Abstract

AI-delivered information is now more persuasive than political ads (and most human beings) for controversial sociopolitical questions. Drawing from a wide range of randomized experiments involving many thousands of human-AI persuasive conversations, I will try to answer some key questions related to this phenomenon, including "why?", "what are the boundary conditions?", "is this a magic bullet?, and "can bad actors use this to spread misinformation"? Given that virtually all published work on this topic involves laboratory studies, I will preview ongoing field studies and consider whether, as AI becomes the default epistemic intermediary for most people, we might see ecosystem-level improvements on matters related to belief accuracy, polarization, and shared reality. I conclude that the answer is potentially yes. However, several unresolved bottlenecks could hinder progress. I then close with a call to build the tools and highlight some practical ideas for doing so. 

Speaker Bio

Thomas Costello is an Assistant Professor of Social and Decision Sciences at Carnegie Mellon University, with affiliations at the Human-Computer Interaction Institute and MIT's Sloan School of Management. He studies where political and social beliefs come from, how they differ from person to person—and, ultimately, why they change—using artificial intelligence and the tools of personality, cognitive, clinical, and political science. He has published over 40 research papers in peer-reviewed outlets, including Science, Nature, Journal of Personality and Social Psychology, Psychological Bulletin, and Trends in Cognitive Sciences. His research has been featured in the New York Times, The Atlantic, The Economist, The Wall Street Journal, and Financial Times, and he has appeared on NBC Nightly News, BBC World News, CBC/Radio-Canada, and CNN. Thomas developed DebunkBot.com, a public tool for combatting conspiracy theories with AI. He has received the AAAS Newcomb Cleveland Prize, the APS Rising Star designation, and the Klarman Fellowship from Cornell University, and his research is currently supported by grants from DARPA, Schmidt Sciences, and FAR.AI, among others. In third grade, he hit a stand-up triple, and it's all been downhill from there. More at thcostello.com.