Kai Zhu

Does AI democratize knowledge production or amplify existing disparities? We investigate this tension by studying the deployment of neural machine translation across more than 100 Wikipedia language communities. Leveraging rich, fine-grained data and exogenous variation from a natural experiment, we uncover the “AI Democratization Paradox,” where the technology simultaneously drives democratizing and concentrating forces. AI lowered barriers, leading to a substantial increase in content creation across diverse target languages without sacrificing quality or readership. However, the benefits were concentrated: well-resourced communities captured disproportionate gains—3–4 times larger than mid-tier editions. While editors actively leveraged AI to address representation gaps, translating female biographies at twice the expected rate, structural constraints limited the impact in high-need areas. We conclude that technological solutions alone cannot overcome structural inequalities; AI’s distributional impact is contingent on the interplay between technological capabilities and existing social structures.

This study investigates the consequences of monetizing a marketplace for product promotion within a digital platform, specifically the Giveaways program on Goodreads.com. Using a natural experiment and fine-grained platform data from 2016 to 2020, we examine how introducing a fixed entry cost for content creators affected both supply and demand in this two-sided market. Our findings reveal significant shifts in marketplace dynamics after monetization: (1) a substantial decrease in overall program participation, particularly among indie publishers and self-published authors, leading to increased market concentration; (2) reduced genre diversity, with popular genres becoming more dominant at the expense of niche categories; and (3) intensified promotional effects, characterized by higher review volume but lower average ratings for participating books. Analysis of review text suggests an increase in consumer-book mismatches as a potential mechanism driving this outcome. Our study advances platform economics by demonstrating how entry costs reshape marketplace composition and affect value creation in two-sided markets. These findings inform platform design and policy, particularly for markets with horizontally differentiated products and heterogeneous consumer preferences.

Social interactions are becoming increasingly central to the day-to-day operations of digital platforms. In this study, we investigate how negative peer feedback shapes user behavior in an online review platform. Leveraging fine-grained digital trace data of users throughout their tenure on the platform, we find interesting insights regarding the impact of negative peer feedback. First, negative feedback improves user retention on the platform relative to users who receive no feedback at all. Moreover, for users who are retained and write their next reviews, we find that negative feedback improves review frequency and quality. These novel findings demonstrate that, contrary to the conventional wisdom, negative peer feedback can benefit the platform’s welfare as it increases user engagement and retention, and that no feedback can be a worse alternative.

Throughout the COVID-19 crisis, as we confronted questions about social distancing, mask wearing, and vaccines, public safety experts warned that the consequences of a misinformed population would be particularly dire due to the serious nature of the threat and necessity of severe collective action to keep the population safe. Thus, the media and the political elites (e.g., President of the United States) who possess the power to set the information agenda around COVID-19 bear a huge responsibility for the general welfare. Through automated text analysis of complete transcripts of national cable, network, and local news, we explore their narratives surrounding the COVID-19 pandemic and we characterize the differences in which topics were covered and how they were covered by various media sources. Our analysis reveals polarized narratives around blame, racial and economic disparities, and scientific conclusions about COVID-19. Among the various agenda-setting mechanisms available to the president is daily press conferences, which provide a unique opportunity to leverage public exposure, accelerated by the state of crisis. We found both resonance and contrast between the narratives of media and President press conferences. However, as online search data revealed, public information-seeking behavior resembles media coverage more than the President’s messages.

Open collaboration platforms have fundamentally changed the way that knowledge is produced, disseminated, and consumed. In these systems, contributions arise organically with little to no central governance. Although such decentralization provides many benefits, a lack of broad oversight and coordination can leave questions of information poverty and skewness to the mercy of the system’s natural dynamics. Unfortunately, we still lack a basic understanding of the dynamics at play in these systems and specifically, how contribution and attention interact and propagate through information networks. We leverage a large-scale natural experiment to study how exogenous content contributions to Wikipedia articles affect the attention that they attract and how that attention spills over to other articles in the network. Results reveal that exogenously added content leads to significant, substantial, and long-term increases in both content consumption and subsequent contributions. Furthermore, we find significant attention spillover to downstream hyperlinked articles. Through both analytical estimation and empirically informed simulation, we evaluate policies to harness this attention contagion to address the problem of information poverty and skewness. We find that harnessing attention contagion can lead to as much as a twofold increase in the total attention flow to clusters of disadvantaged articles. Our findings have important policy implications for open collaboration platforms and information networks.

AI search retrieves web information to ground generated answers. We study this new information intermediary using U.S. desktop clickstream data spanning ChatGPT Search’s rollout. Observing use alongside referrals, we find recorded referrals for 5.2% of ChatGPT conversation sessions versus 31.1% of Google queries. Ad-supported sites account for smaller shares of ChatGPT than Google referrals, while reference and technical sites account for larger shares. Households search Google, Bing, and Yahoo less after adopting ChatGPT or gaining ChatGPT Search access. Together, these findings suggest that AI search may reshape open-web traffic through its own referrals and changes in conventional-search use.

AI systems rely on web content, but access is increasingly governed by voluntary crawler restrictions. We study this access regime among news and media publishers, where content quality matters for factual and civic information. We provide large-scale evidence that publisher access is sharply selective: high-factual outlets restrict AI crawlers at 47%, versus 8% among low-factual outlets. Related composition results show that misinformation sources and outlets at the political extremes become relatively more accessible. A further finding is use-specific selectivity: higher-quality publishers are more likely to restrict training while preserving search-like access, consistent with training offering weaker attribution and referral benefits than search. The evidence shows that publishers are not rejecting AI access wholesale—they distinguish uses. Our findings imply that AI data governance should move beyond binary opt-out toward use-specific access design.

Digital platforms increasingly allow audiences to support creators directly, yet little is known about how expanding access to this support changes creator behavior. We study Twitch’s 2017 Affiliate Program, which extended channel subscriptions and virtual donations to smaller creators. Using a time-shifted difference-in-differences design, we find that the program increased weekly streaming hours by 42%. Relative responses were strongest among smaller creators, broadening the distribution of content production. Per-stream viewership did not detectably decline despite the increase in supply. Monetization access also changed creators’ content strategies. Together, these findings identify expanded access to direct audience support as a platform-design lever for cultivating the long tail.

This paper introduces a novel approach using neural embeddings to extract marketing insights from user-product interactions. Our method adapts the distributional hypothesis to consumer behavior, representing users and products as vectors in a shared latent space. This framework allows for the direct measurement of individual preferences from revealed-preference data. Using the Goodreads Book Graph Dataset, we demonstrate that this method can uncover hidden patterns in consumer tastes and provide new micro-level evidence for canonical marketing theories. This research offers valuable tools for academics studying market dynamics and practitioners developing marketing strategies for experiential goods.