Research

Abstract

Artificial intelligence (AI) synthesizes existing knowledge into refined knowledge, which the human brain then recombines to generate new ideas. By lowering the cognitive burden of information processing, AI can accelerate discovery, but it may also reduce knowledge spillovers by filtering out information that later proves valuable. We show that research productivity varies nonmonotonically with AI efficiency: AI boosts innovation when knowledge is very scarce or very abundant, yet may create a mid-knowledge level AI trap where faster AI progress slows down innovation and lowers productivity. In our endogenous growth model, faster AI raises long-run growth, but its effect on R&D labor share is ambiguous. 

Abstract

A notable phenomenon in the digital economy is the emergence of digital villages: e-commerce platforms support and subsidize suppliers in less developed rural areas to manufacture and sell products online. We argue that, despite platforms' philanthropic claims, these actions are strategically designed to enhance profitability. By providing early subsidies to young sellers, platforms incentivize entry and reduce learning costs, later recouping these investments as sellers gain experience and increase sales. Using a dynamic model of two-sided markets, we analyze the intertemporal and cross-side pricing strategies of a platform with market power. Our findings indicate that sellers' network externalities and learning-by-doing effects reinforce each other, motivating the platform to subsidize them. This study bridges two typically distinct areas of the economy: global online platforms and less developed rural regions. 

Abstract

Many online retail platforms, such as Amazon and JD.com, have recently begun providing sellers with proprietary consumer data. In this paper, we investigate how the sharing of such data can incentivize seller collusion through personalized pricing. We find that the effect of data sharing on collusion sustainability and profitability depends on the mode operated by the platform. When a platform acts as both a host and seller, sharing data leads to more collusive outcomes. However, when a platform only intermediates purchases, sharing data hinders collusion. Our results suggest that imposing a ban on data sharing may be ineffective and harmful to consumer welfare. 

Abstract

Platforms often enforce pricing policies onto sellers such as the minimum margin agreement (MMA). MMAs require that platforms receive a guaranteed profit-margin on seller goods. If the margin is not met then any difference between the actual margin is taken from the seller. I develop a model where a seller sells through a direct channel and serves as a supplier for a separate platform. Under MMAs, a platform can threaten the seller to flood the market demand by pricing low and capturing a guaranteed profit. I show that MMAs potentially lead to an increase in platform facilitated purchases and inflate both direct and intermediated prices. Furthermore, I find that MMAs may be more prevalent when the seller faces competition and platforms can steer buyers towards specific goods.

Abstract

Bernheim and Whinston (1990) famously show that contact across multiple markets may facilitate collusion given that firms or markets are not identical. As pricing decisions are increasingly made by algorithms, antitrust authorities are concerned that algorithms may autonomously collude on supracompetitive prices. In this paper, we test how multimarket contact affects the behavior of pricing facilitated by reinforcement learning algorithms via Q-learning. 

Abstract

Many online platforms, most notably search engines, incorporate some aspect of search. Recently, these platforms have begun incorporating artificial intelligence (AI) search features that scrape content from online publishers and synthesize it into AI-generated content. In this paper, we develop a theoretical model where a platform hosts two publishers who choose the quality of content to produce. The platform has control over two parameters that affect its AI-generated content: (i) a “mixing rule” that determines how much each publisher’s content is used and (ii) a “quality-conversion ability” parameter that represents the extent to which the platform can effectively scrape and create content. We find that AI search features improves user-to-publisher match quality by informing users of their preferences which has an ambiguous effect on content quality. Improving the AI's quality-conversion ability reduces the number of users who directly accesses each publisher’s site, thereby disincentivizing publishers to invest in high quality content. This indirectly reduces the quality of the AI-generated content which may hurt the platform. Therefore, we show that it is optimal for the search platform to degrade the quality of its AI search feature to temper this indirect feedback effect. We also find that enforcing a policy restricting the platform's quality-conversion ability, either through scraping or copyright policies, is socially-optimal. Our analysis includes several extensions such as allowing publishers to block AI scraping and accounting for platform competition.