Measuring and improving AI's societal impacts

Millions of people now turn to large language models for guidance on personal judgments, preferences, and decisions, making how these systems respond to such queries critical to AI safety and societal impact. This talk draws on two case studies to examine the interplay between model behavior and user behavior in these domains: one on how AI sycophancy shapes users' interactions over time, and one on how users may come to (over)rely on AI in subjective and personal settings. On the model side, we examine how alignment training and design choices drive these behaviors. On the user side, we investigate how sustained interaction influences the ways people seek guidance and make decisions. Together, this work points toward the importance of new methods, evaluations, and design paradigms that account for users' long-term autonomy and wellbeing, and toward the broader question of what it means to build AI systems for human flourishing.

Lujain Ibrahim is a DPhil candidate in social data science at the University of Oxford and a research scientist at Google DeepMind. Previously, she was a visiting researcher at the Stanford AI Lab, a Schwarzman Scholar, and a fellow at the Centre for Governance of AI. Her research focuses on understanding and evaluating how language models shape human judgment, beliefs, and relationships, and what this means for the safety of systems deployed at scale. Her work has been published in Nature, ICLR, NeurIPS, AIES, and FAccT, and covered by the BBC, The Telegraph, Le Monde, NBC, Wired, NPR, Mashable, and more.

---

Register now