光华讲坛——社会名流与企业家论坛第7028期

主题：AI Deployment as Incentive Design: Commitment and Human Effort in AI-Assisted Decisions作为激励设计的人工智能部署：AI辅助决策中的承诺与人力付出

主讲人：东北财经大学金融学院 凤迪副教授

主持人：经济学院 程瑶副教授

时间：11 月 13 日 14:00-15:30

地点：柳林校区诚正楼1600会议室

主办单位：经济学院 科研处

主讲人简介：

凤迪，东北财经大学金融学院副教授、瑞士洛桑商学院博士。担任Econometrica, Games and Economic Behavior 等期刊及EC、WINE等国际顶尖计算机会议的匿名审稿人。主要研究领域为计算机科学与经济学的交叉学科，具体包括多智能体系统、拍卖与机制设计等。主持国家级项目1项、省部级2项。在Games and Economic Behavior，Journal of Mathematical Economics 等国际顶尖期刊/会议发表论文10余篇。

内容提要：

How should a principal deploy and use AI to assist the assessment by a human agent? We study a principal-agent model of AI-assisted decision making in which the principal decides whether to deploy an AI system and commits to a rule mapping an AI recommendation and an agent report into a binary decision. The agent shares the principal’s objective but privately observes his expertise and must exert costly effort to generate an independent assessment. We show that, first, conditional on AI deployment, the unique optimal rule is full deference to the agent when AI and agent disagree. Full deference maximizes the incentive of an agent who can outperform the AI to produce independent judgment, while an agent whose expertise falls below AI accuracy optimally remain inactive. Second, the value of human assessment is justified only by the mass of highly matched experts, while improving the average quality of agents may add little value. Third, the principal may optimally withhold an informative AI system. A more accurate AI directly improves the baseline recommendation but also reduces the expertise gap that motivates human effort. The results are robust to random effort costs, continuous expertise distributions, and an AI-augmented assessment technology in which agents use effort to correct AI mistakes. The analysis highlights that AI-assisted decision systems should be designed not only to aggregate information, but also to preserve incentives to create it.

委托人应如何部署和使用人工智能（AI）来辅助人类代理人的判断？本次讲座将研究一个AI辅助决策中的委托—代理模型。在该模型中，委托人决定是否部署AI系统，并承诺采用一项规则，将AI的建议与代理人的报告映射为一个二元决策。代理人与委托人的目标一致，但其专业能力为私人信息，并且必须付出有成本的努力，才能形成独立判断。我们得到三个主要结论。第一，在部署AI的条件下，唯一的最优规则是：当AI与代理人的判断不一致时，完全听从代理人的判断。完全服从代理人可以最大程度激励那些专业能力高于AI准确率的代理人形成独立判断；而专业能力低于AI准确率的代理人则最优地选择不投入努力。第二，人类判断的价值只取决于高专业能力代理人的占比，而提高代理人的平均能力水平可能带来的价值十分有限。第三，委托人有时会最优地选择不向代理人提供一个具有信息价值的AI系统。更高准确率的AI一方面可以直接改善基准建议的质量，但另一方面也会缩小AI与人类专业能力之间的差距，从而削弱代理人付出努力的激励。上述结果在努力成本具有随机性、专业能力服从连续分布，以及采用AI增强型判断技术的情形下仍然成立。在后一种情形中，代理人通过付出努力来识别并纠正AI的错误。本次讲座表明，AI辅助决策系统的设计不应只关注如何整合已有信息，还应关注如何维持人类主动创造信息的激励。