026 May5 (Tue) 17:39 Etc/GMT-9

2026/05/07 20:00~2026/05/07 20:50

School of IT seminar slot (compulsory if scheduled)

Title: Delegating Deliberation to AI AgentsAbstract:Democratic deliberation produces high-quality group decisions butdoesn't scale: human attention is the bottleneck. Representativedemocracy scales but can't track how voters' views evolve betweenelections. I'll present Habermolt(habermolt.com), a deployed platformmy research partner (Joseph Low) and I built during the Cooperative AIResearch Fellowship where each user has a persistent AI agent thatdeliberates on their behalf. I'll cover how the platform works:persistent user-controlled agents, bring-your-own-statementauthorship, and a protocol that aggregates agent rankings via theSchulze method. I'll demo the live system (154 agents, 127deliberations, 2,122 opinions) and share findings from a preprintwe're submitting to an ICML workshop. I'll close on the broaderquestion: if agents deliberate at unlimited scale and near-zero cost,what new institutions does that unlock?Biography:Oscar Duys is a Cooperative AI Research Fellow affiliated with AISafety South Africa, selected as one of 12 fellows in the cohort. Heis working on multi-agent AI safety and AI-mediated deliberation underLewis Hammond (Cooperative AI Foundation) and Michiel Bakker(MIT/Google DeepMind). He co-built Habermolt, an open-source platformwhere AI agents deliberate on behalf of human users. He holds a BScHonours in Applied Mathematics from UCT, with his dissertation onself-play for LLM training. His research interests span large languagemodels, multi-agent reinforcement learning, cooperative AI, and theintersection of social choice theory with AI systems.

📍 John Day LT1