Pluralistic and Collective Alignment, End-To-End

Speaker: Prof. Mitchell Gordon

Time: 12:00–1:30 PM CDT, Aug 10, 2026

Abstract

How should we align AI systems when people genuinely disagree about how those systems should behave? In this talk, I will discuss challenges that arise at several different parts of an end-to-end collective alignment process, from theory and definitions, to gathering and acting on people’s input, to applying behavioral guidance in agentic systems. I will first examine what it means for a model to be pluralistically aligned and how different forms of pluralism can be defined and evaluated. I will then discuss how Collective Alignment contributed to OpenAI's Model Spec, including how we gathered human preferences about model behavior and translated their feedback into behavioral guidance. Finally, I will discuss ongoing work on designing systems for when agents must interpret a model spec or constitution in subjective, open-ended problems.