Working Papers
Working Papers
Generative AI (GenAI) tools have been widely integrated into online matching platforms. Although these tools can improve writing quality, they may also lead to content convergence, resulting in increased screening difficulty and reduced matching efficiency. In this paper, we study the Job Post Generator, a GenAI tool launched by a leading freelance platform. We find evidence that overall writing quality improved after the launch of this tool, but the content has become more homogeneous, increasing by 18% from March 2023 to September 2024. Using a propensity score matching difference-in-differences method, we find that, on average, GenAI improves matching efficiency. However, employers who rely heavily on AI-generated content with minimal modifications experience longer matching times and higher payment cost. In contrast, those who actively edit and customize AI-generated job postings achieve significant improvements in matching efficiency, benefiting from shorter matching times, lower payment costs, and higher satisfaction with freelancers' performance. Our paper shows that the homogeneity of job postings caused by AI-generated writing can negatively affect matching efficiency. These findings suggest that online matching platforms offering generative AI tools should encourage users to modify and personalize AI-generated content to achieve better outcomes.
Presentations:
NUS Applied Economics Student Workshop, NUS Behavioral & Experimental Econ Therapy, NUS Graduate Seminar, ISMS Marketing Science Conference (Washington D.C.), 2025 INFORMS International Conference (Singapore), NUS IO Day, NUS Applied Micro SIG Brown Bag, NTU Economics Brown Bag Seminar (scheduled)
This article examines endogenous consumer reviews and their impacts on seller pricing on Steam, an online video game platform. We find that discounts generate a positive post-promotion effect on sales through review accumulation. We build and estimate a structural model that incorporates endogenous consumer reviews and forward-looking game sellers. Counterfactual results demonstrate that without accounting for the effects of reviews, game prices would be 14% higher. We quantify that the reviews have great externalities. The whole system generates millions in surplus for both parties, particularly benefiting new and indie game sellers, while further fostering market diversity.
Presentations:
NUS Applied Economics Student Workshop, APIOC 2023 (Hong Kong), 22nd Annual Conference of IIOC (Rising star session, Boston), SMU Raising Star Conference*, AMES (Hangzhou), 2025 CES North American Conference (Rising star session, Michigan)
Online freelance platforms have expanded access to flexible work, but persistent inequalities remain. This paper examines whether generative artificial intelligence (GenAI) tools can help reduce disparities faced by disadvantaged workers, using high-frequency panel data from a major digital labor platform. Among various worker characteristics, language proficiency emerges as the most strongly impacted dimension. Following the launch of ChatGPT, we find that the gap in earnings and job count between native and non-native English speakers narrowed. Non-native speakers gained access to more complex, higher-paying jobs, experienced lower job displacement, and connected more frequently with employers from high-income countries. While overall entry and activity gaps remained, we observe shifts in participation across job categories. Together, the findings suggest that GenAI can help reduce language-based barriers and expand opportunity in digital labor markets.
Presentations:
NUS Applied Economics Student Workshop, NUS Real Estate Brown Bag*, NUS Marketing Brown Bag, NUS Applied Micro SIG Brown Bag*
This article studies a major ride-hailing service provider in Singapore, which offers subsidies for both consumer side and driver side. We investigate the firm's optimal subsidy strategy in the presence of real-time externalities. By using a zone-time panel of 329,319 observations we estimate both the customer demand and driver labor supply. We also simulate the causal impact of demand and supply conditions on the surge multiplier, customers' waiting time as well as drivers' roaming time. The estimation results suggest that customers are price elastic and place a high value on time; while drivers are even more elastic to price and time than customers. Our counterfactual analysis indicates that considering real-time externalities, instead of subsidizing drivers and consumers at the same time, it will be more profitable for the platform to subsidize only the customer side. However, in contrast, the merely driver side subsidy is likely to generate the highest social welfare.
Presentations:
NUS Applied Economics Student Workshop, NUS Graduate Seminar, IAAE 2024 International Association for Applied Econometrics (Xiamen), 2024 CES China Conference (Hangzhou), NUS Applied Micro SIG Brown Bag, 51st conference of EARIE (Amsterdam)
Note: * presented by a co-author.