"Better Technology, Worse Motivation: Gen AI's Mediocrity Trap" with Yvonne Jie Chen, Jin Li and Zibo Zhao, October 2025, revise and resubmit at Management Science.
Mentioned in: Economist
Abstract: While generative AI (GenAI) promises productive efficiency, it can paradoxically lead to lower-quality work. We conducted an experiment with professional illustrators and found that AI assistance flattens the quality curve—it accelerates initial gains but sharply diminishes the returns on sustained effort. Faced with this, a significant number of professionals made a strategic choice: they sacrificed the final quality to save time. Our finding highlights a critical challenge for GenAI, which can weaken the motivation required for creative excellence and innovation.
"When Less Is More: Managing AI Adoption with Usage Targets" with Jiayi Hou, Jin Li, Fei Pu and Xinjue Yao.
Abstract: Usage mandates can accelerate AI adoption, but they also risk turning adoption into a metric to be gamed. We study a medical-device company that required roughly 5,000 employees to submit 200 AI queries per month, then reduced the target to 100 through a staggered branch rollout. Complete query logs show that the mandate increased first-time adoption, but 32% of queries are either repeated or non-work queries under the 200-query target. Lowering the target reduced query volume by 30%, improved query quality, and increased sales by 7% among sales employees. Our findings suggest that the design of usage targets shapes both how employees use AI and the returns from that use.
"The Value of Local Knowledge in Service Work" with Xiao'ou Liu, Ang Sun and Huili Zhang, September 2025, revise and resubmit at Management Science.
Abstract: Many service tasks appear mechanical, yet automating them can destroy value. We study this puzzle through the lens of local knowledge—the contextual, real-time information that workers gather through direct observation and interaction. A large restaurant chain experienced an unexpected system failure, forcing its outlets to switch from digital ordering to full human service. We leverage this natural experiment to examine how servers use two types of local knowledge that automated systems cannot easily access: real-time customer characteristics and product knowledge that allows credible, tailored recommendations. We find that human servers increase revenue per order by 6.5% compared to digital systems. These gains come from strategic use of local knowledge: they promote signature and high-margin items, and adjust strategies to context–recommending premium dishes to small groups, expanding quantities for larger groups, scaling back upselling during busy periods, and generating greater revenue in tourist areas where customers are less familiar with the menu. Our findings show that decentralized decision-making based on local knowledge creates value that centralized automated systems cannot match.
"Eliciting Perceived Policy Effects: Evidence From a Traineeship Program for College Graduates" with Jessica Pan and Basit Zafar, Working paper, May 2026.
Abstract: We develop and apply a survey-based measurement design for studying perceived policy effects. The design elicits within-person counterfactual beliefs across hypothetical scenarios to recover subjective treatment effects—objects that we argue are intrinsically informative even absent objective validation, since they govern individual take-up decisions, selection, and labor market behavior. We apply this design to a large-scale traineeship program targeting Singaporean college graduates entering the labor market during the pandemic-induced recession, using novel survey data collected six months after graduation and again two years later. Graduates anticipate large negative labor market effects from the recession persisting to age 35, but expect the traineeship program to substantially mitigate these impacts on full-time employment and income, with these perceived treatment effects persisting through the two-year follow-up. Both participants and non-participants perceive non-negative effects from the program’s existence—trainees attribute this primarily to their own participation, while non-trainees attribute it to option value rather than displacement, suggesting little perceived crowd-out. Finally, perceived participation effects elicited ex post are consistent with selection on perceived gains, with observable characteristics having limited predictive power—underscoring the value of direct belief elicitation.
"Organization DNA: Knowledge Transmission through Mentorship Chains" with Jiang Bian, Feiyong Ke, Jin Li and Zhiwei Shang. (draft coming soon!)
Abstract: How do organizations transmit and replicate the tacit knowledge that underlies competitive advantage across successive generations of workers? We investigate this question by studying formal mentorship in a Chinese commercial bank, where incoming loan officers are quasi-randomly assigned to experienced mentors. Drawing on the knowledge-based view and knowledge transfer literature, we develop predictions about how mentor characteristics shape transfer effectiveness, durability, and cross-generational propagation. We find that mentor productive expertise substantially increases mentee performance, with effects persisting beyond the formal mentorship period, consistent with genuine human capital acquisition. Knowledge cascades through mentorship chains: first-generation mentor quality predicts second-generation mentee performance despite no direct interaction between them. Productive expertise and hierarchical rank interact asymmetrically: rank amplifies the effect of expertise, but rank without expertise is associated with lower mentee performance. Mechanism tests rule out direct assistance and document the transmission of specific work practices. We contribute to the knowledge transfer literature by establishing mentor quality as a micro-foundation of transfer effectiveness, demonstrating that mentorship can function as a self-replicating organizational knowledge mechanism.