主题: Forget Me Not: Algorithmic Inference and Machine Unlearning

题目：勿忘我：算法推测和机器遗忘

时间: May 14^th, 2026 (Thursday)  9:45 am-10:45 am Beijing Time

地点:  第五教学楼5306

主讲人：Xuanqi Chen, PhD Candidate, School of Management, Xi’an Jiaotong University & The Hong Kong Polytechnic University

Bio:

Mr. Xuanqi Chen is a PhD candidate at both Xi’an Jiaotong University and The Hong Kong Polytechnic University. His research interests include data privacy, AI economics, platform economics, and the IS-marketing interface. His work has been published in MIS Quarterly and is under revision at Management Science.

照片：

Abstract: 

The GDPR’s “right to be forgotten” (RTBF) grants customers the ability to request erasure of their personal data. However, traditional data-level RTBF fails to address algorithmic inference: even after data deletion, trained models retain learned patterns that enable ongoing prediction and price discrimination; that is, a “forget-me-not” outcome. To achieve genuine forgetting, regulators have proposed extending RTBF to the algorithm level through machine unlearning, which removes the influence of erased data from trained models. Despite rapid technical advances, the economic implications of this extended right remain unexplored. This paper develops a two-period analytical model to examine how algorithm-level RTBF, enabled by machine unlearning, affects customers and the platform. We characterize equilibrium behaviors and uncover several counterintuitive findings. First, customers’ RTBF exercise exhibits nonmonotonic patterns regardless of whether unlearning is available. Second, mandating machine unlearning can paradoxically enhance algorithmic inference via an opt-in effect, inducing more customers to provide data for algorithm training. Third, machine unlearning can intensify price discrimination via an algorithm-bypass effect, prompting the platform to bypass the algorithm and charge high prices, which leads to more severe price discrimination. Fourth, the platform may actually increase its profit under extended RTBF. Most critically for policymakers, algorithm-level RTBF can ultimately reduce overall customer surplus, particularly in highly personalized service environments. These findings reveal that granting individuals greater control through machine unlearning can subvert its own purpose.