Title:Double Jeopardy: How Mutual Adaptation Compromises Human-LLM Collaboration in Financial Robo-Advisory

时间：4月10日周五上午10 :00-11:30

地点：承泽园333教室

Abstract:

This study examines how mutual adaptation between humans and advanced AI can paradoxically lead to suboptimal user outcomes—a phenomenon we term “double jeopardy.” We introduce an Adaptation-Inertia Duality (AID) framework that identifies two forces: user inertia, arising from legacy habits incompatible with new AI systems, and AI misalignment, occurring when AI systems reinforce existing suboptimal user preferences. This dynamic leads to a paradox of experience: incumbent users, despite their richer experience, may benefit significantly less from an adaptive AI system than new users. We test this theory using millions of conversation-level data from a large-scale randomized experiment at a leading Chinese FinTech firm, during which it upgraded its legacy robo-advisor to an LLM-powered system. While the upgrade boosted average annualized returns by 2.58%, the gains were highly skewed: new users enjoyed a 2.88% increase, whereas incumbent users gained only 1.85%. We develop a theoretical framework to prescribe why and under what conditions this paradox may occur, illuminating the dual role of legacy experience—both an asset for personalization and a potential source of inertia and misalignment. This offers significant insights into the practical deployment of generative AI in high-stakes domains, revealing little-known design considerations crucial for effective AI-assisted human decision-making.

Bio

Yidi Liu is an Assistant Professor of Information Systems at School of Management and Economics, the Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen). He received his Ph.D. in Information Systems from the City University of Hong Kong, and a Bachelor’s degree in Logistics Management from Huazhong University of Science and Technology. His research interest covers digital economy, artificial intelligence, fintech, and network analysis. His research appears on Information Systems Research, Journal of AIS, among others.