Bias and Trust in LLM-Driven Screening Automation:Experimental Insights from FinTech Lending

主讲嘉宾：蔡熙乾

报告人简介:

蔡熙乾，厦门大学经济学院和王亚南经济研究院教授、博士生导师，厦门大学国际合作与交流处副处长。主要研究领域为行为经济学、劳动经济学与中国经济。主持国家自然科学基金优秀青年基金、面上项目、青年基金等项目。主要论文发表在《中国社会科学》，《世界经济》，Review of Economics and Statistics, Management Science, Journal of Development Economics, Psychological Science等国内外权威期刊。多项研究成果获福建省社会科学优秀成果奖，山东省社会科学优秀成果奖。担任期刊China & World Economy副主编， Fundamental Research青年编委。报告论文简介:

The rapid rise of artificial intelligence (AI) has sparked discussions of its adoption in the financial areas, including FinTech lending; however, concerns about fairness in algorithmic decision-making has kept skepticism alive. In this study, we investigate the performance, biases, and trust associated with large language models (LLMs)-GPT-4 and Claude 3 Opus--in peer-to-peer lending decisions. Using 1,095 human participants and 12 task sets, we find that LLMs consistently outperform humans in default judgments, even without prior training on Chinese-language data and when analyzing unstructured information. Nonetheless, both LLMs and humans exhibit a mix of taste-based and statisticaldiscrimination,with LLMs tending to lower thresholds for women but imposing more stringent lending terms on them. In human-LLM hybrid settings, GPT-4’s inputs improve human accuracy, yet additional human input neither enhances human decisions nor benefits the LLM. Notably, human participants display considerable algorithm aversion, which declines as task complexity rises though women initially exhibit stronger aversion than men. These results underscore that performance gains alone do not guarantee effective AI integration: biases within LLMs and persistent human skepticism must be addressed to fully realize AI’s benefits in FinTech lending.

活动时间: 2025年4月29日 星期二 上午9:30

活动地点: 金融学院会议室(劝学楼345小会议室)

主办单位: 金融学院、实验经济学实验室

欢迎广大教师、同学参加!