Why we disagree: decoding the debate between AI “Doomers” and “Boomers” using an ensemble of Large Language Models

Research Seminar

Management & Human Resources

Speaker: Phanish Puranam

INSEAD

Bernard Ramanantsoa room 

Abstract: Using a multi-agent AI architecture - a team of Large Language Model (LLMs) – we develop a procedure to reliably parse arguments into their factual, lexical, causal and evaluative (moral) premises. We map the chains of reasoning for each individual, and compare these chains to pinpoint premises or logical operations that are at the root of disagreements. We apply this methodology to transcripts from the Lex Fridman podcast, in order to develop a better understanding of disagreements between the interviewees regarding the nature and extent of risks pose by Artificial Intelligence technologies. Our approach also indicates “Paths to Possible Agreement” between any randomly selected pair of interviewees based on the detected divergence in their reasoning chains. We draw implications for the applications of AI technologies to improve the quality of deliberation in organizations and other settings.